diff --git a/.claude/skills/ingest-run/SKILL.md b/.claude/skills/ingest-run/SKILL.md
new file mode 100644
index 000000000..c91ae2379
--- /dev/null
+++ b/.claude/skills/ingest-run/SKILL.md
@@ -0,0 +1,117 @@
+---
+name: ingest-run
+description: Ingest one vngcloud/InferenceX GitHub Actions benchmark run into this fork's self-hosted prod DB, over SSH into the deploy host. Use whenever asked to "ingest this run", given a vngcloud/InferenceX actions/runs/<id> URL, or asked to get benchmark data from a workflow run into the dashboard. NOT for the upstream SemiAnalysisAI/InferenceX-app Neon-DB flow (a different agent/env covers that).
+---
+
+# Ingest one benchmark run (this fork, self-hosted host)
+
+This fork does **not** use Neon. Prod is a self-hosted docker-compose stack on a
+single runner host, with its own dockerized Postgres. That changes the ingest
+mechanics vs. the upstream project's docs — follow this skill, not generic
+`admin:db:ingest:run` examples you find elsewhere in the repo/docs.
+
+## Environment (fixed facts about this deployment)
+
+- **Deploy host**: `ssh -p 234 hoanq333@61.28.228.19`
+- **Prod compose project**: `/opt/docker-compose/docker-compose.yml`, services
+  `inferencex-app-1`, `inferencex-postgres-1` (local Postgres, NOT Neon), `inferencex-nginx-1`,
+  `inferencex-certbot-1`.
+- **Source repo for runs**: `vngcloud/InferenceX` (not `SemiAnalysisAI/InferenceX`).
+- **`DATABASE_WRITE_URL`** on this host points at the local `postgres` compose service.
+  It does **not** speak TLS — every `bun run admin:db:*` command below **MUST** pass
+  `--no-ssl`, or you get a confusing `Client network socket disconnected before secure TLS
+connection was established` / `ECONNRESET` error that looks like network flakiness but
+  isn't. `deploy.yml`'s own migration step already does this (`admin:db:migrate --yes --no-ssl`).
+- Dev stack (`inferencex-dev`, port `:8080`) is a separate, isolated compose project on the
+  same host — this skill is about **prod**.
+
+## Step 0 — confirm the run is actually finished
+
+```bash
+gh api repos/vngcloud/InferenceX/actions/runs/<RUN_ID> --jq '{name,status,conclusion}'
+```
+
+If `status` isn't `completed`, wait — do not attempt to ingest a partial run. Re-check later
+(a `ScheduleWakeup`/poll loop, not tight retries).
+
+## Step 1 — check whether the model/GPU/precision/framework is already registered
+
+Read the run's job names and artifact names (`gh run view <RUN_ID> --repo vngcloud/InferenceX`,
+not `--log`) and check whether the model prefix, hardware key, precision, and framework already
+resolve in the codebase (`packages/constants/src/models.ts` `DB_MODEL_TO_DISPLAY`,
+`packages/constants/src/gpu-keys.ts` `HW_REGISTRY`, `packages/db/src/etl/normalizers.ts`
+`PRECISION_ALIASES`/`MODEL_TO_KEY`). **If anything is new, follow
+`docs/adding-entities.md` first** — ingesting before the mapping lands means the DB silently
+skips those rows (no error, no partial write — the ETL just drops what it can't resolve). Commit,
+push to `master`, and wait for `deploy.yml` to finish (`gh run watch <run-id>`) before proceeding —
+the new mapping has to actually be in the container's code for ingest to pick it up.
+
+## Step 2 — download artifacts once, persistently, inside the container
+
+The artifact-download step and the DB write step both go over flaky links from this host
+(GitHub's artifact CDN, and TLS-handshake resets from the app container to the local Postgres
+under load). Don't repeat `--download` end-to-end on every retry — it wastes time re-fetching
+30+ artifacts just to retry the DB half. Split it:
+
+```bash
+ssh -p 234 hoanq333@61.28.228.19 \
+  "cd /opt/docker-compose && docker compose exec -T app sh -c \
+   'mkdir -p /tmp/ingest-<RUN_ID> && cd /tmp/ingest-<RUN_ID> && gh run download <RUN_ID> -R vngcloud/InferenceX && ls | wc -l'"
+```
+
+Confirm the artifact count matches what `gh run view` listed before moving on.
+
+## Step 3 — ingest in CI mode against the local artifacts (retry this, not the download)
+
+```bash
+ssh -p 234 hoanq333@61.28.228.19 \
+  "cd /opt/docker-compose && docker compose exec -T \
+   -e INGEST_RUN_ID=<RUN_ID> -e INGEST_RUN_ATTEMPT=1 \
+   -e INGEST_ARTIFACTS_PATH=/tmp/ingest-<RUN_ID> -e INGEST_REPO=vngcloud/InferenceX \
+   -e UNMAPPED_ENTITIES_OUTPUT=/tmp/unmapped.json \
+   app bun run admin:db:ingest:ci --no-ssl"
+```
+
+If it fails with the TLS/ECONNRESET error above, that's the local-Postgres connection blip, not
+a real problem with the data — just retry this exact command a few times. Do NOT switch to
+`admin:db:ingest:run <id> vngcloud/InferenceX` (the one-shot download+ingest command) as your
+retry loop; that re-downloads everything each time.
+
+Note the positional-arg gotcha if you ever do use the one-shot form: it's
+`admin:db:ingest:run <RUN_ID> <owner/repo>` — **positional**, not `--repo <owner/repo>`. Passing
+`--repo` makes the script try to parse `--repo` itself as a run ID and fail immediately.
+
+Watch the summary block at the end for `+N new, M duplicate` counts per data type (benchmark
+results, run stats, eval results, eval samples, changelog entries). `0 new, N duplicate` for a
+run you know was already ingested is expected, not an error.
+
+## Step 4 — overrides, verify, invalidate
+
+```bash
+ssh -p 234 hoanq333@61.28.228.19 "cd /opt/docker-compose && docker compose exec -T app bun run admin:db:apply-overrides --yes --no-ssl"
+ssh -p 234 hoanq333@61.28.228.19 "cd /opt/docker-compose && docker compose exec -T app bun run admin:db:verify --no-ssl"
+ssh -p 234 hoanq333@61.28.228.19 "cd /opt/docker-compose && docker compose exec -T app bun run admin:cache:invalidate http://localhost:3000"
+```
+
+`apply-overrides` reporting everything as "not in DB, skipping" is normal — most registry entries
+in `run-overrides.ts` target other runs. `db:verify`'s "Sample Rows" section is the fastest way to
+eyeball that the new config/benchmark_results rows look right (correct model key, precision,
+hardware, throughput numbers in a sane range).
+
+## Cleanup
+
+```bash
+ssh -p 234 hoanq333@61.28.228.19 "cd /opt/docker-compose && docker compose exec -T app rm -rf /tmp/ingest-<RUN_ID> /tmp/unmapped.json"
+```
+
+## Don't forget
+
+- If Step 1 required a code change, that same PR/commit should have updated
+  `packages/app/src/lib/api-route-catalog.ts`'s digest for `models.ts` (there's a vitest guard
+  that fails the build otherwise — `bun run test:unit` will tell you which digest to recompute:
+  `python3 -c "import hashlib; print(hashlib.sha256(open('<file>','rb').read()).hexdigest())"`).
+- If the model/GPU should be selectable in the dashboard UI (not just stored), that's a
+  **separate, optional** follow-up: `Model`/GPU registry entries in
+  `packages/app/src/lib/data-mappings.ts`, `compare-slug.ts`, `compare-ssr.ts` — see
+  `docs/adding-entities.md`. Ask the user whether they want that now or the DB write is enough
+  for this pass.
diff --git a/.claude/skills/review-zh-copy/SKILL.md b/.claude/skills/review-zh-copy/SKILL.md
new file mode 100644
index 000000000..80a2473d9
--- /dev/null
+++ b/.claude/skills/review-zh-copy/SKILL.md
@@ -0,0 +1,113 @@
+---
+name: review-zh-copy
+description: Use when reviewing PRs or diffs that add or modify user-visible Simplified Chinese in InferenceX, including refactors and files whose names do not contain zh.
+---
+
+# Review Simplified Chinese copy
+
+## Operating mode: advisory
+
+Every finding produced by this skill is non-blocking. Do not label a Chinese-copy finding
+`BLOCKING`, submit a request-changes review for it, or otherwise hold another contributor's merge.
+Do not mention `@edwingao28` for a clean review, routine coverage, or an ordinary wording or
+naturalness suggestion. Report an ordinary finding without a CC. Mention `@edwingao28` only for a
+high-confidence semantic or factual error, changed attribution or speaker voice, or high-impact
+ambiguity that cannot be resolved from the source and context. Label that finding
+`Needs Chinese maintainer confirmation` and ask for review. It remains a suggestion while the
+Chinese maintainer decides whether any follow-up is needed.
+
+The Chinese maintainer makes the final wording decision and manually reviews every changed
+Chinese passage.
+
+## Editorial source of truth
+
+Read [docs/chinese-copy.md](../../../docs/chinese-copy.md) completely before reviewing. Apply its
+audience, surface-specific register, context-aware terminology, rewrite method, and checklist.
+
+Semantic fidelity and natural Chinese are independent gates. Every changed passage must pass
+both:
+
+- **Fidelity:** facts, scope, modality, metrics, numbers, links, product names, and attribution
+  retain the English source's meaning.
+- **Naturalness:** the result reads as original Chinese technical writing in the register of its
+  surface, without preserving English clause order or noun stacking.
+
+A fluent mistranslation fails fidelity. An accurate sentence shaped like English fails
+naturalness.
+
+## Find the complete review scope
+
+Review every added or modified user-visible Chinese passage in the complete diff against its
+actual base. Do not infer scope from filenames. Chinese may appear in:
+
+- `zh:` dictionaries inside otherwise English TypeScript or TSX files;
+- locale-aware ternaries and registry labels;
+- `/zh` routes, Chinese MDX, JSON, metadata, schema.org content, alt text, and captions;
+- tooltips, modals, controls, option labels, empty/loading/error states, and feature-gated UI.
+
+Include the full UI element or paragraph, not only the changed word or source line. When an
+English/Chinese dictionary entry changes, inspect both siblings even if only one side is in the
+diff. Do not report awkward pre-existing Chinese that the diff did not touch; record it separately
+if it matters to a later rollout.
+
+Exclude identifiers, test data that is not rendered, DB-stored benchmark content, logs, and code
+syntax unless the changed text is presented to a user.
+
+## Build context before judging
+
+For each passage, collect:
+
+1. the complete English source;
+2. the route, component, and UI surface;
+3. the surrounding heading, control, paragraph, or attributed quotation;
+4. the intended meaning in plain language;
+5. facts, qualifications, modality, metrics, links, names, and attribution that cannot change;
+6. rendered desktop and mobile context when width, hierarchy, or interaction affects wording.
+
+If no English source exists, say so. Infer intent from adjacent UI and product behavior, and lower
+confidence rather than inventing a source claim.
+
+## Review workflow
+
+For every changed passage:
+
+1. State the source intent without copying the English sentence structure.
+2. Evaluate fidelity. Reject additions, omissions, changed scope, stronger certainty, altered
+   metric definitions, or a changed speaker voice.
+3. Hide the old Chinese and draft from the stated intent.
+4. Evaluate naturalness without looking at the English. Check information order, sentence
+   structure, register, and established Chinese ML infrastructure usage.
+5. Compare the proposed rewrite with the source again.
+6. Check the complete rendered element or paragraph. A locally natural word can still be wrong in
+   its button, tooltip, metadata, marketing, technical-prose, or quotation context.
+
+Do not use fixed substitutions for contextual questions. In particular:
+
+- sentence length alone is not evidence of translationese;
+- technical English is not wrong merely because a Chinese equivalent exists;
+- one concept may need different wording in a UI action, prose, a unit, or an attributed quote;
+- established English technical terms and phrases that Chinese ML infrastructure engineers
+  normally use in English should remain English in first-party UI and technical prose; judge this
+  from real industry usage and the surface, not from a closed list. `warmup`, `seed`, and `offload`
+  are examples, not the whole category. Chinese-first explanations such as `预热（warmup）` remain
+  valid for a broader audience, and attributed quotations preserve the speaker's wording;
+- `您` is normally omitted in controls and tooltips but may be intentional in a respectful CTA;
+- #819 and other bulk-rewrite outputs are not automatic editorial ground truth; only their
+  separately verified mechanical cases may seed deterministic CI fixtures.
+
+## Report high-signal findings
+
+For each finding provide:
+
+- `file:line` and route or surface;
+- the exact English source and a one-sentence statement of its intent;
+- which gate failed: fidelity, naturalness, or both;
+- the specific defect and its user impact;
+- one complete suggested replacement, not an isolated word swap;
+- confidence and advisory severity.
+
+Do not report only “this reads awkwardly.” Stay silent when the existing wording is accurate,
+natural, and appropriate for its surface, even if another wording is also possible.
+
+End with coverage: list the changed Chinese surfaces reviewed and any surface that could not be
+rendered or whose English source could not be identified.
diff --git a/.claude/skills/write-inferencex-blog/iso_interactivity.py b/.claude/skills/write-inferencex-blog/iso_interactivity.py
index 64058e8ba..fef08a88f 100644
--- a/.claude/skills/write-inferencex-blog/iso_interactivity.py
+++ b/.claude/skills/write-inferencex-blog/iso_interactivity.py
@@ -7,7 +7,8 @@
 packages/app/src/components/inference/hooks/useInterpolatedTrendData.ts.
 
 Pipeline:
-  1. Build the upper-left Pareto frontier on (interactivity, throughput).
+  1. Build the upper-left Pareto frontier on (interactivity, throughput). For a
+     proportional metric, use the total/output/input throughput it scales.
   2. Sort the frontier by interactivity ascending.
   3. Build monotone cubic Hermite slopes (Steffen 1990 method, same as
      d3.curveMonotoneX) over the frontier knots for the metric of interest.
@@ -18,12 +19,18 @@
   6. Clamp the interpolated metric value to [min(ys), max(ys)] of the
      frontier to prevent cubic-spline overshoots above/below the data.
 
-Reciprocal metrics ($/M tok, J/token) are a special case: they are a per-chip
-constant divided by a throughput, so they are NOT splined independently. Two
-independent splines need not preserve `metric * throughput = constant` between
+Reciprocal metrics ($/M tok in blog tables, J/token) are a special case: they are
+a per-chip constant divided by a throughput, so they are NOT splined independently.
+Two independent splines need not preserve `metric * throughput = constant` between
 measured knots. Pass `reciprocal_of='throughput'` (or the matching output/input
-throughput key) to spline that throughput and re-derive the metric, which is what
-the dashboard does. See docs/tco-calculator.md for a reproducible measurement.
+throughput key) to spline that throughput and re-derive the metric. The live
+inference charts also expose tokens/$ and normalized token revenue as separate
+metrics; both are directly proportional to throughput. OpenRouter-priced revenue
+is proportional only when the input/output mix stays constant; otherwise spline
+the already-priced metric directly. Pass
+`proportional_to='throughput'` (or the matching output/input throughput key) so
+they use the same Pareto knots and spline as that throughput before applying the
+recovered multiplier. See docs/tco-calculator.md for a reproducible measurement.
 
 Usage as a module:
     from iso_interactivity import interpolate_metric, pareto_front_upper_left
@@ -182,6 +189,25 @@ def recover_reciprocal_numerator(
     return numerator
 
 
+def recover_proportional_multiplier(
+    values: list[Optional[float]], throughputs: list[Optional[float]]
+) -> Optional[float]:
+    """Recover the constant `k` of a metric defined as `throughput * k`."""
+    relative_tolerance = 1e-3
+    multiplier: Optional[float] = None
+    for value, throughput in zip(values, throughputs):
+        if value is None or throughput is None:
+            continue
+        if not value > 0 or not throughput > 0:
+            continue
+        candidate = value / throughput
+        if multiplier is None:
+            multiplier = candidate
+        elif abs(candidate - multiplier) > abs(multiplier) * relative_tolerance:
+            return None
+    return multiplier
+
+
 def interpolate_metric(
     points: list[dict],
     target_iv: float,
@@ -189,6 +215,7 @@ def interpolate_metric(
     iv_key: str = 'interactivity',
     tput_key: str = 'throughput',
     reciprocal_of: Optional[str] = None,
+    proportional_to: Optional[str] = None,
 ) -> Optional[float]:
     """Interpolate `metric_key` at `target_iv` using the chart's algorithm.
 
@@ -196,16 +223,23 @@ def interpolate_metric(
     the chart code does not extrapolate. Blog tables should render this as
     `_unreachable_` in that row.
 
-    The Pareto frontier is ALWAYS built on (interactivity, throughput) regardless
-    of which metric you're interpolating — this matches the chart, where the
-    frontier is defined by the upper-left throughput envelope and other metrics
-    (cost, energy, TPOT, ...) are derived values at frontier points.
+    The Pareto frontier is built on (interactivity, throughput). For tokens/$ or
+    normalized token revenue (and revenue with a constant token mix),
+    `proportional_to` names the total/output/input
+    throughput it scales; that throughput also supplies the frontier knots so
+    both curves stay aligned.
 
     `reciprocal_of` names the throughput key a metric divides, for metrics of the
     form `constant / throughput` ($/M tok, J/token). Set it for those metrics:
     that throughput is splined and the metric re-derived, matching the dashboard.
     Leave it None for metrics splined directly (throughput itself, tok/s/MW,
     latency, and measured energy — whose numerator varies per point).
+
+    `proportional_to` names the throughput key for metrics of the form
+    `throughput * constant` (tokens/$, normalized token revenue, and revenue
+    with a constant input/output mix). That
+    throughput is splined and the recovered multiplier applied, matching the
+    dashboard.
     """
     if not points:
         return None
@@ -213,7 +247,7 @@ def interpolate_metric(
     frontier = pareto_front_upper_left(
         points,
         get_x=lambda p: p[iv_key],
-        get_y=lambda p: p[tput_key],
+        get_y=lambda p: p[proportional_to or tput_key],
     )
     if not frontier:
         return None
@@ -241,6 +275,19 @@ def interpolate_metric(
             return None
         ys.append(value)
 
+    if proportional_to is not None:
+        tputs: list[float] = []
+        for point in sorted_front:
+            throughput = point.get(proportional_to)
+            if throughput is None:
+                return None
+            tputs.append(throughput)
+        multiplier = recover_proportional_multiplier(ys, tputs)
+        if multiplier is not None:
+            tput_slopes = monotone_slopes(xs, tputs)
+            tput = hermite_interpolate(xs, tputs, tput_slopes, target_iv)
+            return max(0.0, tput) * multiplier
+
     if reciprocal_of is not None:
         tputs: list[float] = []
         for point in sorted_front:
@@ -276,7 +323,8 @@ def interpolate_metric(
 
 def _cli() -> None:
     """Stdin: {"points": [...], "target_iv": N, "metric_key": "...",
-               "reciprocal_of": "throughput" for $/M tok and J/token}
+               "reciprocal_of": "throughput" for $/M tok and J/token,
+               "proportional_to": "throughput" for tokens/$ or revenue with a constant mix}
     Stdout: {"value": N or null}"""
     req = json.loads(sys.stdin.read())
     value = interpolate_metric(
@@ -286,6 +334,7 @@ def _cli() -> None:
         iv_key=req.get('iv_key', 'interactivity'),
         tput_key=req.get('tput_key', 'throughput'),
         reciprocal_of=req.get('reciprocal_of'),
+        proportional_to=req.get('proportional_to'),
     )
     json.dump({'value': value}, sys.stdout)
     sys.stdout.write('\n')
diff --git a/.github/claude/review-prompt.md b/.github/claude/review-prompt.md
index bc3e2665c..2eb0e5104 100644
--- a/.github/claude/review-prompt.md
+++ b/.github/claude/review-prompt.md
@@ -91,6 +91,30 @@ PostHog autocapture tracks all interactions automatically. Named `track()` calls
 
 **Important**: Only flag NEW or MODIFIED interactive elements in the PR diff. Do NOT flag existing code that was not changed in this PR.
 
+## 💡 NON-BLOCKING: Simplified Chinese copy quality
+
+The `/zh` side is hand-authored. **If the complete diff touches any user-visible Chinese string, load `.claude/skills/review-zh-copy/SKILL.md` and follow it.** Search the whole diff rather than relying on `zh` in a filename: Chinese also lives in component-local dictionaries, locale ternaries, metadata, option registries, tooltips, states, and feature-gated UI.
+
+Read `docs/chinese-copy.md` completely. Review the full UI element or paragraph together with its English source, intended meaning, surface, and rendered context. Evaluate semantic fidelity and natural Chinese as independent gates; fluent Chinese that changes the source is still wrong.
+
+**Chinese-copy findings are always non-blocking while the skill says `Operating mode: advisory`.** Do not label them `🔴 BLOCKING`, submit a request-changes review for them, or otherwise hold another contributor's merge. Do not mention `@edwingao28` for a clean review, routine coverage, or an ordinary wording or naturalness suggestion. Mention `@edwingao28` only for a high-confidence semantic or factual error, changed attribution or speaker voice, or high-impact ambiguity that cannot be resolved from the source and context. Label that finding `Needs Chinese maintainer confirmation` and ask for review. It remains a suggestion; the Chinese maintainer decides whether any follow-up is needed.
+
+Worth a suggestion:
+
+- Chinese that follows English clause order or modifier chains closely enough to sound translated.
+- A rendering that changes or loses the source meaning — fluent Chinese can still be a mistranslation, and that is the more expensive error.
+- Register wrong for the surface: marketing voice in a UI label, invented internet slang, or an attributed quote rewritten into a voice its speaker would not use.
+
+Not worth a comment:
+
+- Wording you would have phrased differently when the existing rendering is accurate and natural.
+- Long Chinese sentences per se. Length is not evidence of translationese.
+- Established English technical terms and phrases that Chinese ML infrastructure engineers normally use in English. Keep them English in first-party UI and technical prose; decide from real industry usage and the surface, not from a closed list or capitalization. `warmup`, `seed`, and `offload` are examples, not the whole category. Chinese-first explanations may be appropriate for a broader audience, and attributed quotations preserve the speaker's wording.
+- `您` solely because it appears. Controls normally omit the pronoun, but a respectful CTA may use it intentionally.
+- Awkward Chinese the diff did not touch. Mention it separately from the review.
+
+For each finding, quote the English source, state its intended meaning and failed gate, and give a complete replacement. Never comment "this reads awkwardly" on its own. End with the changed Chinese surfaces covered and any context that could not be rendered or verified.
+
 ## 🔴 BLOCKING: Test coverage enforcement
 
 When reviewing a PR diff, check if new code was added WITHOUT corresponding tests:
diff --git a/.github/workflows/apply-run-overrides.yml b/.github/workflows/apply-run-overrides.yml
index 4ec9fcf26..a0fa1189d 100644
--- a/.github/workflows/apply-run-overrides.yml
+++ b/.github/workflows/apply-run-overrides.yml
@@ -22,7 +22,7 @@ jobs:
     permissions:
       contents: read
     steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -30,14 +30,15 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - name: Install dependencies
diff --git a/.github/workflows/cache-refresh.yml b/.github/workflows/cache-refresh.yml
index 06cd7d680..61b6c2e8d 100644
--- a/.github/workflows/cache-refresh.yml
+++ b/.github/workflows/cache-refresh.yml
@@ -17,7 +17,7 @@ jobs:
     permissions:
       contents: read
     steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -25,14 +25,15 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - name: Install dependencies
diff --git a/.github/workflows/claude.yml b/.github/workflows/claude.yml
index cb3159300..a4dffca59 100644
--- a/.github/workflows/claude.yml
+++ b/.github/workflows/claude.yml
@@ -24,17 +24,17 @@ jobs:
   route:
     name: Claude (route)
     if: >-
-      ((github.event_name == 'issue_comment' || github.event_name == 'pull_request_review_comment')
-      && contains(github.event.comment.body || '', '@claude')
-      && contains(fromJson('["OWNER", "MEMBER", "COLLABORATOR", "BOT"]'), github.event.comment.author_association))
-      ||
-      (github.event_name == 'issues'
-      && (contains(github.event.issue.body || '', '@claude') || contains(github.event.issue.title || '', '@claude'))
-      && contains(fromJson('["OWNER", "MEMBER", "COLLABORATOR", "BOT"]'), github.event.issue.author_association))
-      ||
-      (github.event_name == 'pull_request'
-      && !github.event.pull_request.draft
-      && contains(fromJson('["OWNER", "MEMBER", "COLLABORATOR", "BOT"]'), github.event.pull_request.author_association))
+      ((github.event_name == 'issue_comment' || github.event_name ==
+      'pull_request_review_comment') && contains(github.event.comment.body ||
+      '', '@claude') && contains(fromJson('["OWNER", "MEMBER", "COLLABORATOR",
+      "BOT"]'), github.event.comment.author_association)) || (github.event_name
+      == 'issues' && (contains(github.event.issue.body || '', '@claude') ||
+      contains(github.event.issue.title || '', '@claude')) &&
+      contains(fromJson('["OWNER", "MEMBER", "COLLABORATOR", "BOT"]'),
+      github.event.issue.author_association)) || (github.event_name ==
+      'pull_request' && !github.event.pull_request.draft &&
+      contains(fromJson('["OWNER", "MEMBER", "COLLABORATOR", "BOT"]'),
+      github.event.pull_request.author_association))
     runs-on: ubuntu-latest
     # No repo write/PR access needed — this job only calls the Anthropic API.
     permissions:
@@ -42,7 +42,8 @@ jobs:
     # Latest event wins: cancel an in-flight route for the same PR/issue when a newer
     # event arrives, so rapid pushes don't fan out into parallel downstream reviews.
     concurrency:
-      group: claude-route-${{ github.event.pull_request.number || github.event.issue.number }}
+      group: claude-route-${{ github.event.pull_request.number ||
+        github.event.issue.number }}
       cancel-in-progress: true
     outputs:
       profile: ${{ steps.route.outputs.profile }}
@@ -58,7 +59,9 @@ jobs:
           ROUTER_MODEL: ${{ env.CLAUDE_ROUTER_MODEL }}
           BODY: ${{ github.event.comment.body || github.event.issue.body || '' }}
           TITLE: ${{ github.event.issue.title || '' }}
-          IS_PR: ${{ github.event_name == 'pull_request' || github.event.issue.pull_request != null || github.event.pull_request != null }}
+          IS_PR: ${{ github.event_name == 'pull_request' ||
+            github.event.issue.pull_request != null || github.event.pull_request
+            != null }}
         run: |
           set -euo pipefail
 
@@ -147,7 +150,8 @@ jobs:
   implement:
     name: Claude (implement)
     needs: route
-    if: contains(fromJson('["ui", "code", "docs", "question"]'), needs.route.outputs.profile)
+    if: contains(fromJson('["ui", "code", "docs", "question"]'),
+      needs.route.outputs.profile)
     runs-on: ubuntu-latest
     env:
       GITHUB_TOKEN: ${{ secrets.PAT }}
@@ -165,12 +169,13 @@ jobs:
     # Serialize @claude runs on the same issue/PR (incl. reviews — shared group with the
     # review job) so back-to-back triggers don't clobber each other.
     concurrency:
-      group: claude-${{ github.event.pull_request.number || github.event.issue.number }}
+      group: claude-${{ github.event.pull_request.number || github.event.issue.number
+        }}
       cancel-in-progress: false
 
     steps:
       - name: Checkout repository
-        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           fetch-depth: 0
           token: ${{ secrets.PAT }}
@@ -184,14 +189,15 @@ jobs:
           bun-version-file: package.json
 
       - name: Setup Node.js
-        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
 
@@ -239,7 +245,8 @@ jobs:
           fi
 
       - name: Install Cypress binary
-        if: needs.route.outputs.profile == 'ui' && steps.cypress-cache.outputs.cache-hit != 'true'
+        if: needs.route.outputs.profile == 'ui' && steps.cypress-cache.outputs.cache-hit
+          != 'true'
         working-directory: packages/app
         run: bunx cypress install
 
@@ -324,7 +331,7 @@ jobs:
       # so a dev server that didn't come up does not block this step.
       - name: Run Claude Code
         id: claude
-        uses: anthropics/claude-code-action@536f2c32a39763739000b0e1ac69ca2647d97ce9 # v1.0.170
+        uses: anthropics/claude-code-action@459ad358ae43fea66bfefd0a1f8d840b4b9791fb # v1.0.194
         env:
           GH_TOKEN: ${{ secrets.PAT }}
           GITHUB_TOKEN: ${{ secrets.PAT }}
@@ -361,10 +368,9 @@ jobs:
     # author/draft gate, so this is plain success()-gated: run when route said "review"
     # and there's a PR to review.
     if: >-
-      needs.route.outputs.profile == 'review'
-      && (github.event_name == 'pull_request'
-      || github.event.issue.pull_request
-      || github.event.pull_request)
+      needs.route.outputs.profile == 'review' && (github.event_name ==
+      'pull_request' || github.event.issue.pull_request ||
+      github.event.pull_request)
     runs-on: ubuntu-latest
     env:
       DATABASE_READONLY_URL: ${{ secrets.DATABASE_READONLY_URL }}
@@ -376,7 +382,8 @@ jobs:
       actions: read # Inspect workflow runs and checks
       id-token: write # Authenticate the Claude Code action with OIDC
     concurrency:
-      group: claude-${{ github.event.pull_request.number || github.event.issue.number }}
+      group: claude-${{ github.event.pull_request.number || github.event.issue.number
+        }}
       cancel-in-progress: false
 
     steps:
@@ -399,7 +406,7 @@ jobs:
           echo "Pinned PR #${PR_NUMBER} head sha: $SHA"
 
       - name: Checkout PR head
-        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           ref: ${{ steps.prsha.outputs.sha }}
           fetch-depth: 0
@@ -414,14 +421,15 @@ jobs:
           bun-version-file: package.json
 
       - name: Setup Node.js
-        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
 
@@ -500,7 +508,7 @@ jobs:
 
       - name: Run Claude Code
         id: claude
-        uses: anthropics/claude-code-action@536f2c32a39763739000b0e1ac69ca2647d97ce9 # v1.0.170
+        uses: anthropics/claude-code-action@459ad358ae43fea66bfefd0a1f8d840b4b9791fb # v1.0.194
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
diff --git a/.github/workflows/db-backup.yml b/.github/workflows/db-backup.yml
index 9e623d796..7da94437a 100644
--- a/.github/workflows/db-backup.yml
+++ b/.github/workflows/db-backup.yml
@@ -48,8 +48,8 @@ jobs:
   dump:
     name: Dump and compress
     needs: preflight
-    timeout-minutes: 180
-    runs-on: ubuntu-latest
+    timeout-minutes: 360
+    runs-on: blacksmith-32vcpu-ubuntu-2404
     steps:
       - name: Dump, compress, and split database
         env:
@@ -86,7 +86,7 @@ jobs:
           cat SHA256SUMS
 
       - name: Upload dump parts
-        uses: actions/upload-artifact@ea165f8d65b6e75b540449e92b4886f43607fa02 # v4
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
         with:
           name: db-dump-parts
           path: |
@@ -105,7 +105,7 @@ jobs:
       contents: write # Create the database dump release
     steps:
       - name: Download dump parts
-        uses: actions/download-artifact@d3f86a106a0bac45b974a628896c90dbdf5c8093 # v4
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
         with:
           name: db-dump-parts
 
diff --git a/.github/workflows/ingest-agentic-results.yml b/.github/workflows/ingest-agentic-results.yml
index 398376b9e..e9e273777 100644
--- a/.github/workflows/ingest-agentic-results.yml
+++ b/.github/workflows/ingest-agentic-results.yml
@@ -87,7 +87,8 @@ on:
 permissions: {}
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.event.client_payload.source-run-id || github.event.client_payload.run-id || inputs.run-id }}
+  group: ${{ github.workflow }}-${{ github.event.client_payload.source-run-id ||
+    github.event.client_payload.run-id || inputs.run-id }}
   cancel-in-progress: false
 
 jobs:
@@ -102,19 +103,19 @@ jobs:
     permissions:
       contents: read
     env:
-      REQUESTED_DATABASE_TARGET: ${{ github.event.client_payload.database-target || inputs.database-target || 'production' }}
+      REQUESTED_DATABASE_TARGET: ${{ github.event.client_payload.database-target ||
+        inputs.database-target || 'production' }}
     steps:
       - name: Wait for source run to finish
         if: >-
-          github.event_name == 'repository_dispatch' &&
-          github.event.action == 'ingest-agentic-results' &&
-          (
+          github.event_name == 'repository_dispatch' && github.event.action ==
+          'ingest-agentic-results' && (
             github.event.client_payload.source-run-id == '' ||
             github.event.client_payload.source-run-id == github.event.client_payload.merge-run-id
           )
         run: sleep 300
 
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -122,14 +123,15 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - name: Install dependencies
@@ -144,7 +146,8 @@ jobs:
           DATABASE_WRITE_URL_STAGING: ${{ secrets.DATABASE_STAGING_WRITE_URL }}
           INVALIDATE_SECRET_DEFAULT: ${{ secrets.VERCEL_INVALIDATE_SECRET }}
           PROTECTION_BYPASS_SECRET_STAGING: ${{ secrets.VERCEL_STAGING_BYPASS_SECRET }}
-          STAGING_SITE_URL: ${{ vars.STAGING_SITE_URL || 'https://inferencemax-app-git-staging-semianalysisai.vercel.app' }}
+          STAGING_SITE_URL: ${{ vars.STAGING_SITE_URL ||
+            'https://inferencemax-app-git-staging-semianalysisai.vercel.app' }}
         run: |
           case "$REQUESTED_DATABASE_TARGET" in
             production)
@@ -202,8 +205,10 @@ jobs:
         id: artifacts
         env:
           GH_TOKEN: ${{ secrets.INFX_MAIN_PAT }}
-          SOURCE_RUN_ID: ${{ github.event.client_payload.source-run-id || github.event.client_payload.run-id || inputs.run-id }}
-          MERGE_RUN_ID: ${{ github.event.client_payload.merge-run-id || github.event.client_payload.run-id || inputs.run-id }}
+          SOURCE_RUN_ID: ${{ github.event.client_payload.source-run-id ||
+            github.event.client_payload.run-id || inputs.run-id }}
+          MERGE_RUN_ID: ${{ github.event.client_payload.merge-run-id ||
+            github.event.client_payload.run-id || inputs.run-id }}
           ARTIFACTS_PATH: ${{ github.workspace }}/artifacts
           INGEST_REPO: SemiAnalysisAI/InferenceX
         run: bun run admin:db:prepare:ci
@@ -263,8 +268,9 @@ jobs:
           fi
 
       - name: Notify Slack on unmapped entities
-        if: steps.unmapped.outputs.found == 'true' && env.REQUESTED_DATABASE_TARGET != 'staging'
-        uses: slackapi/slack-github-action@45a88b9581bfab2566dc881e2cd66d334e621e2c # v3.0.3
+        if: steps.unmapped.outputs.found == 'true' && env.REQUESTED_DATABASE_TARGET !=
+          'staging'
+        uses: slackapi/slack-github-action@dcb1066f776dd043e64d0e8ba94ca15cc7e1875d # v4.0.0
         with:
           webhook: ${{ secrets.SLACK_WEBHOOK_URL }}
           webhook-type: incoming-webhook
@@ -275,7 +281,7 @@ jobs:
 
       - name: Notify Slack on failure
         if: failure() && env.REQUESTED_DATABASE_TARGET != 'staging'
-        uses: slackapi/slack-github-action@45a88b9581bfab2566dc881e2cd66d334e621e2c # v3.0.3
+        uses: slackapi/slack-github-action@dcb1066f776dd043e64d0e8ba94ca15cc7e1875d # v4.0.0
         with:
           webhook: ${{ secrets.SLACK_WEBHOOK_URL }}
           webhook-type: incoming-webhook
diff --git a/.github/workflows/ingest-results.yml b/.github/workflows/ingest-results.yml
index be2ed5abd..238d2a7b7 100644
--- a/.github/workflows/ingest-results.yml
+++ b/.github/workflows/ingest-results.yml
@@ -7,7 +7,8 @@ on:
 permissions: {}
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.event.client_payload.source-run-id || github.event.client_payload.run-id }}
+  group: ${{ github.workflow }}-${{ github.event.client_payload.source-run-id ||
+    github.event.client_payload.run-id }}
   cancel-in-progress: false
 
 jobs:
@@ -21,10 +22,11 @@ jobs:
       - name: Wait for source run to finish
         if: >-
           github.event.client_payload.source-run-id == '' ||
-          github.event.client_payload.source-run-id == github.event.client_payload.merge-run-id
+          github.event.client_payload.source-run-id ==
+          github.event.client_payload.merge-run-id
         run: sleep 300
 
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -32,14 +34,15 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - name: Install dependencies
@@ -51,8 +54,10 @@ jobs:
         id: artifacts
         env:
           GH_TOKEN: ${{ secrets.INFX_MAIN_PAT }}
-          SOURCE_RUN_ID: ${{ github.event.client_payload.source-run-id || github.event.client_payload.run-id }}
-          MERGE_RUN_ID: ${{ github.event.client_payload.merge-run-id || github.event.client_payload.run-id }}
+          SOURCE_RUN_ID: ${{ github.event.client_payload.source-run-id ||
+            github.event.client_payload.run-id }}
+          MERGE_RUN_ID: ${{ github.event.client_payload.merge-run-id ||
+            github.event.client_payload.run-id }}
           ARTIFACTS_PATH: ${{ github.workspace }}/artifacts
           INGEST_REPO: SemiAnalysisAI/InferenceX
         run: bun run admin:db:prepare:ci
@@ -115,7 +120,7 @@ jobs:
 
       - name: Notify Slack on unmapped entities
         if: steps.unmapped.outputs.found == 'true'
-        uses: slackapi/slack-github-action@45a88b9581bfab2566dc881e2cd66d334e621e2c # v3.0.3
+        uses: slackapi/slack-github-action@dcb1066f776dd043e64d0e8ba94ca15cc7e1875d # v4.0.0
         with:
           webhook: ${{ secrets.SLACK_WEBHOOK_URL }}
           webhook-type: incoming-webhook
@@ -126,7 +131,7 @@ jobs:
 
       - name: Notify Slack on failure
         if: failure()
-        uses: slackapi/slack-github-action@45a88b9581bfab2566dc881e2cd66d334e621e2c # v3.0.3
+        uses: slackapi/slack-github-action@dcb1066f776dd043e64d0e8ba94ca15cc7e1875d # v4.0.0
         with:
           webhook: ${{ secrets.SLACK_WEBHOOK_URL }}
           webhook-type: incoming-webhook
diff --git a/.github/workflows/lint.yml b/.github/workflows/lint.yml
index e44856e89..69a43a32b 100644
--- a/.github/workflows/lint.yml
+++ b/.github/workflows/lint.yml
@@ -20,7 +20,7 @@ jobs:
     permissions:
       contents: read
     steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -28,14 +28,15 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - name: Install dependencies
@@ -46,3 +47,5 @@ jobs:
         run: bun run lint
       - name: Format check
         run: bun run fmt
+      - name: Typography gate
+        run: bun run check:typography
diff --git a/.github/workflows/stage-results.yml b/.github/workflows/stage-results.yml
index 74155047f..9f91a765c 100644
--- a/.github/workflows/stage-results.yml
+++ b/.github/workflows/stage-results.yml
@@ -40,6 +40,10 @@ on:
 
 permissions: {}
 
+concurrency:
+  group: stage-results
+  cancel-in-progress: false
+
 jobs:
   validate:
     name: Validate staging request
diff --git a/.github/workflows/tests-e2e.yml b/.github/workflows/tests-e2e.yml
index 0ab6b69b4..48f292e70 100644
--- a/.github/workflows/tests-e2e.yml
+++ b/.github/workflows/tests-e2e.yml
@@ -13,6 +13,7 @@ on:
       - 'bunfig.toml'
       - 'bun.lock'
       - 'tsconfig.json'
+      - 'tsconfig.base.json'
 
 permissions: {}
 
@@ -29,7 +30,7 @@ jobs:
     permissions:
       contents: read
     steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -37,7 +38,7 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Restore Bun package cache
@@ -45,7 +46,8 @@ jobs:
         uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - name: Install dependencies
@@ -56,7 +58,8 @@ jobs:
         uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.cache/Cypress
-          key: cypress-binary-${{ runner.os }}-${{ runner.arch }}-${{ hashFiles('bun.lock') }}
+          key: cypress-binary-${{ runner.os }}-${{ runner.arch }}-${{
+            hashFiles('bun.lock') }}
           restore-keys: |
             cypress-binary-${{ runner.os }}-${{ runner.arch }}-
       - name: Install Cypress binary if missing
@@ -84,7 +87,7 @@ jobs:
         # duplicated fixed cost dominates and yields sharply diminishing returns.
         shard: [1, 2, 3, 4]
     steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -92,7 +95,7 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Restore Bun package cache
@@ -100,7 +103,8 @@ jobs:
         uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - uses: browser-actions/setup-firefox@0bc507ddf224827e3b1af68e014d5e42ab93e795 # v1.7.2
@@ -120,7 +124,9 @@ jobs:
           # New cache whenever packages or source files change; fall back to
           # the newest cache for the same lockfile so changed sources still
           # start from a warm compiler cache.
-          key: ${{ runner.os }}-nextjs-${{ hashFiles('bun.lock') }}-${{ hashFiles('packages/**/*.js', 'packages/**/*.jsx', 'packages/**/*.ts', 'packages/**/*.tsx') }}
+          key: ${{ runner.os }}-nextjs-${{ hashFiles('bun.lock') }}-${{
+            hashFiles('packages/**/*.js', 'packages/**/*.jsx',
+            'packages/**/*.ts', 'packages/**/*.tsx') }}
           restore-keys: |
             ${{ runner.os }}-nextjs-${{ hashFiles('bun.lock') }}-
             ${{ runner.os }}-nextjs-
@@ -133,14 +139,16 @@ jobs:
         uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.cache/Cypress
-          key: cypress-binary-${{ runner.os }}-${{ runner.arch }}-${{ hashFiles('bun.lock') }}
+          key: cypress-binary-${{ runner.os }}-${{ runner.arch }}-${{
+            hashFiles('bun.lock') }}
           restore-keys: |
             cypress-binary-${{ runner.os }}-${{ runner.arch }}-
       - name: Install Cypress binary if missing
         working-directory: packages/app
         run: bunx cypress verify || bunx cypress install
       - name: Save Cypress binary cache
-        if: ${{ matrix.browser == 'chrome' && matrix.shard == 1 && steps.cypress-cache.outputs.cache-hit != 'true' }}
+        if: ${{ matrix.browser == 'chrome' && matrix.shard == 1 &&
+          steps.cypress-cache.outputs.cache-hit != 'true' }}
         uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.cache/Cypress
@@ -148,7 +156,8 @@ jobs:
       # The component job only restores this cache. One E2E shard writes it, avoiding
       # concurrent cache reservations across the nine parallel test jobs.
       - name: Save Bun package cache
-        if: ${{ matrix.browser == 'chrome' && matrix.shard == 1 && steps.bun-cache.outputs.cache-hit != 'true' }}
+        if: ${{ matrix.browser == 'chrome' && matrix.shard == 1 &&
+          steps.bun-cache.outputs.cache-hit != 'true' }}
         uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
@@ -158,13 +167,14 @@ jobs:
         env:
           E2E_FIXTURES: '1'
       - name: Save Next.js build cache
-        if: ${{ matrix.browser == 'chrome' && matrix.shard == 1 && steps.nextjs-cache.outputs.cache-hit != 'true' }}
+        if: ${{ matrix.browser == 'chrome' && matrix.shard == 1 &&
+          steps.nextjs-cache.outputs.cache-hit != 'true' }}
         uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: packages/app/.next/cache
           key: ${{ steps.nextjs-cache.outputs.cache-primary-key }}
       - name: Run integration tests
-        uses: cypress-io/github-action@fa4a118725a8f001170d49631ea89e5d66fee626 # v7.4.1
+        uses: cypress-io/github-action@c32f12761482a282d24ca0fd7466d8ae86f54ba8 # v7.4.2
         with:
           working-directory: packages/app
           install: false
diff --git a/.github/workflows/tests-unit.yml b/.github/workflows/tests-unit.yml
index abdfb1041..76f003e6c 100644
--- a/.github/workflows/tests-unit.yml
+++ b/.github/workflows/tests-unit.yml
@@ -10,10 +10,12 @@ on:
       - 'packages/app/**'
       - 'packages/constants/**'
       - 'packages/db/**'
+      - 'packages/mcp/**'
       - 'bunfig.toml'
       - 'bun.lock'
       - 'scripts/run-workspace-script.ts'
       - 'tsconfig.json'
+      - 'tsconfig.base.json'
 
 permissions: {}
 
@@ -30,7 +32,7 @@ jobs:
     permissions:
       contents: read
     steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Setup Bun
@@ -38,14 +40,15 @@ jobs:
         uses: oven-sh/setup-bun@0c5077e51419868618aeaa5fe8019c62421857d6 # v2.2.0
         with:
           bun-version-file: package.json
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
         with:
           node-version: '24'
       - name: Cache Bun packages
         uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
         with:
           path: ~/.bun/install/cache
-          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
+          key: ${{ runner.os }}-bun-${{ runner.arch }}-${{
+            steps.setup-bun.outputs.bun-version }}-${{ hashFiles('bun.lock') }}
           restore-keys: |
             ${{ runner.os }}-bun-${{ runner.arch }}-${{ steps.setup-bun.outputs.bun-version }}-
       - name: Install dependencies
diff --git a/.github/workflows/zizmor.yml b/.github/workflows/zizmor.yml
index 1ce7240ca..389777957 100644
--- a/.github/workflows/zizmor.yml
+++ b/.github/workflows/zizmor.yml
@@ -27,10 +27,11 @@ jobs:
     permissions:
       contents: read
     steps:
-      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
         with:
           persist-credentials: false
       - name: Install uv
-        uses: astral-sh/setup-uv@11f9893b081a58869d3b5fccaea48c9e9e46f990 # v8.3.2
+        uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
       - name: Audit workflows
-        run: uvx --exclude-newer P1D zizmor@latest --strict-collection --pedantic --no-ignores .
+        run: uvx --exclude-newer P1D zizmor@latest --strict-collection --pedantic
+          --no-ignores .
diff --git a/.oxlintrc.json b/.oxlintrc.json
index 5a03a5a0f..8a2e56158 100644
--- a/.oxlintrc.json
+++ b/.oxlintrc.json
@@ -7,7 +7,9 @@
     "nursery": "deny"
   },
   "rules": {
-    "capitalized-comments": "off",
+    "eslint/capitalized-comments": "off",
+    "eslint/one-var": "off",
+    "one-var": "off",
     "curly": "off",
     "func-names": "off",
     "func-style": "off",
@@ -38,6 +40,7 @@
     "unicorn/filename-case": "off",
     "unicorn/max-nested-calls": "off",
     "unicorn/no-array-callback-reference": "off",
+    "unicorn/no-array-sort": "off",
     "unicorn/no-nested-ternary": "off",
     "unicorn/no-null": "off",
     "unicorn/no-useless-undefined": "off",
diff --git a/AGENTS.md b/AGENTS.md
index a50342f68..1f943f877 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -4,7 +4,9 @@ For detailed subsystem docs, see [docs/index.md](./docs/index.md).
 
 > **PR and GitHub-issue titles & descriptions must be bilingual — include a Simplified Chinese version in addition to English.** Title format: `<English title> / <中文标题>` (keep bracket prefixes at the front untranslated). In the PR/issue body, follow the English content with a `## 中文说明` section mirroring the summary; don't translate code blocks, logs, or stack traces — summarize around them. **Commit messages must include a Chinese translation too**: keep the subject line in English (conventional-commit style) and include the Chinese translation of the subject and key points in the commit body (e.g. a trailing `中文：<translation>` paragraph); squash-merge commits inherit the bilingual PR title, which satisfies the subject requirement automatically.
 
-> **Translation quality bar:** write natural technical Chinese, not word-for-word machine translation (style reference: [`vllm-project/vllm-ascend` `README.zh.md`](https://github.com/vllm-project/vllm-ascend/blob/main/README.zh.md)). Preserve product names, hardware SKUs, framework/library names (Next.js, React Query, D3.js, Tailwind ...), flags, and code identifiers in English. Use parenthetical English clarification for acronyms on first use. Preferred terms: benchmark 基准测试, dashboard 仪表板, chart 图表, config 配置, throughput 吞吐量, latency 延迟, single-node/multi-node 单节点/多节点, evaluation 评估, artifact 产物.
+> **Translation quality bar:** write natural technical Chinese, not word-for-word machine translation (style reference: [`vllm-project/vllm-ascend` `README.zh.md`](https://github.com/vllm-project/vllm-ascend/blob/main/README.zh.md)). Preserve product names, hardware SKUs, framework/library names (Next.js, React Query, D3.js, Tailwind ...), flags, and code identifiers in English. Use parenthetical English clarification for acronyms on first use. Preferred terms: benchmark 基准测试, dashboard 仪表板, chart 图表, config 配置, throughput 吞吐量, latency 延迟, single-node/multi-node 单节点/多节点, evaluation 评估, artifact 产物. In first-party ML infrastructure UI and technical prose, keep established English technical terms and phrases that Chinese engineers normally use in English; decide from real industry usage and the surface, not from a closed list. `warmup`, `seed`, and `offload` are examples, not the whole category. Follow `docs/chinese-copy.md` for explanatory and quotation exceptions.
+
+> **Chinese editorial decisions are context-aware.** Follow [docs/chinese-copy.md](./docs/chinese-copy.md) for audience, surface-specific register, terminology exceptions, the two independent fidelity/naturalness gates, the Claude-plus-manual-review workflow, and the PR checklist. Do not turn a preferred wording or pronoun choice into a global mechanical rule unless it is correct in every supported context.
 
 > **The website itself is bilingual too — every indexable page must ship a Simplified Chinese sibling under `/zh`.** See [Chinese Website Pages](#chinese-website-pages-zh--mandatory-for-all-indexable-surfaces) below; a new page, tab, or blog post without its `/zh` version is 🔴 BLOCKING on PR review.
 
@@ -27,6 +29,7 @@ InferenceX App — Next.js 16 dashboard for ML inference benchmark data. DB-back
 ```bash
 bun install               # Install dependencies
 bun run dev                # Dev server with Turbopack (http://localhost:3000)
+bun run mcp                # MCP server exposing read-only benchmark tools
 bun run build              # Production build
 bun run typecheck          # TypeScript type checking (all packages)
 bun run lint               # Lint with oxlint
@@ -52,7 +55,8 @@ packages/
 │       ├── hooks/api/    # React Query hooks (use-benchmarks, use-availability, etc.)
 │       └── lib/          # Utilities, constants, d3-chart/, chart-utils, blog, data-mappings
 ├── constants/            # Shared constants (GPU keys, model mappings, SEO)
-└── db/                   # DB layer, ETL, migrations, queries, ingest scripts
+├── db/                   # DB layer, ETL, migrations, queries, ingest scripts
+└── mcp/                  # MCP server exposing read-only benchmark tools
 ```
 
 **Path alias**: `@/*` → `packages/app/src/`
@@ -123,6 +127,7 @@ Static content routes (no DB):
 - **Linter**: oxlint — `bun run lint` / `bun run lint:fix`
 - **Formatter**: oxfmt — `bun run fmt` / `bun run fmt:fix`
 - **Type checking**: `bun run typecheck` (tsc --noEmit, strict mode)
+- **Typography gate**: `bun run check:typography` (CI + pre-commit) — no new arbitrary font sizes (`text-[11px]`) or letter-spacing (`tracking-[0.16em]`) in class strings, and no quoted font-size literals in `src/lib/d3-chart/`. Use `text-2xs`/`text-3xs`, `tracking-eyebrow`/`tracking-eyebrow-wide`/`tracking-heading`, the `<Heading>`/`<Eyebrow>` components, or `CHART_TYPE` from `@/lib/d3-chart/typography` (chart font sizes must stay in TS — CSS variables don't survive PNG export). Existing offenders live in `packages/app/scripts/typography-allowlist.json` and migrate on touch. See [Typography](./docs/typography.md).
 - **Node**: 24.x
 
 ## Environment Variables
@@ -150,7 +155,7 @@ All interactive elements should have `track()` from `@/lib/analytics` (autocaptu
 
 ## Tab Structure
 
-Order: `inference` → `evaluation` → `historical` → `calculator` → `reliability` → `gpu-specs` (defined in `page-content.tsx` `VALID_TABS`). Tab value = URL hash.
+Dashboard route keys, paths, primary/feature-gated/footer-only navigation, indexability, provider capabilities, `/zh` mirroring, and share-parameter scopes are defined once in `packages/app/src/lib/dashboard-routes.ts`. `TabNav`, `DashboardShell`, metadata, i18n, share URLs, and the sitemap derive from that registry; do not add a second route array.
 
 ## Unofficial Run Support — Mandatory for Inference / Evaluation Features
 
@@ -172,14 +177,18 @@ The site ships a hand-authored Simplified Chinese sibling for every indexable pa
 
 **Every new indexable page, dashboard tab, or blog post MUST ship its Chinese version in the same PR:**
 
-1. **New page** → create `packages/app/src/app/zh/<route>/page.tsx` with fully translated content and metadata. Metadata: `alternates: zhAlternates('<en-path>')` plus `openGraph.locale: ZH_OG_LOCALE`. Switch the English page's `alternates` to `enAlternates('<en-path>')` so both sides carry bidirectional hreflang. Register the route in `ZH_MIRRORED_ROUTES` (`src/lib/i18n.ts`) so the header nav and EN↔中文 toggle link to it, and add it to the sitemap via `localizedPair()` in `src/app/sitemap.ts`.
-2. **New dashboard tab** → add the tab to `ZH_TAB_KEYS`, `TAB_META_ZH`, `TAB_INTRO_ZH`, and `TAB_LABELS_ZH` in `src/lib/tab-meta-zh.ts`, then create `src/app/zh/(dashboard)/<tab>/page.tsx` mirroring the English page with `tabMetadataZh('<tab>')` and a `<ZhTabIntro tab="<tab>" />` block above the chart; the chart's own UI strings must follow rule 5. `tab-meta-zh.test.ts` enforces dictionary completeness.
+1. **New page** → create `packages/app/src/app/zh/<route>/page.tsx` with fully translated content and metadata. Metadata: `alternates: zhAlternates('<en-path>')` plus `openGraph.locale: ZH_OG_LOCALE`. Switch the English page's `alternates` to `enAlternates('<en-path>')` so both sides carry bidirectional hreflang. Register non-dashboard routes in `ZH_MIRRORED_ROUTES` (`src/lib/i18n.ts`) so the header nav and EN↔中文 toggle link to them, and add them to the sitemap via `localizedPair()`.
+2. **New dashboard tab** → add one entry to `src/lib/dashboard-routes.ts`, then add exact entries to `TAB_META_ZH`, `TAB_INTRO_ZH`, and `TAB_LABELS_ZH` in `src/lib/tab-meta-zh.ts`. Create `src/app/zh/(dashboard)/<tab>/page.tsx` mirroring the English page with `tabMetadataZh('<tab>')` and a `<ZhTabIntro tab="<tab>" />` block above the chart; the registry automatically supplies `/zh` route matching and indexable sitemap entries. The chart's own UI strings must follow rule 5. Registry and metadata tests enforce route parity and dictionary completeness.
 3. **New blog post** → the translation `packages/app/content/blog/zh/<same-filename>.mdx` is REQUIRED in the same PR. Translate frontmatter `title`/`subtitle` and the body; keep `date`, `publishDate`, `modifiedDate`, `tags`, and the filename/slug identical (English and Chinese posts pair by filename; visibility gating always follows the English post's `publishDate`). Rewrite internal `/blog/<slug>` links to `/zh/blog/<slug>`; never alter numbers, code blocks, or `<Figure>`/`<JsonLd>` structure. The `/zh/blog` listing, hreflang, and sitemap pick the file up automatically.
-4. **Editing an existing English page or post** → update its Chinese sibling in the same PR. Content drift between languages is a 🔴 BLOCKING review issue.
+4. **Editing an existing English page or post** → update its Chinese sibling in the same PR. Omitting the required Chinese sibling update is a 🔴 BLOCKING review issue. Fidelity or wording issues inside an updated Chinese sibling follow the advisory workflow below; serious findings require Chinese maintainer confirmation.
 5. **ALL user-visible UI strings MUST have a Chinese equivalent** — no carve-outs for "chart internals" or "option labels". This includes: headers/footers, card titles/descriptions, control and filter labels, buttons, toggles (Log Scale, Optimal Only, …), nudges, dropdown OPTION display names (Y-axis metric names, token types, scale modes), searchable-select placeholders ("Search…"), table column headers and action buttons ("Prompts"), modal/drawer chrome, legend footnotes, and empty/loading/error messages. Mechanism: client components call `useLocale()` (`src/lib/use-locale.ts`) and read from a component-local `STRINGS = { en, zh }` dict; server components take an optional `locale` prop passed from the /zh page; registry-defined display names (e.g. `Y_AXIS_METRICS`, legend toggle configs) carry a `labelZh` field resolved through a locale-aware label helper at render time. The `en` values must keep the exact original strings so English pages stay byte-identical.
-6. **What stays English** (only these): brand/product names, hardware SKUs, model/framework/precision names, units (tok/s/user, GB/s, $/M tok), code identifiers and flags — per the translation quality bar — plus DB-stored _content_ (benchmark rows, dataset conversation text, run logs), which is data, not UI.
+6. **What stays English**: brand/product names, hardware SKUs, model/framework/precision names, units (tok/s/user, GB/s, $/M tok), code identifiers and flags — per the translation quality bar — plus established English technical terms and phrases that Chinese ML infrastructure engineers normally use in English, as documented contextually in `docs/chinese-copy.md` (`warmup`, `seed`, and `offload` are non-exhaustive examples), and DB-stored _content_ (benchmark rows, dataset conversation text, run logs), which is data, not UI.
 7. **Compare slug narrative sync**: the per-slug compare pages are mirrored at `/zh/compare/[slug]` and `/zh/compare-per-dollar/[slug]`; their Chinese prose templates live in `src/lib/compare-ssr-zh.ts`, a 1:1 port of the English templates in `compare-ssr.ts`. The variant compare pages (`/zh/compare-precision/[slug]` and `/zh/compare-spec-decode/[slug]`) have their Chinese templates in `src/lib/compare-variant-ssr-zh.ts`, porting `compare-variant-ssr.ts`. Any PR that changes the English narrative templates MUST update the zh port in the same commit.
-8. **Every route gets a /zh sibling — including hidden/feature-gated ones** (`/datasets`, `/ai-chart`, `/current-inferencex-image`, `/feedback`, agentic detail pages). Noindex routes keep their noindex on both sides. The only exceptions: `feed.xml`/`llms.txt` (single-language machine feeds) and per-post OG images (Chinese posts reuse the English post's OG image — the OG renderer's font has no CJK glyphs).
+8. **Every route gets a /zh sibling, including hidden or feature-gated ones** (`/agentx`, `/ai-chart`, `/current-inferencex-image`, `/feedback`, agentic detail pages). Noindex routes keep their noindex on both sides. The only exceptions are `feed.xml` and `llms.txt` (single-language machine feeds) plus per-post OG images. Chinese posts reuse the English post's OG image because the renderer font has no CJK glyphs.
+
+> **Claude reviews every changed user-visible Chinese string, but the Chinese maintainer makes the final decision.** Load `review-zh-copy` before opening or reviewing any PR that touches user-visible Chinese text, including refactors whose filenames do not contain `zh`. The skill evaluates semantic fidelity and natural Chinese as separate gates and follows [docs/chinese-copy.md](./docs/chinese-copy.md). Its findings never block another contributor's merge. Do not mention `@edwingao28` for a clean review, routine coverage, or an ordinary wording or naturalness suggestion. Mention `@edwingao28` only for a high-confidence semantic or factual error, changed attribution or speaker voice, or unresolved high-impact ambiguity, and ask for confirmation. The Chinese maintainer makes the final editorial decision.
+
+> **Chinese copy has a narrow mechanical CI guard.** `packages/app/src/lib/zh-copy.test.ts` checks objective regressions: the standalone English common noun `Chip` in Chinese text (with explicit identifier and unit exemptions), prose-like English `<strong>` labels that bypass an available `zh:` dictionary, and redundant English-first double translations such as `warmup 预热`. It does not globally ban `预热`, `随机种子`, or `卸载`, and deliberately does **not** enforce pronouns, preferred contextual terminology, punctuation style, sentence structure, fluency, tone, or register; those require page context and human review. Every mechanical rule must use the same matcher for the source scan and a hand-checked case in `packages/app/src/lib/zh-copy-mechanical-regressions.jsonl`, including an exemption case whenever the rule has one.
 
 ## Chart Interpolation — TS and Python Helpers MUST Stay in Sync
 
@@ -217,6 +226,7 @@ Authoritative total / active parameter counts for every model in the dashboard.
 | Kimi-K2.6              | 1T    | 32B         | `moonshotai/Kimi-K2.6`              | HF model card                      |
 | Kimi-K2.7-Code         | 1T    | 32B         | `moonshotai/Kimi-K2.7-Code`         | HF model card                      |
 | Qwen3.5-397B-A17B      | 397B  | 17B         | `Qwen/Qwen3.5-397B-A17B`            | HF model card                      |
+| Qwen3.8-Flash-Next     | 176B  | 6B          | `Qwen/Qwen3.8-Flash-Next-FP8`       | HF model card                      |
 | GLM-5                  | 744B  | 40B         | `zai-org/GLM-5`                     | HF model card                      |
 | GLM-5.1                | 744B  | 40B         | `zai-org/GLM-5.1-FP8`               | HF model card (same base as GLM-5) |
 | MiniMax-M2.5           | 230B  | 10B         | `MiniMaxAI/MiniMax-M2.5`            | HF model card                      |
@@ -226,6 +236,7 @@ Authoritative total / active parameter counts for every model in the dashboard.
 
 **Common mislabel traps** (have all bitten this repo at least once — do not repeat):
 
+- **Qwen3.8-Flash-Next is 176B, not 125B.** The model card leads with "125B with 6B activated", but that is the main model only; the 51B n-gram embedding table brings the total to 176B. The separate 4B MTP head sits outside both figures. It is a Qwen4-architecture preview (GatedDeltaNet + Qwen Sparse Attention, 512 experts, 10 routed + 1 shared), not a Qwen3.5 point release, so it gets its own DB bucket.
 - **GLM-5 ≠ 355B.** 355B is GLM-4.5. GLM-5 jumped to 744B / 40B active (256-expert MoE with DSA).
 - **MiniMax-M2.5/M2.7 ≠ 456B.** 456B is the older MiniMax-Text-01 / M1 (32 large experts). The M2 series is a different architecture: 230B / 10B active, 256 small experts.
 - **DeepSeek-R1 is 671B, not 685B.** HF metadata shows 685B because the bundled MTP head adds ~14B; the core MoE is 671B / 37B active.
@@ -248,13 +259,12 @@ Authoritative total / active parameter counts for every model in the dashboard.
 
 ### Add/modify a metric
 
-1. Register in `src/lib/chart-utils.ts`: `Y_AXIS_METRICS`, `calculateRoofline`, `computeAllRooflines`, `markRooflinePoints`
-2. Add TS types: optional field in `InferenceData`, add to `YAxisMetricKey`, add `ChartDefinition` fields
-3. Add chart config: `src/components/inference/inference-chart-config.json`
-4. Add Y-axis dropdown: `ChartControls.tsx`
-5. Add subtitle/disclaimer in `ChartDisplay.tsx` if metric depends on assumed constants
-6. Add disagg caveat banner in `ChartDisplay.tsx` for per-GPU or per-MW metrics (animated amber `border-l-2` banner pattern)
-7. Expose in UI state: `InferenceContext.tsx`
+1. Register the field, bilingual labels/titles, polarity, and custom-source metadata in `src/components/inference/metric-registry.ts` (`METRIC_REGISTRY`).
+2. Add or reuse the derived field in `buildDerivedChartFields` (`src/lib/chart-utils.ts`) and the corresponding optional `InferenceData` field. Historical trends request the same selective builder; do not duplicate formulas.
+3. Add the metric to `METRIC_CONTROL_GROUPS` when it belongs in a selector group. Chart definitions and metric key types derive from the registry.
+4. Add subtitle/disclaimer copy in `ChartDisplay.tsx` if the metric depends on assumed constants.
+5. Add the disaggregated-config caveat banner for per-GPU or per-MW metrics.
+6. Verify both official rows and `?unofficialrun=` overlays, including clipping, CSV/table behavior where applicable, zoom, labels, and Chinese UI.
 
 ### Add a new blog post
 
@@ -282,11 +292,12 @@ See [Blog](./docs/blog.md) for content format, available MDX components, and des
 
 ### Adding a new tab
 
-1. `page-content.tsx`: Add to `VALID_TABS`, add `TabsTrigger` (desktop), `SelectItem` (mobile), `TabsContent`
-2. Create a per-section context provider (see `InferenceContext.tsx`, `EvaluationContext.tsx` for patterns)
-3. Use `ChartLegend` with `variant="sidebar"`, sorted by `HW_REGISTRY` sort order, default expanded
-4. Analytics: all interactive elements use `track()` with `{tabname}_` prefix
-5. Create the Chinese sibling: extend `src/lib/tab-meta-zh.ts` dictionaries and add `src/app/zh/(dashboard)/<tab>/page.tsx` (see [Chinese Website Pages](#chinese-website-pages-zh--mandatory-for-all-indexable-surfaces))
+1. Add the canonical route entry to `packages/app/src/lib/dashboard-routes.ts`, including navigation group, indexability, provider capabilities, locale mirroring, and share scopes.
+2. Add explicit English metadata in `src/lib/tab-meta.ts` and Chinese metadata, intro, and label entries in `src/lib/tab-meta-zh.ts`.
+3. Create the English and Chinese App Router pages. `dashboard-routes.test.ts` enforces structural parity.
+4. Add the tab content and any section-specific provider only when the route actually needs one; declare dashboard-wide provider needs in the registry.
+5. Use `ChartLegend` with `variant="sidebar"`, sorted by `HW_REGISTRY` sort order, default expanded.
+6. Track interactive actions with the `{tabname}_` analytics prefix.
 
 ### Bumping dependencies
 
@@ -303,7 +314,7 @@ Workflow for a periodic dep bump. Branch: `chore/bump-deps-YYYY-MM-DD`. Commit e
 Detailed design rationale (the "why" and "how", not the "what") lives in [docs/](./docs/index.md):
 
 - **[Index](./docs/index.md)** — index of all docs **MUST ALWAYS READ IN CASE OF RELEVANT INFORMATION**
-- **[Architecture](./docs/architecture.md)** — Client-first design, hash routing, caching, color system
+- **[Architecture](./docs/architecture.md)** — Client-first design, route navigation, URL state, caching, color system
 - **[D3 Charts](./docs/d3-charts.md)** — 4-effect architecture, zoom refs, tooltip lifecycle
 - **[Data Pipeline](./docs/data-pipeline.md)** — DB schema reasoning, ETL design, spline interpolation
 - **[Pitfalls](./docs/pitfalls.md)** — Token type bugs, schema evolution, stale closures, zoom loss
diff --git a/README.md b/README.md
index 43b6b968b..b50e3ed8d 100644
--- a/README.md
+++ b/README.md
@@ -21,6 +21,7 @@ Neon PostgreSQL → API routes (/api/v1/*) → React Query hooks → Context pro
 ```
 
 The frontend fetches data from API routes backed by a Neon PostgreSQL read replica. All presentation logic lives in the frontend — API routes return raw DB data.
+The MCP server exposes read-only benchmark tools through the same `db` and `constants` packages, without depending on frontend code.
 
 ### Monorepo Structure
 
@@ -28,7 +29,8 @@ The frontend fetches data from API routes backed by a Neon PostgreSQL read repli
 packages/
 ├── app/          # Next.js frontend
 ├── constants/    # Shared constants (GPU keys, model mappings)
-└── db/           # DB layer, ETL, migrations, queries, ingest scripts
+├── db/           # DB layer, ETL, migrations, queries, ingest scripts
+└── mcp/          # MCP server exposing read-only benchmark tools
 ```
 
 ## Prerequisites
@@ -105,27 +107,29 @@ Open [http://localhost:3000](http://localhost:3000) with your browser.
 These are the main scripts you'll use during development. Admin scripts for database and cache management are listed separately below.
 Some of these may require additional setup or environment variables.
 
-| Script                         | Description                                    |
-| ------------------------------ | ---------------------------------------------- |
-| `bun run dev`                  | Start development server with Turbopack        |
-| `bun run build`                | Production build                               |
-| `bun run start`                | Start production server                        |
-| `bun run preview`              | Build then start production server locally     |
-| `bun run typecheck`            | TypeScript type checking (all packages)        |
-| `bun run lint`                 | Lint with oxlint                               |
-| `bun run lint:fix`             | Auto-fix lint issues                           |
-| `bun run fmt`                  | Format check with oxfmt                        |
-| `bun run fmt:fix`              | Auto-fix formatting                            |
-| `bun run security`             | Security audit                                 |
-| `bun run test`                 | Run all tests (unit + E2E)                     |
-| `bun run test:unit`            | Vitest unit tests                              |
-| `bun run test:unit:coverage`   | Vitest unit tests with coverage                |
-| `bun run test:e2e`             | Curated local Cypress smoke suite              |
-| `bun run test:e2e:full`        | Complete Cypress component + integration suite |
-| `bun run test:e2e:component`   | Complete Cypress component suite only          |
-| `bun run test:e2e:integration` | Complete Cypress integration suite only        |
-| `bun run clean`                | Remove build artifacts                         |
-| `bun run clean:all`            | Remove build artifacts + node_modules          |
+| Script                         | Description                                     |
+| ------------------------------ | ----------------------------------------------- |
+| `bun run dev`                  | Start development server with Turbopack         |
+| `bun run mcp`                  | Start the InferenceX MCP server                 |
+| `bun run build`                | Production build + route bundle budget guard    |
+| `bun run check:bundle-budgets` | Check existing Next.js route bundle diagnostics |
+| `bun run start`                | Start production server                         |
+| `bun run preview`              | Build then start production server locally      |
+| `bun run typecheck`            | TypeScript type checking (all packages)         |
+| `bun run lint`                 | Lint with oxlint                                |
+| `bun run lint:fix`             | Auto-fix lint issues                            |
+| `bun run fmt`                  | Format check with oxfmt                         |
+| `bun run fmt:fix`              | Auto-fix formatting                             |
+| `bun run security`             | Security audit                                  |
+| `bun run test`                 | Run all tests (unit + E2E)                      |
+| `bun run test:unit`            | Vitest unit tests                               |
+| `bun run test:unit:coverage`   | Vitest unit tests with coverage                 |
+| `bun run test:e2e`             | Curated local Cypress smoke suite               |
+| `bun run test:e2e:full`        | Complete Cypress component + integration suite  |
+| `bun run test:e2e:component`   | Complete Cypress component suite only           |
+| `bun run test:e2e:integration` | Complete Cypress integration suite only         |
+| `bun run clean`                | Remove build artifacts                          |
+| `bun run clean:all`            | Remove build artifacts + node_modules           |
 
 This repository uses Vitest and Cypress. Use the `bun run test:*` scripts above; `bun test` invokes Bun's separate built-in test runner and is not supported.
 
diff --git a/README_zh.md b/README_zh.md
index 94c800ae5..502586663 100644
--- a/README_zh.md
+++ b/README_zh.md
@@ -21,6 +21,7 @@ Neon PostgreSQL → API routes (/api/v1/*) → React Query hooks → Context pro
 ```
 
 前端从由 Neon PostgreSQL 只读副本支撑的 API 路由获取数据。所有展示逻辑都在前端 — API 路由只返回原始数据库数据。
+MCP server 通过相同的 `db` 与 `constants` package 提供只读基准测试工具，不依赖前端代码。
 
 ### Monorepo 结构
 
@@ -28,7 +29,8 @@ Neon PostgreSQL → API routes (/api/v1/*) → React Query hooks → Context pro
 packages/
 ├── app/          # Next.js 前端
 ├── constants/    # 共享常量（GPU key、模型映射）
-└── db/           # 数据库层、ETL、迁移、查询、数据摄取脚本
+├── db/           # 数据库层、ETL、迁移、查询、数据摄取脚本
+└── mcp/          # 提供只读基准测试工具的 MCP server
 ```
 
 ## 前置条件
@@ -105,27 +107,29 @@ bun run dev
 以下是开发过程中的常用脚本。数据库与缓存管理的运维脚本单独列在下方。
 部分脚本可能需要额外的配置或环境变量。
 
-| 脚本                           | 说明                              |
-| ------------------------------ | --------------------------------- |
-| `bun run dev`                  | 启动开发服务器（Turbopack）       |
-| `bun run build`                | 生产构建                          |
-| `bun run start`                | 启动生产服务器                    |
-| `bun run preview`              | 本地构建并启动生产服务器          |
-| `bun run typecheck`            | TypeScript 类型检查（所有包）     |
-| `bun run lint`                 | 使用 oxlint 进行 lint             |
-| `bun run lint:fix`             | 自动修复 lint 问题                |
-| `bun run fmt`                  | 使用 oxfmt 检查格式               |
-| `bun run fmt:fix`              | 自动修复格式                      |
-| `bun run security`             | 安全审计                          |
-| `bun run test`                 | 运行所有测试（单元 + E2E）        |
-| `bun run test:unit`            | Vitest 单元测试                   |
-| `bun run test:unit:coverage`   | Vitest 单元测试（含覆盖率）       |
-| `bun run test:e2e`             | 精选的本地 Cypress smoke 测试套件 |
-| `bun run test:e2e:full`        | 完整 Cypress 组件与集成测试套件   |
-| `bun run test:e2e:component`   | 仅运行完整 Cypress 组件测试套件   |
-| `bun run test:e2e:integration` | 仅运行完整 Cypress 集成测试套件   |
-| `bun run clean`                | 清除构建产物                      |
-| `bun run clean:all`            | 清除构建产物 + node_modules       |
+| 脚本                           | 说明                                |
+| ------------------------------ | ----------------------------------- |
+| `bun run dev`                  | 启动开发服务器（Turbopack）         |
+| `bun run mcp`                  | 启动 InferenceX MCP server          |
+| `bun run build`                | 生产构建并检查路由 bundle 预算      |
+| `bun run check:bundle-budgets` | 检查已有的 Next.js 路由 bundle 诊断 |
+| `bun run start`                | 启动生产服务器                      |
+| `bun run preview`              | 本地构建并启动生产服务器            |
+| `bun run typecheck`            | TypeScript 类型检查（所有包）       |
+| `bun run lint`                 | 使用 oxlint 进行 lint               |
+| `bun run lint:fix`             | 自动修复 lint 问题                  |
+| `bun run fmt`                  | 使用 oxfmt 检查格式                 |
+| `bun run fmt:fix`              | 自动修复格式                        |
+| `bun run security`             | 安全审计                            |
+| `bun run test`                 | 运行所有测试（单元 + E2E）          |
+| `bun run test:unit`            | Vitest 单元测试                     |
+| `bun run test:unit:coverage`   | Vitest 单元测试（含覆盖率）         |
+| `bun run test:e2e`             | 精选的本地 Cypress smoke 测试套件   |
+| `bun run test:e2e:full`        | 完整 Cypress 组件与集成测试套件     |
+| `bun run test:e2e:component`   | 仅运行完整 Cypress 组件测试套件     |
+| `bun run test:e2e:integration` | 仅运行完整 Cypress 集成测试套件     |
+| `bun run clean`                | 清除构建产物                        |
+| `bun run clean:all`            | 清除构建产物 + node_modules         |
 
 本项目使用 Vitest 和 Cypress。请使用上表中的 `bun run test:*` 脚本；`bun test` 会调用 Bun 自带的另一套测试运行器，本项目不支持该命令。
 
diff --git a/bun.lock b/bun.lock
index 4bf79debf..654639de0 100644
--- a/bun.lock
+++ b/bun.lock
@@ -6,14 +6,14 @@
       "name": "inferencex",
       "devDependencies": {
         "@babel/core": "^8.0.1",
-        "@types/node": "^26.1.1",
+        "@types/node": "^26.2.0",
         "is-ci": "^4.1.0",
         "lefthook": "^2.1.10",
-        "oxfmt": "^0.60.0",
-        "oxlint": "^1.75.0",
+        "oxfmt": "^0.63.0",
+        "oxlint": "^1.78.0",
         "rimraf": "^6.1.3",
         "typescript": "^6.0.3",
-        "vite": "^8.1.5",
+        "vite": "^8.2.1",
       },
     },
     "packages/app": {
@@ -22,9 +22,8 @@
       "dependencies": {
         "@chenglou/pretext": "^0.0.8",
         "@jpinsonneau/html-to-image": "^1.11.13",
-        "@noble/ciphers": "^2.2.0",
-        "@posthog/nextjs-config": "^1.9.68",
-        "@radix-ui/react-accordion": "^1.2.20",
+        "@noble/ciphers": "^2.3.0",
+        "@posthog/nextjs-config": "^1.9.69",
         "@radix-ui/react-dialog": "^1.1.23",
         "@radix-ui/react-label": "^2.1.15",
         "@radix-ui/react-popover": "^1.1.23",
@@ -33,26 +32,26 @@
         "@radix-ui/react-switch": "^1.3.7",
         "@radix-ui/react-tabs": "^1.1.21",
         "@radix-ui/react-tooltip": "^1.2.16",
-        "@react-three/fiber": "^9.6.1",
+        "@react-three/fiber": "^9.7.0",
         "@semianalysisai/inferencex-constants": "workspace:*",
         "@semianalysisai/inferencex-db": "workspace:*",
-        "@shikijs/rehype": "^4.3.1",
+        "@shikijs/rehype": "^4.4.3",
         "@tailwindcss/typography": "^0.5.20",
         "@tanstack/react-query": "^5.101.4",
         "@vercel/analytics": "^2.0.1",
-        "@vercel/blob": "^2.6.1",
+        "@vercel/blob": "^2.8.0",
         "@vercel/speed-insights": "^2.0.0",
         "adm-zip": "^0.6.0",
         "class-variance-authority": "^0.7.1",
         "clsx": "^2.1.1",
         "d3": "^7.9.0",
-        "dompurify": "^3.4.12",
+        "dompurify": "^3.4.13",
         "gray-matter": "^4.0.3",
         "iwanthue": "^2.1.0",
         "katex": "^0.16.47",
-        "lucide-react": "^1.27.0",
+        "lucide-react": "^1.31.0",
         "mp4-muxer": "^5.2.2",
-        "next": "^16.2.12",
+        "next": "^16.3.1",
         "next-mdx-remote": "^6.0.0",
         "next-themes": "^0.4.6",
         "posthog-js": "^1.407.2",
@@ -61,7 +60,7 @@
         "rehype-katex": "^7.0.1",
         "remark-gfm": "^4.0.1",
         "remark-math": "^6.0.0",
-        "shiki": "^4.3.1",
+        "shiki": "^4.4.3",
         "tailwind-merge": "^3.6.0",
         "three": "^0.185.1",
       },
@@ -71,16 +70,17 @@
         "@types/adm-zip": "^0.5.8",
         "@types/d3": "^7.4.3",
         "@types/mdast": "^4.0.4",
-        "@types/node": "^26.1.1",
-        "@types/react": "^19.2.17",
-        "@types/react-dom": "^19.2.3",
-        "@types/three": "^0.185.1",
+        "@types/node": "^26.2.0",
+        "@types/react": "19.2.17",
+        "@types/react-dom": "19.2.3",
+        "@types/three": "^0.185.4",
         "@vitest/coverage-v8": "^4.1.10",
-        "cypress": "^15.19.0",
+        "axe-core": "^4.13.0",
+        "cypress": "15.19.0",
+        "cypress-axe": "^1.7.0",
         "cypress-split": "^1.25.0",
-        "esbuild": "^0.28.1",
-        "fake-indexeddb": "^6.2.5",
-        "jsdom": "^29.1.1",
+        "esbuild": "^0.28.2",
+        "jsdom": "^30.0.1",
         "rimraf": "^6.1.3",
         "tailwindcss": "^4.3.3",
         "tw-animate-css": "^1.4.0",
@@ -101,7 +101,7 @@
       "version": "0.1.0",
       "dependencies": {
         "@neondatabase/serverless": "^1.1.0",
-        "@noble/ciphers": "^2.2.0",
+        "@noble/ciphers": "^2.3.0",
         "@semianalysisai/inferencex-constants": "workspace:*",
         "postgres": "^3.4.9",
         "stream-chain": "^4.2.5",
@@ -109,7 +109,7 @@
       },
       "devDependencies": {
         "@types/adm-zip": "^0.5.8",
-        "@types/node": "^26.1.1",
+        "@types/node": "^26.2.0",
         "@types/stream-json": "^1.7.8",
         "@vitest/coverage-v8": "^4.1.10",
         "adm-zip": "^0.6.0",
@@ -121,14 +121,14 @@
       "name": "@semianalysisai/inferencex-mcp",
       "version": "1.0.0",
       "dependencies": {
-        "@modelcontextprotocol/sdk": "^1.29.0",
+        "@modelcontextprotocol/sdk": "^1.30.0",
         "@semianalysisai/inferencex-constants": "workspace:*",
         "@semianalysisai/inferencex-db": "workspace:*",
         "postgres": "^3.4.9",
         "zod": "^4.4.3",
       },
       "devDependencies": {
-        "@types/node": "^26.1.1",
+        "@types/node": "^26.2.0",
         "typescript": "^6.0.3",
         "vitest": "^4.1.10",
       },
@@ -142,6 +142,17 @@
     "lefthook",
     "cypress",
   ],
+  "overrides": {
+    "brace-expansion": "5.0.9",
+    "dompurify": "3.4.13",
+    "fast-uri": "4.1.2",
+    "hono": "4.13.2",
+    "ip-address": "10.5.0",
+    "js-yaml": "3.15.1",
+    "jsdom": "30.0.1",
+    "nanoid": "3.3.18",
+    "undici": "8.10.0",
+  },
   "packages": {
     "@actions/core": ["@actions/core@2.0.3", "", { "dependencies": { "@actions/exec": "2.0.0", "@actions/http-client": "3.0.2" } }, "sha512-Od9Thc3T1mQJYddvVPM4QGiLUewdh+3txmDYHHxoNdkqysR1MbCT+rFOtNUxYAz+7+6RIsqipVahY2GJqGPyxA=="],
 
@@ -153,13 +164,9 @@
 
     "@alloc/quick-lru": ["@alloc/quick-lru@5.2.0", "", {}, "sha512-UrcABB+4bUrFABwbluTIBErXwvbsU/V7TZWfmbgJfbkwiBuziS9gxdODUyuiecfdGQ85jglMW6juS3+z5TsKLw=="],
 
-    "@asamuzakjp/css-color": ["@asamuzakjp/css-color@5.1.11", "", { "dependencies": { "@asamuzakjp/generational-cache": "1.0.1", "@csstools/css-calc": "3.2.1", "@csstools/css-color-parser": "4.1.9", "@csstools/css-parser-algorithms": "4.0.0", "@csstools/css-tokenizer": "4.0.0" } }, "sha512-KVw6qIiCTUQhByfTd78h2yD1/00waTmm9uy/R7Ck/ctUyAPj+AEDLkQIdJW0T8+qGgj3j5bpNKK7Q3G+LedJWg=="],
-
-    "@asamuzakjp/dom-selector": ["@asamuzakjp/dom-selector@7.1.1", "", { "dependencies": { "@asamuzakjp/generational-cache": "1.0.1", "@asamuzakjp/nwsapi": "2.3.9", "bidi-js": "1.0.3", "css-tree": "3.2.1", "is-potential-custom-element-name": "1.0.1" } }, "sha512-67RZDnYRc8H/8MLDgQCDE//zoqVFwajkepHZgmXrbwybzXOEwOWGPYGmALYl9J2DOLfFPPs6kKCqmbzV895hTQ=="],
+    "@asamuzakjp/css-color": ["@asamuzakjp/css-color@6.0.7", "", { "dependencies": { "@csstools/css-calc": "^3.3.0", "@csstools/css-color-parser": "^4.1.10", "@csstools/css-parser-algorithms": "^4.0.0", "@csstools/css-tokenizer": "^4.0.0", "lru-cache": "^11.5.2" } }, "sha512-vC/bk1Lz7Tn/EfU9/apOTBk80/8dyGyWMowPoV1tJ52muDGsDqt2HPT2klrFUiY60MQmQv9q8yIht15JnBgDGw=="],
 
-    "@asamuzakjp/generational-cache": ["@asamuzakjp/generational-cache@1.0.1", "", {}, "sha512-wajfB8KqzMCN2KGNFdLkReeHncd0AslUSrvHVvvYWuU8ghncRJoA50kT3zP9MVL0+9g4/67H+cdvBskj9THPzg=="],
-
-    "@asamuzakjp/nwsapi": ["@asamuzakjp/nwsapi@2.3.9", "", {}, "sha512-n8GuYSrI9bF7FFZ/SjhwevlHc8xaVlb/7HmHelnc/PZXBD2ZR49NnN9sMMuDdEGPeeRQ5d0hqlSlEpgCX3Wl0Q=="],
+    "@asamuzakjp/dom-selector": ["@asamuzakjp/dom-selector@8.3.2", "", { "dependencies": { "bidi-js": "^1.0.3", "css-tree": "^3.2.1", "is-potential-custom-element-name": "^1.0.1", "lru-cache": "^11.5.2" } }, "sha512-93Z1N+BQNXysodoicpOIyNh2drHfz/CTf9nnT0FEx72GJcIiwgydD7tGAr78j41LsYn3hlRn+LdGPuBLn1Bl8Q=="],
 
     "@babel/code-frame": ["@babel/code-frame@8.0.0", "", { "dependencies": { "@babel/helper-validator-identifier": "8.0.2", "js-tokens": "10.0.0" } }, "sha512-dYYg153EyN2Ekbqw2zAsbd6/JR+9N2SEoC7YV2GyyqMM7x9bLDTjBD6XBhSMLH0wtIVyJj03jWNriQhaN+eoCw=="],
 
@@ -203,15 +210,15 @@
 
     "@chenglou/pretext": ["@chenglou/pretext@0.0.8", "", {}, "sha512-yqm2GMxnPI7VHcHwe84P8ZF0JK/2d2DMKPqMN+s95jQhwDMYYXKVFVJUMEaVWckQStdsjdLav/0Vu+d9YbtGxA=="],
 
-    "@csstools/color-helpers": ["@csstools/color-helpers@6.1.0", "", {}, "sha512-064IFJdjTfUqnjpCVpMOdbr8FLQBhinbZj6yRv2An2E41O/pLEXqfFRWqGq/SxlE5PEUYTlvWsG2r8MswAVvkg=="],
+    "@csstools/color-helpers": ["@csstools/color-helpers@6.1.1", "", {}, "sha512-gLNsunvwf3mCi5u5o46/Z/JcJMnhbHSaZ69rkgPzNM3J4s8hWwpPUQB6/tt0EDFyCiWzxANlx+2LJwpYj4zS1w=="],
 
-    "@csstools/css-calc": ["@csstools/css-calc@3.2.1", "", { "peerDependencies": { "@csstools/css-parser-algorithms": "4.0.0", "@csstools/css-tokenizer": "4.0.0" } }, "sha512-DtdHlgXh5ZkA43cwBcAm+huzgJiwx3ZTWVjBs94kwz2xKqSimDA3lBgCjphYgwgVUMWatSM0pDd8TILB1yrVVg=="],
+    "@csstools/css-calc": ["@csstools/css-calc@3.3.0", "", { "peerDependencies": { "@csstools/css-parser-algorithms": "^4.0.0", "@csstools/css-tokenizer": "^4.0.0" } }, "sha512-c5ihYsPkdG6JCkU2zTMm4+k6r7RXuGxtWYhu5DHMIiF1FHzrfmHL5so11AoFpUv/tu61xfcmT4AmKoFfMPoqdQ=="],
 
-    "@csstools/css-color-parser": ["@csstools/css-color-parser@4.1.9", "", { "dependencies": { "@csstools/color-helpers": "6.1.0", "@csstools/css-calc": "3.2.1" }, "peerDependencies": { "@csstools/css-parser-algorithms": "4.0.0", "@csstools/css-tokenizer": "4.0.0" } }, "sha512-paQcIaOO53Rk5+YrBaBjm/SgrV4INImjo2BT1DtQRYr+XeTRbeAYlS+jxXp9drqvKmtFnWRJKIalDLhZZDu42A=="],
+    "@csstools/css-color-parser": ["@csstools/css-color-parser@4.2.0", "", { "dependencies": { "@csstools/color-helpers": "^6.1.1", "@csstools/css-calc": "^3.3.0" }, "peerDependencies": { "@csstools/css-parser-algorithms": "^4.0.0", "@csstools/css-tokenizer": "^4.0.0" } }, "sha512-5+5LEmFuY1AjXdYhmgjTJogtQnP1evJ1zrBZGUNZ0thkpwnnmKxcHdAMn/OtFjAb25zA+jKDVYVRl+5G7rjv1A=="],
 
     "@csstools/css-parser-algorithms": ["@csstools/css-parser-algorithms@4.0.0", "", { "peerDependencies": { "@csstools/css-tokenizer": "4.0.0" } }, "sha512-+B87qS7fIG3L5h3qwJ/IFbjoVoOe/bpOdh9hAjXbvx0o8ImEmUsGXN0inFOnk2ChCFgqkkGFQ+TpM5rbhkKe4w=="],
 
-    "@csstools/css-syntax-patches-for-csstree": ["@csstools/css-syntax-patches-for-csstree@1.1.6", "", { "optionalDependencies": { "css-tree": "3.2.1" } }, "sha512-TcJCWFbXLPpJYq6z7bfOyjWYJDiDg2/I4gyUC9pqPNqHFRIey0EB0q0L5cSnQDfWJg8Jd6VadakxdIez/3zkqQ=="],
+    "@csstools/css-syntax-patches-for-csstree": ["@csstools/css-syntax-patches-for-csstree@1.1.8", "", { "peerDependencies": { "css-tree": "^3.2.1" }, "optionalPeers": ["css-tree"] }, "sha512-CpMLjAvwQg3BL5S0IeqsZNMH7EQrEWi0kLKOC13ZBF0ZwERiLWlibNPJr8G1kdU3Ms/r2KiNrF81pUh2HwAHdg=="],
 
     "@csstools/css-tokenizer": ["@csstools/css-tokenizer@4.0.0", "", {}, "sha512-QxULHAm7cNu72w97JUNCBFODFaXpbDg+dP8b/oWFAZ2MTRppA3U00Y2L1HqaS4J6yBqxwa/Y3nMBaxVKbB/NsA=="],
 
@@ -227,61 +234,61 @@
 
     "@emnapi/core": ["@emnapi/core@1.11.1", "", { "dependencies": { "@emnapi/wasi-threads": "1.2.2", "tslib": "2.8.1" } }, "sha512-RSvbQmHzdKzNsLYa/wHrbc3KN4sYLKAdPZxqiM2HATqv/SBk2/ENSHpvXGaLOMcsAyz0poEGqkmmKYG3OWiJEQ=="],
 
-    "@emnapi/runtime": ["@emnapi/runtime@1.11.1", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-vgj7R3y3Wgx24IQaGPA/R6YFXLHVMOZ0uVEyIQPaWs+rd1AzfEMXlAC22FYwO1XkKR6NPsq7mUandH8oIRdZFw=="],
+    "@emnapi/runtime": ["@emnapi/runtime@1.11.2", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-kyOl3X0DuTiT1h2ft8r2fYO8JYtU9a9Xis/zBSiGArNaagCOWx90N1k2wxp18czFDH+OgcWGb5ZP/XMt3dcyPA=="],
 
     "@emnapi/wasi-threads": ["@emnapi/wasi-threads@1.2.2", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-c95qOXkHdydNKhscBTebqEC1CVAZpyqOfVfBzQ1qgzyl3gfeldUjIggDbIZgDKsHLgnsM+igH7TJ/eAasaVuMA=="],
 
-    "@esbuild/aix-ppc64": ["@esbuild/aix-ppc64@0.28.1", "", { "os": "aix", "cpu": "ppc64" }, "sha512-Svl7tq8k/08+p6CXPpRjQ1fKX+1odH/BQbb48fV6fj3CWHhsoIOoY87w1oHXm0qEpkIK3ZfVgp0hed3XBXzXMQ=="],
+    "@esbuild/aix-ppc64": ["@esbuild/aix-ppc64@0.28.2", "", { "os": "aix", "cpu": "ppc64" }, "sha512-XExcO+dvLKvVtNTibSTBej1NCAbaGhWn9Ww1ZPx80qsahhPFe/8jgWP0IchNe0F3HwkU7n8ejhH8bjonqht8mQ=="],
 
-    "@esbuild/android-arm": ["@esbuild/android-arm@0.28.1", "", { "os": "android", "cpu": "arm" }, "sha512-0k2F129Xdio1TdJfzJ8sy1Q47vUD2NnwdhiAf7drUN1EBTfPf4hsFCtmMgu/6m8JSzsBrlmVjudMBQqOfG8usQ=="],
+    "@esbuild/android-arm": ["@esbuild/android-arm@0.28.2", "", { "os": "android", "cpu": "arm" }, "sha512-kXXoiPVVGQcnIYGOeaovwOURpniDBpSq4A03qkQ+BMQqtGG6HYap3xne9C1O1yo4TR3qxlCX5IqqmX6fFo2Lqg=="],
 
-    "@esbuild/android-arm64": ["@esbuild/android-arm64@0.28.1", "", { "os": "android", "cpu": "arm64" }, "sha512-34EGEbCIAgosYz6goLcopX6Mo7NyGv9tfwEM2/7Ce2VcVRk568iSvniGWcUXIy7wEDR1wzolcxcriFVrWYcwBg=="],
+    "@esbuild/android-arm64": ["@esbuild/android-arm64@0.28.2", "", { "os": "android", "cpu": "arm64" }, "sha512-5YfKeeI8qWfBZIX+u2xZC3Zlb3Os/gLS2sbEKM+I4ZOcsWmHS2WLysCcQZDAFRslDUU5Oiq44gf6PYN1vGwG5A=="],
 
-    "@esbuild/android-x64": ["@esbuild/android-x64@0.28.1", "", { "os": "android", "cpu": "x64" }, "sha512-dbwY7ltSMDWsRatcRpCnES4F+im88OCUgGZjy52shC7GqHRE/cYlxNbB4Z4UpJswpcc4Qxd2oE/ufM0p61IKng=="],
+    "@esbuild/android-x64": ["@esbuild/android-x64@0.28.2", "", { "os": "android", "cpu": "x64" }, "sha512-O387ite7SzUyCcy3JQX4P4bLtEA7bLLkx+esve5JHnyYfNTxcVpXZo9jhdB0lTKN44gztELTdU7nS8Nr16Fs1Q=="],
 
-    "@esbuild/darwin-arm64": ["@esbuild/darwin-arm64@0.28.1", "", { "os": "darwin", "cpu": "arm64" }, "sha512-TZbWkQY7kvTAXbXUT7uVACR5cMHsDiSz9z7ZKAX/RTq/WJEk3QyRr0wZpNhBDX+/0CtdqUIJlOiodQcta6tY3Q=="],
+    "@esbuild/darwin-arm64": ["@esbuild/darwin-arm64@0.28.2", "", { "os": "darwin", "cpu": "arm64" }, "sha512-n4KqkOQrraxHJcgjM1RvwbigfQKIKJVpM7xp+KsxiyUSrRdIXnt73VhrPAx0fV44hgfmIVKjxMN9J1t5jySVkw=="],
 
-    "@esbuild/darwin-x64": ["@esbuild/darwin-x64@0.28.1", "", { "os": "darwin", "cpu": "x64" }, "sha512-zfdzgK9ACBNZLI/CyHTOx81SyNbM6YXn7rxSgX97VjyiPl9W1i4Ka4fgKECEoFCKGpvBj5qArWIGgQjOwkgskQ=="],
+    "@esbuild/darwin-x64": ["@esbuild/darwin-x64@0.28.2", "", { "os": "darwin", "cpu": "x64" }, "sha512-uq6suIWYP37qzGddBKPw5QEQPi6HiLGsO7UmkpfyaYNQ3D+rN6w6WfwH+nuqcGXWvawGwxOEroO4YGnFh95azw=="],
 
-    "@esbuild/freebsd-arm64": ["@esbuild/freebsd-arm64@0.28.1", "", { "os": "freebsd", "cpu": "arm64" }, "sha512-wG2EA8ENdEI0qhkSZMjfqrdY+ziCYCPMmtZjjIwOmXFjmyzEHn+UUxk5of+SYsjtfs3VpnlC7QLzSI5hY/rOAw=="],
+    "@esbuild/freebsd-arm64": ["@esbuild/freebsd-arm64@0.28.2", "", { "os": "freebsd", "cpu": "arm64" }, "sha512-n+I0BTSRIoy+d6RPKnEVwql5UwBJolytvY4mAOIEJorKlqgPII8ix6slVVrfZ5Tnj7glIZvloylbB/EJPMWEXw=="],
 
-    "@esbuild/freebsd-x64": ["@esbuild/freebsd-x64@0.28.1", "", { "os": "freebsd", "cpu": "x64" }, "sha512-i7dZ9vQgnvSCzi/rYCXNgtF/U+eKZNJBzu3eTQbRgHnM7tNSizLOkRFAl3qzVc/Op/u5YkHHa4pf/3DOYHthLQ=="],
+    "@esbuild/freebsd-x64": ["@esbuild/freebsd-x64@0.28.2", "", { "os": "freebsd", "cpu": "x64" }, "sha512-78XJTJkvPs0kz2w61301PJjXl4g7q3JqiYMZ/M/yVI73EHBrCRTgkhu9oqG7vPqq+a/yadEW8aD+agKlk5xrmg=="],
 
-    "@esbuild/linux-arm": ["@esbuild/linux-arm@0.28.1", "", { "os": "linux", "cpu": "arm" }, "sha512-qVXBOHQS+d5Y722GwJzJUtOLlX7km3CraOaGormF1pDtPd2C/l1SHRPgjLunLGe51Sh5YYWKMFDyV4SxgMQYTQ=="],
+    "@esbuild/linux-arm": ["@esbuild/linux-arm@0.28.2", "", { "os": "linux", "cpu": "arm" }, "sha512-XlDnu2q5yoqems+xay6wSAcg9DDD7K9RLKZEBOMZm3ckNpJBvOX20tSfby8KfrrhINDyv9V2YVZKY/SpoGJI8w=="],
 
-    "@esbuild/linux-arm64": ["@esbuild/linux-arm64@0.28.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-yHs+0uc8+nvEAfAfxrWQKK5peSNzBc4PegcMO0EJ2hT71uA7vB8Ihg2e77R2P7SG5uYjPbHlLLmve4LLLRCf0g=="],
+    "@esbuild/linux-arm64": ["@esbuild/linux-arm64@0.28.2", "", { "os": "linux", "cpu": "arm64" }, "sha512-pW4AC0P3it8c7do9MVM4p51FzHzdM/TZrerurgRcHJ2WTa1VQ1CIq18xncfpBJw4ojkiZZrKW2yIBWBP92j6Ug=="],
 
-    "@esbuild/linux-ia32": ["@esbuild/linux-ia32@0.28.1", "", { "os": "linux", "cpu": "ia32" }, "sha512-d1z4ZuP0ajrfz/FhGT4vv278rX8KnPPJx8i5+AtK7TYbx9Le9F1hyzurZpkEyjkGa9dUGhQow4C1NmeGvqxN2w=="],
+    "@esbuild/linux-ia32": ["@esbuild/linux-ia32@0.28.2", "", { "os": "linux", "cpu": "ia32" }, "sha512-CYbnj78HsIeA+DhgUKgFCfvNsTHFhMMrinUrMZpDXJXKN8T3XViTZ/+wtHeVxEWY8ewSzTFN+nRmSwO2tZaLUQ=="],
 
-    "@esbuild/linux-loong64": ["@esbuild/linux-loong64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-M5sRjUVZrkm1OAPR3dlOYzNmN+loZKGVi1VUQGrwuqLcbR6qeAz+famMhjASeH3YVKvZz+zT1jlh/keC3Rj/lg=="],
+    "@esbuild/linux-loong64": ["@esbuild/linux-loong64@0.28.2", "", { "os": "linux", "cpu": "none" }, "sha512-buwkd8nsph4R+ajRvw0qM5Hja/TXQow3ptzWO2EbG/cqcIkHloRrdlBtQlshyYGTNFvfkfJ5tpPLVkY4DtsPfQ=="],
 
-    "@esbuild/linux-mips64el": ["@esbuild/linux-mips64el@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-mRObBZeHh2OxcBFPWE/FjylkRgZdYuiTR3vaTozquCGOH14iP9oN4x4Ge81CoIDYQrXmIxpFumJBu5MtZpnQJQ=="],
+    "@esbuild/linux-mips64el": ["@esbuild/linux-mips64el@0.28.2", "", { "os": "linux", "cpu": "none" }, "sha512-ZVykbDyk7519VwiNb9Lcj9m8XM6v5V9uKPvrEMkkEedVewf+0itkhahp4HDpgERXhwLRpWFypsGbG/J8s0QjJA=="],
 
-    "@esbuild/linux-ppc64": ["@esbuild/linux-ppc64@0.28.1", "", { "os": "linux", "cpu": "ppc64" }, "sha512-slScBsMAb3GFDcdrCgLwZtPYRoH2H/youv10QiZyRjmsP48fznoveWytSgCI/R0ZcUgpc0ZhIUEx6LHts8yrfQ=="],
+    "@esbuild/linux-ppc64": ["@esbuild/linux-ppc64@0.28.2", "", { "os": "linux", "cpu": "ppc64" }, "sha512-CAXl+Dtd9UUuJd8pKKdwh6MLm3MUMiqMPmhZ3tTSXPqfyQ3vDl6R5hZdZ/kYojK4ofXtdfSv1tFq8XzWx3heNQ=="],
 
-    "@esbuild/linux-riscv64": ["@esbuild/linux-riscv64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-kw0owk1o0GFETUJyW0jc0G4Yzs0BHZn0JDZ8JRT088vjJYX777BAs1fDGxAC+q831qOs2DTC96mNsG2opdfyyQ=="],
+    "@esbuild/linux-riscv64": ["@esbuild/linux-riscv64@0.28.2", "", { "os": "linux", "cpu": "none" }, "sha512-GeXCej4IQtU1B+QlDV8W/RRvbzI3O/Stss+/bCXv4lZls5WGRtu2a+3JkA3i4qIUlMXpcHebWpF8AkJhATowuA=="],
 
-    "@esbuild/linux-s390x": ["@esbuild/linux-s390x@0.28.1", "", { "os": "linux", "cpu": "s390x" }, "sha512-/lAIjX8aYFRByhh6L5rYtPEDRqa9de/4V/juOXcta5frjvzXO4/sqEtyytse0g3zZFuWu5cDN0MkLz2qRDD2Ag=="],
+    "@esbuild/linux-s390x": ["@esbuild/linux-s390x@0.28.2", "", { "os": "linux", "cpu": "s390x" }, "sha512-3H1weTYZPxt/WOhByszQZybS9w5lKzUn1FDMsgEChbHWQwHYQQRfBxgCcZvPhjHfKyJjIievvMmEUawJrdY9Dg=="],
 
-    "@esbuild/linux-x64": ["@esbuild/linux-x64@0.28.1", "", { "os": "linux", "cpu": "x64" }, "sha512-u/anNYF2mmVOEDwLtnQ1wOr3EZ9sTNGLWrsYGYwHWzGA3Si84IOkHXlbWTD1NB+9/1lcnweYKO54uhxZydNzfA=="],
+    "@esbuild/linux-x64": ["@esbuild/linux-x64@0.28.2", "", { "os": "linux", "cpu": "x64" }, "sha512-4xTZr1FUmSoQW4XIWmit3tzQrUTZM+N3P0XV8xROKYF50XfI7xeO90+1bZvNwxIufQ9hDQVRJH5YhgPVF8A/HQ=="],
 
-    "@esbuild/netbsd-arm64": ["@esbuild/netbsd-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-oks0DYbLwWMmaakTsCb+zL4E+aHRVLom9IJZOAthMQEPiQmydXHkziYEsGYRx0uNV/IjEKGAV941JzH02pflqw=="],
+    "@esbuild/netbsd-arm64": ["@esbuild/netbsd-arm64@0.28.2", "", { "os": "none", "cpu": "arm64" }, "sha512-sSATRjPeDBg3pdgHoQfoYBob11Kk1FGa9lui5RIHZCoCkJa9QKlvl3/vKz2usCmYYjs7ymJR/2Nnsqe+Hjt5nw=="],
 
-    "@esbuild/netbsd-x64": ["@esbuild/netbsd-x64@0.28.1", "", { "os": "none", "cpu": "x64" }, "sha512-aeL6lAnN89Hz43Mlh1G8ARasbuoYvSITDEx0tHh5b7jJnHcssqgjy9Yx430GDpmCa6OyrKoS0aNRjKundRizGg=="],
+    "@esbuild/netbsd-x64": ["@esbuild/netbsd-x64@0.28.2", "", { "os": "none", "cpu": "x64" }, "sha512-lqnzCV+mM0gIADaKihiCg6ifgfU2L3h5E33rNQBN1Y4MaVGnzryzmvvf7UHxprpQdE8hpqLolJ9Rl+SkIRDpyw=="],
 
-    "@esbuild/openbsd-arm64": ["@esbuild/openbsd-arm64@0.28.1", "", { "os": "openbsd", "cpu": "arm64" }, "sha512-MEFJe5C3R8pwXdZ5Y21oo6m7ePiS0d9pWucn99O/wvyJZChoIQKrQDxKrGeW8F5+T0okTHesAmDeiHDTIq0V/Q=="],
+    "@esbuild/openbsd-arm64": ["@esbuild/openbsd-arm64@0.28.2", "", { "os": "openbsd", "cpu": "arm64" }, "sha512-AL2qJILH7lNjrDmCQDvdxMfAUIv8KMNZOvrwAQ8i8//ntL9FflhOyMJ8OZSMBb8/AWXe3/5v5S20y3zCoZWKoQ=="],
 
-    "@esbuild/openbsd-x64": ["@esbuild/openbsd-x64@0.28.1", "", { "os": "openbsd", "cpu": "x64" }, "sha512-i/ZLIOafE0Z8cI/XANJAixoJL/uRAoS2xOA3rb0xN+KK0K177cMAsQYkzHtBrtMXAKuAc7HGgcWiZ/sRC1Nxgw=="],
+    "@esbuild/openbsd-x64": ["@esbuild/openbsd-x64@0.28.2", "", { "os": "openbsd", "cpu": "x64" }, "sha512-QtiuPytchRyC4rwUKhexJdQKvDuZ6hWloi3igqPQNUJCS1/v9EiO3UTOXR6A3FoMo4fnAKbWJdqaIwhOzh8qEw=="],
 
-    "@esbuild/openharmony-arm64": ["@esbuild/openharmony-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-ge+Z7EXFNt2BO1oAMsVpiQ8EwndV9i1xXerAeTIK7AtPs3bKFXQM7nlRxDSIUIMeueR1CNXxqztLzdNeReKBJg=="],
+    "@esbuild/openharmony-arm64": ["@esbuild/openharmony-arm64@0.28.2", "", { "os": "none", "cpu": "arm64" }, "sha512-WkhYDmpTjLvGlScA1rwjRUmhl4k8oXR3cIbtqWmELgU/dFeHHlEllxDvdWcNJV9rbzCexB5vz8gtNewWLgCT7Q=="],
 
-    "@esbuild/sunos-x64": ["@esbuild/sunos-x64@0.28.1", "", { "os": "sunos", "cpu": "x64" }, "sha512-BEjgtECkL3vY+SaSQ6nzVfiALUeFxpawyp8Jmf5PtYhf1Ug40N1h/hxlhts+f1FvSvarEigdxS3BlSMI2PJLcQ=="],
+    "@esbuild/sunos-x64": ["@esbuild/sunos-x64@0.28.2", "", { "os": "sunos", "cpu": "x64" }, "sha512-GPMSkTOtMnv2U2F8gxe4Io6qmVs+YKyp832Etqqxr0hFngmXQ3rzwytelm3GIn7T4VviRUlf3sOgBOiTdvaf7g=="],
 
-    "@esbuild/win32-arm64": ["@esbuild/win32-arm64@0.28.1", "", { "os": "win32", "cpu": "arm64" }, "sha512-lCv9eK/H6ZJWbE7bh2nw54CZ9M2nupBxJcTsdk/QQnWkdSjKGuxmmH8/GWrlT1eMmZfn4dGcCjRte397WqfQXA=="],
+    "@esbuild/win32-arm64": ["@esbuild/win32-arm64@0.28.2", "", { "os": "win32", "cpu": "arm64" }, "sha512-PIhhEkE9uPBleRBrQEJpUn7MBnibZzbGzYWPmY3x+YoVg/95zbjB4CxPPOQ8l5tYYM4mMaCthF8/1DIfBQQyWQ=="],
 
-    "@esbuild/win32-ia32": ["@esbuild/win32-ia32@0.28.1", "", { "os": "win32", "cpu": "ia32" }, "sha512-zvb/mB2bSCoJOpoCBgYKKpX6YM6mJBlBUVUtVj41DlZJVEB6/0CKlRYxP5wWl1C1ILiCoAU5wZZ4q1P3qeS6Eg=="],
+    "@esbuild/win32-ia32": ["@esbuild/win32-ia32@0.28.2", "", { "os": "win32", "cpu": "ia32" }, "sha512-YmJbfTlvU7Sdn9BB+4PRES4oB6pxgS37MAONj+hBr/cpXS1aBPKXxNnDbu+QCWPj0o9dgyxeq79g6c5P8KeuYA=="],
 
-    "@esbuild/win32-x64": ["@esbuild/win32-x64@0.28.1", "", { "os": "win32", "cpu": "x64" }, "sha512-bm4Mowrv+GXMlpWX++EcXw/iLyd1o3+bJkC2DkWXYVvgZCqD/bSj9ctZeAMC3cIxgjRVR2Dufaiu4YPxr5gW1A=="],
+    "@esbuild/win32-x64": ["@esbuild/win32-x64@0.28.2", "", { "os": "win32", "cpu": "x64" }, "sha512-5ebpxr3nWMzrL/rnUI755Jkuee0bHL/Gq0WTF9lvcpv73wAp5eu8MfBUgWK9bhWvZjj7yX8etf/8tI8Ney695g=="],
 
     "@exodus/bytes": ["@exodus/bytes@1.15.1", "", {}, "sha512-S6mL0yNB/Abt9Ei4tq8gDhcczc4S3+vQ4ra7vxnAf+YHC02srtqxKKZghx2Dq6p0e66THKwR6r8N6P95wEty7Q=="],
 
@@ -367,31 +374,31 @@
 
     "@mdx-js/react": ["@mdx-js/react@3.1.1", "", { "dependencies": { "@types/mdx": "2.0.13" }, "peerDependencies": { "@types/react": "19.2.17", "react": "19.2.8" } }, "sha512-f++rKLQgUVYDAtECQ6fn/is15GkEH9+nZPM3MS0RcxVqoTfawHvDlSCH7JbMhAM6uJ32v3eXLvLmLvjGu7PTQw=="],
 
-    "@modelcontextprotocol/sdk": ["@modelcontextprotocol/sdk@1.29.0", "", { "dependencies": { "@hono/node-server": "2.0.11", "ajv": "8.20.0", "ajv-formats": "3.0.1", "content-type": "1.0.5", "cors": "2.8.6", "cross-spawn": "7.0.6", "eventsource": "3.0.7", "eventsource-parser": "3.0.8", "express": "5.2.1", "express-rate-limit": "8.5.2", "hono": "4.12.32", "jose": "6.2.3", "json-schema-typed": "8.0.2", "pkce-challenge": "5.0.1", "raw-body": "3.0.2", "zod-to-json-schema": "3.25.2" }, "peerDependencies": { "zod": "4.4.3" } }, "sha512-zo37mZA9hJWpULgkRpowewez1y6ML5GsXJPY8FI0tBBCd77HEvza4jDqRKOXgHNn867PVGCyTdzqpz0izu5ZjQ=="],
+    "@modelcontextprotocol/sdk": ["@modelcontextprotocol/sdk@1.30.0", "", { "dependencies": { "@hono/node-server": "^1.19.9 || ^2.0.5", "ajv": "^8.17.1", "ajv-formats": "^3.0.1", "content-type": "^1.0.5", "cors": "^2.8.5", "cross-spawn": "^7.0.5", "eventsource": "^3.0.2", "eventsource-parser": "^3.0.0", "express": "^5.2.1", "express-rate-limit": "^8.2.1", "hono": "^4.11.4", "jose": "^6.1.3", "json-schema-typed": "^8.0.2", "pkce-challenge": "^5.0.0", "raw-body": "^3.0.0", "zod": "^3.25 || ^4.0", "zod-to-json-schema": "^3.25.1" }, "peerDependencies": { "@cfworker/json-schema": "^4.1.1" }, "optionalPeers": ["@cfworker/json-schema"] }, "sha512-xKd8OIzlqNzcqcNumGAa6g+PW2kjD5vrpcKOnfldAUPP3j7lnqMPwlTXQm8gF+UwH72z0lqaRbjr9hqGz0eITA=="],
 
     "@napi-rs/wasm-runtime": ["@napi-rs/wasm-runtime@1.1.6", "", { "dependencies": { "@tybys/wasm-util": "0.10.3" }, "peerDependencies": { "@emnapi/core": "1.11.1", "@emnapi/runtime": "1.11.1" } }, "sha512-ZLv/JdUfkvOy9eCnnBaGfiO+XimbjebAeO+MRQqD/B+FR1tnRN0tpKSJHRbE8sFfS6aqsXZ67TQjfwfsxULVbg=="],
 
     "@neondatabase/serverless": ["@neondatabase/serverless@1.1.0", "", {}, "sha512-r3ZZhRjEcfEdKIZnoB1RusNgvHuaBRqfCzV4Gi+5A9yUX0S4HTws/ASWqt13wL4y4I+0rqsWGdA2w7EQXHi3+Q=="],
 
-    "@next/env": ["@next/env@16.2.12", "", {}, "sha512-d0Z5Bc13Fa4nR8pFAKx2jay2yhJM16vlfHbTzYnUQAxlNb6B6lmn4hjt69lYNt4kRtyYP6gEM49lPRHNbIyneg=="],
+    "@next/env": ["@next/env@16.3.1", "", {}, "sha512-35G3xwkQUb2oETSDjFXGrVugknoayLFBh7vSE+yAcl9IP2zT9wyGwq7297AYHR11kJld807t5f8AJBs6WBzXsQ=="],
 
-    "@next/swc-darwin-arm64": ["@next/swc-darwin-arm64@16.2.12", "", { "os": "darwin", "cpu": "arm64" }, "sha512-0W1R0teHWJrqKX0FH20IzzIWAOuGtBxPGuObrxy1lE8hQvCFj49KE8a3WUg0D7sq6rn6zkM4c7YGUnhudBS6oA=="],
+    "@next/swc-darwin-arm64": ["@next/swc-darwin-arm64@16.3.1", "", { "os": "darwin", "cpu": "arm64" }, "sha512-ABMIu2zQ7cnNIHm5ivKGwZwUrm0pAai3yiJ/gK/rF1c1VP9UOnj7XECbMKFdVKp9I9eMYq9NoDs1WXOoowxzJw=="],
 
-    "@next/swc-darwin-x64": ["@next/swc-darwin-x64@16.2.12", "", { "os": "darwin", "cpu": "x64" }, "sha512-Hy5Ls099+aFUmOLmIgPfLqNi6iCwhL3uQCssz5rWk+5Nkc6TUKCE83DY5BbNylfm3+mfwcSFnLRfrZDJhVxdtw=="],
+    "@next/swc-darwin-x64": ["@next/swc-darwin-x64@16.3.1", "", { "os": "darwin", "cpu": "x64" }, "sha512-gNG21e/UnrroeScbY/QndUEdl0mF1FRibW7BBeYUz/5ABCepjqDdEdgr592vpzMtCn/m7FTjYq3TN4TpyDnutw=="],
 
-    "@next/swc-linux-arm64-gnu": ["@next/swc-linux-arm64-gnu@16.2.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-+YqU2h1cQkHsGfvjAsrSmst8UIFBibBGm5x3Xgel8NLMiDQtNOM4sM2GOEMvG5YiOBNeN/Ykk8cQC2S0Xrqljg=="],
+    "@next/swc-linux-arm64-gnu": ["@next/swc-linux-arm64-gnu@16.3.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-6B6Lw016iwNUQuaJoraMMTLh6TwHzFUtxipSScD1F3YyymcrRWkobodRS2ftIOkF5vrs4zNlyUrTC5YZQ9Lz5w=="],
 
-    "@next/swc-linux-arm64-musl": ["@next/swc-linux-arm64-musl@16.2.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-0qjhiYBaKAqF63LA1ZWAAnKTzFUguAaZiRa5etMLGGPj/B6uEVjtIZldIzFEp3wHlB0koK6aTzqPtSdplTCjoA=="],
+    "@next/swc-linux-arm64-musl": ["@next/swc-linux-arm64-musl@16.3.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-JUiPXZKK9wOhjf4MgDiH29GZLxfqOesbLtHq2pDxwH/WwscTRV2ToymnOTh1egzaZf0ueUf8T2+CeYTGHjW0Iw=="],
 
-    "@next/swc-linux-x64-gnu": ["@next/swc-linux-x64-gnu@16.2.12", "", { "os": "linux", "cpu": "x64" }, "sha512-7A3q26W+h7gnA15uqBToNuDqBEFZZcqh0mW2mn4AJh/G5pdg2RVE3n4slzLEliASZFG3NmsbEzng/x2Sh09mBg=="],
+    "@next/swc-linux-x64-gnu": ["@next/swc-linux-x64-gnu@16.3.1", "", { "os": "linux", "cpu": "x64" }, "sha512-Uog9jsrmIRIL/lfvIp9htmskSNC7JcQsMVucXL2V2YY1y/D9IUN3LPEafqy0zRJ2cIU1SQ0V6F6TlffQ+pLAGg=="],
 
-    "@next/swc-linux-x64-musl": ["@next/swc-linux-x64-musl@16.2.12", "", { "os": "linux", "cpu": "x64" }, "sha512-qSjL/uppm+cbh21s72Ss8gkiOhQ4dExWHNGOWy6eZV7STj5WsKehgxT61beSsOj+YYQuTplL376lOCdMQU5T8w=="],
+    "@next/swc-linux-x64-musl": ["@next/swc-linux-x64-musl@16.3.1", "", { "os": "linux", "cpu": "x64" }, "sha512-6yy3FT13KgUFOj5H8bl8w/6nKiJwHIvbtwh1V+1acsu+7y4tJjnemSa6mhsh53BeoVrlozE+fMgZhXH46WmjMA=="],
 
-    "@next/swc-win32-arm64-msvc": ["@next/swc-win32-arm64-msvc@16.2.12", "", { "os": "win32", "cpu": "arm64" }, "sha512-X6hzsOUJac/e7AWSbn9gQ9nzHld1xWP5iyjHpYWvud8pufB679O1xg4JDyKr8Xd69Jvd+kM2Der6uftiZCmjYA=="],
+    "@next/swc-win32-arm64-msvc": ["@next/swc-win32-arm64-msvc@16.3.1", "", { "os": "win32", "cpu": "arm64" }, "sha512-iOoN1QecUoGNZik536U/vtK43YwgyrCsGIkth52yIkl612n+0C9MjSnJbQAikISpb+WYRooBVhaDlUW7iZoKog=="],
 
-    "@next/swc-win32-x64-msvc": ["@next/swc-win32-x64-msvc@16.2.12", "", { "os": "win32", "cpu": "x64" }, "sha512-F6fakeHuFTLOPt0bslQJdf+xtT+WIP9DVn/m4y1w1mRnVPyh3D/cNvzlRkxM444xfm+IvvYNSOrKiA2CDJ0Uxw=="],
+    "@next/swc-win32-x64-msvc": ["@next/swc-win32-x64-msvc@16.3.1", "", { "os": "win32", "cpu": "x64" }, "sha512-d/k+PpAriUPaeMJJOG7HUSdqfEX46FEPWU1p3/nm2ACmXhj9hFEWdFODUBIpkuijXYkfL90qZzTqVPRp4BW/hw=="],
 
-    "@noble/ciphers": ["@noble/ciphers@2.2.0", "", {}, "sha512-Z6pjIZ/8IJcCGzb2S/0Px5J81yij85xASuk1teLNeg75bfT07MV3a/O2Mtn1I2se43k3lkVEcFaR10N4cgQcZA=="],
+    "@noble/ciphers": ["@noble/ciphers@2.3.0", "", {}, "sha512-Clu/xdfgVTf9o7ngLOURaxePwR0j8sjclKEtVij10/jGulwFsPWCvvRgG/XjUVf8Nei+jLG6uwyXzUTGY1DQrw=="],
 
     "@nodelib/fs.scandir": ["@nodelib/fs.scandir@2.1.5", "", { "dependencies": { "@nodelib/fs.stat": "2.0.5", "run-parallel": "1.2.0" } }, "sha512-vq24Bq3ym5HEQm2NKCr3yXDwjc7vTsEThRDnkp2DK9p1uqLR+DHurm/NOTo0KG7HYHU7eppKZj3MyqYuMBf62g=="],
 
@@ -401,108 +408,104 @@
 
     "@opentelemetry/api": ["@opentelemetry/api@1.9.1", "", {}, "sha512-gLyJlPHPZYdAk1JENA9LeHejZe1Ti77/pTeFm/nMXmQH/HFZlcS/O2XJB+L8fkbrNSqhdtlvjBVjxwUYanNH5Q=="],
 
-    "@oxc-project/types": ["@oxc-project/types@0.139.0", "", {}, "sha512-r9gHphtCs+1M7J0pw6Sn/hh/Wpa/iQrOOkrNAlVLF/gHq+/CJmHIWKKUUhdWjcD6CIa8idarspCsASiXCXvFUw=="],
+    "@oxc-project/types": ["@oxc-project/types@0.144.0", "", {}, "sha512-nuhZIOLuI6TFQ32I/WnUx+SCPY7SdSKwgnFHydAuoS1+Z4BRcaP+RRJmGzl9lw+0OFF7UmaESf7KQRXaNLHypg=="],
 
-    "@oxfmt/binding-android-arm-eabi": ["@oxfmt/binding-android-arm-eabi@0.60.0", "", { "os": "android", "cpu": "arm" }, "sha512-1q4q4Jc8FlOMVojEisyFAVyl8h1yawNv6phjgmhGVEDeyeOdsSnSr9x0+D4mOnEKvpO5L4mxKZ/DP9X6U3A/Mw=="],
+    "@oxfmt/binding-android-arm-eabi": ["@oxfmt/binding-android-arm-eabi@0.63.0", "", { "os": "android", "cpu": "arm" }, "sha512-YmRth4ZPGgEXcgmkhvANbC9uD67dxmSobW7DQuyt5tOBOKvPnIpk5SVHBj88E+7wMNRI2FhqaDbOhQFBix+b8A=="],
 
-    "@oxfmt/binding-android-arm64": ["@oxfmt/binding-android-arm64@0.60.0", "", { "os": "android", "cpu": "arm64" }, "sha512-tD41I6nCt9k8SQXft0CSjjU9jg6SwG7uMu7PxodSEHXl+GDW0868oy6tTtoJkyUze8YKFgTpz/k5LuPUnFiGLw=="],
+    "@oxfmt/binding-android-arm64": ["@oxfmt/binding-android-arm64@0.63.0", "", { "os": "android", "cpu": "arm64" }, "sha512-icbahX8X2X3sRamOMecvdYeZXWjPDazRDIfvWfy7Ca1nc/ZDT2Y9k5Nt7s46EqFd7NQPdgk+CM3/SgIT5LPCaQ=="],
 
-    "@oxfmt/binding-darwin-arm64": ["@oxfmt/binding-darwin-arm64@0.60.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-TTpzPug96Zxdyb46KvTyIUQDdsqbumXh2TKG9C23PCT0kF7JkW56Z/quPuG9rqOFKQIi1gpRNZ7DX18LwxXPnw=="],
+    "@oxfmt/binding-darwin-arm64": ["@oxfmt/binding-darwin-arm64@0.63.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-WV+Ze5v5gI2qoj8jpAovt8KBTW8pjEz/AiMXXjeTQS+Bmf/MmZXTS40S8xNPDszX+W8WDv2Bbk6qKrMTtUGu1A=="],
 
-    "@oxfmt/binding-darwin-x64": ["@oxfmt/binding-darwin-x64@0.60.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-CnOoWgQ7L+JL/YQaRJ+NyATciSfcftncm7y3kqyte1cGtFEGnStaCd1TAyrinkfQ7nRBfHrTs1/vTwUJr3WF2Q=="],
+    "@oxfmt/binding-darwin-x64": ["@oxfmt/binding-darwin-x64@0.63.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-CJGSBdDxXOWIpoFXHpverimCvz084KA7L483rqJ44c3jDtzv6d4qOSoR/V9ywSHfV+Ks1lwIj2P49BFhunLNAA=="],
 
-    "@oxfmt/binding-freebsd-x64": ["@oxfmt/binding-freebsd-x64@0.60.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-ychJo7S3hZxdO6eDZ9zM6F2lM9fpJS3EKS5CAUSWyprdLYxTu4gbaUKV/VBPTcMJwQa2Bpo+643y3OJ537pihA=="],
+    "@oxfmt/binding-freebsd-x64": ["@oxfmt/binding-freebsd-x64@0.63.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-BDfKY+KhL2078cgswBBFQPAYuxCy93bS/iC5frdSeSbTLcGrR6VC2hsuPTanoJmg84+wSyWl0wWC1eR+uTnkRg=="],
 
-    "@oxfmt/binding-linux-arm-gnueabihf": ["@oxfmt/binding-linux-arm-gnueabihf@0.60.0", "", { "os": "linux", "cpu": "arm" }, "sha512-36IH5o55T2Fx7E0feDttt+mifxN6yk9pWv4KfhAIsP0dFnUq27331OwbpOsZdoXF9soOLWm7mQUz5+UUmyec4g=="],
+    "@oxfmt/binding-linux-arm-gnueabihf": ["@oxfmt/binding-linux-arm-gnueabihf@0.63.0", "", { "os": "linux", "cpu": "arm" }, "sha512-Ov1cQEXT4mj7cojAokWSS1eoxkoyvbDfAbxNsGIKY2o36kvdAaFzPxRN6NxFRk9fD72B8oCoTTX/NuYTUWlpsg=="],
 
-    "@oxfmt/binding-linux-arm-musleabihf": ["@oxfmt/binding-linux-arm-musleabihf@0.60.0", "", { "os": "linux", "cpu": "arm" }, "sha512-G1Ve7lAa6sFBolVI2LWHfEAqy0YKh4vnioH8uYO9kAEdgM7mR40IksIx9/Zk4+vbYew/sGa4J9Q4tZ3n9gXDHA=="],
+    "@oxfmt/binding-linux-arm-musleabihf": ["@oxfmt/binding-linux-arm-musleabihf@0.63.0", "", { "os": "linux", "cpu": "arm" }, "sha512-0LE7ro3+6L79jcMANycAZfRaC7zxr9YZ2+vEL5uMD9QlEep+rS/r1kSJsnuLl991NXJZD60euh0PC1GHrR20vw=="],
 
-    "@oxfmt/binding-linux-arm64-gnu": ["@oxfmt/binding-linux-arm64-gnu@0.60.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-LTQdRBf6uzj/h7Xk6lKzbGD2hrF/fK4YI9LIN1c0509tPUn8wRa3mCmrFQpEWJPLYGFrLFFMTYW1Ljj6VqW2Hw=="],
+    "@oxfmt/binding-linux-arm64-gnu": ["@oxfmt/binding-linux-arm64-gnu@0.63.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-izPk+2Z4gjuZK32Fqh5qXoMpT/2NXzLh++ob57HiEiVSQZ1iYXu8EKMzb+K5AvWyIEXhdDIt7ADjGGtFhkT9Bw=="],
 
-    "@oxfmt/binding-linux-arm64-musl": ["@oxfmt/binding-linux-arm64-musl@0.60.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-2JMo3XPxMPx3hiqddSZYyaH+fKJm6cz0u8n1naYjP/CdOQOZW34i8lKBUfmbWiuFvd6KoYXLmhAyBuvojsYS7Q=="],
+    "@oxfmt/binding-linux-arm64-musl": ["@oxfmt/binding-linux-arm64-musl@0.63.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-alPmbOuWXFXiSo+lOtv6X71C7SYMEDW2WVvywOvf9BwKgEhSNGhMTLeFVSjKUMCamcjbbgVdsWF8GN1uy8xshg=="],
 
-    "@oxfmt/binding-linux-ppc64-gnu": ["@oxfmt/binding-linux-ppc64-gnu@0.60.0", "", { "os": "linux", "cpu": "ppc64" }, "sha512-L3C+nBD13lr306tr/PjM3RMll+BVqgFrIgUyoeHuai5oueJrRLgO3j+GO5/Cbhtkf5PSlHYTI1JY7iqBd1qa6A=="],
+    "@oxfmt/binding-linux-ppc64-gnu": ["@oxfmt/binding-linux-ppc64-gnu@0.63.0", "", { "os": "linux", "cpu": "ppc64" }, "sha512-BdzCPvolJc4AWZ+YMzgUDJcDzbQWrFjYuqBHoNHNqP1aCaluQRJNs4k3vNU5IG7vTpjf9zeD73D7MFM1TecZpg=="],
 
-    "@oxfmt/binding-linux-riscv64-gnu": ["@oxfmt/binding-linux-riscv64-gnu@0.60.0", "", { "os": "linux", "cpu": "none" }, "sha512-M4MsmvqlxFiPtSRGyBYQSZxchEf463AOyd+Dh4/9xDpjWBsRtDUTDMFN5EdHinjVK1/eDJQ8MLpcYjpYayaCnA=="],
+    "@oxfmt/binding-linux-riscv64-gnu": ["@oxfmt/binding-linux-riscv64-gnu@0.63.0", "", { "os": "linux", "cpu": "none" }, "sha512-7sIgfLzqtNKSkMGsGVyRpHwpjNezRg2XONvUOheFZs95TSZpM0JAuPpA8KrQFsWc4wPU95roX2O69JgH8igOgw=="],
 
-    "@oxfmt/binding-linux-riscv64-musl": ["@oxfmt/binding-linux-riscv64-musl@0.60.0", "", { "os": "linux", "cpu": "none" }, "sha512-OH+9UskYuxRB+GxqdGkVN8f5UpwhqG8YscNo1wl8+KJ62cd7wZdGga6iGLJIf8kibF1WBwvlfDUx3cez/VXwFg=="],
+    "@oxfmt/binding-linux-riscv64-musl": ["@oxfmt/binding-linux-riscv64-musl@0.63.0", "", { "os": "linux", "cpu": "none" }, "sha512-9Tcg0y0WcVa6Mm9AgcgFMseDS+VkFJZpKZ8We9SpDY4gg5jewSwln+0sO04QLcTS1BtfDl9MwR+NfID8L7PUTg=="],
 
-    "@oxfmt/binding-linux-s390x-gnu": ["@oxfmt/binding-linux-s390x-gnu@0.60.0", "", { "os": "linux", "cpu": "s390x" }, "sha512-y7AAFutt9wFWBFOAn6+BHaV39usZmcr3YYH2385f+NHgPNpIF9HpqKp0jgUxPaUOCyG3oaX5VhJduL1Nw164rw=="],
+    "@oxfmt/binding-linux-s390x-gnu": ["@oxfmt/binding-linux-s390x-gnu@0.63.0", "", { "os": "linux", "cpu": "s390x" }, "sha512-qWKC1pEOpx1qYhXaugPhHUeXwSfqEOk2wJH2LqVXGPV5iQYfdAZdt+d2XDiX4DTSWA2QDMUcFB+wEORh3Xn/sA=="],
 
-    "@oxfmt/binding-linux-x64-gnu": ["@oxfmt/binding-linux-x64-gnu@0.60.0", "", { "os": "linux", "cpu": "x64" }, "sha512-yKZ9+CXAI+1RO5nH/4Z/9M6DAsfOzd5bw/gtWk81KB4mpalMaRRSXfouc5/tHxazDmBek55HNPepNYBgaCew0Q=="],
+    "@oxfmt/binding-linux-x64-gnu": ["@oxfmt/binding-linux-x64-gnu@0.63.0", "", { "os": "linux", "cpu": "x64" }, "sha512-S9wXYOiGSqYGS4Fx/TFsY+xDd/7dE5s+rUgbA4TsHiVF9e8J3ZcKmP7dsP/7iqLI9Wz7Ic7TzEr3mdthRCTdrA=="],
 
-    "@oxfmt/binding-linux-x64-musl": ["@oxfmt/binding-linux-x64-musl@0.60.0", "", { "os": "linux", "cpu": "x64" }, "sha512-bCUGaF6hJOYnQzLJdHLZbvGsOd5oSvGAyJhPAKum2uyLYUuXmP8vqg690DWi2hqcnIoYpqSqCrjzE5aiUAgwQg=="],
+    "@oxfmt/binding-linux-x64-musl": ["@oxfmt/binding-linux-x64-musl@0.63.0", "", { "os": "linux", "cpu": "x64" }, "sha512-5eGyTJuMZNwBSHCivXt8Yuta6GeTYksOPXRk2MIhajiyFGQx7bjaHIwY+ZusAoFHhT157A9x6sktLjYo9D5oMQ=="],
 
-    "@oxfmt/binding-openharmony-arm64": ["@oxfmt/binding-openharmony-arm64@0.60.0", "", { "os": "none", "cpu": "arm64" }, "sha512-GrUeZOvzP30ExxfCuQiyofuUGI+OmvAgFwOO5w5p9mGPlxcyuqI+6Sy9fAKFFfLQrqKYWFgc5sYA2Unj/29nPg=="],
+    "@oxfmt/binding-openharmony-arm64": ["@oxfmt/binding-openharmony-arm64@0.63.0", "", { "os": "none", "cpu": "arm64" }, "sha512-Rz7hx+Dv3DoW/S6pwVAyjfFXp7/trdQ1zg+vNmsdsdDNlUccugp4XNqambSuEAeP0DaG9k72AtNyfDXCEg0AGw=="],
 
-    "@oxfmt/binding-win32-arm64-msvc": ["@oxfmt/binding-win32-arm64-msvc@0.60.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-WD4Q954kUl2TDJV/6q7UnE2rlKk047kXLJsr4bJ2mXRaAqNXcmV3nwKUsGCc3mz/jYDBnXtJEaBErJEybK8iQQ=="],
+    "@oxfmt/binding-win32-arm64-msvc": ["@oxfmt/binding-win32-arm64-msvc@0.63.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-T/IuizKN9mr4Xw6YYnptkXRNdLkyIlUZ7c8zfTOBpoytZyJ1BAsMUvsMDEx0X4YvSMpaivm+DR8112rQfzC25g=="],
 
-    "@oxfmt/binding-win32-ia32-msvc": ["@oxfmt/binding-win32-ia32-msvc@0.60.0", "", { "os": "win32", "cpu": "ia32" }, "sha512-HqDekjr8JXzVDUP1YthDZ1Y3CBEcuZT4WX3B+1kaxj8CvZA8Y2YhcEsXqoSop3tVsgjACxjnFQFDkBo0r/jq1Q=="],
+    "@oxfmt/binding-win32-ia32-msvc": ["@oxfmt/binding-win32-ia32-msvc@0.63.0", "", { "os": "win32", "cpu": "ia32" }, "sha512-XjrO5FJ5Wl9vsAxtCP1G/eaeT6y1K2s9CICUHGE42cEjou32/J6S+B1KnrOAboj6E7uhJnwPbRSvznWcxNdA0g=="],
 
-    "@oxfmt/binding-win32-x64-msvc": ["@oxfmt/binding-win32-x64-msvc@0.60.0", "", { "os": "win32", "cpu": "x64" }, "sha512-tz78yhmGPKboTMHCHSaUqXK8JrmoSejgDcWeqAtg2s07ZGKQ3rH5Jn8NuXPGNG33CDbY2e9NoQWXIVEmKO21Rw=="],
+    "@oxfmt/binding-win32-x64-msvc": ["@oxfmt/binding-win32-x64-msvc@0.63.0", "", { "os": "win32", "cpu": "x64" }, "sha512-sgsHCQy432OTQH4Ikk3tZptp3GqwnhwUDuY0loBH41zyHWfMZY9v8Dy78wsnSofHejvFozZGgJgBB1A0LQRwMQ=="],
 
-    "@oxlint/binding-android-arm-eabi": ["@oxlint/binding-android-arm-eabi@1.75.0", "", { "os": "android", "cpu": "arm" }, "sha512-lutovtFzJqlRaqpZrCqSSGaHZzl9nIxxpjLzhSRLunN6dCLylj0uzlCyQGaQDIys7rrv8kVXiFO+R4Zpn0bX7g=="],
+    "@oxlint/binding-android-arm-eabi": ["@oxlint/binding-android-arm-eabi@1.78.0", "", { "os": "android", "cpu": "arm" }, "sha512-Bu819lmAfZMUHErrpe0cEWj3iaefuUODHSU8+UbXy67V/r7/7f4K3FL0NmbD85E+wiFLDYuhP8Zlv0XnVeXshw=="],
 
-    "@oxlint/binding-android-arm64": ["@oxlint/binding-android-arm64@1.75.0", "", { "os": "android", "cpu": "arm64" }, "sha512-hXI0hDgHkw4w5nfru72aG7y+2iQJmC4waH/KV6H/hbgA6yAP5jYNx0P9yug15Hs0tWl/+mda3Jjn/2gmDT48tw=="],
+    "@oxlint/binding-android-arm64": ["@oxlint/binding-android-arm64@1.78.0", "", { "os": "android", "cpu": "arm64" }, "sha512-CDfxZgB61B7buRdY2FJoAYYPPXCZ1EoC1LKscnC5dg3kjobdxiconvAvvN1BmHyW4PyFT3jRLDag/BY/roSNBQ=="],
 
-    "@oxlint/binding-darwin-arm64": ["@oxlint/binding-darwin-arm64@1.75.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-D91BWbK/dMYfCcrghspPIuKs2D9LF4Z/OabVSQjw1AO6PWxArD7teDA48bm0ySFqWDaPVqmQRl5GMWNglTXyrQ=="],
+    "@oxlint/binding-darwin-arm64": ["@oxlint/binding-darwin-arm64@1.78.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-2Y2U9Ahrz+OO0Ej88f9SJYq51/jUBp1Mc7iZu0ukrbeeZ3gpRGfzIFnoqfHDY96xr0GEfNrPUBFEy0nN5aD7HA=="],
 
-    "@oxlint/binding-darwin-x64": ["@oxlint/binding-darwin-x64@1.75.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-02mpwzf12BonZ6PT0TuQoomvEh2kVl2WGBIKWezCyToIS+rYkQZ6GXnARBAl9A4Ovm2V+Xe7M4KretyqmmcnJQ=="],
+    "@oxlint/binding-darwin-x64": ["@oxlint/binding-darwin-x64@1.78.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-rpych6eJq6m9jDRypTEaPD1xysaEW5h9+xuxhGK/QhOg+/xaqPZrCrTNoIl/f3nEjuJeCEmstNDlrE9rJi/3/g=="],
 
-    "@oxlint/binding-freebsd-x64": ["@oxlint/binding-freebsd-x64@1.75.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-qZJgLnDaBsiL5YESx2t/TZ8eXkL9fEkKoXEdzegROhlz9A0lgyGnZ0dAzJrh7LJAHQl2K9RdRueN2s/9N7+odg=="],
+    "@oxlint/binding-freebsd-x64": ["@oxlint/binding-freebsd-x64@1.78.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-IcMGrQT3QizkOESUJd5et+rOhVqSkNDfNik1cvrKDqIbzqx9KMtRswpFgkCuNTSwylCFLKhGUu8KmqY1ZnC0Dg=="],
 
-    "@oxlint/binding-linux-arm-gnueabihf": ["@oxlint/binding-linux-arm-gnueabihf@1.75.0", "", { "os": "linux", "cpu": "arm" }, "sha512-7XlaWA5BJD3XpCfrEqjEe6Zseeb14S7QGa304XfwKignRaKQ+eIj775BQ7nIslggWickl4IsPUFqJ+/gAyNHVg=="],
+    "@oxlint/binding-linux-arm-gnueabihf": ["@oxlint/binding-linux-arm-gnueabihf@1.78.0", "", { "os": "linux", "cpu": "arm" }, "sha512-/uLdoJ0IXE6vo/0f0LKjinQAp+re+VMaCWaNT8ENIv2EOCkSsc8SGaflXAuW0Jua2dq5+GLVWm1NQK7P3UFSNQ=="],
 
-    "@oxlint/binding-linux-arm-musleabihf": ["@oxlint/binding-linux-arm-musleabihf@1.75.0", "", { "os": "linux", "cpu": "arm" }, "sha512-av6Tpv8yrcMMMOadOqENBhlsLRcGFXXwoQ0hzHhsmS9FJ4Wioy8we427GbcMe2XTxmL2e60T67H1Dyr3up+tAA=="],
+    "@oxlint/binding-linux-arm-musleabihf": ["@oxlint/binding-linux-arm-musleabihf@1.78.0", "", { "os": "linux", "cpu": "arm" }, "sha512-7xi4Wb/O8NRJhLoUXmDJMUVpNYvB5kefdhFU1Jb8rtae4QoXlTiLwI14X4YvAXVZLNZChP8m5qO9SQAlWQTbkQ=="],
 
-    "@oxlint/binding-linux-arm64-gnu": ["@oxlint/binding-linux-arm64-gnu@1.75.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-WcUhd8fHT5plrA14lANevl+hOl815mVI5t2hU21oFWrZKFXIVV/Sr4rWQV0NzSvzBupbMLNc5ErEA6Ehxh5jMg=="],
+    "@oxlint/binding-linux-arm64-gnu": ["@oxlint/binding-linux-arm64-gnu@1.78.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-4hFW0+fVXa3OIh1Y4A5SPkmvI4wuuBSrCVKzOyE7PTjhc7yEqZ1pmvEEeS5Lj/MaqvegFxXyF33N+6jkehxdyg=="],
 
-    "@oxlint/binding-linux-arm64-musl": ["@oxlint/binding-linux-arm64-musl@1.75.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-UWzp5wRHFe/ESO3+eEaxXsTkYTGLYjnTsi/I5neEacXSItQ6WNleapfOAeA4x2b8nyhJ4uQxqvtv9pHv8kWJtQ=="],
+    "@oxlint/binding-linux-arm64-musl": ["@oxlint/binding-linux-arm64-musl@1.78.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-oC0mvsgBJjlMijSDEhx9KuvR9zYeHXceA9MjbuXB1F8NSR78Yj2unOBrstEvTVaq+pko+kuue6DajC00eqvTdg=="],
 
-    "@oxlint/binding-linux-ppc64-gnu": ["@oxlint/binding-linux-ppc64-gnu@1.75.0", "", { "os": "linux", "cpu": "ppc64" }, "sha512-XEVRwGMLKCUKrvhLAz4F6AIh8MJrQVdSZtAmPpRZt9tGPsUnamPOcl3dS/ZQzJnar/Ymgc//+xho0L60Emzuxg=="],
+    "@oxlint/binding-linux-ppc64-gnu": ["@oxlint/binding-linux-ppc64-gnu@1.78.0", "", { "os": "linux", "cpu": "ppc64" }, "sha512-XAllT5SUZS+ohjuZ3/5S0cwe0r7eboiuigeStCZ5DXRYx/2KVM2UvQXvAfyzXEimtQjAB7cDQ2YxDe2Zl2WNQQ=="],
 
-    "@oxlint/binding-linux-riscv64-gnu": ["@oxlint/binding-linux-riscv64-gnu@1.75.0", "", { "os": "linux", "cpu": "none" }, "sha512-mAG4DUXqfLC8cTjMD2kt3jDmVzFREYtDyeLNdLdsCcBc4Zbl2EMuiFektGBilQwkNjYnMvCqJs55U+Hyb+b+jw=="],
+    "@oxlint/binding-linux-riscv64-gnu": ["@oxlint/binding-linux-riscv64-gnu@1.78.0", "", { "os": "linux", "cpu": "none" }, "sha512-trucMER/0QtecoXvc1y/UVqE3kwJipDwrx4oHfj+nNm3dq2zjP44WT0CfHNDPM3G1DXIkx/gY6lAD21NSCZVhA=="],
 
-    "@oxlint/binding-linux-riscv64-musl": ["@oxlint/binding-linux-riscv64-musl@1.75.0", "", { "os": "linux", "cpu": "none" }, "sha512-95hrAvriAlI+pekSomTFIn0+bawMDlDwTNVmdjsFusTHyL2JWh7TWvRNG/Lkim72uN8OiCcO9wcaC6omLP5E3w=="],
+    "@oxlint/binding-linux-riscv64-musl": ["@oxlint/binding-linux-riscv64-musl@1.78.0", "", { "os": "linux", "cpu": "none" }, "sha512-cm3O4F/HQbdzOUX5mKHqG5KDL6E5w0pnlZ+fbBy2rmLryPOowkuLagFHTopQsEIpjcaZoPOrL+BmmAytAG9HFg=="],
 
-    "@oxlint/binding-linux-s390x-gnu": ["@oxlint/binding-linux-s390x-gnu@1.75.0", "", { "os": "linux", "cpu": "s390x" }, "sha512-4b6f2+FrtruAESrCqIKcrarzfrSx+wk2QNcp+RT91/Prc+pMQMAfyZ1rG1c3tFQNl8Bc616tx40uNXyxNBRPbQ=="],
+    "@oxlint/binding-linux-s390x-gnu": ["@oxlint/binding-linux-s390x-gnu@1.78.0", "", { "os": "linux", "cpu": "s390x" }, "sha512-33wRf6HqGNsybJ3qX4cGaQN2ODPxNmc1rMa0mrTmx3eFq1VzOnvQooi9bIGVYakW8a/wmqVx1mgsUm8R2xfTiw=="],
 
-    "@oxlint/binding-linux-x64-gnu": ["@oxlint/binding-linux-x64-gnu@1.75.0", "", { "os": "linux", "cpu": "x64" }, "sha512-nshAhrUvXFUWOvqQ2soIw7HFNWvpvEV4o0cYSqPtzLiPF5gKyYTDOOTJ6Rn8g8K/iGvPIrbDA4v8+5MvnjJrrg=="],
+    "@oxlint/binding-linux-x64-gnu": ["@oxlint/binding-linux-x64-gnu@1.78.0", "", { "os": "linux", "cpu": "x64" }, "sha512-rRdISSYegj6VganMZ9tjRjijowfHJ09IZU01i0toBAqr6n5LEtwHq2IeS4FjW2RoskOHlb6efB26H5izYb3GEQ=="],
 
-    "@oxlint/binding-linux-x64-musl": ["@oxlint/binding-linux-x64-musl@1.75.0", "", { "os": "linux", "cpu": "x64" }, "sha512-e4jNxLKnxLC6sYBQRxrI2pgIIxnmMtF8U/VwNYcjTT/CLS+spH624cYVnj07bTKwaEWT37/e025isOs6j/0xqA=="],
+    "@oxlint/binding-linux-x64-musl": ["@oxlint/binding-linux-x64-musl@1.78.0", "", { "os": "linux", "cpu": "x64" }, "sha512-GmsP4rW0xTL6u5CVdcDsaN5Fbc7hBc382Wmar1kttbnwSEviM+rSINKOMQ+UQ6iH+AGwC+8gaAiwu134Tgh6Lg=="],
 
-    "@oxlint/binding-openharmony-arm64": ["@oxlint/binding-openharmony-arm64@1.75.0", "", { "os": "none", "cpu": "arm64" }, "sha512-hZ2lH+1qLf/DiEP9UWuQTK2JWj/BgvMB4jhIV4SmNU1wfEiYYX4TynQyAZXx0j9X4qRYizAL042SKaV+8ynh4w=="],
+    "@oxlint/binding-openharmony-arm64": ["@oxlint/binding-openharmony-arm64@1.78.0", "", { "os": "none", "cpu": "arm64" }, "sha512-sy9yeYuADc8a+n4TLBayzMCZiHPW78DcIFVpOXTmdKHWQeM9xe5uzkqIIZmi326D5hY9XVwacipEB1p7tQjPAg=="],
 
-    "@oxlint/binding-win32-arm64-msvc": ["@oxlint/binding-win32-arm64-msvc@1.75.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-Ilj6PNzGDS3bCU0MSJH7Msh0NhH+T/mRp2shwg+q+GHeVlPwP5LEboW96aW+3kVKFk6zYZy1Xi5pZkqZh6X8KQ=="],
+    "@oxlint/binding-win32-arm64-msvc": ["@oxlint/binding-win32-arm64-msvc@1.78.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-rjc2hF1KfMi8fZj1X/m3AmnHbdsF3rL0v6KQg0Uc880Yb2khjz+3U14sfdZ7jWTpRnN1m1NQa/TT7uU9lJWPrA=="],
 
-    "@oxlint/binding-win32-ia32-msvc": ["@oxlint/binding-win32-ia32-msvc@1.75.0", "", { "os": "win32", "cpu": "ia32" }, "sha512-QVit2nOEOiPhkmsrksPSkoGCdnZRNkspt8fwoYyP09te1VEbnSj4LAxua4rc8FKTmWkySVe05j8iz9GXYfF1AQ=="],
+    "@oxlint/binding-win32-ia32-msvc": ["@oxlint/binding-win32-ia32-msvc@1.78.0", "", { "os": "win32", "cpu": "ia32" }, "sha512-zcuXFVrEFHIafRfkCQT8w/Xe41o07ozl/vwHq7p94vB29xVzsB0sZGYORU1jhcYKv3Lr0J3HbJ2T4fHH5rWmvA=="],
 
-    "@oxlint/binding-win32-x64-msvc": ["@oxlint/binding-win32-x64-msvc@1.75.0", "", { "os": "win32", "cpu": "x64" }, "sha512-DSxnNkBUAYARPwJtR12Ig3deWr8w0H997xP6jy33i+e0SyYJw8FKuz4+cZtpmPEhQmvlPJE3X/2vNxDmLkd/rA=="],
+    "@oxlint/binding-win32-x64-msvc": ["@oxlint/binding-win32-x64-msvc@1.78.0", "", { "os": "win32", "cpu": "x64" }, "sha512-Sb5ocmLSuYeOuXd+CFOToGKp/gjXUEWDnvIGwhnh8aq8wY4TMmEnKnvbogSW7RdMZv77JSARduS7/gv+khYEjA=="],
 
     "@posthog/browser-common": ["@posthog/browser-common@0.2.1", "", { "dependencies": { "@posthog/core": "1.45.1", "@posthog/types": "1.398.0" } }, "sha512-FTVPsRw6GBKWvFwN26TNIQNNYPR9FsUj+B+CKhk/ht63IRzn4yYFtFOjcmW+bfXvNHuRADs6APbZ6Haz1kpoAw=="],
 
-    "@posthog/cli": ["@posthog/cli@0.7.24", "", { "dependencies": { "detect-libc": "2.1.2" }, "bin": { "posthog-cli": "run-posthog-cli.js" } }, "sha512-Jho5QDPYv21gDd8O1JDewPmA8ZEO2VbikszYTtD0rCem6nsDN528tpZnNUPGPYyeuAQHTQasoo7/g3IJVCBmDg=="],
+    "@posthog/cli": ["@posthog/cli@0.11.2", "", { "dependencies": { "detect-libc": "^2.1.2" }, "bin": { "posthog-cli": "run-posthog-cli.js" } }, "sha512-iqYIl/Q2FfE/+vftlHqGYFz4DRcDon7t28PU6kIefFFUMr0Z4bZrypnxU/Wl4xW/0uUw4fiAKs2tkWmwh3lSDA=="],
 
     "@posthog/core": ["@posthog/core@1.45.1", "", { "dependencies": { "@posthog/types": "1.398.0" } }, "sha512-tLtvzomavb2PPWdGYKsusyIzIeL2Px47v348Smibkay7sMy/83TyPk+Ptsp2NdeOgJsbuwSxWkR2+XA0aSCAaA=="],
 
-    "@posthog/nextjs-config": ["@posthog/nextjs-config@1.9.68", "", { "dependencies": { "@posthog/cli": "0.7.24", "@posthog/plugin-utils": "1.1.2", "@posthog/webpack-plugin": "1.5.24", "semver": "7.8.0" }, "peerDependencies": { "next": "16.2.12" } }, "sha512-MmzSrompIThnnrDArCKIOBZuD/b3RErH48ZJyzAuX+WfJTiHBjjTMvGWzr+qlIn8jCyJoglSygAwYVG8lJXybg=="],
+    "@posthog/nextjs-config": ["@posthog/nextjs-config@1.9.69", "", { "dependencies": { "@posthog/cli": "~0.11.1", "@posthog/plugin-utils": "^1.1.2", "@posthog/webpack-plugin": "^1.5.27", "semver": "^7.8.5" }, "peerDependencies": { "next": ">12.1.0" } }, "sha512-l1flhPrRGbJWXNwbbfFfmimQjO39d2pvha6MO9K9yTpHhJR9p8IbmyvsGFsY7l8xntaR+2gRsamto/fY8O086g=="],
 
     "@posthog/plugin-utils": ["@posthog/plugin-utils@1.1.2", "", { "dependencies": { "cross-spawn": "7.0.6" } }, "sha512-CS/14cMmny9NN/LHAb4QZdW3TfxhbKzsfyVYMCjsRdE0IKMTrXMr0cPRZxwczsN6VnWvVoy2Dd5wcG6RrIjlYA=="],
 
     "@posthog/types": ["@posthog/types@1.398.0", "", {}, "sha512-sJMkl4k+u8yS/0fjHsKqE9xTdsAh30a2WvgChiptellnVoE0e8QJKFgqOMD2sk8FaEArPdeFklAhXvmENAt3Sg=="],
 
-    "@posthog/webpack-plugin": ["@posthog/webpack-plugin@1.5.24", "", { "dependencies": { "@posthog/cli": "0.7.24", "@posthog/core": "1.40.1", "@posthog/plugin-utils": "1.1.2" }, "peerDependencies": { "webpack": "5.106.2" } }, "sha512-sdqedI8V/Iafy4+pT2g8glEMEhFSnE8PqW5rbE2qZH4TR57Xx3qfFrKFw6A9IJ29y9L0/kCdgmIWkxdTndY1+g=="],
+    "@posthog/webpack-plugin": ["@posthog/webpack-plugin@1.5.27", "", { "dependencies": { "@posthog/cli": "~0.11.1", "@posthog/core": "^1.48.0", "@posthog/plugin-utils": "^1.1.2" }, "peerDependencies": { "webpack": "^5" } }, "sha512-H5OWzFymU8UfLbFaZ4y6kEZHvgZJwSrTACosId34Gq1UnBymeRuCfOFQRIG5xVqPAB6+vN0YuPpFIzpJ82i28g=="],
 
     "@radix-ui/number": ["@radix-ui/number@1.1.3", "", {}, "sha512-Road2bidD0uu/1BGDOWNdPI06g0lIRy6IF9GZcIrDK2KGItfor8IQwQa+yM2ERgHM1MmHxaxpTzk0/Jp42lNfA=="],
 
     "@radix-ui/primitive": ["@radix-ui/primitive@1.1.7", "", {}, "sha512-rqWnm76nYT8HoNNqEjpgJ7Pw/DrBj5iBTrmEPo6HTX5+VJyBNOqTdv4g89G63HuR5g0AaENoAcH7Is5fF2kZ8Q=="],
 
-    "@radix-ui/react-accordion": ["@radix-ui/react-accordion@1.2.20", "", { "dependencies": { "@radix-ui/primitive": "1.1.7", "@radix-ui/react-collapsible": "1.1.20", "@radix-ui/react-collection": "1.1.15", "@radix-ui/react-compose-refs": "1.1.5", "@radix-ui/react-context": "1.2.2", "@radix-ui/react-direction": "1.1.4", "@radix-ui/react-id": "1.1.4", "@radix-ui/react-primitive": "2.1.10", "@radix-ui/react-use-controllable-state": "1.2.6" }, "optionalDependencies": { "@types/react": "19.2.17", "@types/react-dom": "19.2.3" }, "peerDependencies": { "react": "19.2.8", "react-dom": "19.2.8" } }, "sha512-jDhG9FvAEnlhnjrsINbNXcUa4G+L1KqSkJSunkbKEzFRcAb52jvM0PjPxPRvhe1HNc5F5yc0yzzWeeqlH4yBIg=="],
-
     "@radix-ui/react-arrow": ["@radix-ui/react-arrow@1.1.15", "", { "dependencies": { "@radix-ui/react-primitive": "2.1.10" }, "optionalDependencies": { "@types/react": "19.2.17", "@types/react-dom": "19.2.3" }, "peerDependencies": { "react": "19.2.8", "react-dom": "19.2.8" } }, "sha512-v4zggRcjadnI+ClKDuijlQEW4tw3NoaeHc/PwpKnLoLLKNUG4InLegkstooLcRIUWCs+8L22dGURCVuFfOKfnA=="],
 
-    "@radix-ui/react-collapsible": ["@radix-ui/react-collapsible@1.1.20", "", { "dependencies": { "@radix-ui/primitive": "1.1.7", "@radix-ui/react-compose-refs": "1.1.5", "@radix-ui/react-context": "1.2.2", "@radix-ui/react-id": "1.1.4", "@radix-ui/react-presence": "1.1.10", "@radix-ui/react-primitive": "2.1.10", "@radix-ui/react-use-controllable-state": "1.2.6", "@radix-ui/react-use-layout-effect": "1.1.4" }, "optionalDependencies": { "@types/react": "19.2.17", "@types/react-dom": "19.2.3" }, "peerDependencies": { "react": "19.2.8", "react-dom": "19.2.8" } }, "sha512-mcGesGplBnzN2sbvJETzpCNfSMyPnb29q1GRLU+Ib7bJrpIG2ywmRoh2V5VbA2uNvKikKUlVbAPks7JDjz4A8Q=="],
-
     "@radix-ui/react-collection": ["@radix-ui/react-collection@1.1.15", "", { "dependencies": { "@radix-ui/react-compose-refs": "1.1.5", "@radix-ui/react-context": "1.2.2", "@radix-ui/react-primitive": "2.1.10", "@radix-ui/react-slot": "1.3.3" }, "optionalDependencies": { "@types/react": "19.2.17", "@types/react-dom": "19.2.3" }, "peerDependencies": { "react": "19.2.8", "react-dom": "19.2.8" } }, "sha512-9W+B9NPF0NaaPh/1NJd3+KqsnlLqU9H7T2rvww+fp+T/evVXdNAyYcnfRQZFOjkR1ajQp3yORlqnI8soawLvNA=="],
 
     "@radix-ui/react-compose-refs": ["@radix-ui/react-compose-refs@1.1.5", "", { "optionalDependencies": { "@types/react": "19.2.17" }, "peerDependencies": { "react": "19.2.8" } }, "sha512-+48PbAAbq3didjJxa+OaWY2ZwgAKsNiRGyeHKszblZMQ+kcpd9pAaT11cMkGEie0vsOi3QdeTE6d5Fe3Gn61kA=="],
@@ -565,37 +568,37 @@
 
     "@radix-ui/rect": ["@radix-ui/rect@1.1.3", "", {}, "sha512-JtyZR+mqgBibTo8xea3B6ZRmzZiM/YeVBtUkas6zMuXjAlfIFIW2FgqeM9eLyvEaYX66vr6DJMK+4U6LV0KhNw=="],
 
-    "@react-three/fiber": ["@react-three/fiber@9.6.1", "", { "dependencies": { "@babel/runtime": "7.29.2", "@types/webxr": "0.5.24", "base64-js": "1.5.1", "buffer": "6.0.3", "its-fine": "2.0.0", "react-use-measure": "2.1.7", "scheduler": "0.27.0", "suspend-react": "0.1.3", "use-sync-external-store": "1.6.0", "zustand": "5.0.13" }, "optionalDependencies": { "react-dom": "19.2.8" }, "peerDependencies": { "react": "19.2.8", "three": "0.185.1" } }, "sha512-zF0rsKcVYpcJwbFEnv2HkHX9cvOEgsfQo/X8lwmR2dn13S4qEQJXir9fxf5js2LQFoXqxOY7MDkOkYx2uZ4gSg=="],
+    "@react-three/fiber": ["@react-three/fiber@9.7.0", "", { "dependencies": { "@babel/runtime": "^7.17.8", "@types/webxr": "*", "base64-js": "^1.5.1", "buffer": "^6.0.3", "its-fine": "^2.0.0", "react-use-measure": "^2.1.7", "scheduler": "^0.27.0", "suspend-react": "^0.1.3", "use-sync-external-store": "^1.4.0", "zustand": "^5.0.3" }, "peerDependencies": { "expo": ">=43.0", "expo-asset": ">=8.4", "expo-file-system": ">=11.0", "expo-gl": ">=11.0", "react": ">=19 <19.3", "react-dom": ">=19 <19.3", "react-native": ">=0.78", "three": ">=0.156" }, "optionalPeers": ["expo", "expo-asset", "expo-file-system", "expo-gl", "react-dom", "react-native"] }, "sha512-EWm9FwcaOZQu/ExFW5rggoCMM1NJet5YbxVxKaOE+KSncrjU0Wx7017qSyGFvupviK89nMYGCWU3BIK4dI1clw=="],
 
-    "@rolldown/binding-android-arm64": ["@rolldown/binding-android-arm64@1.1.5", "", { "os": "android", "cpu": "arm64" }, "sha512-lZg8fqIv2v7FF237bwMgzGZEJvGL79/s5knJ/i6FmsGF4XXlzccZ4jb+TrFIxtSSxFtIpdsgrPZeMk1I9AFcyQ=="],
+    "@rolldown/binding-android-arm64": ["@rolldown/binding-android-arm64@1.2.4", "", { "os": "android", "cpu": "arm64" }, "sha512-jHC2cnyKz5xU2fhECtFl8OZ83cYNt13GZQD+0uMJ/X3o+ijmd56okHhTUwxVSHPx1IRVIJEZ1/1pPzeLCU6XKA=="],
 
-    "@rolldown/binding-darwin-arm64": ["@rolldown/binding-darwin-arm64@1.1.5", "", { "os": "darwin", "cpu": "arm64" }, "sha512-51Bnx9pNiMRKSUNtBfySkNJ9vMU9Hh3I1ozDd6gyPPYzaXCfnptUcEZxXGYFn+ul2dtcMUiqGR1Yai2K10uoTw=="],
+    "@rolldown/binding-darwin-arm64": ["@rolldown/binding-darwin-arm64@1.2.4", "", { "os": "darwin", "cpu": "arm64" }, "sha512-Dc5mPD8F5F/FS8i01syd7FTF6yB2fVthH/TRkjwJkzUK6EpoxHtqvZQP5Zwq80/5z19TWYHIg1KOHboCgVx/aQ=="],
 
-    "@rolldown/binding-darwin-x64": ["@rolldown/binding-darwin-x64@1.1.5", "", { "os": "darwin", "cpu": "x64" }, "sha512-Tm+gbfC0aHu1tBA/JvKQh32S0K6YgCHkiAF4/W6xX0K0RmNuc94VeK419dJoE65R5aRxmo+noZQSWrAMF6yb6g=="],
+    "@rolldown/binding-darwin-x64": ["@rolldown/binding-darwin-x64@1.2.4", "", { "os": "darwin", "cpu": "x64" }, "sha512-fpDm4oBo6SqLvWUYCmFhdde3U9KH2fRNNMeAnAPAIwxRL345xutL0EtEUcuoxsoazdJGv/MuDBQHlCDrtbvqOg=="],
 
-    "@rolldown/binding-freebsd-x64": ["@rolldown/binding-freebsd-x64@1.1.5", "", { "os": "freebsd", "cpu": "x64" }, "sha512-JMzDKCCXq93YccG5gz3hvOs1oXRKAf0XYpfOS88e+wZrC8Iugj6j68867vrYZkvpDDpKn/KoKORThmchMpF6TA=="],
+    "@rolldown/binding-freebsd-x64": ["@rolldown/binding-freebsd-x64@1.2.4", "", { "os": "freebsd", "cpu": "x64" }, "sha512-rSJoreDE/HoIzoaib6MTp5jQtCTdMHKIvItAKT/ImS6Y6Ww76oUaeMyp4Vc/fAgd/ehji068IxetHXAnqUwN9A=="],
 
-    "@rolldown/binding-linux-arm-gnueabihf": ["@rolldown/binding-linux-arm-gnueabihf@1.1.5", "", { "os": "linux", "cpu": "arm" }, "sha512-uML21j2K5TfPGutKxub+M+nLjZIrWjXQ5Grx4lCe/nimTj9B4L63zHpjXLl4y0L3mcm2htEQIb06oCG/szerNw=="],
+    "@rolldown/binding-linux-arm-gnueabihf": ["@rolldown/binding-linux-arm-gnueabihf@1.2.4", "", { "os": "linux", "cpu": "arm" }, "sha512-/jm8OGHgn7oGaJu3i/qZI9spUGcJ+y/lk43ttQ/iO1tOd9NissG6o97bighBCiL+BKRngmcDuR6ikfwYdJmVuQ=="],
 
-    "@rolldown/binding-linux-arm64-gnu": ["@rolldown/binding-linux-arm64-gnu@1.1.5", "", { "os": "linux", "cpu": "arm64" }, "sha512-navSiuTMogvnQoZoM/v+l3ZWo50/NTwSHSzheABx/RCnmUPaKwq9qSo4Br2OYRs21+Fz8uFqITZM3H4opOB0/Q=="],
+    "@rolldown/binding-linux-arm64-gnu": ["@rolldown/binding-linux-arm64-gnu@1.2.4", "", { "os": "linux", "cpu": "arm64" }, "sha512-tIP06BeD9EqvECBrPZ+sqdPlYrT+aYaAiu1wYziVx5elRK/ftm33JxVDy2bXGbr6J0CrtirCkR87/X5a2euEng=="],
 
-    "@rolldown/binding-linux-arm64-musl": ["@rolldown/binding-linux-arm64-musl@1.1.5", "", { "os": "linux", "cpu": "arm64" }, "sha512-lAryqH7IteztmCXQXk0etKj4wBQ7Gx5S6LjKhsgp9zb8I5bsuvU/2llH1hDQcjsFeqIsovMVN339/8pUDDBXxA=="],
+    "@rolldown/binding-linux-arm64-musl": ["@rolldown/binding-linux-arm64-musl@1.2.4", "", { "os": "linux", "cpu": "arm64" }, "sha512-Ql1Q0EQqVThvn9VAVlwNzsUvbSFtCMGjLpRRi4pk5i7NZZ4n5ISiLMjHYtus4VQ2PvkSw24zyaCVsiS+sXPj1w=="],
 
-    "@rolldown/binding-linux-ppc64-gnu": ["@rolldown/binding-linux-ppc64-gnu@1.1.5", "", { "os": "linux", "cpu": "ppc64" }, "sha512-fsK/sNBnxzBlL4O1JNrZakVQxPspqpED5dLtNsZS9oOKmtSpdNIzxH2kkol5HYTWJN47sE20ztMJPxfZ89qGOg=="],
+    "@rolldown/binding-linux-ppc64-gnu": ["@rolldown/binding-linux-ppc64-gnu@1.2.4", "", { "os": "linux", "cpu": "ppc64" }, "sha512-GjbjXD4XXfN19D0LZNbmiCBUoDiRACsYHr0yaIbbn8aFsXjHZifcYqu/W5Er5X2X990WjHXFrxarn5chzItorQ=="],
 
-    "@rolldown/binding-linux-s390x-gnu": ["@rolldown/binding-linux-s390x-gnu@1.1.5", "", { "os": "linux", "cpu": "s390x" }, "sha512-gLYb4BIadlfTOYT5gO503n8zQjXflgzpD0FcyKh0Mzx3rqCZKnHoJWV9xe1KXUJ5lx2JfcSHr/mhzS0PC/McAA=="],
+    "@rolldown/binding-linux-s390x-gnu": ["@rolldown/binding-linux-s390x-gnu@1.2.4", "", { "os": "linux", "cpu": "s390x" }, "sha512-p5WR0NOwaRmJ/B1b6IjEFLLivwEsf3PrdBIhRbhTCQisbo2SvHHpG4ELB/+FgQNnB88LTOF86upmJmbvZdQ2lw=="],
 
-    "@rolldown/binding-linux-x64-gnu": ["@rolldown/binding-linux-x64-gnu@1.1.5", "", { "os": "linux", "cpu": "x64" }, "sha512-FjcpEKUyJygHgs1o50VYNvkt5+7Le/VEdYt0AkRpkL33MnyQfwr8l5mXwMmfmTbyMPr5vJLC+8/Gd9gXnwU1QQ=="],
+    "@rolldown/binding-linux-x64-gnu": ["@rolldown/binding-linux-x64-gnu@1.2.4", "", { "os": "linux", "cpu": "x64" }, "sha512-4/GyVjmhR+Tc6HLJvwc1sOhPqAZtySiSMesOZyX6JQ5XBxoTDEMKQzvo07NIK6nTon/SivlZqvhzvuVBNQhObQ=="],
 
-    "@rolldown/binding-linux-x64-musl": ["@rolldown/binding-linux-x64-musl@1.1.5", "", { "os": "linux", "cpu": "x64" }, "sha512-Me+PfPI2TMeOQk0gYWfLQZtTktrmzbr8cDboqX83XKc7UrgAi55gF+2dUkWdxd19n55Essp2yeca+O9N5rBxHg=="],
+    "@rolldown/binding-linux-x64-musl": ["@rolldown/binding-linux-x64-musl@1.2.4", "", { "os": "linux", "cpu": "x64" }, "sha512-l9eeLsCNvPpmSXUej0etw/J1eqV0Jj1D5G/xG6YTijmE6dkv6E2QezgWbTfQk63v952DPqrjOCoiqxq7Bw0YUQ=="],
 
-    "@rolldown/binding-openharmony-arm64": ["@rolldown/binding-openharmony-arm64@1.1.5", "", { "os": "none", "cpu": "arm64" }, "sha512-yc5WrLzXks6zCQfn9Oxr8pORKyl/pF+QjHmW/Qx3qu0oyrrNC+y2JLTU1E2rcWYAmzlnqngWXHQjy51VzW70Vw=="],
+    "@rolldown/binding-openharmony-arm64": ["@rolldown/binding-openharmony-arm64@1.2.4", "", { "os": "none", "cpu": "arm64" }, "sha512-e0F355MSTMm3+UOqtV3L24gFUp2N5m1f8L/7d56deik6va+AXdrt9F8LbzGpeWGWRbZEDq4m8NVnJDeBtf9DZg=="],
 
     "@rolldown/binding-wasm32-wasi": ["@rolldown/binding-wasm32-wasi@1.1.5", "", { "dependencies": { "@emnapi/core": "1.11.1", "@emnapi/runtime": "1.11.1", "@napi-rs/wasm-runtime": "1.1.6" }, "cpu": "none" }, "sha512-VbQGPX2b4r48TAMIM2cjgluIM1HYutm4pcTEJsle7iEP7sB1dFqtPLBVbdLAZCxy1txCcPxf4QFf4v8uvltPqA=="],
 
-    "@rolldown/binding-win32-arm64-msvc": ["@rolldown/binding-win32-arm64-msvc@1.1.5", "", { "os": "win32", "cpu": "arm64" }, "sha512-gHv82k63z4qpV5+Q1y/12KrK0ltWBukVDI8nZcbT7Tt/ZlOIVwppazneq0F93oDxTo3IgAMEDIoQh3E2n6mVsw=="],
+    "@rolldown/binding-win32-arm64-msvc": ["@rolldown/binding-win32-arm64-msvc@1.2.4", "", { "os": "win32", "cpu": "arm64" }, "sha512-AWLi0uBRYh6QlE7OKhiz+phZC0qwtij2QZmhmOdsLdFn64m7oMpooE9ICE3lhm9xMb4SpDo2WbHcxX1iFLFtqw=="],
 
-    "@rolldown/binding-win32-x64-msvc": ["@rolldown/binding-win32-x64-msvc@1.1.5", "", { "os": "win32", "cpu": "x64" }, "sha512-tTZuDBPw85tEN5PQi1pnEBzDy0Z49HtScLAbD5t6hyeU92A95pRWaSMw1GZZi/RwgSgUIl0xrSlXIT/9QzvYSA=="],
+    "@rolldown/binding-win32-x64-msvc": ["@rolldown/binding-win32-x64-msvc@1.2.4", "", { "os": "win32", "cpu": "x64" }, "sha512-UwSDJOg3dqCAejWdxclJjCsh3Qq4vLYMDxmyHqo1btz3stK2VqgwNd3mm5tuIwzSlGIQ/1H9Hr+Zn09mrezNqQ=="],
 
     "@rolldown/pluginutils": ["@rolldown/pluginutils@1.0.1", "", {}, "sha512-2j9bGt5Jh8hj+vPtgzPtl72j0yRxHAyumoo6TNfAjsLB04UtpSvPbPcDcBMxz7n+9CYB0c1GxQFxYRg2jimqGw=="],
 
@@ -607,27 +610,27 @@
 
     "@semianalysisai/inferencex-mcp": ["@semianalysisai/inferencex-mcp@workspace:packages/mcp"],
 
-    "@shikijs/core": ["@shikijs/core@4.3.1", "", { "dependencies": { "@shikijs/primitive": "4.3.1", "@shikijs/types": "4.3.1", "@shikijs/vscode-textmate": "10.0.2", "@types/hast": "3.0.4", "hast-util-to-html": "9.0.5" } }, "sha512-ANMDxuaPsNMdDC1m4vfvhlDmJweMwkE5XitTwrq2rWHx5jM+dlm4MmHt2PP6t0uejfR77SuhrhJ0zEijIF/uhA=="],
+    "@shikijs/core": ["@shikijs/core@4.4.3", "", { "dependencies": { "@shikijs/primitive": "4.4.3", "@shikijs/types": "4.4.3", "@shikijs/vscode-textmate": "^10.0.2", "@types/hast": "^3.0.5", "hast-util-to-html": "^9.0.5" } }, "sha512-QCR4q2ZO/ILJEuwiBMel4wdcTDb1JGwfjKTxPDF6x8ixOaluPrVqIn06C99AcRPhmYlBR56d/Fb+GN58GzExpg=="],
 
-    "@shikijs/engine-javascript": ["@shikijs/engine-javascript@4.3.1", "", { "dependencies": { "@shikijs/types": "4.3.1", "@shikijs/vscode-textmate": "10.0.2", "oniguruma-to-es": "4.3.6" } }, "sha512-JBItcnPuYq7jVJdZo/vMj94r+szT7XEjHFX+mvFDGSEIbVAXAGyHAHzhbWzpGOwYidCZrErJLLgn2PVeiokHnQ=="],
+    "@shikijs/engine-javascript": ["@shikijs/engine-javascript@4.4.3", "", { "dependencies": { "@shikijs/types": "4.4.3", "@shikijs/vscode-textmate": "^10.0.2", "oniguruma-to-es": "^4.3.6" } }, "sha512-FbOjFJp9VLdo1Wevs10BBtVxiTWwNLqZh5Gkhjgda/ioL15YOgeSl9n+6XMa3qRlPQzfhFNe641SrynFHYG0nQ=="],
 
-    "@shikijs/engine-oniguruma": ["@shikijs/engine-oniguruma@4.3.1", "", { "dependencies": { "@shikijs/types": "4.3.1", "@shikijs/vscode-textmate": "10.0.2" } }, "sha512-OXyNMzg0pews+msMj4cHeqT4xiYKKvbnn6VbdAXxfoFl3SSx4fJTc8FadECuc5/H9p3BzhNAoAUXKwAu9rWYhg=="],
+    "@shikijs/engine-oniguruma": ["@shikijs/engine-oniguruma@4.4.3", "", { "dependencies": { "@shikijs/types": "4.4.3", "@shikijs/vscode-textmate": "^10.0.2" } }, "sha512-EcOQkxdxGQrc1Row/cC2c96/v1dbZqGnEVu1qTuT/MJmp6+cXCvQussowVmCv5Tqr3KuY3c7IbM6HTW3LJ1k9w=="],
 
-    "@shikijs/langs": ["@shikijs/langs@4.3.1", "", { "dependencies": { "@shikijs/types": "4.3.1" } }, "sha512-m0l9nsDqgBHvbZbk7A0/kXz/impK3uB/c6rAn6Gpg/uPtdZRQ+alsN/17MU5thb68XTj/4DxkZAotrM0GGSpDQ=="],
+    "@shikijs/langs": ["@shikijs/langs@4.4.3", "", { "dependencies": { "@shikijs/types": "4.4.3" } }, "sha512-ePic0yfAJGOF83D5wBHK/00EjK65oahBYxFk5epgq33WRv7X9UuxLEV8PtR0szC0z8dl7INIpIodB99JRFlR+A=="],
 
-    "@shikijs/primitive": ["@shikijs/primitive@4.3.1", "", { "dependencies": { "@shikijs/types": "4.3.1", "@shikijs/vscode-textmate": "10.0.2", "@types/hast": "3.0.4" } }, "sha512-CXQRQOYy1leqQ8ceTeJdmXv/bsUY++6QyLpXJ94LZAAYj5X2SKRdc5ipguv4NPyGVKItB2PPwUpRNe0Sjh5S1A=="],
+    "@shikijs/primitive": ["@shikijs/primitive@4.4.3", "", { "dependencies": { "@shikijs/types": "4.4.3", "@shikijs/vscode-textmate": "^10.0.2", "@types/hast": "^3.0.5" } }, "sha512-m0wBeLDQDeIxRdUmrCPdQqfuUamDwRL5isCfYbguKD6NiaKpVbsv+3J81DyIKgNW5h4WAIIr8T4EkgQrBBxvaQ=="],
 
-    "@shikijs/rehype": ["@shikijs/rehype@4.3.1", "", { "dependencies": { "@shikijs/types": "4.3.1", "@types/hast": "3.0.4", "hast-util-to-string": "3.0.1", "shiki": "4.3.1", "unified": "11.0.5", "unist-util-visit": "5.1.0" } }, "sha512-oshrlfUF3VPUJfnp5K1lLwsS/SRBKrIxONpdWebSKZXdBE3UsZnxgqpvRUA8UsofS7vmjFOCAHIT71ECbmOxTw=="],
+    "@shikijs/rehype": ["@shikijs/rehype@4.4.3", "", { "dependencies": { "@shikijs/types": "4.4.3", "@types/hast": "^3.0.5", "hast-util-to-string": "^3.0.1", "shiki": "4.4.3", "unified": "^11.0.5", "unist-util-visit": "^5.1.0" } }, "sha512-vkG9jG1aRnrx05R31uAOKQHE8qpY7r1cBXE2sAZgFK2IaPnQHwaP4L1C6amQixmZ8thBsKwfbSsYxMDoo46UWQ=="],
 
-    "@shikijs/themes": ["@shikijs/themes@4.3.1", "", { "dependencies": { "@shikijs/types": "4.3.1" } }, "sha512-dgpoJ4WqNi2yTmizQHBJ5zcX6j2lE6icN/0yt4l1kkf16jrY/pwPLoTb1ETsWMz0OBLf9ZNvwmxft+cH+N9qSA=="],
+    "@shikijs/themes": ["@shikijs/themes@4.4.3", "", { "dependencies": { "@shikijs/types": "4.4.3" } }, "sha512-w8UHjeUnIR965KMWJHUPXOc2mNJUnK3vpVLYLvw5IYU2mnTTJ89E24OrJDBNiJDQ0qzb0tc4l7mrIXx5cFeIyw=="],
 
-    "@shikijs/types": ["@shikijs/types@4.3.1", "", { "dependencies": { "@shikijs/vscode-textmate": "10.0.2", "@types/hast": "3.0.4" } }, "sha512-CHFxE0jztBIZRHH6gxXE7DXUCFXjReEGxZ/j0rfSLGKZuwp2xBYycEP14875DSa9KLL/6700oxIq6oO6ef9K2g=="],
+    "@shikijs/types": ["@shikijs/types@4.4.3", "", { "dependencies": { "@shikijs/vscode-textmate": "^10.0.2", "@types/hast": "^3.0.5" } }, "sha512-UEJxmRR++MAGR6hugn0vgVS2W/6lWAts84FFSrnlH9sP0LNol7E5+NQ792pH8liWUhyMyjhTgSUH3k7iD7tc5g=="],
 
     "@shikijs/vscode-textmate": ["@shikijs/vscode-textmate@10.0.2", "", {}, "sha512-83yeghZ2xxin3Nj8z1NMd/NCuca+gsYXswywDy5bHvwlWL8tpTQmzGeUuHd9FC3E/SBEMvzJRwWEOz5gGes9Qg=="],
 
     "@standard-schema/spec": ["@standard-schema/spec@1.1.0", "", {}, "sha512-l2aFy5jALhniG5HgqrD6jXLi/rUWrKvqN/qJx6yoJsgKhblVd+iqqU4RCXavm/jPityDo5TCvKMnpjKnOriy0w=="],
 
-    "@swc/helpers": ["@swc/helpers@0.5.15", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-JQ5TuMi45Owi4/BIMAJBoSQoOJu12oOk/gADqlcUL9JEdHB8vyjUSsxqeNXnmXHjYKMi2WcYtezGEEhqUI/E2g=="],
+    "@swc/helpers": ["@swc/helpers@0.5.23", "", { "dependencies": { "tslib": "^2.8.0" } }, "sha512-5lSsMOTXURePglDfvuAQUqkGek9Hg2kksOYay2m0+XR++b2NWYL/4sWyuvVBIs8oKnJaxkdi9whaL/sqN13afw=="],
 
     "@tailwindcss/node": ["@tailwindcss/node@4.3.3", "", { "dependencies": { "@jridgewell/remapping": "2.3.5", "enhanced-resolve": "5.24.2", "jiti": "2.7.0", "lightningcss": "1.32.0", "magic-string": "0.30.21", "source-map-js": "1.2.1", "tailwindcss": "4.3.3" } }, "sha512-/T8IKEsf9VTU6tLjgC7+sv2mOPtQxzE2jMw7u4Tt40Tx+QSZxpzh95/H6cMKoja9XuW7iMdLJYBB0o9G1CaAgg=="],
 
@@ -753,7 +756,7 @@
 
     "@types/geojson": ["@types/geojson@7946.0.16", "", {}, "sha512-6C8nqWur3j98U6+lXDfTUWIfgvZU+EumvpHKcYjujKH7woYyLj2sUmff0tRhrqM7BohUw7Pz3ZB1jj2gW9Fvmg=="],
 
-    "@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+    "@types/hast": ["@types/hast@3.0.5", "", { "dependencies": { "@types/unist": "*" } }, "sha512-rp/ezSWaD1m44dPKICGhiskI13nVr7qTloFwDa/IYkhhf5nzwP+zIQcIJh3WIFSBOy/H1PzB40jPjMDksN4F+g=="],
 
     "@types/jsesc": ["@types/jsesc@2.5.1", "", {}, "sha512-9VN+6yxLOPLOav+7PwjZbxiID2bVaeq0ED4qSQmdQTdjnXJSaCVKTR58t15oqH1H5t8Ng2ZX1SabJVoN9Q34bw=="],
 
@@ -767,7 +770,7 @@
 
     "@types/ms": ["@types/ms@2.1.0", "", {}, "sha512-GsCCIZDE/p3i96vtEqx+7dBUGXrc7zeSK3wwPHIaRThS+9OhWIXRqzs4d6k1SVU8g91DrNRWxWUGhp5KXQb2VA=="],
 
-    "@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+    "@types/node": ["@types/node@26.2.0", "", { "dependencies": { "undici-types": "~8.3.0" } }, "sha512-5IviulTZeRNp2vAJ514cc/HUlY5nZ9fCbq9DMyC52BrhFZACo3nI0R7qBxhQmo/d27NFe96ur/b7Wwxklda+kg=="],
 
     "@types/react": ["@types/react@19.2.17", "", { "dependencies": { "csstype": "3.2.3" } }, "sha512-MXfmqaVPEVgkBT/aY0aGCkRWWtByiYQXo3xdQ8r5RzuFrPiRn8Gar2tQdXSUQ2GKV3bkXckek89V8wQBY2Q/Aw=="],
 
@@ -785,7 +788,7 @@
 
     "@types/stream-json": ["@types/stream-json@1.7.8", "", { "dependencies": { "@types/node": "26.1.1", "@types/stream-chain": "2.1.0" } }, "sha512-MU1OB1eFLcYWd1LjwKXrxdoPtXSRzRmAnnxs4Js/ayB5O/NvHraWwuOaqMWIebpYwM6khFlsJOHEhI9xK/ab4Q=="],
 
-    "@types/three": ["@types/three@0.185.1", "", { "dependencies": { "@dimforge/rapier3d-compat": "0.12.0", "@tweenjs/tween.js": "23.1.3", "@types/stats.js": "0.17.4", "@types/webxr": "0.5.24", "fflate": "0.8.3", "meshoptimizer": "1.1.1" } }, "sha512-db1xTb+EgYF2didW+eudSvVPtn75zo+fGsY8ShQrJY/B5ZBmC2Fiaykv3aImHAlCNEGuMPkPGXBJGLwzu5mC7A=="],
+    "@types/three": ["@types/three@0.185.4", "", { "dependencies": { "@dimforge/rapier3d-compat": "~0.12.0", "@tweenjs/tween.js": "~23.1.3", "@types/stats.js": "*", "@types/webxr": ">=0.5.17", "fflate": "~0.8.2", "meshoptimizer": "~1.1.1" } }, "sha512-gAsBIC07NIFrxjbf7tH2t71c38uulFfk/RFoC7FNBSjMRAQ8J1x/RBvusX0N5PJouaYFJawXQqfCQ0RKUx/1nA=="],
 
     "@types/tmp": ["@types/tmp@0.2.6", "", {}, "sha512-chhaNf2oKHlRkDGt+tiKE2Z5aJ6qalm7Z9rlLdBwmOiAAf09YQvvoLXjWK4HWPF1xU/fqvMgfNfpVoBscA/tKA=="],
 
@@ -811,7 +814,7 @@
 
     "@vercel/analytics": ["@vercel/analytics@2.0.1", "", { "optionalDependencies": { "next": "16.2.12", "react": "19.2.8" } }, "sha512-MTQG6V9qQrt1tsDeF+2Uoo5aPjqbVPys1xvnIftXSJYG2SrwXRHnqEvVoYID7BTruDz4lCd2Z7rM1BdkUehk2g=="],
 
-    "@vercel/blob": ["@vercel/blob@2.6.1", "", { "dependencies": { "@vercel/oidc": "3.8.0", "async-retry": "1.3.3", "is-buffer": "2.0.5", "is-node-process": "1.2.0", "throttleit": "2.1.0", "undici": "6.27.0" } }, "sha512-KTJytw85j1XQBxjN5d6UXI7fIWNQe1jotn4nWN+0hePqLs+Qi1B3jHdQcSKFGF0m2rsy9uhPT6GOXMtHe3qNzg=="],
+    "@vercel/blob": ["@vercel/blob@2.8.0", "", { "dependencies": { "@vercel/oidc": "^3.6.1", "async-retry": "^1.3.3", "is-buffer": "^2.0.5", "is-node-process": "^1.2.0", "throttleit": "^2.1.0", "undici": "^6.23.0" } }, "sha512-Nu+HWKpkgovCh/ezlG7wCVwF7RErTzLzZMbGKFBdGBCbTKyK+s5VXPLl+0+TpNEQPH8AVaGzOpIsXUOtkqylCQ=="],
 
     "@vercel/cli-config": ["@vercel/cli-config@0.2.0", "", { "dependencies": { "xdg-app-paths": "5.5.1", "zod": "4.1.11" } }, "sha512-fJRRRB7734BDuXZ89yBEaA2ncYhH7bWX30mk04W80J6VAfQc+4iB8lyzAdaGpFV3/vNlkt9VZt+/uoQoWX6UsQ=="],
 
@@ -943,6 +946,8 @@
 
     "aws4": ["aws4@1.13.2", "", {}, "sha512-lHe62zvbTB5eEABUVi/AwVh0ZKY9rMMDhmm+eeyuuUQbQ3+J+fONVQOZyj+DdrvD4BY33uYniyRJ4UJIaSKAfw=="],
 
+    "axe-core": ["axe-core@4.13.0", "", {}, "sha512-UzGt8zg7Ny8djbYMhxl2zuEevVa7r2gJjYY5Lwr1xM7+XU2nd6CkIWFTVcCIbAP63vSz71NaVyyuSk9lHKcy0A=="],
+
     "bail": ["bail@2.0.2", "", {}, "sha512-0xO6mYd7JB2YesxDKplafRpsiOzPt9V02ddPCLbY1xYGPOX24NTyN50qnUxgCPcSoYMhKpAuBTjQoRZCAkUDRw=="],
 
     "balanced-match": ["balanced-match@4.0.4", "", {}, "sha512-BLrgEcRTwX2o6gGxGOCNyMvGSp35YofuYzw9h1IMTRmKqttAZZVU67bdb9Pr2vUHA8+j3i2tJfjO6C6+4myGTA=="],
@@ -961,7 +966,7 @@
 
     "body-parser": ["body-parser@2.3.0", "", { "dependencies": { "bytes": "3.1.2", "content-type": "2.0.0", "debug": "4.4.3", "http-errors": "2.0.1", "iconv-lite": "0.7.2", "on-finished": "2.4.1", "qs": "6.15.3", "raw-body": "3.0.2", "type-is": "2.1.0" } }, "sha512-2cGmJupaNgg+QUwVLAucDuWuoMZ6EX9iHDRswZ5lsNYEmwPaRknMPCLZz07yTzVq/83p4o/wzbDZbBrTvGGTIw=="],
 
-    "brace-expansion": ["brace-expansion@5.0.8", "", { "dependencies": { "balanced-match": "4.0.4" } }, "sha512-JZyDyq3D4AUifKTPOB7DELf6XsB3WdPuNxCtob1vFXPsSXhdAiHBWJ/tJ8HAc9aH84BK+5JFZLNkJKx3G9kzQg=="],
+    "brace-expansion": ["brace-expansion@5.0.9", "", { "dependencies": { "balanced-match": "^4.0.2" } }, "sha512-ScQ4IuvIEF1TMlP7Zt+vjJ//9zlPb2SDcxWxM3bk8s6t6GGdJ7KO1dCcTidOPJKePW30LE/2cT7wCyPho9/Wxg=="],
 
     "braces": ["braces@3.0.3", "", { "dependencies": { "fill-range": "7.1.1" } }, "sha512-yQbXgO/OSZVD2IsiLlro+7Hf6Q18EJrKSEsdoMzKePKXct3gvD8oLcOQdIzGupr5Fj+EDe8gO/lxc1BzfMpxvA=="],
 
@@ -1061,7 +1066,9 @@
 
     "csstype": ["csstype@3.2.3", "", {}, "sha512-z1HGKcYy2xA8AGQfwrn0PAy+PB7X/GSj3UVJW9qKyn43xWa+gl5nXmU4qqLMRzWVLFC8KusUX8T/0kCiOYpAIQ=="],
 
-    "cypress": ["cypress@15.19.0", "", { "dependencies": { "@cypress/request": "4.0.1", "@cypress/xvfb": "1.2.4", "@types/sinonjs__fake-timers": "8.1.1", "@types/sizzle": "2.3.10", "@types/tmp": "0.2.6", "arch": "2.2.0", "blob-util": "2.0.2", "bluebird": "3.7.2", "buffer": "5.7.1", "cachedir": "2.4.0", "chalk": "4.1.2", "ci-info": "4.4.0", "cli-table3": "0.6.1", "commander": "6.2.1", "common-tags": "1.8.2", "dayjs": "1.11.21", "debug": "4.4.3", "eventemitter2": "6.4.7", "execa": "4.1.0", "executable": "4.1.1", "fs-extra": "9.1.0", "hasha": "5.2.2", "is-installed-globally": "0.4.0", "listr2": "9.0.5", "lodash": "4.18.1", "log-symbols": "4.1.0", "minimist": "1.2.8", "ospath": "1.2.2", "pretty-bytes": "5.6.0", "process": "0.11.10", "proxy-from-env": "1.0.0", "request-progress": "3.0.0", "supports-color": "8.1.1", "systeminformation": "5.31.15", "tmp": "0.2.7", "tree-kill": "1.2.2", "tslib": "1.14.1", "untildify": "4.0.0", "yauzl": "3.4.0" }, "bin": { "cypress": "bin/cypress" } }, "sha512-kjy1u3SWlMRWS5qffH3U+bXx1PqPP7qwhIzEY3UtERcW2r/8zy5uk6uSwa75pYXJyYFVP1SO8mAn/avZUvUbfg=="],
+    "cypress": ["cypress@15.19.0", "", { "dependencies": { "@cypress/request": "^4.0.0", "@cypress/xvfb": "^1.2.4", "@types/sinonjs__fake-timers": "8.1.1", "@types/sizzle": "^2.3.2", "@types/tmp": "^0.2.3", "arch": "^2.2.0", "blob-util": "^2.0.2", "bluebird": "^3.7.2", "buffer": "^5.7.1", "cachedir": "^2.4.0", "chalk": "^4.1.0", "ci-info": "^4.1.0", "cli-table3": "0.6.1", "commander": "^6.2.1", "common-tags": "^1.8.0", "dayjs": "^1.10.4", "debug": "^4.3.4", "eventemitter2": "6.4.7", "execa": "4.1.0", "executable": "^4.1.1", "fs-extra": "^9.1.0", "hasha": "5.2.2", "is-installed-globally": "~0.4.0", "listr2": "^9.0.5", "lodash": "^4.17.23", "log-symbols": "^4.0.0", "minimist": "^1.2.8", "ospath": "^1.2.2", "pretty-bytes": "^5.6.0", "process": "^0.11.10", "proxy-from-env": "1.0.0", "request-progress": "^3.0.0", "supports-color": "^8.1.1", "systeminformation": "^5.31.1", "tmp": "~0.2.4", "tree-kill": "1.2.2", "tslib": "1.14.1", "untildify": "^4.0.0", "yauzl": "^3.3.1" }, "bin": { "cypress": "bin/cypress" } }, "sha512-kjy1u3SWlMRWS5qffH3U+bXx1PqPP7qwhIzEY3UtERcW2r/8zy5uk6uSwa75pYXJyYFVP1SO8mAn/avZUvUbfg=="],
+
+    "cypress-axe": ["cypress-axe@1.7.0", "", { "peerDependencies": { "axe-core": "^3 || ^4", "cypress": "^10 || ^11 || ^12 || ^13 || ^14 || ^15" } }, "sha512-zzJpvAAjauEB3GZl0KYXb8i3w6MztWAt2WM3czYTFyNVC30alDmqCm9E7GwZ4bgkldZJlmHakaVEyu73R5St4w=="],
 
     "cypress-split": ["cypress-split@1.25.0", "", { "dependencies": { "@actions/core": "2.0.3", "arg": "5.0.2", "console.table": "0.10.0", "debug": "4.4.3", "fast-shuffle": "6.1.1", "find-cypress-specs": "1.54.12", "globby": "11.1.0", "humanize-duration": "3.33.2" }, "bin": { "cypress-split-merge": "bin/merge.js", "cypress-split-preview": "bin/preview.js" } }, "sha512-/K8qQz3yeSc5GyHFyqJC6ipz3S47kwtFu7WfuZ+uSAjtfNlCxTBlPV5CTTn0v39zjWEVqrfpu2arZuzZegu/9g=="],
 
@@ -1183,7 +1190,7 @@
 
     "dir-glob": ["dir-glob@3.0.1", "", { "dependencies": { "path-type": "4.0.0" } }, "sha512-WkrWp9GR4KXfKGYzOLmTuGVi1UWFfws377n9cc55/tb6DuqyF6pcQ5AbiHEshaDpY9v6oaSr2XCDidGmMwdzIA=="],
 
-    "dompurify": ["dompurify@3.4.12", "", { "optionalDependencies": { "@types/trusted-types": "2.0.7" } }, "sha512-zQvGet8Z2sWbQhCmfFz/T5QWH2oBmjnqK3qvOjaqaNLrLEF912WamU+ohnTp0TCep/MFVHpdJuCZEdFOdTnEFg=="],
+    "dompurify": ["dompurify@3.4.13", "", { "optionalDependencies": { "@types/trusted-types": "^2.0.7" } }, "sha512-2vmYIoqjze2d+kakP8S/nS5shfsl587kzwEjcGlTdiksUVgFHnFCsLYDVj/JNqJVOQZGSYBTmuycv0PodwmnMQ=="],
 
     "dunder-proto": ["dunder-proto@1.0.1", "", { "dependencies": { "call-bind-apply-helpers": "1.0.2", "es-errors": "1.3.0", "gopd": "1.2.0" } }, "sha512-KIN/nDJBQRcXw0MLVhZE9iQHmG68qAVIBg9CqmUYjmQIhgij9U5MFvrqkUL5FbtyyzZuOeOt0zdeRe4UY7ct+A=="],
 
@@ -1225,7 +1232,7 @@
 
     "esast-util-from-js": ["esast-util-from-js@2.0.1", "", { "dependencies": { "@types/estree-jsx": "1.0.5", "acorn": "8.16.0", "esast-util-from-estree": "2.0.0", "vfile-message": "4.0.3" } }, "sha512-8Ja+rNJ0Lt56Pcf3TAmpBZjmx8ZcK5Ts4cAzIOjsjevg9oSXJnl6SUQ2EevU8tv3h6ZLWmoKL5H4fgWvdvfETw=="],
 
-    "esbuild": ["esbuild@0.28.1", "", { "optionalDependencies": { "@esbuild/aix-ppc64": "0.28.1", "@esbuild/android-arm": "0.28.1", "@esbuild/android-arm64": "0.28.1", "@esbuild/android-x64": "0.28.1", "@esbuild/darwin-arm64": "0.28.1", "@esbuild/darwin-x64": "0.28.1", "@esbuild/freebsd-arm64": "0.28.1", "@esbuild/freebsd-x64": "0.28.1", "@esbuild/linux-arm": "0.28.1", "@esbuild/linux-arm64": "0.28.1", "@esbuild/linux-ia32": "0.28.1", "@esbuild/linux-loong64": "0.28.1", "@esbuild/linux-mips64el": "0.28.1", "@esbuild/linux-ppc64": "0.28.1", "@esbuild/linux-riscv64": "0.28.1", "@esbuild/linux-s390x": "0.28.1", "@esbuild/linux-x64": "0.28.1", "@esbuild/netbsd-arm64": "0.28.1", "@esbuild/netbsd-x64": "0.28.1", "@esbuild/openbsd-arm64": "0.28.1", "@esbuild/openbsd-x64": "0.28.1", "@esbuild/openharmony-arm64": "0.28.1", "@esbuild/sunos-x64": "0.28.1", "@esbuild/win32-arm64": "0.28.1", "@esbuild/win32-ia32": "0.28.1", "@esbuild/win32-x64": "0.28.1" }, "bin": { "esbuild": "bin/esbuild" } }, "sha512-HrJrvZv5ayxBzPfwphOoNzkzOIIlifzk0KJrGK2c8R4+LKpMtpYLQeUdjnwjWv/LZlkH2laZk+4w78pi99D4Vw=="],
+    "esbuild": ["esbuild@0.28.2", "", { "optionalDependencies": { "@esbuild/aix-ppc64": "0.28.2", "@esbuild/android-arm": "0.28.2", "@esbuild/android-arm64": "0.28.2", "@esbuild/android-x64": "0.28.2", "@esbuild/darwin-arm64": "0.28.2", "@esbuild/darwin-x64": "0.28.2", "@esbuild/freebsd-arm64": "0.28.2", "@esbuild/freebsd-x64": "0.28.2", "@esbuild/linux-arm": "0.28.2", "@esbuild/linux-arm64": "0.28.2", "@esbuild/linux-ia32": "0.28.2", "@esbuild/linux-loong64": "0.28.2", "@esbuild/linux-mips64el": "0.28.2", "@esbuild/linux-ppc64": "0.28.2", "@esbuild/linux-riscv64": "0.28.2", "@esbuild/linux-s390x": "0.28.2", "@esbuild/linux-x64": "0.28.2", "@esbuild/netbsd-arm64": "0.28.2", "@esbuild/netbsd-x64": "0.28.2", "@esbuild/openbsd-arm64": "0.28.2", "@esbuild/openbsd-x64": "0.28.2", "@esbuild/openharmony-arm64": "0.28.2", "@esbuild/sunos-x64": "0.28.2", "@esbuild/win32-arm64": "0.28.2", "@esbuild/win32-ia32": "0.28.2", "@esbuild/win32-x64": "0.28.2" }, "bin": { "esbuild": "bin/esbuild" } }, "sha512-HKVLS8dvII+xoKW9kmqxbRKrnWEXfJJr/FZhhJmiqIB0e053QNYFqOBouTMO/k5sID4MvCiUCvv8b9M4h32wIA=="],
 
     "escalade": ["escalade@3.2.0", "", {}, "sha512-WUj2qlxaQtO4g6Pq5c29GTcWGDyd8itL8zTlipgECz3JesAiiOKotd8JU6otB3PACgG6xkJUyVhboMS+bje/jA=="],
 
@@ -1289,15 +1296,13 @@
 
     "extsprintf": ["extsprintf@1.3.0", "", {}, "sha512-11Ndz7Nv+mvAC1j0ktTa7fAb0vLyGGX+rMHNBYQviQDGU0Hw7lhctJANqbPhu9nV9/izT/IntTgZ7Im/9LJs9g=="],
 
-    "fake-indexeddb": ["fake-indexeddb@6.2.5", "", {}, "sha512-CGnyrvbhPlWYMngksqrSSUT1BAVP49dZocrHuK0SvtR0D5TMs5wP0o3j7jexDJW01KSadjBp1M/71o/KR3nD1w=="],
-
     "fast-deep-equal": ["fast-deep-equal@3.1.3", "", {}, "sha512-f3qQ9oQy9j2AhBe/H9VC91wLmKBCCU/gDOnKNAYG5hswO7BLKj09Hc5HYNz9cGI++xlpDCIgDaitVs03ATR84Q=="],
 
     "fast-glob": ["fast-glob@3.3.3", "", { "dependencies": { "@nodelib/fs.stat": "2.0.5", "@nodelib/fs.walk": "1.2.8", "glob-parent": "5.1.2", "merge2": "1.4.1", "micromatch": "4.0.8" } }, "sha512-7MptL8U0cqcFdzIzwOTHoilX9x5BrNqye7Z/LuC7kCMRio1EMSyqRK3BEAUD7sXRq4iT4AzTVuZdhgQ2TCvYLg=="],
 
     "fast-shuffle": ["fast-shuffle@6.1.1", "", { "dependencies": { "pcg": "1.1.0" } }, "sha512-HPxFJxEi18KPmVQuK5Hi5l4KSl3u50jtaxseRrPqrxewqfvU+sTPTaUpP33Hj+NdJoLuJP5ipx3ybTr+fa6dEw=="],
 
-    "fast-uri": ["fast-uri@4.1.1", "", {}, "sha512-YPOs1zD5TG2+EZt+r88LwF6mclA7TPkpwMP7ZN3TO2HiHS8TXvq7QA/17iJsV9dubcLo/f8eEYqMBruyQV21hQ=="],
+    "fast-uri": ["fast-uri@4.1.2", "", {}, "sha512-TyGmBcbDTZXcb2cj5MV89DrF42DKvb3y5DDUNh95iO+IMeAzMkVSxK1PZRrRIpc9yg8U2GhGdbofNa0LS/a4Bw=="],
 
     "fastq": ["fastq@1.20.1", "", { "dependencies": { "reusify": "1.1.0" } }, "sha512-GGToxJ/w1x32s/D2EKND7kTil4n8OVk/9mycTc4VDza13lOvpUZTGX3mFSCtV9ksdGBVzvsyAVLM6mHFThxXxw=="],
 
@@ -1409,7 +1414,7 @@
 
     "hastscript": ["hastscript@9.0.1", "", { "dependencies": { "@types/hast": "^3.0.0", "comma-separated-tokens": "^2.0.0", "hast-util-parse-selector": "^4.0.0", "property-information": "^7.0.0", "space-separated-tokens": "^2.0.0" } }, "sha512-g7df9rMFX/SPi34tyGCyUBREQoKkapwdY/T04Qn9TDWfHhAYt4/I0gMVirzK5wEzeUqIjEB+LXC/ypb7Aqno5w=="],
 
-    "hono": ["hono@4.12.32", "", {}, "sha512-XcuyW9qE2kJn07PkecMOBd5Vq/hMy7mmGw+idz1yblbg9N17ijJODrvPkn7/dwL3Kulj8LcRJ69DLOWf91dRUg=="],
+    "hono": ["hono@4.13.2", "", {}, "sha512-JydRilDRkYBQMt9qR9U92mXxmbGqsqSn/IKOrh4e7/gEbn+0zSr8igTu0obwJoNGN4sez28DIql7FBHWydoJpA=="],
 
     "html-encoding-sniffer": ["html-encoding-sniffer@6.0.0", "", { "dependencies": { "@exodus/bytes": "1.15.1" } }, "sha512-CV9TW3Y3f8/wT0BRFc1/KAVQ3TUHiXmaAb6VW9vtiMFf7SLoMd1PdAc4W3KFOFETBJUb90KatHqlsZMWV+R9Gg=="],
 
@@ -1443,7 +1448,7 @@
 
     "internmap": ["internmap@2.0.3", "", {}, "sha512-5Hh7Y1wQbvY5ooGgPbDaL5iYLAPzMTUrjMulskHLH6wnv/A+1q5rgEaiuqEjB+oxGXIVZs1FF+R/KPN3ZSQYYg=="],
 
-    "ip-address": ["ip-address@10.2.0", "", {}, "sha512-/+S6j4E9AHvW9SWMSEY9Xfy66O5PWvVEJ08O0y5JGyEKQpojb0K0GKpz/v5HJ/G0vi3D2sjGK78119oXZeE0qA=="],
+    "ip-address": ["ip-address@10.5.0", "", {}, "sha512-R5SnVLJmgYYvf2F2ZgwSBnelz5G4q5AxIC277GDfUaNbrZKNANcBC7RHqYYePlszf4kBolVkJauG0ZjHHFh55g=="],
 
     "ipaddr.js": ["ipaddr.js@1.9.1", "", {}, "sha512-0KI/607xoxSToH7GjN1FfSbLoU0+btTicjsQSWQlh/hZykN8KpmMf7uYwPW3R+akZ6R/w18ZlXSHBYXiYUPO3g=="],
 
@@ -1549,11 +1554,11 @@
 
     "js-tokens": ["js-tokens@10.0.0", "", {}, "sha512-lM/UBzQmfJRo9ABXbPWemivdCW8V2G8FHaHdypQaIy523snUjog0W71ayWXTjiR+ixeMyVHN2XcpnTd/liPg/Q=="],
 
-    "js-yaml": ["js-yaml@3.15.0", "", { "dependencies": { "argparse": "1.0.10", "esprima": "4.0.1" }, "bin": { "js-yaml": "bin/js-yaml.js" } }, "sha512-ttBQIIQPDeLjpPOohtUdXuXUVoA2uIB6fEH9HyJ7234s5mBJ5wTx20njxplLZQgLaOfpmPQA7X2t5AX6tIPbog=="],
+    "js-yaml": ["js-yaml@3.15.1", "", { "dependencies": { "argparse": "^1.0.7", "esprima": "^4.0.0" }, "bin": { "js-yaml": "bin/js-yaml.js" } }, "sha512-S99WuO3HlhO3XN41EtYUNl9zzXjoJx7QvmipxsJVxtCBT0YHEFy+iOJhjSvrmV12nYhWpZaM8lPHkJm0yUMbag=="],
 
     "jsbn": ["jsbn@0.1.1", "", {}, "sha512-UVU9dibq2JcFWxQPA6KCqj5O42VOmAY3zQUfEKxU0KpTGXwNoCjkX1e13eHNvw/xPynt6pU0rZ1htjWTNTSXsg=="],
 
-    "jsdom": ["jsdom@29.1.1", "", { "dependencies": { "@asamuzakjp/css-color": "5.1.11", "@asamuzakjp/dom-selector": "7.1.1", "@bramus/specificity": "2.4.2", "@csstools/css-syntax-patches-for-csstree": "1.1.6", "@exodus/bytes": "1.15.1", "css-tree": "3.2.1", "data-urls": "7.0.0", "decimal.js": "10.6.0", "html-encoding-sniffer": "6.0.0", "is-potential-custom-element-name": "1.0.1", "lru-cache": "11.5.2", "parse5": "8.0.1", "saxes": "6.0.0", "symbol-tree": "3.2.4", "tough-cookie": "6.0.2", "undici": "7.28.0", "w3c-xmlserializer": "5.0.0", "webidl-conversions": "8.0.1", "whatwg-mimetype": "5.0.0", "whatwg-url": "16.0.1", "xml-name-validator": "5.0.0" } }, "sha512-ECi4Fi2f7BdJtUKTflYRTiaMxIB0O6zfR1fX0GXpUrf6flp8QIYn1UT20YQqdSOfk2dfkCwS8LAFoJDEppNK5Q=="],
+    "jsdom": ["jsdom@30.0.1", "", { "dependencies": { "@asamuzakjp/css-color": "^6.0.5", "@asamuzakjp/dom-selector": "^8.3.0", "@bramus/specificity": "^2.4.2", "@csstools/css-syntax-patches-for-csstree": "^1.1.7", "@exodus/bytes": "^1.15.1", "css-tree": "^3.2.1", "data-urls": "^7.0.0", "decimal.js": "^10.6.0", "html-encoding-sniffer": "^6.0.0", "is-potential-custom-element-name": "^1.0.1", "lru-cache": "^11.5.2", "parse5": "^8.0.1", "saxes": "^6.0.0", "symbol-tree": "^3.2.4", "tough-cookie": "^6.0.2", "undici": "^8.9.0", "w3c-xmlserializer": "^5.0.0", "webidl-conversions": "^8.0.1", "whatwg-mimetype": "^5.0.0", "whatwg-url": "^17.1.0", "xml-name-validator": "^5.0.0" }, "peerDependencies": { "canvas": "^3.2.3" }, "optionalPeers": ["canvas"] }, "sha512-52v7mUVUfNQVYYqE1lcdaymWL0njO7lTLUog6ZvW2U5KsbiLk/GnZlVJ+qx0xfNJZ6Gn+KSpPNE52vurbxZwrA=="],
 
     "jsesc": ["jsesc@3.1.0", "", { "bin": { "jsesc": "bin/jsesc" } }, "sha512-/sM3dO2FOzXjKQhJuo0Q173wf2KOo8t4I8vHy6lF9poUp7bKT0/NHE8fPX23PwfhnykfqnC2xRxOnVw5XuGIaA=="],
 
@@ -1599,29 +1604,29 @@
 
     "lefthook-windows-x64": ["lefthook-windows-x64@2.1.10", "", { "os": "win32", "cpu": "x64" }, "sha512-c/XH8YZtylG4XaxzqFfXluvq2LXq2W/p54Bnzn3+Z7E5X2Fk3JlFJAibulMbIt2+w8T7UI/r97ok5GqE4kGaeA=="],
 
-    "lightningcss": ["lightningcss@1.32.0", "", { "dependencies": { "detect-libc": "2.1.2" }, "optionalDependencies": { "lightningcss-android-arm64": "1.32.0", "lightningcss-darwin-arm64": "1.32.0", "lightningcss-darwin-x64": "1.32.0", "lightningcss-freebsd-x64": "1.32.0", "lightningcss-linux-arm-gnueabihf": "1.32.0", "lightningcss-linux-arm64-gnu": "1.32.0", "lightningcss-linux-arm64-musl": "1.32.0", "lightningcss-linux-x64-gnu": "1.32.0", "lightningcss-linux-x64-musl": "1.32.0", "lightningcss-win32-arm64-msvc": "1.32.0", "lightningcss-win32-x64-msvc": "1.32.0" } }, "sha512-NXYBzinNrblfraPGyrbPoD19C1h9lfI/1mzgWYvXUTe414Gz/X1FD2XBZSZM7rRTrMA8JL3OtAaGifrIKhQ5yQ=="],
+    "lightningcss": ["lightningcss@1.33.0", "", { "dependencies": { "detect-libc": "^2.0.3" }, "optionalDependencies": { "lightningcss-android-arm64": "1.33.0", "lightningcss-darwin-arm64": "1.33.0", "lightningcss-darwin-x64": "1.33.0", "lightningcss-freebsd-x64": "1.33.0", "lightningcss-linux-arm-gnueabihf": "1.33.0", "lightningcss-linux-arm64-gnu": "1.33.0", "lightningcss-linux-arm64-musl": "1.33.0", "lightningcss-linux-x64-gnu": "1.33.0", "lightningcss-linux-x64-musl": "1.33.0", "lightningcss-win32-arm64-msvc": "1.33.0", "lightningcss-win32-x64-msvc": "1.33.0" } }, "sha512-WkUDrojuJs0xkgGf2udWxa3yGBRxPtxUkB79i6aCZLRgc7PM8fZe9TosfPDcvEpQZbuFASnHYmRLBLUbmLOIIA=="],
 
-    "lightningcss-android-arm64": ["lightningcss-android-arm64@1.32.0", "", { "os": "android", "cpu": "arm64" }, "sha512-YK7/ClTt4kAK0vo6w3X+Pnm0D2cf2vPHbhOXdoNti1Ga0al1P4TBZhwjATvjNwLEBCnKvjJc2jQgHXH0NEwlAg=="],
+    "lightningcss-android-arm64": ["lightningcss-android-arm64@1.33.0", "", { "os": "android", "cpu": "arm64" }, "sha512-gEpRTalKdosp4Bb8qWtc2iOgE5SeIHlpS1up9bFq2wAyYhl1UdTObYiHe98zEM9SQvSoqQZ1IQD0JNpg3Ml5pg=="],
 
-    "lightningcss-darwin-arm64": ["lightningcss-darwin-arm64@1.32.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-RzeG9Ju5bag2Bv1/lwlVJvBE3q6TtXskdZLLCyfg5pt+HLz9BqlICO7LZM7VHNTTn/5PRhHFBSjk5lc4cmscPQ=="],
+    "lightningcss-darwin-arm64": ["lightningcss-darwin-arm64@1.33.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-Sciaz8eenNTKn9b3t7+xr0ipTp9YxKQY4npwQ3mrRuL0BAVHBLyZxofhaKBAVtzmtRZ/zTyo0/to4B1uWG/Djg=="],
 
-    "lightningcss-darwin-x64": ["lightningcss-darwin-x64@1.32.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-U+QsBp2m/s2wqpUYT/6wnlagdZbtZdndSmut/NJqlCcMLTWp5muCrID+K5UJ6jqD2BFshejCYXniPDbNh73V8w=="],
+    "lightningcss-darwin-x64": ["lightningcss-darwin-x64@1.33.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-Z5UPAxzrjlWNNyGy6i65cJzzvgJ5D3T6wMvs+gWpY9d7qRhANrxqAp6LhxIgZhWEw18RfJTGcRxjuLIBr+m8XQ=="],
 
-    "lightningcss-freebsd-x64": ["lightningcss-freebsd-x64@1.32.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-JCTigedEksZk3tHTTthnMdVfGf61Fky8Ji2E4YjUTEQX14xiy/lTzXnu1vwiZe3bYe0q+SpsSH/CTeDXK6WHig=="],
+    "lightningcss-freebsd-x64": ["lightningcss-freebsd-x64@1.33.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-QQM/Ti/hQajJwCY+RiWuCZ9sdtI/XQk7nDK5vC8kkdwixezOlDgvDx7+RT+QjK6FcFT4MpsuoBnHIo/O3StRRg=="],
 
-    "lightningcss-linux-arm-gnueabihf": ["lightningcss-linux-arm-gnueabihf@1.32.0", "", { "os": "linux", "cpu": "arm" }, "sha512-x6rnnpRa2GL0zQOkt6rts3YDPzduLpWvwAF6EMhXFVZXD4tPrBkEFqzGowzCsIWsPjqSK+tyNEODUBXeeVHSkw=="],
+    "lightningcss-linux-arm-gnueabihf": ["lightningcss-linux-arm-gnueabihf@1.33.0", "", { "os": "linux", "cpu": "arm" }, "sha512-N7FVBe6iS24MlM6R/4RBTxGhQheZGs7tiQ9U32UtF75NzP5Q7xWPRqLBCKxlRQRk3rY1jCIPLzx7WzOhuUIRLQ=="],
 
-    "lightningcss-linux-arm64-gnu": ["lightningcss-linux-arm64-gnu@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-0nnMyoyOLRJXfbMOilaSRcLH3Jw5z9HDNGfT/gwCPgaDjnx0i8w7vBzFLFR1f6CMLKF8gVbebmkUN3fa/kQJpQ=="],
+    "lightningcss-linux-arm64-gnu": ["lightningcss-linux-arm64-gnu@1.33.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-j2v/itmy4HlNxlc6voKXYgBqNi0Ng2LShg4z7GufpEgs05P+2suBVyi9I6YHq5uoVFx9ETin3eCEhLVyXGQnKg=="],
 
-    "lightningcss-linux-arm64-musl": ["lightningcss-linux-arm64-musl@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-UpQkoenr4UJEzgVIYpI80lDFvRmPVg6oqboNHfoH4CQIfNA+HOrZ7Mo7KZP02dC6LjghPQJeBsvXhJod/wnIBg=="],
+    "lightningcss-linux-arm64-musl": ["lightningcss-linux-arm64-musl@1.33.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-yiO5ROMuYQgXbC60yjZU5CYSFZGKXL0HFATXt9mHJn1+zW55oCtMI9NfcVhYLMFDL7gV7oBPon/EmMMGg2OvtQ=="],
 
-    "lightningcss-linux-x64-gnu": ["lightningcss-linux-x64-gnu@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-V7Qr52IhZmdKPVr+Vtw8o+WLsQJYCTd8loIfpDaMRWGUZfBOYEJeyJIkqGIDMZPwPx24pUMfwSxxI8phr/MbOA=="],
+    "lightningcss-linux-x64-gnu": ["lightningcss-linux-x64-gnu@1.33.0", "", { "os": "linux", "cpu": "x64" }, "sha512-ar+Ju7LmcN0Jo4FpL4hpFybwNG9/3A/Br5KW2n2jyODg3MEZXaDYADdemoNS+BDNfMgKvylJLj4S5tyRActuAg=="],
 
-    "lightningcss-linux-x64-musl": ["lightningcss-linux-x64-musl@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-bYcLp+Vb0awsiXg/80uCRezCYHNg1/l3mt0gzHnWV9XP1W5sKa5/TCdGWaR/zBM2PeF/HbsQv/j2URNOiVuxWg=="],
+    "lightningcss-linux-x64-musl": ["lightningcss-linux-x64-musl@1.33.0", "", { "os": "linux", "cpu": "x64" }, "sha512-RYiYbkokw0trfKqqzfF55lginwEPrD3OJDfTuJzFs1MK6iFnDenaz1fqLLtX4ITG3OktJQXOeTaw1awrBAlZPw=="],
 
-    "lightningcss-win32-arm64-msvc": ["lightningcss-win32-arm64-msvc@1.32.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-8SbC8BR40pS6baCM8sbtYDSwEVQd4JlFTOlaD3gWGHfThTcABnNDBda6eTZeqbofalIJhFx0qKzgHJmcPTnGdw=="],
+    "lightningcss-win32-arm64-msvc": ["lightningcss-win32-arm64-msvc@1.33.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-1K+MPfLSFVpphzpdbfkhlWk6wBrTObBzS2T6db10PNOZgR9GoVsAWzwNyuhUYYbTp23j+4RrncfujZ4uAzXvwA=="],
 
-    "lightningcss-win32-x64-msvc": ["lightningcss-win32-x64-msvc@1.32.0", "", { "os": "win32", "cpu": "x64" }, "sha512-Amq9B/SoZYdDi1kFrojnoqPLxYhQ4Wo5XiL8EVJrVsB8ARoC1PWW6VGtT0WKCemjy8aC+louJnjS7U18x3b06Q=="],
+    "lightningcss-win32-x64-msvc": ["lightningcss-win32-x64-msvc@1.33.0", "", { "os": "win32", "cpu": "x64" }, "sha512-OlEICDx/Xl0FqSp4bry8zFnCvGpig3Gl4gCquvYwHuqJKEC1+n9NgDniFvqHGmMv1ZkqDJrDqKKSykTDX+ehuA=="],
 
     "listr2": ["listr2@9.0.5", "", { "dependencies": { "cli-truncate": "5.2.0", "colorette": "2.0.20", "eventemitter3": "5.0.4", "log-update": "6.1.0", "rfdc": "1.4.1", "wrap-ansi": "9.0.2" } }, "sha512-ME4Fb83LgEgwNw96RKNvKV4VTLuXfoKudAmm2lP8Kk87KaMK0/Xrx/aAkMWmT8mDb+3MlFDspfbCs7adjRxA2g=="],
 
@@ -1641,7 +1646,7 @@
 
     "lru-cache": ["lru-cache@11.5.2", "", {}, "sha512-4pfM1Ff0x50o0tQwb5ucw/RzNyD0/YJME6IVcStalZuMWxdt3sR3huStTtxz4PUmvZfRguvDejasvQ2kifR11g=="],
 
-    "lucide-react": ["lucide-react@1.27.0", "", { "peerDependencies": { "react": "19.2.8" } }, "sha512-rJicGl/3Fly/E0rOH1YmPZ6e49JCnKknh1ox1vpHnkfjujAkKA6sqUZvH3MTAaXXjgexyUwgNwTJzTtYuAFYJw=="],
+    "lucide-react": ["lucide-react@1.31.0", "", { "peerDependencies": { "react": "^16.5.1 || ^17.0.0 || ^18.0.0 || ^19.0.0" } }, "sha512-G8u2eEtoHUnUa9f8lbvqDhCiORMnYLdUEo06EEG9MQvHQrInKcX3Pa2TH39MM5qyzRcWETxB0+aOwAPI1g1kEg=="],
 
     "magic-string": ["magic-string@0.30.21", "", { "dependencies": { "@jridgewell/sourcemap-codec": "1.5.5" } }, "sha512-vd2F4YUyEXKGcLHoq+TEyCjxueSeHnFxyyjNp80yg0XV4vUhnDer/lvvlqM/arB5bXQN5K2/3oinyCRyx8T2CQ=="],
 
@@ -1797,13 +1802,13 @@
 
     "ms": ["ms@2.1.3", "", {}, "sha512-6FlzubTLZG3J2a/NVCAleEhjzq5oxgHyaCU9yYXvcLsvoVaHJq/s5xXI6/XXP6tz7R9xAOtHnSO/tXtF3WRTlA=="],
 
-    "nanoid": ["nanoid@3.3.16", "", { "bin": { "nanoid": "bin/nanoid.cjs" } }, "sha512-bzlKTyNJ7+LdGIIwy8ijFpIqEQIvafahV7eYykJ8Cvh42EdJeODoJ6gUJXpQJvej1BddH8OqTXZNE/KfbWAu8Q=="],
+    "nanoid": ["nanoid@3.3.18", "", { "bin": { "nanoid": "bin/nanoid.cjs" } }, "sha512-DTg4MJbGMWkfi6VZFdNt2/caMbQy4Ou+Op/hJQvGEWcnVfoA1QA+xzRKAzw9jD6+GVOOeYr/mIcuDSdug6F6+w=="],
 
     "negotiator": ["negotiator@1.0.0", "", {}, "sha512-8Ofs/AUQh8MaEcrlq5xOX0CQ9ypTF5dl78mjlMNfOK08fzpgTHQRQPBxcPlEtIw0yRpws+Zo/3r+5WRby7u3Gg=="],
 
     "neo-async": ["neo-async@2.6.2", "", {}, "sha512-Yd3UES5mWCSqR+qNT93S3UoYUkqAZ9lLg8a7g9rimsWmYGK8cVToA4/sF3RrshdyV3sAGMXVUmpMYOw+dLpOuw=="],
 
-    "next": ["next@16.2.12", "", { "dependencies": { "@next/env": "16.2.12", "@swc/helpers": "0.5.15", "baseline-browser-mapping": "2.10.42", "caniuse-lite": "1.0.30001803", "postcss": "8.5.23", "styled-jsx": "5.1.6" }, "optionalDependencies": { "@next/swc-darwin-arm64": "16.2.12", "@next/swc-darwin-x64": "16.2.12", "@next/swc-linux-arm64-gnu": "16.2.12", "@next/swc-linux-arm64-musl": "16.2.12", "@next/swc-linux-x64-gnu": "16.2.12", "@next/swc-linux-x64-musl": "16.2.12", "@next/swc-win32-arm64-msvc": "16.2.12", "@next/swc-win32-x64-msvc": "16.2.12", "@opentelemetry/api": "1.9.1", "sharp": "0.35.3" }, "peerDependencies": { "react": "19.2.8", "react-dom": "19.2.8" }, "bin": { "next": "dist/bin/next" } }, "sha512-iD59eYQWmbFcEbX7v/acG5DRym9iw1DdaPoD0WTA920naWsE25wShzJW4+UvAs8MK9EC2kBfIH6vtto1H1PHGw=="],
+    "next": ["next@16.3.1", "", { "dependencies": { "@next/env": "16.3.1", "@swc/helpers": "0.5.23", "baseline-browser-mapping": "^2.9.19", "caniuse-lite": "^1.0.30001579", "postcss": "8.5.23", "styled-jsx": "5.1.6" }, "optionalDependencies": { "@next/swc-darwin-arm64": "16.3.1", "@next/swc-darwin-x64": "16.3.1", "@next/swc-linux-arm64-gnu": "16.3.1", "@next/swc-linux-arm64-musl": "16.3.1", "@next/swc-linux-x64-gnu": "16.3.1", "@next/swc-linux-x64-musl": "16.3.1", "@next/swc-win32-arm64-msvc": "16.3.1", "@next/swc-win32-x64-msvc": "16.3.1", "sharp": "^0.35.3" }, "peerDependencies": { "@opentelemetry/api": "^1.1.0", "@playwright/test": "^1.51.1", "babel-plugin-react-compiler": "*", "react": "^18.2.0 || 19.0.0-rc-de68d2f4-20241204 || ^19.0.0", "react-dom": "^18.2.0 || 19.0.0-rc-de68d2f4-20241204 || ^19.0.0", "sass": "^1.3.0" }, "optionalPeers": ["@opentelemetry/api", "@playwright/test", "babel-plugin-react-compiler", "sass"], "bin": { "next": "dist/bin/next" } }, "sha512-hsAp0i7Rh+/dhe7DGIeN2YlpLM1DP4MNxti9EtDMtqcO612X81MvvEj388/oTce9U1EcEIOWDlGq0zRwrBKvuA=="],
 
     "next-mdx-remote": ["next-mdx-remote@6.0.0", "", { "dependencies": { "@babel/code-frame": "7.29.0", "@mdx-js/mdx": "3.1.1", "@mdx-js/react": "3.1.1", "unist-util-remove": "4.0.0", "unist-util-visit": "5.1.0", "vfile": "6.0.3", "vfile-matter": "5.0.1" }, "peerDependencies": { "react": "19.2.8" } }, "sha512-cJEpEZlgD6xGjB4jL8BnI8FaYdN9BzZM4NwadPe1YQr7pqoWjg9EBCMv3nXBkuHqMRfv2y33SzUsuyNh9LFAQQ=="],
 
@@ -1843,9 +1848,9 @@
 
     "ospath": ["ospath@1.2.2", "", {}, "sha512-o6E5qJV5zkAbIDNhGSIlyOhScKXgQrSRMilfph0clDfM0nEnBOlKlH4sWDmG95BW/CvwNz0vmm7dJVtU2KlMiA=="],
 
-    "oxfmt": ["oxfmt@0.60.0", "", { "dependencies": { "tinypool": "2.1.0" }, "optionalDependencies": { "@oxfmt/binding-android-arm-eabi": "0.60.0", "@oxfmt/binding-android-arm64": "0.60.0", "@oxfmt/binding-darwin-arm64": "0.60.0", "@oxfmt/binding-darwin-x64": "0.60.0", "@oxfmt/binding-freebsd-x64": "0.60.0", "@oxfmt/binding-linux-arm-gnueabihf": "0.60.0", "@oxfmt/binding-linux-arm-musleabihf": "0.60.0", "@oxfmt/binding-linux-arm64-gnu": "0.60.0", "@oxfmt/binding-linux-arm64-musl": "0.60.0", "@oxfmt/binding-linux-ppc64-gnu": "0.60.0", "@oxfmt/binding-linux-riscv64-gnu": "0.60.0", "@oxfmt/binding-linux-riscv64-musl": "0.60.0", "@oxfmt/binding-linux-s390x-gnu": "0.60.0", "@oxfmt/binding-linux-x64-gnu": "0.60.0", "@oxfmt/binding-linux-x64-musl": "0.60.0", "@oxfmt/binding-openharmony-arm64": "0.60.0", "@oxfmt/binding-win32-arm64-msvc": "0.60.0", "@oxfmt/binding-win32-ia32-msvc": "0.60.0", "@oxfmt/binding-win32-x64-msvc": "0.60.0" }, "bin": { "oxfmt": "bin/oxfmt" } }, "sha512-fViX6i+gJuZWY+jI/fnR6WRbRj70GZ9RlCd30MygJrHTUNc4DxvKHWw8vBjMjffv3PgU5qWDR0AzmojQByqaZA=="],
+    "oxfmt": ["oxfmt@0.63.0", "", { "dependencies": { "tinypool": "2.1.0" }, "optionalDependencies": { "@oxfmt/binding-android-arm-eabi": "0.63.0", "@oxfmt/binding-android-arm64": "0.63.0", "@oxfmt/binding-darwin-arm64": "0.63.0", "@oxfmt/binding-darwin-x64": "0.63.0", "@oxfmt/binding-freebsd-x64": "0.63.0", "@oxfmt/binding-linux-arm-gnueabihf": "0.63.0", "@oxfmt/binding-linux-arm-musleabihf": "0.63.0", "@oxfmt/binding-linux-arm64-gnu": "0.63.0", "@oxfmt/binding-linux-arm64-musl": "0.63.0", "@oxfmt/binding-linux-ppc64-gnu": "0.63.0", "@oxfmt/binding-linux-riscv64-gnu": "0.63.0", "@oxfmt/binding-linux-riscv64-musl": "0.63.0", "@oxfmt/binding-linux-s390x-gnu": "0.63.0", "@oxfmt/binding-linux-x64-gnu": "0.63.0", "@oxfmt/binding-linux-x64-musl": "0.63.0", "@oxfmt/binding-openharmony-arm64": "0.63.0", "@oxfmt/binding-win32-arm64-msvc": "0.63.0", "@oxfmt/binding-win32-ia32-msvc": "0.63.0", "@oxfmt/binding-win32-x64-msvc": "0.63.0" }, "peerDependencies": { "svelte": "^5.0.0", "vite-plus": "*" }, "optionalPeers": ["svelte", "vite-plus"], "bin": { "oxfmt": "bin/oxfmt" } }, "sha512-kgdDwv35wvVf6554U2Ab8Jnd0zTM+TsEQWwaB70RAjK3gICFAFGO+2Hd3Be27GMoXj3XRL9IKSNRVl7KBQL6iw=="],
 
-    "oxlint": ["oxlint@1.75.0", "", { "optionalDependencies": { "@oxlint/binding-android-arm-eabi": "1.75.0", "@oxlint/binding-android-arm64": "1.75.0", "@oxlint/binding-darwin-arm64": "1.75.0", "@oxlint/binding-darwin-x64": "1.75.0", "@oxlint/binding-freebsd-x64": "1.75.0", "@oxlint/binding-linux-arm-gnueabihf": "1.75.0", "@oxlint/binding-linux-arm-musleabihf": "1.75.0", "@oxlint/binding-linux-arm64-gnu": "1.75.0", "@oxlint/binding-linux-arm64-musl": "1.75.0", "@oxlint/binding-linux-ppc64-gnu": "1.75.0", "@oxlint/binding-linux-riscv64-gnu": "1.75.0", "@oxlint/binding-linux-riscv64-musl": "1.75.0", "@oxlint/binding-linux-s390x-gnu": "1.75.0", "@oxlint/binding-linux-x64-gnu": "1.75.0", "@oxlint/binding-linux-x64-musl": "1.75.0", "@oxlint/binding-openharmony-arm64": "1.75.0", "@oxlint/binding-win32-arm64-msvc": "1.75.0", "@oxlint/binding-win32-ia32-msvc": "1.75.0", "@oxlint/binding-win32-x64-msvc": "1.75.0" }, "bin": { "oxlint": "bin/oxlint" } }, "sha512-m9WzjRcRYA/uqIZDa9tclrieoPJ/ln1QYTKdFx6NUOs8uY5DiHlIwRQoCrHT6OM6O3ww3l2skY5gO7G7ZphE7g=="],
+    "oxlint": ["oxlint@1.78.0", "", { "optionalDependencies": { "@oxlint/binding-android-arm-eabi": "1.78.0", "@oxlint/binding-android-arm64": "1.78.0", "@oxlint/binding-darwin-arm64": "1.78.0", "@oxlint/binding-darwin-x64": "1.78.0", "@oxlint/binding-freebsd-x64": "1.78.0", "@oxlint/binding-linux-arm-gnueabihf": "1.78.0", "@oxlint/binding-linux-arm-musleabihf": "1.78.0", "@oxlint/binding-linux-arm64-gnu": "1.78.0", "@oxlint/binding-linux-arm64-musl": "1.78.0", "@oxlint/binding-linux-ppc64-gnu": "1.78.0", "@oxlint/binding-linux-riscv64-gnu": "1.78.0", "@oxlint/binding-linux-riscv64-musl": "1.78.0", "@oxlint/binding-linux-s390x-gnu": "1.78.0", "@oxlint/binding-linux-x64-gnu": "1.78.0", "@oxlint/binding-linux-x64-musl": "1.78.0", "@oxlint/binding-openharmony-arm64": "1.78.0", "@oxlint/binding-win32-arm64-msvc": "1.78.0", "@oxlint/binding-win32-ia32-msvc": "1.78.0", "@oxlint/binding-win32-x64-msvc": "1.78.0" }, "peerDependencies": { "oxlint-tsgolint": ">=7.0.2001", "vite-plus": "*" }, "optionalPeers": ["oxlint-tsgolint", "vite-plus"], "bin": { "oxlint": "bin/oxlint" } }, "sha512-QgQePuxIqKOzo1KSjG2EnITEeWvWnKAm77eq8nrMtf6AGoA+zyGc4PFYtDNJSD25g/ibOwfQ851hZ4/SPkMVoA=="],
 
     "package-json-from-dist": ["package-json-from-dist@1.0.1", "", {}, "sha512-UEZIS3/by4OC8vL3P2dTXRETpebLI2NiI5vIrjaD/5UtrkFX/tNbwjTSRAGC/+7CAo2pIcBaRgWmcBBHcsaCIw=="],
 
@@ -1885,7 +1890,7 @@
 
     "possible-typed-array-names": ["possible-typed-array-names@1.1.0", "", {}, "sha512-/+5VFTchJDoVj3bhoqi6UeymcD00DAwb1nJwamzPvHEszJ4FpF6SNNbUbOS8yI56qHzdV8eK0qEfOSiodkTdxg=="],
 
-    "postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+    "postcss": ["postcss@8.5.26", "", { "dependencies": { "nanoid": "^3.3.17", "picocolors": "^1.1.1", "source-map-js": "^1.2.1" } }, "sha512-u82N74LFzG8ca+dD8puPnplTXoGH4fTPpVGuIbt36G3qvNlkvfD0lEAZSxaly3KX8TS/L1A1gsCEmvKmBcVbkQ=="],
 
     "postcss-selector-parser": ["postcss-selector-parser@6.0.10", "", { "dependencies": { "cssesc": "3.0.0", "util-deprecate": "1.0.2" } }, "sha512-IQ7TZdoaqbT+LCpShg46jnZVlhWD2w6iQYAcYXfHARZ7X1t/UGhhceQDs5X0cGqKvYlHNOuv7Oa1xmb0oQuA3w=="],
 
@@ -1997,7 +2002,7 @@
 
     "robust-predicates": ["robust-predicates@3.0.3", "", {}, "sha512-NS3levdsRIUOmiJ8FZWCP7LG3QpJyrs/TE0Zpf1yvZu8cAJJ6QMW92H1c7kWpdIHo8RvmLxN/o2JXTKHp74lUA=="],
 
-    "rolldown": ["rolldown@1.1.5", "", { "dependencies": { "@oxc-project/types": "0.139.0", "@rolldown/pluginutils": "1.0.1" }, "optionalDependencies": { "@rolldown/binding-android-arm64": "1.1.5", "@rolldown/binding-darwin-arm64": "1.1.5", "@rolldown/binding-darwin-x64": "1.1.5", "@rolldown/binding-freebsd-x64": "1.1.5", "@rolldown/binding-linux-arm-gnueabihf": "1.1.5", "@rolldown/binding-linux-arm64-gnu": "1.1.5", "@rolldown/binding-linux-arm64-musl": "1.1.5", "@rolldown/binding-linux-ppc64-gnu": "1.1.5", "@rolldown/binding-linux-s390x-gnu": "1.1.5", "@rolldown/binding-linux-x64-gnu": "1.1.5", "@rolldown/binding-linux-x64-musl": "1.1.5", "@rolldown/binding-openharmony-arm64": "1.1.5", "@rolldown/binding-wasm32-wasi": "1.1.5", "@rolldown/binding-win32-arm64-msvc": "1.1.5", "@rolldown/binding-win32-x64-msvc": "1.1.5" }, "bin": { "rolldown": "./bin/cli.mjs" } }, "sha512-t9z29cJjXf/vxQ8dyhCSpt6H6aSwHTk8cT5I3iy6SMXuFpk5mB6PL6XfC8PCwrPTx93udwKUm9HRteAlTGBLiA=="],
+    "rolldown": ["rolldown@1.2.4", "", { "dependencies": { "@oxc-project/types": "=0.144.0", "@rolldown/pluginutils": "^1.0.0" }, "optionalDependencies": { "@rolldown/binding-android-arm64": "1.2.4", "@rolldown/binding-darwin-arm64": "1.2.4", "@rolldown/binding-darwin-x64": "1.2.4", "@rolldown/binding-freebsd-x64": "1.2.4", "@rolldown/binding-linux-arm-gnueabihf": "1.2.4", "@rolldown/binding-linux-arm64-gnu": "1.2.4", "@rolldown/binding-linux-arm64-musl": "1.2.4", "@rolldown/binding-linux-ppc64-gnu": "1.2.4", "@rolldown/binding-linux-s390x-gnu": "1.2.4", "@rolldown/binding-linux-x64-gnu": "1.2.4", "@rolldown/binding-linux-x64-musl": "1.2.4", "@rolldown/binding-openharmony-arm64": "1.2.4", "@rolldown/binding-win32-arm64-msvc": "1.2.4", "@rolldown/binding-win32-x64-msvc": "1.2.4" }, "bin": { "rolldown": "./bin/cli.mjs" } }, "sha512-rSr7irW0K7QRWzjdJXqZowkcRdDtjRduh43rBltnVKd0VFq839l1lJoDvGJb6gl7+4rTTCrPWu+YfujUL8Ug7w=="],
 
     "router": ["router@2.2.0", "", { "dependencies": { "debug": "4.4.3", "depd": "2.0.0", "is-promise": "4.0.0", "parseurl": "1.3.3", "path-to-regexp": "8.4.2" } }, "sha512-nLTrUKm2UyiL7rlhapu/Zl45FwNgkZGaCpZbIHajDYgwlJCOzLSk+cIPAnsEqV955GjILJnKbdQC1nVPz+gAYQ=="],
 
@@ -2041,7 +2046,7 @@
 
     "shelljs": ["shelljs@0.10.0", "", { "dependencies": { "execa": "5.1.1", "fast-glob": "3.3.3" } }, "sha512-Jex+xw5Mg2qMZL3qnzXIfaxEtBaC4n7xifqaqtrZDdlheR70OGkydrPJWT0V1cA1k3nanC86x9FwAmQl6w3Klw=="],
 
-    "shiki": ["shiki@4.3.1", "", { "dependencies": { "@shikijs/core": "4.3.1", "@shikijs/engine-javascript": "4.3.1", "@shikijs/engine-oniguruma": "4.3.1", "@shikijs/langs": "4.3.1", "@shikijs/themes": "4.3.1", "@shikijs/types": "4.3.1", "@shikijs/vscode-textmate": "10.0.2", "@types/hast": "3.0.4" } }, "sha512-oR+qDVi2OjX1tmDpyv+3KviX01KzO6Af+0NNnKnsp9491UEGz2YpxTuJboS/6VhYpTdqzmuJBuiTlrAWWJAssw=="],
+    "shiki": ["shiki@4.4.3", "", { "dependencies": { "@shikijs/core": "4.4.3", "@shikijs/engine-javascript": "4.4.3", "@shikijs/engine-oniguruma": "4.4.3", "@shikijs/langs": "4.4.3", "@shikijs/themes": "4.4.3", "@shikijs/types": "4.4.3", "@shikijs/vscode-textmate": "^10.0.2", "@types/hast": "^3.0.5" } }, "sha512-Mb/GvXPHBAXdgGIcnfU5L3ldpn1XcxrGkPHwqgRx17/I2XRfqlFKk2vGkHWINn1kdXvzJZeuO3is6I9KLPFm0g=="],
 
     "side-channel": ["side-channel@1.1.1", "", { "dependencies": { "es-errors": "1.3.0", "object-inspect": "1.13.4", "side-channel-list": "1.0.1", "side-channel-map": "1.0.1", "side-channel-weakmap": "1.0.2" } }, "sha512-6x6dK6zJdpTzF4sQeNYxwtvBzf6Eg4GtlesS94HOvTudUeyK2WXAaIfmDgsyslYrRBeFIlsi54AYsFGUuhmvrQ=="],
 
@@ -2185,7 +2190,7 @@
 
     "typescript": ["typescript@6.0.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-y2TvuxSZPDyQakkFRPZHKFm+KKVqIisdg9/CZwm9ftvKXLP8NRWj38/ODjNbr43SsoXqNuAisEf1GdCxqWcdBw=="],
 
-    "undici": ["undici@7.28.0", "", {}, "sha512-cRZYrTDwWznlnRiPjggAGxZXanty6M8RV1ff8Wm4LWXBp7/IG8v5DnOm74DtUBp9OONpK75YlPnIjQqX0dBDtA=="],
+    "undici": ["undici@8.10.0", "", {}, "sha512-HvltHd7avK13QIw/oLe4qoOLyoVSoafqJ2jYOrtMRBkbYT31eiBQ8O0ehRKZiEZCMEyLFQNIADpgCWC5fALvYQ=="],
 
     "undici-types": ["undici-types@8.3.0", "", {}, "sha512-j375ScV60dom+YkPFIfTLcOiPxkN/buHz5GobjLhixFuANaNs3C9l4GmrWqejgXWJ7BbJcFYpTEUkS1Ge8bpZQ=="],
 
@@ -2237,7 +2242,7 @@
 
     "vfile-message": ["vfile-message@4.0.3", "", { "dependencies": { "@types/unist": "3.0.3", "unist-util-stringify-position": "4.0.0" } }, "sha512-QTHzsGd1EhbZs4AsQ20JX1rC3cOlt/IWJruk893DfLRr57lcnOeMaWG4K0JrRta4mIJZKth2Au3mM3u03/JWKw=="],
 
-    "vite": ["vite@8.1.5", "", { "dependencies": { "lightningcss": "1.32.0", "picomatch": "4.0.5", "postcss": "8.5.23", "rolldown": "1.1.5", "tinyglobby": "0.2.17" }, "optionalDependencies": { "@types/node": "26.1.1", "esbuild": "0.28.1", "fsevents": "2.3.3", "jiti": "2.7.0", "terser": "5.49.0", "tsx": "4.23.1", "yaml": "2.9.0" }, "bin": { "vite": "bin/vite.js" } }, "sha512-7ULLwsCdYx/nRyrpiEwvqb5TFHrMVZyBt+rg/OAXT7rgj/z+DtTDyKFeLAdDkubDVDKD8jOsndmy7m55XcfUsw=="],
+    "vite": ["vite@8.2.1", "", { "dependencies": { "lightningcss": "^1.33.0", "picomatch": "^4.0.5", "postcss": "^8.5.25", "rolldown": "~1.2.1", "tinyglobby": "^0.2.17" }, "optionalDependencies": { "fsevents": "~2.3.3" }, "peerDependencies": { "@types/node": "^20.19.0 || >=22.12.0", "@vitejs/devtools": "^0.4.0", "esbuild": "^0.27.0 || ^0.28.0", "jiti": ">=1.21.0", "less": "^4.0.0", "sass": "^1.70.0", "sass-embedded": "^1.70.0", "stylus": ">=0.54.8", "sugarss": "^5.0.0", "terser": "^5.16.0", "tsx": "^4.8.1", "yaml": "^2.4.2" }, "optionalPeers": ["@types/node", "@vitejs/devtools", "esbuild", "jiti", "less", "sass", "sass-embedded", "stylus", "sugarss", "terser", "tsx", "yaml"], "bin": { "vite": "bin/vite.js" } }, "sha512-EU/eS7BH3XROHh2YnBefjM6DBKA6ZeMZEYQbj7NLWg5wHYlhB8B/Mayd5XsgWq+NFYccDOTemRpdETWR6Ka/lw=="],
 
     "vitest": ["vitest@4.1.10", "", { "dependencies": { "@vitest/expect": "4.1.10", "@vitest/mocker": "4.1.10", "@vitest/pretty-format": "4.1.10", "@vitest/runner": "4.1.10", "@vitest/snapshot": "4.1.10", "@vitest/spy": "4.1.10", "@vitest/utils": "4.1.10", "es-module-lexer": "2.3.0", "expect-type": "1.4.0", "magic-string": "0.30.21", "obug": "2.1.3", "pathe": "2.0.3", "picomatch": "4.0.5", "std-env": "4.2.0", "tinybench": "2.9.0", "tinyexec": "1.2.4", "tinyglobby": "0.2.17", "tinyrainbow": "3.1.0", "why-is-node-running": "2.3.0" }, "optionalDependencies": { "@opentelemetry/api": "1.9.1", "@types/node": "26.1.1", "@vitest/coverage-v8": "4.1.10", "jsdom": "29.1.1" }, "peerDependencies": { "vite": "8.1.5" }, "bin": { "vitest": "./vitest.mjs" } }, "sha512-R9jUTe5S4Qb0HCd4TNqpC7oGcrMssMRGXLW80ubjWsW9VH5GF8y1Y0SFLY9AbqSk6nt0PnOx4H4WNJYZ13GUPw=="],
 
@@ -2259,7 +2264,7 @@
 
     "whatwg-mimetype": ["whatwg-mimetype@5.0.0", "", {}, "sha512-sXcNcHOC51uPGF0P/D4NVtrkjSU2fNsm9iog4ZvZJsL3rjoDAzXZhkm2MWt1y+PUdggKAYVoMAIYcs78wJ51Cw=="],
 
-    "whatwg-url": ["whatwg-url@16.0.1", "", { "dependencies": { "@exodus/bytes": "1.15.1", "tr46": "6.0.0", "webidl-conversions": "8.0.1" } }, "sha512-1to4zXBxmXHV3IiSSEInrreIlu02vUOvrhxJJH5vcxYTBDAx51cqZiKdyTxlecdKNSjj8EcxGBxNf6Vg+945gw=="],
+    "whatwg-url": ["whatwg-url@17.1.0", "", { "dependencies": { "@exodus/bytes": "^1.15.1", "tr46": "^6.0.0", "webidl-conversions": "^8.0.1" } }, "sha512-3GeworPmc2ZfEEHP7lEbUfBX/L75wdEsi0rLNhXcXxnoN5jyq0SL5gCy06SGW2cyTIZdTvWIDQNQoza++vKeaw=="],
 
     "which": ["which@2.0.2", "", { "dependencies": { "isexe": "2.0.0" }, "bin": { "node-which": "./bin/node-which" } }, "sha512-BLI3Tl1TW3Pvl70l3yq3Y64i+awpwXqsGBYWkkqMtnbXgrMD+yj7rhW0kuEDxzJaYXGjEW5ogapKNMEKNMjibA=="],
 
@@ -2295,7 +2300,7 @@
 
     "zwitch": ["zwitch@2.0.4", "", {}, "sha512-bXE4cR/kVZhKZX/RjPEflHaKVhUVl85noU3v6b8apfQEc1x4A+zBxjZ4lN8LqGd6WZ3dl98pY4o717VFmoPp+A=="],
 
-    "@actions/http-client/undici": ["undici@6.27.0", "", {}, "sha512-YmfV3YnEDzXRC5lZ2jWtWWHKGUm1zIt8AhesR1tens+HTNv+YZlN/dp6G727LOvMJ8xjP9Be7Y2Sdr96LDm+pg=="],
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild": ["esbuild@0.28.1", "", { "optionalDependencies": { "@esbuild/aix-ppc64": "0.28.1", "@esbuild/android-arm": "0.28.1", "@esbuild/android-arm64": "0.28.1", "@esbuild/android-x64": "0.28.1", "@esbuild/darwin-arm64": "0.28.1", "@esbuild/darwin-x64": "0.28.1", "@esbuild/freebsd-arm64": "0.28.1", "@esbuild/freebsd-x64": "0.28.1", "@esbuild/linux-arm": "0.28.1", "@esbuild/linux-arm64": "0.28.1", "@esbuild/linux-ia32": "0.28.1", "@esbuild/linux-loong64": "0.28.1", "@esbuild/linux-mips64el": "0.28.1", "@esbuild/linux-ppc64": "0.28.1", "@esbuild/linux-riscv64": "0.28.1", "@esbuild/linux-s390x": "0.28.1", "@esbuild/linux-x64": "0.28.1", "@esbuild/netbsd-arm64": "0.28.1", "@esbuild/netbsd-x64": "0.28.1", "@esbuild/openbsd-arm64": "0.28.1", "@esbuild/openbsd-x64": "0.28.1", "@esbuild/openharmony-arm64": "0.28.1", "@esbuild/sunos-x64": "0.28.1", "@esbuild/win32-arm64": "0.28.1", "@esbuild/win32-ia32": "0.28.1", "@esbuild/win32-x64": "0.28.1" }, "bin": { "esbuild": "bin/esbuild" } }, "sha512-HrJrvZv5ayxBzPfwphOoNzkzOIIlifzk0KJrGK2c8R4+LKpMtpYLQeUdjnwjWv/LZlkH2laZk+4w78pi99D4Vw=="],
 
     "@cypress/request/tough-cookie": ["tough-cookie@5.1.2", "", { "dependencies": { "tldts": "6.1.86" } }, "sha512-FVDYdxtnj0G6Qm/DhNPSb8Ju59ULcup3tuJxkFb5K8Bv2pUXILbf0xZWU8PX8Ov19OXljbUyveOFwRMwkXzO+A=="],
 
@@ -2307,21 +2312,33 @@
 
     "@emnapi/wasi-threads/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
 
-    "@img/sharp-wasm32/@emnapi/runtime": ["@emnapi/runtime@1.11.2", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-kyOl3X0DuTiT1h2ft8r2fYO8JYtU9a9Xis/zBSiGArNaagCOWx90N1k2wxp18czFDH+OgcWGb5ZP/XMt3dcyPA=="],
+    "@mdx-js/mdx/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
 
     "@mdx-js/mdx/acorn": ["acorn@8.16.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw=="],
 
-    "@posthog/nextjs-config/semver": ["semver@7.8.0", "", { "bin": { "semver": "bin/semver.js" } }, "sha512-AcM7dV/5ul4EekoQ29Agm5vri8JNqRyj39o0qpX6vDF2GZrtutZl5RwgD1XnZjiTAfncsJhMI48QQH3sN87YNA=="],
+    "@napi-rs/wasm-runtime/@emnapi/runtime": ["@emnapi/runtime@1.11.1", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-vgj7R3y3Wgx24IQaGPA/R6YFXLHVMOZ0uVEyIQPaWs+rd1AzfEMXlAC22FYwO1XkKR6NPsq7mUandH8oIRdZFw=="],
 
-    "@posthog/webpack-plugin/@posthog/core": ["@posthog/core@1.40.1", "", { "dependencies": { "@posthog/types": "1.393.0" } }, "sha512-jXuMtZCwA7AMpYlo1wjm6GlC58YBlz/ZxpDIsF9hroUfqYoPPDmQbsQb4iiZAS43+o/kwloxdKJIlE0IiwQ5HQ=="],
+    "@posthog/webpack-plugin/@posthog/core": ["@posthog/core@1.48.1", "", { "dependencies": { "@posthog/types": "^1.404.1" } }, "sha512-mxw31XdYgt/SnlwqLPAcltK67q+QmsiYjVLGQ4GbBc8OJ7O4yRFSDwAXt8QsokHokeyEZtTRWM6jDHL7LYMx1A=="],
 
     "@react-three/fiber/buffer": ["buffer@6.0.3", "", { "dependencies": { "base64-js": "1.5.1", "ieee754": "1.2.1" } }, "sha512-FTiCpNxtwiZZHEZbcbTIcZjERVICn9yq/pDFkTl95/AxzD1naBctN7YO68riM/gLSDY7sdrMby8hofADYuuqOA=="],
 
+    "@rolldown/binding-wasm32-wasi/@emnapi/runtime": ["@emnapi/runtime@1.11.1", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-vgj7R3y3Wgx24IQaGPA/R6YFXLHVMOZ0uVEyIQPaWs+rd1AzfEMXlAC22FYwO1XkKR6NPsq7mUandH8oIRdZFw=="],
+
     "@swc/helpers/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
 
+    "@tailwindcss/node/lightningcss": ["lightningcss@1.32.0", "", { "dependencies": { "detect-libc": "2.1.2" }, "optionalDependencies": { "lightningcss-android-arm64": "1.32.0", "lightningcss-darwin-arm64": "1.32.0", "lightningcss-darwin-x64": "1.32.0", "lightningcss-freebsd-x64": "1.32.0", "lightningcss-linux-arm-gnueabihf": "1.32.0", "lightningcss-linux-arm64-gnu": "1.32.0", "lightningcss-linux-arm64-musl": "1.32.0", "lightningcss-linux-x64-gnu": "1.32.0", "lightningcss-linux-x64-musl": "1.32.0", "lightningcss-win32-arm64-msvc": "1.32.0", "lightningcss-win32-x64-msvc": "1.32.0" } }, "sha512-NXYBzinNrblfraPGyrbPoD19C1h9lfI/1mzgWYvXUTe414Gz/X1FD2XBZSZM7rRTrMA8JL3OtAaGifrIKhQ5yQ=="],
+
+    "@tailwindcss/postcss/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
     "@tybys/wasm-util/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
 
-    "@vercel/blob/undici": ["undici@6.27.0", "", {}, "sha512-YmfV3YnEDzXRC5lZ2jWtWWHKGUm1zIt8AhesR1tens+HTNv+YZlN/dp6G727LOvMJ8xjP9Be7Y2Sdr96LDm+pg=="],
+    "@types/adm-zip/@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+
+    "@types/stream-chain/@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+
+    "@types/stream-json/@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+
+    "@vercel/analytics/next": ["next@16.2.12", "", { "dependencies": { "@next/env": "16.2.12", "@swc/helpers": "0.5.15", "baseline-browser-mapping": "2.10.42", "caniuse-lite": "1.0.30001803", "postcss": "8.5.23", "styled-jsx": "5.1.6" }, "optionalDependencies": { "@next/swc-darwin-arm64": "16.2.12", "@next/swc-darwin-x64": "16.2.12", "@next/swc-linux-arm64-gnu": "16.2.12", "@next/swc-linux-arm64-musl": "16.2.12", "@next/swc-linux-x64-gnu": "16.2.12", "@next/swc-linux-x64-musl": "16.2.12", "@next/swc-win32-arm64-msvc": "16.2.12", "@next/swc-win32-x64-msvc": "16.2.12", "@opentelemetry/api": "1.9.1", "sharp": "0.35.3" }, "peerDependencies": { "react": "19.2.8", "react-dom": "19.2.8" }, "bin": { "next": "dist/bin/next" } }, "sha512-iD59eYQWmbFcEbX7v/acG5DRym9iw1DdaPoD0WTA920naWsE25wShzJW4+UvAs8MK9EC2kBfIH6vtto1H1PHGw=="],
 
     "@vercel/cli-config/zod": ["zod@4.1.11", "", {}, "sha512-WPsqwxITS2tzx1bzhIKsEs19ABD5vmCVa4xBo2tq/SrV4RNZtfws1EnCWQXM6yh8bD08a1idvkB5MZSBiZsjwg=="],
 
@@ -2329,6 +2346,10 @@
 
     "@vercel/oidc/jose": ["jose@5.10.0", "", {}, "sha512-s+3Al/p9g32Iq+oqXxkW//7jk2Vig6FF1CFqzVXoTUXt2qz89YWbL+OwS17NFYEvxC35n0FKeGO2LGYSxeM2Gg=="],
 
+    "@vercel/speed-insights/next": ["next@16.2.12", "", { "dependencies": { "@next/env": "16.2.12", "@swc/helpers": "0.5.15", "baseline-browser-mapping": "2.10.42", "caniuse-lite": "1.0.30001803", "postcss": "8.5.23", "styled-jsx": "5.1.6" }, "optionalDependencies": { "@next/swc-darwin-arm64": "16.2.12", "@next/swc-darwin-x64": "16.2.12", "@next/swc-linux-arm64-gnu": "16.2.12", "@next/swc-linux-arm64-musl": "16.2.12", "@next/swc-linux-x64-gnu": "16.2.12", "@next/swc-linux-x64-musl": "16.2.12", "@next/swc-win32-arm64-msvc": "16.2.12", "@next/swc-win32-x64-msvc": "16.2.12", "@opentelemetry/api": "1.9.1", "sharp": "0.35.3" }, "peerDependencies": { "react": "19.2.8", "react-dom": "19.2.8" }, "bin": { "next": "dist/bin/next" } }, "sha512-iD59eYQWmbFcEbX7v/acG5DRym9iw1DdaPoD0WTA920naWsE25wShzJW4+UvAs8MK9EC2kBfIH6vtto1H1PHGw=="],
+
+    "@vitest/mocker/vite": ["vite@8.1.5", "", { "dependencies": { "lightningcss": "1.32.0", "picomatch": "4.0.5", "postcss": "8.5.23", "rolldown": "1.1.5", "tinyglobby": "0.2.17" }, "optionalDependencies": { "@types/node": "26.1.1", "esbuild": "0.28.1", "fsevents": "2.3.3", "jiti": "2.7.0", "terser": "5.49.0", "tsx": "4.23.1", "yaml": "2.9.0" }, "bin": { "vite": "bin/vite.js" } }, "sha512-7ULLwsCdYx/nRyrpiEwvqb5TFHrMVZyBt+rg/OAXT7rgj/z+DtTDyKFeLAdDkubDVDKD8jOsndmy7m55XcfUsw=="],
+
     "@vue/compiler-core/@babel/parser": ["@babel/parser@7.29.7", "", { "dependencies": { "@babel/types": "7.29.7" }, "bin": "./bin/babel-parser.js" }, "sha512-hnORnjP/1P/zFEndoeX+n+t1RwWRJiJpM/jO7FW32Kn9r5+sJB2JWOdYo4L6k78j15eCwY3Gm/7364B1EMwtNg=="],
 
     "@vue/compiler-core/entities": ["entities@7.0.1", "", {}, "sha512-TWrgLOFUQTH994YUyl1yT4uyavY5nNB5muff+RtWaqNVCAK408b5ZnnbNAUEWLTCpum9w6arT70i1XdQ4UeOPA=="],
@@ -2339,6 +2360,8 @@
 
     "@vue/compiler-sfc/estree-walker": ["estree-walker@2.0.2", "", {}, "sha512-Rfkk/Mp/DL7JVje3u18FxFujQlTNR2q6QfMSMB7AvCBx91NGj/ba3kCfza0f6dVDbw7YlRf/nDrn7pQrCCyQ/w=="],
 
+    "@vue/compiler-sfc/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
     "accepts/mime-types": ["mime-types@3.0.2", "", { "dependencies": { "mime-db": "1.54.0" } }, "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A=="],
 
     "acorn-jsx/acorn": ["acorn@8.16.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw=="],
@@ -2357,12 +2380,16 @@
 
     "d3-dsv/commander": ["commander@7.2.0", "", {}, "sha512-QrWXB+ZQSVPmIWIhtEO9H+gwHaMGYiF5ChvoJ+K9ZGHG/sVsa6yiesAD1GC/x46sET00Xlwo1u49RVVVzvcSkw=="],
 
+    "data-urls/whatwg-url": ["whatwg-url@16.0.1", "", { "dependencies": { "@exodus/bytes": "1.15.1", "tr46": "6.0.0", "webidl-conversions": "8.0.1" } }, "sha512-1to4zXBxmXHV3IiSSEInrreIlu02vUOvrhxJJH5vcxYTBDAx51cqZiKdyTxlecdKNSjj8EcxGBxNf6Vg+945gw=="],
+
     "deep-equal/side-channel": ["side-channel@1.1.0", "", { "dependencies": { "es-errors": "1.3.0", "object-inspect": "1.13.4", "side-channel-list": "1.0.1", "side-channel-map": "1.0.1", "side-channel-weakmap": "1.0.2" } }, "sha512-ZX99e6tRweoUXqR+VBrslhda51Nh5MTQwou5tnUDgbtyM0dBgmhEDtWGP/xbKn6hqfPRHujUNwz5fy/wbbhnpw=="],
 
     "dependency-tree/commander": ["commander@12.1.0", "", {}, "sha512-Vw8qHK3bZM9y/P10u3Vib8o/DdkvA2OtPtZvD871QKjy74Wj1WSKFILMPRPSdUSx5RFK1arlJzEtA4PkFgnbuA=="],
 
     "dependency-tree/typescript": ["typescript@5.9.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw=="],
 
+    "detective-postcss/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
     "esast-util-from-js/acorn": ["acorn@8.16.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw=="],
 
     "escodegen/estraverse": ["estraverse@5.3.0", "", {}, "sha512-MMdARuVEQziNTeJD8DgMqmhwR11BRQ/cBP+pLtYdSTnf3MIO8fFeiINEbX36ZdNlfU/7A9f3gUw49B3oQsvwBA=="],
@@ -2379,16 +2406,44 @@
 
     "find-test-names/@babel/parser": ["@babel/parser@7.29.7", "", { "dependencies": { "@babel/types": "7.29.7" }, "bin": "./bin/babel-parser.js" }, "sha512-hnORnjP/1P/zFEndoeX+n+t1RwWRJiJpM/jO7FW32Kn9r5+sJB2JWOdYo4L6k78j15eCwY3Gm/7364B1EMwtNg=="],
 
+    "hast-util-from-dom/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-from-html/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
     "hast-util-from-html/parse5": ["parse5@7.3.0", "", { "dependencies": { "entities": "^6.0.0" } }, "sha512-IInvU7fabl34qmi9gY8XOVxhYyMyuH2xUNpb2q8/Y+7552KlejkRvqvD19nMoUW/uQGGbqNpA6Tufu5FL5BZgw=="],
 
+    "hast-util-from-html-isomorphic/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-from-parse5/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
     "hast-util-from-parse5/property-information": ["property-information@7.2.0", "", {}, "sha512-IAtzIB6sUiWaJYrX9smp3V46pBGbBeLFRGdh25kg1334VcBlD8HzhPeNIWQH9zhGmo2itIe25EHt9dQP7G5hmg=="],
 
+    "hast-util-is-element/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-parse-selector/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-to-estree/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-to-html/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
     "hast-util-to-html/property-information": ["property-information@7.2.0", "", {}, "sha512-IAtzIB6sUiWaJYrX9smp3V46pBGbBeLFRGdh25kg1334VcBlD8HzhPeNIWQH9zhGmo2itIe25EHt9dQP7G5hmg=="],
 
+    "hast-util-to-jsx-runtime/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-to-string/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-to-text/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hast-util-whitespace/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "hastscript/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
     "hastscript/property-information": ["property-information@7.2.0", "", {}, "sha512-IAtzIB6sUiWaJYrX9smp3V46pBGbBeLFRGdh25kg1334VcBlD8HzhPeNIWQH9zhGmo2itIe25EHt9dQP7G5hmg=="],
 
     "istanbul-lib-report/supports-color": ["supports-color@7.2.0", "", { "dependencies": { "has-flag": "4.0.0" } }, "sha512-qpCAvRl9stuOHveKsn7HncJRvv501qIacKzQlO/+Lwxc9+0q2wLyv4Dfvt80/DPn2pqOBsJdDiogXGR9+OvwRw=="],
 
+    "jest-worker/@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+
     "katex/commander": ["commander@8.3.0", "", {}, "sha512-OkTL9umf+He2DZkUq8f8J9of7yL6RJKI24dVITBmNfZBmri9zYZQrKkuXiKhyfPSu8tUhnVBB1iKXevvnlR4Ww=="],
 
     "log-update/slice-ansi": ["slice-ansi@7.1.2", "", { "dependencies": { "ansi-styles": "6.2.3", "is-fullwidth-code-point": "5.1.0" } }, "sha512-iOBWFgUX7caIZiuutICxVgX1SdxwAVFFKwt1EvMYYec/NWO5meOJ6K5uQxhrYBdQJne4KxiqZc+KptFOWFSI9w=="],
@@ -2399,22 +2454,38 @@
 
     "magicast/@babel/types": ["@babel/types@7.29.7", "", { "dependencies": { "@babel/helper-string-parser": "7.29.7", "@babel/helper-validator-identifier": "7.29.7" } }, "sha512-4zBIxpPzowiZpusoFkyGVwakdRJUyuH5PxQ/PrqghfdFWWasvnCdPfQXHrenDai+gyLARulZjZowCOj6fjT4pA=="],
 
+    "mdast-util-math/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "mdast-util-mdx-expression/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "mdast-util-mdx-jsx/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "mdast-util-mdxjs-esm/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "mdast-util-to-hast/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
     "micromark-extension-mdxjs/acorn": ["acorn@8.16.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw=="],
 
     "micromatch/picomatch": ["picomatch@2.3.2", "", {}, "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA=="],
 
     "module-lookup-amd/commander": ["commander@12.1.0", "", {}, "sha512-Vw8qHK3bZM9y/P10u3Vib8o/DdkvA2OtPtZvD871QKjy74Wj1WSKFILMPRPSdUSx5RFK1arlJzEtA4PkFgnbuA=="],
 
+    "next/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
     "next-mdx-remote/@babel/code-frame": ["@babel/code-frame@7.29.0", "", { "dependencies": { "@babel/helper-validator-identifier": "7.28.5", "js-tokens": "4.0.0", "picocolors": "1.1.1" } }, "sha512-9NhCeYjq9+3uxgdtp20LSiJXJvN0FeCtNGpJxuMFZ1Kv3cWUNb6DOhJwUvcVCzKGR66cw4njwM6hrJLqgOwbcw=="],
 
     "node-source-walk/@babel/parser": ["@babel/parser@7.29.7", "", { "dependencies": { "@babel/types": "7.29.7" }, "bin": "./bin/babel-parser.js" }, "sha512-hnORnjP/1P/zFEndoeX+n+t1RwWRJiJpM/jO7FW32Kn9r5+sJB2JWOdYo4L6k78j15eCwY3Gm/7364B1EMwtNg=="],
 
     "parse-entities/@types/unist": ["@types/unist@2.0.11", "", {}, "sha512-CmBKiL6NNo/OqgmMn95Fk9Whlp2mtvIv+KNpQKN2F4SjvrEesubTRWGYSg+BnWZOnlCaSTU1sMpsBOzgbYhnsA=="],
 
+    "postcss-values-parser/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
     "posthog-js/fflate": ["fflate@0.4.8", "", {}, "sha512-FJqqoDBR00Mdj9ppamLa/Y7vxm+PRmNWA67N846RvsoYVMKB4q3y/de5PA7gUmRMYK/8CMz2GDZQmCRN1wBcWA=="],
 
     "precinct/commander": ["commander@12.1.0", "", {}, "sha512-Vw8qHK3bZM9y/P10u3Vib8o/DdkvA2OtPtZvD871QKjy74Wj1WSKFILMPRPSdUSx5RFK1arlJzEtA4PkFgnbuA=="],
 
+    "precinct/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
     "precinct/typescript": ["typescript@5.9.3", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-jl1vZzPDinLr9eUt3J/t7V6FgNEw9QjvBPdysz9KfQDD41fQrC2Y4vKQdiaUpFT4bXlb1RHhLpp8wtm6M5TgSw=="],
 
     "raw-body/iconv-lite": ["iconv-lite@0.7.2", "", { "dependencies": { "safer-buffer": "2.1.2" } }, "sha512-im9DjEDQ55s9fL4EYzOAv0yMqmMBSZp6G0VvFyTMPKWxiSBHUj9NW/qqLmXUwXrrM7AvqSlTCfvqRb0cM8yYqw=="],
@@ -2427,6 +2498,12 @@
 
     "recma-jsx/acorn": ["acorn@8.16.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-UVJyE9MttOsBQIDKw1skb9nAwQuR5wuGD3+82K6JgJlm/Y+KI92oNsMNGZCYdDsVtRHSak0pcV5Dno5+4jh9sw=="],
 
+    "rehype-katex/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "rehype-recma/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
+    "remark-rehype/@types/hast": ["@types/hast@3.0.4", "", { "dependencies": { "@types/unist": "3.0.3" } }, "sha512-WPs+bbQw5aCj+x6laNGWLH3wviHtoCv/P3+otBhbOhJgG8qtpdAMlTCxLtsTWA7LH1Oh/bFCHsBn0TPS5m30EQ=="],
+
     "request-progress/throttleit": ["throttleit@1.0.1", "", {}, "sha512-vDZpf9Chs9mAdfY046mcPt8fg5QSZr37hEH4TXYBnDF+izxgrbRGUAAaBvIk/fJm9aOFCGFd1EsNg5AZCbnQCQ=="],
 
     "require-and-forget/debug": ["debug@4.3.4", "", { "dependencies": { "ms": "2.1.2" }, "optionalDependencies": { "supports-color": "8.1.1" } }, "sha512-PRWFHuSU3eDtQJPvnNY7Jcket1j0t5OuOsFzPPzsekD52Zl8qUfFIPEiswXqIvHWGVHOgX+7G/vCNNhehwxfkQ=="],
@@ -2441,6 +2518,8 @@
 
     "send/mime-types": ["mime-types@3.0.2", "", { "dependencies": { "mime-db": "1.54.0" } }, "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A=="],
 
+    "sharp/@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+
     "shelljs/execa": ["execa@5.1.1", "", { "dependencies": { "cross-spawn": "7.0.6", "get-stream": "6.0.1", "human-signals": "2.1.0", "is-stream": "2.0.1", "merge-stream": "2.0.0", "npm-run-path": "4.0.1", "onetime": "5.1.2", "signal-exit": "3.0.7", "strip-final-newline": "2.0.0" } }, "sha512-8uSpZZocAZRBAPIEINJj3Lo9HyGitllczc27Eh5YYojjMFMn8yHMDMaUHE2Jqfq05D/wucwI4JGURyXt1vchyg=="],
 
     "slice-ansi/ansi-styles": ["ansi-styles@6.2.3", "", {}, "sha512-4Dj6M28JB+oAH8kFkTLUo+a2jwOFkuqb3yucU0CANcRRUbxS0cP0nZYCGjcc3BNXwRIsUVmDGgzawme7zvJHvg=="],
@@ -2453,6 +2532,14 @@
 
     "terser/commander": ["commander@2.20.3", "", {}, "sha512-GpVkmM8vF2vQUkj2LvZmD35JxeJOLCwJ9cUkugyk2nuhbv3+mJvpLYYt+0+USMxE+oj+ey/lJEnhZw75x/OMcQ=="],
 
+    "terser-webpack-plugin/esbuild": ["esbuild@0.28.1", "", { "optionalDependencies": { "@esbuild/aix-ppc64": "0.28.1", "@esbuild/android-arm": "0.28.1", "@esbuild/android-arm64": "0.28.1", "@esbuild/android-x64": "0.28.1", "@esbuild/darwin-arm64": "0.28.1", "@esbuild/darwin-x64": "0.28.1", "@esbuild/freebsd-arm64": "0.28.1", "@esbuild/freebsd-x64": "0.28.1", "@esbuild/linux-arm": "0.28.1", "@esbuild/linux-arm64": "0.28.1", "@esbuild/linux-ia32": "0.28.1", "@esbuild/linux-loong64": "0.28.1", "@esbuild/linux-mips64el": "0.28.1", "@esbuild/linux-ppc64": "0.28.1", "@esbuild/linux-riscv64": "0.28.1", "@esbuild/linux-s390x": "0.28.1", "@esbuild/linux-x64": "0.28.1", "@esbuild/netbsd-arm64": "0.28.1", "@esbuild/netbsd-x64": "0.28.1", "@esbuild/openbsd-arm64": "0.28.1", "@esbuild/openbsd-x64": "0.28.1", "@esbuild/openharmony-arm64": "0.28.1", "@esbuild/sunos-x64": "0.28.1", "@esbuild/win32-arm64": "0.28.1", "@esbuild/win32-ia32": "0.28.1", "@esbuild/win32-x64": "0.28.1" }, "bin": { "esbuild": "bin/esbuild" } }, "sha512-HrJrvZv5ayxBzPfwphOoNzkzOIIlifzk0KJrGK2c8R4+LKpMtpYLQeUdjnwjWv/LZlkH2laZk+4w78pi99D4Vw=="],
+
+    "terser-webpack-plugin/lightningcss": ["lightningcss@1.32.0", "", { "dependencies": { "detect-libc": "2.1.2" }, "optionalDependencies": { "lightningcss-android-arm64": "1.32.0", "lightningcss-darwin-arm64": "1.32.0", "lightningcss-darwin-x64": "1.32.0", "lightningcss-freebsd-x64": "1.32.0", "lightningcss-linux-arm-gnueabihf": "1.32.0", "lightningcss-linux-arm64-gnu": "1.32.0", "lightningcss-linux-arm64-musl": "1.32.0", "lightningcss-linux-x64-gnu": "1.32.0", "lightningcss-linux-x64-musl": "1.32.0", "lightningcss-win32-arm64-msvc": "1.32.0", "lightningcss-win32-x64-msvc": "1.32.0" } }, "sha512-NXYBzinNrblfraPGyrbPoD19C1h9lfI/1mzgWYvXUTe414Gz/X1FD2XBZSZM7rRTrMA8JL3OtAaGifrIKhQ5yQ=="],
+
+    "terser-webpack-plugin/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
+    "tsx/esbuild": ["esbuild@0.28.1", "", { "optionalDependencies": { "@esbuild/aix-ppc64": "0.28.1", "@esbuild/android-arm": "0.28.1", "@esbuild/android-arm64": "0.28.1", "@esbuild/android-x64": "0.28.1", "@esbuild/darwin-arm64": "0.28.1", "@esbuild/darwin-x64": "0.28.1", "@esbuild/freebsd-arm64": "0.28.1", "@esbuild/freebsd-x64": "0.28.1", "@esbuild/linux-arm": "0.28.1", "@esbuild/linux-arm64": "0.28.1", "@esbuild/linux-ia32": "0.28.1", "@esbuild/linux-loong64": "0.28.1", "@esbuild/linux-mips64el": "0.28.1", "@esbuild/linux-ppc64": "0.28.1", "@esbuild/linux-riscv64": "0.28.1", "@esbuild/linux-s390x": "0.28.1", "@esbuild/linux-x64": "0.28.1", "@esbuild/netbsd-arm64": "0.28.1", "@esbuild/netbsd-x64": "0.28.1", "@esbuild/openbsd-arm64": "0.28.1", "@esbuild/openbsd-x64": "0.28.1", "@esbuild/openharmony-arm64": "0.28.1", "@esbuild/sunos-x64": "0.28.1", "@esbuild/win32-arm64": "0.28.1", "@esbuild/win32-ia32": "0.28.1", "@esbuild/win32-x64": "0.28.1" }, "bin": { "esbuild": "bin/esbuild" } }, "sha512-HrJrvZv5ayxBzPfwphOoNzkzOIIlifzk0KJrGK2c8R4+LKpMtpYLQeUdjnwjWv/LZlkH2laZk+4w78pi99D4Vw=="],
+
     "type-is/content-type": ["content-type@2.0.0", "", {}, "sha512-j/O/d7GcZCyNl7/hwZAb606rzqkyvaDctLmckbxLzHvFBzTJHuGEdodATcP3yIRoDrLHkIATJuvzbFlp/ki2cQ=="],
 
     "type-is/mime-types": ["mime-types@3.0.2", "", { "dependencies": { "mime-db": "1.54.0" } }, "sha512-Lbgzdk0h4juoQ9fCKXW4by0UJqj+nOOrI9MJ1sSj4nI8aI2eo1qmvQEie4VD1glsS250n15LsWsYtCugiStS5A=="],
@@ -2461,6 +2548,8 @@
 
     "use-sidecar/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
 
+    "vitest/@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+
     "webpack/mime-db": ["mime-db@1.54.0", "", {}, "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ=="],
 
     "wrap-ansi/ansi-styles": ["ansi-styles@6.2.3", "", {}, "sha512-4Dj6M28JB+oAH8kFkTLUo+a2jwOFkuqb3yucU0CANcRRUbxS0cP0nZYCGjcc3BNXwRIsUVmDGgzawme7zvJHvg=="],
@@ -2469,16 +2558,146 @@
 
     "wrap-ansi/strip-ansi": ["strip-ansi@7.2.0", "", { "dependencies": { "ansi-regex": "6.2.2" } }, "sha512-yDPMNjp4WyfYBkHnjIRLfca1i6KMyGCtsVgoKe/z1+6vukgaENdgGBZt+ZmKPc4gavvEZ5OgHfHdrazhgNyG7w=="],
 
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/aix-ppc64": ["@esbuild/aix-ppc64@0.28.1", "", { "os": "aix", "cpu": "ppc64" }, "sha512-Svl7tq8k/08+p6CXPpRjQ1fKX+1odH/BQbb48fV6fj3CWHhsoIOoY87w1oHXm0qEpkIK3ZfVgp0hed3XBXzXMQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/android-arm": ["@esbuild/android-arm@0.28.1", "", { "os": "android", "cpu": "arm" }, "sha512-0k2F129Xdio1TdJfzJ8sy1Q47vUD2NnwdhiAf7drUN1EBTfPf4hsFCtmMgu/6m8JSzsBrlmVjudMBQqOfG8usQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/android-arm64": ["@esbuild/android-arm64@0.28.1", "", { "os": "android", "cpu": "arm64" }, "sha512-34EGEbCIAgosYz6goLcopX6Mo7NyGv9tfwEM2/7Ce2VcVRk568iSvniGWcUXIy7wEDR1wzolcxcriFVrWYcwBg=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/android-x64": ["@esbuild/android-x64@0.28.1", "", { "os": "android", "cpu": "x64" }, "sha512-dbwY7ltSMDWsRatcRpCnES4F+im88OCUgGZjy52shC7GqHRE/cYlxNbB4Z4UpJswpcc4Qxd2oE/ufM0p61IKng=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/darwin-arm64": ["@esbuild/darwin-arm64@0.28.1", "", { "os": "darwin", "cpu": "arm64" }, "sha512-TZbWkQY7kvTAXbXUT7uVACR5cMHsDiSz9z7ZKAX/RTq/WJEk3QyRr0wZpNhBDX+/0CtdqUIJlOiodQcta6tY3Q=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/darwin-x64": ["@esbuild/darwin-x64@0.28.1", "", { "os": "darwin", "cpu": "x64" }, "sha512-zfdzgK9ACBNZLI/CyHTOx81SyNbM6YXn7rxSgX97VjyiPl9W1i4Ka4fgKECEoFCKGpvBj5qArWIGgQjOwkgskQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/freebsd-arm64": ["@esbuild/freebsd-arm64@0.28.1", "", { "os": "freebsd", "cpu": "arm64" }, "sha512-wG2EA8ENdEI0qhkSZMjfqrdY+ziCYCPMmtZjjIwOmXFjmyzEHn+UUxk5of+SYsjtfs3VpnlC7QLzSI5hY/rOAw=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/freebsd-x64": ["@esbuild/freebsd-x64@0.28.1", "", { "os": "freebsd", "cpu": "x64" }, "sha512-i7dZ9vQgnvSCzi/rYCXNgtF/U+eKZNJBzu3eTQbRgHnM7tNSizLOkRFAl3qzVc/Op/u5YkHHa4pf/3DOYHthLQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-arm": ["@esbuild/linux-arm@0.28.1", "", { "os": "linux", "cpu": "arm" }, "sha512-qVXBOHQS+d5Y722GwJzJUtOLlX7km3CraOaGormF1pDtPd2C/l1SHRPgjLunLGe51Sh5YYWKMFDyV4SxgMQYTQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-arm64": ["@esbuild/linux-arm64@0.28.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-yHs+0uc8+nvEAfAfxrWQKK5peSNzBc4PegcMO0EJ2hT71uA7vB8Ihg2e77R2P7SG5uYjPbHlLLmve4LLLRCf0g=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-ia32": ["@esbuild/linux-ia32@0.28.1", "", { "os": "linux", "cpu": "ia32" }, "sha512-d1z4ZuP0ajrfz/FhGT4vv278rX8KnPPJx8i5+AtK7TYbx9Le9F1hyzurZpkEyjkGa9dUGhQow4C1NmeGvqxN2w=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-loong64": ["@esbuild/linux-loong64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-M5sRjUVZrkm1OAPR3dlOYzNmN+loZKGVi1VUQGrwuqLcbR6qeAz+famMhjASeH3YVKvZz+zT1jlh/keC3Rj/lg=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-mips64el": ["@esbuild/linux-mips64el@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-mRObBZeHh2OxcBFPWE/FjylkRgZdYuiTR3vaTozquCGOH14iP9oN4x4Ge81CoIDYQrXmIxpFumJBu5MtZpnQJQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-ppc64": ["@esbuild/linux-ppc64@0.28.1", "", { "os": "linux", "cpu": "ppc64" }, "sha512-slScBsMAb3GFDcdrCgLwZtPYRoH2H/youv10QiZyRjmsP48fznoveWytSgCI/R0ZcUgpc0ZhIUEx6LHts8yrfQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-riscv64": ["@esbuild/linux-riscv64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-kw0owk1o0GFETUJyW0jc0G4Yzs0BHZn0JDZ8JRT088vjJYX777BAs1fDGxAC+q831qOs2DTC96mNsG2opdfyyQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-s390x": ["@esbuild/linux-s390x@0.28.1", "", { "os": "linux", "cpu": "s390x" }, "sha512-/lAIjX8aYFRByhh6L5rYtPEDRqa9de/4V/juOXcta5frjvzXO4/sqEtyytse0g3zZFuWu5cDN0MkLz2qRDD2Ag=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/linux-x64": ["@esbuild/linux-x64@0.28.1", "", { "os": "linux", "cpu": "x64" }, "sha512-u/anNYF2mmVOEDwLtnQ1wOr3EZ9sTNGLWrsYGYwHWzGA3Si84IOkHXlbWTD1NB+9/1lcnweYKO54uhxZydNzfA=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/netbsd-arm64": ["@esbuild/netbsd-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-oks0DYbLwWMmaakTsCb+zL4E+aHRVLom9IJZOAthMQEPiQmydXHkziYEsGYRx0uNV/IjEKGAV941JzH02pflqw=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/netbsd-x64": ["@esbuild/netbsd-x64@0.28.1", "", { "os": "none", "cpu": "x64" }, "sha512-aeL6lAnN89Hz43Mlh1G8ARasbuoYvSITDEx0tHh5b7jJnHcssqgjy9Yx430GDpmCa6OyrKoS0aNRjKundRizGg=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/openbsd-arm64": ["@esbuild/openbsd-arm64@0.28.1", "", { "os": "openbsd", "cpu": "arm64" }, "sha512-MEFJe5C3R8pwXdZ5Y21oo6m7ePiS0d9pWucn99O/wvyJZChoIQKrQDxKrGeW8F5+T0okTHesAmDeiHDTIq0V/Q=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/openbsd-x64": ["@esbuild/openbsd-x64@0.28.1", "", { "os": "openbsd", "cpu": "x64" }, "sha512-i/ZLIOafE0Z8cI/XANJAixoJL/uRAoS2xOA3rb0xN+KK0K177cMAsQYkzHtBrtMXAKuAc7HGgcWiZ/sRC1Nxgw=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/openharmony-arm64": ["@esbuild/openharmony-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-ge+Z7EXFNt2BO1oAMsVpiQ8EwndV9i1xXerAeTIK7AtPs3bKFXQM7nlRxDSIUIMeueR1CNXxqztLzdNeReKBJg=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/sunos-x64": ["@esbuild/sunos-x64@0.28.1", "", { "os": "sunos", "cpu": "x64" }, "sha512-BEjgtECkL3vY+SaSQ6nzVfiALUeFxpawyp8Jmf5PtYhf1Ug40N1h/hxlhts+f1FvSvarEigdxS3BlSMI2PJLcQ=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/win32-arm64": ["@esbuild/win32-arm64@0.28.1", "", { "os": "win32", "cpu": "arm64" }, "sha512-lCv9eK/H6ZJWbE7bh2nw54CZ9M2nupBxJcTsdk/QQnWkdSjKGuxmmH8/GWrlT1eMmZfn4dGcCjRte397WqfQXA=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/win32-ia32": ["@esbuild/win32-ia32@0.28.1", "", { "os": "win32", "cpu": "ia32" }, "sha512-zvb/mB2bSCoJOpoCBgYKKpX6YM6mJBlBUVUtVj41DlZJVEB6/0CKlRYxP5wWl1C1ILiCoAU5wZZ4q1P3qeS6Eg=="],
+
+    "@bahmutov/cypress-esbuild-preprocessor/esbuild/@esbuild/win32-x64": ["@esbuild/win32-x64@0.28.1", "", { "os": "win32", "cpu": "x64" }, "sha512-bm4Mowrv+GXMlpWX++EcXw/iLyd1o3+bJkC2DkWXYVvgZCqD/bSj9ctZeAMC3cIxgjRVR2Dufaiu4YPxr5gW1A=="],
+
     "@cypress/request/tough-cookie/tldts": ["tldts@6.1.86", "", { "dependencies": { "tldts-core": "6.1.86" }, "bin": { "tldts": "bin/cli.js" } }, "sha512-WMi/OQ2axVTf/ykqCQgXiIct+mSQDFdH2fkwhPwgEwvJ1kSzZRiinb0zF2Xb8u4+OqPChmyI6MEu4EezNJz+FQ=="],
 
-    "@img/sharp-wasm32/@emnapi/runtime/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
+    "@napi-rs/wasm-runtime/@emnapi/runtime/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
+
+    "@posthog/webpack-plugin/@posthog/core/@posthog/types": ["@posthog/types@1.404.1", "", {}, "sha512-i2Gei6ARfOSBeTN4s2yUP1p97s2UNI+1NWmtLjhnR/V6t3RFOfI1sBWKcJNWHjtoOCCWoAFU+PNPY6SgT2VtEQ=="],
+
+    "@rolldown/binding-wasm32-wasi/@emnapi/runtime/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-android-arm64": ["lightningcss-android-arm64@1.32.0", "", { "os": "android", "cpu": "arm64" }, "sha512-YK7/ClTt4kAK0vo6w3X+Pnm0D2cf2vPHbhOXdoNti1Ga0al1P4TBZhwjATvjNwLEBCnKvjJc2jQgHXH0NEwlAg=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-darwin-arm64": ["lightningcss-darwin-arm64@1.32.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-RzeG9Ju5bag2Bv1/lwlVJvBE3q6TtXskdZLLCyfg5pt+HLz9BqlICO7LZM7VHNTTn/5PRhHFBSjk5lc4cmscPQ=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-darwin-x64": ["lightningcss-darwin-x64@1.32.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-U+QsBp2m/s2wqpUYT/6wnlagdZbtZdndSmut/NJqlCcMLTWp5muCrID+K5UJ6jqD2BFshejCYXniPDbNh73V8w=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-freebsd-x64": ["lightningcss-freebsd-x64@1.32.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-JCTigedEksZk3tHTTthnMdVfGf61Fky8Ji2E4YjUTEQX14xiy/lTzXnu1vwiZe3bYe0q+SpsSH/CTeDXK6WHig=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-linux-arm-gnueabihf": ["lightningcss-linux-arm-gnueabihf@1.32.0", "", { "os": "linux", "cpu": "arm" }, "sha512-x6rnnpRa2GL0zQOkt6rts3YDPzduLpWvwAF6EMhXFVZXD4tPrBkEFqzGowzCsIWsPjqSK+tyNEODUBXeeVHSkw=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-linux-arm64-gnu": ["lightningcss-linux-arm64-gnu@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-0nnMyoyOLRJXfbMOilaSRcLH3Jw5z9HDNGfT/gwCPgaDjnx0i8w7vBzFLFR1f6CMLKF8gVbebmkUN3fa/kQJpQ=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-linux-arm64-musl": ["lightningcss-linux-arm64-musl@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-UpQkoenr4UJEzgVIYpI80lDFvRmPVg6oqboNHfoH4CQIfNA+HOrZ7Mo7KZP02dC6LjghPQJeBsvXhJod/wnIBg=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-linux-x64-gnu": ["lightningcss-linux-x64-gnu@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-V7Qr52IhZmdKPVr+Vtw8o+WLsQJYCTd8loIfpDaMRWGUZfBOYEJeyJIkqGIDMZPwPx24pUMfwSxxI8phr/MbOA=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-linux-x64-musl": ["lightningcss-linux-x64-musl@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-bYcLp+Vb0awsiXg/80uCRezCYHNg1/l3mt0gzHnWV9XP1W5sKa5/TCdGWaR/zBM2PeF/HbsQv/j2URNOiVuxWg=="],
 
-    "@posthog/webpack-plugin/@posthog/core/@posthog/types": ["@posthog/types@1.393.0", "", {}, "sha512-vzWeEJZ7ERQhFRoQYaP5jzN1JvIu46UJyHXsuv+dTGW2r3sMgREOhNxXLZjmFHwZ8/FOHQoyqqQmXTCXZSfMSg=="],
+    "@tailwindcss/node/lightningcss/lightningcss-win32-arm64-msvc": ["lightningcss-win32-arm64-msvc@1.32.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-8SbC8BR40pS6baCM8sbtYDSwEVQd4JlFTOlaD3gWGHfThTcABnNDBda6eTZeqbofalIJhFx0qKzgHJmcPTnGdw=="],
+
+    "@tailwindcss/node/lightningcss/lightningcss-win32-x64-msvc": ["lightningcss-win32-x64-msvc@1.32.0", "", { "os": "win32", "cpu": "x64" }, "sha512-Amq9B/SoZYdDi1kFrojnoqPLxYhQ4Wo5XiL8EVJrVsB8ARoC1PWW6VGtT0WKCemjy8aC+louJnjS7U18x3b06Q=="],
+
+    "@vercel/analytics/next/@next/env": ["@next/env@16.2.12", "", {}, "sha512-d0Z5Bc13Fa4nR8pFAKx2jay2yhJM16vlfHbTzYnUQAxlNb6B6lmn4hjt69lYNt4kRtyYP6gEM49lPRHNbIyneg=="],
+
+    "@vercel/analytics/next/@next/swc-darwin-arm64": ["@next/swc-darwin-arm64@16.2.12", "", { "os": "darwin", "cpu": "arm64" }, "sha512-0W1R0teHWJrqKX0FH20IzzIWAOuGtBxPGuObrxy1lE8hQvCFj49KE8a3WUg0D7sq6rn6zkM4c7YGUnhudBS6oA=="],
+
+    "@vercel/analytics/next/@next/swc-darwin-x64": ["@next/swc-darwin-x64@16.2.12", "", { "os": "darwin", "cpu": "x64" }, "sha512-Hy5Ls099+aFUmOLmIgPfLqNi6iCwhL3uQCssz5rWk+5Nkc6TUKCE83DY5BbNylfm3+mfwcSFnLRfrZDJhVxdtw=="],
+
+    "@vercel/analytics/next/@next/swc-linux-arm64-gnu": ["@next/swc-linux-arm64-gnu@16.2.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-+YqU2h1cQkHsGfvjAsrSmst8UIFBibBGm5x3Xgel8NLMiDQtNOM4sM2GOEMvG5YiOBNeN/Ykk8cQC2S0Xrqljg=="],
+
+    "@vercel/analytics/next/@next/swc-linux-arm64-musl": ["@next/swc-linux-arm64-musl@16.2.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-0qjhiYBaKAqF63LA1ZWAAnKTzFUguAaZiRa5etMLGGPj/B6uEVjtIZldIzFEp3wHlB0koK6aTzqPtSdplTCjoA=="],
+
+    "@vercel/analytics/next/@next/swc-linux-x64-gnu": ["@next/swc-linux-x64-gnu@16.2.12", "", { "os": "linux", "cpu": "x64" }, "sha512-7A3q26W+h7gnA15uqBToNuDqBEFZZcqh0mW2mn4AJh/G5pdg2RVE3n4slzLEliASZFG3NmsbEzng/x2Sh09mBg=="],
+
+    "@vercel/analytics/next/@next/swc-linux-x64-musl": ["@next/swc-linux-x64-musl@16.2.12", "", { "os": "linux", "cpu": "x64" }, "sha512-qSjL/uppm+cbh21s72Ss8gkiOhQ4dExWHNGOWy6eZV7STj5WsKehgxT61beSsOj+YYQuTplL376lOCdMQU5T8w=="],
+
+    "@vercel/analytics/next/@next/swc-win32-arm64-msvc": ["@next/swc-win32-arm64-msvc@16.2.12", "", { "os": "win32", "cpu": "arm64" }, "sha512-X6hzsOUJac/e7AWSbn9gQ9nzHld1xWP5iyjHpYWvud8pufB679O1xg4JDyKr8Xd69Jvd+kM2Der6uftiZCmjYA=="],
+
+    "@vercel/analytics/next/@next/swc-win32-x64-msvc": ["@next/swc-win32-x64-msvc@16.2.12", "", { "os": "win32", "cpu": "x64" }, "sha512-F6fakeHuFTLOPt0bslQJdf+xtT+WIP9DVn/m4y1w1mRnVPyh3D/cNvzlRkxM444xfm+IvvYNSOrKiA2CDJ0Uxw=="],
+
+    "@vercel/analytics/next/@swc/helpers": ["@swc/helpers@0.5.15", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-JQ5TuMi45Owi4/BIMAJBoSQoOJu12oOk/gADqlcUL9JEdHB8vyjUSsxqeNXnmXHjYKMi2WcYtezGEEhqUI/E2g=="],
+
+    "@vercel/analytics/next/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
 
     "@vercel/cli-exec/execa/get-stream": ["get-stream@6.0.1", "", {}, "sha512-ts6Wi+2j3jQjqi70w5AlN8DFnkSwC+MqmxEzdEALB2qXZYV3X/b1CTfgPLGJNMeAWxdPfU8FO1ms3NUfaHCPYg=="],
 
     "@vercel/cli-exec/execa/human-signals": ["human-signals@2.1.0", "", {}, "sha512-B4FFZ6q/T2jhhksgkbEW3HBvWIfDW85snkQgawt07S7J5QXTk6BkNV+0yAeZrM5QpMAdYlocGoljn0sJ/WQkFw=="],
 
+    "@vercel/speed-insights/next/@next/env": ["@next/env@16.2.12", "", {}, "sha512-d0Z5Bc13Fa4nR8pFAKx2jay2yhJM16vlfHbTzYnUQAxlNb6B6lmn4hjt69lYNt4kRtyYP6gEM49lPRHNbIyneg=="],
+
+    "@vercel/speed-insights/next/@next/swc-darwin-arm64": ["@next/swc-darwin-arm64@16.2.12", "", { "os": "darwin", "cpu": "arm64" }, "sha512-0W1R0teHWJrqKX0FH20IzzIWAOuGtBxPGuObrxy1lE8hQvCFj49KE8a3WUg0D7sq6rn6zkM4c7YGUnhudBS6oA=="],
+
+    "@vercel/speed-insights/next/@next/swc-darwin-x64": ["@next/swc-darwin-x64@16.2.12", "", { "os": "darwin", "cpu": "x64" }, "sha512-Hy5Ls099+aFUmOLmIgPfLqNi6iCwhL3uQCssz5rWk+5Nkc6TUKCE83DY5BbNylfm3+mfwcSFnLRfrZDJhVxdtw=="],
+
+    "@vercel/speed-insights/next/@next/swc-linux-arm64-gnu": ["@next/swc-linux-arm64-gnu@16.2.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-+YqU2h1cQkHsGfvjAsrSmst8UIFBibBGm5x3Xgel8NLMiDQtNOM4sM2GOEMvG5YiOBNeN/Ykk8cQC2S0Xrqljg=="],
+
+    "@vercel/speed-insights/next/@next/swc-linux-arm64-musl": ["@next/swc-linux-arm64-musl@16.2.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-0qjhiYBaKAqF63LA1ZWAAnKTzFUguAaZiRa5etMLGGPj/B6uEVjtIZldIzFEp3wHlB0koK6aTzqPtSdplTCjoA=="],
+
+    "@vercel/speed-insights/next/@next/swc-linux-x64-gnu": ["@next/swc-linux-x64-gnu@16.2.12", "", { "os": "linux", "cpu": "x64" }, "sha512-7A3q26W+h7gnA15uqBToNuDqBEFZZcqh0mW2mn4AJh/G5pdg2RVE3n4slzLEliASZFG3NmsbEzng/x2Sh09mBg=="],
+
+    "@vercel/speed-insights/next/@next/swc-linux-x64-musl": ["@next/swc-linux-x64-musl@16.2.12", "", { "os": "linux", "cpu": "x64" }, "sha512-qSjL/uppm+cbh21s72Ss8gkiOhQ4dExWHNGOWy6eZV7STj5WsKehgxT61beSsOj+YYQuTplL376lOCdMQU5T8w=="],
+
+    "@vercel/speed-insights/next/@next/swc-win32-arm64-msvc": ["@next/swc-win32-arm64-msvc@16.2.12", "", { "os": "win32", "cpu": "arm64" }, "sha512-X6hzsOUJac/e7AWSbn9gQ9nzHld1xWP5iyjHpYWvud8pufB679O1xg4JDyKr8Xd69Jvd+kM2Der6uftiZCmjYA=="],
+
+    "@vercel/speed-insights/next/@next/swc-win32-x64-msvc": ["@next/swc-win32-x64-msvc@16.2.12", "", { "os": "win32", "cpu": "x64" }, "sha512-F6fakeHuFTLOPt0bslQJdf+xtT+WIP9DVn/m4y1w1mRnVPyh3D/cNvzlRkxM444xfm+IvvYNSOrKiA2CDJ0Uxw=="],
+
+    "@vercel/speed-insights/next/@swc/helpers": ["@swc/helpers@0.5.15", "", { "dependencies": { "tslib": "2.8.1" } }, "sha512-JQ5TuMi45Owi4/BIMAJBoSQoOJu12oOk/gADqlcUL9JEdHB8vyjUSsxqeNXnmXHjYKMi2WcYtezGEEhqUI/E2g=="],
+
+    "@vercel/speed-insights/next/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
+    "@vitest/mocker/vite/@types/node": ["@types/node@26.1.1", "", { "dependencies": { "undici-types": "8.3.0" } }, "sha512-nxAkRSVkN1Y0JC1W8ky/fTfkGsMmcrRsbx+3XoZE+rMOX71kLYTV7fLXpqud1GpbpP5TuffXFqfX7fH2GgZREw=="],
+
+    "@vitest/mocker/vite/esbuild": ["esbuild@0.28.1", "", { "optionalDependencies": { "@esbuild/aix-ppc64": "0.28.1", "@esbuild/android-arm": "0.28.1", "@esbuild/android-arm64": "0.28.1", "@esbuild/android-x64": "0.28.1", "@esbuild/darwin-arm64": "0.28.1", "@esbuild/darwin-x64": "0.28.1", "@esbuild/freebsd-arm64": "0.28.1", "@esbuild/freebsd-x64": "0.28.1", "@esbuild/linux-arm": "0.28.1", "@esbuild/linux-arm64": "0.28.1", "@esbuild/linux-ia32": "0.28.1", "@esbuild/linux-loong64": "0.28.1", "@esbuild/linux-mips64el": "0.28.1", "@esbuild/linux-ppc64": "0.28.1", "@esbuild/linux-riscv64": "0.28.1", "@esbuild/linux-s390x": "0.28.1", "@esbuild/linux-x64": "0.28.1", "@esbuild/netbsd-arm64": "0.28.1", "@esbuild/netbsd-x64": "0.28.1", "@esbuild/openbsd-arm64": "0.28.1", "@esbuild/openbsd-x64": "0.28.1", "@esbuild/openharmony-arm64": "0.28.1", "@esbuild/sunos-x64": "0.28.1", "@esbuild/win32-arm64": "0.28.1", "@esbuild/win32-ia32": "0.28.1", "@esbuild/win32-x64": "0.28.1" }, "bin": { "esbuild": "bin/esbuild" } }, "sha512-HrJrvZv5ayxBzPfwphOoNzkzOIIlifzk0KJrGK2c8R4+LKpMtpYLQeUdjnwjWv/LZlkH2laZk+4w78pi99D4Vw=="],
+
+    "@vitest/mocker/vite/lightningcss": ["lightningcss@1.32.0", "", { "dependencies": { "detect-libc": "2.1.2" }, "optionalDependencies": { "lightningcss-android-arm64": "1.32.0", "lightningcss-darwin-arm64": "1.32.0", "lightningcss-darwin-x64": "1.32.0", "lightningcss-freebsd-x64": "1.32.0", "lightningcss-linux-arm-gnueabihf": "1.32.0", "lightningcss-linux-arm64-gnu": "1.32.0", "lightningcss-linux-arm64-musl": "1.32.0", "lightningcss-linux-x64-gnu": "1.32.0", "lightningcss-linux-x64-musl": "1.32.0", "lightningcss-win32-arm64-msvc": "1.32.0", "lightningcss-win32-x64-msvc": "1.32.0" } }, "sha512-NXYBzinNrblfraPGyrbPoD19C1h9lfI/1mzgWYvXUTe414Gz/X1FD2XBZSZM7rRTrMA8JL3OtAaGifrIKhQ5yQ=="],
+
+    "@vitest/mocker/vite/postcss": ["postcss@8.5.23", "", { "dependencies": { "nanoid": "3.3.16", "picocolors": "1.1.1", "source-map-js": "1.2.1" } }, "sha512-g50586zr4bZmwFiTlflMu8E0bDTb5I5gertgwAKmsdUlTQIhZtunzUlD1WSzwcVWPoAVpsrA6vlfCD7oXvRwgg=="],
+
+    "@vitest/mocker/vite/rolldown": ["rolldown@1.1.5", "", { "dependencies": { "@oxc-project/types": "0.139.0", "@rolldown/pluginutils": "1.0.1" }, "optionalDependencies": { "@rolldown/binding-android-arm64": "1.1.5", "@rolldown/binding-darwin-arm64": "1.1.5", "@rolldown/binding-darwin-x64": "1.1.5", "@rolldown/binding-freebsd-x64": "1.1.5", "@rolldown/binding-linux-arm-gnueabihf": "1.1.5", "@rolldown/binding-linux-arm64-gnu": "1.1.5", "@rolldown/binding-linux-arm64-musl": "1.1.5", "@rolldown/binding-linux-ppc64-gnu": "1.1.5", "@rolldown/binding-linux-s390x-gnu": "1.1.5", "@rolldown/binding-linux-x64-gnu": "1.1.5", "@rolldown/binding-linux-x64-musl": "1.1.5", "@rolldown/binding-openharmony-arm64": "1.1.5", "@rolldown/binding-wasm32-wasi": "1.1.5", "@rolldown/binding-win32-arm64-msvc": "1.1.5", "@rolldown/binding-win32-x64-msvc": "1.1.5" }, "bin": { "rolldown": "./bin/cli.mjs" } }, "sha512-t9z29cJjXf/vxQ8dyhCSpt6H6aSwHTk8cT5I3iy6SMXuFpk5mB6PL6XfC8PCwrPTx93udwKUm9HRteAlTGBLiA=="],
+
     "@vue/compiler-core/@babel/parser/@babel/types": ["@babel/types@7.29.7", "", { "dependencies": { "@babel/helper-string-parser": "7.29.7", "@babel/helper-validator-identifier": "7.29.7" } }, "sha512-4zBIxpPzowiZpusoFkyGVwakdRJUyuH5PxQ/PrqghfdFWWasvnCdPfQXHrenDai+gyLARulZjZowCOj6fjT4pA=="],
 
     "@vue/compiler-sfc/@babel/parser/@babel/types": ["@babel/types@7.29.7", "", { "dependencies": { "@babel/helper-string-parser": "7.29.7", "@babel/helper-validator-identifier": "7.29.7" } }, "sha512-4zBIxpPzowiZpusoFkyGVwakdRJUyuH5PxQ/PrqghfdFWWasvnCdPfQXHrenDai+gyLARulZjZowCOj6fjT4pA=="],
@@ -2517,6 +2736,132 @@
 
     "shelljs/execa/human-signals": ["human-signals@2.1.0", "", {}, "sha512-B4FFZ6q/T2jhhksgkbEW3HBvWIfDW85snkQgawt07S7J5QXTk6BkNV+0yAeZrM5QpMAdYlocGoljn0sJ/WQkFw=="],
 
+    "terser-webpack-plugin/esbuild/@esbuild/aix-ppc64": ["@esbuild/aix-ppc64@0.28.1", "", { "os": "aix", "cpu": "ppc64" }, "sha512-Svl7tq8k/08+p6CXPpRjQ1fKX+1odH/BQbb48fV6fj3CWHhsoIOoY87w1oHXm0qEpkIK3ZfVgp0hed3XBXzXMQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/android-arm": ["@esbuild/android-arm@0.28.1", "", { "os": "android", "cpu": "arm" }, "sha512-0k2F129Xdio1TdJfzJ8sy1Q47vUD2NnwdhiAf7drUN1EBTfPf4hsFCtmMgu/6m8JSzsBrlmVjudMBQqOfG8usQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/android-arm64": ["@esbuild/android-arm64@0.28.1", "", { "os": "android", "cpu": "arm64" }, "sha512-34EGEbCIAgosYz6goLcopX6Mo7NyGv9tfwEM2/7Ce2VcVRk568iSvniGWcUXIy7wEDR1wzolcxcriFVrWYcwBg=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/android-x64": ["@esbuild/android-x64@0.28.1", "", { "os": "android", "cpu": "x64" }, "sha512-dbwY7ltSMDWsRatcRpCnES4F+im88OCUgGZjy52shC7GqHRE/cYlxNbB4Z4UpJswpcc4Qxd2oE/ufM0p61IKng=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/darwin-arm64": ["@esbuild/darwin-arm64@0.28.1", "", { "os": "darwin", "cpu": "arm64" }, "sha512-TZbWkQY7kvTAXbXUT7uVACR5cMHsDiSz9z7ZKAX/RTq/WJEk3QyRr0wZpNhBDX+/0CtdqUIJlOiodQcta6tY3Q=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/darwin-x64": ["@esbuild/darwin-x64@0.28.1", "", { "os": "darwin", "cpu": "x64" }, "sha512-zfdzgK9ACBNZLI/CyHTOx81SyNbM6YXn7rxSgX97VjyiPl9W1i4Ka4fgKECEoFCKGpvBj5qArWIGgQjOwkgskQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/freebsd-arm64": ["@esbuild/freebsd-arm64@0.28.1", "", { "os": "freebsd", "cpu": "arm64" }, "sha512-wG2EA8ENdEI0qhkSZMjfqrdY+ziCYCPMmtZjjIwOmXFjmyzEHn+UUxk5of+SYsjtfs3VpnlC7QLzSI5hY/rOAw=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/freebsd-x64": ["@esbuild/freebsd-x64@0.28.1", "", { "os": "freebsd", "cpu": "x64" }, "sha512-i7dZ9vQgnvSCzi/rYCXNgtF/U+eKZNJBzu3eTQbRgHnM7tNSizLOkRFAl3qzVc/Op/u5YkHHa4pf/3DOYHthLQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-arm": ["@esbuild/linux-arm@0.28.1", "", { "os": "linux", "cpu": "arm" }, "sha512-qVXBOHQS+d5Y722GwJzJUtOLlX7km3CraOaGormF1pDtPd2C/l1SHRPgjLunLGe51Sh5YYWKMFDyV4SxgMQYTQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-arm64": ["@esbuild/linux-arm64@0.28.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-yHs+0uc8+nvEAfAfxrWQKK5peSNzBc4PegcMO0EJ2hT71uA7vB8Ihg2e77R2P7SG5uYjPbHlLLmve4LLLRCf0g=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-ia32": ["@esbuild/linux-ia32@0.28.1", "", { "os": "linux", "cpu": "ia32" }, "sha512-d1z4ZuP0ajrfz/FhGT4vv278rX8KnPPJx8i5+AtK7TYbx9Le9F1hyzurZpkEyjkGa9dUGhQow4C1NmeGvqxN2w=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-loong64": ["@esbuild/linux-loong64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-M5sRjUVZrkm1OAPR3dlOYzNmN+loZKGVi1VUQGrwuqLcbR6qeAz+famMhjASeH3YVKvZz+zT1jlh/keC3Rj/lg=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-mips64el": ["@esbuild/linux-mips64el@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-mRObBZeHh2OxcBFPWE/FjylkRgZdYuiTR3vaTozquCGOH14iP9oN4x4Ge81CoIDYQrXmIxpFumJBu5MtZpnQJQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-ppc64": ["@esbuild/linux-ppc64@0.28.1", "", { "os": "linux", "cpu": "ppc64" }, "sha512-slScBsMAb3GFDcdrCgLwZtPYRoH2H/youv10QiZyRjmsP48fznoveWytSgCI/R0ZcUgpc0ZhIUEx6LHts8yrfQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-riscv64": ["@esbuild/linux-riscv64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-kw0owk1o0GFETUJyW0jc0G4Yzs0BHZn0JDZ8JRT088vjJYX777BAs1fDGxAC+q831qOs2DTC96mNsG2opdfyyQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-s390x": ["@esbuild/linux-s390x@0.28.1", "", { "os": "linux", "cpu": "s390x" }, "sha512-/lAIjX8aYFRByhh6L5rYtPEDRqa9de/4V/juOXcta5frjvzXO4/sqEtyytse0g3zZFuWu5cDN0MkLz2qRDD2Ag=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/linux-x64": ["@esbuild/linux-x64@0.28.1", "", { "os": "linux", "cpu": "x64" }, "sha512-u/anNYF2mmVOEDwLtnQ1wOr3EZ9sTNGLWrsYGYwHWzGA3Si84IOkHXlbWTD1NB+9/1lcnweYKO54uhxZydNzfA=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/netbsd-arm64": ["@esbuild/netbsd-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-oks0DYbLwWMmaakTsCb+zL4E+aHRVLom9IJZOAthMQEPiQmydXHkziYEsGYRx0uNV/IjEKGAV941JzH02pflqw=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/netbsd-x64": ["@esbuild/netbsd-x64@0.28.1", "", { "os": "none", "cpu": "x64" }, "sha512-aeL6lAnN89Hz43Mlh1G8ARasbuoYvSITDEx0tHh5b7jJnHcssqgjy9Yx430GDpmCa6OyrKoS0aNRjKundRizGg=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/openbsd-arm64": ["@esbuild/openbsd-arm64@0.28.1", "", { "os": "openbsd", "cpu": "arm64" }, "sha512-MEFJe5C3R8pwXdZ5Y21oo6m7ePiS0d9pWucn99O/wvyJZChoIQKrQDxKrGeW8F5+T0okTHesAmDeiHDTIq0V/Q=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/openbsd-x64": ["@esbuild/openbsd-x64@0.28.1", "", { "os": "openbsd", "cpu": "x64" }, "sha512-i/ZLIOafE0Z8cI/XANJAixoJL/uRAoS2xOA3rb0xN+KK0K177cMAsQYkzHtBrtMXAKuAc7HGgcWiZ/sRC1Nxgw=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/openharmony-arm64": ["@esbuild/openharmony-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-ge+Z7EXFNt2BO1oAMsVpiQ8EwndV9i1xXerAeTIK7AtPs3bKFXQM7nlRxDSIUIMeueR1CNXxqztLzdNeReKBJg=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/sunos-x64": ["@esbuild/sunos-x64@0.28.1", "", { "os": "sunos", "cpu": "x64" }, "sha512-BEjgtECkL3vY+SaSQ6nzVfiALUeFxpawyp8Jmf5PtYhf1Ug40N1h/hxlhts+f1FvSvarEigdxS3BlSMI2PJLcQ=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/win32-arm64": ["@esbuild/win32-arm64@0.28.1", "", { "os": "win32", "cpu": "arm64" }, "sha512-lCv9eK/H6ZJWbE7bh2nw54CZ9M2nupBxJcTsdk/QQnWkdSjKGuxmmH8/GWrlT1eMmZfn4dGcCjRte397WqfQXA=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/win32-ia32": ["@esbuild/win32-ia32@0.28.1", "", { "os": "win32", "cpu": "ia32" }, "sha512-zvb/mB2bSCoJOpoCBgYKKpX6YM6mJBlBUVUtVj41DlZJVEB6/0CKlRYxP5wWl1C1ILiCoAU5wZZ4q1P3qeS6Eg=="],
+
+    "terser-webpack-plugin/esbuild/@esbuild/win32-x64": ["@esbuild/win32-x64@0.28.1", "", { "os": "win32", "cpu": "x64" }, "sha512-bm4Mowrv+GXMlpWX++EcXw/iLyd1o3+bJkC2DkWXYVvgZCqD/bSj9ctZeAMC3cIxgjRVR2Dufaiu4YPxr5gW1A=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-android-arm64": ["lightningcss-android-arm64@1.32.0", "", { "os": "android", "cpu": "arm64" }, "sha512-YK7/ClTt4kAK0vo6w3X+Pnm0D2cf2vPHbhOXdoNti1Ga0al1P4TBZhwjATvjNwLEBCnKvjJc2jQgHXH0NEwlAg=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-darwin-arm64": ["lightningcss-darwin-arm64@1.32.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-RzeG9Ju5bag2Bv1/lwlVJvBE3q6TtXskdZLLCyfg5pt+HLz9BqlICO7LZM7VHNTTn/5PRhHFBSjk5lc4cmscPQ=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-darwin-x64": ["lightningcss-darwin-x64@1.32.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-U+QsBp2m/s2wqpUYT/6wnlagdZbtZdndSmut/NJqlCcMLTWp5muCrID+K5UJ6jqD2BFshejCYXniPDbNh73V8w=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-freebsd-x64": ["lightningcss-freebsd-x64@1.32.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-JCTigedEksZk3tHTTthnMdVfGf61Fky8Ji2E4YjUTEQX14xiy/lTzXnu1vwiZe3bYe0q+SpsSH/CTeDXK6WHig=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-linux-arm-gnueabihf": ["lightningcss-linux-arm-gnueabihf@1.32.0", "", { "os": "linux", "cpu": "arm" }, "sha512-x6rnnpRa2GL0zQOkt6rts3YDPzduLpWvwAF6EMhXFVZXD4tPrBkEFqzGowzCsIWsPjqSK+tyNEODUBXeeVHSkw=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-linux-arm64-gnu": ["lightningcss-linux-arm64-gnu@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-0nnMyoyOLRJXfbMOilaSRcLH3Jw5z9HDNGfT/gwCPgaDjnx0i8w7vBzFLFR1f6CMLKF8gVbebmkUN3fa/kQJpQ=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-linux-arm64-musl": ["lightningcss-linux-arm64-musl@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-UpQkoenr4UJEzgVIYpI80lDFvRmPVg6oqboNHfoH4CQIfNA+HOrZ7Mo7KZP02dC6LjghPQJeBsvXhJod/wnIBg=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-linux-x64-gnu": ["lightningcss-linux-x64-gnu@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-V7Qr52IhZmdKPVr+Vtw8o+WLsQJYCTd8loIfpDaMRWGUZfBOYEJeyJIkqGIDMZPwPx24pUMfwSxxI8phr/MbOA=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-linux-x64-musl": ["lightningcss-linux-x64-musl@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-bYcLp+Vb0awsiXg/80uCRezCYHNg1/l3mt0gzHnWV9XP1W5sKa5/TCdGWaR/zBM2PeF/HbsQv/j2URNOiVuxWg=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-win32-arm64-msvc": ["lightningcss-win32-arm64-msvc@1.32.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-8SbC8BR40pS6baCM8sbtYDSwEVQd4JlFTOlaD3gWGHfThTcABnNDBda6eTZeqbofalIJhFx0qKzgHJmcPTnGdw=="],
+
+    "terser-webpack-plugin/lightningcss/lightningcss-win32-x64-msvc": ["lightningcss-win32-x64-msvc@1.32.0", "", { "os": "win32", "cpu": "x64" }, "sha512-Amq9B/SoZYdDi1kFrojnoqPLxYhQ4Wo5XiL8EVJrVsB8ARoC1PWW6VGtT0WKCemjy8aC+louJnjS7U18x3b06Q=="],
+
+    "tsx/esbuild/@esbuild/aix-ppc64": ["@esbuild/aix-ppc64@0.28.1", "", { "os": "aix", "cpu": "ppc64" }, "sha512-Svl7tq8k/08+p6CXPpRjQ1fKX+1odH/BQbb48fV6fj3CWHhsoIOoY87w1oHXm0qEpkIK3ZfVgp0hed3XBXzXMQ=="],
+
+    "tsx/esbuild/@esbuild/android-arm": ["@esbuild/android-arm@0.28.1", "", { "os": "android", "cpu": "arm" }, "sha512-0k2F129Xdio1TdJfzJ8sy1Q47vUD2NnwdhiAf7drUN1EBTfPf4hsFCtmMgu/6m8JSzsBrlmVjudMBQqOfG8usQ=="],
+
+    "tsx/esbuild/@esbuild/android-arm64": ["@esbuild/android-arm64@0.28.1", "", { "os": "android", "cpu": "arm64" }, "sha512-34EGEbCIAgosYz6goLcopX6Mo7NyGv9tfwEM2/7Ce2VcVRk568iSvniGWcUXIy7wEDR1wzolcxcriFVrWYcwBg=="],
+
+    "tsx/esbuild/@esbuild/android-x64": ["@esbuild/android-x64@0.28.1", "", { "os": "android", "cpu": "x64" }, "sha512-dbwY7ltSMDWsRatcRpCnES4F+im88OCUgGZjy52shC7GqHRE/cYlxNbB4Z4UpJswpcc4Qxd2oE/ufM0p61IKng=="],
+
+    "tsx/esbuild/@esbuild/darwin-arm64": ["@esbuild/darwin-arm64@0.28.1", "", { "os": "darwin", "cpu": "arm64" }, "sha512-TZbWkQY7kvTAXbXUT7uVACR5cMHsDiSz9z7ZKAX/RTq/WJEk3QyRr0wZpNhBDX+/0CtdqUIJlOiodQcta6tY3Q=="],
+
+    "tsx/esbuild/@esbuild/darwin-x64": ["@esbuild/darwin-x64@0.28.1", "", { "os": "darwin", "cpu": "x64" }, "sha512-zfdzgK9ACBNZLI/CyHTOx81SyNbM6YXn7rxSgX97VjyiPl9W1i4Ka4fgKECEoFCKGpvBj5qArWIGgQjOwkgskQ=="],
+
+    "tsx/esbuild/@esbuild/freebsd-arm64": ["@esbuild/freebsd-arm64@0.28.1", "", { "os": "freebsd", "cpu": "arm64" }, "sha512-wG2EA8ENdEI0qhkSZMjfqrdY+ziCYCPMmtZjjIwOmXFjmyzEHn+UUxk5of+SYsjtfs3VpnlC7QLzSI5hY/rOAw=="],
+
+    "tsx/esbuild/@esbuild/freebsd-x64": ["@esbuild/freebsd-x64@0.28.1", "", { "os": "freebsd", "cpu": "x64" }, "sha512-i7dZ9vQgnvSCzi/rYCXNgtF/U+eKZNJBzu3eTQbRgHnM7tNSizLOkRFAl3qzVc/Op/u5YkHHa4pf/3DOYHthLQ=="],
+
+    "tsx/esbuild/@esbuild/linux-arm": ["@esbuild/linux-arm@0.28.1", "", { "os": "linux", "cpu": "arm" }, "sha512-qVXBOHQS+d5Y722GwJzJUtOLlX7km3CraOaGormF1pDtPd2C/l1SHRPgjLunLGe51Sh5YYWKMFDyV4SxgMQYTQ=="],
+
+    "tsx/esbuild/@esbuild/linux-arm64": ["@esbuild/linux-arm64@0.28.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-yHs+0uc8+nvEAfAfxrWQKK5peSNzBc4PegcMO0EJ2hT71uA7vB8Ihg2e77R2P7SG5uYjPbHlLLmve4LLLRCf0g=="],
+
+    "tsx/esbuild/@esbuild/linux-ia32": ["@esbuild/linux-ia32@0.28.1", "", { "os": "linux", "cpu": "ia32" }, "sha512-d1z4ZuP0ajrfz/FhGT4vv278rX8KnPPJx8i5+AtK7TYbx9Le9F1hyzurZpkEyjkGa9dUGhQow4C1NmeGvqxN2w=="],
+
+    "tsx/esbuild/@esbuild/linux-loong64": ["@esbuild/linux-loong64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-M5sRjUVZrkm1OAPR3dlOYzNmN+loZKGVi1VUQGrwuqLcbR6qeAz+famMhjASeH3YVKvZz+zT1jlh/keC3Rj/lg=="],
+
+    "tsx/esbuild/@esbuild/linux-mips64el": ["@esbuild/linux-mips64el@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-mRObBZeHh2OxcBFPWE/FjylkRgZdYuiTR3vaTozquCGOH14iP9oN4x4Ge81CoIDYQrXmIxpFumJBu5MtZpnQJQ=="],
+
+    "tsx/esbuild/@esbuild/linux-ppc64": ["@esbuild/linux-ppc64@0.28.1", "", { "os": "linux", "cpu": "ppc64" }, "sha512-slScBsMAb3GFDcdrCgLwZtPYRoH2H/youv10QiZyRjmsP48fznoveWytSgCI/R0ZcUgpc0ZhIUEx6LHts8yrfQ=="],
+
+    "tsx/esbuild/@esbuild/linux-riscv64": ["@esbuild/linux-riscv64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-kw0owk1o0GFETUJyW0jc0G4Yzs0BHZn0JDZ8JRT088vjJYX777BAs1fDGxAC+q831qOs2DTC96mNsG2opdfyyQ=="],
+
+    "tsx/esbuild/@esbuild/linux-s390x": ["@esbuild/linux-s390x@0.28.1", "", { "os": "linux", "cpu": "s390x" }, "sha512-/lAIjX8aYFRByhh6L5rYtPEDRqa9de/4V/juOXcta5frjvzXO4/sqEtyytse0g3zZFuWu5cDN0MkLz2qRDD2Ag=="],
+
+    "tsx/esbuild/@esbuild/linux-x64": ["@esbuild/linux-x64@0.28.1", "", { "os": "linux", "cpu": "x64" }, "sha512-u/anNYF2mmVOEDwLtnQ1wOr3EZ9sTNGLWrsYGYwHWzGA3Si84IOkHXlbWTD1NB+9/1lcnweYKO54uhxZydNzfA=="],
+
+    "tsx/esbuild/@esbuild/netbsd-arm64": ["@esbuild/netbsd-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-oks0DYbLwWMmaakTsCb+zL4E+aHRVLom9IJZOAthMQEPiQmydXHkziYEsGYRx0uNV/IjEKGAV941JzH02pflqw=="],
+
+    "tsx/esbuild/@esbuild/netbsd-x64": ["@esbuild/netbsd-x64@0.28.1", "", { "os": "none", "cpu": "x64" }, "sha512-aeL6lAnN89Hz43Mlh1G8ARasbuoYvSITDEx0tHh5b7jJnHcssqgjy9Yx430GDpmCa6OyrKoS0aNRjKundRizGg=="],
+
+    "tsx/esbuild/@esbuild/openbsd-arm64": ["@esbuild/openbsd-arm64@0.28.1", "", { "os": "openbsd", "cpu": "arm64" }, "sha512-MEFJe5C3R8pwXdZ5Y21oo6m7ePiS0d9pWucn99O/wvyJZChoIQKrQDxKrGeW8F5+T0okTHesAmDeiHDTIq0V/Q=="],
+
+    "tsx/esbuild/@esbuild/openbsd-x64": ["@esbuild/openbsd-x64@0.28.1", "", { "os": "openbsd", "cpu": "x64" }, "sha512-i/ZLIOafE0Z8cI/XANJAixoJL/uRAoS2xOA3rb0xN+KK0K177cMAsQYkzHtBrtMXAKuAc7HGgcWiZ/sRC1Nxgw=="],
+
+    "tsx/esbuild/@esbuild/openharmony-arm64": ["@esbuild/openharmony-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-ge+Z7EXFNt2BO1oAMsVpiQ8EwndV9i1xXerAeTIK7AtPs3bKFXQM7nlRxDSIUIMeueR1CNXxqztLzdNeReKBJg=="],
+
+    "tsx/esbuild/@esbuild/sunos-x64": ["@esbuild/sunos-x64@0.28.1", "", { "os": "sunos", "cpu": "x64" }, "sha512-BEjgtECkL3vY+SaSQ6nzVfiALUeFxpawyp8Jmf5PtYhf1Ug40N1h/hxlhts+f1FvSvarEigdxS3BlSMI2PJLcQ=="],
+
+    "tsx/esbuild/@esbuild/win32-arm64": ["@esbuild/win32-arm64@0.28.1", "", { "os": "win32", "cpu": "arm64" }, "sha512-lCv9eK/H6ZJWbE7bh2nw54CZ9M2nupBxJcTsdk/QQnWkdSjKGuxmmH8/GWrlT1eMmZfn4dGcCjRte397WqfQXA=="],
+
+    "tsx/esbuild/@esbuild/win32-ia32": ["@esbuild/win32-ia32@0.28.1", "", { "os": "win32", "cpu": "ia32" }, "sha512-zvb/mB2bSCoJOpoCBgYKKpX6YM6mJBlBUVUtVj41DlZJVEB6/0CKlRYxP5wWl1C1ILiCoAU5wZZ4q1P3qeS6Eg=="],
+
+    "tsx/esbuild/@esbuild/win32-x64": ["@esbuild/win32-x64@0.28.1", "", { "os": "win32", "cpu": "x64" }, "sha512-bm4Mowrv+GXMlpWX++EcXw/iLyd1o3+bJkC2DkWXYVvgZCqD/bSj9ctZeAMC3cIxgjRVR2Dufaiu4YPxr5gW1A=="],
+
     "type-is/mime-types/mime-db": ["mime-db@1.54.0", "", {}, "sha512-aU5EJuIN2WDemCcAp2vFBfp/m4EAhWJnUNSSw0ixs7/kXbd6Pg64EmwJkNdFhB8aWt1sH2CTXrLxo/iAGV3oPQ=="],
 
     "wrap-ansi/string-width/emoji-regex": ["emoji-regex@10.6.0", "", {}, "sha512-toUI84YS5YmxW219erniWD0CIVOo46xGKColeNQRgOzDorgBi1v4D71/OFzgD9GO2UGKIv1C3Sp8DAn0+j5w7A=="],
@@ -2525,6 +2870,114 @@
 
     "@cypress/request/tough-cookie/tldts/tldts-core": ["tldts-core@6.1.86", "", {}, "sha512-Je6p7pkk+KMzMv2XXKmAE3McmolOQFdxkKw0R8EYNr7sELW46JqnNeTX8ybPiQgvg1ymCoF8LXs5fzFaZvJPTA=="],
 
+    "@vercel/analytics/next/@swc/helpers/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
+
+    "@vercel/speed-insights/next/@swc/helpers/tslib": ["tslib@2.8.1", "", {}, "sha512-oJFu94HQb+KVduSUQL7wnpmqnfmLsOA/nAh6b6EH0wCEoK0/mPeXU6c3wKDV83MkOuHPRHtSXKKU99IBazS/2w=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/aix-ppc64": ["@esbuild/aix-ppc64@0.28.1", "", { "os": "aix", "cpu": "ppc64" }, "sha512-Svl7tq8k/08+p6CXPpRjQ1fKX+1odH/BQbb48fV6fj3CWHhsoIOoY87w1oHXm0qEpkIK3ZfVgp0hed3XBXzXMQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/android-arm": ["@esbuild/android-arm@0.28.1", "", { "os": "android", "cpu": "arm" }, "sha512-0k2F129Xdio1TdJfzJ8sy1Q47vUD2NnwdhiAf7drUN1EBTfPf4hsFCtmMgu/6m8JSzsBrlmVjudMBQqOfG8usQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/android-arm64": ["@esbuild/android-arm64@0.28.1", "", { "os": "android", "cpu": "arm64" }, "sha512-34EGEbCIAgosYz6goLcopX6Mo7NyGv9tfwEM2/7Ce2VcVRk568iSvniGWcUXIy7wEDR1wzolcxcriFVrWYcwBg=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/android-x64": ["@esbuild/android-x64@0.28.1", "", { "os": "android", "cpu": "x64" }, "sha512-dbwY7ltSMDWsRatcRpCnES4F+im88OCUgGZjy52shC7GqHRE/cYlxNbB4Z4UpJswpcc4Qxd2oE/ufM0p61IKng=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/darwin-arm64": ["@esbuild/darwin-arm64@0.28.1", "", { "os": "darwin", "cpu": "arm64" }, "sha512-TZbWkQY7kvTAXbXUT7uVACR5cMHsDiSz9z7ZKAX/RTq/WJEk3QyRr0wZpNhBDX+/0CtdqUIJlOiodQcta6tY3Q=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/darwin-x64": ["@esbuild/darwin-x64@0.28.1", "", { "os": "darwin", "cpu": "x64" }, "sha512-zfdzgK9ACBNZLI/CyHTOx81SyNbM6YXn7rxSgX97VjyiPl9W1i4Ka4fgKECEoFCKGpvBj5qArWIGgQjOwkgskQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/freebsd-arm64": ["@esbuild/freebsd-arm64@0.28.1", "", { "os": "freebsd", "cpu": "arm64" }, "sha512-wG2EA8ENdEI0qhkSZMjfqrdY+ziCYCPMmtZjjIwOmXFjmyzEHn+UUxk5of+SYsjtfs3VpnlC7QLzSI5hY/rOAw=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/freebsd-x64": ["@esbuild/freebsd-x64@0.28.1", "", { "os": "freebsd", "cpu": "x64" }, "sha512-i7dZ9vQgnvSCzi/rYCXNgtF/U+eKZNJBzu3eTQbRgHnM7tNSizLOkRFAl3qzVc/Op/u5YkHHa4pf/3DOYHthLQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-arm": ["@esbuild/linux-arm@0.28.1", "", { "os": "linux", "cpu": "arm" }, "sha512-qVXBOHQS+d5Y722GwJzJUtOLlX7km3CraOaGormF1pDtPd2C/l1SHRPgjLunLGe51Sh5YYWKMFDyV4SxgMQYTQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-arm64": ["@esbuild/linux-arm64@0.28.1", "", { "os": "linux", "cpu": "arm64" }, "sha512-yHs+0uc8+nvEAfAfxrWQKK5peSNzBc4PegcMO0EJ2hT71uA7vB8Ihg2e77R2P7SG5uYjPbHlLLmve4LLLRCf0g=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-ia32": ["@esbuild/linux-ia32@0.28.1", "", { "os": "linux", "cpu": "ia32" }, "sha512-d1z4ZuP0ajrfz/FhGT4vv278rX8KnPPJx8i5+AtK7TYbx9Le9F1hyzurZpkEyjkGa9dUGhQow4C1NmeGvqxN2w=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-loong64": ["@esbuild/linux-loong64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-M5sRjUVZrkm1OAPR3dlOYzNmN+loZKGVi1VUQGrwuqLcbR6qeAz+famMhjASeH3YVKvZz+zT1jlh/keC3Rj/lg=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-mips64el": ["@esbuild/linux-mips64el@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-mRObBZeHh2OxcBFPWE/FjylkRgZdYuiTR3vaTozquCGOH14iP9oN4x4Ge81CoIDYQrXmIxpFumJBu5MtZpnQJQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-ppc64": ["@esbuild/linux-ppc64@0.28.1", "", { "os": "linux", "cpu": "ppc64" }, "sha512-slScBsMAb3GFDcdrCgLwZtPYRoH2H/youv10QiZyRjmsP48fznoveWytSgCI/R0ZcUgpc0ZhIUEx6LHts8yrfQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-riscv64": ["@esbuild/linux-riscv64@0.28.1", "", { "os": "linux", "cpu": "none" }, "sha512-kw0owk1o0GFETUJyW0jc0G4Yzs0BHZn0JDZ8JRT088vjJYX777BAs1fDGxAC+q831qOs2DTC96mNsG2opdfyyQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-s390x": ["@esbuild/linux-s390x@0.28.1", "", { "os": "linux", "cpu": "s390x" }, "sha512-/lAIjX8aYFRByhh6L5rYtPEDRqa9de/4V/juOXcta5frjvzXO4/sqEtyytse0g3zZFuWu5cDN0MkLz2qRDD2Ag=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/linux-x64": ["@esbuild/linux-x64@0.28.1", "", { "os": "linux", "cpu": "x64" }, "sha512-u/anNYF2mmVOEDwLtnQ1wOr3EZ9sTNGLWrsYGYwHWzGA3Si84IOkHXlbWTD1NB+9/1lcnweYKO54uhxZydNzfA=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/netbsd-arm64": ["@esbuild/netbsd-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-oks0DYbLwWMmaakTsCb+zL4E+aHRVLom9IJZOAthMQEPiQmydXHkziYEsGYRx0uNV/IjEKGAV941JzH02pflqw=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/netbsd-x64": ["@esbuild/netbsd-x64@0.28.1", "", { "os": "none", "cpu": "x64" }, "sha512-aeL6lAnN89Hz43Mlh1G8ARasbuoYvSITDEx0tHh5b7jJnHcssqgjy9Yx430GDpmCa6OyrKoS0aNRjKundRizGg=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/openbsd-arm64": ["@esbuild/openbsd-arm64@0.28.1", "", { "os": "openbsd", "cpu": "arm64" }, "sha512-MEFJe5C3R8pwXdZ5Y21oo6m7ePiS0d9pWucn99O/wvyJZChoIQKrQDxKrGeW8F5+T0okTHesAmDeiHDTIq0V/Q=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/openbsd-x64": ["@esbuild/openbsd-x64@0.28.1", "", { "os": "openbsd", "cpu": "x64" }, "sha512-i/ZLIOafE0Z8cI/XANJAixoJL/uRAoS2xOA3rb0xN+KK0K177cMAsQYkzHtBrtMXAKuAc7HGgcWiZ/sRC1Nxgw=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/openharmony-arm64": ["@esbuild/openharmony-arm64@0.28.1", "", { "os": "none", "cpu": "arm64" }, "sha512-ge+Z7EXFNt2BO1oAMsVpiQ8EwndV9i1xXerAeTIK7AtPs3bKFXQM7nlRxDSIUIMeueR1CNXxqztLzdNeReKBJg=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/sunos-x64": ["@esbuild/sunos-x64@0.28.1", "", { "os": "sunos", "cpu": "x64" }, "sha512-BEjgtECkL3vY+SaSQ6nzVfiALUeFxpawyp8Jmf5PtYhf1Ug40N1h/hxlhts+f1FvSvarEigdxS3BlSMI2PJLcQ=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/win32-arm64": ["@esbuild/win32-arm64@0.28.1", "", { "os": "win32", "cpu": "arm64" }, "sha512-lCv9eK/H6ZJWbE7bh2nw54CZ9M2nupBxJcTsdk/QQnWkdSjKGuxmmH8/GWrlT1eMmZfn4dGcCjRte397WqfQXA=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/win32-ia32": ["@esbuild/win32-ia32@0.28.1", "", { "os": "win32", "cpu": "ia32" }, "sha512-zvb/mB2bSCoJOpoCBgYKKpX6YM6mJBlBUVUtVj41DlZJVEB6/0CKlRYxP5wWl1C1ILiCoAU5wZZ4q1P3qeS6Eg=="],
+
+    "@vitest/mocker/vite/esbuild/@esbuild/win32-x64": ["@esbuild/win32-x64@0.28.1", "", { "os": "win32", "cpu": "x64" }, "sha512-bm4Mowrv+GXMlpWX++EcXw/iLyd1o3+bJkC2DkWXYVvgZCqD/bSj9ctZeAMC3cIxgjRVR2Dufaiu4YPxr5gW1A=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-android-arm64": ["lightningcss-android-arm64@1.32.0", "", { "os": "android", "cpu": "arm64" }, "sha512-YK7/ClTt4kAK0vo6w3X+Pnm0D2cf2vPHbhOXdoNti1Ga0al1P4TBZhwjATvjNwLEBCnKvjJc2jQgHXH0NEwlAg=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-darwin-arm64": ["lightningcss-darwin-arm64@1.32.0", "", { "os": "darwin", "cpu": "arm64" }, "sha512-RzeG9Ju5bag2Bv1/lwlVJvBE3q6TtXskdZLLCyfg5pt+HLz9BqlICO7LZM7VHNTTn/5PRhHFBSjk5lc4cmscPQ=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-darwin-x64": ["lightningcss-darwin-x64@1.32.0", "", { "os": "darwin", "cpu": "x64" }, "sha512-U+QsBp2m/s2wqpUYT/6wnlagdZbtZdndSmut/NJqlCcMLTWp5muCrID+K5UJ6jqD2BFshejCYXniPDbNh73V8w=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-freebsd-x64": ["lightningcss-freebsd-x64@1.32.0", "", { "os": "freebsd", "cpu": "x64" }, "sha512-JCTigedEksZk3tHTTthnMdVfGf61Fky8Ji2E4YjUTEQX14xiy/lTzXnu1vwiZe3bYe0q+SpsSH/CTeDXK6WHig=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-linux-arm-gnueabihf": ["lightningcss-linux-arm-gnueabihf@1.32.0", "", { "os": "linux", "cpu": "arm" }, "sha512-x6rnnpRa2GL0zQOkt6rts3YDPzduLpWvwAF6EMhXFVZXD4tPrBkEFqzGowzCsIWsPjqSK+tyNEODUBXeeVHSkw=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-linux-arm64-gnu": ["lightningcss-linux-arm64-gnu@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-0nnMyoyOLRJXfbMOilaSRcLH3Jw5z9HDNGfT/gwCPgaDjnx0i8w7vBzFLFR1f6CMLKF8gVbebmkUN3fa/kQJpQ=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-linux-arm64-musl": ["lightningcss-linux-arm64-musl@1.32.0", "", { "os": "linux", "cpu": "arm64" }, "sha512-UpQkoenr4UJEzgVIYpI80lDFvRmPVg6oqboNHfoH4CQIfNA+HOrZ7Mo7KZP02dC6LjghPQJeBsvXhJod/wnIBg=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-linux-x64-gnu": ["lightningcss-linux-x64-gnu@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-V7Qr52IhZmdKPVr+Vtw8o+WLsQJYCTd8loIfpDaMRWGUZfBOYEJeyJIkqGIDMZPwPx24pUMfwSxxI8phr/MbOA=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-linux-x64-musl": ["lightningcss-linux-x64-musl@1.32.0", "", { "os": "linux", "cpu": "x64" }, "sha512-bYcLp+Vb0awsiXg/80uCRezCYHNg1/l3mt0gzHnWV9XP1W5sKa5/TCdGWaR/zBM2PeF/HbsQv/j2URNOiVuxWg=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-win32-arm64-msvc": ["lightningcss-win32-arm64-msvc@1.32.0", "", { "os": "win32", "cpu": "arm64" }, "sha512-8SbC8BR40pS6baCM8sbtYDSwEVQd4JlFTOlaD3gWGHfThTcABnNDBda6eTZeqbofalIJhFx0qKzgHJmcPTnGdw=="],
+
+    "@vitest/mocker/vite/lightningcss/lightningcss-win32-x64-msvc": ["lightningcss-win32-x64-msvc@1.32.0", "", { "os": "win32", "cpu": "x64" }, "sha512-Amq9B/SoZYdDi1kFrojnoqPLxYhQ4Wo5XiL8EVJrVsB8ARoC1PWW6VGtT0WKCemjy8aC+louJnjS7U18x3b06Q=="],
+
+    "@vitest/mocker/vite/rolldown/@oxc-project/types": ["@oxc-project/types@0.139.0", "", {}, "sha512-r9gHphtCs+1M7J0pw6Sn/hh/Wpa/iQrOOkrNAlVLF/gHq+/CJmHIWKKUUhdWjcD6CIa8idarspCsASiXCXvFUw=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-android-arm64": ["@rolldown/binding-android-arm64@1.1.5", "", { "os": "android", "cpu": "arm64" }, "sha512-lZg8fqIv2v7FF237bwMgzGZEJvGL79/s5knJ/i6FmsGF4XXlzccZ4jb+TrFIxtSSxFtIpdsgrPZeMk1I9AFcyQ=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-darwin-arm64": ["@rolldown/binding-darwin-arm64@1.1.5", "", { "os": "darwin", "cpu": "arm64" }, "sha512-51Bnx9pNiMRKSUNtBfySkNJ9vMU9Hh3I1ozDd6gyPPYzaXCfnptUcEZxXGYFn+ul2dtcMUiqGR1Yai2K10uoTw=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-darwin-x64": ["@rolldown/binding-darwin-x64@1.1.5", "", { "os": "darwin", "cpu": "x64" }, "sha512-Tm+gbfC0aHu1tBA/JvKQh32S0K6YgCHkiAF4/W6xX0K0RmNuc94VeK419dJoE65R5aRxmo+noZQSWrAMF6yb6g=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-freebsd-x64": ["@rolldown/binding-freebsd-x64@1.1.5", "", { "os": "freebsd", "cpu": "x64" }, "sha512-JMzDKCCXq93YccG5gz3hvOs1oXRKAf0XYpfOS88e+wZrC8Iugj6j68867vrYZkvpDDpKn/KoKORThmchMpF6TA=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-linux-arm-gnueabihf": ["@rolldown/binding-linux-arm-gnueabihf@1.1.5", "", { "os": "linux", "cpu": "arm" }, "sha512-uML21j2K5TfPGutKxub+M+nLjZIrWjXQ5Grx4lCe/nimTj9B4L63zHpjXLl4y0L3mcm2htEQIb06oCG/szerNw=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-linux-arm64-gnu": ["@rolldown/binding-linux-arm64-gnu@1.1.5", "", { "os": "linux", "cpu": "arm64" }, "sha512-navSiuTMogvnQoZoM/v+l3ZWo50/NTwSHSzheABx/RCnmUPaKwq9qSo4Br2OYRs21+Fz8uFqITZM3H4opOB0/Q=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-linux-arm64-musl": ["@rolldown/binding-linux-arm64-musl@1.1.5", "", { "os": "linux", "cpu": "arm64" }, "sha512-lAryqH7IteztmCXQXk0etKj4wBQ7Gx5S6LjKhsgp9zb8I5bsuvU/2llH1hDQcjsFeqIsovMVN339/8pUDDBXxA=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-linux-ppc64-gnu": ["@rolldown/binding-linux-ppc64-gnu@1.1.5", "", { "os": "linux", "cpu": "ppc64" }, "sha512-fsK/sNBnxzBlL4O1JNrZakVQxPspqpED5dLtNsZS9oOKmtSpdNIzxH2kkol5HYTWJN47sE20ztMJPxfZ89qGOg=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-linux-s390x-gnu": ["@rolldown/binding-linux-s390x-gnu@1.1.5", "", { "os": "linux", "cpu": "s390x" }, "sha512-gLYb4BIadlfTOYT5gO503n8zQjXflgzpD0FcyKh0Mzx3rqCZKnHoJWV9xe1KXUJ5lx2JfcSHr/mhzS0PC/McAA=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-linux-x64-gnu": ["@rolldown/binding-linux-x64-gnu@1.1.5", "", { "os": "linux", "cpu": "x64" }, "sha512-FjcpEKUyJygHgs1o50VYNvkt5+7Le/VEdYt0AkRpkL33MnyQfwr8l5mXwMmfmTbyMPr5vJLC+8/Gd9gXnwU1QQ=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-linux-x64-musl": ["@rolldown/binding-linux-x64-musl@1.1.5", "", { "os": "linux", "cpu": "x64" }, "sha512-Me+PfPI2TMeOQk0gYWfLQZtTktrmzbr8cDboqX83XKc7UrgAi55gF+2dUkWdxd19n55Essp2yeca+O9N5rBxHg=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-openharmony-arm64": ["@rolldown/binding-openharmony-arm64@1.1.5", "", { "os": "none", "cpu": "arm64" }, "sha512-yc5WrLzXks6zCQfn9Oxr8pORKyl/pF+QjHmW/Qx3qu0oyrrNC+y2JLTU1E2rcWYAmzlnqngWXHQjy51VzW70Vw=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-win32-arm64-msvc": ["@rolldown/binding-win32-arm64-msvc@1.1.5", "", { "os": "win32", "cpu": "arm64" }, "sha512-gHv82k63z4qpV5+Q1y/12KrK0ltWBukVDI8nZcbT7Tt/ZlOIVwppazneq0F93oDxTo3IgAMEDIoQh3E2n6mVsw=="],
+
+    "@vitest/mocker/vite/rolldown/@rolldown/binding-win32-x64-msvc": ["@rolldown/binding-win32-x64-msvc@1.1.5", "", { "os": "win32", "cpu": "x64" }, "sha512-tTZuDBPw85tEN5PQi1pnEBzDy0Z49HtScLAbD5t6hyeU92A95pRWaSMw1GZZi/RwgSgUIl0xrSlXIT/9QzvYSA=="],
+
     "@vue/compiler-core/@babel/parser/@babel/types/@babel/helper-string-parser": ["@babel/helper-string-parser@7.29.7", "", {}, "sha512-Pb5ijPrZ89GDH8223L4UP8i6QApWxs04RbPQJTeWDV0/keR2E36MeKnyr6LYmUUvqRRI+Iv87SuF1W6ErINzYw=="],
 
     "@vue/compiler-core/@babel/parser/@babel/types/@babel/helper-validator-identifier": ["@babel/helper-validator-identifier@7.29.7", "", {}, "sha512-qehxGkRj55h/ff8EMaJ+cYhyaKlHIxqYDn682wQD7RNp9UujOQsHog2uS0r2vzr4pW+sXf90NeeayjcNaX3fFg=="],
diff --git a/docs/architecture.md b/docs/architecture.md
index bcf23b021..0b34558f1 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -14,6 +14,7 @@ Some routes serve a purpose that raw DB rows cannot satisfy:
 
 - CollectiveX assembles stored sweep documents through its shared reader.
 - `tco-feed` performs server-side frontier interpolation for spreadsheet consumers that cannot run the TypeScript transforms.
+- `server-log-search` searches immutable log bundles in bounded database slices and returns only contextual matches. Returning raw rows would require transferring complete files that can exceed 100 MB before the browser could search them.
 - `/api/v1/overview` is a page-owned backend-for-frontend (BFF). It returns the compact `OverviewPageData` shape used by the initial server render, allowing selector changes to update the matrix without downloading every model's raw benchmark history or triggering a React Server Component (RSC) round trip. It is not a reusable public data API.
 - `/api/v1/benchmarks?view=calculator&sequence=...` is a page-owned compact view. It returns only the selected calculator scenario and the metric keys needed for interpolation, while the default endpoint preserves the raw-row contract. This reduces transfer and browser parsing without narrowing the reusable inference response.
 
@@ -24,13 +25,30 @@ Two rules keep that BFF honest, both enforced in `overview-navigation.tsx`:
 
 Because the selector commit uses `History.prototype.pushState` rather than Next's patched version, `useSearchParams()` and `usePathname()` stay at the load-time URL on `/overview`. Anything that needs the live URL there listens for `CLIENT_SEARCH_CHANGE_EVENT`; the provider emits its own `$pageview`. Do not add a `useSearchParams()` consumer to the overview tree.
 
-## Hash-Based Tab Routing (Not Next.js Routes)
+## Route-Based Dashboard Navigation
 
-Tabs use `window.location.hash` instead of Next.js file-based routing because:
+Every dashboard surface is a real Next.js route. The data-only registry in
+`packages/app/src/lib/dashboard-routes.ts` owns canonical paths, navigation groups,
+indexability, provider capabilities, locale mirroring, and share-parameter scopes.
+`TabNav` renders links from that registry, so browser back/forward uses ordinary
+navigation and a newly added route cannot silently disappear from the sitemap or
+Chinese route map.
 
-- The entire app is a single dashboard page. Separate routes would mean separate page loads, losing React state (zoom positions, filter selections, legend toggles).
-- Hash changes don't trigger Next.js navigation, so context providers stay mounted. This is critical — rebuilding D3 charts from scratch on tab switch would cause visible jank.
-- Browser back/forward still works (hashchange event listener updates tab state).
+Chart state survives navigation through the URL-state contract rather than a
+long-lived single-page tab tree. Each route mounts only the providers it consumes.
+
+### Per-Model Tab Routes
+
+`/calculator/<model>` and `/historical/<model>` (plus `/zh` siblings) give every
+dashboard model an indexable URL. `packages/app/src/lib/model-routes.ts` derives
+the slugs from the compare-page model registry (one slug vocabulary site-wide)
+and stays a thin layer on top of `dashboard-routes.ts`, whose prefix matching
+already resolves the child paths — it is not a second route registry.
+`GlobalFilterProvider` seeds the model from the pathname; switching models on
+the page rewrites the pathname with the pristine `History.prototype.replaceState`
+(same trick as `replaceClientSearch`), so the address bar tracks the model
+without an App Router navigation or remount. The bare tab paths remain the
+canonical home of the default model; alias slugs 308-redirect like `/compare`.
 
 ## URL State Persistence
 
@@ -38,23 +56,28 @@ Chart filter state (model, sequence, metric, precisions, date range, GPU selecti
 
 **Why debounced writes (150ms)?** Rapid filter changes (e.g., clicking multiple precision checkboxes) would spam `history.pushState`. Debouncing batches them into a single URL update.
 
-**Why snapshot-and-clear on load?** Initial params are read into React state, then stripped from the URL via `history.replaceState`. This prevents stale params from accumulating across navigation — the URL always reflects current state, written back by the debounced sync.
+**Why snapshot-and-clear on load?** Initial params are read into the URL-state snapshot, then stripped from the visible address bar with `history.replaceState` while preserving the pathname and hash. Route-owned entry points can also pass typed server seeds. Subsequent debounced writes update the in-memory share state.
+
+**Prefix convention**: `g_` (global), `i_` (inference), `e_` (evaluation), `r_` (reliability), and `c_` (calculator). The canonical dashboard route registry declares which scopes belong in each share URL.
 
-**Prefix convention**: `g_` (global), `i_` (inference), `e_` (evaluation), `r_` (reliability). Prevents namespace collisions and allows `buildShareUrl()` to include only tab-relevant params.
+## Provider Ownership
 
-## Provider Nesting Order
+`DashboardShell` reads provider capabilities from the route registry:
 
 ```
-QueryProvider → ThemeProvider → UnofficialRunProvider → GlobalStateProvider
-  → GlobalFilterProvider → InferenceProvider → EvaluationProvider → ReliabilityProvider
+QueryProvider → ThemeProvider → DashboardShell
+  filtered routes: UnofficialRunProvider → GlobalFilterProvider
+    inference / historical: InferenceProvider
+    evaluation: EvaluationProvider
+  calculator: UnofficialRunProvider → route-owned seeded GlobalFilterProvider
+  reliability: route-owned ReliabilityProvider
+  static / internal routes: no data providers
 ```
 
-This isn't arbitrary. Each provider depends on the one above it:
-
-- `GlobalFilterProvider` needs React Query (`useAvailability()`, `useWorkflowInfo()`)
-- `InferenceProvider` needs global model/date selection; gated by `activeTab` to skip heavy work on non-inference tabs
-- Evaluation and Reliability need the hardware config from Inference context
-- TCO Calculator and Historical Trends reuse InferenceContext state (sequence, precisions) without their own providers — local `useState` is sufficient since they don't share state with other tabs
+This keeps availability, workflow, and unofficial-run requests off pages that do not
+consume them. Compare routes sit outside `DashboardShell` and own their seeded
+`GlobalFilterProvider` / `InferenceProvider` pair. Agentic point-detail routes are also
+standalone because they fetch point-owned data rather than dashboard filters.
 
 ## Client-Side Caching (React Query — In-Memory Only)
 
@@ -71,7 +94,9 @@ API route responses are cached at two layers before hitting the CDN.
 | Local (unstable_cache) | Next.js in-process | ~2 MB default      | Small payloads (availability, workflow-info, etc.) |
 | Blob storage           | Vercel Blob        | No practical limit | Large payloads that exceed the 2 MB threshold      |
 
-`cachedQuery()` in `src/lib/api-cache.ts` wraps both tiers. Pass `{ blobOnly: true }` for payloads known to be large (e.g. `/api/v1/benchmarks`, which returns full benchmark rows for a model). The blob path is `{BLOB_CACHE_PREFIX}/{keyPrefix}:{args}.json`.
+`cachedQuery()` in `src/lib/api-cache.ts` wraps both tiers. Pass `{ blobOnly: true }` for payloads known to be large (e.g. `/api/v1/benchmarks`, which returns full benchmark rows for a model). Blob keys encode canonical typed arguments as `{keyPrefix}:v2:{base64url}`; keys that would exceed the pathname limit use `{keyPrefix}:v2:sha256:{digest}` instead. `blob-cache.ts` stores that key beneath `BLOB_CACHE_PREFIX` with a `.json` suffix.
+
+`cachedDerivedData()` is the tagged, `unstable_cache`-only wrapper for compact server payloads. Overview caches its fully assembled selector response. Compare caches one pair-filtered, full-shape hydration row set per model/GPU pair, then recomputes the inexpensive sequence/precision table transform per request; nesting a selector cache around the pair cache would bypass Next.js incremental caching. Raw public benchmark responses keep their existing Blob-backed contract.
 
 `blobSet()` is no-op if the key already exists, making concurrent lambda invocations race-safe — only the first writer wins, subsequent calls skip silently.
 
@@ -79,7 +104,7 @@ API route responses are cached at two layers before hitting the CDN.
 
 `unstable_cache` entries are tagged `'db'`. Calling `revalidateTag('db', { expire: 0 })` evicts all local cache entries in one call. Blob storage has no built-in tag system, so `blobPurge()` walks the paginated blob list and deletes every object under the prefix.
 
-Both are called together by `purgeAll()`, which also writes a new `cache-version` timestamp to blob storage.
+`purgeAll()` calls both mechanisms and also invalidates the separate CollectiveX tag. Blob keys are versioned and encoded deterministically so structured arguments cannot collide.
 
 ### CDN Cache Headers
 
@@ -94,7 +119,7 @@ Vercel-Cache-Tag: db
 
 ### /api/v1/invalidate
 
-`POST /api/v1/invalidate` requires a `Bearer {INVALIDATE_SECRET}` header (compared with `timingSafeEqual` to prevent timing attacks). On success it calls `purgeAll()` — which clears blob storage, bumps the cache-version timestamp, and revalidates the `'db'` tag — then returns `{ invalidated: true, blobsDeleted: N }`. This endpoint is called by the CI ingest pipelines after benchmark runs and by the run-override pipeline after it applies and verifies merged override changes.
+`POST /api/v1/invalidate` requires a `Bearer {INVALIDATE_SECRET}` header, checked through the shared byte-safe constant-time bearer helper. On success it calls `purgeAll()`, which clears blob storage and revalidates the database and CollectiveX tags, then returns `{ invalidated: true, blobsDeleted: N }`. This endpoint is called by the CI ingest pipelines after benchmark runs and by the run-override pipeline after it applies and verifies merged override changes.
 
 ## React Query Configuration
 
diff --git a/docs/chinese-copy.md b/docs/chinese-copy.md
new file mode 100644
index 000000000..0e18c3331
--- /dev/null
+++ b/docs/chinese-copy.md
@@ -0,0 +1,192 @@
+# Chinese Copy Editorial Guide / 中文文案编辑规范
+
+This guide is the editorial source of truth for user-visible Simplified Chinese in
+InferenceX. It complements the route architecture in [i18n.md](./i18n.md): that document
+explains where Chinese pages live; this one defines how their copy is written and reviewed.
+
+本文是 InferenceX 简体中文用户文案的编辑规范。它与 [i18n.md](./i18n.md) 分工：
+后者说明中文页面的技术架构，本文规定中文文案的写法与审核方式。
+
+## Audience and baseline tone / 读者与基本语气
+
+Write for Chinese-speaking ML infrastructure engineers who already understand models,
+hardware, serving frameworks, and performance metrics. The copy should be precise,
+restrained, and easy to read without consulting the English source.
+
+目标读者是熟悉模型、硬件、推理框架和性能指标的中文 ML 基础设施工程师。文案应当
+准确、克制；即使不对照英文，也能独立读懂。
+
+- Lead with the user action, result, or claim. Do not reproduce English modifier chains.
+- Prefer ordinary technical Chinese over slogans, literal metaphors, and stacked nouns.
+- Preserve facts, qualifications, modality, numbers, units, product names, and attribution.
+- Do not add confidence, causality, or endorsement that is absent from the English source.
+- Keep English when Chinese practitioners normally do: product and model names, hardware SKUs,
+  framework names, flags, identifiers, units, and established English technical terms or phrases.
+  Decide from real industry usage and the surface, not from a closed list.
+
+- 先说用户动作、结果或核心判断，不照搬英文的前置修饰链。
+- 使用工程师日常采用的技术中文，避免口号、直译隐喻和名词堆叠。
+- 保留事实、限定条件、语气强弱、数字、单位、产品名和署名。
+- 英文原文没有表达的确定性、因果关系或背书，不得自行添加。
+- 中文技术社区通常直接使用的英文应保留，包括产品与模型名称、硬件 SKU、框架名、flag、
+  标识符、单位，以及行业中已习惯使用英文的技术术语和短语。应结合真实行业用法和界面场景
+  判断，不能依赖封闭词表。
+
+## Register by surface / 不同界面的语域
+
+| Surface                               | Editorial register                                                                                                                  | 中文要求                                                                 |
+| ------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------ |
+| Buttons, navigation, and short labels | Brief and action-led. Omit subjects that the UI already makes obvious.                                                              | 简短、以动作或对象为中心；界面已经明确的主语不重复。                     |
+| Tooltips and dashboard explanations   | Explain what the metric or control means and what changes when it is used. Prefer one to three direct sentences.                    | 说明指标或控件的含义，以及操作后会发生什么；通常用一至三句直述。         |
+| Empty, loading, and error states      | State the current condition, then the next useful action when one exists. Do not blame the user.                                    | 先说明当前状态，再给出可执行的下一步；不要把问题归咎于用户。             |
+| Metadata                              | Standalone, factual, and search-readable. Do not stuff keywords or repeat the title mechanically.                                   | 脱离页面正文也能读懂；以事实为主，不堆关键词，也不机械重复标题。         |
+| Technical prose                       | Preserve definitions and causal structure, but reorganize clauses for Chinese readability. Define an uncommon acronym on first use. | 保留定义和因果关系，但按中文逻辑重组分句；不常见缩写首次出现时补充说明。 |
+| Landing and marketing copy            | Evidence-led and restrained. Concrete capabilities and provenance are stronger than superlatives.                                   | 以事实和证据为主，克制表达；具体能力和数据来源比夸张形容词更可信。       |
+| Attributed quotations                 | Preserve the speaker's claim, degree of certainty, and voice. Do not polish a quotation into house marketing copy.                  | 保留说话者的观点、语气强弱和个人表达，不得改写成官方营销文案。           |
+
+Second-person wording is contextual, not a mechanical ban. Controls and tooltips normally omit
+the pronoun. A deliberate `您` is acceptable in a direct, respectful call to action when the
+surrounding surface uses that register. Do not replace every `您` with `你`, and do not reject an
+otherwise natural sentence solely because it contains `您`.
+
+第二人称必须结合场景判断，不能全局禁用。控件和提示框通常省略第二人称；直接面向读者、
+且有意采用礼貌语气的 CTA 可以使用 `您`。不要把所有 `您` 机械替换为 `你`，也不要仅因
+一句话包含 `您` 就判定文案不合格。
+
+## Rewrite method / 改写方法
+
+For each changed UI element or paragraph:
+
+1. Read the complete English source and the surrounding rendered UI or paragraph.
+2. Write one plain-language sentence stating the intended meaning. Record facts, scope,
+   modality, metrics, links, and attribution that cannot change.
+3. Hide the old Chinese translation and rewrite from the intended meaning.
+4. Compare the rewrite with the source for semantic fidelity.
+5. Read the Chinese alone for naturalness, register, and information order.
+6. Verify the desktop and mobile rendering when width, hierarchy, or interaction affects copy.
+
+逐个 UI 元素或段落执行：
+
+1. 阅读完整英文原文，并查看对应的页面或段落上下文。
+2. 用一句直白的中文写出原意，并记录不得改变的事实、范围、语气强弱、指标、链接和署名。
+3. 隐藏旧译文，根据原意重新写，不在旧句上逐词替换。
+4. 回到英文原文，检查语义是否完整、准确。
+5. 脱离英文单独阅读中文，检查自然度、语域和信息顺序。
+6. 如果宽度、层级或交互会影响文案，在桌面端和移动端实际页面中检查。
+
+Semantic fidelity and natural Chinese are independent gates. A fluent mistranslation fails the
+first gate; an accurate but translation-shaped sentence fails the second. Both must pass.
+
+语义准确和中文自然度是两道独立门槛。中文流畅但翻错原意，第一关不通过；语义准确但仍是
+英文句式，第二关不通过。两关都通过才算完成。
+
+## Context-aware terminology / 上下文相关的术语表
+
+The table records defaults, not blind substitutions. When an exception is needed, document the
+surface and reason in the PR or add a reviewed example here.
+
+Established English technical terms and phrases that Chinese ML infrastructure engineers normally
+use in English should not be translated merely to increase Chinese-language coverage. This is an
+open, context-dependent category: `warmup`, `seed`, and `offload` are examples, not its boundaries.
+
+下表记录默认写法，不是全局替换表。需要例外时，应在 PR 中说明使用场景与理由，或将经过
+审核的案例补充到本文。
+
+中文 ML 基础设施工程师在行业实践中通常直接使用的英文技术术语和短语，不应为了提高中文
+占比而强行翻译。这是一个需要结合上下文判断的开放类别；`warmup`、`seed`、`offload` 只是
+示例，并非完整清单。
+
+| English                | Preferred rendering | Context and exceptions                                                                                                                                                                                                                                                                                                                              |
+| ---------------------- | ------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| benchmark              | 基准测试            | Use for the process or result set. A short UI label may use “测试” only when the surrounding heading already establishes that it is a benchmark. / 表示测试过程或结果集；只有上下文已经明确是基准测试时，短标签才可简写为“测试”。                                                                                                                   |
+| dashboard              | 仪表板              | Use for the interactive data surface. Do not use “完整仪表板” merely to mirror “full”; write “查看完整仪表板” when it is an action. / 表示交互式数据页面；按钮需要表达动作时写“查看完整仪表板”。                                                                                                                                                    |
+| chart                  | 图表                | Use “图” only when naming a specific conventional form such as a scatter plot. / 特指散点图等固定图形时可用“图”。                                                                                                                                                                                                                                   |
+| configuration / config | 配置                | “测试配置” is preferred when the surrounding text could otherwise refer to deployment configuration. / 容易与部署配置混淆时写“测试配置”。                                                                                                                                                                                                           |
+| inference              | 推理                | Keep inside established product names and identifiers. / 产品名和标识符中的英文保持不变。                                                                                                                                                                                                                                                           |
+| throughput             | 吞吐量              | Preserve the token type, denominator, and scope. / 必须保留 token 类型、分母和统计范围。                                                                                                                                                                                                                                                            |
+| latency                | 延迟                | Preserve percentile and stage qualifiers such as TTFT or TPOT. / 必须保留分位数以及 TTFT、TPOT 等阶段限定。                                                                                                                                                                                                                                         |
+| evaluation / eval      | 评估                | Keep `eval` in code identifiers, flags, and filenames. / 代码标识符、flag 和文件名中的 `eval` 不翻译。                                                                                                                                                                                                                                              |
+| artifact               | 产物                | Use for CI or benchmark outputs. “文件” is acceptable only when the object is literally a user-facing file. / 表示 CI 或基准测试输出；确实指用户下载的文件时可写“文件”。                                                                                                                                                                            |
+| warmup                 | `warmup`            | Keep English in first-party ML infrastructure UI and technical prose. Chinese-first `预热（warmup）` is acceptable when introducing the concept to a broader audience. Never write the redundant `warmup 预热`. / 第一方 ML 基础设施 UI 与技术文案默认保留英文；面向更广泛读者首次解释概念时可写 `预热（warmup）`，不得使用重复表达 `warmup 预热`。 |
+| seed                   | `seed`              | Keep English in technical copy and preserve flags such as `--random-seed`. Chinese-first `随机种子（seed）` is acceptable in explanatory copy. / 技术文案默认保留英文，`--random-seed` 等 flag 保持不变；解释性文案可写 `随机种子（seed）`。                                                                                                        |
+| offload                | `offload`           | Keep English in first-party ML infrastructure UI and technical prose. Chinese-first `卸载（offload）` is acceptable in explanatory copy; attributed quotations retain the speaker's wording. / 第一方 ML 基础设施 UI 与技术文案默认保留英文；解释性文案可写 `卸载（offload）`，署名引用保留说话者原有措辞。                                         |
+| chip                   | 芯片                | Keep English inside units and identifiers such as `tok/s/chip`, `$/chip-hour`, and `ChipSKU`. / `tok/s/chip`、`$/chip-hour`、`ChipSKU` 等单位和标识符保持英文。                                                                                                                                                                                     |
+| framework              | 框架                | Keep names such as vLLM, SGLang, and TensorRT-LLM unchanged. / vLLM、SGLang、TensorRT-LLM 等名称不翻译。                                                                                                                                                                                                                                            |
+| agent / agentic        | 智能体 / 智能体相关 | Keep `AgentX` as a product name. Translate according to the role, not by replacing every occurrence mechanically. / `AgentX` 作为产品名保持不变；其他场景根据实际角色翻译，不做全局替换。                                                                                                                                                           |
+| workload               | 工作负载            | Avoid using “场景” when the source refers to the actual request distribution or task mix. / 原文指请求分布或任务组合时，不要弱化成“场景”。                                                                                                                                                                                                          |
+| disaggregated serving  | 分离式推理          | Use “分离式部署” when the sentence is about architecture or deployment rather than the serving process. / 强调架构或部署方式时可写“分离式部署”。                                                                                                                                                                                                    |
+| speculative decoding   | 投机解码            | Do not rewrite the ordinary verb “推测”. / 不影响作为普通动词使用的“推测”。                                                                                                                                                                                                                                                                         |
+| recipe                 | 测试配置 / 配置方案 | Choose from context; do not leave `recipe` in user-facing prose merely because it appears in code. / 根据上下文选择，不因代码中使用 `recipe` 就在用户文案中保留英文。                                                                                                                                                                               |
+| methodology            | 测试方法 / 方法说明 | Use “方法论” only when the content actually discusses a methodology as a field or system of principles. / 只有内容确实讨论一套方法论或原则体系时才使用“方法论”。                                                                                                                                                                                    |
+| industry testimonials  | 业界评价            | Use for navigation and section labels. Preserve the speaker's own wording inside each attributed quotation. / 用于导航和栏目名称；每条署名引用仍需保留说话者自己的表达。                                                                                                                                                                            |
+| Land Acknowledgement   | 原住民传统领地声明  | Use for the dedicated legal-cultural page and its link. Preserve community, nation, and place names exactly. / 用于对应的法律与文化说明页面及其链接；原住民族群、社群和地名必须准确保留。                                                                                                                                                           |
+
+## Review workflow / 审核流程
+
+1. The author supplies the English source, intended meaning, and rendered context for every
+   changed Chinese passage.
+2. Claude loads `review-zh-copy`, evaluates semantic fidelity and natural Chinese separately,
+   and reports complete suggested replacements rather than isolated word swaps.
+3. The skill's findings never block another contributor's merge. Claude must not label a
+   Chinese-copy finding `BLOCKING` or submit a request-changes review for it. Clean reviews,
+   routine coverage, and ordinary wording or naturalness suggestions do not mention the Chinese
+   maintainer. Only a high-confidence semantic or factual error, changed attribution or speaker
+   voice, or unresolved high-impact ambiguity is labeled `Needs Chinese maintainer confirmation`
+   and mentions `@edwingao28`.
+4. The Chinese-speaking maintainer manually reads every changed sentence and makes the final
+   editorial decision. Attributed quotations receive an explicit attribution-and-voice check.
+
+中文步骤：
+
+1. 作者为每段改动的中文提供英文原文、核心意思和实际页面上下文。
+2. Claude 加载 `review-zh-copy`，分别检查语义准确度和中文自然度，并给出完整改写，
+   而不是只替换孤立词语。
+3. skill 的 finding 不阻断其他贡献者合并。Claude 不得把中文文案 finding 标为
+   `BLOCKING`，也不得因此提交 request-changes review。审核通过、例行 coverage 和普通措辞
+   或自然度建议都不提及中文维护者。只有高置信度的语义或事实错误、署名或说话者语气被
+   改变，或无法消除且影响较大的歧义，才标为 `Needs Chinese maintainer confirmation`，并
+   `@edwingao28` 请中文维护者审核。
+4. 中文维护者逐句人工审阅并作出最终编辑决定；署名引用还要单独核对署名、观点和说话者
+   语气。
+
+## Pull request checklist / PR 审核清单
+
+- [ ] Every changed Chinese passage has an identified English source and intended meaning.
+      中文：每段改动的中文都有对应英文来源和明确原意。
+- [ ] Facts, numbers, units, identifiers, links, modality, and attribution are unchanged.
+      中文：事实、数字、单位、标识符、链接、语气强弱和署名均未改变。
+- [ ] The Chinese reads naturally without preserving English word order or noun stacking.
+      中文：中文脱离英文后仍自然通顺，没有照搬英文语序或名词堆叠。
+- [ ] The register matches the surface.
+      中文：文案语域与所在界面一致。
+- [ ] English UI strings remain byte-identical in a Chinese-only rewrite.
+      中文：仅修改中文时，英文界面字符串保持逐字节不变。
+- [ ] The complete UI element or paragraph was reviewed, not only the changed word.
+      中文：审核对象是完整 UI 元素或段落，而不是孤立的替换词。
+- [ ] Visible copy was checked in rendered desktop and mobile context.
+      中文：用户可见文案已在桌面端和移动端实际页面中检查。
+- [ ] The Chinese-speaking maintainer manually reviewed every changed sentence.
+      中文：中文维护者已逐句人工审核所有改动。
+
+## Relationship to automation / 与自动化检查的关系
+
+Deterministic CI may enforce only context-independent rules such as punctuation, protected
+identifiers, dictionary completeness, and terminology with a genuinely universal rendering.
+It must not decide fluency, sentence structure, register, or contextual pronoun choice.
+
+确定性 CI 只适合检查与上下文无关的规则，例如标点、受保护标识符、字典完整性，以及确实
+只有一种正确写法的术语。流畅度、句法、语域和第二人称选择不得交给机械规则判定。
+
+The `review-zh-copy` skill assists the judgment layer and stays non-blocking. It mentions
+`@edwingao28` only for high-confidence semantic or factual risk, attribution risk, or unresolved
+high-impact ambiguity; ordinary findings do not CC the Chinese maintainer. CI does not globally ban
+`预热`, `随机种子`, or `卸载`, because explanations and quotations can use them correctly. It may
+reject context-independent duplication such as `warmup 预热`. Separately verified mechanical cases
+from #819 may enter deterministic CI fixtures in #820, but editorial rewrites are never automatic
+ground truth.
+
+`review-zh-copy` skill 用于辅助判断性审核，并保持非阻断模式。只有高置信度的语义或事实
+风险、署名风险，或无法消除且影响较大的歧义才 `@edwingao28`；普通 finding 不 CC 中文
+维护者。CI 不全局禁用 `预热`、`随机种子` 或 `卸载`，因为这些写法在解释与引用中可能
+正确；`warmup 预热` 这类与上下文无关的重复表达则可由 CI 拦截。#819 中经过单独验证的
+机械案例可以进入 #820 的确定性 CI fixtures，但编辑性改写不得自动视为标准答案。
diff --git a/docs/collectivex.md b/docs/collectivex.md
index 973694f49..8ac376e50 100644
--- a/docs/collectivex.md
+++ b/docs/collectivex.md
@@ -44,6 +44,11 @@ Key invariants:
   `ON CONFLICT (run_id) DO NOTHING`; concurrent first-viewers can't double-ingest or
   expose a partial run. A GitHub re-run (newer `run_attempt`) is replaced through a
   `FOR UPDATE`-guarded refresh statement.
+- **Reads after lazy persistence use the primary**: discovery state checks, writes, and
+  each route's post-discovery query use `DATABASE_COLLECTIVEX_WRITE_URL`. This guarantees
+  that a newly inserted or refreshed run is visible in the same request even when
+  `DATABASE_COLLECTIVEX_READONLY_URL` points at a lagging replica. Ordinary read-only
+  consumers can still use the read URL.
 - **Deletion tombstones** (`cx_runs.deleted_at`, documents freed): discovery must never
   resurrect a deleted run. Re-ingesting via the CLI
   (`bun run admin:db:ingest:collectivex <run-url-or-id>`) clears the tombstone — that CLI is
@@ -65,11 +70,12 @@ Key invariants:
   `s-maxage` (freshness bound for lazy discovery). Run deletion and
   `POST /api/v1/invalidate?scope=collectivex` purge only that tag; the main dashboard's
   blob cache is untouched by CollectiveX operations.
-- **Env**: `DATABASE_COLLECTIVEX_READONLY_URL` (must be the same primary as the write URL
-  — the routes read their own writes), `DATABASE_COLLECTIVEX_WRITE_URL` (direct/unpooled;
-  also used by migrations via `bun run admin:db:migrate:collectivex`),
-  `COLLECTIVEX_ADMIN_SECRET` (delete route Bearer token — deliberately not
-  INVALIDATE_SECRET, since it is remembered in browser localStorage), and `GITHUB_TOKEN`.
+- **Env**: `DATABASE_COLLECTIVEX_READONLY_URL` is the ordinary read-only connection and
+  may point at a replica. `DATABASE_COLLECTIVEX_WRITE_URL` is the primary used by lazy
+  persistence, its consistent reads, deletion, and migrations via
+  `bun run admin:db:migrate:collectivex`. `COLLECTIVEX_ADMIN_SECRET` is the delete route
+  Bearer token (deliberately not `INVALIDATE_SECRET`, since it is remembered in browser
+  localStorage), and `GITHUB_TOKEN` authorizes artifact reads.
 
 ## Multi-run explorer
 
@@ -89,6 +95,8 @@ visible legend labels while retaining them in the legend item's accessible title
 The newest run with measured cases is checked by default; newer incomplete sweeps remain listed but
 cannot blank the initial explorer. Deletion is available per row or as one confirmed action for all
 currently shown runs; both paths keep the same tombstone semantics described above.
+The Runs card's suite filter narrows the table to runs containing EP or KV cases; a run containing
+both appears under both filters, and changing the filter does not alter the checked-run selection.
 
 ## The raw-rows exception
 
diff --git a/docs/d3-charts.md b/docs/d3-charts.md
index 4dc1a4bac..d2a63ae5a 100644
--- a/docs/d3-charts.md
+++ b/docs/d3-charts.md
@@ -1,17 +1,21 @@
 # D3 Chart Design Rationale
 
-## Why 4 Effects in ScatterGraph
+## Four Invalidation Phases
 
-A single useEffect for all D3 operations takes ~500ms (React reconciliation ~200ms + full SVG rebuild ~300ms). Splitting into 4 effects makes the most common interaction — switching Y-axis metric — take ~100ms instead:
+One effect for every D3 operation takes roughly 500ms (React reconciliation ~200ms plus a full SVG update ~300ms). `D3Chart` separates structure, stable-identity data joins, metric coordinates, and display decoration. The most common interaction, switching the Y-axis metric, then takes roughly 100ms:
 
 | Effect            | Trigger                | Cost   | What it does                                    |
 | ----------------- | ---------------------- | ------ | ----------------------------------------------- |
 | 1. Structure      | Mount, resize, theme   | ~5ms   | SVG skeleton, axes groups, defs, clip paths     |
 | 2. Data render    | Data shape changes     | ~500ms | Full D3 join, bindpoints, rooflines, zoom setup |
 | 3. Metric update  | Metric/scale selection | ~100ms | Reposition points in-place, rebuild rooflines   |
-| 4. Display toggle | Legend/label toggles   | <20ms  | Opacity transitions only                        |
+| 4. Display toggle | Legend/label toggles   | <20ms  | Restyle marks or redraw changed decorations     |
 
-**The key insight**: Effect 2 depends on `dataIdentity` (a stable string hash of point keys), NOT on the `data` array reference. When users switch metrics, data points get new y-values (new array reference), but `dataIdentity` stays the same because the set of points hasn't changed. So Effect 2 skips entirely, and only Effect 3 runs.
+**The key insight**: The data phase depends on `dataIdentity`, a stable string built from point join keys, rather than the `data` array reference. Metric changes produce new `x`/`y` values and a new array, but the set of points stays the same. The metric phase mutates bound coordinates by key and skips enter/update/exit joins.
+
+Custom layers that own scale-neutral decorations declare a `displayIdentity` and an `onDisplayUpdate` callback. The renderer snapshots each identity after the data and metric phases, then reruns only the custom layers whose identities changed. This removes stale labels and decorative overlays without rebuilding joins, scales, or unrelated paths. Callbacks receive the current zoomed scales so toggles do not jump decorations back to the unzoomed frame.
+
+`ScatterGraph` further partitions Effect 4 because its controls have disjoint mutation scopes. Mark visibility, palette/shape, trace metadata, point-label visibility/collision, line-label placement, and known-issue annotations each have a dedicated layout effect. It deliberately omits the generic scatter `displayIdentity`: point-label toggles update only `.point-label` nodes, while the unofficial overlay layer's own display callback updates only `.overlay-label` nodes. Recombining these passes causes label toggles to restamp thousands of unchanged shape, visibility, and trace attributes.
 
 ## In-Place Y-Value Mutation
 
@@ -60,14 +64,14 @@ HTML strings bypass React entirely. The tooltip div is a plain DOM element appen
 
 Different metrics need different "optimal" directions:
 
-| Direction   | "Best" means   | Example metric                                            |
-| ----------- | -------------- | --------------------------------------------------------- |
-| upper_right | High x, high y | (not currently used)                                      |
-| upper_left  | Low x, high y  | Interactivity chart: low latency (x), high throughput (y) |
-| lower_right | High x, low y  | Cost chart: high interactivity (x), low cost (y)          |
-| lower_left  | Low x, low y   | (not currently used)                                      |
+| Direction   | "Best" means   | Example metric                                                        |
+| ----------- | -------------- | --------------------------------------------------------------------- |
+| upper_right | High x, high y | (not currently used)                                                  |
+| upper_left  | Low x, high y  | Interactivity chart: low latency (x), high throughput (y)             |
+| lower_right | High x, low y  | Cost/energy chart: high interactivity (x), low $/M tok or J/token (y) |
+| lower_left  | Low x, low y   | (not currently used)                                                  |
 
-The direction is declared per-metric in `inference-chart-config.json`, not computed. This makes the roofline direction a data concern, not a rendering concern.
+The metric registry declares whether higher or lower values are preferable. Chart definitions derive the concrete corner from that polarity and the chart's x-axis direction, which keeps the choice in data rather than rendering code.
 
 ## Gradient Roofline Labels
 
@@ -114,6 +118,11 @@ D3 bar charts measure actual Y-axis label widths using a temporary SVG `<text>`
 
 Opacity animates via inline CSS `transition: opacity 150ms ease` (set on dots, rooflines, and labels in the render path); d3 `.transition()` is reserved for attributes CSS can't animate — the `data-update` entrance transitions on dot `transform` and roofline `d`. Never point both systems at the same property: a d3 transition re-writes the style every animation frame, and each write restarts the CSS transition, emitting `transitionrun`/`transitioncancel` per node per frame (a legend hover across a full chart used to produce tens of thousands of events per session, all of it also observed by PostHog's session-replay MutationObserver). Handlers like legend hover therefore write opacity **once** and let CSS do the animation.
 
+Official coordinate transitions use one 300ms D3 tween for visible points and rooflines. Hidden
+official marks snap directly to their final geometry and are excluded from the transition snapshot,
+avoiding frame-by-frame writes for data the user cannot see. Unofficial overlays retain their
+existing snap behavior. Replay passes a zero duration and remains frame-driven.
+
 ## Batched Label Measurement
 
 Label loops that size a background rect to its text (`.ll-bg`/`.pl-bg`, point-label collision avoidance) must not interleave `getBBox()` with DOM writes — each read after a write forces a synchronous layout, turning N labels into N reflows. The pattern is two passes over the selection: write every label's text first, then measure every bbox (one forced layout for the whole batch), then write the rects. Same rule for `measureLegendRightInset`: it reads `getBoundingClientRect`, so it's skipped entirely when there are no known-issue annotations to place — it would otherwise run on every zoom frame.
diff --git a/docs/data-pipeline.md b/docs/data-pipeline.md
index 9d1ce61cc..176acc32c 100644
--- a/docs/data-pipeline.md
+++ b/docs/data-pipeline.md
@@ -2,24 +2,38 @@
 
 ## DB Schema Decisions
 
-### Why JSONB + Hot Columns (Hybrid)
-
-Benchmark metrics are stored in a JSONB `metrics` column AND extracted into dedicated "hot" columns (`tput_per_gpu`, `median_intvty`, `median_ttft`, `median_e2el`, `p99_ttft`, `median_tpot`).
-
-- **JSONB**: New metrics can be added by CI without schema migrations. Old data doesn't need backfilling — missing fields default to 0 at read time (`m.field ?? 0`).
-- **Hot columns**: The most-queried metrics need B-tree indexes for `DISTINCT ON` queries. JSONB extraction (`metrics->>'field'`) can't use these indexes efficiently.
-- **Trade-off**: ~6 duplicated values per row. Acceptable because benchmark_results is write-once/read-many, and the index speedup on daily queries is orders of magnitude.
+### Why Metrics Stay in JSONB
+
+Benchmark measurements live only in the `benchmark_results.metrics` JSONB column;
+there are no duplicated metric-specific columns on `benchmark_results` or
+`latest_benchmarks`.
+
+- **Schema flexibility**: CI can add metrics without a database migration, and
+  historical rows simply omit keys that their artifact did not produce.
+- **Read contract**: Raw benchmark endpoints pass the JSONB object through.
+  Specialized readers such as MCP select keys explicitly. A missing historical
+  metric remains absent until a presentation consumer applies its own fallback.
+- **Indexes**: The benchmark indexes cover relational selection dimensions such
+  as config, scenario, concurrency, sequence lengths, date, offload mode, and
+  recipe fingerprint. Metric ranking and extraction use JSONB expressions rather
+  than nonexistent "hot" columns.
 
 ### Why Denormalized Dates
 
 `benchmark_results.date`, `workflow_runs.date`, and the `availability` table all store denormalized date values (derived from `workflow_runs.created_at`). This avoids JOINs in the hottest queries:
 
-- `getLatestBenchmarks` uses `DISTINCT ON (config_id, conc, isl, osl) ORDER BY date DESC` — needs date on the same table as benchmark_results for the covering index.
+- `getLatestBenchmarks` orders candidate line snapshots by `date`, `run_started_at`,
+  and workflow-run ID before applying the append-only curve rules. Keeping `date`
+  on each benchmark row avoids another join in that base-table scan.
 - `availability` is a separate denormalized table because the date-picker query (`SELECT DISTINCT model, date`) needs to be fast without scanning benchmark_results.
 
 ### Why a Materialized View (latest_benchmarks)
 
-The `DISTINCT ON` query for "latest benchmark per config" is expensive on the full table (millions of rows). The materialized view pre-computes this, refreshed concurrently after each ingest. API routes use the view when no date filter is specified; date-filtered requests hit the base table.
+Computing the newest logical curve from the full result table is expensive. The
+materialized view pre-computes the newest successful snapshot for each config,
+scenario, sequence, offload mode, recipe fingerprint, and concurrency, including
+the same-image append-only history when applicable. API routes use the view when no
+date filter is specified; date-filtered requests hit the base table.
 
 `REFRESH CONCURRENTLY` allows reads during refresh (no downtime). The trade-off is a brief window where the view is stale after ingest — acceptable since data changes at most daily.
 
@@ -31,16 +45,40 @@ Every INSERT uses `ON CONFLICT DO UPDATE` or `DO NOTHING`. This means:
 - **Partial failures recover**: If ingest crashes mid-batch, re-running picks up where it left off.
 - **No cleanup needed**: No "delete old data first" step that could leave the DB empty on failure.
 
-The unique constraints match natural keys (e.g., `(workflow_run_id, config_id, isl, osl, conc)` for benchmarks), not surrogate keys.
+The unique constraints match natural keys (for benchmarks, workflow/config/scenario,
+concurrency, offload mode, and the nullable recipe fingerprint), not surrogate keys.
+
+### Append-Only Curve Extensions
+
+Normal workflow runs are complete line snapshots: the latest run for a line replaces
+the prior run as a unit, so partial re-sweeps cannot silently stitch points from
+different recipes. A changelog containing only `append-only: true` entries marks the
+one narrow exception. The latest-curve queries then walk backward through consecutive
+append-only runs and include the nearest full snapshot, selecting the newest producer
+for each recipe and concurrency. The producer stamps every new point with a deterministic
+`recipe_fingerprint`, so variants that share the app's normalized topology and concurrency
+remain separate. Historical rows have a null fingerprint and retain their legacy identity.
+
+The chain continues only while the image is identical. Each returned benchmark row
+keeps its original workflow-run ID and run URL, so extending a curve does not erase
+point provenance; `curve_date` and `curve_workflow_run_id` carry the separate logical
+snapshot identity used by charts and history. InferenceX CI separately verifies that
+the generated matrix is strictly additive: existing recipes and points are immutable,
+while newly added concurrency points or complete recipe variants may run as deltas.
+All appended points must retain the target curve's image, and benchmark-affecting code
+changes must be isolated to the new points.
 
 ### Audited Point Purges
 
 `packages/db/src/etl/run-overrides.ts` is the durable audit record for exceptional
 data removal. Use `PURGED_BENCHMARK_POINTS` when a valid workflow run contains a
 specific invalid result, such as a server hang. Each record names `githubRunId`,
-`runAttempt`, `configId`, `benchmarkType`, `isl`, `osl`, `conc`, and `offloadMode`,
-with a dated reason comment. The full identity is required because one run can
-contain multiple serving configurations at the same sequence lengths and concurrency.
+`runAttempt`, `configId`, `benchmarkType`, `isl`, `osl`, `conc`, `offloadMode`, and
+the optional nullable `recipeFingerprint`, with a dated reason comment. A fingerprint
+targets exactly that recipe; omitting it or setting it to null targets only legacy rows
+whose stored fingerprint is null. The full identity is required because one run can
+contain multiple serving configurations and recipes at the same sequence lengths and
+concurrency.
 
 `bun run db:apply-overrides` previews every matching row and requires confirmation
 before deleting it in a transaction. It also removes unreferenced server logs and
@@ -54,14 +92,59 @@ conservative fallback prevents a failed GitHub lookup from restoring a suppresse
 point. It can only suppress an identical point from another attempt while that
 attempt remains unknown.
 
-### Why Two Connection Types
-
-| Connection                      | Library     | Use Case                             | Why                                                                                                  |
-| ------------------------------- | ----------- | ------------------------------------ | ---------------------------------------------------------------------------------------------------- |
-| `@neondatabase/serverless` HTTP | API routes  | Stateless, read-only, scales to zero | Serverless functions can't hold persistent connections; HTTP driver works over Vercel's edge network |
-| `postgres` TCP                  | ETL scripts | Bulk inserts, transactions, COPY     | HTTP driver has per-query overhead that's unacceptable for 10K+ row batches                          |
-
-The read replica (`DATABASE_READONLY_URL`) is used by API routes to isolate read traffic from write load. ETL uses the primary writer (`DATABASE_WRITE_URL`).
+### Audited Run Backfills
+
+Exceptional metadata corrections use the same reviewed, automatically enforced ledger
+as purges. Add a `CHANGELOG_BACKFILLS` or `BENCHMARK_POINT_BACKFILLS` entry in
+`packages/db/src/etl/run-overrides.ts`; never add a one-off production SQL statement.
+Every entry has a stable kebab-case `id`, a human-readable `reason`, an exact run attempt,
+and the target table's complete natural-key selector. The source-control history records
+who approved the correction and why, while the stable ID appears in ingest and apply logs.
+
+Changelog backfills select `(githubRunId, runAttempt, baseRef, headRef)` and can patch
+`configKeys`, `description`, `prLink`, or `appendOnly`. An `appendOnly` correction updates
+both `changelog_entries.append_only` and `workflow_runs.append_only` atomically because
+the materialized benchmark view reads the run-level value.
+
+Benchmark point backfills use the same complete point selector as audited point purges:
+`githubRunId`, `runAttempt`, `configId`, `benchmarkType`, `isl`, `osl`, `conc`,
+`offloadMode`, and nullable `recipeFingerprint`. Their `set` block supports a shallow
+`metricsMerge`, top-level `metricsRemove`, and `offloadMode`. Setting `offloadMode`
+updates both the first-class column and `metrics.offload_mode`; for example, a missed CPU
+KV offload annotation can set `offloadMode: 'on'` and merge `kv_offloading` plus backend
+metadata in one correction. Unrelated metrics remain unchanged.
+
+Run `bun run admin:db:apply-overrides` to preview the exact rows, reasons, and patches;
+the command requires confirmation unless passed `--yes`. It fails before writing when a
+target is missing, when a point's desired identity already exists, or when registry
+validation finds an overlap or duplicate. Writes are verified against the declared state
+and `latest_benchmarks` is refreshed afterward. Merges to `master` run the same command in
+CI, followed by database verification, cache invalidation, and cache warmup.
+
+Point corrections are additionally applied before each CI or GCS benchmark insert. This
+prevents a later idempotent re-ingest from restoring the artifact's original value. The
+ingest tracks source and desired identities across the run and fails on a collision instead
+of silently collapsing two distinct artifact points. When GCS cannot recover an attempt
+number, it follows the purge path's conservative fallback and matches the exact run and
+point across attempts; the log records the applied backfill ID. Changelog corrections are
+likewise applied to artifact metadata before `appendOnly`/`evalsOnly` run semantics are
+resolved and before the changelog is upserted.
+
+### Connection drivers and policy
+
+| Connection                      | Library                  | Use case                             |
+| ------------------------------- | ------------------------ | ------------------------------------ |
+| `@neondatabase/serverless` HTTP | Application API reads    | Stateless serverless queries         |
+| `postgres` TCP                  | ETL, migrations, and MCP | Bulk work, transactions, and MCP SQL |
+
+`resolveDatabaseConnection` in `packages/db/src/connection.ts` is the single URL and TLS
+policy. Each caller selects its environment variable and driver deliberately. An explicit
+URL can override the environment value, `DATABASE_SSL` can override host detection, and
+loopback hosts disable TLS by default. Application reads may select Neon HTTP, while admin
+and MCP clients remain on postgres.js.
+
+Normal API reads use `DATABASE_READONLY_URL` so a replica can isolate read traffic from
+writes. ETL uses the primary writer in `DATABASE_WRITE_URL`.
 
 ### Why CHECK Constraints for Lowercase
 
@@ -73,7 +156,7 @@ All text keys (model, hardware, framework, precision) have `CHECK (field = lower
 
 Phase 1 (parallel 20): ZIP reading + JSON parsing + row mapping. IO-bound (network + disk), so high parallelism.
 
-Phase 2 (parallel 5): DB writes. Connection-limited (max 20 connections), and each write does config lookup + bulk insert. Lower parallelism prevents connection exhaustion.
+Phase 2 (parallel 10): DB writes. Connection-limited (max 20 connections), and each write does config lookup + bulk insert. Half-pool concurrency leaves capacity for metadata and summary queries while improving ingest throughput.
 
 ### Config Cache
 
@@ -89,12 +172,105 @@ AIPerf defines the `server_metrics_export.json` envelope, but labels such as wor
 
 Adapters are selected from the benchmark's canonical framework, and per-worker series are only emitted for disaggregated configs with a recognized adapter. Unknown orchestrators and non-disaggregated configs retain their aggregate-only series; roles are never guessed from ports or metric names. The frontend only consumes the canonical source identity and never interprets orchestrator-native labels.
 
+### Logical Engines vs Raw Series
+
+A raw series in the blob is one `(scrape endpoint × phase block × label set)` tuple, which is **not** the same as one engine. The KV-cache chart needs one entry per _logical engine_ — one KV pool — so `compute-chart-series.ts` groups series by their Prometheus label set (`seriesIdentityKey`) rather than emitting one entry per raw series. Three kinds of duplication collapse there:
+
+- **Phase blocks.** The warmup and profiling blocks each carry their own series for the same engine, so keying by scrape instant unions them into one continuous line. The two blocks' first/last bounds can look overlapping — the profiling series often emits a single boundary sample and then gaps until warmup ends — but they never share a scrape instant, so the union neither drops nor double-counts a sample.
+- **Mirrored API-server frontends.** vLLM run with several API servers exposes the _same_ engine set on every `/metrics` endpoint, a few hundred ms apart. Identity ignores `endpoint_url` (Prometheus treats the label set as the series), and the endpoint covering the most wall-clock wins — merging the mirrors instead would interleave near-duplicate samples and halve the effective span of the frontend's fixed-width rolling average. Endpoints are only fused when their values actually agree; same-label endpoints whose readings diverge are independent replicas behind a router, and are kept as separate engines so none is silently discarded.
+- **Intra-engine shard ranks.** `tp_rank` / `pp_rank` / `ep_rank` / `moe_ep_rank` shard one pool and all report the same utilization, so they are excluded from the identity and averaged. `engine` / `engine_idx` / `dp_rank` are _not_ excluded — those do name distinct pools.
+
+The cluster average is then a mean across those logical engines on the union of their scrape instants, with each engine holding its last sample until its next one and contributing only inside its own observed window (and only while that sample is fresher than 5× the engine's own median scrape gap, so a reporting hole drops the engine out of the mean rather than pinning it to a stale reading). Grouping on an exact `start_ns` instead would average whichever engines happened to share that nanosecond — on a disaggregated run that alternates between "prefill only" and "decode only" and reads as a full-scale sawtooth.
+
+Engines are ordered by role, then numeric rank, then worker — never by the composed display string, which would sort `"decode 10"` before `"decode 2"` and scramble DP ranks. The role is only shown when engines actually differ in role, so an aggregated deployment reads `DP 0…DP 3` rather than `decode 0…decode 3`.
+
+### Summed Series and the Canonical Grid
+
+The same "components aren't scraped in lockstep" problem hits the **summed** series — prefill/decode/prefix-hit rates, queue depth, host KV usage, and the prompt-token source breakdown — but it cannot be solved the same way. A mean is scale-free, so `averageAcrossEngines` can evaluate on the union of scrape instants; a sum is not. `cumulativeUniqueInputTokens` turns these rates into token totals with `sum += value` and `rollingAverage` is a sample-count mean, so both only stay correct at **one point per scrape bucket**.
+
+Summing on an exact `start_ns` emitted one point per component per tick, each holding that component's share alone. Downstream that reads as a comb: the rolling average mixed the real samples with the other components' structural gaps and drew roughly 1/N of the cluster total. Two shapes produced it in the corpus — a disaggregated run puts every worker on its own `/metrics` endpoint and its own sub-second offset (~7× low on the 7-worker rows), and even a single-endpoint SGLang run splits its token counters into `is_streaming="true"`/`"false"` series ~16 ms apart, where the `"false"` half is an all-zero run for a fully streaming benchmark (2.2× low).
+
+Every summed series is therefore evaluated by `sumOntoGrid` on one canonical grid at the blob's native scrape cadence (`canonicalTickNs`, the median gap of the best-sampled series), with each component holding its last sample between its own scrapes and contributing only inside its observed window. The lattice is anchored at `t=0` and shared by every metric, so the pairs that get divided or added downstream (hits/queries, used/total, running+waiting) land on identical `t` values — anchoring each metric at its own first sample instead silently emptied the prefix-cache-hit-rate chart, because `sglang:cached_tokens` starts ~0.18 s off the grid `sglang:prompt_tokens` starts on.
+
+Mirrored endpoints are collapsed here too, on a **relative** tolerance rather than the gauge path's absolute one — a throughput mean is O(10⁵), so an absolute threshold would never fire and every mirror would be counted twice. This is a correction as well as a de-duplication: the two-API-server vLLM rows were double-counting their queue depth and token totals exactly 2× before v14.
+
+Nothing groups on an exact `start_ns` any more; `aggregateByStart` was removed in v14.
+
+The detail page does not average the resulting per-tick cache-hit ratios. Cache-hit and
+prompt-token counters can publish logically related deltas in adjacent scrape buckets, so a
+quiet denominator bucket can make a pointwise ratio exceed 100% even though the run-wide totals
+are valid. The displayed line is instead a centered 50-sample ratio of sums —
+`Σ(cache-hit rate) / Σ(cache-query rate)` — computed in O(n) with prefix sums. The query-rate
+denominator is recovered from the stored pointwise ratio and hit rate, preserving vLLM's
+`prefix_cache_queries` semantics; prompt-token rate is only a proxy weight for zero-hit intervals
+where `hits / ratio` cannot be inverted. Hit-counter ticks with no stored ratio (zero queries) are
+also unioned back into the window rather than dropped. Only after that volume-weighted
+aggregation is the semantic `[0, 1]` bound applied as a guard against residual counter timing
+skew. Older rows without the component rate series retain the stored-ratio fallback.
+
+### Server-Log Artifact Bundles
+
+Server-log artifacts are stored as filename-keyed bundles rather than as a hard-coded list of
+router, worker, or benchmark roles. Ingest recursively retains every regular `.log` and `.out`
+file and preserves its artifact-relative path. This applies to both `server_logs_*` ZIPs and
+`multinode_server_logs_*` artifacts whose files are nested inside
+`multinode_server_logs.tar.gz`.
+
+The primary file remains in `server_logs.server_log` for compatibility and KV-cache metadata
+extraction; `server_log_files` holds the remaining files. `server_logs.file_name` records the
+primary file's original path, and `files_complete` distinguishes fully scanned bundles from
+legacy single-file rows. Run `bun run admin:db:backfill-server-log-files` to use the public GCS
+artifact backup first and fall back to retained GitHub artifacts for recent runs. Pass `--all`
+for complete DB history, `--source gcs` to exercise only the backup path, and `--dry-run` to
+inventory pair counts and compressed download size without writing. `--run <id>`, `--limit <n>`,
+and `--yes` remain available for targeted, idempotent recovery. Bundles already marked
+`files_complete` are skipped after their small benchmark artifact is mapped, avoiding repeated
+large log downloads.
+
+Full-bundle search stays inside PostgreSQL: each file is scanned through overlapping 16 MiB
+character slices, and the API returns at most 50 match contexts rather than transferring the
+complete file. The overlap preserves literal matches that cross a slice boundary. Files are
+processed sequentially so a multinode bundle cannot materialize every large log in one query.
+
 ### Agentic Dataset Provenance
 
 AIPerf exports public-dataset provenance in `metadata.dataset`, including the Hugging Face dataset ID. InferenceX preserves that object as `dataset` on each agentic aggregate benchmark row. During benchmark ingest, `ingest-ci-run.ts` derives the dashboard slug from `hf_dataset_name` (for example, `semianalysisai/cc-traces-weka-062126` becomes `cc-traces-weka-062126`) and upserts `run_datasets` for the workflow run.
 
 Legacy artifacts without provenance leave any existing mapping untouched. A workflow run can map to only one dataset; conflicting dataset IDs fail ingest rather than silently linking the run to an arbitrary dataset.
 
+### Agentic Replay-Lane Identity
+
+AIPerf can sample one dataset conversation into multiple trajectory lanes, and
+those replays may execute concurrently. The source `conversation_id` therefore
+identifies dataset provenance, not a unique execution lane. During trace-replay
+ingest, `computeRequestTimeline()` groups requests by `root_correlation_id`.
+For older exports without that field, it walks `parent_correlation_id` ancestry
+from each per-session `x_correlation_id`, so nested subagents resolve to the
+main session rather than becoming separate replay lanes. The extractor replaces
+the repeated root UUID with a compact, deterministic numeric `ri` ranked by each
+lane's first dispatch.
+
+The request-timeline frontend groups by `(conversation_id, ri)`, keeping all
+main-agent, subagent, warmup, and profiling requests from one replay together.
+Legacy timelines without correlation metadata omit `ri` and retain the older
+conversation-only grouping. `REQUEST_TIMELINE_VERSION` invalidates derived
+JSONB when this extraction changes; `db:backfill-request-timeline` recomputes it
+from the raw gzipped `profile_export.jsonl` already stored in Neon, so no GitHub
+artifact refetch or schema migration is required.
+
+Timeline extraction reads the gzip stream twice instead of materializing the
+entire JSONL. The first pass retains only timeline bounds and one first-dispatch
+timestamp per replay; the second emits compact request records. This bounds
+peak memory for high-throughput traces whose decompressed profiles exceed
+400 MB, at the cost of a second sequential decompression pass during ingest or
+backfill. The backfill also downloads the stored gzip and uploads the derived
+JSONB in 4 MiB chunks, matching trace ingest and avoiding single-value transport
+limits for oversized runs. The request-timeline reader checks only compact
+metadata inline, then reconstructs every current timeline from bounded text
+chunks. It deliberately does not choose by compressed JSONB storage size, so a
+highly compressible timeline cannot cross Neon's 64 MiB response limit or force
+the database to materialize the serialized size just to select a read path.
+
 ### Agentic Full-Response Interactivity
 
 Agentic charts use full-response inter-token latency as their canonical ITL and
@@ -109,6 +285,24 @@ minus TTFT, then divides by `output_sequence_length - 1`. The same helper powers
 the one-time `db:backfill-full-response-interactivity` data migration, keeping
 historical rows and newly ingested rows on one definition.
 
+### Agentic Point-Detail Payloads
+
+The point-detail page deliberately splits its immutable trace data by use case.
+The default charts fetch `request-chart-data`, a dictionary-encoded tuple
+projection containing only conversation, phase, request timing, latency, and
+sequence-length fields. The source-rich `request-timeline` document is fetched
+only after the user opens the Timeline view. Likewise, `trace-server-metrics`
+returns the aggregate series plus source descriptors; a selected endpoint's
+full arrays come from `trace-server-metric-source` on demand.
+
+This split is a scale invariant, not only a transport optimization. A high-
+throughput point can contain more than 150,000 requests and several copies of
+each server series. Reusing either full document for the default view multiplies
+JSON parsing, React Query memory, and cache-transfer cost by data the visible
+charts do not consume. The Timeline view also virtualizes rows, so the number of
+SVG bar/link nodes is bounded by the scroll viewport rather than total request
+count.
+
 ## Frontend Transform Pipeline
 
 ### Why transformBenchmarkRows Exists
diff --git a/docs/data-transforms.md b/docs/data-transforms.md
index b04ae973e..b09c43d06 100644
--- a/docs/data-transforms.md
+++ b/docs/data-transforms.md
@@ -21,7 +21,7 @@ RenderableGraph[]  (consumed by ScatterGraph)
 
 **`AggDataEntry`** (`components/inference/types.ts`) — flattened, fully-typed working representation. All metric keys are promoted to top-level fields with `?? 0` defaults. The display model name is resolved here. `hwKey` starts as an empty string and is filled in by `transformBenchmarkRows` after calling `getHardwareKey`. The `actualDate` field holds the real DB date when the `date` field has been overridden to a user-selected comparison date.
 
-**`InferenceData`** (`components/inference/types.ts`) — chart-ready scatter point. Extends `Partial<AggDataEntry>` and adds: chart coordinates (`x`, `y`), all derived metrics as `{ y: number; roof: boolean }` objects, and narrowed boolean types for the `dp_attention`/`disagg` family. The `roof` flag is always `false` at creation and set to `true` by `markRooflinePoints` during the render phase.
+**`InferenceData`** (`components/inference/types.ts`) is a chart-ready scatter point. It extends `Partial<AggDataEntry>` with chart coordinates (`x`, `y`), derived metric objects shaped as `{ y: number; roof: boolean }`, and narrowed boolean types for the `dp_attention`/`disagg` family. Derived metric objects start with `roof: false`; current frontier membership is computed separately from the selected metric and chart direction.
 
 **`RenderableGraph`** — final output of `useChartData`. Bundles `model`, `sequence`, `chartDefinition`, and `InferenceData[]` for a single scatter chart panel.
 
@@ -44,27 +44,29 @@ RenderableGraph[]  (consumed by ScatterGraph)
 1. Converts every row to `AggDataEntry` once (via `rowToAggDataEntry`).
 2. Calls `getHardwareKey(entry)` and writes the result back into `entry.hwKey`.
 3. Calls `getHardwareConfig(hwKey)` with a per-call cache to build the `HardwareConfig` map (hardware display metadata — label, color, GPU title).
-4. For each `ChartDefinition` in `inference-chart-config.json`, calls `createChartDataPoint` with that definition's `x`/`y` field keys to produce one `InferenceData` array per chart. The same `AggDataEntry` objects are reused across chart definitions — they are not re-created.
+4. Builds canonical derived fields once per prepared entry, then calls `createChartDataPoint` for each chart definition with that definition's `x`/`y` keys and the shared fields. The same `AggDataEntry` and derived objects are reused across both charts.
 
 Returns `{ chartData: InferenceData[][], hardwareConfig: HardwareConfig }`.
 
 ### Step 2: AggDataEntry to InferenceData (`lib/chart-utils.ts`)
 
-**`createChartDataPoint(date, entry, xKey, yKey, hwKey)`** — spreads `entry` first, then overrides with derived fields:
+**`createChartDataPoint(date, entry, xKey, yKey, hwKey, derivedFields?)`** spreads `entry` first, then overrides chart coordinates and metadata. The full transform passes precomputed derived fields; direct callers may omit them.
 
 - `x` / `y`: read directly from `entry[xKey]` and `entry[yKey]` (set per chart definition).
 - `tp`: for disaggregated configs, set to `num_prefill_gpu + num_decode_gpu` instead of `decode_tp`.
 - Boolean narrowing: `dp_attention`, `prefill_dp_attention`, `decode_dp_attention`, and `is_multinode` are coerced from `boolean | string` to `boolean | undefined`.
 - Disagg fields: `num_prefill_gpu` / `num_decode_gpu` are only set when `entry.disagg` is true; otherwise they are dropped.
 
-**Derived metric fields** — each uses `{ y: number; roof: boolean }` so the chart layer can switch between them without re-fetching data:
+**`buildDerivedChartFields(entry, hwKey, requestedMetrics?)`** is the sole owner of inference and historical trend formulas. The full transform requests every benchmark-backed field. History requests only the selected metric and its interpolation dependencies, so it stays lightweight. Each emitted value uses `{ y: number; roof: boolean }`.
 
 - `tpPerGpu`, `outputTputPerGpu`, `inputTputPerGpu` — raw throughput from the entry (tok/s/gpu).
+- `tokenRevenuePerGpuHour` — gross revenue in $/GPU/hr. `buildDerivedChartFields` emits the normalized default (`tputPerGpu × 3,600 ÷ 1,000,000`, with input and output both priced at $1/M). `applyTokenRevenuePricing` can replace it with the selected model's current public OpenRouter prompt/completion prices. Aggregate rows use their measured input/output split; disaggregated rows derive a like-for-like split from fixed ISL:OSL or measured agentic prompt:generation totals because their raw input/output rates use different prefill/decode GPU denominators.
 - `tpPerMw` — `(tputPerGpu * 1000) / hardwarePower` (GPU power in kW, result in tok/s/MW).
-- Cost fields — GPU hourly cost divided by tokens-per-hour (in millions): `costh` / `costn` / `costr` for hyperscaler / neocloud / 3-year-rental pricing respectively. Three token variants exist: combined (`costh`/`costn`/`costr`), output-only (`costhOutput`/`costnOutput`/`costrOutput`), input-only (`costhi`/`costni`/`costri`).
+- Cost-per-million fields — GPU hourly cost divided by tokens-per-hour (in millions): `costh` / `costn` / `costr` for hyperscaler / neocloud / 3-year-rental pricing respectively. Three token variants exist: combined (`costh`/`costn`/`costr`), output-only (`costhOutput`/`costnOutput`/`costrOutput`), and input-only (`costhi`/`costni`/`costri`).
+- Tokens-per-dollar fields — tokens-per-hour divided by the same hourly costs. They are separate Y-axis metrics rather than a display toggle: combined (`tokensPerDollarH`/`N`/`R`), output-only (`outputTokensPerDollarH`/`N`/`R`), and input-only (`inputTokensPerDollarH`/`N`/`R`). This keeps existing `i_metric=y_cost*` links and cost semantics unchanged.
 - Energy fields — `jTotal` / `jOutput` / `jInput`: `(hardwarePower * 1000) / tputPerGpu` (Joules per token, where power in kW is converted to W).
 
-**GPU specs lookup** — `createChartDataPoint` calls `getGpuSpecs(hwKey)` (`lib/constants.ts`) which splits on `[-_]` to extract the base GPU token (e.g. `"b200_trt_mtp"` → `"b200"`) and looks it up in `HW_REGISTRY`. `HW_REGISTRY` stores power (kW per GPU) and three cost tiers ($ per GPU-hour). Missing keys return zeroed specs, which produces `0` cost/energy values rather than crashing.
+**GPU specs lookup** happens inside `buildDerivedChartFields`. `getGpuSpecs(hwKey)` (`lib/constants.ts`) splits on `[-_]` to extract the base GPU token (for example, `"b200_trt_mtp"` becomes `"b200"`) and looks it up in `HW_REGISTRY`. Missing keys return zeroed specs, producing `0` cost/energy values rather than crashing.
 
 ### Step 3: Filtering, memoization, and rendering (`hooks/useChartData.ts`)
 
@@ -78,7 +80,7 @@ The hook runs a 5-step memoized pipeline:
 
 4. **Sort `hardwareConfig`** — the `HardwareConfig` object is sorted by `getModelSortIndex` and stabilized with a ref: if the sorted key string matches the previous render, the same object reference is returned. This prevents D3 Effect 2 (data bind) from firing when a sequence change returns the same GPU set.
 
-5. **Build renderable graphs** — `stableChartDefinitions` is computed in a separate `useMemo` that depends only on metric/axis selections (not on data). This decouples Y-axis changes from data changes so D3 Effect 3 (metric repositioning) does not fire alongside Effect 2 (data bind). Within this memo, the x-axis field is resolved per chart type, roofline directions are flipped when the x-axis polarity reverses (e.g. interactivity → TTFT), and Y-axis labels are looked up. The final `graphs` memo applies GPU filtering, cost-limit clamping, optional user-cost/user-power overrides, and remaps each point's `x`/`y` from the selected metric's `{ y, roof }` object.
+5. **Build renderable graphs** — `stableChartDefinitions` is computed in a separate `useMemo` that depends only on metric/axis selections (not on data). This decouples Y-axis changes from data changes so D3 Effect 3 (metric repositioning) does not fire alongside Effect 2 (data bind). Within this memo, the x-axis field is resolved per chart type, roofline directions are flipped when the x-axis polarity reverses (e.g. interactivity → TTFT), and Y-axis labels are looked up. The final `graphs` memo applies GPU filtering, cost-limit clipping, optional user-cost/user-power overrides, and remaps each point's `x`/`y` from the selected metric's `{ y, roof }` object.
 
 ---
 
@@ -106,7 +108,7 @@ The resulting key's base GPU must exist in `HW_REGISTRY`. Display fields (label,
 
 ## Chart Configuration
 
-`inference-chart-config.json` defines exactly two `ChartDefinition` objects:
+`metric-registry.ts` owns metric fields, English and Chinese labels, titles, polarity, and control ordering. It derives exactly two `ChartDefinition` objects:
 
 | `chartType`     | Default x-axis  | x meaning                                                      |
 | --------------- | --------------- | -------------------------------------------------------------- |
@@ -115,13 +117,13 @@ The resulting key's base GPU must exist in `HW_REGISTRY`. Display fields (label,
 
 Both charts share the same Y-axis options. The `y` field is the default `AggDataEntry` key used for raw Y values; each `y_{metric}` field overrides this with a dotted path into the `InferenceData` derived fields (e.g. `"tpPerGpu.y"`).
 
-**Per-metric Y-axis schema** — for each metric key (e.g. `y_costh`), the config carries:
+**Per-metric Y-axis schema** is owned once in `METRIC_REGISTRY`:
 
-- `y_{metric}`: dotted path for the value (e.g. `"costh.y"`).
-- `y_{metric}_label`: Y-axis label string.
-- `y_{metric}_title`: dropdown/UI title string.
-- `y_{metric}_roofline`: Pareto direction (`upper_left`, `upper_right`, `lower_left`, `lower_right`). Roofline direction differs between chart types for the same metric because the x-axis polarity differs (interactivity: higher-is-better; E2EL: lower-is-better).
+- `field`: dotted path for the value (for example, `"costh.y"`).
+- `label` / `labelZh`: bilingual Y-axis labels.
+- `title` / `titleZh`: bilingual dropdown and chart titles.
+- `polarity`: whether higher or lower values are preferable. The derived chart definitions combine this with x-axis polarity to produce the concrete Pareto corner.
 
-**Input-metric x-axis override** — when the selected Y metric's title contains `"input"`, the interactivity chart switches its x-axis to `p99_ttft` (or the user-overridden x metric). This is detected in `stableChartDefinitions` via `metricTitle.toLowerCase().includes('input')`. The config encodes the default override fields: `y_inputTputPerGpu_x: "p99_ttft"` and `y_inputTputPerGpu_x_label`.
+**Input-metric x-axis override** switches the interactivity chart to TTFT when the selected Y metric is input-related. `inputTputPerGpu` declares `p90_ttft` and its label in the registry; a user-selected TTFT metric may override it.
 
-**Limits** — both charts include `y_cost_limit: 5` (clamp cost-metric y-axis to $5/M tokens) and `y_latency_limit: 60` (filter x-axis outliers beyond 60s when TTFT is on x).
+**Limits** — both charts include `y_cost_limit: 5` (clip cost-per-million metrics above $5/M tokens) and `y_latency_limit: 60` (clip x-axis outliers beyond 60s when TTFT is on x). Tokens-per-dollar metrics are not cost-clipped; clipped cost points remain available to the dashed continuation layer.
diff --git a/docs/i18n.md b/docs/i18n.md
index 2a6c5dde5..4063f5502 100644
--- a/docs/i18n.md
+++ b/docs/i18n.md
@@ -4,7 +4,7 @@ Why the Simplified Chinese site is a hand-authored `/zh` page tree instead of an
 
 ## Why hand-authored pages, not next-intl / `[locale]` routing
 
-- **SEO is the goal, not full UI translation.** The objective is Chinese pages that crawlers can index: Chinese titles, meta descriptions, server-rendered Chinese content, and bidirectional hreflang. The interactive dashboard (charts, filters, tooltips) stays English — model/GPU/framework names are English-first terms in Chinese ML writing anyway, and translating deep chart UI would touch hundreds of client components for little search value.
+- **SEO and a complete Chinese user experience are both goals.** Chinese titles, meta descriptions, server-rendered content, and bidirectional hreflang make the pages indexable; localized controls, tooltips, states, and explanations make them usable without falling back to English. Product names, model and hardware SKUs, framework names, units, flags, and identifiers remain English where Chinese ML infrastructure writing normally keeps them.
 - **A `[locale]` root segment would move every route** and force middleware rewrites to keep existing URLs stable — high blast radius for a two-locale site. A parallel `/zh` tree adds pages without touching English URLs.
 - **No message-catalog indirection.** With exactly two locales and mostly page-level content, colocated `STRINGS = { en, zh }` dictionaries (landing page, quotes content) and dedicated zh files (`tab-meta-zh.ts`, `faq-data-zh.ts`) are easier to review than a parallel key-file hierarchy, and dead strings die with their page.
 
@@ -26,5 +26,6 @@ Why the Simplified Chinese site is a hand-authored `/zh` page tree instead of an
 - **zh OG images reuse the English post meta** — the `next/og` default Satori font has no CJK glyphs, so a Chinese title would render as tofu. Loading a subset CJK font is a known follow-up.
 - **`/zh/inference` canonicalizes to `/zh`**, mirroring the English quirk where `/inference` canonicalizes to `/`.
 - **All UI strings are localized in place** via `useLocale()` + component-local `STRINGS = { en, zh }` dicts (footer, TabNav, dashboard display headings/labels, nudges, preset cards, legend toggles, select placeholders), and registry-defined display names (Y-axis metrics, toggle configs) carry `labelZh` fields resolved through locale-aware helpers. The `en` values keep the exact original strings so English pages are byte-identical. Only brand/product/hardware/framework/precision names, units, code identifiers, and DB-stored content stay English.
+- **Editorial decisions are contextual.** [Chinese Copy Editorial Guide](./chinese-copy.md) defines the audience, surface-specific register, terminology exceptions, rewrite workflow, and manual-review process. It is the source of truth for how Chinese copy should read; this document remains the source of truth for route and rendering architecture.
 - **Compare slug pages are mirrored** at `/zh/compare/[slug]` and `/zh/compare-per-dollar/[slug]`. The Chinese narrative templates live in `compare-ssr-zh.ts` as a 1:1 port of `compare-ssr.ts` (data logic is imported, only sentence templates differ) — the two files must change together.
 - **Sitemap pairs**: `localizedPair()` in `sitemap.ts` emits the EN and zh URL together, both carrying the same `alternates.languages` map. Blog posts without a translation fall back to an English-only entry, so a missing translation degrades gracefully instead of 404-ing crawlers.
diff --git a/docs/index.md b/docs/index.md
index c4a72633b..7a22a0d1b 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -4,7 +4,7 @@ Design rationale and non-obvious conventions. See [CLAUDE.md](../CLAUDE.md) for
 
 ## Docs
 
-- [Architecture](./architecture.md) — Why client-first, hash routing, URL state, provider nesting, server-side caching (unstable_cache + blob), in-memory client cache, color system, analytics enforcement
+- [Architecture](./architecture.md) — Why client-first, route navigation, URL state, provider nesting, server-side caching (unstable_cache + blob), in-memory client cache, color system, analytics enforcement
 - [D3 Charts](./d3-charts.md) — Why 4 effects, in-place mutation, refs for zoom, rAF throttling, HTML tooltips, Pareto directions, gradient labels
 - [Data Pipeline](./data-pipeline.md) — DB schema reasoning, ETL design, transform pipeline, spline method choice, normalizer resolution order (model/GPU/framework)
 - [Pitfalls](./pitfalls.md) — Failure modes: token type consistency, schema evolution, empty objects, zoom loss, stale closures, disaggregated metrics, negative splines, date stamping, ref stability, cost inheritance
@@ -12,8 +12,10 @@ Design rationale and non-obvious conventions. See [CLAUDE.md](../CLAUDE.md) for
 - [TCO Calculator](./tco-calculator.md) — Why interpolation, composite keys, cost matrix, token type bugs, badge logic, state design
 - [Adding Entities](./adding-entities.md) — Step-by-step checklists for adding new models, GPUs, precisions, sequences, frameworks (ingest + constants + frontend), plus featuring a day-0 model (launch banner, modal, Quick Comparisons preset)
 - [Testing](./testing.md) — Requirements, quality standards, pre-commit checklist
+- [Typography](./typography.md) — Type tokens (text-2xs/3xs, named tracking), Heading/Eyebrow components, TS-based chart font sizes, and the check:typography ratchet gate + allowlist burndown
 - [Data Transforms](./data-transforms.md) — Full pipeline from BenchmarkRow to RenderableGraph: type hierarchy, hardware key construction, derived metrics, memoization strategy
 - [State Ownership](./state-ownership.md) — Which context owns which state, availability filtering cascade, comparison date mechanics, URL param sync
 - [Blog](./blog.md) — MDX content system, SEO features (OG images, RSS, llms.txt, JSON-LD), TOC sidebar, reading progress, heading links, analytics events
 - [CollectiveX](./collectivex.md) — lazy ingest-on-read pipeline (separate Neon DB as a durable GitHub-artifact cache), tombstoned deletes, raw-docs storage + shared reader
 - [Chinese Pages (/zh)](./i18n.md) — Why hand-authored /zh pages instead of an i18n framework, hreflang pairing, blog translation pairing, html lang workaround, CJK reading time/slugs
+- [Chinese Copy Editorial Guide](./chinese-copy.md) — Audience, surface-specific register, context-aware terminology, rewrite workflow, manual-review process, and PR checklist
diff --git a/docs/pitfalls.md b/docs/pitfalls.md
index 756de2dde..ffe3af17c 100644
--- a/docs/pitfalls.md
+++ b/docs/pitfalls.md
@@ -4,7 +4,7 @@ Non-obvious failure modes discovered during development. Each entry explains wha
 
 ## Token Type Consistency
 
-**Bug pattern**: Chart title says "Output Token Cost" but bars show total token cost.
+**Bug pattern**: Chart title says "Output Tokens per $1" or "Cost per Million Output Tokens" but values use total-token throughput.
 
 When a UI has a Token Type selector (Total/Input/Output), **every** downstream value must use the selected type: chart title, bar values, table values, tooltips, sorting, comparison text, axis labels. The common mistake is updating the title and sorting but forgetting the tooltip or comparison banner.
 
@@ -40,9 +40,9 @@ D3 event handlers capture variables from their creation scope. If a handler refe
 
 ## Disaggregated Config Metrics
 
-**Bug pattern**: Cost per token is 2x too high for disaggregated setups.
+**Bug pattern**: Tokens per $1 is 2x too low—or cost per token is 2x too high—for disaggregated setups.
 
-Disaggregated configs (MoRI SGLang, Dynamo TRT) have separate prefill and decode GPU counts. Cost and throughput metrics must be calculated per decode/prefill GPU, not per total GPU count. The disclaimer in ChartDisplay.tsx explains this to users.
+Disaggregated configs (MoRI SGLang, Dynamo TRT) have separate prefill and decode GPU counts. Cost, purchasing-power, and throughput metrics must be calculated per decode/prefill GPU, not per total GPU count. The disclaimer in ChartDisplay.tsx explains this to users.
 
 When adding metrics that involve per-GPU calculations, check if the config is disaggregated (`d.disagg === true`) and use the appropriate GPU count.
 
diff --git a/docs/state-ownership.md b/docs/state-ownership.md
index 07e5a2bcf..84f546bbb 100644
--- a/docs/state-ownership.md
+++ b/docs/state-ownership.md
@@ -2,25 +2,32 @@
 
 Reference for agents modifying filter behavior. Explains which context to touch for a given change, how availability cascades, and how URL params are kept in sync.
 
-## Provider Nesting
+## Provider Ownership
+
+`DashboardShell` resolves provider capabilities from
+`packages/app/src/lib/dashboard-routes.ts`:
 
 ```
 QueryProvider
   ThemeProvider
-    UnofficialRunProvider
-      GlobalFilterProvider          ← availability, model/sequence/precision/date/runId
-        InferenceProvider           ← GPU selection, comparison dates, chart UI, y-axis metric
-          inference tab
-          historical tab (no own provider — reads InferenceContext)
-        EvaluationProvider          ← benchmark, eval-specific date, hardware toggle
-          evaluation tab
-        ReliabilityProvider         ← date range, model toggle
-          reliability tab
-        calculator tab              ← no provider; local useState in ThroughputCalculatorDisplay
-        gpu-specs tab               ← no provider; static data
+    DashboardShell
+      inference / historical / evaluation
+        UnofficialRunProvider
+          GlobalFilterProvider
+            route-owned InferenceProvider or EvaluationProvider
+      calculator
+        UnofficialRunProvider
+          route-owned, URL-seeded GlobalFilterProvider
+      reliability
+        route-owned ReliabilityProvider
+      static / internal routes
+        no data providers
 ```
 
-Source: `packages/app/src/components/page-content.tsx` lines 203–239.
+Compare routes live outside `DashboardShell` and own seeded
+`GlobalFilterProvider` / `InferenceProvider` pairs. Agentic point-detail routes are
+standalone. Provider changes must update the canonical route registry and its parity
+tests rather than adding pathname conditionals in the shell.
 
 ---
 
@@ -30,27 +37,40 @@ Source: `packages/app/src/components/page-content.tsx` lines 203–239.
 
 File: `packages/app/src/components/GlobalFilterContext.tsx`
 
-**Selection state** (user-settable, URL-initialised):
+**Requested selection state** (user or URL intent):
+
+- `selectedModel` (`g_model`)
+- `selectedSequence` (`i_seq`)
+- `selectedPrecisions` (`i_prec`)
+- requested run date and run ID (`g_rundate`, `g_runid`)
 
-- `selectedModel` — active model (`g_model`)
-- `selectedSequence` — active ISL/OSL sequence (`i_seq`; owned here because it gates availability for all tabs)
-- `selectedPrecisions` — active precision list (`i_prec`; same reason)
-- `selectedRunDate` / `selectedRunId` — active benchmark run date and run ID (`g_rundate`, `g_runid`)
+**Effective values** (purely derived):
 
-**Effective (auto-corrected) values** (derived, not settable directly):
+- `effectiveSequence` resolves the requested scenario against availability
+- `effectivePrecisions` resolves requested precisions against available curves
+- `effectiveRunDate` resolves the requested date or latest valid date
+- `selectedRunId` resolves requested run intent against `availableRuns`
 
-- `effectiveSequence` — `selectedSequence` if valid for current model, else first available
-- `effectivePrecisions` — subset of `selectedPrecisions` that are available; falls back to `[availablePrecisions[0]]`
-- `effectiveRunDate` — latest available date unless user explicitly picked one
+`GlobalFilterProvider` owns the state, but it does not publish one response-shaped
+context value. Consumers subscribe through five independently memoized domains:
 
-**Derived availability** (memos over `availabilityRows`):
+- `useGlobalFilterSelection`: requested model, sequence, and precision state plus
+  their effective values and `sequenceResolved`
+- `useGlobalFilterActions`: stable setters for selection, run date, and run ID
+- `useGlobalFilterRun`: effective date and run ID selectors plus the manual date revision
+- `useGlobalFilterAvailability`: availability options, raw rows, settled state, and errors
+- `useGlobalFilterWorkflow`: the run map plus workflow loading and error state
 
-- `availableModels`, `availableSequences`, `availablePrecisions`, `availableDates`
-- `availabilityRows` — raw `AvailabilityRow[]` from `useAvailability()`, passed down to InferenceProvider for GPU filtering
+Workflow query updates therefore do not notify selection or action consumers. Selection
+updates do not notify workflow-only consumers. Components should subscribe only to the
+domains they read.
 
-**Workflow / run info** (derived from `useWorkflowInfo(effectiveRunDate)`):
+Availability owns explicit settled and error states. Consumers must distinguish an
+unresolved request from a settled empty result. Inference fetches are gated on
+`sequenceResolved` and render `availabilityError` rather than an indefinite skeleton.
 
-- `availableRuns`, `workflowInfo`, `workflowLoading`, `workflowError`
+There is no response-shaped workflow adapter. `RunInfo` and `availableRuns` are the
+canonical run contract.
 
 **Why here, not InferenceProvider**: Model, sequence, and precision are cross-tab. EvaluationContext consumes `selectedModel` and `availableModels` directly. If these lived in InferenceProvider, EvaluationProvider would need an indirect coupling or duplicate state.
 
@@ -60,55 +80,55 @@ File: `packages/app/src/components/GlobalFilterContext.tsx`
 
 File: `packages/app/src/components/inference/InferenceContext.tsx`
 
-Depends on: `GlobalFilterProvider` (reads all filter state and availability, including `availabilityRows`).
-
-**GPU comparison state** (inference-only, URL-initialised):
-
-- `selectedGPUs` — hardware keys selected for GPU filter/comparison (`i_gpus`)
-- `selectedDates` — discrete comparison dates (`i_dates`)
-- `selectedDateRange` — `{startDate, endDate}` for range comparisons (`i_dstart`, `i_dend`)
-- `activeDates` — `Set<string>` toggle controlling visible comparison overlays (keyed by `${date}_${gpuKey}`)
+Depends on `GlobalFilterProvider` through its narrow selection, run, availability,
+workflow, and action hooks.
 
-**Chart axis / display state** (URL-initialised):
+The provider owns inference-specific effects and URL synchronization, then exposes
+four independently memoized domains. Consumers must subscribe only to the domains
+they read.
 
-- `selectedYAxisMetric` (`i_metric`), `selectedXAxisMetric` (`i_xmetric`), `selectedE2eXAxisMetric` (`i_e2e_xmetric`)
-- `scaleType` — `auto | linear | log` (`i_scale`)
-- `hideNonOptimal` (`i_optimal`), `showPointLabels` (`i_label`), `logScale` (`i_log`)
-- `highContrast` (`i_hc`), `isLegendExpanded` (`i_legend`)
-- `useAdvancedLabels` (`i_advlabel`), `showGradientLabels` (`i_gradlabel`)
-- `colorShuffleSeed` — no URL param; ephemeral
+**`useInferenceData`**
 
-**Derived availability** (GPU-level, computed from `availabilityRows` inherited from GlobalFilterContext):
+Fetched and derived benchmark results:
 
-- `availableGPUs` — hardware configs that have data for the current model + sequence + precisions AND have a known base GPU in `HW_REGISTRY`
-- `dateRangeAvailableDates` — dates available for the current filter combination, further narrowed by `selectedGPUs`
-- `hwTypesWithData` — `Set<string>` of GPU keys currently present in fetched chart data
+- `graphs`, `hardwareConfig`, `loading`, and `error`
+- `hwTypesWithData`
+- GPU, date, precision, sequence, model, run, and quick-filter availability
 
-**Hardware toggle set**:
+**`useInferenceFilters`**
 
-- `activeHwTypes` — subset of `hwTypesWithData` that are visible (managed by `useChartDataFilter`)
+Workflow, filter, and date selection state:
 
-**Tracked configs / presets**:
+- effective model, sequence, and precision selections read from global filter state
+- `selectedGPUs`, `selectedDates`, `selectedDateRange`, and `activeDates`
+- `activeHwTypes`, `bestPerSku`, and `quickFilters`
+- effective run selection, custom cost and power values, preset state, and compare scope
 
-- `trackedConfigs` — up to 6 pinned data points for cross-chart comparison
-- `activePresetId`, `pendingHwFilter` — active favourite preset and its deferred GPU filter
+This domain owns inference-only selection state. GPU selections and comparison dates
+do not belong in global filter state because evaluation and reliability do not use
+them.
 
-**User overrides**:
+**`useInferenceDisplay`**
 
-- `userCosts`, `userPowers` — per-GPU cost/power overrides for custom cost metric; reset when `selectedYAxisMetric` changes away from `y_costUser`/`y_powerUser`
+Axis and presentation state:
 
-**Run filtering** (inference-local, not written back to GlobalFilterContext):
+- selected x-axis and y-axis metrics, percentile, and effective x-axis mode
+- token-revenue price source (`i_revenue`): normalized $1/M or the selected model's live OpenRouter catalog prices
+- scale, optimal-point, label, contrast, legend, and overlay controls
 
-- `filteredAvailableRuns` — `availableRuns` filtered to runs matching `selectedModel` + `effectivePrecisions`
-- `effectiveSelectedRunId` — validated run ID within `filteredAvailableRuns`; intentionally NOT synced back to GlobalFilterContext to avoid full-tree re-renders on precision change
+Display changes stay in this domain. A contrast or label toggle therefore does not
+notify filter-only or data-only consumers.
 
-**Charts data** (from `useChartData`):
+**`useInferenceActions`**
 
-- `graphs` — `RenderableGraph[]` used by all D3 charts
-- `hardwareConfig` — config map derived from benchmark rows
-- `loading`, `error`
+All inference commands and setters. The context value and every exposed action keep
+stable identities. The provider routes each stable action to the latest implementation,
+so new availability or chart data cannot invalidate action-only consumers.
 
-**Why not in GlobalFilterContext**: GPU selection and comparison dates are meaningless outside the inference/historical tabs. Putting them in the global context would pollute the interface for evaluation and reliability.
+Run filtering remains inference-local. `filteredAvailableRuns` narrows global run
+availability to the effective model and precisions. `effectiveSelectedRunId` is
+validated within that set but is not written back through the global run action, so a
+precision-specific fallback cannot replace the user's global run intent.
 
 ---
 
@@ -116,12 +136,16 @@ Depends on: `GlobalFilterProvider` (reads all filter state and availability, inc
 
 File: `packages/app/src/components/evaluation/EvaluationContext.tsx`
 
-Depends on: `GlobalFilterProvider` (reads `selectedModel`, `setSelectedModel`, `selectedRunDate`, `selectedRunDateRev`, `setSelectedRunDate`, `availableModels`, `availableDates`).
+Depends on `GlobalFilterProvider` for model and global date intent.
 
 **Selection state**:
 
-- `selectedRunDate` — evaluation-specific date; initialised from `e_rundate` URL param or `globalRunDate`. Bidirectionally synced: when the user picks a date, it calls `setGlobalRunDate` (if the date exists in inference availability). When `globalRunDate` changes (from another tab), the effect at line 124 applies it here.
-- `selectedBenchmark` — active eval task (`e_bench`)
+- requested evaluation date plus a reducer-derived nearest valid `selectedRunDate`
+- `selectedBenchmark` (`e_bench`)
+
+Explicit evaluation date actions synchronize the global requested date only when that
+date exists in inference availability. Global changes are reducer inputs rather than a
+pair of bidirectional mirroring effects.
 
 **UI state**:
 
@@ -165,15 +189,19 @@ Does NOT consume `GlobalFilterProvider`. Fully standalone — reliability data h
 
 ---
 
-### Tabs without providers
+### Route-Owned Providers
 
-**TCO Calculator** (`calculator` tab): All state is local `useState` inside `ThroughputCalculatorDisplay`. It reads `effectiveSequence` and `effectivePrecisions` from `useGlobalFilters()` for the initial GPU list but does not share state back.
+**TCO Calculator**: the route parses a typed URL seed and mounts the sole
+`GlobalFilterProvider` for the calculator. Calculator-specific visibility, target, and
+bar selection transitions live in one reducer. `visibleHwKeys` remains the single
+source of truth for official and unofficial bars.
 
-It reads `unofficialBenchmarkRows` / `unofficialRunInfos` / `runIndexByUrl` from `useUnofficialRun()` to render unofficial-run overlay bars, but it neither reads nor writes the shared `activeOverlayHwTypes` that the inference and evaluation tabs use: the calculator's local `visibleHwKeys` governs its official and overlay bars alike. Two visibility sets behind one legend can only drift — see [TCO Calculator › Unofficial-Run Overlays](./tco-calculator.md#unofficial-run-overlays-unofficialrun).
+**Historical Trends**: mounts `InferenceProvider` and shares the global filter provider
+declared by its route capability.
 
-**Historical Trends** (`historical` tab): Rendered inside `InferenceProvider` (shares the `inference` + `historical` `isActive` gate). It reads `useInference()` directly — no additional provider. Uses InferenceContext's model/sequence/precision/date state.
+**Reliability**: mounts only `ReliabilityProvider`.
 
-**GPU Specs**: Static data, no provider.
+**GPU Specs and other static/internal pages**: mount no data providers.
 
 ---
 
@@ -190,14 +218,16 @@ GlobalFilterProvider
   → selectedModel     (user pick)
   → modelRows         = availabilityRows filtered to selectedModel (internal memo)
   → availableSequences = unique sequences in modelRows
-  → effectiveSequence  = selectedSequence if in availableSequences, else availableSequences[0]
+  → effectiveSequence  = Agentic (if unchosen and in availableSequences),
+                         else selectedSequence if in availableSequences,
+                         else 8k/1k if available, else availableSequences[0]
   → availablePrecisions = unique precisions in modelRows where sequence = effectiveSequence
   → effectivePrecisions = selectedPrecisions ∩ availablePrecisions; falls back to [availablePrecisions[0]]
   → availableDates     = unique dates in modelRows where sequence = effectiveSequence
                          AND precision ∈ effectivePrecisions
   → effectiveRunDate   = latest of availableDates (unless user explicitly picked a date)
 
-InferenceProvider (receives availabilityRows from GlobalFilterContext)
+InferenceProvider (receives availability rows from `useGlobalFilterAvailability`)
   → availableGPUs     = availabilityRows filtered to (model, effectiveSequence, effectivePrecisions)
                         → hwKey extracted via buildAvailabilityHwKey()
                         → filtered by isKnownGpu() (base GPU in HW_REGISTRY)
@@ -235,62 +265,82 @@ Source files: `packages/app/src/lib/url-state.ts`, `packages/app/src/hooks/useUr
 
 | Prefix | Scope                                                                               |
 | ------ | ----------------------------------------------------------------------------------- |
-| `g_`   | GlobalFilterContext — model, run date, run ID                                       |
-| `i_`   | InferenceProvider — sequence, precision, GPUs, dates, metrics, display toggles      |
+| `g_`   | Global filter run and selection contexts: model, run date, run ID                   |
+| `i_`   | Global selection plus InferenceProvider: sequence, precision, GPUs, metrics         |
 | `e_`   | EvaluationProvider — eval date (only when it differs from globalRunDate), benchmark |
 | `r_`   | ReliabilityProvider — date range, display toggles                                   |
 
-Note: `i_seq` and `i_prec` are written by `GlobalFilterProvider` (not InferenceProvider) because they live in GlobalFilterContext.
+Note: `i_seq` and `i_prec` are written by `GlobalFilterProvider` because selection state is owned by `useGlobalFilterSelection`, not InferenceProvider.
+
+### Snapshot, external subscriptions, and writes
+
+`url-state.ts` snapshots known share parameters at module load and removes them from the
+visible address bar after hydration while preserving the pathname and hash. Route-owned
+entry points pass typed seeds where they render URL-controlled state on the server.
 
-### Snapshot-and-clear on load
+Live address-bar consumers use the shared `useClientSearch` external store. It subscribes
+once to `popstate` and `CLIENT_SEARCH_CHANGE_EVENT` through `useSyncExternalStore`;
+components do not mirror search strings through independent effects.
 
-`url-state.ts` runs at module-load time (before any React render). It reads all known keys from `window.location.search` into `_initialParams` and copies them into `currentState`. A `setTimeout(0)` callback then strips those keys from the URL via `history.replaceState`. This ensures the URL is clean after load — subsequent writes reflect only current state.
+`writeUrlParams` batches state serialization for 150 ms. `useUrlStateSync` and explicit
+state actions write requested/user state, not derived fallbacks.
 
-`useUrlState` caches `_initialParams` in a ref via `readUrlParams()` on first call. All `getUrlParam()` calls return from this snapshot, so providers initialize from the original share URL even if the cleanup timer has already run.
+The mutable remount snapshot and explicit URL intent are separate. Debounced
+filter writes update the remount snapshot so providers preserve current state,
+but they do not become explicit share-link choices that block automatic
+unofficial-run model selection. A real navigation rebuilds the explicit-intent
+set from the destination URL.
 
-### Debounced write-back (150 ms)
+### Carrying state across a full-document navigation
 
-`writeUrlParams(params)` merges incoming params into `pendingParams` and sets a 150 ms debounce timer. On flush, params matching their default in `PARAM_DEFAULTS` are deleted from `currentState`; non-defaults are stored. This keeps share URLs short.
+The two effects above combine into a trap: filter changes reach only the
+in-memory `currentState`, and the address bar is stripped clean, so **the URL of
+the `/inference` history entry describes nothing**. Any full-document navigation
+away from the chart therefore destroys the state entirely — the module is torn
+down, and Back returns to a bare `/inference` that rebuilds from defaults.
 
-`useUrlStateSync` (used by InferenceProvider, EvaluationProvider, ReliabilityProvider) skips the first render via `isMountedRef` to avoid overwriting the just-snapshotted URL params. GlobalFilterProvider uses a manually-written equivalent (lines 288–308 in `GlobalFilterContext.tsx`).
+The agentic point-detail links (`/inference/agentic/<id>`, rendered by
+`tooltipUtils.viewChartsButtonHTML` and `legend-points-table.pointDetailHref`)
+are exactly that: plain `<a href>`, deliberately, so browsers can offer
+open-in-new-tab. Three helpers in `url-state.ts` bridge the gap:
+
+| Helper                      | Used by                                                      | Why                                                                           |
+| --------------------------- | ------------------------------------------------------------ | ----------------------------------------------------------------------------- |
+| `rememberChartStateInUrl()` | the point-click handlers (Scatter / GPU graph, legend table) | `history.replaceState`s the chart state onto the entry Back will return to    |
+| `withChartState(href)`      | `agenticDetailHref()`                                        | appends it to the outbound link so the detail page can link back to that view |
+| `currentChartSearch()`      | both of the above                                            | resolves the tab through the `/zh` prefix, flushes pending writes, filters    |
+
+The detail page reads the state back through `withChartState` (its own URL was
+snapshotted into `currentState` at load, then stripped as usual), which is what
+makes its "Inference chart" link land on the chart the reader left rather than
+on defaults.
+
+### Re-hydration on client-side navigation
+
+`useUrlState` calls `refreshUrlParamsOnNavigation(pathname)` **during render**,
+guarded by a module-level last-pathname so it runs at most once per navigation.
+The refresh first flushes pending debounced writes, then applies explicit
+destination parameters over that current state. This prevents a rapid filter
+change followed by a retained-provider tab switch from replaying the previous
+selection.
+
+That ordering also matters because providers read `getUrlParam` in `useState`
+initialisers and mount effects, both of which run before any parent's layout
+effect. With the hook-level refresh, every provider sees the params of the page
+it actually landed on. The once-per-pathname guard keeps a late-mounting
+component from replaying a stale URL over subsequent filter changes.
 
 ### Share URL construction
 
-`buildShareUrl()` flushes pending writes, reads `currentState`, and filters to only the param prefixes relevant to the current tab (defined in `TAB_PARAM_PREFIXES`). Inference share URLs include `g_` + `i_` params; evaluation includes `g_` + `e_`; reliability includes only `r_`.
-
-Historical Trends and TCO Calculator share the inference tab's URL path (`/inference` and `/calculator` respectively) but have no dedicated param prefix — they inherit whichever `i_` params are relevant.
-
-### Full parameter list
-
-| Param           | Owner               | Default                           |
-| --------------- | ------------------- | --------------------------------- |
-| `g_model`       | GlobalFilterContext | `DeepSeek-R1-0528`                |
-| `g_rundate`     | GlobalFilterContext | `''`                              |
-| `g_runid`       | GlobalFilterContext | `''`                              |
-| `i_seq`         | GlobalFilterContext | `8k/1k`                           |
-| `i_prec`        | GlobalFilterContext | `fp4`                             |
-| `i_metric`      | InferenceProvider   | `y_tpPerGpu`                      |
-| `i_xmetric`     | InferenceProvider   | `p99_ttft`                        |
-| `i_e2e_xmetric` | InferenceProvider   | `''`                              |
-| `i_scale`       | InferenceProvider   | `auto`                            |
-| `i_gpus`        | InferenceProvider   | `''`                              |
-| `i_dates`       | InferenceProvider   | `''`                              |
-| `i_dstart`      | InferenceProvider   | `''`                              |
-| `i_dend`        | InferenceProvider   | `''`                              |
-| `i_optimal`     | InferenceProvider   | `''` (truthy = hide non-optimal)  |
-| `i_label`       | InferenceProvider   | `''` (truthy = show point labels) |
-| `i_nolabel`     | InferenceProvider   | `''` (legacy, read-only)          |
-| `i_hc`          | InferenceProvider   | `''`                              |
-| `i_log`         | InferenceProvider   | `''`                              |
-| `i_legend`      | InferenceProvider   | `''`                              |
-| `i_advlabel`    | InferenceProvider   | `''`                              |
-| `i_gradlabel`   | InferenceProvider   | `''`                              |
-| `e_rundate`     | EvaluationProvider  | `''`                              |
-| `e_bench`       | EvaluationProvider  | `''`                              |
-| `e_hc`          | EvaluationProvider  | `''`                              |
-| `e_labels`      | EvaluationProvider  | `''`                              |
-| `e_legend`      | EvaluationProvider  | `''`                              |
-| `r_range`       | ReliabilityProvider | `last-3-months`                   |
-| `r_pct`         | ReliabilityProvider | `''`                              |
-| `r_hc`          | ReliabilityProvider | `''`                              |
-| `r_legend`      | ReliabilityProvider | `''`                              |
+`buildShareUrl()` preserves the exact current pathname, locale prefix, dynamic slug, and
+hash. It filters parameters using `shareParamScopes` from the canonical dashboard route
+registry and carries the canonical `unofficialruns` value from the live URL.
+
+### Parameter registry
+
+`UrlStateKey`, `URL_STATE_KEYS`, and `PARAM_DEFAULTS` in
+`packages/app/src/lib/url-state.ts` are the exhaustive parameter source of truth.
+Dashboard scope membership is declared by `shareParamScopes` in
+`packages/app/src/lib/dashboard-routes.ts`. Tests enforce completeness and route-specific
+share behavior, so this document deliberately does not duplicate a manually maintained
+parameter table.
diff --git a/docs/tco-calculator.md b/docs/tco-calculator.md
index 2002753d3..d25a1c1ce 100644
--- a/docs/tco-calculator.md
+++ b/docs/tco-calculator.md
@@ -53,16 +53,42 @@ Badges change based on metric because showing power badges when the metric is "C
 - **Cost metric**: TCO $/GPU/hr badges (assumed hourly rates per GPU, sourced from SemiAnalysis AI Cloud TCO Model)
 - **tok/s/MW metric**: Power/GPU badges (assumed power draw per GPU, sourced from SemiAnalysis Datacenter Industry Model)
 
-## Why No Separate Context Provider
+## Provider and State Ownership
 
-The calculator reuses `GlobalStateContext` (model, run date) and `InferenceChartContext` (sequence, precisions). Calculator-specific state (cost provider, token type, bar metric, target interactivity, selected bars) is local `useState`.
+The calculator route parses its URL seed on the server and mounts the sole
+`GlobalFilterProvider` for calculator model, sequence, precision, date, and run state.
+`DashboardShell` deliberately does not mount a second global provider on this route.
 
-Adding another context provider to the nesting hierarchy would increase re-render surface for unrelated tabs. Since calculator state doesn't need to be shared, local state is simpler and more performant.
+Calculator-specific visibility, target/input, metric, and selected-bar transitions are
+owned by one local reducer. Derived ranges and valid selections are computed from current
+data rather than copied back through reconciliation effects.
 
 ## Bar Selection & Comparison
 
 Click-to-compare uses `resultKey` (not hwKey) because multi-precision mode produces multiple bars per GPU. Comparison ratios use the lower value as denominator (ratio >= 1.0). Both metric and token type are reflected in the comparison text to avoid ambiguity.
 
+## Folding sections away (`CollapsibleSection`)
+
+The page is long — chart, cost target, lifecycle — and most readers want one of them
+at a time. The chart section and Interactivity Within a Cost Target each carry a
+chevron toggle (`components/ui/collapsible-section.tsx`).
+
+Two decisions in that component:
+
+- **Folded means unmounted, not hidden.** These bodies hold D3 charts and tables, and
+  leaving them mounted under `display: none` keeps them measuring and re-rendering
+  off-screen for a reader who said they are not interested. The trade is that chart
+  zoom and a frozen tooltip do not survive a fold.
+- **`titleWhenOpen`.** The two fleet-planner sections own their `<h2>`, so the header
+  shows the title always. The chart section's title lives in the chart's own caption
+  (and is part of the PNG export), so its header shows the title **only while
+  folded** — two copies would be worse than none. That is why `getChartTitle` is
+  hoisted out of the caption in `ThroughputCalculatorDisplay`: the fold header needs
+  the same words, and they carry the current metric and target.
+
+State is local `useState`, not a URL param: it is a reading preference, not a view of
+the data worth sharing in a link.
+
 ## Unofficial-Run Overlays (`?unofficialrun=`)
 
 A loaded unofficial run contributes an extra bar per (hardware × run) to the bar chart, in the run's palette color (`overlayRunColor`) and labeled `B300 (✕ my-branch)`. The label keeps the branch inside the same paren group as the precision so the `twoRowYAxisLabels({ split: 'parens' })` y-axis customizer still splits it into two rows.
@@ -102,7 +128,7 @@ The reset's early-out guards on the **merged** list, not the official one. An em
 - **Clamped values.** `interpolateForGPU` clamps the target into each series' measured range and always returns a value, so a bar can be showing its nearest edge point rather than an interpolation. This is pre-existing across GPUs with different ranges, but widening the slider to cover overlay operating points makes it reachable for every official bar at once — which would turn a side-by-side overlay delta into a real-vs-clamped comparison. Results carry a `clamped` flag and the tooltip says so. (Narrowing the slider back is not the fix: it only moves the clamping onto the overlay bars, and an overlay-only model loses its bounds entirely.)
 - **Escaping.** The tooltip is a hand-built HTML string injected with `.html()`, and branch names and run URLs come from the GitHub API for whatever run id the user pasted. Everything untrusted goes through `escapeHtml` (`lib/utils`). The y-axis tick labels render the same branch but go through d3 `.text()`, and the legend entry is React — both already safe.
 
-The calculator supports both fixed-sequence and Agentic Traces scenarios through
+The calculator supports both fixed-sequence and Agentic scenarios through
 the shared `rowToSequence` classifier. Agentic rows carry null `isl`/`osl`, so
 filtering them by numeric sequence lengths would silently drop every point.
 
@@ -141,6 +167,28 @@ at every interpolated point.
 compare interpolation models on a fixed snapshot, but they must not be presented
 as permanent impact figures for the changing live dataset.
 
+The `/inference` page also exposes tokens-per-dollar and normalized token revenue
+as separate Y-axis metrics; neither replaces the cost-per-million metrics.
+Historical trends for these proportional metrics select Pareto knots from the
+matching throughput, spline that throughput, and apply the constant multiplier.
+Tokens-per-dollar uses `3600 / $/GPU-hr`; normalized token revenue uses
+`3600 / 1,000,000` at the axis's explicit `$1/M tok` sale price. Reusing the
+matching throughput frontier is essential: total-throughput knots are not
+necessarily the output- or input-throughput Pareto knots.
+
+The revenue axis defaults to pricing every input and output token at the same
+`$1/M tok` rate, making it a model-independent SLA comparison. Its OpenRouter
+option fetches the selected model's current public prompt/completion prices from
+`https://openrouter.ai/api/v1/models` and prices the streams separately. Aggregate
+rows use their measured input/output split. Disaggregated rows cannot use the raw
+per-prefill/per-decode rates together, so they apply the fixed ISL:OSL shape or
+the measured agentic prompt:generation mix to total tok/s/GPU. The OpenRouter
+option uses standard prompt/completion prices and does not apply cache discounts.
+
+Neither mode is the fleet lifecycle section's realized-revenue model: it does not
+model availability, rollout, or a user-supplied cached-input discount. Those
+business assumptions remain in Fleet Lifecycle below.
+
 ### The consistency guard
 
 `recoverReciprocalNumerator` returns the constant only if **every** usable point
@@ -154,3 +202,799 @@ The rate is recovered across **all three token types at once** (`recoverCostRate
 Checking one family alone and falling back to another would recover a rate from
 output tokens and then apply it to total throughput; the existing
 `maxInteractivityAtCost` tests caught exactly that mistake.
+
+## Fleet Lifecycle (`FleetLifecycle.tsx`)
+
+> **Now on its own page.** The section no longer renders on `/calculator`; it is
+> the body of the `/fleet` route (and `/zh/fleet`), hosted by
+> `FleetLifecycleDisplay.tsx`, which owns the same model/scenario/precision/
+> cost-tier/target controls the calculator page does and renders the chip
+> legend in its controls card, where it stays reachable from the table view and
+> when nothing is plottable. The calculator keeps a pointer card linking to `/fleet`, and
+> its cost-target panel now renders the MW budget input (`c_mw` is shared
+> between the two pages, so a budget set on one seeds the other). Everything
+> below about the section's internals still applies unchanged.
+
+Everything above answers "which chip is cheapest at this interactivity, right
+now?". This section answers "what has a fleet of it earned and cost since the
+model shipped?" — and the answer is measured, not assumed.
+
+### Why it absorbed Fleet Projection, and why the cost target moved instead
+
+The page used to carry three fleet sections: **Fleet Projection** (a MW budget →
+chips, fleet tok/s, concurrent users, $/hr, $/mo), **Interactivity Within a Cost
+Target**, and this one. Two of those three are now one.
+
+Fleet Projection was folded in here because it was half of this section already.
+Its MW input was this section's _primary_ input — the empty state read "enter a
+facility power budget in the Fleet Projection section above", which is a control
+in one section gating a table in another. Its `Fleet $/hr` and `$/mo` are this
+section's `Cost $/day` in other units, and its `Fleet tok/s` is `tok/s/MW now`
+times the budget. What it had that this did not was the physical sizing: **chip
+count** and **concurrent users**. Those are now columns, so the budget, the sizing
+it produces and the economics that rest on it read across one row.
+
+Two consequences worth knowing:
+
+- **The MW input renders outside `body()`.** Every other control here only means
+  something once a fleet exists, but this is the control that brings one into
+  being, so it must be present in the empty state — otherwise the section
+  deadlocks on its own precondition.
+- **"Nothing to plot" needed splitting.** Fleet Projection had a dedicated message
+  for a budget too small to power one chip. Falling back to this section's
+  `noneMeasured` would have blamed the interactivity slider for a budget problem
+  and sent the reader to the wrong control, so an unsizeable budget (`unplottable`
+  non-empty with no rows) says so in its own words.
+
+The cost target was **not** folded in, because it is not a slice of this section at
+any point on the axis: it is the _inverse_ of the target-interactivity slider. The
+slider fixes a speed and the cost falls out; the cost target fixes a cost ceiling
+and the speed falls out. Nothing on a lifecycle answers "how fast can I go under
+$0.50/M tok". So it moved rather than merged — out of the fleet-economics stack and
+up to sit directly beneath the slider it inverts (`CostTargetPanel.tsx`), which is
+also why it is deliberately independent of the slider's current value: it reads the
+unfiltered frontier, so a chip that interpolates to nothing at the present target
+still gets an answer.
+
+### Why it reads the full run history
+
+The rest of the calculator reads one run date, so a chip is credited with
+whatever its latest sweep found. Measured against production history for dsr1
+8k/1k (7,715 rows, 134 dates), roughly 37% of configs have their best read at a
+typical target on an **earlier** date, worth up to ~3.6× — usually because a
+later sweep explored a different part of the space rather than because the chip
+got worse. For a lifecycle projection the honest number is the best config the
+chip has ever demonstrated, so this section is the one place that reads
+`/api/v1/benchmarks/history` instead of `/api/v1/benchmarks`.
+
+### One line per chip, not per config
+
+`hwKey` encodes the software, so a single piece of silicon appears in the history
+many times over: `b200_trtllm`, `b200_sglang`, `b200_sglang_mtp` and so on are all
+one B200. Drawing a line each says a fleet operator ran seven fleets, when they ran
+one and kept re-deploying it onto whichever config was ahead. `mergeProgressionsByChip`
+therefore collapses a chip's hwKeys into their **upper envelope**, and the winning
+config becomes part of what each step reports (`Config Now` in the table, named per
+rung in the tooltip).
+
+Three consequences worth knowing:
+
+- **Disagg competes for the same line.** Disaggregated configs report throughput
+  per decode or prefill chip, so a step won by one is not sized on quite the same
+  basis as a step won by an aggregated config; an earlier iteration split them into
+  two lines for that reason. They are pooled by product decision — the operator's
+  question is what the silicon can be made to do — so the caveat travels with the
+  step instead: `ChipProgression.disagg` flags a chip whose progression involves
+  any, the amber note explains the basis, and the config named on each step says
+  which kind won it. Per-line dashing was dropped when this changed, because a
+  single line can now contain both and a dashed stroke would assert something
+  untrue of half of it.
+- **The legend still filters hwKeys**, one level below the lines — hiding a config
+  removes it from candidacy. Because the legend isolates on click, isolating a
+  single config leaves at most one chip, and none at all when that config was never
+  measured at the target. That is a real state, and `calculator-lifecycle.cy.ts`
+  asserts it rather than assuming a row always survives.
+- **Colour resolves from the winning hwKey**, not the base GPU. The palette is
+  built over the _active_ hwKeys, so a bare base key falls through to the fallback
+  grey — which silently drained the colour out of every series when first wired.
+
+Cost stays flat across the merge because power and $/chip/hr come from the base
+GPU: it is the same silicon whichever framework is in front.
+
+`bestSoFarProgression` supplies the per-hwKey input: each hwKey's **running
+maximum** over run dates — the sequence of dates whose
+interpolated read at the target beat every date before it. That progression is
+the section's subject. A sweep that failed to beat the incumbent is not a step,
+because the fleet kept serving the config it already had, and the last rung of
+the staircase is by construction the same all-time best the table reports
+(`selectBestFromGroups` and `bestSoFarProgression` share one selection basis, so
+the headline figure can never disagree with the plotted line).
+
+That is only defensible with two rules, both in `historical-best.ts`:
+
+- **Clamped reads are discarded, not clamped.** `interpolateForGPU` always
+  returns a value, clamping the target into each frontier's measured range.
+  Searching every date multiplies the chance of picking up such an edge read —
+  at target 20 tok/s/user, 63% of naive winners are clamped, which would credit
+  a sweep's peak throughput at an interactivity it never served. This section
+  therefore applies the no-extrapolation rule (the same one
+  `useInterpolatedTrendData` uses) and lists the hwKeys it consequently has no
+  read for, with their measured ranges. Chips are never silently dropped: at a
+  low or extreme target that empty list _is_ the finding.
+- **Provenance travels with every number.** Each row renders its winning date
+  and links its run, because the run date stamped above the chart no longer
+  describes these numbers. A row whose winner beat the latest date is marked.
+  There is no per-row trust annotation in the data — `PURGED_RUNS` is
+  destructive so retracted runs cannot win, but a merely-superseded run can now
+  win permanently — so the run link is the audit trail.
+
+Precisions are pooled into one frontier per hwKey: the question is what the
+chip's best config does, and precision is part of the config.
+
+### Two-stage memoization
+
+`groupHistoryByHwKeyAndDate` (rows → one sweep per hwKey per date) is separate
+from `selectBestFromGroups` (read every frontier at the target) so moving the
+interactivity slider re-reads without rebuilding ~790 frontiers, each of which
+costs ten splines. For the same reason the hook computes **every** hwKey and the
+component filters by `visibleHwKeys` for display — filtering in the data layer
+would rebuild every frontier on each legend toggle.
+
+### `view=calculator` on the history route
+
+The history response is ~8.6 MB per model. `useHistoricalBest` gates the fetch
+on a facility power budget being set, and requests `view=calculator` to trim it
+to the calculator's metric allowlist (~24% smaller).
+
+**That trim must stay opt-in.** `CALCULATOR_METRIC_KEYS` excludes the
+measured-power metrics Historical Trends plots, so applying it to every history
+response would silently blank those charts. `measured-power-overlay.cy.ts` is
+the regression guard.
+
+### Lifecycle math (`lifecycle.ts`)
+
+Pure and React-free, like `fleet.ts`. It takes a `ThroughputStep[]` — one entry
+per measured improvement — and turns it into a revenue staircase over calendar
+time against a flat cost line:
+
+```
+revenue │              ,──── each config rolls out to its own numbers
+        │         ,───╯
+        │    ,───╯
+────────┼──╱──────────────────────────────  cost: flat, the racks
+        │ ╱ the first config rolls out from zero
+        └──────────────────────────────────▶  months since model release
+```
+
+An earlier iteration modelled a TTFI delay, a ramp to a flat plateau and a
+decommissioning taper, with no measured steps at all. That collapsed each chip's
+optimisation history into a single number held constant for years, so every chip
+drew the same assumed shape. The steps are measured; only the ramp survived, and
+it survived as an explicit user assumption.
+
+**A config is rolled out, not switched on.** So every config gets its own ramp:
+when it lands, the fleet climbs from whatever it was already serving to that
+config's numbers over `rampMonths`. The first config climbs from zero — nothing is
+being served before it lands.
+
+A rollout starts from the level the previous one **actually reached**, not from the
+previous config's nominal rate. With a ramp longer than the gap between sweeps that
+matters: the second config picks up mid-climb instead of jumping, so the line stays
+continuous no matter how the ramp length and the sweep cadence interact.
+
+Four conventions the numbers depend on:
+
+- **Cost is flat, including through every rollout.** It is `chips × $/chip/hr`, and
+  neither term moves when a config rolls out — racks bill from the moment they are
+  energised, not from the moment they are loaded. So the opening rollout is paid for
+  in full while it earns its way up from zero, which is what puts the first months
+  below the rule and gives payback its meaning.
+- **The ramp is an assumption; the steps are not.** It defaults to half a month
+  (`c_ramp`) and moves in quarter-month increments; 0 means every config takes
+  effect instantly.
+- **Markers sit at the release instant** — the foot of each rollout, not its top —
+  so every dot on the chart is a sweep the user can open.
+- **Interrupts are an availability haircut, not drawn events.** A 24-day MTBI
+  over a multi-year window is thousands of interruptions; at any sane chart width
+  each is far under a pixel, so drawing them yields aliasing noise rather than
+  information. They scale revenue via `mtbi / (mtbi + recovery)`.
+
+A blank or zero MTBI means "no interruptions modelled", not "always down" — the
+input is an optional refinement and a hostile default would be worse than none.
+
+### Anchoring and the x axis
+
+Time is measured from the model's release date, from `MODEL_RELEASE_DATES` in
+`@semianalysisai/inferencex-constants` (DeepSeek-V4-Pro: 2026-04-24, confirmed
+against `/api/v1/availability`, whose first `dsv4` row is that date). A model with
+no release date on file falls back to its earliest measured run, and the caption
+states which anchor is in use. A chip's line starts at its **own** first measured
+run, not at the anchor: before that there is no data, so there is no line.
+
+The horizon defaults to the measured window rather than a fixed number of months,
+and stops re-seeding once the user edits it. A fixed 60-month default left ~80% of
+the chart as flat tail extrapolated from the last sweep.
+
+The time scale passes `nice: false`. `buildScale` niceing is on by default
+(`scale-builders.ts:35`), which rounds a multi-year span out to whole years and
+leaves dead space before the release date and after the horizon; both ends of this
+domain are meaningful.
+
+The zero rule is a `type:'custom'` layer, and it removes `.lifecycle-zero-rule`
+before appending. The chart re-renders into the same zoom group, so appending
+unconditionally leaves a stale rule and a duplicate "break-even" label behind at
+the previous y-scale on every data change.
+
+### Y-axis selector (margin / revenue / cumulative revenue)
+
+`c_ly` switches the chart between margin/day (default), revenue/day, and cumulative
+revenue. The table is unchanged — it already carries both rates and cumulative
+margin — so this only affects the plot.
+
+Three things move with it, and they all matter for honesty:
+
+- **The break-even rule is drawn only for margin.** Zero is break-even on a margin
+  axis; on a revenue axis it is just the bottom of the scale, and each chip breaks
+  even at its **own** cost line rather than at zero revenue. Labelling the zero
+  gridline "break-even" there would be false, so `renderZeroRule` returns early.
+  Per-chip cost rules would be the honest equivalent but add one horizontal line
+  per series, which is why they are not drawn. Cumulative revenue is the same case:
+  a running total that starts at zero and only grows.
+- **The readout's value column follows the plotted quantity**, with cumulative
+  margin beside it, so it never depends on the axis to be read.
+- **Cumulative revenue is accumulated, not derived.** `LifecyclePoint` carries a
+  second running total, `cumulativeRevenue`, integrated by the same trapezoid over
+  the same intervals as `cumulative` (cumulative margin). It is tempting to
+  reconstruct it as `cumulative + costPerDay × elapsed`, and that identity holds
+  only when every interval carries the same flat cost — which is true today but is
+  a property of the assumptions, not of the shape. Accumulating both keeps the two
+  curves exactly one flat cost apart by construction, which is what makes them
+  comparable, and `isCumulative(metric)` is what the axis formatting and the
+  break-even suppression key off rather than a per-metric special case.
+
+Revenue mode is for comparing rollout _shapes_ across chips whose costs differ a
+lot; a chip sitting higher there does not mean it is more profitable. Cumulative
+revenue compounds that caveat — a chip with the largest area under its revenue
+curve may still never have covered its cost. The control's tooltip says so.
+
+### One line layer, interpolated linearly
+
+Because every riser is now a sampled rollout curve and every plateau is flat, the
+shape is already in the points: `curveLinear` draws exactly what was computed. A
+step curve would flatten the rollouts back into stairs, and a spline would overshoot
+on the near-vertical ones (a zero-length ramp emits two samples at the same month to
+force a true vertical).
+
+An intermediate version drew the ramp and the steps as two `line` layers with
+different curves. **That does not work, and the reason is worth recording:**
+`useD3ChartRenderer` renders every layer into one shared `renderGroup`, and
+`renderLines` does a keyed join on `.line-path`. Two line layers with the same
+series keys join against each other's paths, and the second silently overwrites the
+first — one of the two segments just vanishes. Before adding a second layer of an
+existing type to a `D3Chart`, remember that the shared render group makes layer keys
+non-isolating; a `type:'custom'` layer with its own class is the way out.
+
+### End-of-line chip labels
+
+Each line is named by its chip at its right end, in the line's own colour, so a
+series can be identified without crossing back to the sidebar legend. Two details
+matter:
+
+- The labels sit **past the plot width** (`x = width + 6`, with `CHART_MARGIN.right`
+  widened to hold them), and `clipContent` defaults to true — layers render into a
+  clipped zoom group, which would erase them entirely. They are drawn into
+  `ctx.layout.g`, the unclipped parent, instead.
+- End values cluster, so labels are placed by a greedy pass that pushes each label
+  down to keep `LABEL_MIN_GAP`, then slides the whole block up if it has run off the
+  bottom. A line ending at the very top or (after a zoom) outside the visible range
+  is pinned to the nearest edge rather than allowed to escape the SVG.
+
+Under an x zoom or pan, a line's last data point can sit well outside the plot, so
+the label is placed at the value where the line **crosses the visible right edge**
+(interpolated on the straddling segment), not at its final value. A line that ends
+inside the plot is labelled at its own end instead, and one panned entirely out of
+view is not labelled at all. Getting this wrong is not cosmetic: the two chips that
+are still climbing at the horizon read several million dollars a day lower at a
+zoomed-in edge than at their plateau.
+
+Both the labels and the break-even rule are `type:'custom'` layers, and **both
+declare `onZoom` as well as `render`**. A custom layer without `onZoom` stays pinned
+to the base scales while the lines move, so a zoomed chart shows break-even — or a
+chip's name — at the wrong height. This only matters because the chart passes
+`zoom={{enabled: true, axes: 'x'}}`; it shipped for a while with the wrapper's
+default caption promising shift+scroll zoom and no `zoom` prop at all, which made
+both `onZoom` handlers dead code and the caption a lie. A Cypress test now asserts
+the axis actually moves. Zoom is x-only on purpose: rescaling y would slide
+break-even under the reader.
+
+The custom layers read the current metric and line data through a **ref**, not
+through the closure they were created in. The zoom behaviour installed by a render
+captures that render's callbacks, and d3's double-click reset is a transition, so its
+trailing events can fire after a later render has already redrawn — a stale callback
+then re-appends what the new render removed. Switch to the revenue axis during a
+reset animation and the break-even rule used to come back and stay, because nothing
+renders again afterwards.
+
+**Known limitation, in the shared renderer rather than here:** those same trailing
+events also repaint the axes and grid from the scales captured when the transition
+started, so changing the metric inside the ~750ms reset window leaves a stale y axis
+until the next interaction. Fixing it means giving `useD3ChartRenderer` a
+latest-context ref, which every zoomable chart shares; it is deliberately not done as
+part of this section. The Cypress spec waits out the transition for this reason.
+
+### The hover readout: one rule, every chip
+
+Hovering anywhere in the plot draws a vertical rule and reads **every** plotted chip
+at that instant into one popup; clicking freezes it, and clicking again releases it.
+This replaced per-dot tooltips, which answered "what is this dot?" — the wrong
+question for a chart whose lines are only interesting against each other. Comparing
+two chips at one date meant hovering twice and holding the first number in your head.
+
+It is built on `TooltipConfig`'s `proximityHover` + `getDataX`, so the shared
+renderer's overlay rect owns hover for the plot area. Three consequences worth
+knowing:
+
+- **The hover positions are a uniform grid** (`HOVER_SLICE_COUNT` slices across the
+  window), not the sample points. Sample density varies by two orders of magnitude
+  along a line — 24 samples across each rollout ramp, one across a multi-month
+  plateau — so snapping to samples makes the cursor crawl through a riser and then
+  jump a year. Values come from linear interpolation between the bracketing samples,
+  which is exactly what `curveLinear` draws, so the number always matches the pixel.
+  A chip has no row before its own first measurement rather than a fabricated one.
+- **Steps match by proximity, not equality** (`STEP_MATCH_SLICES`). A slice is a
+  fraction of a pixel, so a step's instant essentially never coincides with one;
+  matching exactly would make the run links unreachable by pointing at the dot,
+  which is the only place a reader looks for them.
+- **A step's config detail is frozen-only.** Hovering is for scanning, and the popup
+  keeps one shape wherever the cursor is; a step block appearing as the cursor
+  crosses a dot reflows the rows under the reader and buries the comparison they
+  came for. A click means "tell me about this instant", which is where the config,
+  its gain over the first run, and its run links belong.
+- **The dots no longer own hover.** The overlay rect sits above them, so a Cypress
+  test cannot click a `.dot-group` — it clicks the overlay at the dot's x instead.
+
+Freeze is a toggle, which the shared overlay handler cannot express on its own: it
+pins on every click, so a second click would just re-pin. A capture-phase listener on
+the SVG records whether the readout was already frozen _before_ the pin lands, and
+`onPointClick` dismisses when it was. Reading the pinned state afterwards would
+always say "frozen".
+
+Two testing notes, both about the portal tooltip rather than this chart. It is keyed
+`data-chart-tooltip="<chartId>"` so a page with several charts can be asserted
+per-chart. And `:visible` does not work on an **unfrozen** readout: Cypress treats a
+fixed-position element as hidden when something else answers `elementFromPoint` at
+its centre, and an unfrozen readout sets `pointer-events: none`, so the plot answers
+instead. Assert `display` instead. Relatedly, any toast nudge re-renders the chart
+tree and rewrites the portal's inline styles, wiping an open hover readout — app-wide
+behaviour, which the spec sidesteps by suppressing the timed nudge.
+
+Every rung links its run in the frozen readout, not just the opening and closing
+sweeps the table links. Intermediate rungs are exactly where a superseded-but-never-
+purged anomalous run would sit, so leaving them unlinked would have removed the audit
+trail at the one place the trust argument depends on it.
+
+The rule's group is `.lower()`ed on every draw. `renderLines` keeps its paths across
+re-renders through a data join while this layer removes and re-appends, so without
+that the dashed rule climbs above the lines after the first repaint.
+
+### Known visual limitation: the palette is not colourblind-safe here
+
+An adversarial review ran the series colours through a CVD validator and the result
+is worth recording rather than hiding. With the default fleet the chart draws four
+NVIDIA chips plus one AMD chip, which under the site-wide vendor convention
+(`dynamic-colors.ts` gives NVIDIA the green hue zone and AMD the red one) means four
+greens and a red. Seven of ten adjacent pairs fail: worst are MI355X↔B200 at ΔE 1.2
+for protanopia — indistinguishable, and those two lines cross mid-plot — and
+GB200↔GB300 at ΔE 7.7 for normal vision.
+
+This is a property of the vendor convention every chart and legend shares, not of
+this section, so it is deliberately **not** worked around locally: giving this one
+chart its own palette would desync it from the sidebar legend and from every other
+calculator chart. Two things are unavailable as local mitigations too — marker shape
+already encodes _precision_ app-wide and there are only four shapes for five-plus
+chips, and `POINT_SIZE` is a shared constant with no per-layer override, so the dots
+cannot be enlarged to the 8px spec here alone. What the chart does carry is a
+non-colour decoder: every line is named at its right end. Fixing the palette
+properly means re-stepping the vendor zones in `dynamic-colors.ts` for all charts.
+
+Related, and also left alone: the rollout ramps and the post-last-sweep hold are
+assumptions drawn in the same stroke as the measured window, with only the step dots
+marking measurements. The ramp input and the notes say so in words; distinguishing
+them visually was considered and declined to keep one line per chip readable.
+
+### Token price defaults to break-even
+
+Break-even is per-chip, so a single global price input needs one anchor: the
+**cheapest visible chip's** break-even, i.e. the competitive floor at which that
+chip earns exactly nothing and everything pricier is underwater. It re-seeds when
+the tier, token type or target changes, and stops the moment the user edits the
+field (or arrives with `c_price` in the URL); a reset link restores it. The
+figure comes from the TCO model — everything above that line is the user's
+assumption, and the section says so.
+
+It is derived as `costPerHour / fleetTokPerSec` for that fleet, so the plotted
+margin is exactly zero at the default — for that chip, at its **latest** config.
+Under the default price every other series therefore sits below the rule, and each
+one's own earlier steps sit lower still. That is the competitive floor being read
+correctly, not a scaling bug; raising the price lifts the whole fleet.
+
+That is deliberately **not** the `$/M tok` the calculator's cost bars show for
+the same config, and the reason is worth recording because it is a property of
+the existing calculator rather than of this section.
+
+`interpolateForGPU` splines every metric independently. Per point,
+`cost = $/GPU-hr / (tput x 3600)` — cost is a **convex** function of throughput —
+so splining the cost values directly lands above the curve implied by splining
+throughput (Jensen), by a gap that widens with knot spacing. Measured over the
+captured fixture across every frontier and every date:
+
+| target sits…         | n   | median ratio | p95    | max     | share biased high |
+| -------------------- | --- | ------------ | ------ | ------- | ----------------- |
+| exactly on a knot    | 470 | 1.0000       | 1.0000 | 1.0000  | 0%                |
+| midway between knots | 307 | 1.0568       | 3.9661 | 25.3237 | 73.6%             |
+
+(ratio = splined `result.cost` / cost derived from the splined throughput.)
+
+Agreement at every knot and one-sided divergence between them identifies this
+as interpolation bias, not a modelling disagreement.
+
+Scored against the oracle, the splined read is simply **less accurate**.
+`/inference` plots cost only at measured points, as `specs.costh / tokensPerHour`
+(`lib/chart-utils.ts:380`, `roof: false`). Holding out each interior frontier
+knot in turn, rebuilding the frontier from the rest and predicting the held-out
+point's real cost:
+
+| method                          | mean err | p50   | p90    | p99     | closer to oracle |
+| ------------------------------- | -------- | ----- | ------ | ------- | ---------------- |
+| splined `result.cost`           | 162.8%   | 86.3% | 528.1% | 1402.1% | 36 / 144         |
+| derived `k / interpolated tput` | 42.2%    | 23.0% | 72.0%  | 245.3%  | **108 / 144**    |
+
+The structural argument is stronger still. That oracle construction makes
+`cost x throughput = $/GPU-hr` an identity at every real point. Splining the two
+independently breaks it: the calculator's pair violates it by a median 71.6% and
+up to 2026%, i.e. it reports an operating point no real config could occupy.
+Deriving cost from the interpolated throughput satisfies the identity by
+construction.
+
+This section therefore derives break-even from throughput, which also keeps it
+consistent with the fleet planner directly above (fleet cost is
+`gpus x $/GPU-hr`, a constant with no interpolation in it, so no other price can
+zero the plotted margin).
+
+**Follow-up worth taking.** The same defect affects the calculator's own cost
+bars and table, not just this section: `interpolateForGPU` should build `cost`
+from its interpolated throughput rather than splining the cost values. That is a
+small change, but it lives in a Python-synced function (`iso_interactivity.py`)
+and would move published cost numbers, so it needs its own change. Until then the
+price tooltip states the direction of the divergence.
+
+### Input and output are priced separately
+
+Revenue is `input x inputPrice + output x outputPrice`, not one blended price
+against one throughput. The two streams are wildly unequal and the split is
+measured, not assumed — `input_tput_per_gpu` and `output_tput_per_gpu` are
+already on every row, the same figures `/inference` plots:
+
+| scenario       | input : output throughput                                                    |
+| -------------- | ---------------------------------------------------------------------------- |
+| fixed 8k/1k    | 8.0 : 1 (p10 7.9, p90 8.0 — it is the sequence shape, not per-config spread) |
+| agentic traces | ~130 : 1 (p10 114, p90 145)                                                  |
+
+At a 4x output premium on 8k/1k, output is a ninth of the tokens and a third of
+the revenue, so a blended price is not a rounding difference. Three consequences:
+
+- **Revenue no longer depends on the Input/Output/Total cost-type selector.** The
+  fleet sells everything it produces; the selector still drives the cost matrix,
+  the bar chart, and which config wins the staircase. This also retires a live
+  trap — on `costType=output` the old model counted revenue from a ninth of the
+  tokens while cost covered the whole rack.
+- **Break-even is a line, not a point.** Any (input, output) pair on it zeroes the
+  margin. The section resolves that by fixing the ratio currently in the two
+  fields and solving the input price against `effectiveTokPerSec` —
+  `billableInput + multiple x output` — so a reset scales both and lands back on
+  the line at the ratio the user chose. Unseeded, the pair starts at 4x, which is
+  roughly where the major vendors price (DeepSeek publishes $0.27 / $1.10).
+- **Gain stays a token multiple.** `improvementFactor` counts both streams and no
+  price, so it remains a statement about the hardware. The fixture that pins this
+  grows the streams _disproportionately_ — input flat, output tripled — because a
+  proportional one cannot tell "counted both" from "counted input only".
+
+Cached input tokens are discounted into the input stream before pricing (see
+below), so the input price is the fresh-token price.
+
+`c_price` is the input price — unchanged, so a link written before the split
+still seeds the field it always seeded — and `c_oprice` the output price. When a
+legacy link has `c_price` alone, the missing output price is derived at the
+default 4x ratio without overwriting the explicit input price; the symmetric
+rule completes a manually authored output-only link.
+
+### The disaggregated token-split trap
+
+`input_tput_per_gpu` is per **prefill** chip, `output_tput_per_gpu` per **decode** chip,
+and `tput_per_gpu` per chip **overall**. Three denominators. On an aggregated run they
+coincide and the two rates sum to the total on all 937 production rows. On a
+disaggregated run they do not, and the gap is large: across 308 disagg rows the two
+rates sum to between 1.00× and **16.11×** the total, tracking
+`(prefill + decode) × (isl/prefill + osl/decode) / (isl + osl)` to four decimals.
+
+Pricing the two streams off those rates therefore billed a disaggregated config for
+tokens its chips never served. The visible symptom was a **dip**: MI355X on
+DeepSeek-V4-Pro 8k/1k stepped from `mori-sglang` (disaggregated, 2026-07-22) to
+`atom_mtp` (aggregated, 2026-07-28) with a _higher_ total tok/s/MW — a legitimate
+best-so-far rung — and the margin line fell, because the outgoing config had been
+credited with ~2× its tokens. Eleven such dips existed across 215 step transitions on
+eight interactivity targets; every one was a handover where the prefill:decode split
+changed or disappeared.
+
+`tput_per_gpu` itself is **not** affected: it is per chip overall for both kinds, which
+reproduces exactly from the absolute rates and the chip counts — on the first disagg row
+in production, `(54.521×8 + 6.857×8) / 16 = 30.68886 = tput_per_gpu` to five decimals.
+So the ranking, the sizing and the cost line were always comparable across an
+aggregated→disaggregated handover; it was only the two per-stream rates that were not.
+The section's disagg note used to claim otherwise and has been corrected.
+
+One place had the same defect and is fixed alongside: **concurrent users**, which divided
+the per-_decode_-chip output rate into a whole-fleet chip count and so overstated the
+stream count by `(prefill + decode) / decode` — median 2×, up to 7× in production history.
+It now uses `outputTokPerChip`, which re-bases onto chips overall and is a bit-for-bit
+no-op wherever no share is known.
+
+What remains an assumption rather than a measurement: the winning config's prefill:decode
+ratio is taken to apply across the whole fleet from the moment it rolls out. A switch from
+aggregated to disaggregated serving is a redeployment, and the ramp window is what stands
+in for it. Chip counts also will not generally divide into the config's ratio, a remainder
+bounded by one config-unit that the model ignores.
+
+The fix is to use a **share of the per-chip total** wherever a token-type figure is
+needed, never the two rates directly, with the share recovered as:
+
+1. `input / (input + output)` when those agree with `tput_per_gpu` to within 1% — every
+   aggregated row, where the measurement is self-consistent and is the truth;
+2. otherwise `isl / (isl + osl)` — for a fixed sequence the mix is the sequence shape,
+   and no config can change it;
+3. otherwise the run's own `total_prompt_tokens : total_generation_tokens` — agentic
+   traces, which have no ISL/OSL;
+4. otherwise nothing is charged as input, which understates revenue rather than
+   inventing a mix.
+
+After the fix: 0 dips across the same 215 transitions, and the streams sum to the rate
+the fleet was sized and costed on by construction.
+
+**Selection needed the same correction, and it is a second bug.** Fixing revenue alone
+left the _ranking_ — which config wins each step — reading `inputTpPerMw` /
+`outputTpPerMw` raw, so a disaggregated config could win a rung purely for having been
+divided by fewer chips. Measured across seven interactivity targets: on total pricing the
+basis makes no difference (46 of 46 chip-winners identical, confirming `tput_per_gpu` is
+genuinely per total chip), but on **output** pricing 10 of 46 winners change and 5 are
+handed to a disagg config a comparable basis rejects; on **input**, 17 of 46 change.
+`mi355x_mooncake-atom@2026-06-12` won the output ranking at 25 and 30 tok/s/user on the
+raw basis; on a comparable one `mi355x_atom_mtp@2026-07-28` wins.
+
+So `getComparableTpPerMwForType` splits `tpPerMw` by the same share, and the
+lifecycle selection (`useHistoricalBest`) uses it.
+For an aggregated config it is an exact identity — the rates already sum to the total, so
+`tpPerMw × share` _is_ `inputTpPerMw` — so only disaggregated rows move.
+
+The bar chart and cost matrix above still show the raw per-prefill / per-decode figures
+with their existing disclaimer, deliberately: a chart the reader is told how to interpret
+can carry a footnote, and changing the headline chart's numbers is a separate decision.
+What is no longer acceptable is a _ranking_ that silently picks a winner on that basis. A corollary worth stating: on a fixed
+sequence the input/output split changes the **level** and the break-even price, but it
+cannot reorder chips, because the mix is the same for all of them. On agentic traces the
+mix is measured per config, so there it can.
+
+### Margin and revenue per megawatt
+
+Two y-axis metrics normalize the daily rates by provisioned power:
+`marginPerMw` (`c_ly=marginPerMw`) plots `revenue − TCO` in `$/MW/day`, while
+`revenuePerMw` (`c_ly=revenuePerMw`) plots revenue before TCO in the same unit.
+Two things about them are worth being explicit, because a reader could reasonably
+expect more than they deliver:
+
+- **They re-rank almost nothing.** Every chip in this section is sized to the same
+  power budget, so each per-MW rate is its unnormalized rate divided by very nearly
+  the same number for every series. The only spread comes from how completely each
+  chip's power density fills the budget. What the metrics buy is a figure that does
+  not move when the budget does — the unit a power-constrained plan is written in —
+  not a new ordering.
+- **The denominator is provisioned power, not the budget.** `chips × kW/chip ÷ 1000`,
+  because chip counts are whole and the remainder of the budget is stranded. Using
+  the typed budget would credit each chip with power it never provisioned. An
+  unusable figure (no registered power, a budget too small for one chip) leaves the
+  metrics at zero rather than dividing by it.
+
+Because the rescale is by a positive constant, zero on `marginPerMw` is still
+break-even. That is why `isBreakEvenAnchored` includes it and the chart retains
+the dashed rule. `revenuePerMw`, like unnormalized revenue, does not draw the
+rule: zero revenue is not break-even for a fleet with non-zero cost.
+
+### Agentic traces, and cached input tokens
+
+The section refused Agentic Traces until it was shown that the refusal's stated
+reason — "run history is keyed by input/output sequence length" — was no longer
+true. `/api/v1/benchmarks/history` takes `benchmarkType=agentic_traces` and drops
+the ISL/OSL filter server-side; `buildGpuGroups` already maps agentic rows and
+already takes a percentile. Only the fetch in `useHistoricalBest` had to change.
+
+What did need real work is the revenue term. Measured against production history
+(`inferencex.semianalysis.com`, DeepSeek-V4-Pro and GLM-5.2, August 2026):
+
+|                                  | agentic                        | fixed 8k/1k                    |
+| -------------------------------- | ------------------------------ | ------------------------------ |
+| median input:output token ratio  | **133:1**                      | ~8:1                           |
+| median GPU prefix-cache hit rate | **0.92** (min 0.005, p90 0.96) | metric **absent on every row** |
+| rows carrying a hit rate         | 217/223 DSV4, 45/45 GLM-5.2    | 0/1245, 0/541                  |
+
+Charging the input price against the raw input rate therefore bills ~13.5k tok/s
+of which ~92% are cache _reads_, which providers charge a fraction of the fresh
+input rate for — overstating agentic margin by close to an order of magnitude. At
+8:1 with no prefix reuse the distinction is immaterial; at 133:1 with 92% hits it
+is the dominant term.
+
+So the fleet is sized on the physical rate and the **input stream is discounted
+before it is priced**:
+
+```
+billableInputTokPerSec = inputTput × (1 − clamp01(cacheHitRate) × (1 − cacheReadRatio))
+```
+
+Selling `hit` of the stream at `ratio` of the price is exactly the same revenue as
+selling `hit × ratio` of it at full price, which is why one factor on the rate
+does the whole job and the input price stays the fresh-token price.
+
+Three properties are worth stating, because they are what make this safe:
+
+- **It is a subtraction, not a recomposition.** Writing it as
+  `output + billableInput` would re-derive the total from two independently
+  splined series and drift against it. As a subtraction the result is _exactly_
+  `base` whenever there is nothing to discount.
+- **Fixed sequences are unaffected by construction.** They carry no cache metric
+  on any row, so `cacheHitRate` is `undefined` and the term is zero — not "close
+  to zero", identical. That is asserted with `toBe`, not `toBeCloseTo`.
+- **A partly-measured frontier opts out.** `interpolateForGPU` splines the rate
+  only when _every_ frontier point carries one; substituting 0 for the missing
+  points would invent a dip in the cached fraction and overstate the billable
+  rate. Opting out bills every input token at full price instead — wrong in the
+  direction that understates margin.
+
+`cacheReadRatio` is the one user input here, labelled `Cached input (% of price)`
+and defaulting to 10% (the ratio DeepSeek and Anthropic both publish). The cached
+_fraction_ it multiplies is measured per config, not assumed. The control is
+hidden for fixed sequences rather than shown as a no-op. There is no circularity
+with the break-even seed: billable throughput depends only on the ratio, and
+break-even then solves the price given that throughput.
+
+**What counts as cached.** Three tiers are reported and they do not all stack, so
+the rule is conditional rather than a sum. Measured across 326 production agentic
+rows, each checked against its own `theoretical_cache_hit_rate` ceiling:
+
+| row shape                    | rows | `gpu+ext` > ceiling | `gpu+ext+cpu` > ceiling |
+| ---------------------------- | ---- | ------------------- | ----------------------- |
+| external absent, CPU absent  | 92   | 8                   | 8                       |
+| external absent, **CPU > 0** | 26   | 0                   | **0**                   |
+| external > 0, CPU > 0        | 106  | 1                   | **56**                  |
+| external 0, CPU 0            | 96   | 0                   | 0                       |
+
+The middle two rows carry the decision. Where an external rate is reported,
+adding CPU on top breaches the ceiling 56 times out of 106 — the router-side
+external figure already contains the offload tier, so summing double-counts it.
+Where no external rate is reported the CPU tier is real and disjoint: 0 breaches
+across all 26, every one of them an offload-on row. Hence `cacheHitRateOf` adds
+external when present and CPU only in its absence.
+
+Dropping the CPU tier outright, which this did until 2026-08-19, understated the
+cached share on those 26 rows by a median 5.54pp (p90 39.7pp, max 76.9pp — the
+median CPU rate is 0.055 but the tail is long). An understated cached share bills
+more input at the fresh price, so it **overstates** revenue: on those rows the
+input revenue leg was too high by a median 27%, p90 67%. That is the opposite
+direction from the all-or-nothing rule below, which deliberately errs
+conservative, and the reason this was worth a branch. Fleet-wide the median hit
+rate moves only 0.9255 → 0.9279, because 294 of 320 rows are unaffected.
+
+Two honest caveats on the rule. A reported external `0` is treated as a
+measurement that suppresses CPU rather than as an absence, on the reasoning that a
+router reporting zero external hits has still accounted for the offload tier —
+but **no production row currently has external `0` alongside a non-zero CPU rate**,
+so that branch is a deliberate choice about an unobserved shape, covered by a test
+rather than by data. Separately, `gpu` alone already breaches the ceiling on 8
+rows that report neither of the other tiers; that is a data-quality question about
+those rows, partly masked by the `[0,1]` clamp, and this rule neither causes nor
+fixes it.
+
+The clamp to `[0,1]` is not decorative — the GPU figure alone reaches 1.185 on
+rows without an external one.
+
+**Interpolating the cached fraction is safe, and this is the evidence.** The rate
+is splined along the frontier like any other metric, which raises the fair
+objection that two adjacent frontier points can describe different serving
+regimes. Measured by leave-one-out over 141 interior points of 26 production
+frontiers — drop the point, read its interactivity from its neighbours, compare to
+what was measured there:
+
+| metric                                 | median | p90    | max     |
+| -------------------------------------- | ------ | ------ | ------- |
+| cache hit rate, absolute error         | 0.0076 | 0.0631 | 0.4890  |
+| throughput, relative error _(control)_ | 8.60%  | 34.85% | 245.79% |
+
+At the median the hit rate is off by 0.8% relative, against 8.6% for the
+throughput the whole calculator already rests on — so the cached fraction is about
+an order of magnitude better behaved than the number nobody questions. Adjacent
+frontier points differ by under one percentage point at the median (p90 0.079,
+max 0.510). The wide 0.005–0.96 spread quoted above is across _all_ of a chip's
+configs, not along a single Pareto frontier, which is what the spline walks.
+
+**A limitation the UI discloses.** Agentic run history is roughly a week deep:
+DSV4 has 7 run dates with 4 of 19 configs measured on more than one; GLM-5.2 has
+4 dates and **0 of 9** configs on more than one. Fixed 8k/1k has 61 dates over
+3.7 months with 30 of 45 configs on several. Since this section plots a
+best-so-far staircase, agentic lines are often a single rung. When no visible
+chip has two, the section says so — a flat line here is a gap in the history, not
+a finding about the hardware. This resolves itself as runs accumulate.
+
+### Overlay exemption
+
+This section is official-only, and unlike the fleet planner that is not a policy
+choice: unofficial runs are not ingested, so `/api/v1/benchmarks/history` cannot
+serve them. Taking AGENTS.md's documented exemption; the section states the
+exclusion in its own note rather than leaving a silent gap.
+
+### What is still not apples-to-apples, and why it was left alone
+
+The bar chart and the cost matrix still show
+`input_tput_per_gpu` / `output_tput_per_gpu` on their reported per-prefill / per-decode
+basis, and their `$/M tok` with them. On the Input and Output token types a disaggregated
+config therefore reads **faster per chip and cheaper per token than it is** — a median 2×,
+up to 18× on input and 7× on output across run history. Total-token figures are unaffected.
+
+That was a deliberate decision not to move published numbers, taken after establishing
+that the two cannot be fixed independently: **per-token-type cost is not stored data.**
+`interpolateForGPU` recovers the constant `$/chip-hr × 1e6/3600` from the points and
+re-derives cost as `rate ÷ throughput` — verified on all 37 disaggregated sweeps in the
+fixture, where the rate is recovered every time and `cost × throughput` holds to 1e-6 on
+the interpolated result for all three token types. So:
+
+- re-basing the point's `costhi` has no effect, because `costInput` is recomputed from
+  `inputTputValue` downstream;
+- re-basing the divisor instead would break the identity `recoverReciprocalNumerator`
+  licenses, which `maxInteractivityAtCost` depends on — the cost-cap card would answer
+  against a different cost curve than the bars, the one thing its comment says must never
+  happen;
+- and dividing the displayed `$/chip-hr` by the displayed throughput would stop giving the
+  displayed cost, for disaggregated rows only.
+
+Fixing cost alone is therefore not available. The choice is all of it or none of it, and
+the notes now carry the correction instead: which figures are affected (Input and Output,
+never Total), in which direction (flattering), by how much, and that the Fleet Lifecycle
+section next to them is deliberately on a different basis.
+
+### Chart and table are tabs, with the figure's own header
+
+The section renders as a `<figure>` in the same shape as the bar chart above it and
+the `/inference` charts: a `ChartButtons` row carrying a Chart/Table `SegmentedToggle`
+and the PNG/CSV export menu, and a caption with the title plus the provenance a reader
+needs — model, scenario, target interactivity, power budget, source. The caption is
+passed to the chart as its `caption` and rendered as a `<figcaption>` in the table view,
+so switching tabs never loses the heading.
+
+PNG export is disabled on the table tab: a picture of a paginated HTML table is not an
+artefact anyone wants. The CSV is the export for that view, and it carries the
+assumptions in its notes preamble — a CSV read six months later cannot be reconstructed
+from the rows alone.
+
+The table's search box is off (`searchable={false}` on `DataTable`). It is one row per
+chip, and every chip is already named in the legend and on the chart; a search field
+over five rows is furniture. The prop is new and defaults to true, so no other table
+changed.
+
+### URL params
+
+`c_price`, `c_life` (horizon), `c_ramp`, `c_cache` (cached-input price, % — agentic
+only), `c_mtbi`, `c_rec`, `c_ly` (y-axis metric —
+`margin` | `marginPerMw` | `revenue` | `revenuePerMw` | `cumulativeRevenue`, parsed
+against that allowlist so a stale or hand-edited link falls back to `margin` rather
+than seeding an unknown metric).
+The first two default to
+`''` in `PARAM_DEFAULTS` because their real defaults are derived, not constant — see
+the comment there. The MW budget is `c_mw`, defaults to 10 MW, and is shared by
+the calculator and Fleet Lifecycle page so a budget set on either seeds the
+other.
diff --git a/docs/testing.md b/docs/testing.md
index dfdcd1224..3303b8607 100644
--- a/docs/testing.md
+++ b/docs/testing.md
@@ -3,16 +3,18 @@
 ## Structure
 
 **Unit tests**: Vitest, colocated with source (`*.test.ts` next to `*.ts`). Config: `packages/app/vitest.config.ts`.
-**E2E tests**: Cypress, in `packages/app/cypress/e2e/`. Config: `packages/app/cypress.config.ts`.
+**Component tests**: Cypress, in `packages/app/cypress/component/`. Use these for isolated UI behavior with controlled providers, routes, and API intercepts.
+**Integration tests**: Cypress, in `packages/app/cypress/e2e/`. Use these for page routing, browser history, API fixture mode, and cross-component workflows. Config: `packages/app/cypress.config.ts`.
 
 ## Requirements (Mandatory)
 
 Enforced by `@pr-claude` — missing/low-quality tests are flagged 🔴 BLOCKING.
 
 1. New utility functions → colocated unit test
-2. New UI features → E2E test in `cypress/e2e/<feature>.cy.ts`
-3. Bug fixes → regression test reproducing the bug
-4. Run `bun run test:unit` and the local smoke suite, `bun run test:e2e`, before considering a task complete. The full E2E suite runs in CI and is available locally as `bun run test:e2e:full`.
+2. Component-local UI behavior → Cypress component test in `cypress/component/`
+3. Route, navigation, or cross-component behavior → Cypress integration test in `cypress/e2e/`
+4. Bug fixes → regression test at the narrowest layer that reproduces the bug
+5. Run `bun run test:unit` and the local smoke suite, `bun run test:e2e`, before considering a task complete. The full E2E suite runs in CI and is available locally as `bun run test:e2e:full`.
 
 ## Pre-commit Checklist
 
@@ -29,6 +31,26 @@ The local `bun run test:e2e` command is a curated smoke suite across the core pa
 
 The complete suite is `bun run test:e2e:full`. It runs all Cypress component and integration specs. GitHub Actions runs that same coverage as one component job plus four integration shards per browser, Chrome and Firefox. The CI workflow is the merge gate for the full E2E suite.
 
+`packages/app/timings.json` is the committed `cypress-split` baseline. Its unit guard requires one positive timing for every integration spec and rejects removed entries. Regenerate the baseline from an observed full integration run when specs are added, removed, or materially rebalanced.
+
+With an `E2E_FIXTURES=1` app server running on port 3000, run this from the repository root:
+
+```bash
+E2E_FIXTURES=1 \
+  SPLIT=1 \
+  SPLIT_INDEX1=1 \
+  SPLIT_FILE=timings.json \
+  SPLIT_OUTPUT_FILE=timings.json \
+  SPLIT_SUMMARY=false \
+  bun run --cwd packages/app test:e2e:integration
+```
+
+`SPLIT=1` intentionally runs every spec in one chunk so one process records the complete baseline. CI uses `SPLIT=4` only for parallel execution and writes each shard's timing output to a throwaway file.
+
+`E2E_FIXTURES=1` serves the committed API snapshots under `packages/app/cypress/fixtures/api/`. Refresh them with `bun run --cwd packages/app capture:fixtures`. The capture script updates `_manifest.json`, which records the fixture shape, byte length, checksum, source, and capture timestamp. The manifest guard rejects partial or hand-edited snapshots; benchmark history must also contain at least two dates so replay tests cannot silently skip their substantive path.
+
+Server-cache performance regressions must assert source-reader call counts for repeated identical inputs. Do not use elapsed-time thresholds: CI scheduling and cold-start variance make wall-clock assertions nondeterministic.
+
 ## Quality Standards
 
 1. **No tautological tests** — every test must verify a real transformation
diff --git a/docs/typography.md b/docs/typography.md
new file mode 100644
index 000000000..bae6881d4
--- /dev/null
+++ b/docs/typography.md
@@ -0,0 +1,49 @@
+# Typography
+
+How text is sized and spaced in the app, and the ratchet gate that keeps it that way.
+
+## Tokens
+
+Defined in the `@theme` block of `packages/app/src/app/globals.css`:
+
+| Token                     | Class                   | Value            | Use for                                           |
+| ------------------------- | ----------------------- | ---------------- | ------------------------------------------------- |
+| `--text-2xs`              | `text-2xs`              | 0.6875rem (11px) | Micro labels one step below `text-xs`             |
+| `--text-3xs`              | `text-3xs`              | 0.625rem (10px)  | Densest labels: table meta, tooltips, log gutters |
+| `--tracking-eyebrow`      | `tracking-eyebrow`      | 0.16em           | Uppercase eyebrow/kicker labels                   |
+| `--tracking-eyebrow-wide` | `tracking-eyebrow-wide` | 0.2em            | Extra-wide uppercase labels                       |
+| `--tracking-heading`      | `tracking-heading`      | -0.04em          | Tightened display/hero headings                   |
+
+Rules of thumb:
+
+- Body and heading sizes use the standard Tailwind scale (`text-xs` … `text-6xl`). Do **not** add wrapper components for body text.
+- Never write `text-[11px]`, `text-[0.65rem]`, `tracking-[0.16em]`, etc. — the gate (below) blocks new arbitrary font sizes and letter-spacing values. `tracking-[0.1em]` is exactly `tracking-widest`; use the built-in.
+- The micro size tokens deliberately define **no paired line-height**: the arbitrary values they replaced set only `font-size` and inherited line-height, and dense tables depend on that. Don't add `--text-2xs--line-height` without re-checking those surfaces.
+- The custom tracking tokens are registered with tailwind-merge in `src/lib/utils.ts` (`extendTailwindMerge`); if you add a token, add it there too or `cn()` will keep conflicting classes.
+
+## Components
+
+- `<Heading>` (`src/components/ui/heading.tsx`) — cva-based visual heading levels (`display`, `page`, `section`, `card`, `label`) with the rendered tag chosen via `as`, so outline and visual size stay independent. New pages should use it instead of retyping size/weight/tracking recipes; existing headings migrate file-by-file as pages get touched — no sweep.
+- `<Eyebrow>` (`src/components/ui/eyebrow.tsx`) — the uppercase mono kicker label (`tone="brand" | "muted"`, `wide` for 0.2em tracking).
+
+Blog/article typography is owned by `@tailwindcss/typography` (`prose`) and is out of scope — don't migrate MDX content styles to these tokens.
+
+## Chart text
+
+Chart font sizes live in TypeScript, not CSS variables: `CHART_TYPE` in `src/lib/d3-chart/typography.ts`, with the `px()` helper for `.attr('font-size', …)`. The PNG export path (`useChartExport`) serializes the chart with html-to-image, which cannot resolve `var(--*)`; its `resolveCssVarsForExport()` only bakes color-type attributes, so a CSS-variable font-size silently collapses in exports. The shared export font stacks (`CHART_FONT_SANS`, `CHART_FONT_MINECRAFT`) live in the same module.
+
+OG-image renderers (Satori: `opengraph-image.tsx`, `og-image-render.tsx`, `compare-og.tsx`) draw on a fixed canvas and intentionally hardcode display sizes — they are exempt from all of this.
+
+## The ratchet gate
+
+`bun run check:typography` (wired into CI's lint workflow and lefthook pre-commit) scans `packages/app/src` for:
+
+1. `arbitrary-text-size` — `text-[<number>…]` in any `.ts`/`.tsx` (hex colors like `text-[#ed1c24]` are out of scope);
+2. `arbitrary-tracking` — any `tracking-[…]`;
+3. `chart-font-literal` — quoted `font-size` literals inside `src/lib/d3-chart/`.
+
+Per-file counts are compared against `packages/app/scripts/typography-allowlist.json`:
+
+- **A count above the allowlist (or a hit in an unlisted file) fails immediately** — fix it with a token, `CHART_TYPE`, or the standard scale.
+- **A count below the allowlist also fails**, telling you to run `bun run check:typography --update` — so the allowlist only ever shrinks. It doubles as the migration burndown; there is no deadline, offenders migrate when their file is next touched.
+- Matching logic lives in `src/lib/typography-gate.ts` (unit-tested); exemptions (OG renderers, tests, the gate itself) are defined there.
diff --git a/lefthook.yml b/lefthook.yml
index 09753abf7..af93e4507 100644
--- a/lefthook.yml
+++ b/lefthook.yml
@@ -11,3 +11,6 @@ pre-commit:
 
     - name: typecheck
       run: bun run typecheck
+
+    - name: typography
+      run: bun run check:typography
diff --git a/package.json b/package.json
index 65db68a61..20d593638 100644
--- a/package.json
+++ b/package.json
@@ -4,7 +4,7 @@
   "private": true,
   "repository": {
     "type": "git",
-    "url": "https://github.com/SemiAnalysisAI/inferencemax-app"
+    "url": "https://github.com/SemiAnalysisAI/InferenceX-app"
   },
   "workspaces": [
     "packages/*"
@@ -24,6 +24,7 @@
     "lint:fix": "oxlint --fix --no-error-on-unmatched-pattern",
     "fmt": "oxfmt --check --no-error-on-unmatched-pattern",
     "fmt:fix": "oxfmt --write --no-error-on-unmatched-pattern",
+    "check:typography": "bun packages/app/scripts/check-typography.ts",
     "test": "bun run --cwd packages/app test",
     "test:e2e": "bun run --cwd packages/app test:e2e",
     "test:e2e:full": "bun run --cwd packages/app test:e2e:full",
@@ -43,19 +44,31 @@
     "admin:db:migrate:collectivex": "bun run --cwd packages/db db:migrate:collectivex",
     "admin:db:apply-overrides": "bun run --cwd packages/db db:apply-overrides",
     "admin:db:backfill-full-response-interactivity": "bun run --cwd packages/db db:backfill-full-response-interactivity",
+    "admin:db:backfill-server-log-files": "bun run --cwd packages/db db:backfill-server-log-files",
     "admin:db:reset": "bun run --cwd packages/db db:reset",
     "admin:db:verify": "bun run --cwd packages/db db:verify"
   },
   "devDependencies": {
     "@babel/core": "^8.0.1",
-    "@types/node": "^26.1.1",
+    "@types/node": "^26.2.0",
     "is-ci": "^4.1.0",
     "lefthook": "^2.1.10",
-    "oxfmt": "^0.60.0",
-    "oxlint": "^1.75.0",
+    "oxfmt": "^0.63.0",
+    "oxlint": "^1.78.0",
     "rimraf": "^6.1.3",
     "typescript": "^6.0.3",
-    "vite": "^8.1.5"
+    "vite": "^8.2.1"
+  },
+  "overrides": {
+    "brace-expansion": "5.0.9",
+    "dompurify": "3.4.13",
+    "fast-uri": "4.1.2",
+    "hono": "4.13.2",
+    "ip-address": "10.5.0",
+    "js-yaml": "3.15.1",
+    "jsdom": "30.0.1",
+    "nanoid": "3.3.18",
+    "undici": "8.10.0"
   },
   "engines": {
     "bun": ">=1.3.14",
diff --git a/packages/app/content/blog/agentic-benchmark-agent-benchmark-guide.mdx b/packages/app/content/blog/agentic-benchmark-agent-benchmark-guide.mdx
new file mode 100644
index 000000000..24a7dd667
--- /dev/null
+++ b/packages/app/content/blog/agentic-benchmark-agent-benchmark-guide.mdx
@@ -0,0 +1,289 @@
+---
+title: 'Agentic Benchmark for LLM Inference: Metrics and Methodology'
+subtitle: 'How an agent benchmark replays long-context, multi-turn workloads to measure latency, throughput, cache behavior, and serving cost'
+seoDescription: 'Agentic benchmark guide for LLM inference: learn how an agent benchmark measures latency, throughput, prefix caching, concurrency, and cost.'
+date: '2026-08-19'
+publishDate: '2026-08-19'
+tags:
+  - benchmark
+  - agents
+  - agentic
+  - inference
+  - latency
+  - throughput
+---
+
+An **agentic benchmark for LLM inference** measures how a model-serving system behaves under the traffic generated by AI agents. It does not grade whether an agent fixed a bug or completed a workflow. It measures the infrastructure serving each model call: latency, throughput, cache reuse, concurrency, and cost.
+
+That distinction changes the workload. A conventional inference benchmark can send independent prompts with fixed input and output lengths. An agent benchmark must represent sessions in which context grows across turns, requests share long prefixes, tools insert pauses, and subagents create parallel branches. The next request often cannot start until an earlier request finishes.
+
+AgentX is InferenceX's open benchmark for this workload. It converts privacy-preserving shapes from opt-in coding-agent traces into deterministic AIPerf replays, then compares complete hardware and software serving configurations across concurrent agent clients.
+
+<DashboardCTA href="/inference?i_seq=agentic-traces">
+  Compare live AgentX inference results →
+</DashboardCTA>
+
+## What is an agentic inference benchmark?
+
+An agentic inference benchmark is a systems benchmark for multi-turn agent traffic. The system under test includes the accelerator, server topology, inference engine, model checkpoint, precision, parallelism strategy, cache configuration, and request scheduler.
+
+It answers questions such as:
+
+- How many concurrent agent sessions can the deployment serve?
+- How long does each agent wait for its first token?
+- How quickly does each active agent receive output tokens?
+- How much output throughput does each chip deliver at a target latency?
+- Do repeated prefixes hit HBM, host memory, or recompute paths?
+- What does a million served tokens cost at a usable operating point?
+
+The word “benchmark” is also used for agent task evaluations, so the scope must be stated. The two test types answer different questions:
+
+| Question                                     | Test type                   | Typical output                                      |
+| -------------------------------------------- | --------------------------- | --------------------------------------------------- |
+| Can the agent complete the task correctly?   | Agent task evaluation       | Success rate, policy compliance, or grader score    |
+| Can the inference stack serve agent traffic? | Agentic inference benchmark | Latency, throughput, cache behavior, capacity, cost |
+
+AgentX covers the second row. Semantic task quality requires a separate evaluation.
+
+## Why fixed-sequence inference benchmarks miss agent traffic
+
+Fixed-sequence benchmarks remain useful for controlled comparisons. An 8K input and 1K output workload isolates a common request shape and makes configurations easy to reproduce. It does not reproduce a long-running agent session.
+
+| Workload property | Fixed-sequence inference                    | Agentic inference                                      |
+| ----------------- | ------------------------------------------- | ------------------------------------------------------ |
+| Request shape     | One fixed input and output length           | Length changes on every turn                           |
+| Arrival pattern   | Independent requests at a set rate or batch | Dependent requests triggered by earlier completions    |
+| Prefix reuse      | Optional or synthetic                       | Large shared prefixes grow through the session         |
+| Tool time         | Usually absent                              | Uneven pauses between model calls                      |
+| Parallelism       | Request-level concurrency                   | Main-agent and subagent branches can overlap           |
+| Cache lifetime    | Short and regular                           | Long-lived state competes for HBM and offload capacity |
+| Useful result     | Performance for one request distribution    | Capacity and responsiveness for complete sessions      |
+
+These differences can change which system wins. More HBM can keep long-lived KV cache resident. Better prefix caching can avoid reprocessing hundreds of thousands of prompt tokens. Scheduler and router costs that are small for independent prompts can grow with the number and length of live sessions.
+
+## An agent benchmark is a request graph
+
+An agent session is not a flat prompt list. Main-agent turns form a dependency chain. A turn can spawn one or more subagents, and the main chain may wait for those branches before it continues. Auxiliary calls can run without joining the main path.
+
+<Figure
+  src="/images/agentx-methodology/replay-multi-subagent.png"
+  alt="Agentic inference benchmark replay graph with a main-agent request chain, four parallel subagent branches, two join points, and an independent auxiliary request"
+  caption="AgentX represents a session as a dependency graph with main-agent turns, parallel subagents, join points, auxiliary calls, and recorded delays."
+/>
+
+AgentX replays this graph in a [closed loop](/glossary/closed-loop-benchmark). Each client sends its next eligible request only after the required earlier work completes. Faster servers advance through a session sooner and create later work sooner. The instantaneous HTTP request count can exceed client concurrency when subagents fan out, then fall when branches finish.
+
+This is why AgentX concurrency means **live agent clients**, not a fixed request batch. A run at concurrency 50 can place more than 50 requests in flight if several clients have active subagents.
+
+## How AgentX builds a reproducible workload
+
+AgentX turns recorded session structure into a content-free replay in four stages.
+
+### 1. Capture request shape and timing
+
+An opt-in HTTP proxy records request arrival and completion times, input and output token counts, conversation IDs, and subagent IDs. It does not publish prompts, source code, tool arguments, or tool results.
+
+### 2. Preserve prefixes without preserving content
+
+Each input becomes a chain of session-scoped hashes in 64-token blocks. Matching block IDs retain the prefix relationships between turns. AIPerf replaces those blocks with deterministic synthetic coding and tool-use tokens before replay.
+
+### 3. Reconstruct dependencies
+
+Main-agent requests become a linear chain. Subagent requests become separate chains with spawn and join dependencies. Recorded inter-turn gaps reproduce the time spent using tools or waiting for other work. The replay keeps client-visible ordering and overlap without claiming access to provider-internal events.
+
+### 4. Replay the same scenario across systems
+
+A fixed seed controls session sampling, replay start points, and synthetic payloads. Each hardware and software configuration receives the same scenario definition, so differences in the resulting curve come from the serving system rather than a hand-written prompt batch.
+
+AgentX v1.0 contains 393 published sessions selected from opt-in Claude Code traces. Each selected session has at least 20 requests and no more than 10 concurrent subagents.
+
+| AgentX v1.0 property             |   Value |
+| -------------------------------- | ------: |
+| Published sessions               |     393 |
+| Median input tokens per request  | 142,016 |
+| Median output tokens per request |     444 |
+| Sessions containing subagents    |     44% |
+| Full dataset maximum context     |      1M |
+| Context-limited dataset variant  |    256K |
+
+<Figure
+  src="/images/agentx-methodology/request-distributions.png"
+  alt="Log-scale AgentX v1.0 distributions for inter-turn delay, input sequence length, and output sequence length"
+  caption="AgentX v1.0 request distributions. The median request has 142,016 input tokens and 444 output tokens."
+/>
+
+The full variant retains contexts up to 1M tokens. The 256K variant removes requests above its context cap while preserving the timing and subagent overlap of the requests that remain. The [AgentX methodology](/agentx/methodology) publishes the dataset rules, figures, and source links.
+
+## Benchmark controls that make comparisons valid
+
+Long-running agent workloads are sensitive to starting state, cache contents, speculative decoding, and host memory. AgentX locks those variables with explicit controls.
+
+### Cache warmup and measurement window
+
+A fixed seed selects a starting point between 25% and 75% of each recorded session. A `max_tokens=1` primer materializes active main-agent and subagent prefixes. Each replay lane then completes 10 additional warmup requests before measurement starts.
+
+Only the following one-hour profiling window contributes to reported metrics. Each recycled replay receives a unique cache-bust marker so unrelated sessions cannot accumulate a shared prefix.
+
+### Synthetic payloads and speculative decoding
+
+Synthetic tokens preserve length and prefix structure, but they do not naturally reproduce draft-token acceptance. AgentX therefore uses an acceptance length measured from the coding category of [SPEED-Bench](https://arxiv.org/abs/2604.09557) for each model, speculative method, draft length, and thinking mode. The value is recorded in versioned golden files and applied through engine controls.
+
+This keeps the comparison focused on serving performance instead of accidental differences in how synthetic text interacts with a speculator.
+
+### DRAM and KV-cache offload
+
+Host-memory capacity can change whether a long prefix is reloaded or recomputed. Servers without a standardized DRAM configuration are capped at 3 TB. Standard systems such as GB200 NVL72, GB300 NVL72, and TPUv7 can use their installed capacity. A benchmark configuration can access only the host-memory share proportional to its GPU allocation.
+
+The replay format and dependency rules are documented in the [AIPerf WEKA trace guide](https://github.com/ai-dynamo/aiperf/blob/main/docs/tutorials/weka-trace.md).
+
+## Metrics for an agentic benchmark
+
+One number cannot describe agent serving. Report throughput, latency, and interactivity together across a concurrency sweep.
+
+| Metric                            | What it reveals                                                       |
+| --------------------------------- | --------------------------------------------------------------------- |
+| Output throughput per chip        | Generated tokens completed by each accelerator per second             |
+| P90 interactivity                 | Per-client generation speed derived from p90 inter-token latency      |
+| Time to first token (TTFT)        | Delay before an agent starts receiving a response                     |
+| Inter-token latency (ITL)         | Time between streamed output tokens                                   |
+| Input throughput                  | Prompt tokens processed per second                                    |
+| Cache-source distribution         | Prompt tokens served from HBM, host memory, or recomputation          |
+| Queue and in-flight request depth | Scheduler pressure as branches spawn and join                         |
+| Cost per million tokens           | Serving economics under a stated latency or interactivity requirement |
+
+Percentiles matter. A low mean TTFT can coexist with a poor p90 or p99 experience if a subset of agent branches waits behind long prefills. Aggregate output throughput can rise while each individual client receives tokens more slowly.
+
+## Read the full throughput-interactivity curve
+
+AgentX sweeps concurrent agent clients and plots the resulting operating points. Low concurrency usually favors per-client responsiveness. Higher concurrency can improve batching and aggregate throughput until queueing, memory pressure, or another bottleneck reduces interactivity.
+
+<Figure
+  src="/images/agentx-methodology/pareto-b200-minimax-m3.png"
+  alt="AgentX Pareto curve for MiniMax-M3 on NVIDIA B200 showing output throughput per chip versus p90 interactivity at different client concurrency levels"
+  caption="Each point is a client-concurrency setting. Compare systems at the same p90 interactivity instead of comparing unmatched peak-throughput points."
+/>
+
+The useful result is the [Pareto frontier](/glossary/pareto-frontier), not the single point with maximum throughput. Hardware A may produce the highest tok/s/chip at low interactivity while hardware B leads at the response speed required by an interactive coding product.
+
+Closed-loop replay also means a faster configuration can progress farther through its sampled sessions during the same profiling hour. The exact request mix can vary slightly, especially at low concurrency. Results should include the full curve and enough concurrency to reduce sampling noise.
+
+## What agent traffic stresses in an inference stack
+
+Agentic inference exposes several bottlenecks that fixed prompt batches can understate.
+
+- **KV-cache capacity:** Long contexts and many live sessions keep large caches resident for longer.
+- **Prefix-cache correctness:** A false hit produces wrong state; a missed hit repeats expensive prefill work.
+- **Offload bandwidth:** Moving a long prefix from host memory can be faster than recomputing it, but transfer granularity and descriptor overhead matter.
+- **Scheduler fairness:** Parallel subagents can raise throughput while an individual branch waits too long.
+- **Router and frontend CPU:** Prefix matching, expiry tracking, serialization, and streamed response handling scale with live session state.
+- **Long-context kernels:** Attention, sparse attention, normalization, and graph-capture choices can behave differently across varying sequence lengths.
+- **Speculative decoding:** Acceptance controls and scheduler lifetime affect both throughput and interactivity.
+
+The [AgentX optimization tracker](/agentx/optimizations) documents concrete changes in vLLM, SGLang, TensorRT-LLM, ATOM, Dynamo, and KV-cache systems that were exposed by this workload.
+
+## How to compare AgentX results
+
+Use the same scenario contract on both sides of a comparison:
+
+1. Match the model, checkpoint, precision, and thinking mode.
+2. Match the AgentX dataset variant and context limit.
+3. Record the inference engine, container image, parallelism, and speculative-decoding method.
+4. Apply the same golden acceptance-length rule.
+5. Match DRAM capacity, KV-cache offload policy, and the proportional memory rule.
+6. Use the same seed, warmup procedure, profiling duration, and client-concurrency sweep.
+7. Compare per-chip and whole-system results when system sizes differ.
+8. Compare at the same p90 interactivity or latency target.
+9. Include failed and unstable configurations instead of reporting only surviving points.
+
+An iso-interactivity comparison is usually more useful than peak throughput. It answers a deployment question directly: at the response speed users need, which system serves more agent work per chip or per dollar?
+
+## What this agent benchmark can and cannot answer
+
+AgentX can measure serving capacity, responsiveness, cache reuse, memory pressure, throughput, and cost under a trace-derived agent workload. It can show whether a software change improves long-context multi-turn serving on the same hardware, or whether a different hardware topology changes the useful frontier.
+
+AgentX cannot tell whether the model wrote correct code, selected the right tool, followed a business policy, or completed the source task. The synthetic payload deliberately removes those semantics. Use a task evaluation for agent quality and AgentX for inference performance. Keep the scores separate because they describe different systems questions.
+
+<DashboardCTA href="/inference?i_seq=agentic-traces">
+  Explore the AgentX throughput-interactivity frontier →
+</DashboardCTA>
+
+## Frequently asked questions
+
+### What is an agentic benchmark for LLM inference?
+
+It is a systems benchmark that measures how an inference stack serves multi-turn agent traffic. It replays growing contexts, shared prefixes, tool delays, and subagent branches, then reports latency, throughput, interactivity, cache behavior, capacity, and cost.
+
+### How is an agentic benchmark different from a fixed-sequence inference benchmark?
+
+A fixed-sequence benchmark sends independent requests with a specified input and output length. An agentic benchmark replays dependent sessions whose request lengths, timing, prefix reuse, and parallel branches change over time. The latter measures the serving conditions created by working agents.
+
+### What does the AgentX benchmark measure?
+
+AgentX measures complete inference configurations under trace-derived coding-agent traffic. It sweeps concurrent agent clients and reports metrics including output throughput per chip, p90 interactivity, TTFT, ITL, cache behavior, and serving cost.
+
+### Why does AgentX use closed-loop concurrency?
+
+An agent cannot issue every turn in advance because later actions depend on earlier results. In closed-loop replay, each client sends its next eligible request after its dependencies complete. This preserves the relationship between server speed and how quickly a session creates more work.
+
+### Which metrics matter most for agent inference?
+
+Use output throughput, p90 interactivity, TTFT, ITL, cache-source distribution, queue depth, and cost together. Compare full concurrency curves at the same latency or interactivity target. A maximum-throughput point alone does not describe an interactive deployment.
+
+### Does an agent benchmark measure model or agent quality?
+
+Some benchmarks do, but AgentX does not. AgentX measures inference-system performance with synthetic payloads that preserve workload shape. Task correctness, tool selection, and policy compliance require a separate agent evaluation.
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "What is an agentic benchmark for LLM inference?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "It is a systems benchmark that measures how an inference stack serves multi-turn agent traffic. It replays growing contexts, shared prefixes, tool delays, and subagent branches, then reports latency, throughput, interactivity, cache behavior, capacity, and cost."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "How is an agentic benchmark different from a fixed-sequence inference benchmark?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "A fixed-sequence benchmark sends independent requests with a specified input and output length. An agentic benchmark replays dependent sessions whose request lengths, timing, prefix reuse, and parallel branches change over time."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What does the AgentX benchmark measure?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "AgentX measures complete inference configurations under trace-derived coding-agent traffic. It sweeps concurrent agent clients and reports output throughput per chip, p90 interactivity, time to first token, inter-token latency, cache behavior, and serving cost."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does AgentX use closed-loop concurrency?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "An agent cannot issue every turn in advance because later actions depend on earlier results. In closed-loop replay, each client sends its next eligible request after its dependencies complete. This preserves the relationship between server speed and how quickly a session creates more work."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Which metrics matter most for agent inference?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Use output throughput, p90 interactivity, time to first token, inter-token latency, cache-source distribution, queue depth, and cost together. Compare full concurrency curves at the same latency or interactivity target."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Does an agent benchmark measure model or agent quality?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Some benchmarks do, but AgentX does not. AgentX measures inference-system performance with synthetic payloads that preserve workload shape. Task correctness, tool selection, and policy compliance require a separate agent evaluation."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/agentx-inferencexv3-does-cuda-moat.mdx b/packages/app/content/blog/agentx-inferencexv3-does-cuda-moat.mdx
new file mode 100644
index 000000000..9877fef69
--- /dev/null
+++ b/packages/app/content/blog/agentx-inferencexv3-does-cuda-moat.mdx
@@ -0,0 +1,1031 @@
+---
+title: 'AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?'
+subtitle: '$3 Million USD dataset open sourced, 1 Mil+ Context Length, Multiturn, Sub Agents 95%+ KVCache HitRate, GB300 NVL72, MI355X, B200'
+seoDescription: 'AgentX 1.0 is the first fully open source multi-turn agentic coding inference benchmark at 1M context. Agentic results for DeepSeek V4 Pro, Kimi K3, MiniMax M3, Qwen3.5, and GLM 5.3 across GB300 NVL72, B300, B200, MI355X, and H200, plus the 70+ upstream PRs the benchmark drove.'
+date: '2026-08-24'
+publishDate: '2026-08-24'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - gpu
+  - nvidia
+  - amd
+  - announcement
+---
+
+Since the [Claude Code inflection point](https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point) in November 2025, long-context, multi-turn agentic workloads have grown rapidly. They now dominate traffic for production inferencing. In April 2026, OpenAI’s Enterprise agentic spending overtook ChatGPT spending.
+
+[Claude Code is the Inflection Point](https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point) — 4% of GitHub public commits are being authored by Claude Code right now. At the current trajectory, we believe that Claude Code will be 20%+ of all daily commits by the end of 2026. While you blinked, AI consumed all of software development.
+
+Agentic workflows have decisively taken the baton. **Today, we announce AgentX 1.0 - the world’s first fully open source, multi-turn agentic coding inference benchmark at 1 million context, released under Apache 2.0. [Our full dashboard is available here.](https://inferencex.semianalysis.com/)**
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/864512bf-1feb-445d-a03a-bbf9f4f1b791_1928x1234.png"
+  caption="Source: [SemiAnalysis](https://openai.com/signals/enterprise-data/)"
+/>
+
+In the past most measured performance based on fixed sequence length prefill and decode workloads, but this is an inaccurate way to measure workloads. Reality is multi-turn, long context, high prefill reuse, with sub agent bursts, KVCache offload, and numerous tool calls. As such we aimed to build the correct way for the industry to measure AI hardware and software performance.
+
+[We have spent more than $3M building this dataset. Today, we open source everything.](https://inferencex.semianalysis.com/) InferenceXv3 implements AgentX, a new realistic scenario in addition to the existing “fixed sequence length” scenarios (8k1k, 1k1k, 1k8k). It improves the benchmark scenarios by using agentic coding traffic instead of the previous single-turn traffic of 8k input and 1k output tokens.
+
+The full matrix runs on ~2MW of continuously operated compute across over 1000 chips spanning a wide range of SKUs, featuring the MI355X, GB300 NVL72, GB200 NVL72, B300, B200, MI325, MI300X, H200, and RTX Pro Servers. Rubin arrives later this month, and TPUs and Mi455X UALoE72 arrive later this year. [Please drop a star if you found our free open source work valuable.](https://github.com/SemiAnalysisAI/InferenceX)
+
+It is great to see amazing performance from both NVIDIA and AMD on agentic workloads. NVIDIA does very good on a lot of frontier models while AMD also does well on some frontier models for specific comparsions.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/8a065d23-de1d-4e2f-a3ae-7af724409c9d_1350x653.png"
+  caption="Source: [SemiAnalysis GitHub](https://github.com/SemiAnalysisAI/InferenceX)"
+/>
+
+The most valuable thing AgentX produced in its first months was not the initial results. It was the massive industry impact the benchmark is already having. **Over 70+ upstream PRs** for optimizing real world production agentic workloads across vLLM, SGLang, TensorRT-LLM, ATOM, AITER, Dynamo, LMCache, and Mooncake, uses AgentX as the north star benchmark proxy. Most of these optimization improvements are transferable to production traffic. We deep dive into each of these optimizations later in the article.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/193e605e-579b-4ed2-9637-5bfdbee5cf71_1286x1515.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Open source is a core principle for InferenceX and thus, we open more of the stack than most people who use that word. That includes an open frontend, a public database served through an easily consumable REST API **that multiple tier 1 AI lab’s capacity planning teams already consume**, public GitHub Actions CI provenance, [logs](https://inferencex.semianalysis.com/inference/agentic/440193?i_seq=agentic-traces&i_xmode=interactivity&view=logs), and accuracy validation on every single point. Crucially, our benchmark configs mainly track [recipes.vllm.ai](http://recipes.vllm.ai) and [SGLang cookbook](https://docs.sglang.io/cookbook/intro) on upstream images such that we are measuring the performance actual customers are experiencing instead of measuring benchmax’ed images.
+
+In three to four weeks, we will release an AgentX update article. It will cover further optimizations to agentic workloads, plus updated performance results from AMD and Nvidia. It is important to understand that the profile of agentic workloads is updating fast. InferenceX will continue to move swiftly to benchmark the relevant workloads.
+
+InferenceX is 100% committed to being open-source - this would not be possible without the contributions and support from our OSS partners. We would like to thank the following people that have made massive contributions to the AgentX 1.0 release:
+
+- **Inferact/vLLM**: Roger Wang, Yifan Qiao, Simon Mo, Jeff Ma, and many others
+- **RedHat/llm-d**: Michael Goin, Robert Shaw, Tyler Michael Smith
+- **RadixArk/SGLang**: Baizhou Zhang, Yuwei An, Mingyi Lu, and many others
+- **LMCache/TensorMesh**: Samuel Shen
+- **Weka**: Callan Fox, Val Bercovici
+- **MoonCake Maintainers:** Teng Ma, Xu Wenjie, Ke Yang
+- **AMD**: Thomas Wang, HaiShaw, Andy Luo, Seungrok Jung, Chun Fang, Parth Panchal, Bill He, Theresa Shan, Hongxia, Fangzhou, Gilbert Lei, Yanfei Wang, Duyi Wang, Peng Sun, Lingpeng Jin, Simon Danielsson, Xiaohu Guo, Haichen Zhang, Chang Liu, Doug Lehr, Poovaiah Palangappa, and many others in the AMD Shanghai Development Centre
+- **Nvidia**: Xin Li, Anthony Casagrande, Kedar Potdar, Ankur Singh, Ishan Dhanani, Nick Comly, Nvidia Shanghai TensorRT-LLM team, and many others
+- **Anthropic staff, for promptly fixing multiple bugs that made implementing AgentX possible**
+- **GitHub:** Austen Stone for helping with reliability of GitHub Actions that AgentX uses
+- And many others
+
+[In addition, we are thankful to all who support our open source InferenceX initiative, including Meta, Microsoft, Oracle, OpenAI, MiniMax, Moonshot Kimi, Alibaba Qwen, and Zhipu GLM.](https://inferencex.semianalysis.com/quotes)
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e97525a9-216a-4624-91c6-373ee7a3850e_1123x437.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/quotes)"
+/>
+
+## A Brief Overview of Agentic Workloads
+
+At a high level, an agentic workload is characterized by four elements:
+
+1. **Multi-turn:** a session includes many user / assistant interactions (tens or hundreds) compared to a handful in a chatbot scenario. Multi-turn, long context, high prefill reuse, with sub agent bursts and numerous tool calls.
+2. **Long context:** system prompts, tool definitions, and the large number of turns make context accumulate quickly.
+3. **High prefix reuse:** since the conversation progresses linearly, where output from turn n-1 is concatenated to turn n (typically), most context can be served from KV cache rather than recomputed (this depends on amount of storage available to store KV tensors). As n grows, the ratio of cached input relative to uncached typically tends towards 1.
+4. **Sub-agent bursts:** a session launches multiple short-lived sub-agents with fresh context, which create bursty KVCache patterns.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/72b64a02-8a1f-4a19-929c-85c0ed4803cd_2048x909.png"
+  caption="Source: DeepSeek, SemiAnalysis"
+/>
+
+Considering the characteristics above, benchmarking these workloads is fundamentally different from the existing fixed sequence length benchmarks. Namely, agentic inference is inherently a _systems problem_. Because of extremely high prefix reuse, KV tensors must be efficiently transferred across nodes/ranks (NIXL, MORI-IO, Mooncake). Additionally, different conversations should be routed to different nodes/ranks depending on where the appropriate prefix resides in order to maximize cache hit rate (LLM-d, Dynamo, vLLM/SGLang router). Long context conversations stress the HBM capacity for KV cache and necessitate offloading KV tensors to different tiers of memory (DRAM, SSD), a process that needs to be carried out efficiently (Mooncake Store, LMCache, vLLM Simple Offloading, SGLang HiCache).
+
+This is in contrast to fixed sequence length, single turn workloads where prefix reuse is not relevant and inference performance is largely reflective of baseline chip/kernel performance. This is not to say that the plethora of fixed sequence length data on InferenceX is not important. In fact, stripping away the complexities of agentic serving shows clearly how low-level inference performance optimizations are progressing. It also provides an important baseline for AgentX results.
+
+In an attempt to make the AgentX workloads as realistic as possible, we collected an initial corpus of [393 internal SemiAnalysis anonymous Claude Code traces](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126) to replay. To anonymize the content while keeping the original prefix reuse pattern, we use a method similar to the [Qwen-Bailian dataset](https://github.com/alibaba-edu/qwen-bailian-usagetraces-anon), one of the earliest corpora of production traces. We then use [AIPerf](https://github.com/ai-dynamo/aiperf) to reconstruct the traces according to the original schedule of requests at varying levels of concurrent clients. We worked with Anthropic to ship two Claude Code features to make the AgentX dataset possible. We thank the Anthropic staff for their help.
+
+1. [https://github.com/anthropics/claude-code/issues/49207](https://github.com/anthropics/claude-code/issues/49207)
+2. [https://github.com/anthropics/claude-code/issues/66761](https://github.com/anthropics/claude-code/issues/66761)
+
+This brief introduction to agentic workloads should give the reader enough context to understand the results in the next section. In a later section, we will provide a deeper technical dive into the methodology, replay harness, and dataset.
+
+## Agentic Coding Inference Performance
+
+When looking at inference coding performance, OpenAI, Anthropic, xAI, and other frontier labs focus on three things. They look at performance per dollar versus interactivity (TPOT), TTFT (time to first token), and overall end-to-end task completion. Performance per megawatt is also important, considering that terrestrial datacenter power is a critical constraint (money is a social construct and it appears the labs have an unlimited supply, but power is physically hard to come by in this day and age). [Our datacenter model has estimates of quarter by quarter build up of power demand and supply.](https://semianalysis.com/datacenter-industry-model/)
+
+In this section, we highlight some of the overall agentic performance themes across frontier models. We strongly encourage the reader to use this as a guide to [investigate the results for themselves](https://inferencex.semianalysis.com/). **All of the data is open source and the community has the opportunity to draw their own conclusions on the current state of real world inference performance.**
+
+## DeepSeek V4 Pro 0813
+
+DeepSeek V4 Pro 0813 is an ultra popular frontier open weight model from China. It has ~1.6 Trillion parameters with 49 Billion active parameters.
+
+As of August 21, the following graph shows the best performance per SKU for all submissions, normalized by total cost of ownership (TCO).
+
+The ISL/OSL distribution of _all_ requests among all DeepSeek v4 runs was as follows: ISL p50=88k, p90=272k, p95=404k, p99=675k and OSL p50=413, p90=2.2k, p95=3.7k, p99=8.6k.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/d8337140-b620-476a-abd3-59d6cc447d30_2048x1256.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/b5aee8b4-d11a-48d8-823c-1ca53412e689_2048x1253.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&i_seq=agentic-traces&g_runid=32433563482&i_xmode=interactivity&i_label=0&i_e2e_xmetric=)"
+/>
+
+In general, **it is important to consider both tokens per second per user** (TPS - also known as interactivity) and TTFT, **since these often come at the expense of one another**. For instance, in the graph above, some SKUs achieve very high throughput at decent interactivity, however TTFT is severely degraded. What is an “acceptable” p90 TTFT varies heavily depending on the application. For most production systems serving agentic workloads, you can expect p90 TTFT to be anywhere from 200-5,000ms. Anything over 5-10s is pushing the boundary of what can be considered “online inference.” There are still practical applications for the ultra-high throughput sector of the curve, where latency does not matter and peak system utilization is desirable (batch processing, _very_ long running agents, etc).
+
+In terms of single node performance, MI355X open-source performance (vLLM) trails behind vendor specific ATOM (AMDs equivalent of TensorRT LLM). We think it’s great that AMD is pushing the frontier quickly with ATOM, however we encourage them to make a higher priority of upstreaming these improvements into vLLM.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/d96697d9-bf24-4093-a7db-16d27a52aa08_2048x1200.png"
+  caption="Source: InferenceX"
+/>
+
+AMD’s distributed inference (DI) team has made great progress on 8k1k scenarios over the past six months. The team still has some way to go before DI is a viable solution for realistic workloads. In terms of throughput per GPU vs. interactivity, we observe the 1xDEP8+1xDEP8 disagg config is only able to realize slight performance gains in the high throughput scenarios, while actually performing worse in low latency scenarios.
+
+To make matters worse, any increase in throughput at the middle-to-high interactivity configs is overshadowed by the significant spike in p90 TTFT. One of the reasons for this is the use of SGLang’s --enable-prefill-delayer argument above concurrency 64, [which postpones prefill admission so DP ranks can form fuller batches](https://github.com/sgl-project/sglang/blob/v0.5.17/python/sglang/srt/server_args.py#L3180-L3194) (for up to 30 forward passes). Additionally, these points also increase chunked prefill size from 8,192 to 65,536.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/81475383-510e-4823-9373-751b8a3722e3_2048x1202.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/6254047b-173b-4b1b-9f43-2057eb720b05_2048x1200.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=mi355x_atom%2Cmi355x_mori-sglang&i_vendor=AMD&i_linelabel=0&i_metric=y_tpPerGpu)"
+/>
+
+On e2e latency, ATOM MI355X beats B200 vLLM (it does not beat B300 or B200 SGLang though). The issue with this is that most AI labs in China or the west do not want to use ATOM in production besides 1 small advertising business unit at Alibaba Corp due to tons of missing features. The main Qwen LLM org at baba does not use ATOM in production.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e8d915c0-c962-4277-b62c-a1b08b8e7179_1924x1262.png"
+  caption="Source: InferenceX"
+/>
+
+Before August 21, 2026, AMD’s MI355X strong SGLang development team was matching B200 vLLM on performance per dollar on end to end (e2e) performance.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/360a051c-6117-4860-a704-b50198f01521_1978x1246.png"
+  caption="Source: InferenceX"
+/>
+
+However, B300 vLLM and B200 SGLang still beat AMD’s MI355X.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9ec4708c-148e-42ee-a7c5-3bf4d9cbdfaa_1938x1246.png"
+  caption="Source: InferenceX"
+/>
+
+After August 21, 2026, due to optimizations in vLLM from Inferact and Nvidia, the performance per dollar of Nvidia’s B200 has surpassed that of the MI355X. This is a close race and we are excited to see the performance optimizations over the next couple weeks. We will be publishing an AgentX update article very soon.
+
+[AMD has listed their DeepSeekv4 vLLM optimization and includes lots of exciting things that they can do to improve their performance](https://github.com/vllm-project/vllm/issues/52911).
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9302da60-5bc1-46e1-9392-74e912b5c200_2024x1236.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Now turning to Nvidia. Their most competitive solutions are GB300 Dynamo TRTLLM and GB200 Dynamo vLLM. Both configs rely on PD disagg to achieve high throughput at reasonable interactivity. Additionally, GB300 configs employ wide-EP (DEP32) decode instances in order to achieve higher throughput at the middle of the frontier.
+
+Note that the 2xDEP8+1xDEP12 GB200 point is significantly closer to the 3xDEP8+1xDEP16 GB300 point in terms of TPS compared to TTFT. Again, TTFT is, in general, more sensitive to the “spikiness” of the workload. Since the GB300 point achieves much higher overall concurrency, it incurs more subagent traffic and hence more cold prefills. We can see this in the TTFT chart for the point:
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/96f9c9f9-eb86-4f87-8332-c85f23040836_2048x916.png"
+  caption="Source: InferenceX"
+/>
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f79e53f5-c8ff-4ed0-a8a3-c05b9e0b4bab_2048x1191.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/512c6225-5816-493e-839d-e4f539e5b1c4_2048x1215.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_metric=y_tpPerGpu&i_xmode=interactivity&i_linelabel=0&i_best=0&i_active=gb200_dynamo-vllm%2Cgb300_dynamo-trt)"
+/>
+
+When normalized by TCO, B300 vLLM versus B200 vLLM aggregated performance is quite similar. The main difference being that B300 can “squeeze” out extra throughput, given its 50% increase in HBM capacity over B200.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/939b9197-321d-4869-ba13-f0b800343b7b_2048x1300.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=b200_vllm%2Cb300_vllm)"
+/>
+
+We can further visualize this difference using our server metric visualizations, which are new to AgentX.
+
+Under the load of 384 concurrent agentic traces, B300 vLLM DEP8 w/ 3TB DRAM via vLLM simple offloading achieved a 91% HBM cache hit rate with an additional 1.36% DRAM cache hit rate. This is because the **HBM KV cache working set size** is approximately 43M tokens with this configuration, and the load barely exceeds this number of tokens in flight at any given time.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/b7ff2d61-bdab-4c2f-96a0-ec5a52469489_2048x884.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f86888c4-c1a7-47b0-a580-d1a74876cb9c_2048x898.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference/agentic/439522)"
+/>
+
+With B200 concurrency 196 (all other parameters stay the same), we see only 73% HBM cache hit rate and rely more heavily on DRAM with an offload cache hit rate of nearly 20%. We observe that the HBM KV cache working set size is 22M tokens, roughly half that of B300.
+
+DRAM KV offloading is typically implemented as a write-through cache, meaning every prefix written to the HBM cache is also written to the DRAM cache. Therefore, it is most effective when the amount of DRAM available for offloading is significantly bigger (a multiple of 1.5-3) than HBM KV cache capacity.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/61956823-2928-44ec-b00c-8bb5581b8a13_2048x892.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/70a7a7c7-7918-459c-a4e0-f7fc1edca816_2048x887.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference/agentic/439967)"
+/>
+
+H200 SGLang FP8 is able to serve DeepSeek v4 at low concurrency, and is even competitive with B200/MI355X SGLang from a perf/$ standpoint. However, it cannot compete with the newer SKUs in high throughput scenarios due to lack of HBM.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/770b205f-12a1-4d34-81a7-fb235742d005_2048x1305.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=b200_sglang%2Ch200_dynamo-sglang%2Cmi355x_sglang)"
+/>
+
+Furthermore, the reliance on DRAM KV offloading at higher concurrencies leads to unreasonable latency as the number of users scales.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/91dfd9ca-d13d-43e9-a57e-3be0fcc4a5cf_2048x1315.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=h200_dynamo-sglang&i_optimal=0)"
+/>
+
+Overall, MI355X performs decently well compared to its main competitors B200 and B300. Performance is most comparable at the lower throughput / lower latency parts of the curve, where only tensor parallelism and more rudimentary kernels are deployed. AMD needs to work on optimizing DEP kernels on MI355X to be more competitive in the high throughput scenarios, especially given the 1.5x HBM over B200.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/ba6532b3-aa46-4477-be84-8979b1cce35d_2048x1297.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/29cfd340-d61b-46c4-8d87-e081d937b80c_2048x1303.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_vllm%2Cb300_vllm%2Cmi355x_atom%2Cmi355x_vllm&i_label=0)"
+/>
+
+## Kimi K3 2.8 Trillion Parameters
+
+Kimi K3 is another frontier open weight model from China that has 2.8 Trillion total parameters. This is in the same range in terms of number of parameters vs Claude’s Mythos/Fable5 model architecture. We use this as an open weights proxy model architecture. The Kimi K3 model is so big that it does not even fit on a single B200 server and requires using wide EP/wide TP or pipeline parallelism in order to fit all of the weights. On vLLM, speculation decoding/DSpark did not compose at all with pipeline parallelism until very recently, so B200 performance on Kimi K3 was horrible and was getting mogged by MI355X since B200 was unable to use speculative decoding with pipeline parallelism.
+
+MI355X vLLM worked out of the box on day 0 for short context single turn workloads, but for long context multi turn workloads, MI355X AITER and Triton kernels suffered a massive panic attack on the first week and upstream vLLM was completely unusable for MI355X on realistic workloads.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a1440922-e1ea-4a74-ab62-f1263ae31ca0_2048x1473.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+Hopper struggles to serve the AgentX workload for Kimi K3 since Kimi is a massive model and because vLLM maintainers/NVIDIA have not been focusing on optimizing Hopper for Kimi K3. Hopper (SM90) requires custom tuned kernels for K3 along with TP32/EP32 tuned shapes for serving at high interactivity.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ce2b58fc-b7b7-46c2-ae05-baa14419d233_2048x1459.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+We think it is great that AMD is quickly pushing K3 performance forward with ATOM. However, we encourage AMD to further prioritize upstreaming these improvements into vLLM. ATOM is currently AMD's best-performing engine, but vLLM remains the more relevant comparison for customers using an upstream open-source serving stack.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/80eb810f-a9ea-4f01-a41b-e19e3409ce74_2048x1459.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/73ded698-c86e-411b-8e7c-f9bb125487b7_2048x1459.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+On part of the curve between 40 to 60 second e2e latency, MI355X ATOM beats even GB300 NVL72 vLLM on performance per dollar.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/27a3f80e-7959-4651-90eb-f498989908a3_1934x1284.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+## MiniMax M3
+
+Nvidia absolutely destroys all competitors on MiniMax M3 432B. AMD software performance is horrible on MiniMax especially at high context length due to AMD engineering leadership incentivizing tuning only for short context single turn workloads and ignoring long context multi turn workloads.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/c5396c3d-1823-4818-83de-d89cd8242451_2048x1286.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/5179bd8e-c147-46c8-aeb4-b347049b28a7_2048x1303.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft)"
+/>
+
+B300 TRT-LLM TP2 owns the M3 crown. There is a lack of DP-attention points as it is non-optimal on M3 since KV cache locality becomes a routing constraint. This is further explained later on. For GB200 at concurrency 40, TP4/EP4/DPA gets 0.60x the throughput of plain TP4 at a >3x p90 TTFT. At concurrency 32 it hits 28.8% of cache vs 96.0% theoretical. Each DP rank owns a private quarter of the pool; a 300k-token session re-landing on the wrong rank recomputes everything. No decode config with EP appears on the M3 frontier, likely as the concurrency is not high enough to balance the loads on all experts.
+
+B200/B300 also completely beat their rack-scale counterparts for MiniMax M3 on TCO-normalized throughput. On AgentX, the rack-scale advantage isn’t as pronounced as the Dynamo router can become the bottleneck because its work scales with the number and length of live prefixes. Optimizations on this and the several fixes which moved throughput by double-digit percentages are discussed later on in the article. Also, there are no well tuned kernels for wideEP, wide DCP, nor wide TP. And since GB200/300 have higher TCO so without wide ep/wide DCP, it shows up as worse perf per TCO.
+
+With that being said, we also expect further optimizations from Nvidia on their rack scale solutions for this SKU. We will make sure to highlight these in our follow up article.
+
+No submission currently runs context parallelism, despite P90 ISL of 317k. With 4 KV heads, DCP caps at 2 even at TP8, and the MSA indexer needs its own context-parallel handling (a vLLM PR is opened), see the Context Parallelism section for more discussion on this topic.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/aec34e63-34fd-4c42-87a1-97e5c14661ed_2048x1302.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2b40624a-223c-44db-a403-0be563b2af82_2048x1298.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_hc=1&i_active=mi355x_vllm)"
+/>
+
+All of Nvidia’s Pareto optimal points include KV offload above concurrency 20, but for AMD none of the Pareto optimal points use KV offload to DRAM. AMD also uses KV offload less than Nvidia on the other models. The reason for this is that GPU-to-CPU transfers for CPU KVCache offloading are highly inefficient on AMD vLLM. The hipMemcpyBatchAsync API was missing until ROCm 7.14. Without hipMemcpyBatchAsync, vLLM’s native Simple CPUOffloading requires doing serialized Memcpy from CPU to GPU instead of batching them into larger message sizes.
+
+It is also worth mentioning that vLLM performance is very comparable to TRT-LLM in terms of throughput versus p90 interactivity. Additionally, vLLM performs better in terms of throughput versus p90 TTFT.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/68f7ad68-3886-4980-9b01-390582728e47_2048x1292.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/729199d9-66c0-42b5-b1b8-f228503c1585_2048x1289.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_trt%2Cb200_vllm%2Cb300_trt%2Cb300_vllm&i_hc=1&i_label=0)"
+/>
+
+## Qwen3.5 397B
+
+Qwen3.5 397B uses GatedDeltaNet instead of vanilla attention for every couple of layers. GatedDeltaNet was invented at MIT/Nvidia Research and has a theoretically constant state storage requirement instead of vanilla attention’s linear storage requirements. This means that it has lower storage requirements compared to an equivalent dense attention model. Unlike end to end model training research like the Nemotron disaster, Nvidia Research is great at fundamental research like GDN and LatentMoE which is used on frontier models.
+
+Note that this model’s native max context length is 262k tokens, so we use the [truncated dataset](https://inferencex.semianalysis.com/agentx/cc-traces-weka-062126-256k). This simulates a workload on a smaller model where the max context length would be frequently reached with many compactions, how users would actually use this model.
+
+Qwen3.5 397B is a strong hold for NVIDIA on SGLang versus SGLang, with over 20x better performance at 90 tok/s/user. There is currently zero competition from AMD for Qwen3.5 SGLang.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/c92da639-dcff-4e39-b9b5-49742677e93c_2048x1289.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/519a7cd0-ac51-4a5c-9c45-ba7c41f6d058_2048x1300.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_hc=1)"
+/>
+
+Again, we observe Nvidia over optimizing for interactivity at the cost of TTFT, especially in the case of TRT-LLM. In the graph above, all of the Nvidia SGLang submissions have much lower p90 TTFT when compared to TRT-LLM.
+
+Compared to H100, on Qwen3.5, B300 FP4 has 12x better performance per dollar.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/cc2cf6c7-aa58-4a76-a0a2-ee0457d9243c_2048x1298.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=interactivity&i_hc=1&i_best=0&i_prec=fp4%2Cfp8&i_active=b300_sglang%2Ch100_sglang%2Ch200_sglang)"
+/>
+
+## GLM 5.3
+
+GLM 5.3 builds on top of GLM5.2 744B with additional post training. This is a frontier level model.
+
+In terms of OSS SGLang performance, this is another model where Nvidia again beats AMD on realistic agentic inference performance. At 150 tok/s/user p90 interactivity, Nvidia has up to 5x better cost efficiency, With the current state of AMD software, at 150 tok/s/user, Nvidia’s performance advantage is so great that even if the competitor chip hardware was sold for free (but with providers still of course paying for datacenter hosting and power and other operating costs), cost per token would still be cheaper when using Nvidia.
+
+We look forward to AMD’s performance optimizations in the upcoming AgentX update article in a couple of weeks, which will also include some other very exciting results.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f85bdb45-792c-4330-9745-674b38b03425_2048x1289.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/4a21d140-b588-4b33-941d-7c8090ac42b5_2048x1286.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=GLM-5.2&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_prec=fp4&i_active=b200_sglang%2Cb300_sglang%2Cgb200_dynamo-sglang%2Cgb300_dynamo-trt%2Cmi355x_atom%2Cmi355x_sglang)"
+/>
+
+When looking at ATOM, AMD has better performance per dollar than GB300 NVL72 SGLang and even TRTLLM for some parts of the range of p90 E2E Normalized Interactivity. Great work to the AMD team on these results. Again, we look forward to AMD porting over these optimizations to SGLang. We also are looking forward to NVIDIA quickly optimizing GB300 NVL72 in the coming weeks.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9337ed5b-a320-49f7-b31b-cc4b5751e0ff_2048x1290.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+We take a moment to introduce an experimental metric which we call **E2E Normalized Interactivity**. At a high level, this metric is supposed to evaluate how fast a user experiences responsiveness when considering both TTFT as well as TPS. It is defined by OSL/E2EL. Substituting the fact that E2EL equals TTFT plus OSL times TPOT (in reality only OSL - 1 tokens are decoded), we get the following equation.
+
+This is effectively interactivity (the 1/TPOT portion) plus an additional penalty proportional to TTFT.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9e472a49-d228-4d8d-876a-6bc0d59f9d97_746x594.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Please note that this metric is experimental and is not perfect. For instance, it heavily penalizes high TTFT and doesn’t capture all the nuances of certain optimizations such as PD disaggregation. All submissions for AgentX v1.0 optimize for both regular interactivity and TTFT separately. We will continue working on new north star metrics that reflect all nuances of modern agentic inference.
+
+## AgentX Industry Impact - Optimizations for Agentic Workloads
+
+The most impactful result from AgentX in its first months was not producing open source datasets, instead, it has been the **industry impact of 50+ upstream PRs** created by AgentX partners to optimize real world agentic workloads using AgentX as the north star. AgentX’s real agentic traffic benchmarks not only the raw prefill and decode kernels, but also tests the entire end to end token generation process from KV cache lifecycle, hybrid-attention cache correctness, CPU KV offload, transfer progress, routing affinity, to incremental tokenization, request serialization, and scheduler bookkeeping. All of these steps matter for every production agentic deployment.
+
+This is just a continuation of our ongoing mission to help the ecosystem accelerate improvement and deliver light speed improvement in software. A great example is SemiAnalysis’s multi-year collaboration with AMD’s software development team in which we have been providing continual feedback and input to help to modernize their software development principles. This has not only led to many changes that have accelerated AMD’s progress, but also has been instrumental towards getting AMD open source closer to first class on agentic workloads.
+
+## A Brief Introduction to the Distributed Inference Ecosystem
+
+As mentioned, agentic inference is inherently a system-wide problem as opposed to just a chip/kernel level problem. Additionally, when there are large _distributed systems_ handling hundreds of thousands of agentic requests, the scheduling of requests and management of KV cache becomes non-trivial and has legitimate performance implications. For example, sub-agents give bursty KVCache patterns where not properly optimizing, it will improperly evict the main agents cache.
+
+The following diagram illustrates the stack at a high level. At the top, routers (sometimes referred to as “frontends”) route requests to different workers. For instance, in the case a server is running data parallel attention, there are separate KV caches for each DP rank. In order not to thrash any one of the KV caches, requests are routed according to different policies such as [consistent hash](https://github.com/vllm-project/router/blob/main/src/policies/consistent_hash.rs), where requests in the same session/subagent are routed by their unique ID.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/58ac4a3a-cc05-4ff6-9d1e-b28dc0d0f760_1814x2048.png"
+  caption="Source: SemiAnalysis"
+/>
+
+For most routing policies, there is nothing significantly different about each router implementation. Some are separate components such as [vLLM router](https://github.com/vllm-project/router) and [llm-d router](https://github.com/llm-d/llm-d-router) while others are integrated in the engine such as [SGLang model gateway](https://github.com/sgl-project/sglang/tree/main/sgl-model-gateway) and [ATOM Mesh](https://github.com/ROCm/ATOM/tree/main/atom/mesh).
+
+After a request is routed, it is handled by the scheduler of an inference engine such as vLLM, SGLang, etc. The engine is responsible for actually performing the inference and returning the result over an API. Additionally, each engine has an interface for connecting the engine’s internal KV cache to external KV cache managers. This allows for a “pluggable” ecosystem where different KV cache managers can integrate with a variety of inference engines.
+
+A simple deployment, used in the current AgentX results, [runs Mooncake](https://pypi.org/project/mooncake-transfer-engine/) alongside vLLM on the same node. Each vLLM worker embeds a Mooncake Store client and contributes a portion of host DRAM to the external KV-cache pool. vLLM connects to this pool through the MooncakeStoreConnector interface which loads reusable KV blocks into GPU memory and saves newly computed blocks back to host memory. Mooncake Store manages the external cache, including placement and eviction, while Mooncake Transfer Engine performs the actual movement of data between GPU and CPU memory.
+
+Different KV cache managers may use different transfer engines to physically move bytes between memory tiers or machines, such as between prefill and decode workers. Mooncake Store, for example, uses [Mooncake Transfer Engine](https://github.com/kvcache-ai/Mooncake/tree/main/mooncake-transfer-engine) to move KV blocks between GPU memory, host DRAM, and remote nodes.
+
+A deployment can use Mooncake Store to offload reusable KV blocks to host DRAM while simultaneously using NIXL to transfer request-specific KV directly from prefill GPUs to decode GPUs. Mooncake TE handles movement for the Mooncake Store path, while NIXL handles the separate prefill-decode path using UCX and GPUDirect RDMA where supported. Multiple KV-management and transfer paths can therefore coexist within the same inference engine.
+
+The ecosystem consists of many independent components, including inference engines, routers, KV-cache managers, data-transfer libraries, and cluster controllers. Platforms such as Nvidia Dynamo, llm-d, and AMD Infera “package” selected combinations of these components into complete software distributions. They publish compatible container images, connectors, deployment manifests, and orchestration logic that allow the components to be deployed and operated as one system. The resulting product is usually a collection of coordinated containers rather than a single monolithic service (for instance: Dynamo, llm-d, and Infera are typically deployed on k8s and co-ordinate large, distributed systems).
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/af5c4776-221b-43b4-b2fd-d46e46f1ab38_1772x960.png"
+  caption="Source: RedHat / llm-d"
+/>
+
+## Context Parallelism
+
+Long context benefits parallelism techniques that a fixed 8k prompt cannot strongly exercise, because at 8k there is little to divide and TTFT is already short. Moreover, parallelism strategies like TP and DP attention are not optimal at longer context lengths, TP can result in the full KV being replicated on each rank. Although KV is shared for DP attention, it can get hung up on longer contexts, as long context workloads also result in a higher variance of possible context lengths.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/1e2801db-53ff-4029-8b55-01b3c1a7037d_1632x1140.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Context parallelism is a parallelism technique that splits query tokens across GPUs. It comes in two forms: PCP for prefill context parallelism and DCP for decode context parallelism. In PCP, each rank prefills its query chunk (KV ring-passed), since prefill tends to be compute-bound, this parallelizes FLOPs resulting in faster prefill with no giant-prompt prefill spike on one rank. For DCP, each rank scans its KV shard, and the partial attention is then merged flash-decode style. Since decode is memory-BW-bound, parallel KV reads can result in faster tok/s.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/a9544ced-f418-44e0-8910-0cbc63fa26ec_2048x768.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/512c3a96-2ab4-452f-be6a-4fa60900575c_2048x651.png"
+  caption="Source: SemiAnalysis"
+/>
+
+This parallelism technique was invented partially by Nvidia Research. Nvidia Research is great at fundamental research like this while for end to end training research, they are embarrassing America with their horrible Nemotron3 Ultra model which is currently getting massively beaten by even tiny Qwen3.8 27B model. DCP/PCP forms part of CUDA moat as the AMD implementation of DCP/PCP isn’t optimized yet. In the [vLLM support matrix](https://docs.vllm.ai/en/latest/design/attention_backends/#standard-attention-mha-mqa-gqa-backends), every single AMD backend is unsupported.
+
+A few of the changes mentioned in the next few sections focus on DCP/PCP.
+
+## vLLM Agentic Optimizations
+
+Working alongside vLLM maintainers from Inferact, Red Hat, NVIDIA, and AMD, we used AgentX’s realistic replayer as a north star, with the resulting fixes landing upstream where most of the optimizations are highly transferable to production. A few examples follow:
+
+vLLM improved hybrid-attention prefix caching so short-lived sliding-window allocations do not evict useful long-context checkpoints. [Selective retention](https://github.com/vllm-project/vllm/pull/43447) preserves sparse replay boundaries and reported a prefix-cache hit rate above 95% with fourteen concurrent requests and contexts up to one million tokens. The same reachability policy [was applied to Mooncake](https://github.com/vllm-project/vllm/pull/44774), and [unreachable sliding-window lookups were removed](https://github.com/vllm-project/vllm/pull/45444). Earlier follow-up work also [stopped offloading sliding-window blocks that could never be reused](https://github.com/vllm-project/vllm/pull/42258) and [kept the speculative lookahead block in the retained prefix](https://github.com/vllm-project/vllm/pull/44082).
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/df1a505d-b8c3-45f9-b03c-a8798fba54c0_2048x632.png"
+  caption="Source: SemiAnalysis and vLLM GitHub PR for Agentic Workloads"
+/>
+
+Agentic Workloads for high concurrent agents require offloading. Thanks to the aforementioned influence of AgentX, there are already workstreams on vLLM that aim to allow CPU KV offload for hybrid models rather than only for uniform full-attention models. This distinction matters because a uniform model has one KV layout per token, so a connector can describe what to save with a single block geometry. A hybrid model carries several cache groups at once, each with a different shape and a different lifetime, and a connector that assumes one uniform layout cannot express which group a given block belongs to. Offload was therefore unavailable for the models whose long sessions needed it most. The general [SimpleCPU connector](https://github.com/vllm-project/vllm/pull/37160) came first, was [enabled on ROCm](https://github.com/vllm-project/vllm/pull/40549), and was then [extended to DeepSeek-V4 hybrid attention](https://github.com/vllm-project/vllm/pull/42296), reporting 81.7% higher output throughput and 46.6% lower mean e2e latency against recomputing the prefix once it no longer fits in HBM.
+
+Mooncake has also gained [equivalent hybrid-memory allocation support](https://github.com/vllm-project/vllm/pull/42828). The same layout problem resurfaces in disaggregated serving, where a [pending change](https://github.com/vllm-project/vllm/pull/51052) transfers Kimi-K3’s conv+ssm recurrent state alongside the attention KV over MoRI-IO for 1P1D prefill/decode splits; without it the decode side starts from an uninitialized recurrent state. The recurrent-state slot rides the existing remote-block-ids channel, so the disaggregation router needs no model-specific changes, and the path was exercised end-to-end on MI355X, TP8 per leg over cross-node RDMA, with DSpark speculative decoding on both legs.
+
+When profiling realistic workloads, vLLM maintainers noticed that during offload, the cost moved to the store path, which was writing too much and too often. Three new fixes have now addressed this issue:
+
+A store is now [skipped while an identical transfer is already in flight](https://github.com/vllm-project/vllm/pull/41289), so concurrent sessions sharing a prefix pay for it once rather than once each. A store [covers only newly generated KV ranges](https://github.com/vllm-project/vllm/pull/46412), so a session that extends its history writes the delta instead of rewriting the whole prefix on every turn. Finally, a store no longer [depends on whether the same blocks still sit in HBM](https://github.com/vllm-project/vllm/pull/46906), so work already scheduled is not discarded when an eviction lands underneath it.
+
+The load path was tuned separately, because lookups happen on every scheduling decision rather than only when data actually moves. Making [lookups asynchronous in the scheduler path](https://github.com/vllm-project/vllm/pull/45659) keeps the connector off the step’s critical path, so a step no longer waits on CPU-side cache queries before it can admit work. [Compact zero-copy lookup keys](https://github.com/vllm-project/vllm/pull/45969), [parallel receive-side loading](https://github.com/vllm-project/vllm/pull/45971), and [prebuilt Mooncake key strings](https://github.com/vllm-project/vllm/pull/46188) then removed the CPU and transport overhead that remained.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2c2cfea4-a217-47ad-934b-dbd226bb248c_2048x837.png"
+  caption="Source: SemiAnalysis and vLLM GitHub PR for Agentic Workloads"
+/>
+
+Long-lived hybrid state also forced correctness and accounting fixes that fixed-shape requests rarely reach. vLLM [emits cache events per hybrid cache group](https://github.com/vllm-project/vllm/pull/44103), [strides distributed-context stores correctly](https://github.com/vllm-project/vllm/pull/45371), and [computes lookup prefixes correctly under distributed context and prefill](https://github.com/vllm-project/vllm/pull/46855). A related [context-parallel accounting change](https://github.com/vllm-project/vllm/pull/45340) aligns cache ownership with sharded token ranges. Speculative state is now [propagated across merged Mooncake groups](https://github.com/vllm-project/vllm/pull/49069) and [through the SimpleCPU coordinator](https://github.com/vllm-project/vllm/pull/49071), preventing the repeated-turn cache from silently losing EAGLE state.
+
+Looking at the ROCm side of optimizing Agentic workloads in vLLM, the work continues below the cache layer, where the remaining cost is per-layer rather than per-request. Once the prefix survives and arrives on time, what is left is the decode step itself, and a decode step that runs thousands of times per session pays for every avoidable copy and every mismatched kernel.
+
+Three open changes attack that layer:
+
+- A Kimi-K3 change [writes KDA decode results directly into the layer output buffer](https://github.com/vllm-project/vllm/pull/51183), removing one device copy per KDA layer; the saving is small in isolation and repeated for every layer of every decoded token.
+
+- A second change [selects an AITER sparse-MLA decode kernel](https://github.com/vllm-project/vllm/pull/51714) in place of the generic path, and reported 5.22% higher AgentX output throughput with substantially lower inter-token latency.
+- The third is a useful illustration of how much the measurement shape matters. A companion change [routes full-graph attention projections through tuned AITER GEMMs](https://github.com/vllm-project/vllm/pull/51713) and reached a 2.3% gain on fixed sequences at low concurrency. This shows a kernel-level change can show a clean gain on uniform shapes and then be swamped, on an agentic trace, by the cache and scheduling variance that the trace introduces. A [pending change](https://github.com/vllm-project/vllm/pull/52882) turns that shape sensitivity into the dispatch criterion itself: it replaces the DeepSeek V4 C4A selector’s ROCm top-k bottleneck with a hybrid AITER/native path on gfx950, routing short and medium contexts through AITER and long contexts through a graph-safe tuned native fallback. It reports end-to-end selector speedups of 1.21x to 1.76x, with decode-kernel geomeans of 1.2x to 2.9x across an 84-shape matrix.
+
+## SGLang Agentic Optimizations
+
+The AgentX team has been working closely alongside SGLang maintainers from RadixArk, Meta, Nvidia, and AMD to drive optimizations to Agentic workloads run using SGLang, resulting in massive improvements in production inference performance. Let’s discuss these optimizations in more depth.
+
+We will start by explaining how, from the allocator side, SGLang’s sliding-window work addresses the same conflict vLLM’s retention policy does. Window pages and prefix pages are drawn from one pool, and the window is the greedier consumer: it turns over constantly while the prefix sits still and so, under pressure, the transient allocation displaces the durable one.
+
+Three design improvements attack this problem that from different angles. One [proactively frees pages as they leave the window](https://github.com/sgl-project/sglang/pull/26907) rather than waiting for eviction pressure to find them, so dead window state stops competing for pages it can no longer use. Another [caps compute locks to a single window](https://github.com/sgl-project/sglang/pull/27210), bounding how much of the pool an in-flight request can hold pinned at once. A third [removes stale full-KV entries](https://github.com/sgl-project/sglang/pull/29369) that outlive their usefulness.
+
+Proactive freeing has a fork-shaped blind spot, though: a request branching from a shared prefix can still hold reusable full-KV while the window state at the branch point has already been released, and the whole prefix gets recomputed for want of the cheap half. [Open work](https://github.com/sgl-project/sglang/pull/34565) preserves the SWA state at those branch points so forks inherit the window instead of rebuilding it.
+
+Alongside those, [the ROCm ring-cache fix](https://github.com/sgl-project/sglang/pull/30339) is a correctness rather than a capacity change: a ring buffer reuses slots by construction, and reusing one whose old contents are still referenced yields wrong output rather than slow output. None of this is visible on a single 8k prompt, where the window never laps the prefix and the pool is never contended. On a multi-turn hybrid session, these changes are what decide whether the expensive full-attention history is still there on the next turn.
+
+HiCache is SGLang’s first class in-tree offloading mechanism. It faced the same hybrid problem vLLM’s connectors did, and solved it with an asymmetry: [offload the full-attention cache and reconstruct the short sliding-window tail on the way back](https://github.com/sgl-project/sglang/pull/29417). Only the expensive half is worth moving across the bus, and the cheap half can be rebuilt more quickly than it can be fetched. On AMD, [staged write-back](https://github.com/sgl-project/sglang/pull/28534) keeps that movement from blocking the engine while it happens. Recurrent state was the remaining gap, because it cannot be rebuilt from neighbouring tokens the way a window tail can; [FlashInfer GDN checkpoints](https://github.com/sgl-project/sglang/pull/29735) let it participate in prefix reuse at all, and raised throughput from 47,771 to 53,004 tok/s/GPU at a 92.4% cache-hit rate.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/078ad449-7657-4468-9ff9-3480a7999d89_2048x720.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Two further changes address how variable-length traffic affects kernel pipeline. AgentX realistic sessions tend to arrive at continuously varying context lengths, a similar pattern observed for production traffic. A naive runtime that specializes on length will compile a fresh kernel for nearly every request it sees. SGLang maintainers solved this by passing [context length as a runtime scalar](https://github.com/sgl-project/sglang/pull/30255) instead collapsing it into one compilation, improving AgentX concurrency 384 output throughput by 26.75% and mean TTFT by 36.25%, by removing compilation, not from computing anything faster.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/3a807d67-383b-4aa9-93af-e8f3a68313f4_855x634.png"
+  caption="[Source: GitHub](https://github.com/sgl-project/sglang/pull/30255)"
+/>
+
+In the same spirit, [removing a per-step device-to-host sequence-length synchronization](https://github.com/sgl-project/sglang/pull/30365) eliminates a decode bubble that exists only because the host wanted to know a length the device already had.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/b29e003e-db43-45d4-93be-bac0be04d869_1209x646.png"
+  caption="[Source: GitHub](https://github.com/sgl-project/sglang/pull/30365)"
+/>
+
+Variable length also bites inside the attention kernel itself. On GB300, a mixed-context decode batch pays a tail tax: a matched profile attributed 8.4 ms of a 9.8 ms decode-step delta to attention, with the longest requests dragging out the persistent kernel’s shared wave. [Open work ](https://github.com/sgl-project/sglang/pull/34888)splits TRTLLM MHA decode batches into KV-length-sorted groups so the short requests stop waiting on the longest ones.
+
+Decode can starve one level up as well, at the scheduler rather than in a kernel. Under DP attention, every rank joins the same MoE collective while scheduling attention work locally, so a rank fed a stream of chunked-prefill continuations keeps winning the prefill-first decision while peer ranks’ running batches sit waiting, observed on AgentX runs. [A configurable decode interval](https://github.com/sgl-project/sglang/pull/35017) after prefill forces decode rounds between prefills; on AgentX DSv4 Pro, output throughput rose 141% and p99 inter-token latency fell 97.3%, at the cost of median TTFT rising from 36.5 to 59 seconds, trading off first-token wait for stream smoothness.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/47ce130f-13f2-4732-8431-17e26184b8c3_1846x806.png"
+  caption="Source: SemiAnalysis"
+/>
+
+When a request carries no reusable history such as the start of a subagent, any worker will do just fine, and load balancing is the only question worth asking. When the request carries a MB of cached prefix, sending it to an idle worker that does not hold that prefix is the expensive choice, and the router needs to know where the state already lives. SGLang added [DP cache affinity](https://github.com/sgl-project/sglang/pull/26091), so a session is sticky to the rank holding its cache. In this PR, [DP-aware prefill and decode routing](https://github.com/sgl-project/sglang/pull/26245) is both implemented so that both halves of a disaggregated deployment make that decision consistently, and [cache balance as a routing signal](https://github.com/sgl-project/sglang/pull/26293) so affinity does not degenerate into one hot worker. A router can only act on what it is told, so hybrid cache events also become [radix-cache aware](https://github.com/sgl-project/sglang/pull/26387) and [sliding-window aware](https://github.com/sgl-project/sglang/pull/26579).
+
+Speculative decoding receives special attention, because MTP adds a second, smaller piece of per-request state that has to survive everything the main cache survives. SGLang [fixed draft-window transfer in disaggregated serving](https://github.com/sgl-project/sglang/pull/30461) so that state crosses the prefill-to-decode boundary intact, [added overlap scheduling for high-concurrency online decoding](https://github.com/sgl-project/sglang/pull/30497), [removed a no-op EAGLE renormalization](https://github.com/sgl-project/sglang/pull/31294), and [avoided host synchronizations during EAGLE prefill](https://github.com/sgl-project/sglang/pull/33662). The open [resource-lease scheduling work](https://github.com/sgl-project/sglang/pull/32042) and [data-parallel graph-metadata fix](https://github.com/sgl-project/sglang/pull/32196) continue the same effort, which is to make overlap safe when requests can be retracted and resumed rather than simply run to completion.
+
+When looking at Agentic workloads with Heterogeneous prefill and decode topologies, prefix-aware staging is needed, and this is where prefix caching and disaggregation interact badly. When the two sides are not sharded identically, KV cannot be copied across as one contiguous stream; it has to be split on a transfer grid and reassembled at the offsets the decode side expects. A prefix hit makes that harder, not easier, because the prefill worker now sends only the uncached remainder while the decode side still expects a complete, correctly positioned cache. [Radix-cache support in the staging buffer](https://github.com/sgl-project/sglang/pull/30545) splits cached sends on that grid and scatters them at the correct decode offsets.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/88c2d49d-2612-4334-90c2-8a76ab55fd60_1418x1004.png"
+  caption="Source: SemiAnalysis"
+/>
+
+However, this results in correctness problems. A 127,500-token shared-prefix test went from 2 correct needles out of 128 to 128 out of 128, meaning the cache had been silently landing in the wrong places, which a throughput benchmark would have scored as a fast, confident, but wrong answer. The AgentX comparison additionally raised median per-user output throughput by 9.6% at nearly unchanged total throughput per GPU. The transfer itself also carried dead weight: decode-side PREBUILT batches never enter a model forward, yet every transferred prompt was still flattened and copied into a CUDA input tensor that the first decode step reconstructs from relay metadata anyway. [Dropping that unused prompt transfer](https://github.com/sgl-project/sglang/pull/35070) is the largest single win, driving +18.0% per-user output throughput and +12.7% decode throughput per GPU on AgentX GB300. A [follow-up](https://github.com/sgl-project/sglang/pull/35071) moved the prefill DP-rank bootstrap query off the decode scheduler’s critical path, overlapping an HTTP round trip that had been paid synchronously at result consumption, driving a further +1.36% per-user output throughput on the same deployment. Open work continues along the same seam: [multi-pool DeepSeek-V4 support in UMBP](https://github.com/sgl-project/sglang/pull/30762), [unified-KV HiSparse state carried over MoRI](https://github.com/sgl-project/sglang/pull/32368), and [preserving the prefill-owned token when decode terminates without visible content](https://github.com/sgl-project/sglang/pull/34216). The HiSparse work should be read as a capacity and correctness enabler for long contexts rather than as a throughput win at high concurrency, which it is not yet.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ed24f1b9-b591-4674-bc94-06394962d03f_2048x544.png"
+  caption="Source: SemiAnalysis"
+/>
+
+## TensorRT-LLM Agentic Optimizations
+
+Next, we shall move onto explaining a couple of the recent optimizations from TensorRT-LLM. First, we look at TRTLLM’s distinct frontend optimization for repeated chat turns, a cost that only exists because the workload is multiturn. Every turn of a conversation re-sends the entire history plus a little more, and the naive implementation re-tokenizes all of it. Tokenization is cheap per kilobyte, but is ruinous when the same 100,000 tokens are tokenized again on every turn.
+
+The obvious fix, to tokenize only the new suffix, is wrong in a way that is easy to miss, because byte-pair encoding is not position-independent. Tokens can merge across the join, so splitting the text at the boundary and concatenating the two token sequences can produce a different sequence than tokenizing the whole string, which quietly diverges from the sequence the prefix cache was built against.
+
+TRTLLM implements [Boundary-aware incremental tokenization](https://github.com/NVIDIA/TensorRT-LLM/pull/17462) which handles this by finding the rendered-text common prefix, rolling back one complete token so any merge that spans the join is recomputed, and tokenizing only the changed suffix from there. On the Qwen3.5 AgentX trace, it matched full tokenization on all 1,087 transitions — the correctness claim, tested rather than assumed — and reduced mean processing time from 185.1 ms to 11.3 ms.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/30a51cf2-e628-4c20-a4a5-5978234ca713_952x541.png"
+  caption="[Source: GitHub](https://github.com/NVIDIA/TensorRT-LLM/pull/17462)"
+/>
+
+A fixed 8k1k request has no prior rendered turn to reuse, so none of this appears there. Relatedly, [chat-template rendering was moved into the input-processing pool](https://github.com/NVIDIA/TensorRT-LLM/pull/16231), so a long template no longer serializes the main request loop behind it.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/00b459b8-9c4c-4c8c-8659-07e14bf5699f_2048x784.png"
+  caption="Source: SemiAnalysis"
+/>
+
+The MiniMax-M3 work focuses on disaggregated KV movement, where the failure is one of granularity. When prefill and decode do not agree on head layout, the KV for one logical request stops being a few large contiguous regions and becomes thousands of small strided pieces, each of which turns into its own transfer descriptor. The bytes moved are unchanged; the per-descriptor overhead is what explodes, and it explodes in the worst way on exactly the long prompts that matter! [Corrected multi-pool mapping and a chunked NIXL bounce path](https://github.com/NVIDIA/TensorRT-LLM/pull/17518) coalesce those pieces through a bounded reusable arena, trading an extra staging copy for orders of magnitude fewer descriptors. Its AgentX diagnostic reduced request-critical KV p99 from 26.74 seconds to 125 ms at concurrency five, and from 10.15 seconds to 288 ms at concurrency forty.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/52bd05ab-914f-4dc5-9f16-a0ea1b495383_959x501.png"
+  caption="Source: [GitHub](https://github.com/NVIDIA/TensorRT-LLM/pull/17518)"
+/>
+
+[Nonblocking context-transfer polling](https://github.com/NVIDIA/TensorRT-LLM/pull/17428) protects the same path by reaping completed transfers even when scheduling stalls. This breaks a feedback loop in which finished KV blocks stay pinned and prevent new admissions. The stalls themselves also had a [removable cause](https://github.com/NVIDIA/TensorRT-LLM/pull/16734): avoiding implicit device-scalar syncs in DeepSeek-V4’s context sparse-attention metadata eliminated 18 four-byte device reads per step that each forced a cudaStreamSynchronize, on a GB300 disaggregated context worker. The fix threads host-side counts through as plain Python ints, so the executor thread no longer holds the GIL for most of a step while the KV-transfer and response threads wait behind it.
+
+TensorRT-LLM also moved irregular long-context work onto more efficient execution paths. [Context graph producers for MiniMax-M3](https://github.com/NVIDIA/TensorRT-LLM/pull/17473) capture stable sparse producers while leaving request-dependent attention eager, and per-user output throughput improved by 12.58 percent in its AgentX test. An open [native KV-event production change](https://github.com/NVIDIA/TensorRT-LLM/pull/16876) reduces allocation and conversion work on the KV-aware routing path.
+
+AgentX also exposed kernel-selection and scheduler-lifetime failures that only appear at scale and duration. Two are about which kernel gets picked. MiniMax-M3 [added CuTeDSL choices to MXFP8 autotuning](https://github.com/NVIDIA/TensorRT-LLM/pull/17316), widening the candidate set and improving output throughput per GPU by roughly 7 to 10% at low-concurrency aggregate points. In the opposite direction, TensorRT-LLM [disabled corrupt split-K MoE tactics](https://github.com/NVIDIA/TensorRT-LLM/pull/17105) after it crashed five of seven AgentX runs, with no crashes in seven matched runs afterwards. A tactic that is fast and wrong is worse than one that is merely slow, and an autotuner will select it enthusiastically unless it is removed from the pool. Selection is not the only way a kernel goes wrong: MiniMax-M3’s legacy sparse-attention path for short queries could hand the SM100 kernel a non-contiguous, head-major block-index view that it read as contiguous, selecting the wrong KV pages and producing incorrect or non-finite output. [Honoring the block-index strides](https://github.com/NVIDIA/TensorRT-LLM/pull/17285) for q_len ≤ 32 fixed the indexing without materializing the tensor or adding a kernel, and five matched full AgentX pairs on GB300 afterwards completed with zero serving errors and no non-finite markers.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f48290ac-3361-4b0c-a085-ff84dd3dce51_2048x1347.png"
+  caption="Source: [TRT Github](https://github.com/NVIDIA/TensorRT-LLM/pull/17105)"
+/>
+
+The other two are lifetime bugs, which are the characteristic failure of long runs rather than large ones. [Sequence-slot headroom and consistent slot-indexed buffer sizing](https://github.com/NVIDIA/TensorRT-LLM/pull/16279) handle the transient overlap where a completing request and a newly admitted one both need a slot, a window that a steady stream of arrivals and departures hits constantly and a fixed batch never hits at all. A later [attention-data-parallel dummy-request fix](https://github.com/NVIDIA/TensorRT-LLM/pull/17278) kept nine Qwen3.5 disaggregated cells alive where most earlier cells had failed within minutes, the difference between a configuration that benchmarks and one that survives a session.
+
+Two open transfer changes target very long disaggregated prompts, and together they show how a fix can create the next bottleneck. In the default arrangement, a decode worker cannot start until the entire prompt has been prefilled and then transferred, so two expensive phases run back to back even though the first produces its output incrementally. [Pipelined KV transfer](https://github.com/NVIDIA/TensorRT-LLM/pull/15727) begins sending each completed prefill chunk as it lands, so transfer overlaps prefill compute and only the final chunk is on the critical path.
+
+That change makes chunk handling frequent, which exposes work that used to happen once. Its follow-up [retrieves only the block IDs belonging to the current chunk](https://github.com/NVIDIA/TensorRT-LLM/pull/17526) rather than the whole prompt’s block list each time. For a 128,000-token prompt split into 1,024-token chunks, that is the difference between building a 4,096-entry list once and rebuilding it 128 times for every layer group. A per-chunk cost that scales with total prompt length is a cost scaling shape that eats the gain the pipelining just bought.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e8eeb92e-1aad-4a97-b3c1-28409005809e_2048x769.png"
+  caption="Source: SemiAnalysis"
+/>
+
+## AMD ATOM Agentic Optimizations
+
+AMD’s ATOM engine was originally designed only for single turn workloads instead of real world agentic multi-turn production workloads, so there were a lot of changes needed to the core fundamental ATOM engine and kernels to enable good support for long context multi turn workloads. ATOM still has a long way to go to support agentic workloads relative to where vLLM/SGLang are at present. AgentX is used as the realistic north star target for ATOM’s refactor to support agentic workloads. The first optimization we will chat about that ATOM implemented is smartly using sparse checkpoint retention for DeepSeek-V4 paged sliding-window attention. [The merged implementation](https://github.com/ROCm/ATOM/pull/1640) keeps selected window tails alive so branch and replay requests can resume at useful boundaries. Its measurements separate the two effects cleanly: on the same AgentX trace at concurrency 48, the actual prefix hit rate rose from 5.6% to 96.45%, and losses at the sliding-window gate fell from 91.35% to 0.16%. The second number is the mechanism behind the first. Nine out of ten prefix matches were being found and then discarded for want of a window tail, so the cache was not missing but being overruled.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/5c880335-e1e3-44d9-a123-8e316c9c3808_822x555.png"
+  caption="Source: GitHub"
+/>
+
+Two earlier cache-manager fixes had to land before any of this could be measured, and both fixes are worth noting as examples of a cache that reports itself healthy while doing nothing. One [stopped free-pool hits from destroying shared cache entries](https://github.com/ROCm/ATOM/pull/902); the other, [a deferred-output fix](https://github.com/ROCm/ATOM/pull/939), restored prefix hashing in the default scheduler mode and moved repeated long prompts from zero cached tokens to reuse of every complete prefix block. A separate change lets prefix-hit prefill [stay on the optimized sink attention kernel](https://github.com/ROCm/ATOM/pull/1345) rather than falling back to the generic path, so a cache hit does not quietly cost part of what it saves.
+
+Hybrid models also carry a recurrent or compressor state, which differs from ordinary KV in one decisive way: it cannot be reconstructed from the tokens around it. A window tail can be recomputed from neighboring context, but recurrent state is the accumulated result of everything that came before, so if it is dropped, the only way back is to replay the sequence. ATOM [gave this per-request state a content-addressed checkpoint lifecycle](https://github.com/ROCm/ATOM/pull/1771), letting generated turns leave reusable resume points without reserving a separate protected cache for them. In one test, a request reused 512 generated tokens and computed only a two-token suffix.
+
+The tuning detail matters as much as the feature. Publishing a checkpoint unconditionally costs 17.5% throughput on zero-hit traffic, the price paid by every session that never comes back, in order to help the ones that do. Spacing checkpoints by token interval avoided that penalty, and fixed 1k1k throughput stayed within measurement noise, which is the relevant safety property: a feature aimed at agentic reuse should not tax workloads that will never use it.
+
+ATOM’s AgentX-relevant CPU path starts from the arithmetic that justifies offloading at all. [Standalone LMCache offload](https://github.com/ROCm/ATOM/pull/1318) reloads a 32,000-token prefix from CPU in about 0.32 seconds against roughly 2.5 seconds to recompute it, an eight-fold margin. This makes crossing the bus worth doing at these context lengths and would not hold for a short prompt.
+
+The rest of the path is about ownership and index placement rather than bandwidth. ATOM copied vLLM’s [multi-connector](https://github.com/ROCm/ATOM/pull/1406) design which lets a prefill worker send KV to a remote decode worker and save the same prefix to CPU at once, without freeing the blocks until both consumers are finished; two independent readers of the same blocks is a situation where single-turn won’t get.
+
+[Promoting restored blocks back into the GPU prefix index](https://github.com/ROCm/ATOM/pull/1725) fixes a subtler waste: without it, a prefix loaded from CPU is used and then not registered as resident, so the next turn fetches the same hot prefix across the bus again, paying the transfer repeatedly for a cache that was already in HBM. Follow-up work [fixed asynchronous save ordering, packed-KV geometry, unaligned handoffs, and remote request accounting](https://github.com/ROCm/ATOM/pull/1807) together, eliminating reload corruption across a two-round, 2,638-request validation. This bug surfaces only when the same blocks are saved, evicted, and restored many times over.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a7bb7f74-e5a3-4da3-b954-89905e54c4da_2048x420.png"
+  caption="Source: SemiAnalysis"
+/>
+
+The distributed path repeats, in a different codebase, and the pattern is already visible in SGLang and Dynamo: routing has to know where state lives. ATOM’s router called ATOMesh is a fork of SGLang’s router with most of the features removed. Unfortunately, ATOMesh required SGLang’s cache-aware routing feature, so that feature had to be added back. ATOM gained KV lifecycle events for cache-aware routers, so the router can know where state lives at all. It also gained multi-node prefill and decode routing, and session-sticky data-parallel routing. The sticky policy is a two-sided compromise worth stating explicitly: a conversation returns to the healthy worker that owns its state, but idle assignments expire so that stickiness does not permanently unbalance the cluster on behalf of sessions that have gone away.
+
+Disaggregation then has to move whatever the model actually keeps, which is not always one uniform cache. DeepSeek-V4 [transfers both buffers of its mixed FP8 and BF16 cache layout](https://github.com/ROCm/ATOM/pull/1737), and EAGLE disaggregation [moves the draft model’s independent KV cache](https://github.com/ROCm/ATOM/pull/1331) alongside the target cache, the same second-cache problem TensorRT-LLM and SGLang each had to solve. [Remote-KV admission and backpressure](https://github.com/ROCm/ATOM/pull/1647) closes the loop by stopping the decode side from accepting more parked transfers than it can safely resume, which is the disaggregated form of accepting work you cannot finish.
+
+On ATOM, [PCP reported 35 to 43% lower mean TTFT](https://github.com/ROCm/ATOM/pull/1220), with total throughput gains of up to about 49% at a 64,000-token input - a gain that grows with input length rather than with batch size. Making that usable in practice required it to compose with everything else a session relies on, so DCP was [made compatible with prefix caching, chunked prefill, and FP8 KV](https://github.com/ROCm/ATOM/pull/1701) and then [extended to MTP](https://github.com/ROCm/ATOM/pull/1746). Parallelism that cannot coexist with the prefix cache would trade one long-context win for another. The same scarcity of parallelism exists inside a single GPU: a batch-1 MLA decode has no head or query dimension to spread, only the KV walk, and a hardcoded split budget of 16 left that walk running on 16 of a gfx950’s 256 CUs. A [still-open change](https://github.com/ROCm/ATOM/pull/1911) stops overriding the kernel’s own split derivation, so Aiter cuts the walk into as many parts as the machine has clusters.
+
+[Chunked pipeline-parallel prefill](https://github.com/ROCm/ATOM/pull/1552) attacks the same problem from the memory side, replacing repeated tensor-parallel collectives with streamed layer-stage handoffs. Its GLM-5.2 result at high load is the most complete in this section: output throughput doubled, median time to first token fell from 28.6 seconds to 8.7 seconds, and each prefill GPU held 3.68 times as many KV blocks. That last figure is the one to read first, because capacity per prefill GPU is what decides how many long sessions can be in flight before the deployment hits the HBM cliff at all.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/8184b9e6-883b-446c-884b-226b5e9fcca1_2048x549.png"
+  caption="Source: SemiAnalysis"
+/>
+
+## ROCm AITER Agentic Optimizations
+
+ATOM/AMD vLLM/AMD SGLang’s long-context execution depends on matching lower-level AITER kernels, because a parallelism strategy at the engine layer is only real if the kernels can express it. [Prefill context-parallel process groups](https://github.com/ROCm/aiter/pull/3728) provide the extra query-sharding dimension that prefill context parallelism needs, and also widen fused-kernel row indexing for prompts above 131,000 tokens. [Decode context parallelism](https://github.com/ROCm/aiter/pull/3267) (DCP) shards KV across the tensor-parallel GPUs already present, so a longer sequence or a larger batch fits without replicating the whole cache on every rank.
+
+Large caches also exposed a class of failure that short fixed requests essentially never reach: address width. A 32-bit offset is entirely adequate until a single cache pool crosses the boundary, at which point the arithmetic wraps and the kernel addresses the wrong row without any error being raised. AITER added [runtime 64-bit dispatch for batch prefill above 4 GB](https://github.com/ROCm/aiter/pull/2893), [64-bit MLA offsets above 2 GB](https://github.com/ROCm/aiter/pull/4474), and [64-bit addressing throughout DeepSeek-V4’s unified cache paths](https://github.com/ROCm/aiter/pull/4680), the last preventing silent reads and writes to the wrong row in pools of roughly 150 million rows.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ce356c05-883b-4c63-a359-8cfd72dddb63_1408x1046.png"
+  caption="Source: SemiAnalysis"
+/>
+
+DeepSeek-V4 decode also gained [a persistent MLA kernel for 64-head and 128-head MTP packings](https://github.com/ROCm/aiter/pull/3459). Those two head counts are what ordinary decoding and speculative verification actually produce, so this gives the engine a dedicated long-context path for its common shapes instead of treating them as incidental variants of a kernel written for short contexts. It is the same argument as the vLLM AITER sparse-MLA selection above: in long context, the generic path is not a modest compromise, it is the wrong kernel.
+
+## Dynamo Agentic Workload Optimizations
+
+A good chunk of Nvidia submissions use the Dynamo Inference Orchestration and Router Systems. Dynamo’s AgentX series shows that the distributed serving layer can become the bottleneck once engine kernels improve. The router’s work is proportional to the number and length of live prefixes rather than to the number of tokens generated, so a workload of many long, overlapping, long-lived sessions loads it in a way that fixed-shape traffic never does. The first series of PRs reduced the cost of each routing decision: [less work on the lookup hot path](https://github.com/ai-dynamo/dynamo/pull/10540), [no redundant suffix invalidation](https://github.com/ai-dynamo/dynamo/pull/10836), and finally [batched KV matching, registration, ownership, and terminal dereferences](https://github.com/ai-dynamo/dynamo/pull/11095), which reported a 22.2% median output-throughput gain at concurrency 512. Batching helps here for the same reason it helps in an engine: the per-item overhead was dominating the item.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/7af8faf5-300e-4ee5-bc0c-040b7630e923_1016x671.png"
+  caption="Source: GitHub"
+/>
+
+The second series of PRs changed how ownership is represented, which is the harder problem underneath. Every cached block needs to be attributed to the requests relying on it, so it is not freed while still in use and not pinned after everyone has finished. With thousands of concurrent sessions sharing overlapping prefixes, the bookkeeping itself becomes significant. Dynamo moved from [shared block chains](https://github.com/ai-dynamo/dynamo/pull/11503) to [arena-level ownership counts](https://github.com/ai-dynamo/dynamo/pull/11508) and finally to [backend-specific request leases](https://github.com/ai-dynamo/dynamo/pull/12329), each step coarsening the unit being tracked. The lease design reduced AgentX replay time by 23.7% for the vLLM backend and 22.0% for SGLang, and lowered peak memory at the same time, a sign that the previous representation was the problem rather than the traffic.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e3039950-98e0-4cd0-b7fe-929a120bd7d2_2048x576.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Further router profiles removed costs with the same shape, where a periodic sweep or a full recomputation had been acceptable only because live state used to be small. [Bucketed expiry pruning](https://github.com/ai-dynamo/dynamo/pull/10521) replaced a scan proportional to everything tracked and improved high-churn AgentX throughput by 13.7%. [Delta-only suffix cleanup](https://github.com/ai-dynamo/dynamo/pull/10676) processes only what changed and absorbed about 28 times as many store and remove events in the same window. [Compressed prompt paths](https://github.com/ai-dynamo/dynamo/pull/11644) cut front-end CPU by 35.3% and materially improved tail time to first token, which matters because prompts in this workload are long and largely repeated. Overload state is now [tracked incrementally](https://github.com/ai-dynamo/dynamo/pull/10645) rather than recomputed.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/878dd41d-fc0f-407e-8421-ad05ad9e740c_1087x624.png" />
+
+One routing change is a deliberate trade rather than a pure win. Dynamo can now [charge active decode requests in its routing score](https://github.com/ai-dynamo/dynamo/pull/12158), so a worker already committed to long-running decodes looks more expensive than its queue depth alone suggests. That improved median AgentX latency at a small throughput cost in the reported tuning point, which is the kind of choice that only becomes visible when requests occupy a worker for a long time. An [open follow-up](https://github.com/ai-dynamo/dynamo/pull/13447) packages that trade into an new agentic router preset that pushes further in the same direction, crediting prefix overlap at 2, scaling prefill load by 4, and weighting active decode requests at 64. At that tuning point, the trade stops costing throughput: on an 8xH200 AgentX run, the preset improved fixed-window completed-output throughput by 8.26% over the default cost function, cut run-level p95 time to first token by 43.1% and p95 inter-token latency by 22.6%, and completed one more full trajectory.
+
+The request plane was optimized next, because an agentic trace does not send one request and one response. It sends many related requests carrying largely identical prompts, and streams every token back as its own frame, so serialization and copying are paid per turn and per token rather than once. Switching to [MessagePack request payloads](https://github.com/ai-dynamo/dynamo/pull/10437) improved throughput by 8.1% and reduced average time to first token by 9.7% in its AgentX test, and [direct Python transcoding](https://github.com/ai-dynamo/dynamo/pull/11104) removed an intermediate value tree from that path entirely.
+
+What followed is a sequence of changes that all remove a copy rather than speed one up: not copying [MessagePack event payloads](https://github.com/ai-dynamo/dynamo/pull/11539), not copying [received ZeroMQ frames](https://github.com/ai-dynamo/dynamo/pull/11574), and not paying full [inter-token-latency metrics overhead](https://github.com/ai-dynamo/dynamo/pull/11569) on every token. [The chat streaming hot path](https://github.com/ai-dynamo/dynamo/pull/10433) was shortened for the same reason. Individually, these are unremarkable; multiplied by every streamed token of every concurrent session, they are what determines how many requests per second a frontend can sustain.
+
+High-concurrency profiling then found costs that had nothing to do with moving data. [Static logging filters](https://github.com/ai-dynamo/dynamo/pull/11820) removed a shared span-matcher lock, a contention point rather than a volume problem, and raised reported frontend throughput from 932 to 1,133 requests per second. [Simpler positional radix buckets](https://github.com/ai-dynamo/dynamo/pull/12161) reduced peak memory in the mocker by 5.51 GiB in a 32-worker run. An open change [flushes detokenization metrics once per response](https://github.com/ai-dynamo/dynamo/pull/12999) rather than updating cumulative counters on every streamed chunk, approximately halving frontend CPU time in its matched diagnostic profile. That last one is the clearest example of the category: the instrumentation was cheap per call and ruinous at one call per token.
+
+## LMCache Agentic Optimizations
+
+LMCache is an open-source KV cache layer that sits under inference engines like vLLM, storing reusable KV chunks keyed by prefix hash across CPU DRAM, local NVMe, and remote backends (Mooncake, Redis, S3). LMCache can be used as an alternative to vLLM’s native offloading connectors.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/945da9bc-59ec-4084-9f3a-da58da04b3c5_2048x638.png"
+  caption="Source: SemiAnalysis"
+/>
+
+LMCache’s multiprocess path was changed for the volume and shape of agentic cache movement, beginning with a failure that is not a slowdown but a stop. When each of many requests with contexts above 100,000 tokens reserves the blocks for its whole load before starting, the pool is exhausted by requests that are all waiting and none progressing. [Chunked external-cache loading](https://github.com/LMCache/LMCache/pull/3382) reserves per chunk instead, so loads interleave and drain. At concurrency 32, the validation completed 120 requests where the old path deadlocked after 28, and concurrency 48 kept running with the KV pool 98.5 percent full.
+
+The other changes reduce how much is moved and how often the runtime gets in the way. Storing only the useful portions of DeepSeek-V4’s hybrid groups cut [storage per token by almost twenty times](https://github.com/LMCache/LMCache/pull/3635), and sliding-window prefetch now [loads only the live window](https://github.com/LMCache/LMCache/pull/3869) rather than window state that will never be read, the same reachability argument vLLM applied to offload, approached from the storage side. [One native transfer call per object group](https://github.com/LMCache/LMCache/pull/3908) then removes repeated Python lock handoffs across staging copies and kernel launches, which is overhead proportional to the number of pieces rather than to the bytes in them.
+
+Two current LMCache changes are especially specific to AgentX but remain open. [The hybrid lock-accounting fix](https://github.com/LMCache/LMCache/pull/4524) stops one request from releasing another request’s read locks on shared sliding-window or recurrent-state chunks. Several requests must share the same chunks, the accounting must be per-chunk rather than per-holder, and eviction must actually start. Sustained Kimi-K3 runs with DRAM offload supplied all three and produced tens of thousands of warnings, corrupt generations, and eventually GPU crashes once eviction began. Anything short of a long, shared, memory-pressured run leaves it dormant.
+
+A parallel line of LMCache work made all of the above reachable on AMD Instinct hardware. CacheBlend’s non-prefix reuse depended on flashinfer, which is CUDA-only, so [a Triton block-sparse attention backend](https://github.com/LMCache/LMCache/pull/3092) reimplements the three kernels it needs: block-sparse attention with CSR indices and log-sum-exp output, causal prefill, and log-sum-exp output blending. It then routes to them automatically when ROCm is detected or flashinfer is missing. [ROCm Dockerfiles](https://github.com/LMCache/LMCache/pull/3101) mirror the CUDA build and lightweight images. [An AMD hipFile backend](https://github.com/LMCache/LMCache/pull/3843) extends the GDS L1 slab-file tier, which reached storage only through NVIDIA cuFile, by binding ROCm’s hipFile through ctypes and dispatching on torch.version.hip; the cuFile path is unchanged.
+
+Distribution was the remaining gap. CUDA users installed a prebuilt wheel; AMD users built it from source. We worked together with AMD to publish a [prebuilt gfx942 and gfx950 wheel](https://github.com/LMCache/LMCache/pull/4273) which closes that. It installs into the upstream image and passes all 56 KV-transfer kernel tests on MI350X, and it publishes to a GitHub release rather than PyPI so a plain pip install lmcache stays the CUDA build. [A one-line follow-up](https://github.com/LMCache/LMCache/pull/4363) marks the bind-mounted repository as a git safe directory, which only fails in CI because the container runs as root over a runner-owned checkout and the version introspection in setup.py refuses to read it.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/89c2f79c-a704-4144-a900-ddd69ce306dd_1112x667.png"
+  caption="Source: GitHub"
+/>
+
+[DCP-aware CPU offload](https://github.com/LMCache/LMCache/pull/3561) resolves a straightforward incompatibility between two features that long contexts make mandatory together. With decode context parallelism enabled, each rank holds only a stride of the KV, so what any one rank could save is not a usable prefix; the fix gathers the strided shards before saving and redistributes them after loading. Without it, enabling context parallelism silently disables CPU cache hits for exactly the long prefixes that motivated both features. Its validation recorded more than 30,000 CPU hit events, with single-request loads reaching hundreds of thousands of tokens.
+
+## Mooncake Agentic Optimizations
+
+Mooncake serves Moonshot’s Kimi production traffic along with production traffic at many labs, and is a transfer engine underneath disaggregated vLLM and SGLang configurations. Until recently, Mooncake’s AMD support stopped short of both RDMA registration and offering installable packages.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/32c734aa-1f52-47af-a98e-334978b4e6fa_1778x1386.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Registering GPU memory for RDMA on Nvidia either uses the nvidia-peermem kernel module or exports a dmabuf file descriptor. AMD has no nvidia-peermem equivalent, so GPU-direct RDMA had no path at all and deployments fall back to staging KV through host DRAM. [A HIP dmabuf registration branch](https://github.com/kvcache-ai/Mooncake/pull/2225) adds the mirror of the existing CUDA dmabuf path, exporting through ROCm instead of the CUDA handle call, and resolving the true allocation base first because caching allocators pack tensors at an offset inside a larger allocation. Host memory still registers directly.
+
+Support that cannot be installed is not support. Mooncake published CUDA and MUSA wheels but no ROCm package, so AMD users built the engine from source inside every image. [A ROCm wheel, CI, and release path](https://github.com/kvcache-ai/Mooncake/pull/3184) publishes mooncake-transfer-engine-rocm to PyPI alongside them. This workstream from Andy Luo, AMD engineer, was due to noticing a pattern when dogfooding agentic workloads with AgentX that building MoonCake from source in ROCm is not an first class citizen pattern.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/12cd94ca-2aa7-4e06-a4bd-3153982234c5_1220x680.png"
+  caption="Source: GitHub"
+/>
+
+The transfer engine has no device kernels and does not depend on torch, so one architecture-agnostic wheel covers gfx942 and gfx950, and the ROCm runtime is bound at load time rather than vendored, which means the same wheel works unmodified in both the upstream vLLM ROCm image and the SGLang ROCm image. That was verified as a full cross product: MI300X and MI355X, each under vllm/vllm-openai-rocm and lmsysorg/sglang, running the master binary and a HIP buffer transfer test with data verification. The pull request adds a tag-triggered publish across Python 3.10 through 3.13. An open follow-up [adds a self-hosted two-node MI350X external prefill and decode tier](https://github.com/kvcache-ai/Mooncake/pull/3338) so the ROCm disaggregated path is exercised on real hardware rather than only compiled.
+
+Together, these PRs mean an AMD AgentX run can now install the transfer engine and the KV cache layer from published artifacts into stock upstream images, and move KV directly between GPU memory and the fabric.
+
+## Other Optimizations
+
+The changes above address long-context costs: a prefix that has to survive, a hybrid cache that has to stay correct, a transfer that has to keep up. But there are a whole host of day-zero enablement and correctness bugs that break requests just as badly as a million-token session.
+
+MiniMax-M3 tested whether that ROCm work compounds into day-zero readiness, and the [Advancing AI writeup](https://newsletter.semianalysis.com/p/can-amd-break-the-cuda-moat-amd-advancing) draws the comparison directly: AMD’s first public disaggregated recipe, MI355X FP4, reached InferenceX in January months behind Nvidia, while M3 FP4 disaggregation landed on day zero. This is an improvement from the DeepSeek-R1 period, when parity took months. Three vLLM fixes sat on that day-zero path, and each was a correctness failure rather than a performance one.
+
+Disaggregation was blocked first. NixlConnector’s handshake asserted that the SPLIT-region block_len scales with the prefill-to-decode TP ratio, but block_len follows per-rank KV heads. M3 has 4 KV heads, so a TP4 prefill paired with a TP8 decode is GQA-capped to one head per rank on both sides and the two lengths are equal where the assertion demanded a factor of two. The handshake was rejected, no KV moved, decode regenerated everything from scratch, and gsm8k scored 0. [Validating against the actual head ratio](https://github.com/vllm-project/vllm/pull/45879) fixed this.
+
+The other two were platform splits. M3’s sparse-attention backend read the byte-backed FP8 cache as float8_e4m3fn for every E4M3 configuration. But gfx942’s platform dtype is e4m3fnuz, and the two encodings differ. K and V were therefore altered before the kernels consumed them. The prefill and decode wrappers had also omitted the FNUZ types from their FP8 checks. [Using the platform dtype for the cache view](https://github.com/vllm-project/vllm/pull/45720) fixed both halves. Separately, M3 ships as separate NVIDIA and AMD model files, and only the NVIDIA one implemented the EAGLE3 interface, so speculative decoding aborted at engine init on ROCm with a model-does-not-support error. [Bringing the AMD model to parity](https://github.com/vllm-project/vllm/pull/45546) restored it, with MI355X gsm8k matching both the non-EAGLE3 MI355X run and B200.
+
+The TensorRT-LLM section above covers the M3 work that is long-context specific: descriptor explosion in disaggregated KV transfer, context graph capture, sparse block strides, autotuner candidates, and the corrupt split-K MoE tactics that had to be removed from the pool.
+
+The local AgentX matrix combines session-aware or KV-aware routing, long and variable conversation histories, MTP, hybrid attention, aggregate and disaggregated serving, and concurrency sweeps that cross the HBM capacity cliff. It includes GPU-resident comparisons and CPU DRAM offload through vLLM SimpleCPU, Mooncake, LMCache, and SGLang HiCache. That combination is what activates the upstream work above. The old fixed-sequence matrix usually creates one prompt, performs one prefill, decodes one fixed continuation, and discards the request. It therefore does not measure cache survival across turns, repeated tokenization, session affinity, cache-event traffic, offload churn, transfer progress during scheduler stalls, or long-lived ownership bookkeeping.
+
+The allowed optimization policy treats CPU KV offload as optional. A vendor may use vLLM connectors, LMCache, SGLang HiCache, Mooncake, Dynamo KVBM, or another CPU DRAM connector, or disable offload when the resulting latency and throughput point is better. NVMe offload is deferred. CPU DRAM must scale with the fraction of GPUs used, including the 3 TB cap for non-standardized-DRAM systems. Standardized-DRAM systems have no hard cap but retain the same proportionality rule. The local generator currently applies the 3 TB cap to every runner, so it does not yet implement the standardized-DRAM exception.
+
+The net new optimization surface is not simply longer attention. It is the preservation, movement, routing, reconstruction, and repeated processing of a growing session state. AgentX made those costs large enough to drive generic upstream changes across vLLM, SGLang, TensorRT-LLM, ATOM, AITER, Dynamo, and LMCache. Direct searches of NIXL and Mooncake did not identify additional AgentX-tagged runtime PRs, so their relevant effects remain represented through the engine connector changes above.
+
+## AgentX Methodology Deep Dive
+
+AgentX is a massive shift in open source real world long context multi turn agentic trace replaying and we collected traces over $3M worth of tokens within our own dataset consisting of real world traffic from Claude Code, OpenAI Codex, etc. In addition to the dataset, we developed a comprehensive methodology for replaying the traffic patterns fairly. The goal is to stay as authentic to the organic traffic as possible while being equitable about GPU resource requirements.
+
+We will deep dive into the agentic trace datasets, replay methodology, and agentic behavior in general. It is recommended reading for readers that would like to better understand the overall shape of agentic workloads as well as how harnesses orchestrate requests under the hood.
+
+## $3M USD Agentic Traffic Trace Collector
+
+When initially designing AgentX, our north star goal was to make the benchmark as realistic as possible in terms of KV workload shape and KV reuse patterns. We began experimenting with replaying some existing datasets, such as SWE-bench, Qwen-Bailian, and other random Claude Code traces from HuggingFace. At the time, these datasets did not include significant use of subagents, 1M context, compactions, dynamic workflows, or many other of the recent defining characteristics of an agentic trace. At SemiAnalysis, most of the team are AI power users and use agents for a broad variety of tasks including coding, analyst research, excel modeling, social media operations, and many more. Therefore, we decided that the most achievable and realistic traces could be captured in house.
+
+To collect a large number of traces, we created a proxy that intercepts HTTP requests to Claude / Codex. Then, users that wished to upload traces simply changed the base URL in their Claude / Codex setup to point to the proxy. At the time of writing, we have collected over 8,000 sessions, 3.4 million requests, and 610 billion tokens. Together these represent more than $3M USD in spend. [We open sourced a representative subset of these sessions for the AgentX v1.0 benchmark](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126).
+
+Although agentic harnesses can appear complex, they ultimately orchestrate a sequence of HTTP requests. Each request contains some combination of system instructions, tool definitions, and accumulated conversation history. As a session progresses, this history grows and is repeatedly sent back to the model, creating the long contexts and high prefix reuse that AgentX is designed to reproduce.
+
+Our proxy records these requests and responses as they occur. It also extracts metadata/HTTP headers such as timestamps, conversation IDs, and subagent IDs, which lets us recover the _structure_ of the conversation (request ordering, concurrent branches, and the approximate parent/child structure of each session). This metadata is what allows us to replay the traces approximately how they would have been seen by the original Anthropic API server.
+
+To protect employee privacy, the replay dataset contains no original prompts, source code, tool arguments, or tool results. Instead, we tokenize each requests’ content and then group it into 64-token blocks, finally replacing each block with a session-scoped chained hash. Matching prompt prefixes therefore produce matching hash prefixes without revealing their contents ([this paper](https://arxiv.org/abs/2506.02634) talks more about this strategy). During replay, these hash blocks can then be replaced with tokens from, say, a coding dataset. So we preserve the approximate context growth and conversation KV-reuse patterns of the original workload.
+
+It’s worth noting that this process is necessarily imperfect, mostly due to the fact that when using a frontier model provider’s API, much of the content that the end LLM server actually sees is hidden. For instance, thinking/reasoning content for SOTA models are now encrypted in HTTP requests and replaced with a deterministic hash in an attempt to hinder distillation attacks. However, it seems like [this didn’t work out exactly as well as the big labs had planned…](https://arxiv.org/html/2608.09867v1)
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/767f2f58-fa9d-40d3-8711-751d243cae9a_500x487.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Furthermore, while we have access to all the raw _content_ of the user/assistant messages, system prompts, tool usage, etc., API providers apply additional chat templating server side that aren’t transparent. We also cannot observe Anthropic’s proprietary tokenizer or context introduced by server-side tools. Images and documents also do not have a straightforward correspondence between their wire representation and the number of tokens processed by the model. We use deterministic placeholders and empirically calibrated, model-specific padding to bring reconstructed prompt lengths to best estimate the content that is actually seen by the server.
+
+We can’t _perfectly_ capture and replay Claude Code / Codex traces as they are _actually seen_ by the Anthropic / OpenAI servers due to incomplete information, but we can get pretty close. The chart below shows the ratio of hash tokens (after our approximations/processing) and the true API provider token count across all request lengths and models.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/7123b5de-9942-4298-8ccd-1b69b8a71df0_2048x1799.png"
+  caption="Source: SemiAnalysis"
+/>
+
+To summarize, collecting real Claude Code traces in the exact way that they would have been replayed against the original server is not easy due to incomplete information. However, we have enough context to collect and replay traces with extremely high fidelity to match original traffic patterns, timing, prefix caching, and DAG patterns.
+
+## AgentX $3Mil Dataset
+
+The dataset used for AgentX v1.0 can be found on [HuggingFace](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126). It is a 393 session subset of 8.3k session proxy corpus mentioned in the previous section. Additionally, we applied some post-processing to clean up anomalies, such as:
+
+- Removing Claude Code security monitor (auto mode) requests and title generation requests as these are specific to Claude Code and not necessarily representative of general agentic traffic
+- Removing requests with a reconstructed input length greater than 990k tokens (where our approximation overcounted)
+- Remove duplicate requests (sometimes the proxy received identical requests if the connection was dropped)
+
+Additionally, each conversation is formatted into the WEKA trace format, proposed by Callan Fox as part of his [kv-cache-tester](https://github.com/callanjfox/kv-cache-tester) project. We chose this format for storing trace information primarily because we worked closely with Callan to develop the benchmark and found it intuitive for storing per-session traces. All in all, the trace format is quite arbitrary and our proxy dataset could be mapped to other formats such as [Mooncake](https://docs.nvidia.com/aiperf/benchmark-modes/trace-replay-with-mooncake-traces).
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/48045991-20d7-4685-863d-c9aa1d6d3142_1812x2048.png"
+  caption="Source: SemiAnalysis"
+/>
+
+After these are applied, we get the following dataset. Note that not all X-axes are the same.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f22c9c7e-44ba-4ffd-bdc3-58fbbef8f30d_2048x2004.png"
+  caption="Source: SemiAnalysis"
+/>
+
+The distributions for ISL/OSL and inter-turn latency (in agentic work, this is mainly time taken for tool use) are relatively log-normal. The median ISL is 142k tokens and the median OSL is 444 tokens. The median inter-turn latency (or “tool use time”) was 3.84 seconds. Only ~10% of inter-turn latency was greater than 1 minute. These are likely made up of gaps where the harness is waiting on an actual response from a human.
+
+One thing worth mentioning is that these request distributions will look different depending on which harness is being used, since different amounts/types of context are injected (for instance Pi is known to be minimalist in terms of harness-injected context while Claude Code is known for the opposite. Additionally, the ISL/OSL distributions will depend on the model, as different models have tokenizers that can produce either more/fewer tokens. However, given a significant portion of the world’s agentic coding traffic goes through Claude Code, we believe this is rather representative.
+
+The dataset also has 175 sessions with at least one subagent (~44% of all sessions). There are 1,697 total subagent rollouts in the dataset, with a median of 4 per session. The median wall-clock time for a subagent (beginning of first request to end of last request) is 2.27 minutes. This distribution again follows a relatively log-normal distribution.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f4792809-087e-431b-acce-aecdd3855b3b_2048x1390.png"
+  caption="Source: SemiAnalysis"
+/>
+
+This dataset includes context up to 1M context, meant to test the more recent frontier open-weight models. Additionally, we have [a truncated 256k context length dataset](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k) which we replay against models with a max context length of 256k or less.
+
+## Agentic Traffic Trace Replayer
+
+Rather than build a replay solution from the ground up, we decided to partner with [AIPerf](https://github.com/ai-dynamo/aiperf), a vendor agnostic HTTP replayer tool from Nvidia that is adopted by many in the industry including tenstorrent, AWS, AMD, etc. While the intention is to integrate AgentX features into the upstream repo, we maintain a separate [fork](https://github.com/SemiAnalysisAI/aiperf) to be even more vendor neutral such that we have control over allowing even more 3rd more contributions. Again, thank you to the AIPerf team, especially Anthony Casagrande, for the help and dedication to building a realistic and representative agentic benchmark.
+
+An agentic session is naturally described as a directed acyclic graph (DAG). Each request is a node, and an edge means the request at its head cannot be issued until the one at its tail has completed. Every edge additionally carries a delay, specifying how long to wait once that precondition is met.
+
+The simplest session is completely linear, with no subagents and no parallel requests, each request depends on exactly one predecessor. The graph degenerates to a line and the only thing an edge encodes is the inter-turn latency (aka, tool use time or “think” time), which is the client’s local work rather than the model’s.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/20662be0-c311-49e9-8e39-0973d8a1865d_2048x1169.png"
+  caption="Source: SemiAnalysis"
+/>
+
+In agentic traces, subagents can also be spawned. A subagent is a separate stream of requests that has its own context, typically to do a focused task. Multiple subagents can run in parallel to do more aggregate work, and subagents can run in parallel to the main agent in some cases. The main agent then waits on groups of subagents to finish, and then incorporates their outputs back into the main agent’s context (while this is not _always_ the case, this is the most common pattern).
+
+This behavior is responsible for turning the linear chain of requests above into a DAG, where certain requests are dependent on others. When a group of requests belonging to one subagent is identified, AIPerf finds the most recent main agent predecessor and designates it as the “spawning” request. Similarly, the “join” request is identified by the subsequent main agent request after the duration of the subagent group completes.
+
+In the example below, the subagent group consists of a single subagent (001), which runs two requests. Its first request goes out as soon as the main agent’s opening request completes. When that request completes, a 2.2-second inter-turn delay stands in for tool-use wall-clock time, and then the subagent’s second request is sent. The main agent’s second request is the join point for subagent 001. It goes out once both conditions are met: at least 17 seconds have elapsed since the main agent’s first response _and_ subagent 001’s second request has completed.
+
+One small limitation is that HTTP timestamps reveal timing, but not always causality. In the example below, if subagent 001 finishes with seven seconds remaining before main-agent request 2’s recorded start, we cannot tell whether those seven seconds represent work performed after the subagent returned or independent work already underway. AIPerf therefore preserves both constraints: request 2 waits for its recorded main-path delay and for subagent 001 to finish. This reproduces the observed timing and workload topology, but not any dependencies hidden inside the harness. These are things we hope to improve on in subsequent versions of AgentX.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/18adec21-15d9-40cd-af7d-0772da4a9af0_2048x1729.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Multiple subagents can also be spawned from a single request. In the example below, subagents 001 and 002 both identify their spawning parent as main agent request 1, and then join at main agent request 2.
+
+AIPerf can also identify “auxiliary” requests, which are one off requests that do not share context with any other requests in the stream. These branch off of the main agent and never join back. In practical terms, these are requests like Claude Code’s “summarize this session” requests that are unrelated to the conversations context. Another good example is Claude Code’s “/btw” feature.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/3d5f01ad-1736-496b-91b4-31672002bf47_1920x2048.png"
+  caption="Source: SemiAnalysis"
+/>
+
+The example below brings it all together. This is the type of trace snippet you’d see in the actual dataset. We have five parallel streams leaving a single main agent request, separated into groups defined by what main agent request they join upon.
+
+Subagents 001 and 002 finish at 20 and 23 seconds, so the first main agent request starting after that, at 25 seconds, is their join. Subagents 003 and 004 are spawned in the same gap but run much longer, finishing at 46 and 50 seconds, so they join at 52 seconds instead. AIPerf keys each subagent by the pair (spawning request, join request), which means these four streams collapse into two branches even though all four leave the same node. The branch takes the name of its first member, which is why the join edges are labelled with subagent 001 and subagent 003.
+
+This is also the first case where the main agent overlaps its own subagents. The request at 25 seconds goes out while 003 and 004 are still running: the main agent is blocked only on the group that joins it, not on every subagent in flight.
+
+The auxiliary chain attaches to whichever main agent request most recently preceded it, which here is the request at 52 seconds rather than the one that opened the session. That node therefore does two things at once — it receives the second subagent group’s join, and it spawns the one-off. Of course the auxiliary request never joins back.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/d2f465f9-372c-4ee1-b42f-e79fa9fe352b_1677x2048.png"
+  caption="Source: SemiAnalysis"
+/>
+
+## Further AgentX Deep Dive
+
+### Pareto Frontiers Sweeping
+
+In the AgentX workload, in order to generate a Pareto frontier, we sweep over the number of concurrent Claude Code sessions against a single deployment. Since each conversation has realistic inter-turn delays and subagent usage, we get a spikier, more realistic traffic pattern. The example below is an example of replaying 40 concurrent clients against a B200 TP4 vLLM server running MiniMax M3.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/058757ed-531f-4cfb-a1ad-a06b1499ecbd_1360x612.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference/agentic/439239)"
+/>
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/43184e50-b71a-4a48-a569-98d6906cb259_2048x1167.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-13&g_runid=31723588108&i_seq=agentic-traces&i_xmode=interactivity&g_model=MiniMax-M3&i_best=0&i_active=b200_vllm)"
+/>
+
+A final point worth discussing is what metrics are important to consider when evaluating agentic workloads. We believe interactivity (TPS - tokens per second) and time-to-first-token (TTFT) are still important, and these are the industry standard for evaluating SLOs. When viewing AgentX results, it is _extremely important_ to consider both TPS and TTFT together, since there are inference optimizations that can improve one at the cost of the other.
+
+We are currently working on defining a new metric that combines TPS and TTFT in a meaningful way. This should also take into account that in agentic workloads, people often care more about the end-to-end speed at which a _task_ finishes rather than how fast they receive tokens or TTFT.
+
+Finally, it’s worth mentioning that the end-to-end latency, in its current form, is now less meaningful because end-to-end latency is directly proportional to OSL. Therefore, P90 E2E latency is heavily affected by the 10% tail of longest output sequence lengths. While it can still be good to holistically compare the overall performance of certain configurations, we recommend instead looking at a combination of TPS and TTFT.
+
+### Warmup, Timing, Determinism, and Conversation Reuse
+
+The goal of AgentX is to benchmark systems that are already in a steady state. With agentic workloads, this means that profiling should start from a point where some context trajectories are already cached. To mimic a steady state, it is also desirable that not all conversations start at turn 0, which may cause a “[thundering herd](https://en.wikipedia.org/wiki/Thundering_herd_problem)” effect.
+
+Warmup proceeds in two stages. First, AIPerf uses a fixed random seed to select a wall-clock point between 25% and 75% of each conversation. At that point, it identifies every active request stream, including the main agent and any active subagents, and sends the most recent request before the selected point for each stream. These primer requests reconstruct the conversation state at that point and are dispatched together. AIPerf waits for them to drain before continuing.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/963dce62-c49b-4deb-afee-e8dd448526c5_2048x1322.png"
+  caption="Source: SemiAnalysis"
+/>
+
+In the second phase, each replay lane is advanced by 10 additional requests to give additional opportunity for the KV cache to materialize. All warmup requests omit inter-turn delay and use a maximum output length of one token, substantially reducing warmup time.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/85116f5f-be0d-405b-a5d9-6a9237b44cc3_2048x895.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/fef10e0d-261e-4978-8650-b10304ec04d0_2048x897.png"
+  caption="Source: SemiAnalysis"
+/>
+
+When warmup is complete, profiling begins and lasts one hour. All metrics are collected strictly over this duration. For reproducibility, AIPerf accepts a seed that ensures each run samples conversations deterministically, conversations start at the same point, and that each conversation is reconstructed with the same synthetic content run-to-run. During profiling, we impose a 5 minute idle time cap on each stream, so that long inter turn gaps don’t “acquire” a worker lane for the duration of the benchmark. We enforce this so that we can effectively run the benchmark in 1 hour. In later versions of AgentX where we include NVMe offloading, we may choose to increase this so that we can measure a longer TTL. For now, 5 minutes is reasonable as this is [Anthropic’s default KV cache TTL](<https://platform.claude.com/docs/en/build-with-claude/prompt-caching#:~:text=Check%20that%20calls%20are%20made%20within%20the%20cache%20lifetime%20(5%20minutes%20by%20default)>).
+
+This level of determinism ensures that runs using the same inference engine, hardware, concurrency, and server settings are reproducible. However, because AgentX is a [closed-loop benchmark](https://notpeerreviewed.com/blog/tail-latency/), different configurations will complete different numbers of requests at different rates, introducing some natural variation in the workloads they encounter. This is most noticeable at lower concurrency, where fewer requests are completed (naturally) and the workload has less opportunity to converge toward the dataset’s overall distribution.
+
+When a conversation completes during profiling, its replay lane selects another conversation from the dataset sampler. Each replay receives a unique, deterministic cache-bust marker that is prepended to every independent prefix chain, including the main-agent chain and any fresh-context subagent or one-off chains. Forked subagents inherit the marker from their parents. The marker stays the same within a replay, preserving its KV-reuse patterns, but changes between replays to prevent artificially high cache-hit rates. This also enables scenarios to run where concurrency is greater than the number of conversations in the dataset (393).
+
+### AgentX’s Fair Speculative Decoding Methodology
+
+As mentioned, the dataset is anonymized upon collection. This means that the 64-token hash blocks must be synthetically filled in before replay. AIPerf accomplishes this by deterministically sampling from a synthetic coding/tool-use token pool.
+
+Importantly, the KV reuse patterns as well as request timing are maintained, however the synthetic request data does lead to some additional considerations. Namely, running speculative decoding methods on synthetic data may lead to the speculator rejecting/accepting an abnormal number of tokens when compared to non-synthetic data (since the speculator is not trained on synthetic data).
+
+We talked about this shortcoming in our [InferenceX v2 article](https://newsletter.semianalysis.com/i/188090866/multi-token-prediction-mtp), and have since then improved on our methodology. We have worked closely with the community to ensure a mechanism exists in most OSS inference engines that allows users to force how many draft tokens to accept from the speculator (aka, “acceptance length” or “acceptance rate”). Then, for each (model, speculator, draft length, and thinking mode) combination, [we collect the average AL](https://github.com/SemiAnalysisAI/InferenceX/tree/main/golden_al_distribution) on the [SPEED-Bench](https://huggingface.co/datasets/nvidia/SPEED-Bench) agentic coding dataset, a “unified benchmark designed to evaluate speculative decoding (SD) across diverse semantic domains and realistic serving regimes.”
+
+Then, at runtime we apply these realistic Speculative decoding acceptance lengths to AgentX to ensure vendor neutral fairness.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c15f47a8-fcb3-4220-80ee-4b95bfa7b0b0_2048x1376.png"
+  caption="Source: GitHub"
+/>
+
+## Exploring Agentic Workloads Detailed Telemetry Tutorial
+
+AgentX required more than a new benchmark harness and dataset. [We also spent some time rebuilding parts of the InferenceX visualization to make agentic results easier to explore and digest](https://inferencex.semianalysis.com/inference/agentic/439903). A single AgentX datapoint represents thousands of requests across growing conversations, subagents, warmup periods, cache states, and dynamically changing in-flight load. Due to this, having a single point on a Pareto curve can hide a lot of useful information. As we have said many times - there is never just a one size fits all solution for inference serving.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/46159be7-8b5b-49d7-990f-2ce1cf1217d0_1110x687.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference/agentic/439903)"
+/>
+
+One of our major changes is how we construct the curves themselves. In previous versions of InferenceX, configurations with speculative decoding enabled and disabled were often displayed as separate curves. However, we are now moving away from this approach. The frontend now combines allowed inference optimizations and displays the best available curve for each model, SKU, and inference engine combination. Due to this, individual points along a single curve may use different optimization techniques and configurations, including speculative decoding, disaggregation, or KV cache offload.
+
+Our goal is to show the best production performance available from each hardware and software stack, rather than creating a separate curve for every possible combination of optimizations. However, we still expose the underlying configuration and provenance for every point. Clicking a point shows a tooltip with a detailed view showing exactly which configuration produced it, along with the run metadata, links to the publicly viewable CI provenance, and AgentX specific statistics. From there, the “View charts” link opens the full point-detail page with AgentX specific statistics.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/add2d50c-a0be-4771-a00b-e6b4f1fe8502_2048x1256.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+The detailed point view provides a much deeper look into the selected AgentX run. It includes input and output sequence length distributions, interactivity and TTFT over time, KV cache utilization, request queue depth, prefix cache hit rate, input and decode throughput, prompt-token source breakdown, and unique input tokens over time. These metrics make it easier to understand why two points with similar aggregate throughput may behave differently throughout the replay.
+
+The page also separates warmup and profiling data. Readers can switch between the two phases to inspect how the system behaves while its cache state is being established and during the profiling period used for the benchmark run.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/91abef64-d639-4da8-8d7e-3499be750011_1001x1541.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+Points using KV cache offload are surrounded by an additional dotted circle on the main chart, which is used to distinguish points with KV offload enabled. When one of these points is selected, the detail page shows the offload type, KV offload engine, chip cache-hit rate, and CPU cache-hit rate. This makes it possible to see where KV offload contributes to the best curve without creating a separate curve for every offload configuration.
+
+Another new feature is the request timeline. This view shows the individual requests replayed during a selected AgentX run and can be organized either by conversation or by worker. The conversation view groups subagents underneath their corresponding root conversation, making it easy to see when conversations and subagents overlap. Warmup and profiling requests can also still be viewed separately.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c8fc6236-52bd-45ff-ac2e-539899c00d87_2048x1097.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+Each request in the timeline is clickable and links directly to the corresponding conversation and turns on the InferenceX datasets page. This allows readers to move from an aggregate point on the Pareto curve to the exact anonymized request that was replayed.
+
+[The AgentX page](https://inferencex.semianalysis.com/datasets) also includes a flamegraph for visualizing the structure of an individual conversation. Each bar represents one turn and is scaled relative to the largest turn in that conversation. The bar is divided into cached prefix tokens, uncached input tokens, and generated output tokens. This gives a visual representation of how the context grows throughout a conversation and how much of each request can be reused from KV cache.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a0e991d9-6ab6-43e7-9cc9-93ac9ca5b7bd_2048x879.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+## InferenceX/AgentX Future Next Steps
+
+We are excited to continue our goal of making AgentX the most realistic and representative benchmark. For the foreseeable future, we will continue making [small bug fixes to the current v1.0.x harness](https://github.com/SemiAnalysisAI/aiperf/releases), but have plans to make bigger changes to the v1.1 harness. Submissions for each distinct model will always run the same minor version to ensure all results are comparable.
+
+As a fast follow, we will add SSD/NVMe KV offloading. This will allow an even larger KV cache working set size than DRAM allows, which will allow the high throughput left side of the pareto curve.
+
+We will also capture a larger, more diverse, and more recent dataset of agentic traces across a wider variety of models and harnesses. Rather than representing each request as one contiguous list of hash IDs, the next dataset will preserve the boundaries between system instructions, user and assistant messages, tool calls, and tool results. This will allow AgentX to evaluate workload-aware serving techniques that use information available to the agent harness but normally hidden from the inference engine. For example, a router could direct low-reuse tool traffic to dedicated prefill workers, retain an agent’s prefix during a long tool call, or prefetch and share prefixes when subagents fork. The current format preserves request sizes and KV-reuse patterns, but it lacks the structure needed to evaluate these optimizations.
+
+The following [SGLang RFC](https://github.com/sgl-project/sglang/issues/27574) by Ishan provides a concrete example of why this richer trace structure matters. It proposes a router-initiated hint interface that uses information such as session lifecycles, shared-prefix boundaries, tool-call duration, and subagent state to tell the engine when KV should be shared, prefetched, demoted, pinned, or retained. Capturing this structure would allow future AgentX versions to evaluate these workload-aware cache policies instead of only replaying flat token prefixes.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/36899201-da3b-42ec-888f-2443c63980aa_2048x1260.png"
+  caption="Source: [GitHub](https://github.com/sgl-project/sglang/issues/27574)"
+/>
+
+Lastly, we are capturing fine-grained and coarse grain power telemetry data to have an even more accurate view of the efficiency of the Joules per intelligence of different software and hardware stacks.
+
+We have so much data and so many possible visualizations. Please let us know any visualizations you would like to see, as well as any more general feature requests!
+
+[Suggest Frontend Feature!](https://github.com/SemiAnalysisAI/InferenceX-app/issues/new/choose)
+
+## Performance Throughout the Model Lifecycle
+
+In the following sections, we turn to our historical single turn data (8k1k), which covers each model from launch day through the point we retired it from active testing. There are strong results here from both Nvidia and AMD, including several fixed sequence length configurations where MI355X comes out ahead.
+
+AgentX better represents today’s agentic inference workloads. However, the historical fixed-sequence InferenceX results still remain useful for tracking the performance over time. Workloads such as 8k1k and 1k1k strip away most session-level behavior, including prefix reuse, persistent KV cache state, and routing affinity. This makes them less representative of current production traffic, but still useful for tracking how inference performance improves as software support matures.
+
+<Blur>
+
+Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+
+</Blur>
+
+---
+
+_This article continues on our Substack. [Subscribe to SemiAnalysis](https://newsletter.semianalysis.com/subscribe) to read the complete article._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "What is AgentX 1.0?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "AgentX 1.0 is the world's first fully open source, multi-turn agentic coding inference benchmark at 1 million context, released under Apache 2.0. It replays real anonymized Claude Code and Codex traces instead of fixed sequence length prompts, and runs on roughly 2MW of continuously operated compute across more than 1000 chips including MI355X, GB300 NVL72, GB200 NVL72, B300, B200, MI325X, MI300X, H200, and RTX Pro servers."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What makes an agentic workload different from a fixed sequence length benchmark?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Agentic workloads are multi-turn, long context, have very high prefix reuse, and launch bursty short-lived subagents. That makes agentic inference a systems problem rather than a kernel problem: KV tensors must be transferred across nodes, requests must be routed to the rank holding the right prefix, and long contexts force KV offload to DRAM or SSD. Fixed sequence length workloads such as 8k1k strip all of that away and mostly measure baseline chip and kernel performance."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Does the CUDA moat hold up on agentic workloads?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "It depends on the model. On DeepSeek V4 Pro the race is close: MI355X SGLang matched B200 vLLM on performance per dollar until August 21, 2026, when vLLM optimizations from Inferact and NVIDIA pushed B200 ahead, although B300 vLLM and B200 SGLang were already beating MI355X. On MiniMax M3, Qwen3.5, and GLM 5.3, NVIDIA leads by a wide margin, with over 20x better performance at 90 tok/s/user on Qwen3.5 SGLang and up to 5x better cost efficiency on GLM 5.3 at 150 tok/s/user. Context parallelism is a concrete part of the moat: every AMD backend is listed as unsupported in the vLLM attention backend support matrix."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What is the AgentX dataset made of?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "SemiAnalysis collected traces through a proxy that intercepts HTTP requests to Claude and Codex, gathering over 8,000 sessions, 3.4 million requests, and 610 billion tokens, representing more than 3 million USD of spend. The open sourced AgentX v1.0 subset contains 393 sessions with a median input sequence length of 142k tokens, a median output sequence length of 444 tokens, and a median inter-turn latency of 3.84 seconds. 175 sessions, roughly 44 percent, contain at least one subagent, with 1,697 subagent rollouts in total. Content is anonymized into 64-token blocks replaced by session-scoped chained hashes, so prefix reuse patterns survive but prompts do not."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What is E2E Normalized Interactivity?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "E2E Normalized Interactivity is an experimental metric defined as output sequence length divided by end-to-end latency. Because end-to-end latency equals TTFT plus output length times TPOT, the metric works out to interactivity plus a penalty proportional to TTFT, capturing both responsiveness dimensions in one number. It is not perfect: it heavily penalizes high TTFT and does not capture every nuance of optimizations such as prefill-decode disaggregation, so AgentX v1.0 submissions still optimize for interactivity and TTFT separately."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What industry impact has AgentX had so far?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Over 70 upstream pull requests across vLLM, SGLang, TensorRT-LLM, ATOM, AITER, Dynamo, LMCache, and Mooncake used AgentX as the north star benchmark. Examples include hybrid-attention prefix caching in vLLM that reported over 95 percent prefix cache hit rate at 1M context, CPU KV offload extended to DeepSeek V4 for 81.7 percent higher output throughput and 46.6 percent lower mean end-to-end latency, boundary-aware incremental tokenization in TensorRT-LLM that cut mean processing time from 185.1 ms to 11.3 ms on Qwen3.5, and sparse checkpoint retention in AMD ATOM that lifted the DeepSeek V4 prefix hit rate from 5.6 percent to 96.45 percent."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/brief-overview-of-agentic-workloads.mdx b/packages/app/content/blog/brief-overview-of-agentic-workloads.mdx
new file mode 100644
index 000000000..5555973b7
--- /dev/null
+++ b/packages/app/content/blog/brief-overview-of-agentic-workloads.mdx
@@ -0,0 +1,136 @@
+---
+title: 'A Brief Overview of Agentic Workloads'
+subtitle: 'Multi-turn sessions, long contexts, and near-total prefix reuse make agentic inference a systems problem, and change what a benchmark has to measure'
+seoDescription: 'What defines an agentic workload: multi-turn sessions, long context, high prefix reuse, and bursty subagent fan-out, and why benchmarking them is a systems problem.'
+date: '2026-08-19'
+publishDate: '2026-08-19'
+tags:
+  - agentic
+  - agents
+  - agentx
+  - benchmark
+  - inference
+---
+
+At a high level, an [agentic workload](/glossary/agentic-coding-workload) is characterized by four things:
+
+- **Multi-turn.** A session includes tens or hundreds of user and assistant interactions, compared to a handful in a chatbot scenario.
+- **Long context.** Between system prompts, tool definitions, and the sheer number of turns, context accumulates quickly.
+- **High [prefix reuse](/glossary/prefix-caching).** The conversation progresses linearly, with the output of turn _n-1_ concatenated into the input of turn _n_, so most context can be served from [KV cache](/glossary/kv-cache) rather than recomputed. How much depends on the storage available to hold KV tensors. As _n_ grows, the ratio of cached to uncached input tends toward 1.
+- **Bursty [subagent](/glossary/subagent) fan-out.** A turn can launch several short-lived subagents at once, which creates spiky KV cache allocation and eviction patterns.
+
+<Figure
+  src="/images/brief-overview-of-agentic-workloads/context-accumulation.png"
+  alt="Diagram of four agentic turns showing context accumulating downward: each turn's input contains the previous turn's full input plus its output, with the shared portion highlighted as cached from the previous turn"
+  caption="Context accumulation across four turns of an agentic session. Everything highlighted is served from cache; only the newest output and tool results are uncached."
+/>
+
+## Agentic inference is a systems problem
+
+Considering the characteristics above, benchmarking these workloads is fundamentally different from the existing fixed-sequence-length benchmarks.
+
+Because prefix reuse is so high, KV tensors have to move efficiently across nodes and ranks — the job of transfer layers like NIXL, MORI-IO, and [Mooncake](/agentx/optimizations/mooncake). Conversations then have to be routed to the node or rank where their prefix already resides in order to maximize cache hit rate, which is what llm-d, [Dynamo](/agentx/optimizations/dynamo), and the vLLM and SGLang routers do. Long-context conversations stress HBM capacity for KV cache and force offload to slower memory tiers such as DRAM and SSD, which has to be done efficiently — Mooncake Store, [LMCache](/agentx/optimizations/lmcache), vLLM's CPU offloading, and SGLang HiCache all attack this layer.
+
+This is in contrast to fixed-sequence-length, single-turn workloads, where prefix reuse is not relevant and inference performance is largely reflective of baseline chip and kernel performance. That does not make the plethora of fixed-sequence-length data on InferenceX less important. Stripping away the complexities of agentic serving gives a clear picture of how low-level inference optimizations are progressing, and provides the baseline that AgentX results are read against.
+
+## Building a realistic replay corpus
+
+To make the AgentX workloads as realistic as possible, we collected an initial corpus of 393 internal SemiAnalysis Claude Code traces to replay. To anonymize the content while keeping the original prefix-reuse pattern intact, we use a method similar to [Qwen-Bailian](https://github.com/alibaba-edu/qwen-bailian-usagetraces-anon), one of the earliest corpora of production traces. We then use [AIPerf](https://github.com/ai-dynamo/aiperf) to reconstruct the traces according to the original schedule of requests, at varying levels of concurrent clients.
+
+Reconstructing subagent structure required changes in Claude Code itself. We worked with Anthropic to ship two features that make the AgentX dataset possible, both adding agent-instance identity to outgoing requests so that parallel branches can be told apart at the proxy:
+
+- [anthropics/claude-code #49207](https://github.com/anthropics/claude-code/issues/49207) — add an agent instance ID (`x-claude-code-agent-id`) and its parent ID (`x-claude-code-parent-agent-id`) to API request headers, so interleaved subagent requests are attributable instead of collapsing into one session-level stream.
+- [anthropics/claude-code #66761](https://github.com/anthropics/claude-code/issues/66761) — extend those same headers to subagents spawned by the `Workflow` tool's fan-out, which previously carried only the parent conversation's session ID.
+
+Without those headers, a session that fans out N concurrent subagents is indistinguishable from N unrelated conversations, and the spawn and join dependencies that shape the [replay graph](/glossary/trace-replay) cannot be recovered.
+
+The full dataset rules, distributions, and source links are published in the [AgentX methodology](/agentx/methodology), and the concrete engine changes this workload has already exposed are tracked in the [AgentX optimization tracker](/agentx/optimizations).
+
+## Acknowledgments
+
+Thanks to the Anthropic staff who shipped the agent-identification headers in Claude Code, and to the Alibaba and NVIDIA teams behind Qwen-Bailian and AIPerf, whose work the AgentX replay pipeline builds on.
+
+<DashboardCTA href="/inference?i_seq=agentic-traces">
+  Compare live AgentX inference results →
+</DashboardCTA>
+
+## Frequently asked questions
+
+### What is an agentic workload?
+
+An agentic workload is inference traffic generated by AI agents rather than by a person typing into a chat box. It is characterized by multi-turn sessions of tens to hundreds of interactions, long contexts built from system prompts and tool definitions, very high prefix reuse across turns, and bursts of short-lived subagents.
+
+### Why does prefix reuse matter so much for agentic inference?
+
+Each turn's input contains the previous turn's input plus its output, so nearly all of the context has already been processed. If the KV tensors for that prefix are still resident, the prefill work is skipped. As a session grows, the ratio of cached to uncached input tends toward 1, so cache capacity and cache hit rate dominate the cost of serving.
+
+### How is benchmarking an agentic workload different from a fixed-sequence-length benchmark?
+
+A fixed-sequence-length benchmark sends independent requests with a set input and output length, so prefix reuse is not a factor and the result mostly reflects chip and kernel performance. An agentic benchmark replays dependent sessions where context grows, prefixes are shared, and subagents fan out, which makes KV transfer, prefix-aware routing, and offload tiering part of what is being measured.
+
+### What does bursty subagent fan-out do to the KV cache?
+
+A single turn can spawn several subagents at once. Each one allocates its own KV cache, runs briefly, and finishes, so cache pressure arrives in spikes rather than a steady state. Schedulers and eviction policies tuned for uniform request streams behave differently under that pattern.
+
+### How was the AgentX replay corpus built?
+
+We collected 393 internal SemiAnalysis Claude Code traces and anonymized them with a block-hash method similar to Qwen-Bailian, preserving prefix relationships without preserving content. AIPerf then reconstructs the traces on their original request schedule across varying levels of concurrent clients.
+
+### Do fixed-sequence-length benchmarks still matter?
+
+Yes. Fixed-sequence-length results isolate low-level chip and kernel performance without the complexity of agentic serving, which makes them the baseline that AgentX results are interpreted against.
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "What is an agentic workload?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "An agentic workload is inference traffic generated by AI agents rather than by a person typing into a chat box. It is characterized by multi-turn sessions of tens to hundreds of interactions, long contexts built from system prompts and tool definitions, very high prefix reuse across turns, and bursts of short-lived subagents."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does prefix reuse matter so much for agentic inference?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Each turn's input contains the previous turn's input plus its output, so nearly all of the context has already been processed. If the KV tensors for that prefix are still resident, the prefill work is skipped. As a session grows, the ratio of cached to uncached input tends toward 1, so cache capacity and cache hit rate dominate the cost of serving."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "How is benchmarking an agentic workload different from a fixed-sequence-length benchmark?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "A fixed-sequence-length benchmark sends independent requests with a set input and output length, so prefix reuse is not a factor and the result mostly reflects chip and kernel performance. An agentic benchmark replays dependent sessions where context grows, prefixes are shared, and subagents fan out, which makes KV transfer, prefix-aware routing, and offload tiering part of what is being measured."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What does bursty subagent fan-out do to the KV cache?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "A single turn can spawn several subagents at once. Each one allocates its own KV cache, runs briefly, and finishes, so cache pressure arrives in spikes rather than a steady state. Schedulers and eviction policies tuned for uniform request streams behave differently under that pattern."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "How was the AgentX replay corpus built?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "We collected 393 internal SemiAnalysis Claude Code traces and anonymized them with a block-hash method similar to Qwen-Bailian, preserving prefix relationships without preserving content. AIPerf then reconstructs the traces on their original request schedule across varying levels of concurrent clients."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Do fixed-sequence-length benchmarks still matter?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Yes. Fixed-sequence-length results isolate low-level chip and kernel performance without the complexity of agentic serving, which makes them the baseline that AgentX results are interpreted against."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set.mdx b/packages/app/content/blog/deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set.mdx
new file mode 100644
index 000000000..4f1834ae7
--- /dev/null
+++ b/packages/app/content/blog/deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set.mdx
@@ -0,0 +1,86 @@
+---
+title: 'DeepSeek V4 Pro on AgentX: B200 vs B300 and the KV Cache Working Set'
+subtitle: '50% more HBM squeezes out extra throughput, and the per-point telemetry shows exactly where it comes from'
+seoDescription: 'B200 against B300 on DeepSeek V4 Pro AgentX runs. TCO-normalized throughput is close, but the telemetry shows a 91% HBM cache hit rate on B300 against 73% on B200, and a KV working set of 43M tokens against 22M.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - deepseek
+  - b200
+  - b300
+  - h200
+  - nvidia
+---
+
+When normalized by [total cost of ownership](/glossary/total-cost-of-ownership), B300 vLLM and B200 vLLM aggregated performance on DeepSeek V4 Pro is quite similar. The main difference is that B300 can squeeze out extra throughput, given its 50% increase in HBM capacity over B200.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/939b9197-321d-4869-ba13-f0b800343b7b_2048x1300.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=b200_vllm%2Cb300_vllm)"
+/>
+
+The interesting part is not the summary number, it is what the AgentX server telemetry shows underneath it.
+
+## 43 million tokens against 22 million
+
+Under the load of 384 concurrent agentic traces, B300 vLLM DEP8 with 3TB of DRAM using vLLM simple offloading achieved a 91% HBM cache hit rate, with an additional 1.36% DRAM cache hit rate. The reason is that the HBM [KV cache](/glossary/kv-cache) working set size is approximately 43 million tokens with this configuration, and the load barely exceeds that number of tokens in flight at any given time.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f86888c4-c1a7-47b0-a580-d1a74876cb9c_2048x898.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference/agentic/439522)"
+/>
+
+With B200 at concurrency 196, all other parameters unchanged, the picture inverts. The HBM cache hit rate falls to 73% and the deployment leans much harder on DRAM, with an [offload](/glossary/kv-cache-offload) cache hit rate of nearly 20%. The HBM KV cache working set size here is 22 million tokens, roughly half that of B300.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/70a7a7c7-7918-459c-a4e0-f7fc1edca816_2048x887.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference/agentic/439967)"
+/>
+
+## Why the offload tier needs headroom
+
+DRAM KV offloading is typically implemented as a write-through cache, meaning every prefix written to the HBM cache is also written to the DRAM cache. It is therefore most effective when the amount of DRAM available for offloading is significantly bigger than HBM KV cache capacity, by a multiple of roughly 1.5 to 3.
+
+That property is what makes the extra HBM on B300 compound rather than merely add. A larger resident working set keeps more of the session in the fast tier, which leaves the slower tier serving a genuine remainder rather than absorbing a constant stream of evictions.
+
+## The H200 counterexample
+
+H200 SGLang FP8 is able to serve DeepSeek V4 at low concurrency, and is even competitive with B200 and MI355X SGLang from a performance per dollar standpoint. It cannot compete with the newer SKUs in high throughput scenarios, because it lacks the HBM to do so.
+
+The failure mode is instructive: the reliance on DRAM KV offloading at higher concurrencies leads to unreasonable latency as the number of users scales. Offload buys capacity, but it is not free, and a deployment that depends on it to cover a structural memory shortfall pays for that at the tail.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/91dfd9ca-d13d-43e9-a57e-3be0fcc4a5cf_2048x1315.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=h200_dynamo-sglang&i_optimal=0)"
+/>
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "How much better is B300 than B200 on DeepSeek V4 Pro?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "When normalized by total cost of ownership, aggregated B300 vLLM and B200 vLLM performance is quite similar. B300 squeezes out extra throughput thanks to a 50% increase in HBM capacity, which shows up in the telemetry as a 43 million token HBM KV cache working set at concurrency 384 with a 91% cache hit rate, against 22 million tokens and 73% on B200 at concurrency 196."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does H200 fall behind at high concurrency?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "H200 SGLang FP8 serves DeepSeek V4 well at low concurrency and is competitive with B200 and MI355X SGLang on performance per dollar there. It lacks the HBM to compete in high throughput scenarios, so it leans on DRAM KV offloading as concurrency rises, and that reliance leads to unreasonable latency as the number of users scales."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/deepseek-v4-pro-agentx-gb200-vs-gb300-disagg.mdx b/packages/app/content/blog/deepseek-v4-pro-agentx-gb200-vs-gb300-disagg.mdx
new file mode 100644
index 000000000..bdff8ae75
--- /dev/null
+++ b/packages/app/content/blog/deepseek-v4-pro-agentx-gb200-vs-gb300-disagg.mdx
@@ -0,0 +1,72 @@
+---
+title: 'DeepSeek V4 Pro on AgentX: GB200 vs GB300 Rack-Scale Disaggregation'
+subtitle: 'Both lean on PD disagg, GB300 adds DEP32 wide-EP decode, and the gap shows up in first-token latency rather than token rate'
+seoDescription: 'NVIDIA rack-scale results for DeepSeek V4 Pro on AgentX. GB300 Dynamo TRTLLM and GB200 Dynamo vLLM both use prefill-decode disaggregation, and the GB300 point trails on TTFT because higher concurrency brings more subagent traffic and more cold prefills.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - deepseek
+  - gb200
+  - gb300
+  - nvl72
+  - disagg
+  - wide-ep
+  - nvidia
+---
+
+On DeepSeek V4 Pro, NVIDIA's most competitive solutions are GB300 Dynamo TRTLLM and GB200 Dynamo vLLM. Both configurations rely on prefill-decode [disaggregation](/glossary/disaggregated-inference) to achieve high throughput at reasonable [interactivity](/glossary/interactivity). GB300 configurations additionally employ [wide expert parallelism](/glossary/wide-expert-parallelism) decode instances at DEP32 in order to achieve higher throughput in the middle of the frontier.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/512c6225-5816-493e-839d-e4f539e5b1c4_2048x1215.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_metric=y_tpPerGpu&i_xmode=interactivity&i_linelabel=0&i_best=0&i_active=gb200_dynamo-vllm%2Cgb300_dynamo-trt)"
+/>
+
+## The two points are closer on token rate than on first token
+
+The 2xDEP8+1xDEP12 GB200 point is significantly closer to the 3xDEP8+1xDEP16 GB300 point in terms of tokens per second per user than it is in terms of time to first token. Reading only the interactivity axis would therefore understate how differently these two configurations behave.
+
+TTFT is, in general, more sensitive to the spikiness of the workload. The GB300 point achieves much higher overall concurrency, so it incurs more subagent traffic and hence more cold prefills. A [subagent](/glossary/subagent) starts with fresh context and no reusable prefix, which means its first request is a full prefill rather than a cache hit, and enough of them arriving together is what moves the [tail](/glossary/tail-latency).
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/96f9c9f9-eb86-4f87-8332-c85f23040836_2048x916.png"
+  caption="Source: InferenceX"
+/>
+
+## Why this is an agentic effect rather than a hardware one
+
+None of this appears on a fixed-sequence workload. At 8k1k every request is independent and identically shaped, so concurrency scales the batch without changing the mix of cached and uncached work. On AgentX, raising concurrency changes the composition of the traffic itself: more sessions means more parallel subagent branches, and those branches arrive in bursts rather than as a smooth stream.
+
+The practical reading is that a rack-scale configuration tuned to sit further right on the throughput curve is also accepting a burstier prefill load, and the cost of that decision lands on first-token latency rather than on the token rate the same chart advertises.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f79e53f5-c8ff-4ed0-a8a3-c05b9e0b4bab_2048x1191.png" />
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "Which NVIDIA rack-scale configuration is fastest on DeepSeek V4 Pro?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "GB300 Dynamo TRTLLM and GB200 Dynamo vLLM are the most competitive NVIDIA configurations on this model, and both rely on prefill-decode disaggregation to reach high throughput at reasonable interactivity. GB300 additionally runs DEP32 wide expert parallelism decode instances to lift throughput in the middle of the frontier."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does the GB300 point have worse TTFT than GB200?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "The GB300 point achieves much higher overall concurrency, which brings more subagent traffic and therefore more cold prefills. Because a subagent starts with fresh context and no reusable prefix, its first request is a full prefill, and bursts of those requests move the tail. TTFT is generally more sensitive to workload spikiness than the per-user token rate is."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/deepseek-v4-pro-agentx-mi355x-vs-b200-august.mdx b/packages/app/content/blog/deepseek-v4-pro-agentx-mi355x-vs-b200-august.mdx
new file mode 100644
index 000000000..32c0cfe0c
--- /dev/null
+++ b/packages/app/content/blog/deepseek-v4-pro-agentx-mi355x-vs-b200-august.mdx
@@ -0,0 +1,84 @@
+---
+title: 'DeepSeek V4 Pro on AgentX: MI355X vs B200, and the August 21 Flip'
+subtitle: 'AMD matched B200 vLLM on performance per dollar for end-to-end latency, then upstream vLLM work moved the line'
+seoDescription: 'AgentX 1.0 results for DeepSeek V4 Pro 0813 on MI355X against B200. MI355X SGLang matched B200 vLLM on performance per dollar until August 21, 2026, when vLLM optimizations from Inferact and NVIDIA put B200 ahead.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - deepseek
+  - mi355x
+  - b200
+  - amd
+  - nvidia
+---
+
+DeepSeek V4 Pro 0813 is an ultra popular frontier open weight model from China, with roughly 1.6 trillion parameters and 49 billion active. It is also the model where the AMD and NVIDIA race on [AgentX](/glossary/agentx) was closest, and where the ranking changed inside a single week.
+
+The workload matters here. AgentX replays real agentic coding sessions rather than a fixed prompt, and across all DeepSeek V4 runs the input length distribution came out at p50 88k, p90 272k, p95 404k, and p99 675k tokens, against an output distribution of p50 413, p90 2.2k, p95 3.7k, and p99 8.6k. That is a different machine problem from 8k1k.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/360a051c-6117-4860-a704-b50198f01521_1978x1246.png"
+  caption="Source: InferenceX"
+/>
+
+Before August 21, 2026, AMD's strong MI355X SGLang development team was matching B200 vLLM on performance per dollar for end to end performance. That is a real result and worth stating plainly.
+
+It was also not the whole board. B300 vLLM and B200 SGLang were already beating MI355X at the same time.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9ec4708c-148e-42ee-a7c5-3bf4d9cbdfaa_1938x1246.png"
+  caption="Source: InferenceX"
+/>
+
+After August 21, optimizations in vLLM from Inferact and NVIDIA pushed B200 performance per dollar past MI355X. This is a close race rather than a settled one, and [AMD has published its own list of DeepSeek V4 vLLM optimizations](https://github.com/vllm-project/vllm/issues/52911) with plenty left on it.
+
+## The ATOM asterisk
+
+On end to end latency, ATOM on MI355X beats B200 vLLM. It does not beat B300 or B200 SGLang, and the more important caveat is deployment: most AI labs in China or the west do not want to run ATOM in production, so a win that exists only under the vendor engine does not describe what a customer on an upstream stack receives. Single node open-source performance on MI355X, meaning vLLM, still trails ATOM.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e8d915c0-c962-4277-b62c-a1b08b8e7179_1924x1262.png"
+  caption="Source: InferenceX"
+/>
+
+## Where MI355X is actually competitive
+
+Overall MI355X performs decently against B200 and B300. Performance is most comparable at the lower throughput and lower latency part of the curve, where only tensor parallelism and more rudimentary kernels are deployed. The gap opens at high throughput, where AMD needs to optimize its DEP kernels, and it opens despite MI355X carrying 1.5x the HBM of B200.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/29cfd340-d61b-46c4-8d87-e081d937b80c_2048x1303.png"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_vllm%2Cb300_vllm%2Cmi355x_atom%2Cmi355x_vllm&i_label=0)"
+/>
+
+Read both axes together. A configuration can look strong on throughput against [interactivity](/glossary/interactivity) and give that back on [p90](/glossary/tail-latency) time to first token, which is why the comparison above is plotted against TTFT rather than interactivity alone.
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "Did MI355X beat B200 on DeepSeek V4 Pro?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "For a period before August 21, 2026, MI355X SGLang matched B200 vLLM on performance per dollar for end to end performance, and ATOM on MI355X beat B200 vLLM on end to end latency. B300 vLLM and B200 SGLang were beating MI355X over the same period, and after August 21 vLLM optimizations from Inferact and NVIDIA put B200 performance per dollar ahead of MI355X."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does MI355X lose ground at high throughput despite more HBM?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "MI355X carries 1.5x the HBM of B200, but on AgentX its performance is most comparable to B200 and B300 at the lower throughput and lower latency part of the curve, where only tensor parallelism and more rudimentary kernels are used. The high throughput end depends on DEP kernels that AMD still needs to optimize, so the extra memory capacity is not converted into frontier throughput."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72.mdx b/packages/app/content/blog/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
new file mode 100644
index 000000000..ffea55d49
--- /dev/null
+++ b/packages/app/content/blog/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
@@ -0,0 +1,75 @@
+---
+title: 'GLM 5.3 on AgentX: MI355X ATOM Beats GB300 NVL72 on Part of the Curve'
+subtitle: 'Where AMD’s vendor engine wins on performance per dollar, and what E2E Normalized Interactivity actually measures'
+seoDescription: 'On GLM 5.3, AMD ATOM has better performance per dollar than GB300 NVL72 SGLang and even TRTLLM for parts of the p90 E2E Normalized Interactivity range. An explainer on what that experimental metric measures and where it misleads.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - glm5
+  - mi355x
+  - gb300
+  - nvl72
+  - trtllm
+  - sglang
+  - amd
+  - nvidia
+---
+
+When looking at ATOM, AMD has better [performance per dollar](/glossary/performance-per-dollar) than GB300 NVL72 SGLang and even TRTLLM for some parts of the range of p90 E2E Normalized Interactivity. Great work to the AMD team on these results.
+
+The usual two requests apply to both vendors. We look forward to AMD porting these optimizations over to SGLang, and we are also looking forward to NVIDIA quickly optimizing GB300 NVL72 in the coming weeks.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9337ed5b-a320-49f7-b31b-cc4b5751e0ff_2048x1290.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+## What E2E Normalized Interactivity measures
+
+This result is plotted against an experimental metric worth understanding before drawing conclusions from it. [E2E Normalized Interactivity](/glossary/e2e-normalized-interactivity) is meant to evaluate how fast a user experiences responsiveness when considering both time to first token and tokens per second. It is defined as output sequence length divided by end to end latency.
+
+Substituting the fact that end to end latency equals TTFT plus output length times time per output token, and noting that in reality only output length minus one tokens are decoded, gives the following.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9e472a49-d228-4d8d-876a-6bc0d59f9d97_746x594.png"
+  caption="Source: SemiAnalysis"
+/>
+
+This is effectively [interactivity](/glossary/interactivity), the reciprocal-TPOT portion, plus an additional penalty proportional to TTFT.
+
+## Where the metric misleads
+
+Please note that this metric is experimental and is not perfect. For instance, it heavily penalizes high TTFT and does not capture all the nuances of certain optimizations such as prefill-decode [disaggregation](/glossary/disaggregated-inference). All submissions for AgentX v1.0 optimize for both regular interactivity and TTFT separately, and we will continue working on new north star metrics that reflect all nuances of modern agentic inference.
+
+That caveat cuts in both directions here. A configuration that trades first-token latency for streaming speed is penalized on this axis whether or not that trade suits the product, so a win measured only against E2E Normalized Interactivity should be checked against the separate interactivity and TTFT curves before it becomes a procurement argument.
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "What is E2E Normalized Interactivity?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "It is an experimental metric defined as output sequence length divided by end to end latency, intended to evaluate how fast a user experiences responsiveness when considering both TTFT and tokens per second. Because end to end latency equals TTFT plus output length times time per output token, it works out to interactivity plus a penalty proportional to TTFT."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Where does AMD ATOM beat GB300 NVL72 on GLM 5.3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "For some parts of the p90 E2E Normalized Interactivity range, ATOM has better performance per dollar than GB300 NVL72 SGLang and even TRTLLM. The metric is experimental and heavily penalizes high TTFT, so the result should be checked against the separate interactivity and TTFT curves."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/glm-5-3-agentx-nvidia-vs-amd-sglang-150-toks.mdx b/packages/app/content/blog/glm-5-3-agentx-nvidia-vs-amd-sglang-150-toks.mdx
new file mode 100644
index 000000000..d3fd0ad64
--- /dev/null
+++ b/packages/app/content/blog/glm-5-3-agentx-nvidia-vs-amd-sglang-150-toks.mdx
@@ -0,0 +1,69 @@
+---
+title: 'GLM 5.3 on AgentX: NVIDIA Is Up to 5x Cheaper per Token at 150 tok/s/user'
+subtitle: 'At this operating point, free AMD silicon would still not close the gap'
+seoDescription: 'GLM 5.3 AgentX results on open source SGLang. At 150 tok/s/user p90 interactivity, NVIDIA has up to 5x better cost efficiency, a gap so large that free competitor hardware would still leave NVIDIA cheaper per token.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - glm5
+  - b200
+  - b300
+  - mi355x
+  - sglang
+  - nvidia
+  - amd
+---
+
+GLM 5.3 builds on top of GLM 5.2 744B with additional post training. This is a frontier level model.
+
+In terms of open source SGLang performance, this is another model where NVIDIA again beats AMD on realistic agentic inference performance. At 150 tok/s/user p90 [interactivity](/glossary/interactivity), NVIDIA has up to 5x better cost efficiency.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/4a21d140-b588-4b33-941d-7c8090ac42b5_2048x1286.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=GLM-5.2&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_prec=fp4&i_active=b200_sglang%2Cb300_sglang%2Cgb200_dynamo-sglang%2Cgb300_dynamo-trt%2Cmi355x_atom%2Cmi355x_sglang)"
+/>
+
+## The free-hardware test
+
+There is a useful way to size a gap like this. With the current state of AMD software, at 150 tok/s/user, NVIDIA's performance advantage is so great that even if the competitor chip hardware was sold for free, with providers still of course paying for datacenter hosting, power, and other operating costs, cost per token would still be cheaper when using NVIDIA.
+
+That framing matters because hardware price is the lever most often reached for in a procurement conversation, and at this operating point it is not a lever that reaches far enough. The constraint is software, not the bill of materials.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f85bdb45-792c-4330-9745-674b38b03425_2048x1289.png" />
+
+## This is a snapshot, not a verdict
+
+The gap is measured against the current state of AMD software, and that state has been moving quickly elsewhere in these results. We look forward to AMD's performance optimizations in the upcoming AgentX update article in a couple of weeks, which will also include some other very exciting results.
+
+It is also worth being precise about scope: this is the open source SGLang comparison. AMD's vendor engine tells a different story on parts of this same model, which is [covered separately](/blog/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72).
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "How much cheaper is NVIDIA than AMD on GLM 5.3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "On open source SGLang at 150 tok/s/user p90 interactivity, NVIDIA has up to 5x better cost efficiency. With the current state of AMD software at that operating point, the advantage is large enough that even free competitor hardware, with providers still paying for hosting, power, and other operating costs, would leave cost per token cheaper on NVIDIA."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Does AMD win anywhere on GLM 5.3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Yes, but not on this comparison. The 5x figure is the open source SGLang result. With ATOM, AMD has better performance per dollar than GB300 NVL72 SGLang and even TRTLLM for parts of the p90 E2E Normalized Interactivity range."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72.mdx b/packages/app/content/blog/kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
new file mode 100644
index 000000000..0c18cb56e
--- /dev/null
+++ b/packages/app/content/blog/kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
@@ -0,0 +1,70 @@
+---
+title: 'Kimi K3 on AgentX: MI355X ATOM Beats GB300 NVL72 on Part of the Curve'
+subtitle: 'AMD’s vendor engine wins a real slice of the performance per dollar frontier, while Hopper struggles to serve K3 at all'
+seoDescription: 'On Kimi K3 AgentX runs, MI355X ATOM beats GB300 NVL72 vLLM on performance per dollar between 40 and 60 seconds of end-to-end latency, while Hopper struggles because K3 needs custom SM90 kernels and tuned TP32 and EP32 shapes.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - kimi
+  - mi355x
+  - gb300
+  - nvl72
+  - h200
+  - amd
+  - nvidia
+---
+
+On the part of the curve between 40 and 60 seconds of end to end latency, MI355X ATOM beats even GB300 NVL72 vLLM on performance per dollar. On a 2.8 trillion parameter model, against a rack-scale NVIDIA system, that is a result AMD earned.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/27a3f80e-7959-4651-90eb-f498989908a3_1934x1284.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+## The upstream caveat
+
+It is great that AMD is quickly pushing K3 performance forward with [ATOM](/glossary/atom). The encouragement that goes with that is to prioritize upstreaming these improvements into vLLM. ATOM is currently AMD's best-performing engine, but vLLM remains the more relevant comparison for customers using an upstream open-source serving stack, and a curve that only exists under a vendor runtime does not describe what those customers can deploy.
+
+## Hopper cannot really play
+
+Hopper struggles to serve the AgentX workload for Kimi K3. Part of that is simply that Kimi is a massive model, and part of it is that vLLM maintainers and NVIDIA have not been focusing on optimizing Hopper for K3.
+
+The specifics are concrete rather than vague: Hopper, at SM90, requires custom tuned kernels for K3, along with TP32 and EP32 tuned shapes for serving at high [interactivity](/glossary/interactivity). Neither exists yet, so the hardware is not the binding constraint so much as the absence of anyone working on it.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ce2b58fc-b7b7-46c2-ae05-baa14419d233_2048x1459.png"
+  caption="Source: SemiAnalysis InferenceX"
+/>
+
+This is a useful reminder that a generation gap in a benchmark chart is partly a generation gap in engineering attention. Hopper served DeepSeek R1 well for years because that model received sustained kernel work. K3 has not received the same, and the curve reflects the difference.
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "Does AMD beat NVIDIA anywhere on Kimi K3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Yes, on part of the curve. Between 40 and 60 seconds of end to end latency, MI355X ATOM beats even GB300 NVL72 vLLM on performance per dollar. The caveat is that ATOM is AMD's vendor engine rather than an upstream open-source stack, so vLLM remains the more relevant comparison for most customers."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does Hopper struggle on Kimi K3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Kimi K3 is a massive model, and vLLM maintainers and NVIDIA have not focused on optimizing Hopper for it. Serving K3 at high interactivity on SM90 requires custom tuned kernels plus TP32 and EP32 tuned shapes, and that work has not been done."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/kimi-k3-the-manos-the-mythos-the.mdx b/packages/app/content/blog/kimi-k3-the-manos-the-mythos-the.mdx
index 43e244de7..f4d11a898 100644
--- a/packages/app/content/blog/kimi-k3-the-manos-the-mythos-the.mdx
+++ b/packages/app/content/blog/kimi-k3-the-manos-the-mythos-the.mdx
@@ -686,7 +686,7 @@ This benchmark is a big step up from our previous 8k1k/1k1k benchmark, as it tru
 <Figure
   src="/images/kimi-k3-the-manos-the-mythos-the/inferencex-agentic-trace-distributions.png"
   alt="Distribution histograms of the agentic trace dataset: input tokens per turn, output tokens per turn, uncached input tokens per request, and turns per conversation"
-  caption="Source: [InferenceX](https://inferencex.semianalysis.com/datasets)"
+  caption="Source: [InferenceX](https://inferencex.semianalysis.com/agentx)"
 />
 
 For Kimi K3, Day 0 bringup was easier than DSv4 due to better documentation and preparation ahead of weights release. Appropriate images and a speculative decoder model were released at the same time as the weights.
diff --git a/packages/app/content/blog/minimax-m3-agentx-b200-b300-vs-rack-scale.mdx b/packages/app/content/blog/minimax-m3-agentx-b200-b300-vs-rack-scale.mdx
new file mode 100644
index 000000000..7007b1598
--- /dev/null
+++ b/packages/app/content/blog/minimax-m3-agentx-b200-b300-vs-rack-scale.mdx
@@ -0,0 +1,75 @@
+---
+title: 'MiniMax M3 on AgentX: Why B200 and B300 Beat Their Rack-Scale GB200 NVL72 & GB300 NVL72 Counterparts'
+subtitle: 'The Dynamo router becomes the bottleneck, no submission runs context parallelism, and AMD leaves KV offload on the table'
+seoDescription: 'On MiniMax M3, B200 and B300 beat GB200 and GB300 on TCO-normalized throughput. The Dynamo router can bottleneck because its work scales with live prefixes, no submission runs context parallelism despite a 317k p90 input, and AMD Pareto points use no KV offload.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - minimax
+  - b200
+  - b300
+  - gb200
+  - gb300
+  - nvl72
+  - dynamo
+  - amd
+  - rocm
+---
+
+B200 and B300 completely beat their rack-scale counterparts for MiniMax M3 on TCO-normalized throughput. For a model where NVIDIA sweeps the field, that is a surprising internal result, and the explanation is not about the fabric.
+
+On AgentX the rack-scale advantage is not as pronounced, because the Dynamo router can become the bottleneck: its work scales with the number and length of live prefixes rather than with tokens generated. There are also no well tuned kernels for wide expert parallelism, wide DCP, or wide tensor parallelism on this model. Since GB200 and GB300 carry higher TCO, without wide EP or wide DCP that shows up directly as worse performance per TCO.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/aec34e63-34fd-4c42-87a1-97e5c14661ed_2048x1302.png" />
+
+We also expect further optimizations from NVIDIA on their rack-scale solutions for this SKU, and will highlight them in the follow up article.
+
+## Nobody is running context parallelism
+
+No submission currently runs [context parallelism](/glossary/context-parallelism), despite a p90 input sequence length of 317k tokens. That is exactly the regime the technique exists for, so the absence needs an explanation.
+
+M3 has 4 KV heads, so DCP caps at 2 even at TP8, and the MSA indexer needs its own context-parallel handling, for which a vLLM pull request is open. The ceiling is therefore architectural rather than a tuning oversight, and it bounds how much a long-context parallelism strategy can help on this particular model.
+
+## The AMD offload gap
+
+All of NVIDIA's Pareto optimal points include [KV offload](/glossary/kv-cache-offload) above concurrency 20. For AMD, none of the Pareto optimal points use KV offload to DRAM, and AMD uses offload less than NVIDIA on the other models too.
+
+The reason is mechanical. GPU-to-CPU transfers for CPU KV cache offloading are highly inefficient on AMD vLLM, because the hipMemcpyBatchAsync API was missing until ROCm 7.14. Without it, vLLM's native simple CPU offloading has to perform serialized memcpy operations from CPU to GPU instead of batching them into larger message sizes.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2b40624a-223c-44db-a403-0be563b2af82_2048x1298.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_hc=1&i_active=mi355x_vllm)"
+/>
+
+A missing batched-copy API is not the kind of thing that shows up in a specification sheet, and on a fixed-sequence benchmark it would never be exercised at all. On long agentic sessions it decides whether an entire tier of memory is usable.
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "Why do B200 and B300 beat GB200 and GB300 on MiniMax M3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "On TCO-normalized throughput, the rack-scale advantage is muted on AgentX because the Dynamo router can become the bottleneck, since its work scales with the number and length of live prefixes rather than with tokens generated. There are also no well tuned kernels for wide expert parallelism, wide DCP, or wide tensor parallelism on this model, and the higher TCO of GB200 and GB300 is not earned back without them."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does AMD use no KV offload on MiniMax M3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "GPU-to-CPU transfers for CPU KV cache offloading are highly inefficient on AMD vLLM because the hipMemcpyBatchAsync API was missing until ROCm 7.14. Without it, native simple CPU offloading performs serialized memcpy operations instead of batching them into larger messages, so no AMD Pareto optimal point uses DRAM offload while every NVIDIA one above concurrency 20 does."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/minimax-m3-agentx-b300-trtllm-tp2-vs-the-field.mdx b/packages/app/content/blog/minimax-m3-agentx-b300-trtllm-tp2-vs-the-field.mdx
new file mode 100644
index 000000000..69d8f62d0
--- /dev/null
+++ b/packages/app/content/blog/minimax-m3-agentx-b300-trtllm-tp2-vs-the-field.mdx
@@ -0,0 +1,78 @@
+---
+title: 'MiniMax M3 on AgentX: B300 TRT-LLM TP2 Owns the Crown'
+subtitle: 'NVIDIA sweeps the 432B model, and the missing DP-attention points explain why cache locality became a routing constraint'
+seoDescription: 'MiniMax M3 432B AgentX results. B300 TRT-LLM TP2 leads the frontier, DP attention is absent because KV cache locality becomes a routing constraint, and a GB200 TP4/EP4/DPA point reaches 0.60x the throughput of plain TP4 at over 3x the p90 TTFT.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - minimax
+  - b300
+  - b200
+  - gb200
+  - trtllm
+  - vllm
+  - nvidia
+---
+
+NVIDIA absolutely destroys all competitors on MiniMax M3 432B. AMD software performance is horrible on MiniMax, especially at high context length, which we attribute to AMD engineering leadership incentivizing tuning only for short context single turn workloads and ignoring long context multi turn workloads.
+
+Within the NVIDIA field, B300 TRT-LLM at TP2 owns the M3 crown.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/5179bd8e-c147-46c8-aeb4-b347049b28a7_2048x1303.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft)"
+/>
+
+## The missing DP-attention points
+
+There is a lack of [DP-attention](/glossary/dp-attention) points on the M3 frontier, and the absence is informative. Data-parallel attention is non-optimal on M3 because KV cache locality becomes a routing constraint.
+
+The numbers make the mechanism concrete. For GB200 at concurrency 40, TP4/EP4/DPA gets 0.60x the throughput of plain TP4 at more than 3x the p90 TTFT. At concurrency 32 it hits 28.8% of cache against 96.0% theoretical. Each DP rank owns a private quarter of the pool, so a 300k-token session re-landing on the wrong rank recomputes everything.
+
+That gap between 28.8% measured and 96.0% theoretical is the whole story of [prefix reuse](/glossary/prefix-cache-hit-rate) under data parallelism: the cache was large enough, and the session simply was not sent to the rank that held it.
+
+No decode config with expert parallelism appears on the M3 frontier either, likely because the concurrency is not high enough to balance the loads across all experts.
+
+## vLLM against TRT-LLM
+
+It is worth mentioning that vLLM performance is very comparable to TRT-LLM in terms of throughput versus p90 interactivity. On throughput versus p90 TTFT, vLLM performs better.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/729199d9-66c0-42b5-b1b8-f228503c1585_2048x1289.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_trt%2Cb200_vllm%2Cb300_trt%2Cb300_vllm&i_hc=1&i_label=0)"
+/>
+
+Which of the two you prefer therefore depends on which axis your product is written against, which is the recurring argument for reading interactivity and first-token latency together rather than picking whichever chart flatters a conclusion.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/68f7ad68-3886-4980-9b01-390582728e47_2048x1292.png" />
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "Which configuration is fastest on MiniMax M3?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "B300 TRT-LLM at TP2 owns the M3 crown. NVIDIA leads the model overall, while AMD software performance is poor on MiniMax, especially at high context length."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why is data-parallel attention missing from the M3 frontier?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "DP attention is non-optimal on M3 because KV cache locality becomes a routing constraint. On GB200 at concurrency 40, TP4/EP4/DPA reaches 0.60x the throughput of plain TP4 at more than 3x the p90 TTFT, and at concurrency 32 the measured cache hit rate is 28.8% against a 96.0% theoretical ceiling, because each DP rank owns a private quarter of the pool and a session landing on the wrong rank recomputes everything."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/openai-jalapeno-better-than-nvidia.mdx b/packages/app/content/blog/openai-jalapeno-better-than-nvidia.mdx
new file mode 100644
index 000000000..20fa26656
--- /dev/null
+++ b/packages/app/content/blog/openai-jalapeno-better-than-nvidia.mdx
@@ -0,0 +1,383 @@
+---
+title: 'OpenAI Jalapeño: Better Than Nvidia Blackwell'
+subtitle: 'OpenAI’s self-designed ASIC compared with Rubin, Jalapeño’s TCO, throughput per MW, and spicy deets'
+seoDescription: 'OpenAI’s first-generation inference ASIC, built with Broadcom and benchmarked on InferenceX. Perf per MW against Blackwell and Vera Rubin, the HBM4 and MXFP4 specs, the Gluon software stack, why OpenAI skipped prefill-decode disaggregation, and the Katsu, Vindaloo, and Chana rack architecture.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - inference
+  - benchmark
+  - gpu
+  - nvidia
+  - openai
+  - asic
+  - rubin
+---
+
+OpenAI has spent the past couple years quietly building “Jalapeño,” an inference chip just announced at Hot Chips. Rumors of a successful tapeout had been swirling for a while. But now we have details. OpenAI invited us to look at their chip, go to their labs to check out how real it is, and [benchmark](https://openai.com/index/jalapeno-first-results/) it with our [InferenceX](https://inferencex.semianalysis.com/) suite.
+
+In June, [OpenAI unveiled the chip program](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/) in partnership with Broadcom, built from a blank slate exclusively for LLM inference. [Design work began in the middle of 2024](https://newsletter.semianalysis.com/p/openai-chip-team-is-now-serious), going from initial team hiring to manufacturing tape-out in ~16 months, an extremely fast ASIC development cycle.
+
+In general first generation chips are not competitive, but OpenAI bucks the trend by being industry leading and beating every Nvidia, AMD, and Google chip we have been able to test on multiple top open source models. OpenAI does this with extreme hardware software codesign. Surprisingly, OpenAI is not over specialization on any specific part of model inference, but instead by focusing on being a general chip that delivers high performance in all scenarios.
+
+In this article, we will go into architectural details, software details and performance results for Jalapeño on InferenceX.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c028f756-4748-4583-91b2-5c4a63893cb5_2048x1153.png"
+  caption="Source: [OpenAI](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/)"
+/>
+
+## A generalized inference chip
+
+Everyone says that OpenAI’s chip is specialized for OpenAI models, but that’s wrong, OpenAI made a generalized chip for AI inference.
+
+The timelines are insane. It shows that claims that use of AI is being used to accelerate chip design are real. Regardless of the quick timelines,Open AI spent a bunch of money, made pragmatic design decisions and their team is cracked, so this comes as no surprise.
+
+Just looking at the specs, it is an immediate contender:
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e346a4e5-76cb-4fd9-be95-00321116b605_1846x510.png"
+  caption="Source: SemiAnalysis"
+/>
+
+And the use of HBM4 makes it stand out as comparable to flagship GPUs from NVIDIA and AMD:
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/327f95f7-ab6b-44ae-9b8f-17c1ff5d296b_1712x1300.png"
+  caption="Source: OpenAI"
+/>
+
+A lot of the media coverage of this chip has followed a few throwaway comments from OpenAI that claim the chip will be optimized for their models in a way that other chips are not. This is wrong. Jalapeño is a generalized inference chip capable of running all sorts of models, and all sorts of workloads, including our benchmark InferenceX, where we ran the benchmark with OpenAI engineers in the lab. As a joke, OpenAI even showed us it running Doom, which was ported to their chip with just Codex prompts.
+
+The following is our headline perf/W result, looking at token throughput per All-in utility MW. **Jalapeño smokes every other chip**. All this is done without Multi Token Prediction (MTP), while the other chips on the chart are the best performing configs of each respective SKU, all with MTP.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/19a7f45a-df8e-436e-ba04-df5d8610f3da_2048x1330.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Jalapeño beats Blackwell on perf/W across almost all scenarios without being tuned for any specific point in the curve. It excels not only in low-latency scenarios but also in high-throughput scenarios. A more apples to apples comparison is against Single Token Prediction results, it knocks every competitor out of the water. At low concurrency scenarios, Jalapeño demonstrates remarkable interactivity, hitting over 700 tokens per sec per user at concurrency 1 on the DeepSeek R1 model.
+
+Incredibly, this is all achieved with single-token prediction (STP), no speculative decoding and no prefill-decode disaggregation. In addition to DeepSeek R1, we also got to see some other models, including Kimi-K2.5 and GPT-OSS which ran at approximately 1,400 tok/sec/user. For all models, we confirmed that Jalapeño’s GSM8k evals attained results on par with Nvidia chips.
+
+Some caveats on this. First, all numbers are provided to us by OpenAI. We verified the InferenceX runs in person in the lab, but we did not run the full suite of [InferenceX](https://inferencex.semianalysis.com/) benchmarks nor have we seen [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) results. AgentX is our preferred suite for comparing chip performance due to the datasets’ long context and multi-turn characteristics that reflect the cache behavior of realistic production workflows. Frameworks that perform well on 8k1k may perform worse on AgentX as real production loads stress components like routers, prefix cache mechanisms, cache management, offload infrastructure, etc. These are not tested by single turn 8k1k. Read more about this in out AgentX article.
+
+[AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) — Since the Claude Code inflection point in November 2025, long-context, multi-turn agentic workloads have grown rapidly. They now dominate traffic for production inferencing. In April 2026, OpenAI’s Enterprise agentic spending overtook ChatGPT spending.
+
+Second, we believe that comparison to Blackwell is somewhat incomplete and unfair. Jalapeño is really competing against chips like Rubin that also use HBM4. Vera Rubin systems are starting to ship to customers right now, while it will still be some time before OpenAI has anything beyond engineering samples of Jalapeño.
+
+Thus, performance should really be compared against Rubin, not Blackwell, and in some sense we expect a custom chip like Jalapeño to outperform Blackwell. Vera Rubin NVL72 delivers 5.4x the perf/MW of GB200 NVL72 [as we described in our article analyzing the NVIDIA performance claims in their launch with CoreWeave last month](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference). We will compare Jalapeño to Vera Rubin’s July performance figures later below.
+
+[Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference) — Vera Rubin NVL72 is the second generation of Nvidia’s rack-scale Oberon architecture, and its gains on inference come from extreme co-design. Early results from engineering samples are encouraging. Vera Rubin NVL72 running DeepSeek R1 delivers 5.4x performance per MW and 5x performance per dollar over GB200 NVL72 today, and the gap is even wider against GB200 NVL72 during its early bringup in 2025. Vera Rubin is still in the early bringup stage now, so we expect the gap to continue widen. Rubin's inference performance will keep improving as software matures, the same pattern we demonstrated for Blackwell in our
+
+Third, the models being tested are not on the open frontier. NVIDIA and AMD have published results on larger models such as DeepSeek V4 Pro and Kimi K3, using AgentX. The larger the model and the more recent the release, the more complicated it is to bring up on a new chip. With that said the models OpenAI has working on Jalapeno aren’t exactly small either.
+
+## Performance Analysis
+
+OpenAI designs for perf/W. The reason is simple: OpenAI is currently limited by datacenter power, not by budget or floorspace, and thus tokens per MW is paramount. At Computex 2026, Jensen said that perf/W, reliability and long lifetime are the core features of future GPUs. To quote: “If you have 1 gigawatt of power, then throughput per watt is revenue”. He also mentioned that choosing the wrong architecture just because the chips are cheaper doesn’t make sense.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/69dc13f6-ffb1-4d56-8df7-b75b57280382_1980x1254.png"
+  caption="Source: Computex 2026 keynote"
+/>
+
+This was emphasized by Nvidia during the Vera talk at Hot Chips 2026 while showing the same revenue graph: “The data center is power limited today.” Power matters and drives revenue.
+
+Operators cannot simply obtain more MW because adding GPUs and adding grid capacity happen on very different timescales. Datacenter power envelopes have constraints such as their utility interconnection, infrastructure, cooling capacity, and UPS/backup-generation design. Grid delays repeatedly outpace hardware and construction timelines, driving the need for BtM (behind-the-meter) power capacity: gas turbines and on-site generators built and located at the data center itself. This capacity sits behind the utility’s meter rather than being drawn from the public grid. It lets an operator power a facility without waiting on grid interconnection and utility upgrades, which is exactly why xAI’s Colossus 2 relies so heavily on BtM while its actual grid connection lags far behind. Find out more in our [Energy model](https://semianalysis.com/energy-model/).
+
+As we wrote in an X post, tok/s/MW reduces to tokens per joule since a watt is a joule per second. This makes tok/s/MW representative of a system’s efficiency and ability to convert energy into tokens.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c1a1da73-425c-4800-b396-01c7d2151be9_1200x1432.png"
+  caption="Source: [SemiAnalysis](https://x.com/SemiAnalysis_/status/2091208991271420076)"
+/>
+
+On this front, even when compared with Rubin, Jalapeño wins. OpenAI’s Jalapeño has STP output token throughput per MW surpassing [Vera Rubin’s MTP results that NVIDIA and CoreWeave published in July. It also far exceeds GB200’s 2025 MTP results.](https://www.coreweave.com/blog/nvidia-vera-rubin-nvl72-on-coreweave-10x-more-tokens-per-megawatt-than-blackwell) As mentioned in our Vera Rubin article, VR was compared to 2025 GB200 results because that was a similar stage of early bring-up, and comparing to GB200 in 2025 holds software maturity constant. Following this logic, we compare Vera Rubin’s latest July 2026 results, GB200 2025 results, and today’s Jalapeño results. This is a very valid comparison as these are the best public Rubin numbers, and OpenAI taped out their chip after Rubin. Both OpenAI and Rubin are still immature thus performance will continue to rise.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/6e8f9fd2-ec7f-45fa-80b9-dc132e32c661_2048x1450.png"
+  caption="Source: OpenAI, SemiAnalysis"
+/>
+
+On perf/TCO, Vera Rubin and Jalapeño are head-to-head, producing almost the same number of output tokens per $. However, as previously mentioned, **Jalapeño’s results are obtained without speculative decoding** and Vera Rubin’s results use speculative decoding. Speculative decoding leads to a ~3-5x reduction in cost per token. When speculative decoding is implemented on Jalapeño, this will enable Jalapeño to serve tokens even more cost effectively. Of course, part of this TCO advantage comes from trading Nvidia’s high margins for Broadcom’s lower (though still high) margins. But this is not all of it. For example, Meta and Microsoft’s AI ASIC programs not getting off the ground despite being at it for much longer shows that cost is only one part of the equation. For Jalapeño’s full TCO breakdown, see the [SemiAnalysis AI Cloud TCO model](https://semianalysis.com/ai-cloud-tco-model/).
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a6c58331-c05a-4823-9c4b-aae90705bb53_2048x1485.png"
+  caption="Source: OpenAI, SemiAnalysis"
+/>
+
+Architecturally, OpenAI chose not to disaggregate prefill and decode (PD) across separate chip pools. The draft model and main model share the same chips and fabric, a design philosophy that trades some theoretical efficiency for practical operations. The motivation is that the workload mix changes over time, for example the ratio of input to cache write to cache read to output tokens has changed significantly as we have moved through the three eras of models ([knowledge, reasoning, and agentic, as discussed in our recent article](https://newsletter.semianalysis.com/p/are-open-models-catching-up)). Therefore, picking a fixed amount of heterogenous prefill silicon and decode silicon up front can lead to inefficiencies over time. OpenAI chooses a homogenous pool in this architecture and tries to make the chip perform well on everything.
+
+And it does. On Kimi K2.5 (which Cursor Composer 2.5 is based on), Jalapeño reaches nearly 700tok/s/user and more than 9x the next best performing chip at 100tok/s/user.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/3088f64c-9bee-46d9-a445-6818248ba573_2048x1366.png"
+  caption="Source: OpenAI, SemiAnalysis"
+/>
+
+On GPT-OSS, it’s another bloodbath. Jalapeño’s iso-interactivity throughput per MW is nearly double GB200’s highest throughput point and more than 50x GB200’s concurrency 1 point. The higher concurrency Jalapeño points use EP8.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9f245dd7-c6b7-48a3-9215-3976633bb77f_2048x1366.png"
+  caption="Source: OpenAI, SemiAnalysis"
+/>
+
+These results are impressive! However, we have to nitpick: they’re just 8k1k, a much easier workload to tune for, and there are no [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) runs yet. As mentioned in our [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) article, multiturn, long context workloads stress much more aspects of the serving stack, such as routers and prefix cache. Many more optimizations are needed to excel in agentic workloads. Read more about this in the [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) article.
+
+[AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) — Since the Claude Code inflection point in November 2025, long-context, multi-turn agentic workloads have grown rapidly. They now dominate traffic for production inferencing. In April 2026, OpenAI’s Enterprise agentic spending overtook ChatGPT spending.
+
+## Digging into the specs and architecture
+
+All these results were gathered on the A0 stepping of Jalapeño, just 9 months into the program. But there is already a B0 stepping that is currently in the fab! B0 has optimizations that deliver roughly a 25% perf-per-watt improvement over the earlier A0 silicon. Specifically, the B0 stepping delivers 13.4 PFLOPs of MXFP4 on a single reticle-sized compute die that is manufactured on TSMC’s N3P. This compares to 17.5 PFLOPs of dense Rubin NVFP4 for a single Rubin compute die that is similar size and on the same node.
+
+This is more respectable considering Jalapeño’s TDP is only 700W compared to Rubin’s at 900-1,150W per compute die. As Jalapeño is geared towards inference rather than training, it is understandable that OpenAI doesn’t need to push TDPs higher to maximize FLOPs, but regardless the above shows that Jalapeño delivers respectable peak theoretical FLOPs.
+
+When compared directly to other accelerators, Jalapeño has the highest HBM bandwidth per watt, and the highest FLOPs per watt, comparable to the 1,800W Rubin Max-Q configuration:
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c5795c45-d23b-403c-8bc4-f13d5ba6fb45_2048x394.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Off-package I/O is provided by an N3E I/O chiplet with 32 lanes of 800G SerDes, for the compute fabric, with 24 lanes (600GB/s) being used for local scale-up within the rack, and 8 lanes (200GB/s) for global scale-up which is the 2,048 XPU multi-rack domain. PCIe Gen 5 is used for system I/O to connect to the x86 host CPU.
+
+Jalapeño will ship with HBM4, making this chip one of the relatively early adopters after Nvidia and AMD, even beating the established TPU and Trainium programs. As one of the key architectural principles behind Jalapeño is getting the most out of HBM bandwidth, settling for anything but the best HBM would run counter to that goal. This results in 15.4TB/s of memory bandwidth per package which bests all the other accelerators shipping that are using HBM3E. The 15.4TB/s bandwidth shows its HBM4 can hit 10Gbps pin speeds, which would give it a slight edge over the 9.6Gbps Nvidia is getting out of its HBM4 in Rubin. The HBM is likely provided by Samsung.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/0d7d181a-2bb9-45d4-bd67-2f7e525ce6d1_1426x376.png"
+  caption="Source: OpenAI"
+/>
+
+OpenAI taped out Jalapeño in November 2025, or more specifically, this was a tape out of the CoWoS design, not just the top die silicon. Within 9 months of that Nov 2025 tapeout, and with only 3 months of bring-up on actual silicon, OpenAI has already delivered very good results with Jalapeño. This is all the more impressive as the team is starting from zero on the software stack.
+
+Meanwhile, Rubin’s CoWoS tape out was completed in October 2025, a month earlier, and yet the only early results we have seen are from CoreWeave’s engineering samples. Nvidia has not let us test and release benchmarks in the same way that OpenAI has, indicating their chip software is still immature. The CUDA moat is potentially dead given how fast OpenAI can bring up new models on their silicon.
+
+They are still far from optimized and we can see that generally Jalapeño has delivered better numbers. We don’t think that Nvidia hardware is inferior, but more so that Jalapeño’s software bring-up has progressed more quickly than Nvidia’s. This speaks to the power of hardware/software co-design, which is the main area where a cracked frontier lab ASIC team can excel over more established merchant silicon players. Counterintuitively, starting from scratch may also have benefited OpenAI as it could make clean-sheet architectural decisions without worrying about backwards compatibility or older software versions.
+
+While OpenAI has engineering samples of Jalapeño, production is currently scheduled to gradually ramp over 2027 with most of the output currently scheduled for the end of next year. [For more details of unit volumes and ASPs, see the SemiAnalysis Accelerator Model](https://semianalysis.com/accelerator-hbm-model/).
+
+Suffice to say, OpenAI Jalapeno is a real high volume ASIC.
+
+When compared against Rubin’s timeline, Jalapeño’s is shockingly quick. As shown earlier, Jalapeño’s results beat Rubin’s despite Rubin’s head start.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/bb59cacc-1f16-4c6a-98a5-0f0db74dada7_2048x411.png"
+  caption="Source: SemiAnalysis"
+/>
+
+## Jalapeno Architecture
+
+Digging into the architecture now, the chip’s matrix engine uses MXFP numerical formats and a weight stationary systolic array, similar to TPU. But when compared directly to TPU, it has support for smaller shapes / dimensions, meaning that it doesn’t have weird performance cliffs that get exposed by awkwardly shaped matmuls on bigger systolics.
+
+It also has 64-bit scalar cores and FP32/INT32 vector cores. OpenAI has also invested in redundancy at the tray level and has yield harvesting built in at the core and channel level. They claim that AI assistance in chip design delivered an 8% reduction in SIMD area and a 10% reduction in matrix-engine area during design. While they did not clarify the exact process/voltage/temperature (PVT) conditions, they also mentioned the AI-assisted blocks improved timing and power over the initial blocks.
+
+The Jalapeño architecture design focuses on eliminating memory movement of KVCache and weights as well as fixed latencies and overheads in order to make it possible to get closer to the raw peak flops/bandwidth even for small batches or shapes as compared to other accelerators.
+
+The cores and the HBM are divided into slices, where each core slice has a low-latency local view on its own slice of HBM. Synchronization between slices occurs on a high-bandwidth dedicated collective network. This minimal memory hierarchy already gives Jalapeño a big potential advantage over GPUs, where memory accesses must traverse a complicated memory system, resulting in large latencies that must be amortized or hidden over larger shapes.
+
+This choice is feasible because with careful placement of weights and KVs, synchronization between cores can be restricted to limited, known high-bandwidth comms such as tensor-parallel communication that can be overlapped with compute.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/871631ee-8f0d-4fb6-869c-a6fb298e800e_2048x832.png"
+  caption="Source: OpenAI"
+/>
+
+There is also an additional general NoC which is used for general comms and to access the scale-up network. In general OpenAI saves huge power and gets big performance gains with a simplified NOC and memory subsystem vs Nvidia and Google.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/1f93df86-2ea2-4904-8fd1-e0e5eff4b410_1638x854.png"
+  caption="Source: OpenAI"
+/>
+
+At the core level, OpenAI describes an out-of-order (OoO) core with an L1 cache. This is a large divergence from the pattern we have seen in other accelerators, all of which instead use software-managed scratchpad commonly paired with some async DMA support. Again, the argument being made here is that this allows Jalapeño to avoid fixed overheads such as barrier latencies, which on other accelerators (such as GPUs) need to be hidden or amortized over with higher work per core, and make it harder to get close to the raw peak bandwidth/flops.
+
+The tradeoff is that Jalapeño therefore relies on good prefetching to ensure timely arrivals of memory requests, which is less predictable and more difficult to reason about. However, with Codex in a good harness with access to detailed tracing, it is likely that finding the optimal kernel with the best prefetching for a given shape requires little human intervention. We think that is exactly what OpenAI has done to bring up DeepSeek R1, Kimi K2.5, and GPT-OSS so quickly.
+
+The cores also have support for “small” matrix dimensions, which (depending on how small) should make it more general across different model and batch dimensions less sensitive to matrix dimension alignment, padding overhead, and tiling inefficiency. For instance, TPUs, Trainium, and Etched chips have very large systolic arrays which can require large batches or exactly-divisible model dimensions to avoid tiling inefficiencies.
+
+With Jalapeño, OpenAI has focused on eliminating fixed latencies in the system to allow for as-close-to-roofline performance as possible across all areas of the pareto curve. In theory, this could give them advantages over the GPU at multiple operating points:
+
+- Much better upper-bound performance on low-latency/small-batch inference, which on GPUs is limited by many fixed overheads such as launch latencies, barrier latencies, memory system latency
+- Some potential to achieve closer to the hardware roofline even for large-batch or long-context
+
+This comes with the caveat that even if the upper-bound performance is available in theory, it may be more difficult to realize that performance for real kernels. So it seems the approach is:
+
+1. Design for the highest upper-bound performance across all workload shapes
+2. Let Codex do the tedious work of finding the kernels that achieve that upper bound
+
+Judging by the extremely fast turnaround for the OpenAI team to bring up InferenceX workloads on Jalapeño, we are optimistic about this approach.
+
+If Jalapeño is a success, it will be a strong signal that the industry’s obsession over programming models and perfect, universal compilers are invalidated by frontier AI models.
+
+## Software
+
+OpenAI writes Jalapeño kernels like assembly. Each kernel gets hand-tuned code, some running to ~3,000 lines, backed by correctness checks and a custom sanitizer. Early kernel work was human-in-the-loop rather than fully automated, but this shifted with a more scaled-up, internal version of Codex, one which OpenAI plans to pitch to enterprise customers. The internal serving engine is called “Teacup”. Interestingly, **OpenAI had no internal implementation of MLA kernels until they benchmarked DeepSeek with InferenceX**. The ability for Codex to write functional and efficient kernels so quickly (without any of OpenAI’s kernel engineering team intervening) shows the software pipeline’s developmental ability.
+
+OpenAI programs Jalapeño with Gluon. Gluon is OpenAI’s kernel programming language. Built on top of Triton, Gluon preserves Triton’s SPMD (Single Program Multiple Data) programming model, but it exposes **low-level programming abstractions**. For example, for NVIDIA GPUs, it offers APIs that map to PTX instructions, including MMA instructions, TMA instructions, mbarrier mechanisms, and many more. The most unique abstraction Gluon provides is **the layout**. Generally speaking, a layout defines a mapping between a hardware resource (e.g. 5th register of warp 9) and a tensor element (e.g. tensor element on row 6 column 7). Gluon’s layout abstraction is based on [Linear Layouts](https://arxiv.org/abs/2505.23819), a type of layout algebra OpenAI invented. Linear Layouts mathematically formalizes what a layout is and provides tools to operate on layouts. This enables many features, such as provably correct layout conversions and optimal memory swizzling.
+
+In terms of Jalapeño’s programming model, each Gluon program maps to a persistent thread. We believe this hints that Jalapeño suits the **persistent kernel programming pattern**, where each program executes on multiple tiles, and the programmer, rather than the hardware scheduler, assigns the work. OpenAI mentioned TensorInfo, an abstraction that explicitly encodes layouts. This is likely the set of layouts designed for Jalapeño, which will be powered by Linear Layouts. Finally, each core offers data prefetching and decoupled out-of-order units. For example, a user might program a wait on a prefetched data, which is locked behind a semaphore.
+
+In a weird twist of fate, OpenAI models like GPT 5.6 Sol, which currently run on NVIDIA GPUs, have been used to design a chip that poses a real threat to the CUDA moat - NVIDIA’s own GPUs are helping usher in their potential successor in real time.
+
+Comparing across time, we can also see Jalapeño’s developmental pace, achieving more than 2x throughput improvements at certain interactivities in less than 2 weeks. Each tarball we get from the Jalapeño team has a world of wonders inside.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/b38d958b-6d53-4ebe-b92f-c2cb58de05c2_2048x1485.png"
+  caption="Source: OpenAI, SemiAnalysis"
+/>
+
+Not only did kernel performance improve, in the span of 8 days, the Jalapeño team enabled TP32, building on the previous TP8 configs and expanding beyond a single system to get a full rack-scale config running on a large model. This is a really impressive pace of development.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a18d5611-f7c1-4e34-8919-99b1c683054f_2048x1485.png"
+  caption="Source: OpenAI, SemiAnalysis"
+/>
+
+To validate performance before committing to real hardware runs, OpenAI also has a simulator “chilisim” accurate to within 5% of measured hardware, using a fixed-width trace bus. Tracing on A0 was limited but has improved substantially on B0, likely with inputs from actual runs on A0 silicon. Engineers have demoed the Codex CLI running an internal model, nicknamed “Raiku” or “5.3 Codex Spark”, at 1.2ms TPOT.
+
+The team also showed off Codex-written demos running directly on the chip: Doom at 36 FPS, an FP32 fluid-dynamics simulation, and a “Liquid Light” mouse-drag visualization.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/d410878f-5f19-40bd-ad2c-d92ad45d65a9_1252x1232.png"
+  caption="Source: SemiAnalysis"
+/>
+
+On the model side, OpenAI’s internal megakernel approach, nicknamed “gigakernel”, is built around a single megakernel that loops on-device to reduce CPU overhead and launch time. The team is also leaning further into test-time compute strategies, with internal interest specifically in how to coherently use 1 million rollouts.
+
+## To Disagg or Not to Disagg, ‘tis the Question
+
+We mentioned earlier that OpenAI is not using prefill decode disaggregation on these chips. This came as a surprise to us, as NVIDIA and AMD GPU performance benefits significantly from PDD, even on homogenous hardware. Let’s dig into why the Jalapeño team went this way.
+
+Prefill-decode disaggregation (PDD) looks attractive when the workload is frozen. Prefill and decode stress hardware differently, so assigning each phase to a separately tuned pool can improve efficiency at one chosen input/output ratio. Production traffic, however, does not stay at that ratio. Input and output sequence lengths, concurrency, cache-hit rates, speculative-acceptance rates, and latency targets all move throughout the day.
+
+Once devices are divided into prefill and decode pools, too much prefill demand leaves decode chips idle while requests queue. But too much decode demand does the opposite. The operator must continuously predict the right split, provision spare capacity on both sides, and rebalance a system whose ideal ratio is always moving.
+
+In a unified system, some resources may be underused during a particular phase, but every device remains available to serve the next request. In a disaggregated system, an entire chip can sit idle simply because it belongs to the wrong pool. Local utilization looks better, but global utilization can be bad.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/92d363ab-c391-4b9a-9f67-516dac22f534_1364x938.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Disaggregation also breaks locality. The prefill worker produces a large KV cache that the decode worker immediately needs, so the system must transfer that state across the network before generation can continue. That adds bandwidth consumption, synchronization, queueing, and another failure domain. The cost also rises with input sequence length because KV cache grows. However, avoiding the movement of KVs is largely a power and latency optimization; being willing to move some KVs around can allow for increased hardware utilization at the expense of some power and per-request latency.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/5789e0ea-2d0d-4924-b338-f0559c7e7ee9_1928x1292.png" />
+
+A fungible fleet shifts capacity between latency-sensitive requests and throughput-oriented batches, while a fixed split strands hardware whenever the traffic mix changes. Moreover, context length changes the balance between attention and FFN work, making any fixed hardware ratio efficient only near its design point.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9f3a37f8-0996-4dea-9631-02ade61d7fae_1358x938.png"
+  caption="Source: SemiAnalysis"
+/>
+
+The same constraint applies to speculative decoding. A draft model has to feed candidate tokens to the verifier with extremely low latency. Separating the two across specialized pools turns a tightly coupled decoding loop into a distributed protocol. The extra communication and coordination can consume the latency saved by drafting. Keeping both models on the same devices and low-latency fabric preserves the locality that makes speculation worthwhile in the first place.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2eb19374-5a08-4d82-84a3-9875b3fc0f5b_1980x1220.png"
+  caption="Source: SemiAnalysis"
+/>
+
+However, disaggregation can still win where demand is sufficiently large, stable, and predictable, particularly when conventional GPUs need large phase-specific batches to reach good throughput. But it is not free lunch.
+
+## From Japanese to Indian (Mild to Spicy): Katsu, Vindaloo, and Chana - How are the curry dishes put together into a rack system?
+
+The Jalapeño System at the rack unit level consists of a CPU host rack and an ASIC rack. The host rack houses 16 host CPU trays named “Katsu,” each corresponding to one of the 16 ASIC trays, named “Vindaloo,” to the right of the Katsu. Each host houses two Turin-class AMD EPYC CPUs with 1.5TB of DRAM, 2x E1.S, and 2x M.2 SSDs per rack. Each tray is also specced with 400G (2x200G) frontend networking. Each Katsu tray connects to each Vindaloo tray via 8 external PCIe DAC cables that run horizontally across the rack at the front. The system level design is done in partnership with Celestica.
+
+The ASIC rack consists of 16 Vindaloo trays and 8 scale up switch trays (6 for local + 2 for global), named “Chana.” Each Vindaloo tray consists of 8 Jalapeño ASICs, making up a total of 128 Jalapeño ASICs per rack. The ASICs are connected to each of the Chana switch trays via a copper cable backplane, just like that of Nvidia’s Oberon. The scale up topology is split into a local domain of 128 ASICs within the rack and a global domain connecting up to 16 racks or 2,048 ASICs. We will explain the bandwidth and the topology in more detail below.
+
+Power provisioning to a sidecar host rack draws roughly 50kW provisioned (31kW in production), and the ASIC rack draws 130kW, making the total two rack system roughly 160kW. That’s basically a double-wide GB300 rack in terms of power draw.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/596ca531-3bde-44fa-8d1b-34d04b8cc6b8_1304x1382.png"
+  caption="Source: SemiAnalysis, OpenAI"
+/>
+
+OpenAI can connect up to 2,048 Jalapeño XPUs within a single scale-up network. The scale-up network consists of two domains, a local domain connecting all 128 XPUs over backplane within the rack, as well as a global domain connecting 2,048 XPUs over 16 racks using a hybrid of copper and optical interconnect. Each rack consists of 8 Chana switch trays. Six Chana switches in the middle are for the local domain, which come with one 102.4T Tomahawk 6 switch ASIC each. Two Chana switches at the top and bottom of the local switches are for the global domain, which we think could consist of 2x 102.4T Tomahawk 6 switches making up to 204.8T per switch tray.
+
+In the local domain, each of the 128 Jalapeño chips has a per XPU uni-directional bandwidth of 4.8Tb/s and is connected on an all-to-all basis to 6x 102.4Tb/s Tomahawk 6 ASICs. This would amount to 48-differential pair (DP) male and female connector pairs per XPU translating to a total of 6,144DPs worth of passive copper cables per rack used for local scale-up.
+
+For the global domain, 16 racks totaling 2,048 XPUs are connected together via a combination of copper backplane, electrical 204.8T TH6 switch, 1.6T transceivers, and optical circuit switch. Each XPU has a uni-directional bandwidth of 1.6Tb/s for the global link, which is 16-differential pair (DP) male and female connector pairs per XPU for the backplane between the XPU and the global switch. Bandwidth exiting each global switch tray of 2 ASICs each is split between the backplane and front panel optics.
+
+Between local domain and global domain, backplane connector count per rack comes up to 64 DPs per XPU and a total of 8,192 DPs worth of passive copper cables per rack.
+
+The global domain adopts a rail-only architecture consisting of 8-rails across the global domain. We think OpenAI routes optical links in the global domain via Optical Circuit Switches (OCS) installed in every rack. For every XPU, 1.6Tb/s of global bandwidth will travel to the global switch tray over the copper backplane. This then exits the switch through the front panel via 1.6T transceivers, which go to the passive optical switch before exiting the rack. This expands the scale-up world size to 2,048 XPUs combining 16 racks of 128 XPUs each.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ae9b26b8-471a-43b2-a60c-cd99ce2b9b07_3480x1342.png"
+  caption="Source: SemiAnalysis"
+/>
+
+Because scale-up networking is only about 10% of total system cost, that flexibility buys valuable optionality for future 10–20 trillion parameter models or 2–4 million token context windows. On deployment, OpenAI is partnering with neoclouds and is gathering reliability data with datacenter partners through January while optimizing dock-to-rack rollout time.
+
+## What’s Next
+
+Next, we talk about the future of Jalapeño, whose first production token is coming soon. The next goal is 100MW, and the hurdles will mostly be hardware: How much can they produce, how well can they deploy and operate datacenters, how do they handle monitoring, and resiliency, etc. The software is already proven, and with internal models, every software headstart is easily caught up to. Behind the paywall we will discuss implications for NVIDIA, AMD, Cerebras, and other chip companies who have signed deals with OpenAI in the coming years.
+
+[We also cover production volumes, units, and timelines for the next generation chip in our Accelerator model here](https://semianalysis.com/accelerator-hbm-model/).
+
+## Direct Implications for NVIDIA, AMD and Cerebras
+
+<Blur>
+
+Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+
+</Blur>
+
+---
+
+_This article continues on our Substack. [Subscribe to SemiAnalysis](https://newsletter.semianalysis.com/subscribe) to read the complete article._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "What is OpenAI Jalapeño?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Jalapeño is OpenAI's first self-designed inference ASIC, built with Broadcom from a blank slate for LLM inference and announced at Hot Chips. Design work began in the middle of 2024 and the CoWoS design taped out in November 2025, roughly a 16 month cycle from team hiring to tape-out. The B0 stepping now in the fab delivers 13.4 PFLOPs of MXFP4 on a single reticle-sized N3P compute die at a 700W TDP, with HBM4 providing 15.4 TB/s of memory bandwidth per package."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "How does Jalapeño compare to NVIDIA Blackwell and Vera Rubin?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "On token throughput per all-in utility megawatt, Jalapeño beats Blackwell across almost all scenarios, and its single-token-prediction results surpass the Vera Rubin MTP figures NVIDIA and CoreWeave published in July 2026. On performance per dollar of TCO, Jalapeño and Vera Rubin are roughly head to head. Blackwell is the weaker comparison because Jalapeño competes against HBM4 parts like Rubin, and Jalapeño achieves these results without speculative decoding while the NVIDIA configurations use it."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why did OpenAI skip prefill-decode disaggregation?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Disaggregation is attractive when the workload mix is frozen, but production traffic is not: input and output lengths, concurrency, cache hit rates, speculative acceptance, and latency targets all move through the day. Once chips are split into prefill and decode pools, an entire chip can sit idle because it belongs to the wrong pool, so local utilization improves while global utilization suffers. A unified pool also preserves the locality that speculative decoding depends on, since a draft model separated from its verifier turns a tight decoding loop into a distributed protocol."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "How is a Jalapeño rack put together?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "The system is two racks. A host rack holds 16 CPU trays named Katsu, each with two AMD EPYC Turin-class CPUs and 1.5TB of DRAM, connected by external PCIe DAC cables to the ASIC rack. The ASIC rack holds 16 Vindaloo trays of 8 Jalapeño ASICs each, for 128 ASICs per rack, plus 8 Chana scale-up switch trays built on Tomahawk 6. Scale-up runs as a 128-ASIC local domain over a copper backplane and a global domain reaching 2,048 ASICs across 16 racks. The two-rack system draws roughly 160kW."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What software stack runs on Jalapeño?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "OpenAI programs the chip with Gluon, its own kernel language built on Triton, which keeps Triton's SPMD model while exposing low-level abstractions. Its distinguishing feature is the layout abstraction, based on Linear Layouts, a layout algebra OpenAI invented that enables provably correct layout conversions and optimal memory swizzling. The internal serving engine is called Teacup, a simulator called chilisim tracks measured hardware to within 5 percent, and a scaled-up internal version of Codex writes much of the kernel code."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "What are the caveats on these Jalapeño results?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "All numbers were provided by OpenAI. SemiAnalysis verified the InferenceX runs in person in the lab but did not run the full benchmark suite, and there are no AgentX results yet. The published runs are 8k1k single-turn, which is easier to tune for than long-context multi-turn traffic: agentic workloads stress routers, prefix caching, cache management, and offload infrastructure that a single-turn 8k1k run never exercises. The models tested are also not the newest frontier open-weight releases."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/qwen3-5-397b-agentx-b300-fp4-vs-h100.mdx b/packages/app/content/blog/qwen3-5-397b-agentx-b300-fp4-vs-h100.mdx
new file mode 100644
index 000000000..b5c3ac749
--- /dev/null
+++ b/packages/app/content/blog/qwen3-5-397b-agentx-b300-fp4-vs-h100.mdx
@@ -0,0 +1,66 @@
+---
+title: 'Qwen3.5 397B on AgentX: B300 FP4 Delivers 12x the Performance per Dollar of H100'
+subtitle: 'What four years of hardware and a 4-bit format buy on a long-context agentic workload'
+seoDescription: 'On Qwen3.5 397B AgentX runs, B300 FP4 has 12x better performance per dollar than H100. The comparison spans four hardware generations and two numerical formats on a 256k truncated agentic trace replay.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - qwen
+  - b300
+  - h100
+  - h200
+  - fp4
+  - fp8
+  - sglang
+  - nvidia
+---
+
+Compared to H100, on Qwen3.5, B300 FP4 has 12x better [performance per dollar](/glossary/performance-per-dollar).
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/cc2cf6c7-aa58-4a76-a0a2-ee0457d9243c_2048x1298.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=interactivity&i_hc=1&i_best=0&i_prec=fp4%2Cfp8&i_active=b300_sglang%2Ch100_sglang%2Ch200_sglang)"
+/>
+
+## What the number is and is not
+
+Two things are changing at once here, and both belong in the reading. The first is four years of hardware. The second is numerical format: the comparison spans [FP4](/glossary/fp4) and [FP8](/glossary/fp8) rather than holding precision constant, which is the honest way to compare what an operator can actually deploy today on each part, but not a controlled study of the silicon alone.
+
+The workload matters as much. This is not 8k1k. Qwen3.5 has a native max context length of 262k tokens, so AgentX replays the truncated 256k agentic dataset against it, which means frequent context compaction and heavy prefix reuse. A ratio measured on a long-context multi-turn replay is a different claim from the same ratio measured on independent short prompts, and it is the one closer to production coding traffic.
+
+## Read it at your own operating point
+
+A 12x aggregate ratio is a headline, not a deployment plan. The curve is what carries the decision: pick the [interactivity](/glossary/interactivity) your product needs, read each system there, and check the first-token latency at the same point before concluding anything.
+
+That is also why the per-model AgentX charts default to showing the whole frontier rather than a single number. Older hardware can remain competitive at low concurrency on performance per dollar while falling away entirely at high throughput, and a single ratio cannot express both halves of that shape.
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "How much better is B300 than H100 on Qwen3.5 397B?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "On the AgentX Qwen3.5 runs, B300 FP4 has 12x better performance per dollar than H100. The comparison spans both a hardware generation gap and a numerical format difference, since it covers FP4 and FP8 rather than holding precision constant."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Is the 12x figure measured on a fixed sequence length workload?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "No. Qwen3.5 has a native max context length of 262k tokens, so AgentX replays the truncated 256k agentic trace dataset against it. That workload involves frequent context compaction and heavy prefix reuse, which is a different measurement from independent short prompts."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/qwen3-5-397b-agentx-nvidia-vs-amd-sglang.mdx b/packages/app/content/blog/qwen3-5-397b-agentx-nvidia-vs-amd-sglang.mdx
new file mode 100644
index 000000000..03c32b0c4
--- /dev/null
+++ b/packages/app/content/blog/qwen3-5-397b-agentx-nvidia-vs-amd-sglang.mdx
@@ -0,0 +1,67 @@
+---
+title: 'MI355X versus GB300 NVL72 Inference Performance: 20x Gap on Qwen3.5 SGLang'
+subtitle: 'GatedDeltaNet, a 262k native context, and no AMD competition at all on the same engine'
+seoDescription: 'Qwen3.5 397B AgentX results. On SGLang against SGLang, NVIDIA is over 20x better at 90 tok/s/user with zero competition from AMD, and NVIDIA SGLang submissions hold much lower p90 TTFT than TRT-LLM.'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - qwen
+  - sglang
+  - trtllm
+  - nvidia
+  - amd
+---
+
+Qwen3.5 397B uses GatedDeltaNet instead of vanilla attention for every couple of layers. GatedDeltaNet was invented at MIT and NVIDIA Research and has a theoretically constant state storage requirement instead of vanilla attention's linear storage requirement, which means lower storage than an equivalent dense attention model. Unlike end to end model training research such as the Nemotron disaster, NVIDIA Research is great at fundamental research like GDN and LatentMoE, both of which are used in frontier models.
+
+One methodology note before the results. This model's native max context length is 262k tokens, so AgentX replays the [truncated 256k dataset](https://inferencex.semianalysis.com/agentx/cc-traces-weka-062126-256k) against it. That simulates a workload on a smaller model where the max context length would be frequently reached with many compactions, which is how users would actually use this model.
+
+## The gap
+
+Qwen3.5 397B is a strong hold for NVIDIA on SGLang versus SGLang, with over 20x better performance at 90 tok/s/user. There is currently zero competition from AMD for Qwen3.5 SGLang.
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/519a7cd0-ac51-4a5c-9c45-ba7c41f6d058_2048x1300.png"
+  caption="Source: [SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_hc=1)"
+/>
+
+Holding the engine constant is what makes this comparison worth stating. A cross-engine result can always be argued about, since the software stacks differ in maturity and in who is paid to optimize them. Here both sides are SGLang, and the gap is still 20x.
+
+## NVIDIA is trading TTFT for interactivity
+
+Within the NVIDIA field, we again observe over-optimization for [interactivity](/glossary/interactivity) at the cost of time to first token, especially in the case of TRT-LLM. In the chart above, all of the NVIDIA SGLang submissions have much lower p90 TTFT when compared to TRT-LLM.
+
+That is a real tradeoff rather than a defect, but it is one a reader has to see. A recipe that streams faster after making the user wait longer will look better on an interactivity axis and worse against a first-token [service level objective](/glossary/service-level-objective), and only reading both together tells you which one fits your product.
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/c92da639-dcff-4e39-b9b5-49742677e93c_2048x1289.png" />
+
+---
+
+_These results are one slice of AgentX 1.0. The full analysis, the replay methodology, and the 70+ upstream PRs the benchmark drove are in [AgentX - InferenceXv3: Does the CUDA Moat Hold Up in Agentic Inferencing?](/blog/agentx-inferencexv3-does-cuda-moat). Every point is explorable on the [free dashboard](https://inferencex.semianalysis.com/)._
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "How much faster is NVIDIA than AMD on Qwen3.5 397B?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "On SGLang against SGLang, NVIDIA is over 20x better at 90 tok/s/user, and there is currently zero competition from AMD for Qwen3.5 SGLang. Holding the engine constant on both sides is what makes the comparison meaningful."
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Why does AgentX replay a truncated dataset for Qwen3.5?",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "The model has a native max context length of 262k tokens, so AgentX uses the truncated 256k context dataset. That simulates a workload on a smaller model where the max context length is frequently reached with many compactions, which is how users would actually use this model."
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/ultra-high-interactivity-on-nvidia.mdx b/packages/app/content/blog/ultra-high-interactivity-on-nvidia.mdx
index ef532f0e5..b6f9f3ef6 100644
--- a/packages/app/content/blog/ultra-high-interactivity-on-nvidia.mdx
+++ b/packages/app/content/blog/ultra-high-interactivity-on-nvidia.mdx
@@ -3,6 +3,7 @@ title: 'Ultra-High Interactivity on NVIDIA GPUs? TileRT on InferenceX'
 subtitle: 'Can TileRT software on NVIDIA GPUs compete with Cerebras, Groq LPU, and SambaNova? Batch size 1, disaggregated engine, high-throughput prefill engine, high-interactivity decode engine'
 date: '2026-08-10'
 publishDate: '2026-08-10'
+modifiedDate: '2026-08-18'
 tags:
   - benchmark
   - gpu
@@ -13,6 +14,8 @@ tags:
   - tilert
   - vllm
   - glm5
+  - agentx
+  - agentic
 ---
 
 _Originally published on the [SemiAnalysis newsletter](https://newsletter.semianalysis.com/p/ultra-high-interactivity-on-nvidia) on August 10, 2026._
@@ -235,9 +238,9 @@ Work is being done to simplify this, especially as software development can be a
 
 AI coding agents accelerate tuning within known templates, but novel transformations still require expert judgment. A monolithic persistent kernel also reduces the usefulness of conventional per-kernel profiler timelines, making automated feedback loops more difficult.
 
-## Next steps with TileRT and InferenceX
+## Next steps with TileRT
 
-We are actively working on moving TileRT benchmarking from InferenceX's single-turn 8k/1k to our new agentic coding benchmark, which we call AgentX. This scenario replays real Claude Code and Codex traces with long-context, multi-turn requests, realistic subagent activity, and dynamic tool-use delays. Its median input length is 140k tokens, while the theoretical median cache-hit rate roofline reaches 99.2%.
+AgentX, InferenceX's long-context, multi-turn coding scenario, is now live for supported models and serving stacks. It replays privacy-preserving workload shapes derived from real Claude Code and Codex sessions, with long-context, multi-turn requests, realistic subagent activity, and dynamic tool-use delays. Its median input length is 140k tokens, while the theoretical median cache-hit rate roofline reaches 99.2%. Extending TileRT benchmarking from its current single-turn 8k/1k workload to AgentX remains the next step for this serving stack.
 
 <Figure
   src="/images/ultra-high-interactivity-on-nvidia/agentx-benchmark-characteristics.png"
diff --git a/packages/app/content/blog/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx b/packages/app/content/blog/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx
index ccc0e0782..1036d017f 100644
--- a/packages/app/content/blog/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx
+++ b/packages/app/content/blog/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx
@@ -223,7 +223,7 @@ The second nuance in CoreWeave’s performance results is that they are using De
 
 Oddly enough, CoreWeave’s choice of using DeepSeek R1 671B is theoretically more favourable towards the Blackwell baseline, and not Rubin. Rubin’s main advantages lie in a higher HBM capacity, higher CPU DRAM capacity, and greater HBM bandwidth, meaning that Rubin is more optimized for multi-trillion parameter models like Fable 5, Gemini Pro, Kimi K3, and Qwen3.8 2.4T.
 
-The third noteworthy item is that CoreWeave uses only single turn 8k/1k input/output tokens. Theoretically, multi-turn long context workloads like Agentic Coding should do better on Rubin, due to Rubin’s higher HBM capacity and bandwidth, but this would not be captured on a simple single-turn benchmark. [Our upcoming AgentX benchmark scenario created in collaboration with Weka, LMCache, the vLLM/SGLang community, Nvidia, AMD, and many others in the community will provide a realistic agentic workload to benchmark inference performance.](https://inferencex.semianalysis.com/datasets/cc-traces-weka-062126) We encourage everyone to adopt this inference benchmark!
+The third noteworthy item is that CoreWeave uses only single turn 8k/1k input/output tokens. Theoretically, multi-turn long context workloads like Agentic Coding should do better on Rubin, due to Rubin’s higher HBM capacity and bandwidth, but this would not be captured on a simple single-turn benchmark. [Our upcoming AgentX benchmark scenario created in collaboration with Weka, LMCache, the vLLM/SGLang community, Nvidia, AMD, and many others in the community will provide a realistic agentic workload to benchmark inference performance.](https://inferencex.semianalysis.com/agentx/cc-traces-weka-062126) We encourage everyone to adopt this inference benchmark!
 
 Finally, we note that CoreWeave’s testing was done on a pre-production rack without a scale-out fabric. Specifically, CoreWeave used a Dell Engineering Sample (ES) rack. We do believe these results are valuable as they use wide EP and PD disagg, which uses the NVL72 scale-up backplane and proves that it is working well. This backplane faced many reliability challenges during the ramp of GB200 NVL72 Oberon, [as we have noted in our Accelerator model.](https://semianalysis.com/accelerator-hbm-model/)
 
diff --git a/packages/app/content/blog/zh/agentic-benchmark-agent-benchmark-guide.mdx b/packages/app/content/blog/zh/agentic-benchmark-agent-benchmark-guide.mdx
new file mode 100644
index 000000000..6d8bf31a0
--- /dev/null
+++ b/packages/app/content/blog/zh/agentic-benchmark-agent-benchmark-guide.mdx
@@ -0,0 +1,287 @@
+---
+title: '面向 LLM 推理的 Agentic Benchmark：指标与方法'
+subtitle: 'Agent Benchmark 如何回放长上下文、多轮工作负载，并测量延迟、吞吐量、cache 行为与推理服务成本'
+seoDescription: '面向 LLM 推理的 Agentic Benchmark 指南：了解 Agent Benchmark 如何测量延迟、吞吐量、prefix caching、并发与成本。'
+date: '2026-08-19'
+publishDate: '2026-08-19'
+tags:
+  - benchmark
+  - agents
+  - agentic
+  - inference
+  - latency
+  - throughput
+---
+
+**面向 LLM 推理的 Agentic Benchmark** 衡量模型服务系统在 AI agent 流量下的表现。它不判断 agent 是否修复了 bug 或完成了工作流，而是测量承载每次模型调用的基础设施，包括延迟、吞吐量、cache 复用、并发能力与成本。
+
+这一目标会改变 workload。传统推理基准测试可以发送 input 和 output 长度固定的独立 prompt；Agent Benchmark 则必须表示完整会话，其中上下文随轮次增长，请求共享长 prefix，tool 执行带来停顿，subagent 产生并行分支。后续请求通常要等前序请求完成后才能开始。
+
+AgentX 是 InferenceX 面向这类 workload 的开放基准测试。它把自愿提供的编码 agent trace 转换为保护隐私的请求形状，再通过 AIPerf 进行确定性回放，用相同场景比较完整的软硬件推理配置。
+
+<DashboardCTA href="/zh/inference?i_seq=agentic-traces">对比实时 AgentX 推理结果 →</DashboardCTA>
+
+## 什么是 Agentic Inference Benchmark？
+
+Agentic Inference Benchmark 是面向多轮 agent 流量的系统基准测试。被测系统包括 accelerator、服务器拓扑、推理引擎、模型 checkpoint、精度、并行策略、cache 配置与请求调度器。
+
+它回答以下问题：
+
+- 一套部署能同时服务多少个 agent 会话？
+- 每个 agent 需要等待多久才能收到首 token？
+- 每个活跃 agent 接收 output token 的速度是多少？
+- 在目标延迟下，每颗芯片能提供多少 output 吞吐量？
+- 重复 prefix 会命中 HBM、host memory，还是触发重新计算？
+- 在可用的响应速度下，每百万 token 的成本是多少？
+
+“Benchmark”也常用于 agent 任务评估，因此必须说明测试范围。这两类测试回答不同问题：
+
+| 问题                        | 测试类型                    | 常见结果                             |
+| --------------------------- | --------------------------- | ------------------------------------ |
+| Agent 能否正确完成任务？    | Agent 任务评估              | 成功率、policy 合规率或评分器分数    |
+| 推理栈能否承载 agent 流量？ | Agentic Inference Benchmark | 延迟、吞吐量、cache 行为、容量与成本 |
+
+AgentX 对应第二类。任务语义与完成质量需要单独评估。
+
+## 为什么固定序列推理基准测试无法代表 Agent 流量？
+
+固定序列测试仍然适合做受控比较。例如，8K input 和 1K output workload 能隔离常见请求形状，并让不同配置容易复现，但它无法表示持续运行的 agent 会话。
+
+| Workload 特征  | 固定序列推理                    | Agentic 推理                       |
+| -------------- | ------------------------------- | ---------------------------------- |
+| 请求形状       | 固定 input 与 output 长度       | 每一轮长度都不同                   |
+| 到达模式       | 按固定速率或 batch 发送独立请求 | 前序请求完成后才触发后续依赖请求   |
+| Prefix 复用    | 可选或人工构造                  | 大型共享 prefix 随会话不断增长     |
+| Tool 时间      | 通常不存在                      | 模型调用之间存在不均匀停顿         |
+| 并行性         | 请求级并发                      | 主 agent 与 subagent 分支可能重叠  |
+| Cache 生命周期 | 短且规律                        | 长时间状态争用 HBM 与 offload 容量 |
+| 有效结论       | 一种请求分布下的性能            | 完整 agent 会话的容量与响应速度    |
+
+这些差异可能改变系统排序。更大的 HBM 能让长时间存活的 KV cache 保持驻留；更有效的 prefix caching 可以避免重新处理数十万 prompt token；对独立 prompt 影响较小的 scheduler 与 router 开销，也可能随活跃会话的数量和长度增长。
+
+## Agent Benchmark 是一张请求依赖图
+
+Agent 会话不是扁平的 prompt 列表。主 agent 轮次形成依赖链；某一轮可以启动一个或多个 subagent，主链随后可能等待这些分支完成；辅助调用也可以独立运行而不汇入主链。
+
+<Figure
+  src="/images/agentx-methodology/replay-multi-subagent.png"
+  alt="Agentic 推理基准测试回放图，包含主 agent 请求链、四条并行 subagent 分支、两个汇合点与一条独立辅助请求"
+  caption="AgentX 把会话表示为依赖图，保留主 agent 轮次、并行 subagent、汇合点、辅助调用与记录中的等待时间。"
+/>
+
+AgentX 通过[闭环模式](/zh/glossary/closed-loop-benchmark)回放这张图。每个客户端只有在依赖的前序工作完成后才发送下一个可执行请求。服务器越快，会话越早推进并产生后续工作。当 subagent 扇出时，瞬时 HTTP 请求数可以高于客户端 concurrency；分支完成后，请求数会再次下降。
+
+因此，AgentX concurrency 表示**活跃 agent 客户端数量**，并不是固定 request batch。若多个客户端同时存在活跃 subagent，concurrency 50 的运行可能拥有超过 50 个 in-flight 请求。
+
+## AgentX 如何构建可复现 Workload？
+
+AgentX 通过四个阶段，把已记录的会话结构转换为不含原始内容的回放。
+
+### 1. 采集请求形状与时间
+
+自愿启用的 HTTP proxy 会记录请求到达与完成时间、input 和 output token 数、conversation ID 与 subagent ID。公开数据不包含 prompt、源代码、tool argument 或 tool result。
+
+### 2. 保留 Prefix，而不保留内容
+
+每段 input 会被转换为以 64-token block 为单位的会话内 hash 链。相同 block ID 会保留不同轮次间的 prefix 关系。回放前，AIPerf 使用确定性的合成编码与 tool-use token 替换这些 block。
+
+### 3. 重建依赖关系
+
+主 agent 请求形成线性链；subagent 请求形成带有 spawn 和 join 依赖的独立链；记录中的轮次间停顿会重现 tool 执行与等待其他工作的时间。回放保留客户端可见的顺序与重叠关系，但不会推断 provider 内部事件。
+
+### 4. 在不同系统上回放同一场景
+
+固定 seed 控制会话采样、回放起点与合成 payload。每套软硬件配置都接收相同场景定义，因此曲线差异来自推理系统，而不是人工编写的 prompt batch。
+
+AgentX v1.0 从自愿提供的 Claude Code trace 中筛选出 393 个公开会话。每个会话至少包含 20 个请求，并且同时运行的 subagent 不超过 10 个。
+
+| AgentX v1.0 特征           |    数值 |
+| -------------------------- | ------: |
+| 公开会话                   |     393 |
+| 单请求 input token 中位数  | 142,016 |
+| 单请求 output token 中位数 |     444 |
+| 包含 subagent 的会话       |     44% |
+| Full 数据集最大上下文      |      1M |
+| 限制上下文的数据集变体     |    256K |
+
+<Figure
+  src="/images/agentx-methodology/request-distributions.png"
+  alt="AgentX v1.0 的对数分布图，展示轮次间延迟、input sequence length 与 output sequence length"
+  caption="AgentX v1.0 请求分布。单请求 input token 中位数为 142,016，output token 中位数为 444。"
+/>
+
+Full 变体保留最高 1M token 的上下文。256K 变体会移除超过上限的请求，同时保留其余请求的时间关系与 subagent 重叠。[AgentX 方法论](/zh/agentx/methodology)公开了数据集规则、图表与来源链接。
+
+## 保证比较有效的 Benchmark 控制
+
+长时间运行的 agent workload 对初始状态、cache 内容、speculative decoding 与 host memory 十分敏感。AgentX 通过明确规则锁定这些变量。
+
+### Cache Warmup 与测量窗口
+
+固定 seed 会在每段会话 25% 至 75% 的位置选择起点。`max_tokens=1` 的 primer 会建立活跃主 agent 与 subagent 的 prefix，随后每条回放 lane 再完成 10 个 warmup 请求，之后才开始测量。
+
+对外结果只统计随后一小时的 profiling 窗口。每次循环回放都会加入唯一 cache-bust 标记，避免无关会话逐步形成共享 prefix。
+
+### 合成 Payload 与 Speculative Decoding
+
+合成 token 会保留长度和 prefix 结构，却无法自然复现 draft token acceptance。AgentX 因此针对每组模型、speculative 方法、draft length 与 thinking mode，使用 [SPEED-Bench](https://arxiv.org/abs/2604.09557) 编码类别测得的 acceptance length。该数值记录在带版本的 golden 文件中，并通过推理引擎控制项应用。
+
+这样可以把推理服务性能与合成文本偶然造成的 speculator acceptance 差异分开。
+
+### DRAM 与 KV Cache Offload
+
+Host memory 容量会影响长 prefix 应该重新加载还是重新计算。没有标准化 DRAM 配置的服务器上限为 3 TB；GB200 NVL72、GB300 NVL72 与 TPUv7 等标准系统可以使用实际装机容量。每种基准测试配置只能按 GPU 占比使用对应的 host memory。
+
+回放格式与依赖规则记录在 [AIPerf WEKA trace 指南](https://github.com/ai-dynamo/aiperf/blob/main/docs/tutorials/weka-trace.md)中。
+
+## Agentic Benchmark 应该测量哪些指标？
+
+单个数字无法描述 agent 服务。应该在完整 concurrency sweep 中同时报告吞吐量、延迟与交互性。
+
+| 指标                             | 说明                                                |
+| -------------------------------- | --------------------------------------------------- |
+| 每颗芯片的 output 吞吐量         | 每颗 accelerator 每秒完成的生成 token 数            |
+| P90 interactivity                | 根据 p90 inter-token latency 得到的单客户端生成速度 |
+| 首 token 延迟（TTFT）            | Agent 开始收到响应前的等待时间                      |
+| Inter-token latency（ITL）       | 相邻 streamed output token 之间的时间               |
+| Input 吞吐量                     | 每秒处理的 prompt token 数                          |
+| Cache 来源分布                   | Prompt token 来自 HBM、host memory 还是重新计算     |
+| Queue 与 in-flight request depth | 分支启动和汇合时的 scheduler 压力                   |
+| 每百万 token 成本                | 在明确延迟或交互性要求下的推理服务经济性            |
+
+百分位数不能省略。如果部分 agent 分支需要在长 prefill 后排队，较低的平均 TTFT 可能与很差的 p90 或 p99 同时出现。总 output 吞吐量上升时，单个客户端接收 token 的速度也可能下降。
+
+## 应该读取完整的吞吐量与交互性曲线
+
+AgentX 扫描不同的并发 agent 客户端数量，并绘制相应运行点。低并发通常有利于单客户端响应速度；提高并发可以改善 batching 与总吞吐量，直到排队、内存压力或其他瓶颈开始降低交互性。
+
+<Figure
+  src="/images/agentx-methodology/pareto-b200-minimax-m3.png"
+  alt="MiniMax-M3 在 NVIDIA B200 上的 AgentX Pareto 曲线，展示不同客户端并发下每颗芯片的 output 吞吐量与 p90 交互性"
+  caption="每个点对应一个客户端 concurrency 设置。比较系统时应使用相同 p90 interactivity，而不是比较两个不匹配的峰值吞吐量点。"
+/>
+
+有效结果是 [Pareto frontier](/zh/glossary/pareto-frontier)，而不是唯一的最大吞吐量点。硬件 A 可能在低交互性下给出最高 tok/s/chip，硬件 B 却可能在交互式编码产品所需的响应速度下领先。
+
+闭环回放还意味着更快的配置能在相同 profiling 小时内推进到采样会话的更后位置，因此实际请求组合可能略有变化，低并发时尤其明显。结果应包含完整曲线，并使用足够并发降低采样噪声。
+
+## Agent 流量会考验推理栈的哪些部分？
+
+Agentic 推理会暴露固定 prompt batch 容易低估的瓶颈。
+
+- **KV-cache 容量：**长上下文与大量活跃会话会让大型 cache 驻留更久。
+- **Prefix-cache 正确性：**错误命中会产生错误状态，漏掉命中则会重复昂贵的 prefill。
+- **Offload 带宽：**从 host memory 移动长 prefix 可能比重新计算更快，但传输粒度与 descriptor 开销会影响结果。
+- **Scheduler 公平性：**并行 subagent 可以提高总吞吐量，同时让某条分支等待过久。
+- **Router 与 frontend CPU：**Prefix matching、过期项清理、序列化与 streamed response 处理都会随活跃会话状态增长。
+- **长上下文 kernel：**Attention、sparse attention、normalization 与 graph capture 路径可能随 sequence length 变化而表现不同。
+- **Speculative decoding：**Acceptance 控制与 scheduler 生命周期会同时影响吞吐量和交互性。
+
+[AgentX 优化追踪页](/zh/agentx/optimizations)记录了 vLLM、SGLang、TensorRT-LLM、ATOM、Dynamo 与 KV-cache 系统中由该 workload 暴露的具体改进。
+
+## 如何比较 AgentX 结果？
+
+比较双方应使用同一套场景约束：
+
+1. 对齐模型、checkpoint、精度与 thinking mode。
+2. 对齐 AgentX 数据集变体与上下文上限。
+3. 记录推理引擎、container image、并行策略与 speculative decoding 方法。
+4. 应用相同的 golden acceptance-length 规则。
+5. 对齐 DRAM 容量、KV-cache offload policy 与按比例分配内存的规则。
+6. 使用相同 seed、warmup 流程、profiling 时长与客户端 concurrency sweep。
+7. 当系统规模不同时，同时比较每颗芯片与整套系统结果。
+8. 在相同 p90 interactivity 或延迟目标下比较。
+9. 保留失败与不稳定配置，而不是只报告成功运行点。
+
+相同交互性下的比较通常比峰值吞吐量更有用。它直接回答部署问题：在用户需要的响应速度下，哪套系统每颗芯片或每美元能够服务更多 agent 工作？
+
+## 这项 Agent Benchmark 能回答什么，不能回答什么？
+
+AgentX 可以测量 trace 衍生 agent workload 下的推理容量、响应速度、cache 复用、内存压力、吞吐量与成本。它可以显示软件更新是否改善了相同硬件上的长上下文多轮服务，也可以比较不同硬件拓扑如何改变有效 frontier。
+
+AgentX 无法判断模型是否写出了正确代码、选择了正确工具、遵守了业务规则，或完成了原始任务。合成 payload 会有意移除这些语义。Agent 质量应使用任务评估，推理性能则使用 AgentX。两类分数描述不同的系统问题，不应混为一个总分。
+
+<DashboardCTA href="/zh/inference?i_seq=agentic-traces">
+  查看 AgentX 吞吐量与交互性 frontier →
+</DashboardCTA>
+
+## 常见问题
+
+### 什么是面向 LLM 推理的 Agentic Benchmark？
+
+它是一项系统基准测试，用于衡量推理栈如何服务多轮 agent 流量。测试会回放不断增长的上下文、共享 prefix、tool 等待与 subagent 分支，并报告延迟、吞吐量、交互性、cache 行为、容量与成本。
+
+### Agentic Benchmark 与固定序列推理基准测试有何区别？
+
+固定序列测试会发送 input 和 output 长度明确的独立请求。Agentic Benchmark 回放的是相互依赖的会话，其中请求长度、时间、prefix 复用与并行分支会持续变化，因此更接近真实 agent 产生的服务条件。
+
+### AgentX Benchmark 测量什么？
+
+AgentX 使用从编码 agent trace 衍生的流量，测量完整推理配置。它扫描并发 agent 客户端数量，并报告每颗芯片的 output 吞吐量、p90 interactivity、TTFT、ITL、cache 行为与推理服务成本。
+
+### AgentX 为什么使用闭环并发？
+
+Agent 无法提前发出所有轮次，因为后续动作依赖前序结果。在闭环回放中，每个客户端会在依赖完成后发送下一个可执行请求，从而保留服务器速度与会话产生后续工作的时间关系。
+
+### Agent 推理最重要的指标有哪些？
+
+应结合 output 吞吐量、p90 interactivity、TTFT、ITL、cache 来源分布、queue depth 与成本，并在相同延迟或交互性目标下比较完整 concurrency 曲线。单个最大吞吐量点无法描述交互式部署。
+
+### Agent Benchmark 会衡量模型或 Agent 质量吗？
+
+部分 Benchmark 会，但 AgentX 不会。AgentX 使用保留 workload 形状的合成 payload 来测量推理系统性能。任务正确性、工具选择与 policy 合规性需要单独的 agent 评估。
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "什么是面向 LLM 推理的 Agentic Benchmark？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "它是一项系统基准测试，用于衡量推理栈如何服务多轮 agent 流量。测试会回放不断增长的上下文、共享 prefix、tool 等待与 subagent 分支，并报告延迟、吞吐量、交互性、cache 行为、容量与成本。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Agentic Benchmark 与固定序列推理基准测试有何区别？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "固定序列测试会发送 input 和 output 长度明确的独立请求。Agentic Benchmark 回放的是相互依赖的会话，其中请求长度、时间、prefix 复用与并行分支会持续变化。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AgentX Benchmark 测量什么？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "AgentX 使用从编码 agent trace 衍生的流量，测量完整推理配置。它扫描并发 agent 客户端数量，并报告每颗芯片的 output 吞吐量、p90 interactivity、TTFT、ITL、cache 行为与推理服务成本。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AgentX 为什么使用闭环并发？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Agent 无法提前发出所有轮次，因为后续动作依赖前序结果。在闭环回放中，每个客户端会在依赖完成后发送下一个可执行请求，从而保留服务器速度与会话产生后续工作的时间关系。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Agent 推理最重要的指标有哪些？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "应结合 output 吞吐量、p90 interactivity、TTFT、ITL、cache 来源分布、queue depth 与成本，并在相同延迟或交互性目标下比较完整 concurrency 曲线。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Agent Benchmark 会衡量模型或 Agent 质量吗？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "部分 Benchmark 会，但 AgentX 不会。AgentX 使用保留 workload 形状的合成 payload 来测量推理系统性能。任务正确性、工具选择与 policy 合规性需要单独评估。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/agentx-inferencexv3-does-cuda-moat.mdx b/packages/app/content/blog/zh/agentx-inferencexv3-does-cuda-moat.mdx
new file mode 100644
index 000000000..2ca701034
--- /dev/null
+++ b/packages/app/content/blog/zh/agentx-inferencexv3-does-cuda-moat.mdx
@@ -0,0 +1,1031 @@
+---
+title: 'AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？'
+subtitle: '开源价值 300 万美元的数据集、100 万以上上下文长度、多轮、子智能体、95%+ KVCache 命中率、GB300 NVL72、MI355X、B200'
+seoDescription: 'AgentX 1.0 是首个完全开源、1M 上下文的多轮 agentic 编码推理基准测试。文中给出 DeepSeek V4 Pro、Kimi K3、MiniMax M3、Qwen3.5 和 GLM 5.3 在 GB300 NVL72、B300、B200、MI355X 和 H200 上的 agentic 结果，以及该基准测试推动的 70 多个上游 PR。'
+date: '2026-08-24'
+publishDate: '2026-08-24'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - gpu
+  - nvidia
+  - amd
+  - announcement
+---
+
+自 2025 年 11 月的 [Claude Code 拐点](https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point)以来，长上下文、多轮的 agentic 工作负载快速增长，如今已经主导了生产推理的流量。2026 年 4 月，OpenAI 的企业级 agentic 支出超过了 ChatGPT 支出。
+
+[Claude Code is the Inflection Point](https://newsletter.semianalysis.com/p/claude-code-is-the-inflection-point) — 当前 GitHub 公开 commit 中有 4% 由 Claude Code 编写。按现在的增长曲线，我们认为到 2026 年底 Claude Code 将占每日全部 commit 的 20% 以上。一眨眼的工夫，AI 已经吞掉了整个软件开发。
+
+agentic 工作流已经明确接过了接力棒。**今天，我们发布 AgentX 1.0 —— 全球首个完全开源、1M 上下文的多轮 agentic 编码推理基准测试，以 Apache 2.0 协议发布。[完整仪表板在此。](https://inferencex.semianalysis.com/)**
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/864512bf-1feb-445d-a03a-bbf9f4f1b791_1928x1234.png"
+  caption="来源：[SemiAnalysis](https://openai.com/signals/enterprise-data/)"
+/>
+
+过去，业界大多用固定序列长度的 prefill 和 decode 工作负载来衡量性能，但这种衡量方式并不准确。真实世界是多轮、长上下文、高 prefill 复用的，还伴随子智能体（sub agent）突发、KVCache offload 和大量工具调用。因此，我们的目标是为整个行业建立一套正确衡量 AI 硬件与软件性能的方法。
+
+[我们投入了超过 300 万美元构建这个数据集。今天，我们把一切都开源。](https://inferencex.semianalysis.com/)InferenceXv3 在既有的“固定序列长度”场景（8k1k、1k1k、1k8k）之外，新增了 AgentX 这一贴近现实的场景。它用 agentic 编码流量替代了此前 8k 输入、1k 输出 token 的单轮流量，让基准测试场景更接近真实。
+
+完整测试矩阵运行在约 2MW 的持续运行算力上，覆盖 1000 多颗芯片和大量 SKU，包括 MI355X、GB300 NVL72、GB200 NVL72、B300、B200、MI325、MI300X、H200 和 RTX Pro 服务器。Rubin 将于本月晚些时候到位，TPU 和 MI455X UALoE72 会在今年内加入。[如果你觉得我们的免费开源工作有价值，欢迎点个 star。](https://github.com/SemiAnalysisAI/InferenceX)
+
+很高兴看到 NVIDIA 和 AMD 在 agentic 工作负载上都拿出了出色的性能。NVIDIA 在很多前沿模型上表现非常好，AMD 在部分前沿模型的特定对比中同样表现不俗。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/8a065d23-de1d-4e2f-a3ae-7af724409c9d_1350x653.png"
+  caption="来源：[SemiAnalysis GitHub](https://github.com/SemiAnalysisAI/InferenceX)"
+/>
+
+AgentX 头几个月产出的最有价值的东西，并不是首批测试结果，而是这个基准测试已经带来的巨大行业影响。**超过 70 个上游 PR** 以 AgentX 作为北极星基准代理，围绕真实生产 agentic 工作负载在 vLLM、SGLang、TensorRT-LLM、ATOM、AITER、Dynamo、LMCache 和 Mooncake 上做优化。这些优化中的大部分都可以迁移到生产流量上。本文后面会逐项深入分析。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/193e605e-579b-4ed2-9637-5bfdbee5cf71_1286x1515.png"
+  caption="来源：SemiAnalysis"
+/>
+
+开源是 InferenceX 的核心原则，因此我们开放的部分比多数把“开源”挂在嘴边的项目都要多：开放的前端、通过易用 REST API 提供服务的公开数据库（**已有多家一线 AI 实验室的容量规划团队在使用**）、公开的 GitHub Actions CI 溯源信息、[运行日志](https://inferencex.semianalysis.com/inference/agentic/440193?i_seq=agentic-traces&i_xmode=interactivity&view=logs)，以及每一个数据点的精度验证。关键在于，我们的测试配置主要跟随 [recipes.vllm.ai](http://recipes.vllm.ai) 和 [SGLang cookbook](https://docs.sglang.io/cookbook/intro) 使用上游镜像，衡量的是客户实际能拿到的性能，而不是针对 benchmark 特调过的镜像。
+
+三到四周后，我们会发布一篇 AgentX 更新文章，涵盖针对 agentic 工作负载的进一步优化，以及 AMD 和 NVIDIA 的最新性能结果。需要理解的是，agentic 工作负载的画像变化很快，InferenceX 会持续快速跟进，测试真正有代表性的工作负载。
+
+InferenceX 100% 坚持开源，这离不开开源伙伴的贡献与支持。我们要感谢以下为 AgentX 1.0 发布做出巨大贡献的人：
+
+- **Inferact/vLLM**：Roger Wang、Yifan Qiao、Simon Mo、Jeff Ma 等多位贡献者
+- **RedHat/llm-d**：Michael Goin、Robert Shaw、Tyler Michael Smith
+- **RadixArk/SGLang**：Baizhou Zhang、Yuwei An、Mingyi Lu 等多位贡献者
+- **LMCache/TensorMesh**：Samuel Shen
+- **Weka**：Callan Fox、Val Bercovici
+- **MoonCake 维护者：** Teng Ma、Xu Wenjie、Ke Yang
+- **AMD**：Thomas Wang、HaiShaw、Andy Luo、Seungrok Jung、Chun Fang、Parth Panchal、Bill He、Theresa Shan、Hongxia、Fangzhou、Gilbert Lei、Yanfei Wang、Duyi Wang、Peng Sun、Lingpeng Jin、Simon Danielsson、Xiaohu Guo、Haichen Zhang、Chang Liu、Doug Lehr、Poovaiah Palangappa，以及 AMD 上海研发中心的众多同事
+- **NVIDIA**：Xin Li、Anthony Casagrande、Kedar Potdar、Ankur Singh、Ishan Dhanani、Nick Comly、NVIDIA 上海 TensorRT-LLM 团队等多位工程师
+- **Anthropic 团队，感谢他们及时修复了多个 bug，让 AgentX 得以实现**
+- **GitHub：** Austen Stone，帮助提升了 AgentX 所依赖的 GitHub Actions 的可靠性
+- 以及其他众多贡献者
+
+[此外，我们也感谢所有支持 InferenceX 开源计划的伙伴，包括 Meta、Microsoft、Oracle、OpenAI、MiniMax、月之暗面 Kimi、阿里巴巴 Qwen 和智谱 GLM。](https://inferencex.semianalysis.com/quotes)
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e97525a9-216a-4624-91c6-373ee7a3850e_1123x437.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/quotes)"
+/>
+
+## agentic 工作负载简介
+
+从宏观上看，agentic 工作负载有四个特征：
+
+1. **多轮：** 一个会话包含大量用户 / 助手交互（几十到上百轮），而聊天机器人场景只有寥寥数轮。多轮、长上下文、高 prefill 复用，并伴随子智能体突发和大量工具调用。
+2. **长上下文：** 系统提示词、工具定义，加上轮次本身很多，上下文会迅速累积。
+3. **高前缀复用：** 会话是线性推进的，第 n-1 轮的输出通常会拼接进第 n 轮，因此大部分上下文可以从 KV cache 读取而不必重算（具体取决于可用于存放 KV 张量的存储容量）。随着 n 增大，已缓存输入与未缓存输入之比通常趋近于 1。
+4. **子智能体突发：** 一个会话会启动多个带全新上下文的短生命周期子智能体，形成突发式的 KVCache 访问模式。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/72b64a02-8a1f-4a19-929c-85c0ed4803cd_2048x909.png"
+  caption="来源：DeepSeek、SemiAnalysis"
+/>
+
+考虑到上述特征，对这类工作负载做基准测试和现有的固定序列长度基准测试有本质区别：agentic 推理本质上是一个*系统问题*。由于前缀复用率极高，KV 张量必须在节点 / rank 之间高效传输（NIXL、MORI-IO、Mooncake）。此外，不同会话应当根据对应前缀所在的位置被路由到不同节点 / rank，以最大化缓存命中率（LLM-d、Dynamo、vLLM/SGLang router）。长上下文会话会挤占 KV cache 的 HBM 容量，需要把 KV 张量 offload 到不同层级的内存（DRAM、SSD），而这个过程本身也必须高效（Mooncake Store、LMCache、vLLM Simple Offloading、SGLang HiCache）。
+
+相比之下，固定序列长度的单轮工作负载不涉及前缀复用，推理性能基本反映的是芯片 / kernel 的基础性能。这并不是说 InferenceX 上大量的固定序列长度数据不重要。恰恰相反，剥离 agentic 服务的复杂性之后，底层推理优化的进展会显得格外清晰，这些数据也为 AgentX 结果提供了重要基线。
+
+为了让 AgentX 工作负载尽可能贴近真实，我们采集了 [393 条 SemiAnalysis 内部匿名 Claude Code trace](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126) 作为初始语料用于回放。为在保留原始前缀复用模式的同时对内容做匿名化，我们采用了与 [Qwen-Bailian 数据集](https://github.com/alibaba-edu/qwen-bailian-usagetraces-anon)类似的方法，后者是最早的生产 trace 语料之一。随后我们用 [AIPerf](https://github.com/ai-dynamo/aiperf) 在不同并发客户端数量下，按原始请求时序重建这些 trace。为让 AgentX 数据集成为可能，我们与 Anthropic 合作推动了两项 Claude Code 功能上线，在此感谢 Anthropic 团队的帮助。
+
+1. [https://github.com/anthropics/claude-code/issues/49207](https://github.com/anthropics/claude-code/issues/49207)
+2. [https://github.com/anthropics/claude-code/issues/66761](https://github.com/anthropics/claude-code/issues/66761)
+
+这段简介应该足以让读者理解下一节的测试结果。后文我们会更深入地介绍测试方法、回放框架和数据集。
+
+## agentic 编码推理性能
+
+在评估编码推理性能时，OpenAI、Anthropic、xAI 等前沿实验室主要关注三点：每美元性能与交互性（TPOT）的关系、TTFT（首 token 延迟），以及端到端任务完成情况。每兆瓦性能同样重要，因为地面数据中心电力是关键约束（钱是社会构建物，而且看起来实验室们有的是钱；电力在当下则是实打实地难搞）。[我们的数据中心模型给出了逐季度的电力需求与供给爬坡估算。](https://semianalysis.com/datacenter-industry-model/)
+
+本节我们梳理各前沿模型上 agentic 性能的整体规律。我们强烈建议读者把它当作一份导览，[自己去研究这些结果](https://inferencex.semianalysis.com/)。**所有数据都是开源的，社区完全可以就当下真实推理性能的现状得出自己的结论。**
+
+## DeepSeek V4 Pro 0813
+
+DeepSeek V4 Pro 0813 是一款来自中国、极受欢迎的前沿开放权重模型，总参数约 1.6 万亿，激活参数 490 亿。
+
+下图为截至 8 月 21 日、按总拥有成本（TCO）归一化后各 SKU 全部提交结果中的最佳性能。
+
+所有 DeepSeek v4 运行中*全部*请求的 ISL/OSL 分布如下：ISL p50=88k、p90=272k、p95=404k、p99=675k；OSL p50=413、p90=2.2k、p95=3.7k、p99=8.6k。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/d8337140-b620-476a-abd3-59d6cc447d30_2048x1256.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/b5aee8b4-d11a-48d8-823c-1ca53412e689_2048x1253.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&i_seq=agentic-traces&g_runid=32433563482&i_xmode=interactivity&i_label=0&i_e2e_xmetric=)"
+/>
+
+总体而言，**同时关注每用户每秒 token 数**（TPS，也就是交互性）**和 TTFT 很重要，因为这两者往往此消彼长**。例如上图中，某些 SKU 在还算不错的交互性下拿到了很高的吞吐量，但 TTFT 严重恶化。多高的 p90 TTFT 算“可接受”，很大程度上取决于应用场景。对多数服务 agentic 工作负载的生产系统而言，p90 TTFT 通常在 200–5,000ms 之间。超过 5–10s 就已经在“在线推理”的边界上了。曲线上超高吞吐的那一段仍有实际用途——在延迟不重要、追求系统利用率峰值的场景里（批处理、*超*长时间运行的 agent 等）。
+
+单节点性能方面，MI355X 的开源方案（vLLM）落后于厂商自有的 ATOM（AMD 版的 TensorRT-LLM）。我们认为 AMD 用 ATOM 快速推进性能前沿是好事，但仍建议他们把这些改进更优先地上游到 vLLM。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/d96697d9-bf24-4093-a7db-16d27a52aa08_2048x1200.png"
+  caption="来源：InferenceX"
+/>
+
+AMD 的分布式推理（DI）团队在过去半年里于 8k1k 场景上取得了很大进展，但要让 DI 成为真实工作负载下可用的方案，还有一段路要走。从每 GPU 吞吐量与交互性的关系看，1xDEP8+1xDEP8 的 disagg 配置只在高吞吐区间取得了小幅提升，在低延迟区间反而更差。
+
+更糟的是，中高交互性配置下的吞吐量提升，被 p90 TTFT 的大幅飙升所抵消。原因之一是并发超过 64 时启用了 SGLang 的 `--enable-prefill-delayer` 参数，[它会推迟 prefill 准入，让 DP rank 攒出更满的 batch](https://github.com/sgl-project/sglang/blob/v0.5.17/python/sglang/srt/server_args.py#L3180-L3194)（最多等 30 个 forward pass）。此外，这些点还把 chunked prefill 大小从 8,192 提高到了 65,536。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/81475383-510e-4823-9373-751b8a3722e3_2048x1202.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/6254047b-173b-4b1b-9f43-2057eb720b05_2048x1200.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=mi355x_atom%2Cmi355x_mori-sglang&i_vendor=AMD&i_linelabel=0&i_metric=y_tpPerGpu)"
+/>
+
+在端到端延迟上，ATOM MI355X 胜过 B200 vLLM（但并未胜过 B300 或 B200 SGLang）。问题在于，除了阿里巴巴内部一个小型广告业务部门之外，中国和西方的多数 AI 实验室都不愿在生产中使用 ATOM——它缺失的功能太多。阿里巴巴主力的 Qwen 大模型团队在生产中并不使用 ATOM。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e8d915c0-c962-4277-b62c-a1b08b8e7179_1924x1262.png"
+  caption="来源：InferenceX"
+/>
+
+在 2026 年 8 月 21 日之前，AMD 实力很强的 MI355X SGLang 开发团队在端到端（e2e）性能的每美元表现上已经追平了 B200 vLLM。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/360a051c-6117-4860-a704-b50198f01521_1978x1246.png"
+  caption="来源：InferenceX"
+/>
+
+不过，B300 vLLM 和 B200 SGLang 仍然领先 AMD 的 MI355X。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9ec4708c-148e-42ee-a7c5-3bf4d9cbdfaa_1938x1246.png"
+  caption="来源：InferenceX"
+/>
+
+2026 年 8 月 21 日之后，得益于 Inferact 和 NVIDIA 在 vLLM 上的优化，NVIDIA B200 的每美元性能已经超过 MI355X。这是一场胶着的竞赛，我们很期待接下来几周的性能优化，也会很快发布 AgentX 更新文章。
+
+[AMD 已经列出了他们的 DeepSeekv4 vLLM 优化清单，里面有不少值得期待的提升空间。](https://github.com/vllm-project/vllm/issues/52911)
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9302da60-5bc1-46e1-9392-74e912b5c200_2024x1236.png"
+  caption="来源：SemiAnalysis"
+/>
+
+再看 NVIDIA。他们最有竞争力的方案是 GB300 Dynamo TRTLLM 和 GB200 Dynamo vLLM，两者都依靠 PD disagg 在合理交互性下取得高吞吐。此外，GB300 配置采用了 wide-EP（DEP32）decode 实例，以在前沿曲线中段获得更高吞吐。
+
+注意，2xDEP8+1xDEP12 的 GB200 点与 3xDEP8+1xDEP16 的 GB300 点在 TPS 上比在 TTFT 上接近得多。同样，TTFT 总体上对工作负载的“突发性”更敏感。由于 GB300 那个点的总并发高得多，它会遇到更多子智能体流量，因而有更多冷 prefill。这一点在该点的 TTFT 图上可以看出来：
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/96f9c9f9-eb86-4f87-8332-c85f23040836_2048x916.png"
+  caption="来源：InferenceX"
+/>
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f79e53f5-c8ff-4ed0-a8a3-c05b9e0b4bab_2048x1191.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/512c6225-5816-493e-839d-e4f539e5b1c4_2048x1215.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_metric=y_tpPerGpu&i_xmode=interactivity&i_linelabel=0&i_best=0&i_active=gb200_dynamo-vllm%2Cgb300_dynamo-trt)"
+/>
+
+按 TCO 归一化后，B300 vLLM 与 B200 vLLM 的聚合式性能相当接近，主要差别在于 B300 凭借比 B200 多 50% 的 HBM 容量，能“挤”出额外的吞吐量。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/939b9197-321d-4869-ba13-f0b800343b7b_2048x1300.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=b200_vllm%2Cb300_vllm)"
+/>
+
+我们可以用 AgentX 新增的服务端指标可视化进一步观察这个差异。
+
+在 384 条并发 agentic trace 的负载下，B300 vLLM DEP8 配 3TB DRAM、使用 vLLM simple offloading，取得了 91% 的 HBM 缓存命中率，外加 1.36% 的 DRAM 缓存命中率。原因是该配置下 **HBM KV cache 的工作集规模**约为 4300 万 token，而任一时刻在途的 token 数只是略微超过这个量。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/b7ff2d61-bdab-4c2f-96a0-ec5a52469489_2048x884.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f86888c4-c1a7-47b0-a580-d1a74876cb9c_2048x898.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference/agentic/439522)"
+/>
+
+在 B200、并发 196（其余参数不变）的情况下，HBM 缓存命中率只有 73%，需要更多依赖 DRAM，offload 缓存命中率接近 20%。此时 HBM KV cache 工作集规模为 2200 万 token，大约是 B300 的一半。
+
+DRAM KV offload 通常实现为写穿（write-through）缓存，也就是写入 HBM 缓存的每个前缀同时也会写入 DRAM 缓存。因此，只有当可用于 offload 的 DRAM 显著大于（1.5–3 倍）HBM KV cache 容量时，它才最有效。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/61956823-2928-44ec-b00c-8bb5581b8a13_2048x892.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/70a7a7c7-7918-459c-a4e0-f7fc1edca816_2048x887.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference/agentic/439967)"
+/>
+
+H200 SGLang FP8 能在低并发下服务 DeepSeek v4，从每美元性能看甚至可以和 B200/MI355X SGLang 一较高下。但受限于 HBM 容量，它在高吞吐场景下无法与更新的 SKU 竞争。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/770b205f-12a1-4d34-81a7-fb235742d005_2048x1305.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=b200_sglang%2Ch200_dynamo-sglang%2Cmi355x_sglang)"
+/>
+
+此外，随着用户数增加，高并发下对 DRAM KV offload 的依赖会带来无法接受的延迟。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/91dfd9ca-d13d-43e9-a57e-3be0fcc4a5cf_2048x1315.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=h200_dynamo-sglang&i_optimal=0)"
+/>
+
+总体来看，MI355X 相对主要竞争对手 B200 和 B300 表现尚可。在曲线上低吞吐 / 低延迟那一段两者最为接近，因为那里只用到张量并行和较为基础的 kernel。AMD 需要优化 MI355X 上的 DEP kernel，才能在高吞吐场景下更有竞争力——尤其是它的 HBM 还是 B200 的 1.5 倍。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/ba6532b3-aa46-4477-be84-8979b1cce35d_2048x1297.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/29cfd340-d61b-46c4-8d87-e081d937b80c_2048x1303.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_vllm%2Cb300_vllm%2Cmi355x_atom%2Cmi355x_vllm&i_label=0)"
+/>
+
+## Kimi K3：2.8 万亿参数
+
+Kimi K3 是另一款来自中国的前沿开放权重模型，总参数 2.8 万亿，参数量级与 Claude 的 Mythos/Fable5 模型架构相当。我们把它作为开放权重的架构代理来使用。K3 大到单台 B200 服务器都装不下，必须用 wide EP / wide TP 或流水线并行才能放下全部权重。在 vLLM 上，投机解码 / DSpark 直到最近都完全无法与流水线并行组合，因此 B200 在 Kimi K3 上的表现非常糟糕，被 MI355X 压着打——因为 B200 在流水线并行下用不了投机解码。
+
+MI355X vLLM 在短上下文单轮工作负载上 day 0 即开箱可用，但在长上下文多轮工作负载上，MI355X 的 AITER 和 Triton kernel 在第一周就彻底崩了，上游 vLLM 在真实工作负载下对 MI355X 完全不可用。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a1440922-e1ea-4a74-ab62-f1263ae31ca0_2048x1473.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+Hopper 很难承接 Kimi K3 的 AgentX 工作负载：一方面 Kimi 模型体量巨大，另一方面 vLLM 维护者 / NVIDIA 一直没有把精力放在为 K3 优化 Hopper 上。Hopper（SM90）需要为 K3 定制调优 kernel，并针对高交互性服务调优 TP32/EP32 的 shape。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ce2b58fc-b7b7-46c2-ae05-baa14419d233_2048x1459.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+我们认为 AMD 用 ATOM 快速推进 K3 性能是件好事，但仍希望 AMD 进一步优先把这些改进上游到 vLLM。ATOM 目前是 AMD 性能最好的引擎，但对使用上游开源服务栈的客户来说，vLLM 才是更有参考价值的对比对象。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/80eb810f-a9ea-4f01-a41b-e19e3409ce74_2048x1459.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/73ded698-c86e-411b-8e7c-f9bb125487b7_2048x1459.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+在端到端延迟 40 到 60 秒的那段曲线上，MI355X ATOM 的每美元性能甚至超过了 GB300 NVL72 vLLM。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/27a3f80e-7959-4651-90eb-f498989908a3_1934x1284.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+## MiniMax M3
+
+在 MiniMax M3 432B 上，NVIDIA 完胜所有对手。AMD 的软件性能非常糟糕，长上下文下尤其如此，根源在于 AMD 工程管理层只激励团队针对短上下文单轮工作负载做调优，而忽视了长上下文多轮工作负载。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/c5396c3d-1823-4818-83de-d89cd8242451_2048x1286.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/5179bd8e-c147-46c8-aeb4-b347049b28a7_2048x1303.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft)"
+/>
+
+B300 TRT-LLM TP2 拿下了 M3 的桂冠。曲线上缺少 DP-attention 的数据点，因为在 M3 上这种做法并不划算——KV cache 局部性会变成路由约束，后文会进一步说明。以并发 40 的 GB200 为例，TP4/EP4/DPA 的吞吐量只有纯 TP4 的 0.60 倍，而 p90 TTFT 高出 3 倍以上。并发 32 时缓存命中率仅 28.8%，理论值则是 96.0%。每个 DP rank 独占缓存池的四分之一，一个 30 万 token 的会话一旦落到错误的 rank 上，就要全部重算。M3 的前沿曲线上没有出现任何带 EP 的 decode 配置，可能是因为并发还不足以让所有专家的负载均衡起来。
+
+在 MiniMax M3 上，按 TCO 归一化的吞吐量，B200/B300 也完全胜过对应的机架级方案。在 AgentX 上，机架级的优势不像以往那么明显，因为 Dynamo router 可能成为瓶颈——它的工作量随在线前缀的数量和长度增长。相关优化以及若干带来两位数百分比提升的修复会在后文讨论。此外，wideEP、wide DCP 和 wide TP 都还没有调优良好的 kernel。而 GB200/300 的 TCO 更高，在没有 wide EP / wide DCP 的情况下，每 TCO 性能就显得更差。
+
+话虽如此，我们也预期 NVIDIA 会在该 SKU 的机架级方案上做进一步优化，届时会在后续文章中重点介绍。
+
+尽管 P90 ISL 达到 317k，目前还没有任何提交使用上下文并行。M3 只有 4 个 KV head，即使 TP8 下 DCP 也最多只能到 2，而且 MSA indexer 还需要单独的上下文并行处理（vLLM 上已有一个 PR）。更多讨论见后文的上下文并行一节。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/aec34e63-34fd-4c42-87a1-97e5c14661ed_2048x1302.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2b40624a-223c-44db-a403-0be563b2af82_2048x1298.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_hc=1&i_active=mi355x_vllm)"
+/>
+
+NVIDIA 所有 Pareto 最优点在并发 20 以上都启用了 KV offload，而 AMD 的 Pareto 最优点没有一个使用 DRAM KV offload。在其他模型上，AMD 使用 KV offload 的比例也低于 NVIDIA。原因是 AMD vLLM 上 CPU KVCache offload 所需的 GPU 到 CPU 传输效率极低：`hipMemcpyBatchAsync` API 直到 ROCm 7.14 才提供。没有它，vLLM 原生的 Simple CPUOffloading 只能做串行 memcpy，无法把多次拷贝合并成更大的消息。
+
+另外值得一提的是，从吞吐量与 p90 交互性的关系看，vLLM 的表现与 TRT-LLM 非常接近；而在吞吐量与 p90 TTFT 的关系上，vLLM 表现更好。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/68f7ad68-3886-4980-9b01-390582728e47_2048x1292.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/729199d9-66c0-42b5-b1b8-f228503c1585_2048x1289.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_trt%2Cb200_vllm%2Cb300_trt%2Cb300_vllm&i_hc=1&i_label=0)"
+/>
+
+## Qwen3.5 397B
+
+Qwen3.5 397B 每隔几层就用 GatedDeltaNet 替代常规注意力。GatedDeltaNet 由 MIT / NVIDIA Research 提出，其状态存储需求在理论上是常数，而常规注意力是线性增长的，因此相比同等规模的稠密注意力模型，它的存储需求更低。与 Nemotron 那样惨败的端到端模型训练研究不同，NVIDIA Research 在 GDN、LatentMoE 这类被前沿模型采用的基础研究上做得非常出色。
+
+注意，该模型原生最大上下文长度为 262k token，因此我们使用[截断后的数据集](https://inferencex.semianalysis.com/agentx/cc-traces-weka-062126-256k)。这模拟的是在上下文上限较小的模型上频繁触顶、频繁压缩上下文的工作负载，也就是用户实际使用这类模型的方式。
+
+在 SGLang 对 SGLang 的对比中，Qwen3.5 397B 是 NVIDIA 的强势领地：90 tok/s/user 时性能领先 20 倍以上。在 Qwen3.5 SGLang 上，AMD 目前完全没有竞争力。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/c92da639-dcff-4e39-b9b5-49742677e93c_2048x1289.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/519a7cd0-ac51-4a5c-9c45-ba7c41f6d058_2048x1300.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_hc=1)"
+/>
+
+我们再次看到 NVIDIA 以牺牲 TTFT 为代价过度优化交互性，TRT-LLM 尤其明显。上图中所有 NVIDIA SGLang 提交的 p90 TTFT 都远低于 TRT-LLM。
+
+在 Qwen3.5 上，B300 FP4 的每美元性能是 H100 的 12 倍。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/cc2cf6c7-aa58-4a76-a0a2-ee0457d9243c_2048x1298.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=interactivity&i_hc=1&i_best=0&i_prec=fp4%2Cfp8&i_active=b300_sglang%2Ch100_sglang%2Ch200_sglang)"
+/>
+
+## GLM 5.3
+
+GLM 5.3 在 GLM5.2 744B 基础上做了额外的后训练，是一款前沿级别的模型。
+
+在开源 SGLang 的性能上，这又是一个 NVIDIA 在真实 agentic 推理性能上胜过 AMD 的模型。在 150 tok/s/user 的 p90 交互性下，NVIDIA 的成本效率最高领先 5 倍。以 AMD 软件当前的状态，在 150 tok/s/user 时 NVIDIA 的性能优势大到这种程度：即便竞品芯片硬件白送（服务商当然仍要承担数据中心托管、电力和其他运营成本），用 NVIDIA 的单 token 成本依然更低。
+
+我们期待 AMD 在几周后的 AgentX 更新文章中带来性能优化，那篇文章还会包含其他一些非常有意思的结果。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f85bdb45-792c-4330-9745-674b38b03425_2048x1289.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/4a21d140-b588-4b33-941d-7c8090ac42b5_2048x1286.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=GLM-5.2&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_prec=fp4&i_active=b200_sglang%2Cb300_sglang%2Cgb200_dynamo-sglang%2Cgb300_dynamo-trt%2Cmi355x_atom%2Cmi355x_sglang)"
+/>
+
+看 ATOM 的话，在 p90 E2E 归一化交互性的部分区间内，AMD 的每美元性能优于 GB300 NVL72 SGLang，甚至优于 TRTLLM。这一结果值得肯定，也再次期待 AMD 把这些优化移植到 SGLang。同时我们也期待 NVIDIA 在接下来几周里尽快优化 GB300 NVL72。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9337ed5b-a320-49f7-b31b-cc4b5751e0ff_2048x1290.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+这里插入介绍一个我们称之为 **E2E 归一化交互性（E2E Normalized Interactivity）** 的实验性指标。它的目标是在同时考虑 TTFT 和 TPS 的前提下，衡量用户实际感受到的响应速度，定义为 OSL/E2EL。把 E2EL 等于 TTFT 加上 OSL 乘以 TPOT 代入（实际上只解码 OSL - 1 个 token），可得下式。
+
+它本质上等于交互性（1/TPOT 那部分）再加上一项与 TTFT 成正比的惩罚。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9e472a49-d228-4d8d-876a-6bc0d59f9d97_746x594.png"
+  caption="来源：SemiAnalysis"
+/>
+
+需要说明的是，这个指标是实验性的，并不完美。例如它对高 TTFT 的惩罚很重，也无法体现 PD 分离等优化的全部细节。AgentX v1.0 的所有提交仍然分别针对常规交互性和 TTFT 做优化。我们会继续打磨能反映现代 agentic 推理各种细节的新北极星指标。
+
+## AgentX 的行业影响：面向 agentic 工作负载的优化
+
+AgentX 头几个月最有影响力的成果不是产出开源数据集，而是 **AgentX 合作伙伴以 AgentX 为北极星、为优化真实 agentic 工作负载而提交的 50 多个上游 PR** 所带来的行业影响。AgentX 的真实 agentic 流量不只考验原始的 prefill 与 decode kernel，还会检验从 KV cache 生命周期、混合注意力缓存正确性、CPU KV offload、传输进度、路由亲和性，到增量分词、请求序列化和调度器记账在内的完整端到端 token 生成链路。这些环节对每一个生产 agentic 部署都很重要。
+
+这只是我们既有使命的延续：帮助生态加速改进，推动软件以极快的速度进步。一个很好的例子是 SemiAnalysis 与 AMD 软件研发团队多年的合作——我们持续提供反馈和建议，帮助他们现代化软件开发方法。这不仅带来了许多加速 AMD 进展的改变，也在推动 AMD 开源方案在 agentic 工作负载上向一流体验靠拢。
+
+## 分布式推理生态简介
+
+如前所述，agentic 推理本质上是全系统问题，而不只是芯片 / kernel 层面的问题。此外，当大型*分布式系统*要处理数十万条 agentic 请求时，请求调度和 KV cache 管理就不再是小事，会实打实地影响性能。例如子智能体会产生突发式的 KVCache 访问模式，若不加优化，就会错误地把主智能体的缓存挤出去。
+
+下图从宏观上展示了这一技术栈。最上层是 router（有时也叫“frontend”），负责把请求分发给不同 worker。例如当服务端启用数据并行注意力时，每个 DP rank 都有各自的 KV cache。为了避免打乱其中任何一个缓存，请求会按不同策略路由，例如[一致性哈希](https://github.com/vllm-project/router/blob/main/src/policies/consistent_hash.rs)，让同一会话 / 子智能体的请求按唯一 ID 路由到同一处。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/58ac4a3a-cc05-4ff6-9d1e-b28dc0d0f760_1814x2048.png"
+  caption="来源：SemiAnalysis"
+/>
+
+对大多数路由策略来说，各家 router 的实现差异不大。有些是独立组件，例如 [vLLM router](https://github.com/vllm-project/router) 和 [llm-d router](https://github.com/llm-d/llm-d-router)；有些则集成在引擎里，例如 [SGLang model gateway](https://github.com/sgl-project/sglang/tree/main/sgl-model-gateway) 和 [ATOM Mesh](https://github.com/ROCm/ATOM/tree/main/atom/mesh)。
+
+请求路由完成后，会交给 vLLM、SGLang 等推理引擎的调度器处理。引擎负责实际执行推理并通过 API 返回结果。此外，每个引擎都提供接口，把引擎内部的 KV cache 与外部 KV cache 管理器连接起来，从而形成一个“可插拔”的生态：不同的 KV cache 管理器可以与多种推理引擎集成。
+
+当前 AgentX 结果所使用的一种简单部署方式是[在同一节点上把 Mooncake](https://pypi.org/project/mooncake-transfer-engine/) 与 vLLM 一起运行。每个 vLLM worker 内嵌一个 Mooncake Store 客户端，并贡献一部分主机 DRAM 加入外部 KV cache 池。vLLM 通过 MooncakeStoreConnector 接口连接这个池，把可复用的 KV block 载入 GPU 显存，并把新计算出的 block 存回主机内存。Mooncake Store 负责管理外部缓存，包括放置与淘汰；Mooncake Transfer Engine 则负责 GPU 与 CPU 内存之间数据的实际搬运。
+
+不同的 KV cache 管理器可能使用不同的传输引擎，在内存层级之间或机器之间（例如 prefill 与 decode worker 之间）实际搬运字节。以 Mooncake Store 为例，它使用 [Mooncake Transfer Engine](https://github.com/kvcache-ai/Mooncake/tree/main/mooncake-transfer-engine) 在 GPU 显存、主机 DRAM 和远端节点之间搬运 KV block。
+
+一套部署可以一边用 Mooncake Store 把可复用的 KV block offload 到主机 DRAM，一边用 NIXL 把请求专属的 KV 从 prefill GPU 直接传到 decode GPU。Mooncake TE 负责 Mooncake Store 这条路径的搬运，NIXL 则通过 UCX 和（在支持的平台上）GPUDirect RDMA 负责独立的 prefill-decode 路径。也就是说，同一个推理引擎内部可以并存多条 KV 管理与传输路径。
+
+这套生态由许多独立组件构成，包括推理引擎、router、KV cache 管理器、数据传输库和集群控制器。NVIDIA Dynamo、llm-d、AMD Infera 这类平台把其中若干组件“打包”成完整的软件发行版，发布相互兼容的容器镜像、connector、部署清单和编排逻辑，让这些组件可以作为一个系统来部署和运维。最终交付的产品通常是一组协同工作的容器，而不是单体服务（例如 Dynamo、llm-d 和 Infera 通常部署在 k8s 上，用于协调大型分布式系统）。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/af5c4776-221b-43b4-b2fd-d46e46f1ab38_1772x960.png"
+  caption="来源：RedHat / llm-d"
+/>
+
+## 上下文并行
+
+长上下文让某些并行技术真正有了用武之地，而固定 8k 的 prompt 很难体现出来——8k 下没什么可切分的，TTFT 本来就很短。此外，TP 和 DP attention 这类并行策略在更长的上下文下并不理想：TP 会让完整 KV 在每个 rank 上都复制一份；DP attention 虽然共享 KV，但在长上下文下容易卡住，因为长上下文工作负载的上下文长度方差也更大。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/1e2801db-53ff-4029-8b55-01b3c1a7037d_1632x1140.png"
+  caption="来源：SemiAnalysis"
+/>
+
+上下文并行是一种把 query token 切分到多个 GPU 上的并行技术，分为两种形式：用于 prefill 的 PCP 和用于 decode 的 DCP。PCP 中每个 rank 处理自己的 query 分块（KV 以环形方式传递），由于 prefill 通常是计算受限的，这样可以并行化 FLOPs，使 prefill 更快，也不会让某个 rank 单独扛下超大 prompt 的 prefill 尖峰。DCP 中每个 rank 扫描自己那一份 KV 分片，随后以 flash-decode 的方式合并部分注意力结果。由于 decode 是显存带宽受限的，并行读取 KV 可以提高 tok/s。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/a9544ced-f418-44e0-8910-0cbc63fa26ec_2048x768.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/512c3a96-2ab4-452f-be6a-4fa60900575c_2048x651.png"
+  caption="来源：SemiAnalysis"
+/>
+
+这项并行技术部分源自 NVIDIA Research。NVIDIA Research 在这类基础研究上很强，但在端到端训练研究上就相当难堪了——他们的 Nemotron3 Ultra 模型表现糟糕，如今连体量小得多的 Qwen3.8 27B 都能大幅胜过它。DCP/PCP 构成了 CUDA 护城河的一部分，因为 AMD 的 DCP/PCP 实现尚未优化。在 [vLLM 支持矩阵](https://docs.vllm.ai/en/latest/design/attention_backends/#standard-attention-mha-mqa-gqa-backends)里，AMD 的每一个后端都是不支持状态。
+
+接下来几节提到的部分改动就聚焦在 DCP/PCP 上。
+
+## vLLM 的 agentic 优化
+
+我们与来自 Inferact、Red Hat、NVIDIA 和 AMD 的 vLLM 维护者一起，把 AgentX 的真实回放器当作北极星，相关修复最终都进入了上游，其中大部分优化都可以很好地迁移到生产环境。举几个例子：
+
+vLLM 改进了混合注意力的前缀缓存，使短生命周期的滑动窗口分配不会挤掉有价值的长上下文检查点。[选择性保留](https://github.com/vllm-project/vllm/pull/43447)保住了稀疏回放边界，在 14 条并发请求、上下文最长 100 万 token 的条件下，前缀缓存命中率超过 95%。同样的可达性策略[也应用到了 Mooncake](https://github.com/vllm-project/vllm/pull/44774)，并[移除了不可达的滑动窗口查找](https://github.com/vllm-project/vllm/pull/45444)。更早的后续工作还[停止 offload 那些永远不会被复用的滑动窗口 block](https://github.com/vllm-project/vllm/pull/42258)，并[把投机解码的 lookahead block 保留在被保留的前缀中](https://github.com/vllm-project/vllm/pull/44082)。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/df1a505d-b8c3-45f9-b03c-a8798fba54c0_2048x632.png"
+  caption="来源：SemiAnalysis 与 vLLM 中面向 agentic 工作负载的 GitHub PR"
+/>
+
+高并发 agent 的 agentic 工作负载需要 offload。得益于前面提到的 AgentX 的推动，vLLM 上已经有工作流在推进：让 CPU KV offload 支持混合模型，而不只是统一的全注意力模型。这个区别之所以重要，是因为统一模型每个 token 只有一种 KV 布局，connector 用单一 block 几何结构就能描述要保存什么；而混合模型同时携带多个 cache group，每组的形状和生命周期都不同，假设单一布局的 connector 无法表达某个 block 属于哪一组。结果就是，最需要 offload 的长会话模型反而用不上它。通用的 [SimpleCPU connector](https://github.com/vllm-project/vllm/pull/37160) 先落地，随后[在 ROCm 上启用](https://github.com/vllm-project/vllm/pull/40549)，再[扩展到 DeepSeek-V4 的混合注意力](https://github.com/vllm-project/vllm/pull/42296)：相比前缀放不下 HBM 后重新计算，输出吞吐量提高 81.7%，平均端到端延迟降低 46.6%。
+
+Mooncake 也获得了[同等的混合内存分配支持](https://github.com/vllm-project/vllm/pull/42828)。同样的布局问题在分离式服务中再次出现：一个[待合入的改动](https://github.com/vllm-project/vllm/pull/51052)会在 1P1D 的 prefill/decode 拆分下，通过 MoRI-IO 把 Kimi-K3 的 conv+ssm 循环状态与注意力 KV 一起传输；没有它，decode 侧会从未初始化的循环状态开始。循环状态槽位复用既有的 remote-block-ids 通道，因此分离式路由器无需任何模型专属改动。该路径已在 MI355X 上端到端验证：每条链路 TP8、跨节点 RDMA，两侧均开启 DSpark 投机解码。
+
+在对真实工作负载做性能剖析时，vLLM 维护者发现 offload 期间的开销转移到了 store 路径上——写得太多、太频繁。现在有三处修复解决了这个问题：
+
+[当相同的传输已在进行中时跳过 store](https://github.com/vllm-project/vllm/pull/41289)，这样共享同一前缀的并发会话只需付出一次代价，而不是每个会话各付一次。store 现在[只覆盖新生成的 KV 区间](https://github.com/vllm-project/vllm/pull/46412)，因此延续历史的会话只写增量，而不是每轮重写整个前缀。最后，store [不再依赖相同 block 是否仍驻留在 HBM](https://github.com/vllm-project/vllm/pull/46906)，因此已排入队列的工作不会因为底下发生淘汰而被丢弃。
+
+load 路径则单独做了调优，因为查找发生在每次调度决策时，而不只是在真正搬运数据时。[把调度路径上的查找改为异步](https://github.com/vllm-project/vllm/pull/45659)可以让 connector 离开每步的关键路径，这样一步调度不必等待 CPU 侧的缓存查询就能接纳新工作。随后，[紧凑的零拷贝查找 key](https://github.com/vllm-project/vllm/pull/45969)、[接收端并行加载](https://github.com/vllm-project/vllm/pull/45971)和[预构建的 Mooncake key 字符串](https://github.com/vllm-project/vllm/pull/46188)进一步消除了剩余的 CPU 与传输开销。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2c2cfea4-a217-47ad-934b-dbd226bb248c_2048x837.png"
+  caption="来源：SemiAnalysis 与 vLLM 中面向 agentic 工作负载的 GitHub PR"
+/>
+
+长生命周期的混合状态还带来了固定形状请求几乎遇不到的正确性与记账问题。vLLM 现在[按混合 cache group 分别发出 cache 事件](https://github.com/vllm-project/vllm/pull/44103)，[正确计算分布式上下文 store 的 stride](https://github.com/vllm-project/vllm/pull/45371)，并[在分布式上下文与 prefill 下正确计算查找前缀](https://github.com/vllm-project/vllm/pull/46855)。相关的[上下文并行记账改动](https://github.com/vllm-project/vllm/pull/45340)让缓存归属与分片后的 token 区间对齐。投机状态现在也能[在合并后的 Mooncake group 之间传递](https://github.com/vllm-project/vllm/pull/49069)，并[贯穿 SimpleCPU 协调器](https://github.com/vllm-project/vllm/pull/49071)，避免重复轮次的缓存悄悄丢失 EAGLE 状态。
+
+从 ROCm 侧优化 vLLM 的 agentic 工作负载来看，工作继续下沉到缓存层之下，那里剩余的开销是按层而不是按请求计算的。当前缀既能存活又能及时到达之后，剩下的就是 decode 步骤本身；而一个会话里要跑上千次的 decode 步骤，会为每一次可以避免的拷贝、每一个不匹配的 kernel 付出代价。
+
+有三个尚未合入的改动针对这一层：
+
+- 一个 Kimi-K3 的改动[把 KDA decode 结果直接写入层输出缓冲区](https://github.com/vllm-project/vllm/pull/51183)，为每个 KDA 层省下一次设备内拷贝；单看收益很小，但它会在每个解码 token 的每一层上重复发生。
+
+- 第二个改动[选用 AITER 的稀疏 MLA decode kernel](https://github.com/vllm-project/vllm/pull/51714) 替代通用路径，AgentX 输出吞吐量提高 5.22%，token 间延迟显著降低。
+- 第三个改动很好地说明了测量形态有多重要。配套改动[把 full-graph 注意力投影路由到调优过的 AITER GEMM](https://github.com/vllm-project/vllm/pull/51713)，在低并发的固定序列上取得 2.3% 提升。这说明 kernel 层面的改动可以在统一 shape 上拿到干净的收益，却在 agentic trace 上被缓存与调度带来的波动淹没。一个[待合入的改动](https://github.com/vllm-project/vllm/pull/52882)干脆把这种 shape 敏感性变成了分发依据：它用 gfx950 上的 AITER / native 混合路径取代 DeepSeek V4 C4A selector 的 ROCm top-k 瓶颈，短、中上下文走 AITER，长上下文走 graph-safe 的调优 native 回退路径。据报告，selector 的端到端加速为 1.21x 到 1.76x，在 84 种 shape 的矩阵上 decode kernel 的几何平均加速为 1.2x 到 2.9x。
+
+## SGLang 的 agentic 优化
+
+AgentX 团队与来自 RadixArk、Meta、NVIDIA 和 AMD 的 SGLang 维护者密切合作，推动了 SGLang 上 agentic 工作负载的优化，使生产推理性能大幅提升。下面详细介绍这些优化。
+
+先从分配器一侧说起：SGLang 的滑动窗口工作解决的是与 vLLM 保留策略相同的冲突。窗口页和前缀页取自同一个池子，而窗口是更贪婪的一方——它不断翻转，前缀却静止不动，于是在压力之下，短暂的分配会把持久的分配挤出去。
+
+有三处设计改进从不同角度攻击这个问题。其一是[在页离开窗口时主动释放](https://github.com/sgl-project/sglang/pull/26907)，而不是等淘汰压力来找它们，让失效的窗口状态不再争抢自己已经用不到的页。其二是[把 compute lock 限制在单个窗口内](https://github.com/sgl-project/sglang/pull/27210)，限定一条在途请求最多能同时钉住多少池容量。其三是[清除失去作用后仍然残留的 full-KV 条目](https://github.com/sgl-project/sglang/pull/29369)。
+
+不过主动释放有一个与分叉相关的盲区：从共享前缀分叉出来的请求可能仍持有可复用的 full-KV，而分叉点处的窗口状态已被释放，结果为了缺失的那一半廉价数据，整个前缀都要重算。[进行中的工作](https://github.com/sgl-project/sglang/pull/34565)会在这些分叉点保留 SWA 状态，让分叉继承窗口而不是重建它。
+
+与之并行的 [ROCm ring-cache 修复](https://github.com/sgl-project/sglang/pull/30339)属于正确性而非容量问题：环形缓冲区在设计上就会复用槽位，而复用一个旧内容仍被引用的槽位，产生的是错误输出，而不是变慢的输出。这些在单个 8k prompt 上都看不出来——窗口不会绕过前缀，池子也不会有争用。但在多轮混合注意力会话中，这些改动决定了昂贵的全注意力历史在下一轮是否还在。
+
+HiCache 是 SGLang 内置的一流 offload 机制。它面对与 vLLM connector 相同的混合模型问题，解法是一种不对称处理：[offload 全注意力缓存，回读时再重建那一小段滑动窗口尾部](https://github.com/sgl-project/sglang/pull/29417)。只有昂贵的那一半值得跨总线搬运，廉价的那一半重建比取回更快。在 AMD 上，[分阶段写回](https://github.com/sgl-project/sglang/pull/28534)让这种搬运不会在进行中阻塞引擎。剩下的缺口是循环状态，因为它无法像窗口尾部那样从相邻 token 重建；[FlashInfer GDN 检查点](https://github.com/sgl-project/sglang/pull/29735)让它得以参与前缀复用，在 92.4% 缓存命中率下把吞吐量从 47,771 提升到 53,004 tok/s/GPU。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/078ad449-7657-4468-9ff9-3480a7999d89_2048x720.png"
+  caption="来源：SemiAnalysis"
+/>
+
+另有两处改动针对变长流量对 kernel 流水线的影响。AgentX 的真实会话往往以持续变化的上下文长度到达，生产流量也是类似形态。一个按长度做特化的运行时，几乎会为看到的每个请求重新编译一份 kernel。SGLang 维护者的解法是[把上下文长度作为运行时标量传入](https://github.com/sgl-project/sglang/pull/30255)，而不是把它固化进某一次编译；这让 AgentX 并发 384 的输出吞吐量提升 26.75%、平均 TTFT 改善 36.25%，收益来自消除编译，而不是把计算做得更快。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/3a807d67-383b-4aa9-93af-e8f3a68313f4_855x634.png"
+  caption="[来源：GitHub](https://github.com/sgl-project/sglang/pull/30255)"
+/>
+
+同样的思路下，[移除每步一次的 device-to-host 序列长度同步](https://github.com/sgl-project/sglang/pull/30365)消除了一个 decode 气泡——它之所以存在，只是因为 host 想知道一个 device 早就知道的长度。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/b29e003e-db43-45d4-93be-bac0be04d869_1209x646.png"
+  caption="[来源：GitHub](https://github.com/sgl-project/sglang/pull/30365)"
+/>
+
+变长同样会在注意力 kernel 内部造成损耗。在 GB300 上，混合上下文的 decode batch 要付“长尾税”：一次对照 profile 把 9.8 ms decode 步骤差异中的 8.4 ms 归因于注意力，最长的那些请求拖长了持久化 kernel 的共享 wave。[进行中的工作](https://github.com/sgl-project/sglang/pull/34888)把 TRTLLM MHA 的 decode batch 按 KV 长度分组，让短请求不再等待最长的那些请求。
+
+decode 也可能在更上一层被饿死——不是在 kernel 里，而是在调度器上。在 DP attention 下，每个 rank 都要参与同一个 MoE 集合通信，但注意力工作是本地调度的；于是某个 rank 一旦被喂进一串 chunked-prefill 续段，就会不断赢下“prefill 优先”的决策，而同伴 rank 正在运行的 batch 只能干等——这是在 AgentX 运行中观察到的现象。[可配置的 decode 间隔](https://github.com/sgl-project/sglang/pull/35017)在 prefill 之后强制插入 decode 轮次；在 AgentX DSv4 Pro 上，输出吞吐量提升 141%，p99 token 间延迟下降 97.3%，代价是 TTFT 中位数从 36.5 秒升到 59 秒——用首 token 等待时间换取输出流的平滑度。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/47ce130f-13f2-4732-8431-17e26184b8c3_1846x806.png"
+  caption="来源：SemiAnalysis"
+/>
+
+当一条请求不带任何可复用历史时（例如子智能体刚启动），随便哪个 worker 都可以，唯一值得考虑的就是负载均衡。但当请求带着数 MB 的已缓存前缀时，把它发给一个并不持有该前缀的空闲 worker 就是昂贵的选择，router 必须知道状态究竟在哪里。SGLang 为此加入了 [DP 缓存亲和性](https://github.com/sgl-project/sglang/pull/26091)，让会话粘在持有其缓存的 rank 上。在这个 PR 中，还实现了 [DP 感知的 prefill 与 decode 路由](https://github.com/sgl-project/sglang/pull/26245)，让分离式部署的两侧做出一致的决策；以及[把缓存均衡作为路由信号](https://github.com/sgl-project/sglang/pull/26293)，避免亲和性退化成单个热点 worker。router 只能依据它被告知的信息行动，因此混合缓存事件也变得[对 radix cache 可见](https://github.com/sgl-project/sglang/pull/26387)、[对滑动窗口可见](https://github.com/sgl-project/sglang/pull/26579)。
+
+投机解码得到了特别关注，因为 MTP 为每条请求引入了第二份、更小的状态，它必须挺过主缓存所经历的一切。SGLang [修复了分离式服务中的 draft window 传输](https://github.com/sgl-project/sglang/pull/30461)，让这份状态完整跨越 prefill 到 decode 的边界；[为高并发在线解码加入 overlap 调度](https://github.com/sgl-project/sglang/pull/30497)；[移除了一处无实际作用的 EAGLE 重归一化](https://github.com/sgl-project/sglang/pull/31294)；并[避免 EAGLE prefill 期间的 host 同步](https://github.com/sgl-project/sglang/pull/33662)。尚未合入的[资源租约调度工作](https://github.com/sgl-project/sglang/pull/32042)和[数据并行 graph metadata 修复](https://github.com/sgl-project/sglang/pull/32196)延续同一目标：当请求可能被撤回再恢复、而不是一路跑到结束时，让 overlap 仍然安全。
+
+在 prefill 与 decode 拓扑异构的 agentic 工作负载下，需要前缀感知的暂存，而这正是前缀缓存与分离式服务相互冲突的地方。当两侧的分片方式不一致时，KV 无法作为一整段连续流拷贝过去，必须按传输网格切分，再按 decode 侧期望的偏移重新组装。前缀命中会让这件事更难而不是更容易，因为 prefill worker 此时只发送未缓存的剩余部分，而 decode 侧仍然期待一份完整且位置正确的缓存。[在 staging buffer 中支持 radix cache](https://github.com/sgl-project/sglang/pull/30545) 会按该网格切分已缓存的发送内容，并散布到正确的 decode 偏移上。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/88c2d49d-2612-4334-90c2-8a76ab55fd60_1418x1004.png"
+  caption="来源：SemiAnalysis"
+/>
+
+然而这也带来了正确性问题。一项 127,500 token 的共享前缀测试，从 128 根针中只答对 2 根变成了 128 根全对——说明缓存此前一直悄悄落在错误的位置，而这种情况在吞吐量基准里只会被记成一个又快又自信的错误答案。在 AgentX 对比中，每用户输出吞吐量中位数还提高了 9.6%，而每 GPU 总吞吐量几乎不变。传输本身也带着无用负载：decode 侧的 PREBUILT batch 从不进入模型前向，但每条传过来的 prompt 仍会被展平并拷进一个 CUDA 输入张量，而第一次 decode 步骤反正会用中继元数据重建它。[去掉这次无用的 prompt 传输](https://github.com/sgl-project/sglang/pull/35070)是单项收益最大的改动，在 AgentX GB300 上带来每用户输出吞吐量 +18.0%、每 GPU decode 吞吐量 +12.7%。一个[后续改动](https://github.com/sgl-project/sglang/pull/35071)把 prefill DP-rank 的引导查询移出 decode 调度器的关键路径，让原本在消费结果时同步付出的一次 HTTP 往返得以重叠，在同一部署上再带来 +1.36% 的每用户输出吞吐量。同一条缝隙上还有进行中的工作：[UMBP 中的多池 DeepSeek-V4 支持](https://github.com/sgl-project/sglang/pull/30762)、[通过 MoRI 承载的统一 KV HiSparse 状态](https://github.com/sgl-project/sglang/pull/32368)，以及[在 decode 无可见内容即结束时保留 prefill 侧持有的 token](https://github.com/sgl-project/sglang/pull/34216)。HiSparse 这项工作应当被理解为长上下文的容量与正确性使能，而不是高并发下的吞吐量收益——目前它还不是。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ed24f1b9-b591-4674-bc94-06394962d03f_2048x544.png"
+  caption="来源：SemiAnalysis"
+/>
+
+## TensorRT-LLM 的 agentic 优化
+
+接下来介绍 TensorRT-LLM 近期的几项优化。首先看 TRTLLM 针对重复对话轮次的前端优化，这项开销完全源自工作负载是多轮的。每一轮对话都会把整段历史再加一点新内容重新发过来，而朴素实现会把这一整段重新分词。按每 KB 计算分词很便宜，但当同样的 10 万 token 每轮都要重新分词一次时，代价就不可接受了。
+
+显而易见的修法——只对新增后缀分词——其实是错的，而且很容易被忽略，因为字节对编码（BPE）并不是位置无关的。token 可能跨越拼接点发生合并，因此在边界处切分文本、再把两段 token 序列拼起来，可能得到与整段分词不同的序列，从而与前缀缓存所依据的序列悄然分叉。
+
+TRTLLM 实现了[边界感知的增量分词](https://github.com/NVIDIA/TensorRT-LLM/pull/17462)：先找出渲染文本的公共前缀，回退一个完整 token 以便重算任何跨拼接点的合并，然后只对变化的后缀分词。在 Qwen3.5 的 AgentX trace 上，它在全部 1,087 次转换中都与完整分词结果一致——这个正确性结论是实测而非假设——并把平均处理时间从 185.1 ms 降到 11.3 ms。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/30a51cf2-e628-4c20-a4a5-5978234ca713_952x541.png"
+  caption="[来源：GitHub](https://github.com/NVIDIA/TensorRT-LLM/pull/17462)"
+/>
+
+固定的 8k1k 请求没有可复用的上一轮渲染结果，因此这些开销根本不会出现。与之相关地，[chat template 渲染被移入输入处理线程池](https://github.com/NVIDIA/TensorRT-LLM/pull/16231)，长 template 不再让主请求循环排在它后面串行执行。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/00b459b8-9c4c-4c8c-8659-07e14bf5699f_2048x784.png"
+  caption="来源：SemiAnalysis"
+/>
+
+MiniMax-M3 相关的工作聚焦在分离式 KV 搬运上，问题出在粒度。当 prefill 与 decode 对 head 布局理解不一致时，一条逻辑请求的 KV 就不再是几块大的连续区域，而变成数千个小的跨步片段，每一片都要生成自己的传输描述符。搬运的字节数没变，爆炸的是每描述符的开销，而且恰恰在最要紧的长 prompt 上爆得最厉害。[修正多池映射并引入分块 NIXL 中转路径](https://github.com/NVIDIA/TensorRT-LLM/pull/17518)，通过一块有界可复用的 arena 把这些碎片合并，用一次额外的暂存拷贝换取数量级更少的描述符。其 AgentX 诊断显示：并发 5 时请求关键路径上的 KV p99 从 26.74 秒降到 125 ms，并发 40 时从 10.15 秒降到 288 ms。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/52bd05ab-914f-4dc5-9f16-a0ea1b495383_959x501.png"
+  caption="来源：[GitHub](https://github.com/NVIDIA/TensorRT-LLM/pull/17518)"
+/>
+
+[非阻塞的上下文传输轮询](https://github.com/NVIDIA/TensorRT-LLM/pull/17428)保护了同一条路径，即使调度停滞也会回收已完成的传输，从而打破一个反馈环：已完成的 KV block 一直被钉住，阻止新请求准入。这些停滞本身也有[可以消除的原因](https://github.com/NVIDIA/TensorRT-LLM/pull/16734)：在 GB300 分离式 context worker 上，避免 DeepSeek-V4 上下文稀疏注意力元数据中的隐式 device 标量同步，消除了每步 18 次四字节 device 读取，而每一次都会强制触发 cudaStreamSynchronize。修复方式是把 host 侧计数以普通 Python int 传递，这样执行线程不再在一步的大部分时间里持有 GIL，让 KV 传输线程和响应线程排队等待。
+
+TensorRT-LLM 还把不规则的长上下文工作迁到了更高效的执行路径上。[MiniMax-M3 的 context graph producer](https://github.com/NVIDIA/TensorRT-LLM/pull/17473) 捕获稳定的稀疏 producer，同时让依赖请求的注意力保持 eager 执行，在其 AgentX 测试中每用户输出吞吐量提升 12.58%。一个进行中的[原生 KV 事件生成改动](https://github.com/NVIDIA/TensorRT-LLM/pull/16876)减少了 KV 感知路由路径上的分配与转换开销。
+
+AgentX 还暴露出只有在规模和时长都够大时才会出现的 kernel 选择与调度器生命周期问题。其中两个与选哪个 kernel 有关。MiniMax-M3 [为 MXFP8 自动调优加入了 CuTeDSL 候选](https://github.com/NVIDIA/TensorRT-LLM/pull/17316)，扩大候选集，在低并发聚合点上把每 GPU 输出吞吐量提高约 7% 到 10%。反方向上，TensorRT-LLM 在[禁用有缺陷的 split-K MoE tactic](https://github.com/NVIDIA/TensorRT-LLM/pull/17105) 之前，七次 AgentX 运行中有五次崩溃，禁用后七次对照运行零崩溃。一个又快又错的 tactic 比单纯慢的 tactic 更糟，而自动调优器只要它还在候选池里就会兴高采烈地选中它。选择并不是 kernel 出错的唯一方式：MiniMax-M3 针对短 query 的旧稀疏注意力路径，可能把一个非连续、head-major 的 block index 视图交给 SM100 kernel，而后者按连续内存解读，于是选中错误的 KV 页，产生错误或非有限的输出。[在 q_len ≤ 32 时遵循 block index 的 stride](https://github.com/NVIDIA/TensorRT-LLM/pull/17285)修复了索引问题，既不需要物化张量，也不需要新增 kernel；此后在 GB300 上完成的五组完整 AgentX 对照运行零服务错误、无非有限值标记。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f48290ac-3361-4b0c-a085-ff84dd3dce51_2048x1347.png"
+  caption="来源：[TRT Github](https://github.com/NVIDIA/TensorRT-LLM/pull/17105)"
+/>
+
+另外两个是生命周期 bug，这类问题是长时间运行（而非大规模运行）的典型故障。[序列槽位余量与一致的按槽位索引缓冲区尺寸](https://github.com/NVIDIA/TensorRT-LLM/pull/16279)处理的是一个瞬时重叠：一条即将完成的请求和一条新准入的请求同时需要槽位。持续有请求到达和离开的流量会不断撞上这个窗口，而固定 batch 则永远撞不上。之后的[注意力数据并行 dummy request 修复](https://github.com/NVIDIA/TensorRT-LLM/pull/17278)让九个 Qwen3.5 分离式测试格存活下来，而此前多数测试格都在几分钟内失败——这正是“能跑完基准测试的配置”与“能撑住一整个会话的配置”之间的差别。
+
+两个进行中的传输改动针对超长的分离式 prompt，二者合起来展示了一个修复如何制造出下一个瓶颈。默认安排下，decode worker 必须等到整段 prompt 完成 prefill 并传输完毕才能开始，于是两个昂贵阶段串行执行，尽管第一个阶段本来就是增量出结果的。[流水线化的 KV 传输](https://github.com/NVIDIA/TensorRT-LLM/pull/15727)会在每个 prefill 分块完成时就开始发送，让传输与 prefill 计算重叠，只有最后一个分块留在关键路径上。
+
+这个改动让分块处理变得频繁，于是暴露出原本只做一次的工作。它的后续改动[每次只取当前分块对应的 block ID](https://github.com/NVIDIA/TensorRT-LLM/pull/17526)，而不是每次都取整段 prompt 的 block 列表。对一段 128,000 token、按 1,024 token 分块的 prompt 来说，这就是“构建一次 4,096 项列表”与“为每个层组重建 128 次”的差别。一项随 prompt 总长增长的按分块开销，其增长形态会把流水线刚换来的收益吃掉。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e8eeb92e-1aad-4a97-b3c1-28409005809e_2048x769.png"
+  caption="来源：SemiAnalysis"
+/>
+
+## AMD ATOM 的 agentic 优化
+
+AMD 的 ATOM 引擎最初只面向单轮工作负载设计，而不是真实世界的 agentic 多轮生产工作负载，因此要让它较好地支持长上下文多轮场景，核心引擎和 kernel 都需要大量改动。相比 vLLM/SGLang 目前的水平，ATOM 在支持 agentic 工作负载上还有很长的路要走。AgentX 正被用作 ATOM 这轮重构的真实北极星目标。我们要讲的第一项 ATOM 优化，是对 DeepSeek-V4 分页滑动窗口注意力巧妙使用稀疏检查点保留。[已合入的实现](https://github.com/ROCm/ATOM/pull/1640)保留了选定的窗口尾部，让分叉和回放请求能够从有意义的边界处恢复。它的测量把两种效应分得很清楚：在并发 48 的同一条 AgentX trace 上，实际前缀命中率从 5.6% 升到 96.45%，滑动窗口关口处的损失从 91.35% 降到 0.16%。第二个数字是第一个数字背后的机制——十次前缀匹配里有九次是找到了却因为缺少窗口尾部而被丢弃，也就是说缓存并没有失配，而是被否决了。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/5c880335-e1e3-44d9-a123-8e316c9c3808_822x555.png"
+  caption="来源：GitHub"
+/>
+
+在这一切能被测量之前，还有两处更早的缓存管理器修复必须先落地，它们都是“缓存自报健康、实则什么也没做”的典型案例。其一[阻止空闲池命中破坏共享缓存条目](https://github.com/ROCm/ATOM/pull/902)；其二是[延迟输出的修复](https://github.com/ROCm/ATOM/pull/939)，恢复了默认调度模式下的前缀哈希，让重复的长 prompt 从零缓存 token 变为复用每一个完整前缀 block。另一处改动让前缀命中的 prefill [继续走优化过的 sink attention kernel](https://github.com/ROCm/ATOM/pull/1345)，而不是回退到通用路径，这样缓存命中不会悄悄吃掉它本该节省的一部分开销。
+
+混合模型还带有循环状态或压缩状态，它与普通 KV 有一个决定性区别：无法从周围的 token 重建。窗口尾部可以从相邻上下文重算，但循环状态是此前全部内容累积的结果，一旦丢弃，唯一的办法就是重放整个序列。ATOM [为这类按请求的状态引入了内容寻址的检查点生命周期](https://github.com/ROCm/ATOM/pull/1771)，让已生成的轮次留下可复用的恢复点，而无需为它们单独预留受保护的缓存。在一次测试中，一条请求复用了 512 个已生成 token，只计算了两个 token 的后缀。
+
+调优细节和功能本身同样重要。无条件发布检查点会让零命中流量损失 17.5% 的吞吐量——这是所有一去不回的会话为帮助那些会回来的会话所付出的代价。按 token 间隔来安排检查点避免了这一惩罚，固定 1k1k 的吞吐量保持在测量噪声范围内，而这正是关键的安全性质：一个面向 agentic 复用的功能，不应该让永远用不到它的工作负载付出代价。
+
+ATOM 与 AgentX 相关的 CPU 路径，要从“offload 到底值不值”的算术说起。[独立的 LMCache offload](https://github.com/ROCm/ATOM/pull/1318) 从 CPU 重新载入一段 32,000 token 的前缀约需 0.32 秒，而重算约需 2.5 秒，相差八倍。这说明在这种上下文长度下跨总线是值得的，但对短 prompt 就不成立。
+
+这条路径的其余部分关乎归属与索引位置，而不是带宽。ATOM 借鉴了 vLLM 的 [multi-connector](https://github.com/ROCm/ATOM/pull/1406) 设计，让 prefill worker 可以同时把 KV 发给远端 decode worker 并把同一前缀保存到 CPU，在两个消费方都完成之前不释放这些 block；同一批 block 有两个独立读者，这种情况在单轮场景中根本不会出现。
+
+[把恢复回来的 block 重新登记进 GPU 前缀索引](https://github.com/ROCm/ATOM/pull/1725)修复了一种更隐蔽的浪费：没有它，从 CPU 载入的前缀用过之后不会被登记为常驻，于是下一轮又要跨总线取一次同样的热前缀，为一份其实已经在 HBM 里的缓存反复付出传输代价。后续工作一并[修复了异步保存顺序、打包 KV 的几何结构、未对齐的交接以及远端请求记账](https://github.com/ROCm/ATOM/pull/1807)，在两轮、2,638 条请求的验证中消除了重载损坏。这个 bug 只有在同一批 block 被反复保存、淘汰、恢复很多次时才会显现。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a7bb7f74-e5a3-4da3-b954-89905e54c4da_2048x420.png"
+  caption="来源：SemiAnalysis"
+/>
+
+分布式路径的故事在另一个代码库里重演，而这个模式在 SGLang 和 Dynamo 中已经出现过：路由必须知道状态在哪里。ATOM 的 router 叫 ATOMesh，是 SGLang router 的一个删掉了大部分功能的分支。可惜的是，ATOMesh 恰恰需要 SGLang 的缓存感知路由能力，于是这项功能又不得不加回来。ATOM 随后获得了面向缓存感知 router 的 KV 生命周期事件，router 由此才能知道状态在哪里；此外还加入了多节点 prefill 与 decode 路由，以及会话粘性的数据并行路由。这个粘性策略是一个值得明说的双向折中：会话会回到持有其状态的健康 worker，但空闲的分配会过期，避免为已经离开的会话长期打破集群均衡。
+
+分离式部署接下来必须搬运模型实际保留的所有状态，而这未必只有一份统一缓存。DeepSeek-V4 需要[同时传输其 FP8 与 BF16 混合缓存布局的两个缓冲区](https://github.com/ROCm/ATOM/pull/1737)，EAGLE 分离式部署则要在传输目标缓存之外[搬运 draft 模型独立的 KV cache](https://github.com/ROCm/ATOM/pull/1331)——这正是 TensorRT-LLM 和 SGLang 各自都解决过的“第二份缓存”问题。[远端 KV 准入与背压](https://github.com/ROCm/ATOM/pull/1647)补上了闭环，防止 decode 侧接收超出其安全恢复能力的挂起传输，这是“接了做不完的活”在分离式场景下的形式。
+
+在 ATOM 上，[PCP 报告平均 TTFT 降低 35% 到 43%](https://github.com/ROCm/ATOM/pull/1220)，在 64,000 token 输入下总吞吐量最高提升约 49%——这种收益随输入长度增长，而不是随 batch 大小增长。要让它在实践中可用，就必须与会话依赖的其他一切组合起来，因此 DCP 被[做到与前缀缓存、chunked prefill 和 FP8 KV 兼容](https://github.com/ROCm/ATOM/pull/1701)，随后又[扩展支持 MTP](https://github.com/ROCm/ATOM/pull/1746)。一种无法与前缀缓存共存的并行方式，只是用一项长上下文收益换掉另一项。同样的并行度匮乏也存在于单块 GPU 内部：batch 为 1 的 MLA decode 没有 head 维度或 query 维度可以铺开，只剩 KV 扫描，而硬编码为 16 的 split 预算让这次扫描只用上了 gfx950 的 256 个 CU 中的 16 个。一个[尚未合入的改动](https://github.com/ROCm/ATOM/pull/1911)不再覆盖 kernel 自身推导出的 split 数，让 Aiter 按机器实际的 cluster 数量切分扫描。
+
+[分块流水线并行 prefill](https://github.com/ROCm/ATOM/pull/1552) 从内存一侧攻击同一问题，用逐层阶段的流式交接取代反复的张量并行集合通信。它在高负载下的 GLM-5.2 结果是本节中最完整的一组：输出吞吐量翻倍，首 token 延迟中位数从 28.6 秒降到 8.7 秒，每块 prefill GPU 能容纳的 KV block 数是原来的 3.68 倍。最后这个数字最值得先看，因为每块 prefill GPU 的容量决定了在部署撞上 HBM 悬崖之前，究竟能有多少条长会话同时在跑。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/8184b9e6-883b-446c-884b-226b5e9fcca1_2048x549.png"
+  caption="来源：SemiAnalysis"
+/>
+
+## ROCm AITER 的 agentic 优化
+
+ATOM / AMD vLLM / AMD SGLang 的长上下文执行依赖底层 AITER kernel 的配套支持，因为引擎层的并行策略只有在 kernel 能表达时才算数。[Prefill 上下文并行进程组](https://github.com/ROCm/aiter/pull/3728)提供了 prefill 上下文并行所需的额外 query 分片维度，同时把融合 kernel 的行索引扩宽到支持 131,000 token 以上的 prompt。[Decode 上下文并行](https://github.com/ROCm/aiter/pull/3267)（DCP）把 KV 分片到已有的张量并行 GPU 上，使更长的序列或更大的 batch 无需在每个 rank 上复制整份缓存即可放下。
+
+大容量缓存还暴露出一类短固定请求基本碰不到的故障：地址位宽。在单个缓存池跨过边界之前，32 位偏移完全够用；一旦越界，算术就会回绕，kernel 会在不报任何错误的情况下访问错误的行。AITER 为此加入了[大于 4 GB 时批量 prefill 的运行时 64 位分发](https://github.com/ROCm/aiter/pull/2893)、[大于 2 GB 时的 64 位 MLA 偏移](https://github.com/ROCm/aiter/pull/4474)，以及[DeepSeek-V4 统一缓存路径全面使用 64 位寻址](https://github.com/ROCm/aiter/pull/4680)，最后一项避免了在约 1.5 亿行的池中静默读写错误的行。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ce356c05-883b-4c63-a359-8cfd72dddb63_1408x1046.png"
+  caption="来源：SemiAnalysis"
+/>
+
+DeepSeek-V4 的 decode 还获得了[面向 64 head 与 128 head MTP 打包的持久化 MLA kernel](https://github.com/ROCm/aiter/pull/3459)。这两个 head 数正是普通解码和投机验证实际产生的形态，因此引擎在常见 shape 上有了专门的长上下文路径，而不是把它们当成为短上下文编写的 kernel 的附带变体。这与前面 vLLM AITER 稀疏 MLA 选择的道理相同：在长上下文下，通用路径不是一个小小的妥协，而是选错了 kernel。
+
+## Dynamo 的 agentic 工作负载优化
+
+NVIDIA 的相当一部分提交使用了 Dynamo 推理编排与 router 系统。Dynamo 的 AgentX 系列工作表明，一旦引擎 kernel 变快，分布式服务层就可能成为瓶颈。router 的工作量正比于在线前缀的数量和长度，而不是生成的 token 数，因此大量长、相互重叠、生命周期又长的会话给它带来的压力，是固定形状流量永远制造不出来的。第一组 PR 降低了每次路由决策的开销：[减少查找热路径上的工作](https://github.com/ai-dynamo/dynamo/pull/10540)、[消除冗余的后缀失效](https://github.com/ai-dynamo/dynamo/pull/10836)，最后是[批量化 KV 匹配、注册、归属和终止解引用](https://github.com/ai-dynamo/dynamo/pull/11095)，在并发 512 下报告输出吞吐量中位数提升 22.2%。批量化在这里奏效的原因和它在引擎里奏效的原因一样：每项开销已经盖过了项目本身。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/7af8faf5-300e-4ee5-bc0c-040b7630e923_1016x671.png"
+  caption="来源：GitHub"
+/>
+
+第二组 PR 改变了归属关系的表示方式，这是底下更难的问题。每个被缓存的 block 都要归属到依赖它的请求，既不能在仍被使用时释放，也不能在所有人用完后继续钉住。当数千条并发会话共享相互重叠的前缀时，记账本身就成了显著开销。Dynamo 从[共享 block 链](https://github.com/ai-dynamo/dynamo/pull/11503)转向[arena 级归属计数](https://github.com/ai-dynamo/dynamo/pull/11508)，最终走到[后端专属的请求租约](https://github.com/ai-dynamo/dynamo/pull/12329)，每一步都在放粗被追踪的粒度。租约设计让 vLLM 后端的 AgentX 回放时间缩短 23.7%、SGLang 缩短 22.0%，同时还降低了峰值内存——这说明问题出在此前的表示方式，而不是流量本身。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e3039950-98e0-4cd0-b7fe-929a120bd7d2_2048x576.png"
+  caption="来源：SemiAnalysis"
+/>
+
+进一步的 router profile 消除了同一形态的开销：某个周期性扫描或整体重算之所以过去可以接受，只是因为在线状态一直很小。[分桶式过期清理](https://github.com/ai-dynamo/dynamo/pull/10521)取代了正比于全部被追踪对象的扫描，让高频变动（churn）场景下的 AgentX 吞吐量提升 13.7%。[仅处理增量的后缀清理](https://github.com/ai-dynamo/dynamo/pull/10676)只处理变化的部分，在同样的时间窗口内吸收了约 28 倍的 store 与 remove 事件。[压缩 prompt 路径](https://github.com/ai-dynamo/dynamo/pull/11644)把前端 CPU 占用降低 35.3%，并显著改善了尾部首 token 延迟——这一点很重要，因为这类工作负载中的 prompt 很长且大量重复。过载状态现在也改为[增量跟踪](https://github.com/ai-dynamo/dynamo/pull/10645)，而不是每次重算。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/878dd41d-fc0f-407e-8421-ad05ad9e740c_1087x624.png" />
+
+有一处路由改动是有意为之的取舍，而非纯粹的收益。Dynamo 现在可以[在路由打分中计入活跃的 decode 请求](https://github.com/ai-dynamo/dynamo/pull/12158)，这样一个已经承担了长时间 decode 的 worker，看起来会比单看队列深度更“贵”。在所报告的调优点上，这改善了 AgentX 延迟中位数，代价是少量吞吐量——这种选择只有在请求会长时间占用 worker 时才会显现出来。一个[进行中的后续改动](https://github.com/ai-dynamo/dynamo/pull/13447)把这个取舍打包成新的 agentic router 预设，并在同一方向上走得更远：前缀重叠按 2 计分、prefill 负载按 4 缩放、活跃 decode 请求按 64 加权。在这个调优点上，这笔交易不再牺牲吞吐量：在一次 8xH200 的 AgentX 运行中，该预设相比默认代价函数把固定窗口内完成的输出吞吐量提高 8.26%，运行级 p95 首 token 延迟降低 43.1%、p95 token 间延迟降低 22.6%，并且多完成了一条完整轨迹。
+
+接下来优化的是请求平面，因为一条 agentic trace 不是只发一次请求、收一次响应。它会发出许多携带几乎相同 prompt 的相关请求，并把每个 token 作为独立帧流式返回，于是序列化和拷贝的开销要按轮、按 token 反复付出，而不是只付一次。改用 [MessagePack 请求负载](https://github.com/ai-dynamo/dynamo/pull/10437)在其 AgentX 测试中把吞吐量提高 8.1%、平均首 token 延迟降低 9.7%，[直接 Python 转码](https://github.com/ai-dynamo/dynamo/pull/11104)则把这条路径上的中间值树整个去掉了。
+
+随后的一系列改动都是去掉一次拷贝，而不是把某次拷贝做得更快：不再拷贝 [MessagePack 事件负载](https://github.com/ai-dynamo/dynamo/pull/11539)、不再拷贝[收到的 ZeroMQ 帧](https://github.com/ai-dynamo/dynamo/pull/11574)、不再为每个 token 支付完整的 [token 间延迟指标开销](https://github.com/ai-dynamo/dynamo/pull/11569)。[聊天流式热路径](https://github.com/ai-dynamo/dynamo/pull/10433)也因同样的理由被缩短。单看每一项都平平无奇，但乘上每条并发会话的每个流式 token，它们决定了前端每秒能撑住多少请求。
+
+高并发剖析随后找出了与数据搬运无关的开销。[静态日志过滤](https://github.com/ai-dynamo/dynamo/pull/11820)移除了一个共享的 span matcher 锁——这是争用问题而非数据量问题——把前端吞吐量从 932 提升到 1,133 请求/秒。[更简单的位置 radix 分桶](https://github.com/ai-dynamo/dynamo/pull/12161)在 32 worker 的运行中把 mocker 的峰值内存降低了 5.51 GiB。一个进行中的改动改为[每条响应刷新一次去分词指标](https://github.com/ai-dynamo/dynamo/pull/12999)，而不是每个流式分块都更新累计计数器，在对照诊断 profile 中把前端 CPU 时间大约减半。最后这一项是这类问题最清晰的例子：这套埋点单次调用很便宜，但按每 token 一次调用就是灾难。
+
+## LMCache 的 agentic 优化
+
+LMCache 是一个开源 KV cache 层，位于 vLLM 等推理引擎之下，按前缀哈希把可复用的 KV 分块存放在 CPU DRAM、本地 NVMe 和远端后端（Mooncake、Redis、S3）上。它可以替代 vLLM 原生的 offload connector。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/945da9bc-59ec-4084-9f3a-da58da04b3c5_2048x638.png"
+  caption="来源：SemiAnalysis"
+/>
+
+LMCache 的多进程路径针对 agentic 缓存搬运的量级与形态做了改造，起点是一个不是变慢、而是直接卡死的故障。当许多上下文超过 100,000 token 的请求各自在开始前就为整次加载预留 block 时，池子会被一群全在等待、谁也无法推进的请求耗尽。[分块的外部缓存加载](https://github.com/LMCache/LMCache/pull/3382)改为按分块预留，让加载可以交错进行并逐步排空。在并发 32 下，验证完成了 120 条请求，而旧路径在 28 条时就死锁；并发 48 时在 KV 池 98.5% 占满的情况下仍能继续运行。
+
+其余改动则减少搬运量以及运行时的干扰次数。只存储 DeepSeek-V4 混合分组中真正有用的部分，把[每 token 的存储量减少了近二十倍](https://github.com/LMCache/LMCache/pull/3635)；滑动窗口预取现在[只加载存活窗口](https://github.com/LMCache/LMCache/pull/3869)，而不是那些永远不会被读取的窗口状态——这与 vLLM 在 offload 上采用的可达性论证相同，只是换到存储一侧来看。[每个对象组一次原生传输调用](https://github.com/LMCache/LMCache/pull/3908)随后消除了暂存拷贝与 kernel 启动之间反复的 Python 锁交接，这类开销正比于分片数量而不是字节数。
+
+有两项 LMCache 改动与 AgentX 尤其相关，但目前仍未合入。[混合锁记账修复](https://github.com/LMCache/LMCache/pull/4524)阻止一条请求释放另一条请求在共享滑动窗口分块或循环状态分块上的读锁。要触发它需要三个条件同时成立：多条请求共享同一批分块、记账必须按分块而非按持有者进行、并且淘汰确实开始发生。带 DRAM offload 的长时间 Kimi-K3 运行同时满足了三者，产生了数万条告警、生成内容损坏，并在淘汰开始后最终导致 GPU 崩溃。只要不是长时间、共享且内存吃紧的运行，这个 bug 就一直潜伏着。
+
+另一条并行的 LMCache 工作线让上述一切在 AMD Instinct 硬件上成为可能。CacheBlend 的非前缀复用依赖仅支持 CUDA 的 flashinfer，因此[一个 Triton block-sparse 注意力后端](https://github.com/LMCache/LMCache/pull/3092)重新实现了它所需的三个 kernel：带 CSR 索引和 log-sum-exp 输出的 block-sparse 注意力、causal prefill，以及 log-sum-exp 输出混合。当检测到 ROCm 或缺少 flashinfer 时，会自动路由到这些实现。[ROCm Dockerfile](https://github.com/LMCache/LMCache/pull/3101) 对齐了 CUDA 的构建与轻量镜像。[AMD hipFile 后端](https://github.com/LMCache/LMCache/pull/3843)扩展了此前只能通过 NVIDIA cuFile 访问存储的 GDS L1 slab-file 层，通过 ctypes 绑定 ROCm 的 hipFile 并按 torch.version.hip 分发；cuFile 路径保持不变。
+
+最后剩下的缺口是分发。CUDA 用户装的是预编译 wheel，AMD 用户则要从源码构建。我们与 AMD 合作发布了[预编译的 gfx942 与 gfx950 wheel](https://github.com/LMCache/LMCache/pull/4273)来补上这一块。它可以装进上游镜像，并在 MI350X 上通过全部 56 项 KV 传输 kernel 测试；它发布在 GitHub release 而不是 PyPI 上，因此直接 pip install lmcache 装到的仍是 CUDA 构建。[一行的后续修复](https://github.com/LMCache/LMCache/pull/4363)把 bind mount 进来的仓库标记为 git 安全目录，这个问题只在 CI 中出现，因为容器以 root 身份运行在 runner 所属的 checkout 上，setup.py 里的版本探测会拒绝读取它。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/89c2f79c-a704-4144-a900-ddd69ce306dd_1112x667.png"
+  caption="来源：GitHub"
+/>
+
+[DCP 感知的 CPU offload](https://github.com/LMCache/LMCache/pull/3561) 解决了两个在长上下文下必须同时启用的功能之间的直接冲突。启用 decode 上下文并行后，每个 rank 只持有 KV 的一段 stride，因此任何单个 rank 能保存的内容都不构成可用前缀；修复方式是在保存前汇聚这些跨步分片，加载后再重新分发。没有它，启用上下文并行会悄悄让 CPU 缓存命中失效，而失效的恰恰是催生这两项功能的长前缀。其验证记录了超过 30,000 次 CPU 命中事件，单请求加载量达到数十万 token。
+
+## Mooncake 的 agentic 优化
+
+Mooncake 承载着月之暗面 Kimi 的生产流量以及多家实验室的生产流量，也是分离式 vLLM 与 SGLang 配置底层的传输引擎。直到不久前，Mooncake 的 AMD 支持既没有 RDMA 注册，也没有可直接安装的软件包。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/32c734aa-1f52-47af-a98e-334978b4e6fa_1778x1386.png"
+  caption="来源：SemiAnalysis"
+/>
+
+在 NVIDIA 上为 RDMA 注册 GPU 显存，要么用 nvidia-peermem 内核模块，要么导出 dmabuf 文件描述符。AMD 没有 nvidia-peermem 的对应物，因此 GPU-direct RDMA 完全无路可走，部署只能把 KV 经主机 DRAM 中转。[一个 HIP dmabuf 注册分支](https://github.com/kvcache-ai/Mooncake/pull/2225)加入了与现有 CUDA dmabuf 路径对应的实现，改为通过 ROCm 导出，并先解析出真实的分配基址——因为缓存型分配器会把张量打包在更大一块分配内部的某个偏移处。主机内存仍然直接注册。
+
+装不上的支持等于没有支持。Mooncake 发布过 CUDA 和 MUSA 的 wheel，但没有 ROCm 包，因此 AMD 用户只能在每个镜像里从源码构建引擎。[ROCm wheel、CI 与发布流程](https://github.com/kvcache-ai/Mooncake/pull/3184)把 mooncake-transfer-engine-rocm 一并发布到 PyPI。这项工作来自 AMD 工程师 Andy Luo，他在用 AgentX 试用 agentic 工作负载时注意到一个规律：在 ROCm 上从源码构建 MoonCake 并不是一流公民的做法。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/12cd94ca-2aa7-4e06-a4bd-3153982234c5_1220x680.png"
+  caption="来源：GitHub"
+/>
+
+传输引擎没有设备 kernel，也不依赖 torch，因此一个与架构无关的 wheel 就能同时覆盖 gfx942 和 gfx950；ROCm 运行时是在加载时绑定而非打包进去的，这意味着同一个 wheel 在上游 vLLM ROCm 镜像和 SGLang ROCm 镜像中都能原样使用。这一点是按完整交叉矩阵验证过的：MI300X 和 MI355X，各自在 vllm/vllm-openai-rocm 与 lmsysorg/sglang 下运行 master 二进制以及带数据校验的 HIP buffer 传输测试。该 PR 还加入了按 tag 触发的发布流程，覆盖 Python 3.10 到 3.13。一个进行中的后续改动[增加了自建的双节点 MI350X 外部 prefill 与 decode 层](https://github.com/kvcache-ai/Mooncake/pull/3338)，让 ROCm 分离式路径在真实硬件上得到验证，而不只是能编译通过。
+
+综合这些 PR，AMD 上的 AgentX 运行现在可以把传输引擎和 KV cache 层作为已发布产物直接装进原生上游镜像，并在 GPU 显存与网络 fabric 之间直接搬运 KV。
+
+## 其他优化
+
+上面这些改动针对的是长上下文带来的开销：必须存活的前缀、必须保持正确的混合缓存、必须跟上节奏的传输。但除此之外，还有大量 day-0 使能与正确性 bug，它们让请求失败的程度不亚于百万 token 会话。
+
+MiniMax-M3 检验了这些 ROCm 工作能否累积成 day-0 就绪能力，而 [Advancing AI 那篇文章](https://newsletter.semianalysis.com/p/can-amd-break-the-cuda-moat-amd-advancing)直接给出了对比：AMD 首个公开的分离式配置方案 MI355X FP4 于 1 月才进入 InferenceX，比 NVIDIA 晚了数月；而 M3 FP4 的分离式支持是 day 0 就位的。相比 DeepSeek-R1 时期要花几个月才追平，这已经是明显进步。day-0 路径上有三个 vLLM 修复，每一个都是正确性问题而非性能问题。
+
+首先被卡住的是分离式部署。NixlConnector 的握手断言 SPLIT 区域的 block_len 会随 prefill 与 decode 的 TP 比例缩放，但 block_len 实际跟随的是每 rank 的 KV head 数。M3 只有 4 个 KV head，因此 TP4 prefill 搭配 TP8 decode 时，两侧都被 GQA 限制为每 rank 一个 head，两个长度相等，而断言却要求相差两倍。握手被拒，KV 没有搬运，decode 只能全部重新生成，gsm8k 得分为 0。[改为按实际 head 比例校验](https://github.com/vllm-project/vllm/pull/45879)修复了这个问题。
+
+另外两个是平台差异。M3 的稀疏注意力后端在所有 E4M3 配置下都把以字节存放的 FP8 缓存读作 float8_e4m3fn，但 gfx942 的平台 dtype 是 e4m3fnuz，两种编码并不相同，于是 K 和 V 在被 kernel 消费之前就已经被改变。prefill 和 decode 的 wrapper 也没有把 FNUZ 类型纳入 FP8 检查。[改用平台 dtype 来解释缓存视图](https://github.com/vllm-project/vllm/pull/45720)同时修复了这两处。另外，M3 以 NVIDIA 和 AMD 两个独立模型文件发布，而只有 NVIDIA 那份实现了 EAGLE3 接口，因此在 ROCm 上启用投机解码会在引擎初始化时报“模型不支持”并中止。[把 AMD 模型补齐到同等实现](https://github.com/vllm-project/vllm/pull/45546)恢复了该能力，MI355X 的 gsm8k 分数与未启用 EAGLE3 的 MI355X 运行以及 B200 一致。
+
+上面的 TensorRT-LLM 一节涵盖了 M3 上与长上下文相关的工作：分离式 KV 传输中的描述符爆炸、context graph 捕获、稀疏 block stride、自动调优候选，以及必须从候选池中移除的有缺陷 split-K MoE tactic。
+
+本地的 AgentX 测试矩阵把会话感知或 KV 感知路由、长且长度多变的对话历史、MTP、混合注意力、聚合式与分离式服务，以及跨越 HBM 容量悬崖的并发扫描组合在一起。它既包含全 GPU 驻留的对比，也包含通过 vLLM SimpleCPU、Mooncake、LMCache 和 SGLang HiCache 的 CPU DRAM offload。正是这种组合激活了上述上游工作。旧的固定序列矩阵通常只创建一个 prompt、做一次 prefill、解码一段固定长度的续写，然后丢弃请求。因此它无法衡量跨轮次的缓存存活、重复分词、会话亲和性、缓存事件流量、offload 抖动、调度停滞期间的传输进度，以及长生命周期的归属记账。
+
+允许的优化策略把 CPU KV offload 视为可选项。厂商可以使用 vLLM connector、LMCache、SGLang HiCache、Mooncake、Dynamo KVBM 或其他 CPU DRAM connector，也可以在关闭 offload 后延迟与吞吐更优时选择关闭。NVMe offload 暂缓。CPU DRAM 必须按所用 GPU 比例配置，其中非标准化 DRAM 系统适用 3 TB 上限；标准化 DRAM 系统没有硬性上限，但同样遵循按比例配置的规则。目前本地生成器对所有 runner 一律应用 3 TB 上限，也就是说尚未实现标准化 DRAM 的例外规则。
+
+新增的优化面并不只是“注意力更长了”。它是一份不断增长的会话状态在保留、搬运、路由、重建和反复处理上的开销。AgentX 把这些开销放大到足以驱动 vLLM、SGLang、TensorRT-LLM、ATOM、AITER、Dynamo 和 LMCache 上的通用上游改动。我们直接检索 NIXL 和 Mooncake 未发现更多标记为 AgentX 的运行时 PR，因此它们的相关效果仍然通过上述引擎 connector 改动体现。
+
+## AgentX 测试方法深入解析
+
+AgentX 在开源的真实世界长上下文多轮 agentic trace 回放上是一次重大转变。我们采集了价值超过 300 万美元 token 的 trace，构成自有数据集，内容来自 Claude Code、OpenAI Codex 等真实流量。除数据集外，我们还建立了一套公平回放这些流量模式的完整方法，目标是在尽可能忠实于真实流量的同时，对 GPU 资源需求保持公平。
+
+下面我们会深入介绍 agentic trace 数据集、回放方法以及 agentic 行为本身。希望更好理解 agentic 工作负载整体形态、以及各类 harness 在底层如何编排请求的读者，值得读一读这部分。
+
+## 价值 300 万美元的 agentic 流量 trace 采集器
+
+在最初设计 AgentX 时，我们的北极星目标是让基准测试在 KV 工作负载形态和 KV 复用模式上尽可能真实。我们先尝试回放了一些现有数据集，例如 SWE-bench、Qwen-Bailian 以及 HuggingFace 上的一些 Claude Code trace。当时这些数据集都没有大量使用子智能体、1M 上下文、上下文压缩、动态工作流等近来定义 agentic trace 的特征。在 SemiAnalysis，团队里大多数人都是 AI 重度用户，会用 agent 做各种事情：写代码、分析师研究、Excel 建模、社交媒体运营等等。因此我们判断，最可行也最真实的 trace 只能自己采集。
+
+为了采集足够多的 trace，我们做了一个代理来拦截发往 Claude / Codex 的 HTTP 请求。愿意上传 trace 的用户只需把 Claude / Codex 配置中的 base URL 指向这个代理即可。截至撰稿时，我们已采集超过 8,000 个会话、340 万条请求、6,100 亿 token，合计支出超过 300 万美元。[我们为 AgentX v1.0 基准测试开源了其中具有代表性的一个子集。](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126)
+
+agentic harness 看起来复杂，但归根结底就是在编排一串 HTTP 请求。每条请求都包含系统指令、工具定义和累积对话历史的某种组合。随着会话推进，这段历史不断增长并被反复发回模型，从而形成 AgentX 想要复现的长上下文与高前缀复用。
+
+我们的代理会实时记录这些请求和响应，同时提取时间戳、conversation ID、subagent ID 等元数据 / HTTP 头，从而还原对话的*结构*（请求顺序、并发分支，以及每个会话大致的父子关系）。正是这些元数据，让我们能够按照原始 Anthropic API 服务端所看到的样子近似回放这些 trace。
+
+为保护员工隐私，回放数据集不包含任何原始 prompt、源代码、工具参数或工具返回结果。我们的做法是先对每条请求的内容分词，再按 64 个 token 分块，最后把每一块替换成一个会话内链式哈希。这样，相同的 prompt 前缀会产生相同的哈希前缀，同时不暴露内容（[这篇论文](https://arxiv.org/abs/2506.02634)对该策略有更多讨论）。回放时，这些哈希块可以替换成来自比如某个编码数据集的 token。由此我们保留了原始工作负载中大致的上下文增长曲线和会话 KV 复用模式。
+
+需要说明的是，这个过程必然不完美，主要原因是使用前沿模型厂商 API 时，末端 LLM 服务器实际看到的内容有很大一部分是隐藏的。例如 SOTA 模型的思考 / 推理内容如今在 HTTP 请求中是加密的，并被替换成确定性哈希，以阻止蒸馏攻击。不过看起来[这招的效果并不像大实验室预期的那么好……](https://arxiv.org/html/2608.09867v1)
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/767f2f58-fa9d-40d3-8711-751d243cae9a_500x487.png"
+  caption="来源：SemiAnalysis"
+/>
+
+此外，虽然我们能拿到用户 / 助手消息、系统提示词、工具使用等全部原始*内容*，但 API 厂商还会在服务端应用额外的、不透明的 chat template。我们也无法观察 Anthropic 的专有分词器，或服务端工具引入的上下文。图片和文档在传输表示与模型实际处理的 token 数之间也没有直接对应关系。我们使用确定性占位符以及按模型经验校准的填充，把重建后的 prompt 长度调整到最接近服务端实际看到的内容。
+
+由于信息不完整，我们无法*完全*按照 Anthropic / OpenAI 服务端*实际看到*的样子采集并回放 Claude Code / Codex 的 trace，但可以做到相当接近。下图展示了在所有请求长度和模型上，哈希 token 数（经我们近似 / 处理之后）与 API 厂商真实 token 计数的比值。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/7123b5de-9942-4298-8ccd-1b69b8a71df0_2048x1799.png"
+  caption="来源：SemiAnalysis"
+/>
+
+总结来说，由于信息不完整，要完全按原始服务端所见的方式采集并回放真实 Claude Code trace 并不容易。但我们掌握的信息足以采集并回放出与原始流量模式、时序、前缀缓存和 DAG 结构高度吻合的 trace。
+
+## AgentX 的 300 万美元数据集
+
+AgentX v1.0 使用的数据集可在 [HuggingFace](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126) 获取。它是上一节所述 8.3k 会话代理语料中的 393 会话子集。此外我们还做了一些后处理来清理异常，例如：
+
+- 移除 Claude Code 安全监控（auto mode）请求和标题生成请求，因为它们是 Claude Code 特有的，不一定能代表通用 agentic 流量
+- 移除重建后输入长度超过 990k token 的请求（这些是我们的近似方法高估了）
+- 去除重复请求（连接断开时，代理有时会收到完全相同的请求）
+
+此外，每个对话都被整理成 WEKA trace 格式，该格式由 Callan Fox 在其 [kv-cache-tester](https://github.com/callanjfox/kv-cache-tester) 项目中提出。我们选择这个格式，主要是因为我们与 Callan 密切合作开发了这套基准测试，觉得它用来存放按会话组织的 trace 很直观。总体而言 trace 格式的选择相当随意，我们的代理数据集也可以映射到其他格式，例如 [Mooncake](https://docs.nvidia.com/aiperf/benchmark-modes/trace-replay-with-mooncake-traces)。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/48045991-20d7-4685-863d-c9aa1d6d3142_1812x2048.png"
+  caption="来源：SemiAnalysis"
+/>
+
+应用这些处理后，得到如下数据集。注意各子图的 X 轴并不一致。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f22c9c7e-44ba-4ffd-bdc3-58fbbef8f30d_2048x2004.png"
+  caption="来源：SemiAnalysis"
+/>
+
+ISL/OSL 和轮次间延迟（在 agentic 工作中，这主要是工具调用耗时）的分布大致呈对数正态。ISL 中位数为 142k token，OSL 中位数为 444 token，轮次间延迟中位数为 3.84 秒，只有约 10% 的轮次间延迟超过 1 分钟，这部分很可能是 harness 在等待真人回应的空档。
+
+值得一提的是，这些请求分布会随所用 harness 而不同，因为各家注入的上下文数量 / 类型不一样（例如 Pi 以 harness 注入上下文极少著称，Claude Code 则相反）。此外 ISL/OSL 分布还取决于模型，不同模型的分词器产出的 token 数会更多或更少。不过考虑到全球相当大一部分 agentic 编码流量都经过 Claude Code，我们认为这份数据具有代表性。
+
+数据集中有 175 个会话至少包含一个子智能体（约占全部会话的 44%），共有 1,697 次子智能体 rollout，每个会话中位数为 4 次。子智能体的墙钟时长中位数（从第一条请求开始到最后一条请求结束）为 2.27 分钟，这一分布同样接近对数正态。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f4792809-087e-431b-acce-aecdd3855b3b_2048x1390.png"
+  caption="来源：SemiAnalysis"
+/>
+
+该数据集包含最长 1M 的上下文，用于测试较新的前沿开放权重模型。此外我们还提供[一个截断到 256k 上下文长度的数据集](https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k)，用于回放最大上下文长度不超过 256k 的模型。
+
+## agentic 流量 trace 回放器
+
+我们没有从零造回放方案，而是选择与 [AIPerf](https://github.com/ai-dynamo/aiperf) 合作。它是 NVIDIA 出品、与厂商无关的 HTTP 回放工具，已被 tenstorrent、AWS、AMD 等众多机构采用。虽然我们的目标是把 AgentX 的功能合入上游仓库，但仍维护一个独立的 [fork](https://github.com/SemiAnalysisAI/aiperf) 以保持更强的厂商中立性，从而能自主接纳更多第三方贡献。再次感谢 AIPerf 团队，特别是 Anthony Casagrande，为构建真实且有代表性的 agentic 基准测试所付出的努力。
+
+一个 agentic 会话天然可以用有向无环图（DAG）描述。每条请求是一个节点，一条边表示：边头部的请求必须等边尾部的请求完成后才能发出。每条边还带有一个延迟，表示前置条件满足后需要再等多久。
+
+最简单的会话是完全线性的：没有子智能体、没有并行请求，每条请求恰好依赖一个前驱。此时图退化成一条线，边所编码的只有轮次间延迟（也就是工具调用时间或“思考”时间），那是客户端本地的工作，不是模型的耗时。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/20662be0-c311-49e9-8e39-0973d8a1865d_2048x1169.png"
+  caption="来源：SemiAnalysis"
+/>
+
+在 agentic trace 中还会派生出子智能体。子智能体是一条拥有独立上下文的请求流，通常用来完成某个聚焦任务。多个子智能体可以并行执行以完成更多工作，某些情况下子智能体也可以与主智能体并行。主智能体随后等待一组子智能体完成，再把它们的输出并入自己的上下文（虽然并非*总是*如此，但这是最常见的模式）。
+
+正是这种行为把上面那条线性请求链变成了 DAG，其中某些请求依赖于另一些请求。当识别出属于同一个子智能体的一组请求时，AIPerf 会找到最近的一条主智能体前驱请求，把它定为“派生”请求；类似地，“汇合”请求则是该子智能体组持续时间结束之后的下一条主智能体请求。
+
+在下面的例子中，子智能体组只包含一个子智能体（001），它发出两条请求。第一条请求在主智能体的首条请求完成后立即发出；该请求完成后，2.2 秒的轮次间延迟代表工具调用的墙钟时间，随后发出子智能体的第二条请求。主智能体的第二条请求是子智能体 001 的汇合点，它在两个条件同时满足后才发出：距主智能体首次响应已过去至少 17 秒，*并且*子智能体 001 的第二条请求已经完成。
+
+一个小的局限是：HTTP 时间戳能反映时序，却未必能反映因果。在下面的例子中，如果子智能体 001 在主智能体请求 2 的记录起始时间之前 7 秒就结束了，我们无法判断这 7 秒是子智能体返回之后才开始的工作，还是原本就在并行进行的独立工作。因此 AIPerf 同时保留两个约束：请求 2 既要等待它记录的主路径延迟，也要等待子智能体 001 完成。这样能复现观察到的时序和工作负载拓扑，但无法复现 harness 内部隐藏的依赖关系。这些是我们希望在后续版本的 AgentX 中改进的地方。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/18adec21-15d9-40cd-af7d-0772da4a9af0_2048x1729.png"
+  caption="来源：SemiAnalysis"
+/>
+
+一条请求也可以同时派生出多个子智能体。在下面的例子中，子智能体 001 和 002 的派生父节点都是主智能体请求 1，并在主智能体请求 2 处汇合。
+
+AIPerf 还能识别“辅助（auxiliary）”请求，即与流中其他请求不共享上下文的一次性请求。它们从主智能体分叉出去，并且不会汇合回来。实际场景中，这类请求例如 Claude Code 的“总结本次会话”请求，与对话上下文无关；另一个很好的例子是 Claude Code 的 `/btw` 功能。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/3d5f01ad-1736-496b-91b4-31672002bf47_1920x2048.png"
+  caption="来源：SemiAnalysis"
+/>
+
+下面的例子把这些要素放在一起，也是实际数据集中会看到的 trace 片段类型。图中有五条并行流从同一条主智能体请求出发，按各自汇合到哪条主智能体请求被分成不同组。
+
+子智能体 001 和 002 分别在第 20 秒和第 23 秒结束，因此其后第一条在第 25 秒发起的主智能体请求就是它们的汇合点。子智能体 003 和 004 在同一个间隙中派生，但运行时间长得多，分别在第 46 秒和第 50 秒结束，因此它们汇合在第 52 秒。AIPerf 用（派生请求，汇合请求）这一组合作为子智能体的键，因此这四条流虽然都从同一节点出发，却被归并为两条分支。分支以其第一个成员命名，这也是汇合边被标为子智能体 001 和子智能体 003 的原因。
+
+这也是第一次出现主智能体与自己的子智能体重叠的情况：第 25 秒发出的请求，是在 003 和 004 仍在运行时发出的——主智能体只被与它汇合的那一组阻塞，而不是被所有在途的子智能体阻塞。
+
+辅助链会挂在它之前最近的那条主智能体请求上，这里是第 52 秒那条，而不是开启会话的那条。因此该节点同时做了两件事：接收第二组子智能体的汇合，并派生这条一次性请求。当然，辅助请求不会再汇合回来。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/d2f465f9-372c-4ee1-b42f-e79fa9fe352b_1677x2048.png"
+  caption="来源：SemiAnalysis"
+/>
+
+## AgentX 进一步解析
+
+### Pareto 前沿扫描
+
+在 AgentX 工作负载中，为了生成 Pareto 前沿，我们针对单套部署扫描并发 Claude Code 会话数。由于每个对话都带有真实的轮次间延迟和子智能体使用，得到的流量模式更尖锐、也更真实。下面这个例子是对运行 MiniMax M3 的 B200 TP4 vLLM 服务端回放 40 个并发客户端。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/058757ed-531f-4cfb-a1ad-a06b1499ecbd_1360x612.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference/agentic/439239)"
+/>
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/43184e50-b71a-4a48-a569-98d6906cb259_2048x1167.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-13&g_runid=31723588108&i_seq=agentic-traces&i_xmode=interactivity&g_model=MiniMax-M3&i_best=0&i_active=b200_vllm)"
+/>
+
+最后值得讨论的是：评估 agentic 工作负载时应该看哪些指标。我们认为交互性（TPS，每秒 token 数）和首 token 延迟（TTFT）依然重要，它们也是评估 SLO 的行业标准。查看 AgentX 结果时，*必须*把 TPS 和 TTFT 放在一起看，因为有些推理优化会以牺牲一者为代价改善另一者。
+
+我们正在定义一个能把 TPS 和 TTFT 有意义地结合起来的新指标。它还应当体现这样一个事实：在 agentic 工作负载中，人们往往更关心一个*任务*端到端多快完成，而不是收到 token 的速度或 TTFT 本身。
+
+最后要说明的是，端到端延迟在当前形式下参考价值有所下降，因为它与 OSL 直接成正比。因此 P90 端到端延迟会被输出序列长度最长的那 10% 尾部严重影响。它仍可用于整体比较不同配置的综合表现，但我们更建议结合 TPS 和 TTFT 一起看。
+
+### warmup、计时、确定性与对话复用
+
+AgentX 的目标是测量已经进入稳态的系统。对 agentic 工作负载而言，这意味着性能采集应当从部分上下文轨迹已经被缓存的状态开始。为模拟稳态，也不希望所有对话都从第 0 轮开始，否则会造成“[惊群](https://en.wikipedia.org/wiki/Thundering_herd_problem)”效应。
+
+warmup 分两个阶段。第一阶段，AIPerf 用固定的随机 seed 在每个对话 25% 到 75% 之间选取一个墙钟时间点。在该时间点上，它识别出所有活跃的请求流（包括主智能体和任何活跃的子智能体），并为每条流发送该时间点之前最近的一条请求。这些引导请求会重建该时刻的对话状态，并一起发出，AIPerf 会等它们全部完成后再继续。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/963dce62-c49b-4deb-afee-e8dd448526c5_2048x1322.png"
+  caption="来源：SemiAnalysis"
+/>
+
+第二阶段，每条回放通道再前进 10 条请求，让 KV cache 有更多机会成形。所有 warmup 请求都不带轮次间延迟，且最大输出长度为 1 个 token，从而大幅缩短 warmup 时间。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/85116f5f-be0d-405b-a5d9-6a9237b44cc3_2048x895.png" />
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/fef10e0d-261e-4978-8650-b10304ec04d0_2048x897.png"
+  caption="来源：SemiAnalysis"
+/>
+
+warmup 结束后开始性能采集，持续一小时，所有指标都严格在这段时间内统计。为保证可复现，AIPerf 接受一个 seed，确保每次运行采样的对话相同、对话起始点相同，且每个对话每次都用相同的合成内容重建。采集期间，我们对每条流设置 5 分钟的空闲上限，避免过长的轮次间空档在整个测试期间“占住”一条 worker 通道。这样做是为了能在 1 小时内完成测试。在后续加入 NVMe offload 的 AgentX 版本中，我们可能会调高这个上限，以便测量更长的 TTL。就目前而言，5 分钟是合理的，因为这也是 [Anthropic 默认的 KV cache TTL](<https://platform.claude.com/docs/en/build-with-claude/prompt-caching#:~:text=Check%20that%20calls%20are%20made%20within%20the%20cache%20lifetime%20(5%20minutes%20by%20default)>)。
+
+这种程度的确定性保证了在推理引擎、硬件、并发和服务端设置相同的情况下，运行结果是可复现的。不过由于 AgentX 是[闭环基准测试](https://notpeerreviewed.com/blog/tail-latency/)，不同配置在不同速率下完成的请求数不同，因而各自遇到的工作负载会有一些自然差异。这在低并发下最明显：完成的请求本来就少，工作负载向数据集整体分布收敛的机会也更少。
+
+当一个对话在采集期间结束时，它所在的回放通道会从数据集采样器中选取另一个对话。每次回放都会获得一个唯一且确定的 cache-bust 标记，并加在每条独立前缀链的最前面，包括主智能体链，以及任何全新上下文的子智能体链或一次性链。分叉出来的子智能体继承父级的标记。这个标记在一次回放内保持不变，从而保留其 KV 复用模式，但在不同回放之间会变化，以避免出现虚高的缓存命中率。这也使得并发数可以超过数据集中的对话总数（393）。
+
+### AgentX 的公平投机解码方法
+
+如前所述，数据集在采集时已做匿名化，因此那些 64 token 的哈希块必须在回放前用合成内容填充。AIPerf 的做法是从一个合成的编码 / 工具调用 token 池中确定性地采样。
+
+重要的是，KV 复用模式和请求时序都被保留了下来，但合成的请求内容确实带来了一些额外考量：在合成数据上运行投机解码方法，可能导致 speculator 接受 / 拒绝的 token 数与真实数据下明显不同（因为 speculator 并未在合成数据上训练过）。
+
+我们在 [InferenceX v2 文章](https://newsletter.semianalysis.com/i/188090866/multi-token-prediction-mtp)中讨论过这个不足，此后改进了方法。我们与社区密切合作，确保大多数开源推理引擎都提供一种机制，让用户可以强制指定从 speculator 接受多少 draft token（即“接受长度”或“接受率”）。随后，对每一种（模型、speculator、draft 长度、思考模式）组合，[我们在 SPEED-Bench 上采集平均 AL](https://github.com/SemiAnalysisAI/InferenceX/tree/main/golden_al_distribution)。[SPEED-Bench](https://huggingface.co/datasets/nvidia/SPEED-Bench) 是一个 agentic 编码数据集，被描述为“用于在多样语义领域和真实服务条件下评估投机解码（SD）的统一基准”。
+
+然后在运行时，我们把这些真实的投机解码接受长度应用到 AgentX 上，以确保厂商中立的公平性。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c15f47a8-fcb3-4220-80ee-4b95bfa7b0b0_2048x1376.png"
+  caption="来源：GitHub"
+/>
+
+## agentic 工作负载详细遥测数据探索教程
+
+AgentX 需要的不只是新的基准测试 harness 和数据集。[我们还花了一些时间重建 InferenceX 可视化的部分模块，让 agentic 结果更易于探索和理解。](https://inferencex.semianalysis.com/inference/agentic/439903)一个 AgentX 数据点代表着数千条请求，跨越不断增长的对话、子智能体、warmup 阶段、缓存状态和动态变化的在途负载。因此，Pareto 曲线上的单个点会掩盖大量有用信息。正如我们反复强调的：推理服务从来没有一刀切的方案。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/46159be7-8b5b-49d7-990f-2ce1cf1217d0_1110x687.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference/agentic/439903)"
+/>
+
+其中一项重要改动是曲线本身的构建方式。在之前版本的 InferenceX 中，启用和未启用投机解码的配置常常被画成两条独立曲线；现在我们不再这样做。前端会合并允许启用的推理优化，为每个模型、SKU 和推理引擎组合展示可获得的最佳曲线。因此，同一条曲线上的不同点可能使用了不同的优化技术和配置，包括投机解码、分离式服务或 KV cache offload。
+
+我们的目标是展示每套软硬件栈在生产中能达到的最佳性能，而不是为每种优化组合都单独画一条曲线。不过，我们仍会公开每个点的底层配置与溯源信息。点击某个点会弹出详细信息，显示究竟是哪套配置产生了它，以及运行元数据、可公开查看的 CI 溯源链接和 AgentX 专属统计。从那里点击“View charts”可以打开完整的点详情页，查看 AgentX 专属统计。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/add2d50c-a0be-4771-a00b-e6b4f1fe8502_2048x1256.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+点详情页对所选 AgentX 运行提供了深入得多的视角，包含输入和输出序列长度分布、交互性与 TTFT 随时间的变化、KV cache 利用率、请求队列深度、前缀缓存命中率、输入与 decode 吞吐量、prompt token 来源拆分，以及唯一输入 token 随时间的变化。有了这些指标，就更容易理解为什么两个聚合吞吐量相近的点在整个回放过程中表现却不同。
+
+该页面还把 warmup 与采集阶段的数据分开。读者可以在两个阶段之间切换，观察系统在缓存状态建立过程中的行为，以及在用于基准测试的采集期内的行为。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/91abef64-d639-4da8-8d7e-3499be750011_1001x1541.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+使用 KV cache offload 的点在主图上会多一圈虚线，用来与其他点区分。选中这类点时，详情页会显示 offload 类型、KV offload 引擎、芯片缓存命中率和 CPU 缓存命中率。这样就能看出 KV offload 在最佳曲线中的贡献，而不必为每种 offload 配置单独画一条曲线。
+
+另一个新功能是请求时间线。该视图展示所选 AgentX 运行中回放的各条请求，可以按对话或按 worker 组织。对话视图会把子智能体归到其根对话之下，便于看清对话与子智能体何时重叠。warmup 与采集期的请求同样可以分开查看。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c8fc6236-52bd-45ff-ac2e-539899c00d87_2048x1097.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+时间线中的每条请求都可点击，直接跳转到 InferenceX 数据集页面上对应的对话和轮次。这样，读者可以从 Pareto 曲线上的一个聚合点，一路追到实际被回放的那条匿名化请求。
+
+[AgentX 页面](https://inferencex.semianalysis.com/datasets)还提供了火焰图，用于可视化单个对话的结构。每个条形代表一轮，长度相对该对话中最长的一轮缩放，并被拆分为已缓存前缀 token、未缓存输入 token 和生成的输出 token。这直观展示了上下文如何在对话中增长，以及每条请求中有多少可以从 KV cache 复用。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a0e991d9-6ab6-43e7-9cc9-93ac9ca5b7bd_2048x879.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+## InferenceX/AgentX 的下一步
+
+我们会继续朝着让 AgentX 成为最真实、最有代表性的基准测试这一目标前进。在可预见的一段时间内，我们会继续为当前的 v1.0.x harness [做小修复](https://github.com/SemiAnalysisAI/aiperf/releases)，同时计划在 v1.1 harness 上做更大的改动。同一模型的所有提交将始终运行同一个次版本，以保证结果可比。
+
+作为快速跟进，我们会加入 SSD/NVMe KV offload。这将支持比 DRAM 更大的 KV cache 工作集，从而打开 Pareto 曲线左侧的高吞吐区间。
+
+我们还会采集规模更大、更多样、更新的 agentic trace 数据集，覆盖更多模型和 harness。下一版数据集不再把每条请求表示为一整串连续的哈希 ID，而会保留系统指令、用户与助手消息、工具调用和工具返回结果之间的边界。这样 AgentX 就能评估那些利用 agent harness 已知、但通常对推理引擎不可见的信息的工作负载感知服务技术。例如，router 可以把低复用的工具流量导向专用 prefill worker、在长时间工具调用期间保留 agent 的前缀，或在子智能体分叉时预取并共享前缀。当前格式保留了请求规模和 KV 复用模式，但缺少评估这些优化所需的结构信息。
+
+Ishan 提交的这份 [SGLang RFC](https://github.com/sgl-project/sglang/issues/27574) 就是说明这种更丰富 trace 结构为何重要的一个具体例子。它提出了一套由 router 发起的提示接口，利用会话生命周期、共享前缀边界、工具调用时长和子智能体状态等信息，告诉引擎何时应当共享、预取、降级、钉住或保留 KV。捕获这些结构信息，将使未来版本的 AgentX 能够评估这类工作负载感知的缓存策略，而不只是回放扁平的 token 前缀。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/36899201-da3b-42ec-888f-2443c63980aa_2048x1260.png"
+  caption="来源：[GitHub](https://github.com/sgl-project/sglang/issues/27574)"
+/>
+
+最后，我们正在采集细粒度和粗粒度的功耗遥测数据，以便更准确地衡量不同软硬件栈“每单位智能所耗焦耳”的效率。
+
+我们手上有大量数据，也有非常多可能的可视化方式。欢迎告诉我们你希望看到哪些可视化，以及更广泛的功能需求！
+
+[提交前端功能建议！](https://github.com/SemiAnalysisAI/InferenceX-app/issues/new/choose)
+
+## 模型全生命周期中的性能表现
+
+接下来几节我们转向历史单轮数据（8k1k），它覆盖每个模型从发布当天到我们停止对其持续测试为止的表现。这里 NVIDIA 和 AMD 都有亮眼结果，包括若干 MI355X 胜出的固定序列长度配置。
+
+AgentX 更能代表当下的 agentic 推理工作负载。不过，历史上的固定序列 InferenceX 结果对追踪性能随时间的变化仍然有用。8k1k、1k1k 这类工作负载剥离了大部分会话级行为，包括前缀复用、持久 KV cache 状态和路由亲和性，因此它们对当前生产流量的代表性较弱，但仍然适合用来观察推理性能如何随软件支持的成熟而提升。
+
+<Blur>
+
+Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+
+</Blur>
+
+---
+
+_本文完整版发布在我们的 Substack 上。[订阅 SemiAnalysis](https://newsletter.semianalysis.com/subscribe) 阅读完整文章。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "什么是 AgentX 1.0？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "AgentX 1.0 是全球首个完全开源、1M 上下文的多轮 agentic 编码推理基准测试，以 Apache 2.0 协议发布。它回放真实的匿名化 Claude Code 与 Codex trace，而不是固定序列长度的 prompt；测试运行在约 2MW 的持续运行算力上，覆盖 1000 多颗芯片，包括 MI355X、GB300 NVL72、GB200 NVL72、B300、B200、MI325X、MI300X、H200 和 RTX Pro 服务器。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "agentic 工作负载与固定序列长度基准测试有什么区别？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "agentic 工作负载是多轮、长上下文的，前缀复用率极高，并会突发启动短生命周期的子智能体。这让 agentic 推理成为系统问题而非 kernel 问题：KV 张量必须跨节点传输，请求必须路由到持有对应前缀的 rank，长上下文还迫使 KV offload 到 DRAM 或 SSD。8k1k 这类固定序列长度工作负载把这些全部剥离，衡量的基本是芯片和 kernel 的基础性能。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "CUDA 护城河在 agentic 工作负载上还站得住吗？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "取决于模型。在 DeepSeek V4 Pro 上竞争非常胶着：MI355X SGLang 在每美元性能上一度追平 B200 vLLM，直到 2026 年 8 月 21 日 Inferact 和 NVIDIA 在 vLLM 上的优化让 B200 反超，而 B300 vLLM 和 B200 SGLang 此前已经领先 MI355X。在 MiniMax M3、Qwen3.5 和 GLM 5.3 上，NVIDIA 大幅领先：Qwen3.5 SGLang 在 90 tok/s/user 下领先 20 倍以上，GLM 5.3 在 150 tok/s/user 下成本效率最高领先 5 倍。上下文并行是护城河中很具体的一环：在 vLLM 的注意力后端支持矩阵中，AMD 的每一个后端都标为不支持。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AgentX 数据集包含什么？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "SemiAnalysis 通过拦截发往 Claude 和 Codex 的 HTTP 请求的代理采集 trace，累计超过 8,000 个会话、340 万条请求和 6,100 亿 token，合计支出超过 300 万美元。开源的 AgentX v1.0 子集包含 393 个会话，输入序列长度中位数 142k token，输出序列长度中位数 444 token，轮次间延迟中位数 3.84 秒；其中 175 个会话（约 44%）至少包含一个子智能体，共 1,697 次子智能体 rollout。内容被匿名化为 64 token 分块并替换为会话内链式哈希，因此前缀复用模式得以保留，而原始 prompt 不会保留。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "什么是 E2E 归一化交互性？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "E2E 归一化交互性是一个实验性指标，定义为输出序列长度除以端到端延迟。由于端到端延迟等于 TTFT 加上输出长度乘以 TPOT，该指标实际等于交互性加上一项与 TTFT 成正比的惩罚，用一个数字同时刻画两个维度的响应性。它并不完美：对高 TTFT 的惩罚很重，也无法体现 prefill-decode 分离等优化的全部细节，因此 AgentX v1.0 的提交仍然分别针对交互性和 TTFT 做优化。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AgentX 目前带来了哪些行业影响？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "vLLM、SGLang、TensorRT-LLM、ATOM、AITER、Dynamo、LMCache 和 Mooncake 上已有 70 多个上游 PR 以 AgentX 为北极星基准。例如：vLLM 的混合注意力前缀缓存在 1M 上下文下前缀缓存命中率超过 95%；CPU KV offload 扩展到 DeepSeek V4，输出吞吐量提高 81.7%、平均端到端延迟降低 46.6%；TensorRT-LLM 的边界感知增量分词在 Qwen3.5 上把平均处理时间从 185.1 ms 降到 11.3 ms；AMD ATOM 的稀疏检查点保留把 DeepSeek V4 的前缀命中率从 5.6% 提升到 96.45%。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/brief-overview-of-agentic-workloads.mdx b/packages/app/content/blog/zh/brief-overview-of-agentic-workloads.mdx
new file mode 100644
index 000000000..91987ef6d
--- /dev/null
+++ b/packages/app/content/blog/zh/brief-overview-of-agentic-workloads.mdx
@@ -0,0 +1,134 @@
+---
+title: 'Agentic 工作负载简述'
+subtitle: '多轮会话、长上下文与近乎完全的 prefix 复用，让 agentic 推理成为一个系统问题，也改变了基准测试需要测量的内容'
+seoDescription: 'Agentic 工作负载的四个特征：多轮会话、长上下文、高 prefix 复用与突发式 subagent 扇出，以及为什么对其做基准测试是一个系统问题。'
+date: '2026-08-19'
+publishDate: '2026-08-19'
+tags:
+  - agentic
+  - agents
+  - agentx
+  - benchmark
+  - inference
+---
+
+概括来说，[agentic 工作负载](/zh/glossary/agentic-coding-workload)有四个特征：
+
+- **多轮。** 一次会话包含几十到几百次用户与助手交互，而 chatbot 场景通常只有寥寥数轮。
+- **长上下文。** system prompt、tool 定义，再加上轮次本身的数量，上下文会迅速累积。
+- **高 [prefix 复用](/zh/glossary/prefix-caching)。** 对话呈线性推进，第 _n-1_ 轮的输出会拼接进第 _n_ 轮的输入，因此大部分上下文可以直接由 [KV cache](/zh/glossary/kv-cache) 提供，而无需重新计算——具体比例取决于可用于存放 KV tensor 的存储容量。随着 _n_ 增大，命中 cache 的输入与未命中输入之比会趋近于 1。
+- **突发式 [subagent](/zh/glossary/subagent) 扇出。** 一轮可能同时启动多个短生命周期的 subagent，由此产生尖峰式的 KV cache 分配与淘汰模式。
+
+<Figure
+  src="/images/brief-overview-of-agentic-workloads/context-accumulation.png"
+  alt="四轮 agentic 会话的上下文累积示意图：每一轮的输入都包含上一轮的完整输入及其输出，其中共享部分被标记为来自上一轮的 cache"
+  caption="四轮 agentic 会话中的上下文累积。高亮部分均由 cache 提供，只有最新的输出与 tool 结果未命中 cache。"
+/>
+
+## Agentic 推理是一个系统问题
+
+考虑到上述特征，对这类工作负载做基准测试，与现有的固定序列长度基准测试有本质区别。
+
+由于 prefix 复用比例极高，KV tensor 必须在节点与 rank 之间高效传输，这正是 NIXL、MORI-IO 和 [Mooncake](/zh/agentx/optimizations/mooncake) 这类传输层要解决的问题。随后还需要把不同对话路由到其 prefix 所在的节点或 rank，以最大化 cache 命中率，llm-d、[Dynamo](/zh/agentx/optimizations/dynamo) 以及 vLLM 与 SGLang 的 router 承担的就是这项工作。长上下文对话会给 KV cache 的 HBM 容量带来压力，迫使 KV tensor offload 到 DRAM、SSD 等更慢的存储层级，而这一过程同样需要高效实现——Mooncake Store、[LMCache](/zh/agentx/optimizations/lmcache)、vLLM 的 CPU offloading 与 SGLang HiCache 都在攻克这一层。
+
+与之相对的是固定序列长度的单轮工作负载：prefix 复用在其中并不相关，推理性能主要反映芯片与 kernel 的基础表现。这并不意味着 InferenceX 上大量的固定序列长度数据不重要。剥离 agentic 推理服务的复杂性之后，这些数据能清晰呈现底层推理优化的进展，也构成了解读 AgentX 结果时所依据的基线。
+
+## 构建真实的回放语料
+
+为了让 AgentX 的工作负载尽可能贴近真实场景，我们收集了 393 条 SemiAnalysis 内部的 Claude Code trace 用于回放。为在匿名化内容的同时保留原始的 prefix 复用模式，我们采用了与 [Qwen-Bailian](https://github.com/alibaba-edu/qwen-bailian-usagetraces-anon) 类似的方法——后者是最早公开的生产环境 trace 语料之一。随后我们使用 [AIPerf](https://github.com/ai-dynamo/aiperf) 按照原始的请求时间表，在不同并发客户端数量下重建这些 trace。
+
+重建 subagent 结构需要 Claude Code 本身做出改动。我们与 Anthropic 合作推动了两项功能落地，它们都为出站请求补充了 agent 实例身份信息，使并行分支能在 proxy 层被区分开：
+
+- [anthropics/claude-code #49207](https://github.com/anthropics/claude-code/issues/49207) —— 在 API 请求 header 中加入 agent 实例 ID（`x-claude-code-agent-id`）及其父级 ID（`x-claude-code-parent-agent-id`），让交错的 subagent 请求可被归因，而不是坍缩成单一的会话级流量。
+- [anthropics/claude-code #66761](https://github.com/anthropics/claude-code/issues/66761) —— 把同样的 header 扩展到由 `Workflow` tool 扇出产生的 subagent，此前这些请求只携带父级对话的 session ID。
+
+缺少这些 header，一个扇出 N 个并发 subagent 的会话就与 N 个互不相关的对话无法区分，塑造[回放图](/zh/glossary/trace-replay)的 spawn 与 join 依赖关系也就无从还原。
+
+完整的数据集规则、分布图与来源链接发布在 [AgentX 方法论](/zh/agentx/methodology)页面；该工作负载已经暴露出的具体引擎改进，则记录在 [AgentX 优化追踪页](/zh/agentx/optimizations)。
+
+## 致谢
+
+感谢 Anthropic 团队在 Claude Code 中落地 agent 身份标识 header，也感谢 Qwen-Bailian 与 AIPerf 背后的阿里巴巴与 NVIDIA 团队——AgentX 的回放流水线正建立在他们的工作之上。
+
+<DashboardCTA href="/zh/inference?i_seq=agentic-traces">对比实时 AgentX 推理结果 →</DashboardCTA>
+
+## 常见问题
+
+### 什么是 agentic 工作负载？
+
+Agentic 工作负载指由 AI agent 而非真人在对话框中输入所产生的推理流量。它的特征是几十到几百轮交互的多轮会话、由 system prompt 与 tool 定义堆叠出的长上下文、跨轮次极高的 prefix 复用，以及成批出现的短生命周期 subagent。
+
+### 为什么 prefix 复用对 agentic 推理如此关键？
+
+每一轮的输入都包含上一轮的输入加上其输出，因此绝大部分上下文此前已经处理过。只要该 prefix 的 KV tensor 仍然驻留，prefill 工作就可以跳过。随着会话增长，命中 cache 的输入与未命中输入之比会趋近于 1，于是 cache 容量与命中率主导了推理服务的成本。
+
+### 对 agentic 工作负载做基准测试，与固定序列长度基准测试有何不同？
+
+固定序列长度基准测试发送 input 与 output 长度固定的独立请求，prefix 复用不构成变量，结果主要反映芯片与 kernel 性能。Agentic 基准测试回放的是彼此依赖的会话，其中上下文不断增长、prefix 被共享、subagent 会扇出，因此 KV 传输、prefix 感知路由与 cache 分层 offload 都成为被测量的对象。
+
+### 突发式 subagent 扇出会对 KV cache 造成什么影响？
+
+单轮可能一次性拉起多个 subagent，每个都要分配自己的 KV cache，短暂运行后结束，因此 cache 压力以尖峰而非稳态形式到来。面向均匀请求流调优的调度器与淘汰策略，在这种模式下的表现会明显不同。
+
+### AgentX 的回放语料是如何构建的？
+
+我们收集了 393 条 SemiAnalysis 内部的 Claude Code trace，采用与 Qwen-Bailian 类似的分块哈希方法完成匿名化，在不保留内容的前提下保留 prefix 关系。随后由 AIPerf 按原始请求时间表，在不同并发客户端数量下重建这些 trace。
+
+### 固定序列长度基准测试还有意义吗？
+
+有。固定序列长度结果剥离了 agentic 推理服务的复杂性，单独呈现底层芯片与 kernel 性能，因而是解读 AgentX 结果时所依据的基线。
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "什么是 agentic 工作负载？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Agentic 工作负载指由 AI agent 而非真人在对话框中输入所产生的推理流量。它的特征是几十到几百轮交互的多轮会话、由 system prompt 与 tool 定义堆叠出的长上下文、跨轮次极高的 prefix 复用，以及成批出现的短生命周期 subagent。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "为什么 prefix 复用对 agentic 推理如此关键？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "每一轮的输入都包含上一轮的输入加上其输出，因此绝大部分上下文此前已经处理过。只要该 prefix 的 KV tensor 仍然驻留，prefill 工作就可以跳过。随着会话增长，命中 cache 的输入与未命中输入之比会趋近于 1，于是 cache 容量与命中率主导了推理服务的成本。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "对 agentic 工作负载做基准测试，与固定序列长度基准测试有何不同？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "固定序列长度基准测试发送 input 与 output 长度固定的独立请求，prefix 复用不构成变量，结果主要反映芯片与 kernel 性能。Agentic 基准测试回放的是彼此依赖的会话，其中上下文不断增长、prefix 被共享、subagent 会扇出，因此 KV 传输、prefix 感知路由与 cache 分层 offload 都成为被测量的对象。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "突发式 subagent 扇出会对 KV cache 造成什么影响？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "单轮可能一次性拉起多个 subagent，每个都要分配自己的 KV cache，短暂运行后结束，因此 cache 压力以尖峰而非稳态形式到来。面向均匀请求流调优的调度器与淘汰策略，在这种模式下的表现会明显不同。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AgentX 的回放语料是如何构建的？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "我们收集了 393 条 SemiAnalysis 内部的 Claude Code trace，采用与 Qwen-Bailian 类似的分块哈希方法完成匿名化，在不保留内容的前提下保留 prefix 关系。随后由 AIPerf 按原始请求时间表，在不同并发客户端数量下重建这些 trace。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "固定序列长度基准测试还有意义吗？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "有。固定序列长度结果剥离了 agentic 推理服务的复杂性，单独呈现底层芯片与 kernel 性能，因而是解读 AgentX 结果时所依据的基线。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set.mdx b/packages/app/content/blog/zh/deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set.mdx
new file mode 100644
index 000000000..8af550d2b
--- /dev/null
+++ b/packages/app/content/blog/zh/deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set.mdx
@@ -0,0 +1,86 @@
+---
+title: 'DeepSeek V4 Pro 的 AgentX 结果：B200 对比 B300 与 KV cache 工作集'
+subtitle: '多出 50% 的 HBM 能挤出额外吞吐量，逐点遥测数据说明了它究竟从何而来'
+seoDescription: 'DeepSeek V4 Pro AgentX 运行中 B200 与 B300 的对比。按 TCO 归一化后的吞吐量相当接近，但遥测数据显示 B300 的 HBM 缓存命中率为 91%、B200 为 73%，KV 工作集分别为 4300 万与 2200 万 token。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - deepseek
+  - b200
+  - b300
+  - h200
+  - nvidia
+---
+
+按[总拥有成本](/zh/glossary/total-cost-of-ownership)归一化之后，B300 vLLM 与 B200 vLLM 在 DeepSeek V4 Pro 上的聚合式性能相当接近。主要差别在于 B300 能挤出额外的吞吐量，因为它的 HBM 容量比 B200 多 50%。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/939b9197-321d-4869-ba13-f0b800343b7b_2048x1300.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_active=b200_vllm%2Cb300_vllm)"
+/>
+
+真正有意思的不是这个汇总数字，而是 AgentX 服务端遥测在它下面揭示的内容。
+
+## 4300 万 token 对 2200 万 token
+
+在 384 条并发 agentic trace 的负载下，B300 vLLM DEP8 配 3TB DRAM、使用 vLLM simple offloading，取得了 91% 的 HBM 缓存命中率，外加 1.36% 的 DRAM 缓存命中率。原因是该配置下 HBM 中 [KV cache](/zh/glossary/kv-cache) 的工作集规模约为 4300 万 token，而任一时刻在途的 token 数只是略微超过这个量。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/f86888c4-c1a7-47b0-a580-d1a74876cb9c_2048x898.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference/agentic/439522)"
+/>
+
+换成并发 196 的 B200、其余参数不变，情况就反过来了：HBM 缓存命中率降到 73%，部署明显更依赖 DRAM，[offload](/zh/glossary/kv-cache-offload) 缓存命中率接近 20%。此时 HBM 中 KV cache 的工作集规模为 2200 万 token，大约是 B300 的一半。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/70a7a7c7-7918-459c-a4e0-f7fc1edca816_2048x887.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference/agentic/439967)"
+/>
+
+## offload 层为什么需要余量
+
+DRAM KV offload 通常实现为写穿缓存，也就是写入 HBM 缓存的每个前缀同时也会写入 DRAM 缓存。因此，只有当可用于 offload 的 DRAM 显著大于 HBM 中 KV cache 的容量时（大约 1.5 到 3 倍）它才最有效。
+
+正是这一性质，让 B300 额外的 HBM 产生的是叠加效应而不只是简单相加：更大的常驻工作集把更多会话状态留在快速层，慢速层因此只需服务真正的剩余部分，而不必不断吸收被淘汰的数据。
+
+## H200 这个反例
+
+H200 SGLang FP8 能在低并发下服务 DeepSeek V4，从每美元性能看甚至可以和 B200、MI355X SGLang 一较高下。但它无法在高吞吐场景中与更新的 SKU 竞争，因为 HBM 容量不够。
+
+它的失效方式很有教育意义：在更高并发下对 DRAM KV offload 的依赖，会随着用户数增加带来无法接受的延迟。offload 能换来容量，但并不免费；一套需要靠它来弥补结构性显存不足的部署，最终会在尾部延迟上付出代价。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/91dfd9ca-d13d-43e9-a57e-3be0fcc4a5cf_2048x1315.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=h200_dynamo-sglang&i_optimal=0)"
+/>
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 DeepSeek V4 Pro 上，B300 比 B200 好多少？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "按总拥有成本归一化后，B300 vLLM 与 B200 vLLM 的聚合式性能相当接近。B300 凭借多出 50% 的 HBM 容量挤出了额外吞吐量，这在遥测数据中体现为：并发 384 时 HBM 中 KV cache 工作集约 4300 万 token、缓存命中率 91%；而 B200 在并发 196 时为 2200 万 token、命中率 73%。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "H200 为什么在高并发下落后？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "H200 SGLang FP8 在低并发下能很好地服务 DeepSeek V4，在该区间的每美元性能上甚至可与 B200、MI355X SGLang 一较高下。但它缺少足以支撑高吞吐场景的 HBM，因此随着并发上升只能更多依赖 DRAM KV offload，而这种依赖会随用户数增长带来无法接受的延迟。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/deepseek-v4-pro-agentx-gb200-vs-gb300-disagg.mdx b/packages/app/content/blog/zh/deepseek-v4-pro-agentx-gb200-vs-gb300-disagg.mdx
new file mode 100644
index 000000000..81e1985ff
--- /dev/null
+++ b/packages/app/content/blog/zh/deepseek-v4-pro-agentx-gb200-vs-gb300-disagg.mdx
@@ -0,0 +1,72 @@
+---
+title: 'DeepSeek V4 Pro 的 AgentX 结果：GB200 与 GB300 的机架级分离式部署'
+subtitle: '两者都依赖 PD 分离，GB300 额外使用 DEP32 宽专家并行解码，而差距体现在首 token 延迟而非 token 速率上'
+seoDescription: 'DeepSeek V4 Pro 在 AgentX 上的 NVIDIA 机架级结果。GB300 Dynamo TRTLLM 与 GB200 Dynamo vLLM 都采用 prefill-decode 分离；GB300 那个点在 TTFT 上落后，因为更高的并发带来更多子智能体流量和更多冷 prefill。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - deepseek
+  - gb200
+  - gb300
+  - nvl72
+  - disagg
+  - wide-ep
+  - nvidia
+---
+
+在 DeepSeek V4 Pro 上，NVIDIA 最有竞争力的方案是 GB300 Dynamo TRTLLM 和 GB200 Dynamo vLLM。两者都依靠 prefill-decode [分离](/zh/glossary/disaggregated-inference)，在合理的[交互性](/zh/glossary/interactivity)下取得高吞吐。GB300 配置还额外采用了 DEP32 的[宽专家并行](/zh/glossary/wide-expert-parallelism)解码实例，以在前沿曲线中段获得更高吞吐。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/512c6225-5816-493e-839d-e4f539e5b1c4_2048x1215.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_metric=y_tpPerGpu&i_xmode=interactivity&i_linelabel=0&i_best=0&i_active=gb200_dynamo-vllm%2Cgb300_dynamo-trt)"
+/>
+
+## 两个点在 token 速率上比在首 token 延迟上更接近
+
+以每用户每秒 token 数衡量时，2xDEP8+1xDEP12 的 GB200 点与 3xDEP8+1xDEP16 的 GB300 点相当接近；但以首 token 延迟衡量时，两者差距明显更大。因此只看交互性坐标轴，会低估这两种配置行为上的差异。
+
+总体而言，TTFT 对工作负载的「突发性」更敏感。GB300 那个点的总并发要高得多，因而会遇到更多子智能体流量，也就有更多冷 prefill。[子智能体](/zh/glossary/subagent)带着全新上下文启动，没有可复用的前缀，因此它的第一条请求是完整 prefill 而不是缓存命中；当足够多这样的请求同时到达时，[尾部延迟](/zh/glossary/tail-latency)就会被推高。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/96f9c9f9-eb86-4f87-8332-c85f23040836_2048x916.png"
+  caption="来源：InferenceX"
+/>
+
+## 为什么这是 agentic 效应而不是硬件效应
+
+在定长工作负载上完全看不到这些。在 8k1k 场景中，每条请求彼此独立且形状相同，因此提高并发只是放大 batch，并不改变已缓存与未缓存工作的构成。而在 AgentX 上，提高并发会改变流量本身的组成：会话更多意味着并行的子智能体分支更多，而这些分支是成簇到达的，不是平滑的请求流。
+
+实际的读法是：一套为了在吞吐曲线上更靠右而调优的机架级配置，同时也接受了更具突发性的 prefill 负载；这个选择的代价落在首 token 延迟上，而不是同一张图所展示的 token 速率上。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f79e53f5-c8ff-4ed0-a8a3-c05b9e0b4bab_2048x1191.png" />
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 DeepSeek V4 Pro 上，NVIDIA 哪种机架级配置最快？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "GB300 Dynamo TRTLLM 和 GB200 Dynamo vLLM 是 NVIDIA 在该模型上最有竞争力的配置，两者都依靠 prefill-decode 分离，在合理交互性下达到高吞吐。GB300 还额外运行 DEP32 的宽专家并行解码实例，以提升前沿曲线中段的吞吐量。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "为什么 GB300 那个点的 TTFT 比 GB200 更差？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "GB300 那个点的总并发高得多，因而带来更多子智能体流量，也就有更多冷 prefill。子智能体带着全新上下文启动、没有可复用前缀，它的第一条请求是完整 prefill；这类请求成簇到达就会推高尾部延迟。总体而言，TTFT 比每用户 token 速率对工作负载的突发性更敏感。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/deepseek-v4-pro-agentx-mi355x-vs-b200-august.mdx b/packages/app/content/blog/zh/deepseek-v4-pro-agentx-mi355x-vs-b200-august.mdx
new file mode 100644
index 000000000..1b2b14dca
--- /dev/null
+++ b/packages/app/content/blog/zh/deepseek-v4-pro-agentx-mi355x-vs-b200-august.mdx
@@ -0,0 +1,84 @@
+---
+title: 'DeepSeek V4 Pro 的 AgentX 结果：MI355X 对比 B200，以及 8 月 21 日的反转'
+subtitle: '在端到端延迟的每美元性能上，AMD 一度追平 B200 vLLM，随后上游 vLLM 的优化改变了格局'
+seoDescription: 'DeepSeek V4 Pro 0813 在 MI355X 与 B200 上的 AgentX 1.0 结果。在 2026 年 8 月 21 日之前，MI355X SGLang 在每美元性能上追平了 B200 vLLM；此后来自 Inferact 与 NVIDIA 的 vLLM 优化让 B200 反超。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - deepseek
+  - mi355x
+  - b200
+  - amd
+  - nvidia
+---
+
+DeepSeek V4 Pro 0813 是一款来自中国、极受欢迎的前沿开放权重模型，总参数约 1.6 万亿，激活参数 490 亿。它也是 [AgentX](/zh/glossary/agentx) 上 AMD 与 NVIDIA 竞争最胶着的模型，而且排名在一周之内就发生了变化。
+
+工作负载在这里很关键。AgentX 回放的是真实的 agentic 编码会话，而不是固定提示词；在所有 DeepSeek V4 运行中，输入长度分布为 p50 88k、p90 272k、p95 404k、p99 675k token，输出分布为 p50 413、p90 2.2k、p95 3.7k、p99 8.6k。这与 8k1k 是完全不同的系统问题。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/360a051c-6117-4860-a704-b50198f01521_1978x1246.png"
+  caption="来源：InferenceX"
+/>
+
+在 2026 年 8 月 21 日之前，AMD 实力很强的 MI355X SGLang 开发团队在端到端性能的每美元表现上追平了 B200 vLLM。这是实打实的结果，值得直说。
+
+但这并不是全部。同一时期，B300 vLLM 和 B200 SGLang 已经领先 MI355X。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9ec4708c-148e-42ee-a7c5-3bf4d9cbdfaa_1938x1246.png"
+  caption="来源：InferenceX"
+/>
+
+8 月 21 日之后，来自 Inferact 和 NVIDIA 的 vLLM 优化让 B200 的每美元性能超过了 MI355X。这仍是一场胶着的竞赛，而不是已成定局；[AMD 也公布了自己的 DeepSeek V4 vLLM 优化清单](https://github.com/vllm-project/vllm/issues/52911)，上面还有不少待办事项。
+
+## ATOM 这个前提
+
+在端到端延迟上，MI355X 上的 ATOM 胜过 B200 vLLM，但没有胜过 B300 或 B200 SGLang。更重要的前提在于部署：中国和西方的多数 AI 实验室都不愿在生产中使用 ATOM，因此只在厂商引擎下成立的优势，并不能代表使用上游软件栈的客户实际拿到的性能。在单节点上，MI355X 的开源方案（vLLM）依然落后于 ATOM。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e8d915c0-c962-4277-b62c-a1b08b8e7179_1924x1262.png"
+  caption="来源：InferenceX"
+/>
+
+## MI355X 真正有竞争力的区间
+
+总体来看，MI355X 相对 B200 和 B300 表现尚可。在曲线上低吞吐、低延迟的那一段两者最为接近，因为那里只用到张量并行和较为基础的 kernel。差距出现在高吞吐区间，AMD 需要在那里优化自己的 DEP kernel；而且这个差距是在 MI355X 拥有 B200 的 1.5 倍 HBM 的前提下出现的。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/29cfd340-d61b-46c4-8d87-e081d937b80c_2048x1303.png"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/inference?g_rundate=2026-08-21&g_runid=32433563482&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_vllm%2Cb300_vllm%2Cmi355x_atom%2Cmi355x_vllm&i_label=0)"
+/>
+
+两个坐标轴要一起看。某个配置可能在吞吐量对[交互性](/zh/glossary/interactivity)上表现亮眼，却在 [p90](/zh/glossary/tail-latency) 首 token 延迟上把优势还回去，这也是上图选择以 TTFT 为横轴、而不是只看交互性的原因。
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 DeepSeek V4 Pro 上，MI355X 胜过 B200 了吗？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在 2026 年 8 月 21 日之前的一段时间里，MI355X SGLang 在端到端性能的每美元表现上追平了 B200 vLLM，MI355X 上的 ATOM 在端到端延迟上也胜过 B200 vLLM。但同期 B300 vLLM 和 B200 SGLang 都领先 MI355X；8 月 21 日之后，来自 Inferact 和 NVIDIA 的 vLLM 优化让 B200 的每美元性能反超 MI355X。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "MI355X 显存更大，为什么在高吞吐区间反而落后？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "MI355X 的 HBM 是 B200 的 1.5 倍，但在 AgentX 上，它与 B200、B300 最接近的是曲线上低吞吐、低延迟的那一段，那里只用到张量并行和较为基础的 kernel。高吞吐一端依赖 AMD 仍需优化的 DEP kernel，因此额外的显存容量没有转化为前沿吞吐量。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/deepseekv4-16t-day-0-to-day-43-performance.mdx b/packages/app/content/blog/zh/deepseekv4-16t-day-0-to-day-43-performance.mdx
index 7bb0d216f..bceea89d9 100644
--- a/packages/app/content/blog/zh/deepseekv4-16t-day-0-to-day-43-performance.mdx
+++ b/packages/app/content/blog/zh/deepseekv4-16t-day-0-to-day-43-performance.mdx
@@ -348,10 +348,10 @@ vLLM 的计划在 [DeepSeek V4 路线图 issue (#40902)](https://github.com/vllm
 - **核心模型支持**：持续 MegaMoE 工作（PR #40833）和 NVFP4 支持。
 - **运行时与并行**：Model Runner V2 集成、MTP 优化、prefill/decode (PD) 优化、pipeline parallelism 支持。
 - **Kernel 集成**：paged prefill kernel、fast top-k kernel、更多水平融合、DeepEP V2，以及与 DeepSeek 自研 TileKernels 的集成。
-- **KV cache**：KV cache 卸载，涵盖 PD + CPU 卸载（PR #39654）和分布式 KV 卸载。
+- **KV cache**：KV cache offload，涵盖 PD + CPU offload（PR #39654）和分布式 KV offload。
 - **硬件支持**：在已完成 Hopper 支持的基础上，SM120 和 AMD 支持仍是关键待办项。
 
-这里的核心主题聚焦在周边系统：新的模型运行器、pipeline parallelism、KV 卸载，以及更广泛的硬件覆盖。
+这里的核心主题聚焦在周边系统：新的模型运行器、pipeline parallelism、KV offload，以及更广泛的硬件覆盖。
 
 即将到来的 InferenceX 更新——SemiAnalysis 的开源公共 EcosystemX 仪表板——将可视化所有主要 ML 开源库在所有主要 AI 芯片上的软件演进、CI 覆盖率和队列时间：NVIDIA、AMD、TPU、Trainium、Huawei 等。
 
diff --git a/packages/app/content/blog/zh/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72.mdx b/packages/app/content/blog/zh/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
new file mode 100644
index 000000000..d61a5be8c
--- /dev/null
+++ b/packages/app/content/blog/zh/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
@@ -0,0 +1,75 @@
+---
+title: 'GLM 5.3 的 AgentX 结果：MI355X ATOM 在部分曲线上胜过 GB300 NVL72'
+subtitle: 'AMD 厂商引擎在每美元性能上胜出的区间，以及端到端归一化交互性究竟衡量什么'
+seoDescription: '在 GLM 5.3 上，于 p90 端到端归一化交互性的部分区间内，AMD ATOM 的每美元性能优于 GB300 NVL72 SGLang，甚至优于 TRTLLM。本文解释这一实验性指标衡量什么，以及它在哪些地方会产生误导。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - glm5
+  - mi355x
+  - gb300
+  - nvl72
+  - trtllm
+  - sglang
+  - amd
+  - nvidia
+---
+
+看 ATOM 的话，在 p90 端到端归一化交互性的部分区间内，AMD 的[每美元性能](/zh/glossary/performance-per-dollar)优于 GB300 NVL72 SGLang，甚至优于 TRTLLM。这一结果值得肯定，AMD 团队做得很好。
+
+对两家厂商的建议还是那两条：我们期待 AMD 把这些优化移植到 SGLang，也期待 NVIDIA 在接下来几周里尽快优化 GB300 NVL72。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9337ed5b-a320-49f7-b31b-cc4b5751e0ff_2048x1290.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+## 端到端归一化交互性衡量什么
+
+这个结果所对应的横轴是一个实验性指标，在据此下结论之前值得先弄清楚它的含义。[端到端归一化交互性](/zh/glossary/e2e-normalized-interactivity)的目标，是在同时考虑首 token 延迟和每秒 token 数的前提下，衡量用户实际感受到的响应速度，其定义为输出序列长度除以端到端延迟。
+
+把端到端延迟展开为 TTFT 加上输出长度乘以每输出 token 时间（实际上只解码输出长度减一个 token），可得下式。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9e472a49-d228-4d8d-876a-6bc0d59f9d97_746x594.png"
+  caption="来源：SemiAnalysis"
+/>
+
+它本质上等于[交互性](/zh/glossary/interactivity)（即 1/TPOT 那部分），再加上一项与 TTFT 成正比的惩罚。
+
+## 这个指标在哪里会误导
+
+需要说明的是，该指标是实验性的，并不完美。例如它对高 TTFT 的惩罚很重，也无法体现 prefill-decode [分离](/zh/glossary/disaggregated-inference)等优化的全部细节。AgentX v1.0 的所有提交仍然分别针对常规交互性和 TTFT 做优化；我们也会继续打磨能反映现代 agentic 推理各种细节的新北极星指标。
+
+这个前提在这里对双方都成立。一套用首 token 延迟换取流式速度的配置，在这条坐标轴上一定会被惩罚，无论这种取舍是否契合具体产品；因此仅凭端到端归一化交互性得出的胜负，应当再对照独立的交互性曲线和 TTFT 曲线核实，然后才能作为采购论据。
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "什么是端到端归一化交互性？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "它是一个实验性指标，定义为输出序列长度除以端到端延迟，目的是在同时考虑 TTFT 和每秒 token 数的前提下衡量用户感受到的响应速度。由于端到端延迟等于 TTFT 加上输出长度乘以每输出 token 时间，该指标实际等于交互性加上一项与 TTFT 成正比的惩罚。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "在 GLM 5.3 上，AMD ATOM 在哪里胜过 GB300 NVL72？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在 p90 端到端归一化交互性的部分区间内，ATOM 的每美元性能优于 GB300 NVL72 SGLang，甚至优于 TRTLLM。该指标是实验性的，且对高 TTFT 惩罚很重，因此这一结果应当再对照独立的交互性曲线和 TTFT 曲线核实。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/glm-5-3-agentx-nvidia-vs-amd-sglang-150-toks.mdx b/packages/app/content/blog/zh/glm-5-3-agentx-nvidia-vs-amd-sglang-150-toks.mdx
new file mode 100644
index 000000000..97065e908
--- /dev/null
+++ b/packages/app/content/blog/zh/glm-5-3-agentx-nvidia-vs-amd-sglang-150-toks.mdx
@@ -0,0 +1,69 @@
+---
+title: 'GLM 5.3 的 AgentX 结果：150 tok/s/user 下 NVIDIA 单 token 成本最多低 5 倍'
+subtitle: '在这个运行点上，即便 AMD 硬件白送也仍然填不平差距'
+seoDescription: 'GLM 5.3 在开源 SGLang 上的 AgentX 结果。在 150 tok/s/user 的 p90 交互性下，NVIDIA 的成本效率最高领先 5 倍；差距之大，即便竞品硬件免费，NVIDIA 的单 token 成本依然更低。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - glm5
+  - b200
+  - b300
+  - mi355x
+  - sglang
+  - nvidia
+  - amd
+---
+
+GLM 5.3 在 GLM 5.2 744B 基础上做了额外的后训练，是一款前沿级别的模型。
+
+在开源 SGLang 的性能上，这又是一个 NVIDIA 在真实 agentic 推理性能上胜过 AMD 的模型。在 150 tok/s/user 的 p90 [交互性](/zh/glossary/interactivity)下，NVIDIA 的成本效率最高领先 5 倍。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/4a21d140-b588-4b33-941d-7c8090ac42b5_2048x1286.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=GLM-5.2&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_prec=fp4&i_active=b200_sglang%2Cb300_sglang%2Cgb200_dynamo-sglang%2Cgb300_dynamo-trt%2Cmi355x_atom%2Cmi355x_sglang)"
+/>
+
+## 「硬件白送」检验
+
+衡量这种差距有一个很实用的方法。以 AMD 软件当前的状态，在 150 tok/s/user 时，NVIDIA 的性能优势大到这种程度：即便竞品芯片硬件白送（服务商当然仍要承担数据中心托管、电力和其他运营成本），用 NVIDIA 的单 token 成本依然更低。
+
+这个说法之所以重要，是因为在采购谈判中最常被拿来用的杠杆就是硬件价格；而在这个运行点上，这根杠杆的力量不够。真正的约束是软件，而不是物料成本。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/f85bdb45-792c-4330-9745-674b38b03425_2048x1289.png" />
+
+## 这是一次快照，不是最终结论
+
+这个差距是相对 AMD 软件当前状态测得的，而在这批结果的其他部分中，这一状态一直在快速变化。我们期待 AMD 在几周后的 AgentX 更新文章中带来性能优化，那篇文章还会包含其他一些非常有意思的结果。
+
+范围也要说清楚：这里比较的是开源 SGLang。在同一个模型的部分区间上，AMD 的厂商引擎讲的是另一个故事，我们[另文说明](/zh/blog/glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72)。
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 GLM 5.3 上，NVIDIA 比 AMD 便宜多少？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在开源 SGLang、150 tok/s/user 的 p90 交互性下，NVIDIA 的成本效率最高领先 5 倍。以 AMD 软件在该运行点的当前状态，这一优势大到即便竞品硬件免费（服务商仍需承担托管、电力和其他运营成本），NVIDIA 的单 token 成本依然更低。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AMD 在 GLM 5.3 上有胜出的地方吗？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "有，但不在这组对比里。5 倍这个数字是开源 SGLang 的结果。使用 ATOM 时，在 p90 端到端归一化交互性的部分区间内，AMD 的每美元性能优于 GB300 NVL72 SGLang，甚至优于 TRTLLM。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/inferencex-v2-nvidia-blackwell-vs-amd-vs-hopper.mdx b/packages/app/content/blog/zh/inferencex-v2-nvidia-blackwell-vs-amd-vs-hopper.mdx
index 0b71fc1b5..e9051378a 100644
--- a/packages/app/content/blog/zh/inferencex-v2-nvidia-blackwell-vs-amd-vs-hopper.mdx
+++ b/packages/app/content/blog/zh/inferencex-v2-nvidia-blackwell-vs-amd-vs-hopper.mdx
@@ -188,7 +188,7 @@ B200 SGLang 自我们去年 10 月首次发布以来，FP4 和 FP8 场景均呈
 
 ## 分离式推理框架
 
-NVIDIA 使用 Dynamo 作为其分离式推理方案。[Dynamo](https://docs.nvidia.com/dynamo/design-docs/overall-architecture) 是一个专为多节点分布式推理设计的推理框架，具备预填充-解码分离、请求路由和 KV 缓存卸载等技术。它与推理引擎无关，允许我们在基准测试中使用 SGLang 和 TRT-LLM 作为后端。对于 AMD，我们使用 SGLang 配合两种不同的 KV 缓存传输框架：MoRI 和 Mooncake。[MoRI](https://github.com/rocm/mori) 是一个高性能通信接口，专注于 RDMA 和 GPU 集成，提供网络集合操作和专家并行内核等应用。Mooncake [最近加入了 PyTorch 生态系统](https://pytorch.org/blog/mooncake-joins-pytorch-ecosystem/)，支持预填充-解码分离以及多种容错多节点功能。
+NVIDIA 使用 Dynamo 作为其分离式推理方案。[Dynamo](https://docs.nvidia.com/dynamo/design-docs/overall-architecture) 是一个专为多节点分布式推理设计的推理框架，具备预填充-解码分离、请求路由和 KV 缓存 offload 等技术。它与推理引擎无关，允许我们在基准测试中使用 SGLang 和 TRT-LLM 作为后端。对于 AMD，我们使用 SGLang 配合两种不同的 KV 缓存传输框架：MoRI 和 Mooncake。[MoRI](https://github.com/rocm/mori) 是一个高性能通信接口，专注于 RDMA 和 GPU 集成，提供网络集合操作和专家并行内核等应用。Mooncake [最近加入了 PyTorch 生态系统](https://pytorch.org/blog/mooncake-joins-pytorch-ecosystem/)，支持预填充-解码分离以及多种容错多节点功能。
 
 ## DeepSeek Disagg + WideEP 结果深入分析
 
@@ -493,7 +493,7 @@ AMD 在开源分布式推理、宽专家并行和 FP4 可组合性方面落后
 
 ## AMD ATOM 引擎
 
-AMD 推出了名为 ATOM 的新推理引擎。ATOM 能提供略好的单节点性能，但在许多关键功能上完全缺失，导致无法用于真实工作负载。例如，它不支持 NVMe 或 CPU KVCache 卸载、工具解析、宽专家并行或分离式服务。这导致生产环境中没有任何客户使用它。与 NVIDIA 的 TRTLLM 不同——后者在 TogetherAI 等公司全球范围内每小时生成数十亿 token，[并且支持工具解析和其他功能](https://nvidia.github.io/TensorRT-LLM/commands/trtllm-serve/trtllm-serve.html#cmdoption-trtllm-serve-serve-tool_parser)——由于上述功能缺失，目前没有任何 token 工厂在使用 ATOM。
+AMD 推出了名为 ATOM 的新推理引擎。ATOM 能提供略好的单节点性能，但在许多关键功能上完全缺失，导致无法用于真实工作负载。例如，它不支持 NVMe 或 CPU KV cache offload、工具解析、宽专家并行或分离式服务。这导致生产环境中没有任何客户使用它。与 NVIDIA 的 TRTLLM 不同——后者在 TogetherAI 等公司全球范围内每小时生成数十亿 token，[并且支持工具解析和其他功能](https://nvidia.github.io/TensorRT-LLM/commands/trtllm-serve/trtllm-serve.html#cmdoption-trtllm-serve-serve-tool_parser)——由于上述功能缺失，目前没有任何 token 工厂在使用 ATOM。
 
 此外，vLLM 等开源推理引擎的维护者对 AMD 感到失望，原因是 AMD 提供的工程和 GPU 资源不足。例如，vLLM 首席维护者 Simon Mo 在 GitHub RFC 中表示，他仍然没有可用的 MI355X 来添加到 vLLM CI 中，因此用户体验不佳。vLLM 上目前没有任何 MI355X 测试，而 NVIDIA 的 B200 在 vLLM 上有大量测试。同样，vLLM 上的 MI300X CI 机器数量仍然不够。上游 vLLM 至少还需要 20 台 MI300 机器、20 台 MI325 机器和 20 台 MI355X 机器才能达到与 CUDA 相同的可用性水平。
 
@@ -847,7 +847,7 @@ Microsoft/GitHub 自己也意识到了这个问题，已停止在其状态页面
 
 目前，InferenceX 使用完全随机的 token 作为基准测试的输入。然后我们在 [ISL*0.8, ISL] 分布下均匀变化 ISL/OSL，OSL 类似。由于使用随机数据，我们在所有基准测试中禁用了前缀缓存，因为完全随机数据的前缀缓存命中率期望值为 0%。此外，所有随机数据都是单轮的，意味着每个对话只包含一个提示和一个回答。虽然这提供了良好的基线帕累托前沿，但它不是模拟真实世界生产推理工作负载的实际基准测试方案。
 
-在近期，我们将使用类似 [allenai/WildChat-4.8M](https://huggingface.co/datasets/allenai/WildChat-4.8M) 这样的数据集创建基础多轮基准测试，该数据集记录了真实用户的多轮对话。除了在所有场景中启用前缀缓存外，我们还将启用 KV 缓存 CPU 卸载，因为这是我们在生产工作负载中看到的做法。这将更准确地评估每款芯片的优缺点。例如，MI355X 拥有 288GB HBM3e 而 B200 仅有 192GB。因此，我们预期 MI355X 在高并发多轮场景中表现更好，因为更多内存可以分配给 KV 缓存。另一方面，在 GPU KV 缓存紧张、块被卸载到 CPU 的场景中，我们预期 GB 系列表现更优，因为这些芯片拥有 900 GB/s 双向 CPU-GPU 带宽，相比之下 HGX 使用 PCIe 5.0 和 6.0 分别只有 128 GB/s 和 256 GB/s。此外，我们目前看到 AMD 的 CPU 卸载软件表现不佳，这可能在相同场景中对性能产生负面影响。
+在近期，我们将使用类似 [allenai/WildChat-4.8M](https://huggingface.co/datasets/allenai/WildChat-4.8M) 这样的数据集创建基础多轮基准测试，该数据集记录了真实用户的多轮对话。除了在所有场景中启用前缀缓存外，我们还将启用 KV 缓存 CPU offload，因为这是我们在生产工作负载中看到的做法。这将更准确地评估每款芯片的优缺点。例如，MI355X 拥有 288GB HBM3e 而 B200 仅有 192GB。因此，我们预期 MI355X 在高并发多轮场景中表现更好，因为更多内存可以分配给 KV 缓存。另一方面，在 GPU KV 缓存紧张、块被 offload 到 CPU 的场景中，我们预期 GB 系列表现更优，因为这些芯片拥有 900 GB/s 双向 CPU-GPU 带宽，相比之下 HGX 使用 PCIe 5.0 和 6.0 分别只有 128 GB/s 和 256 GB/s。此外，我们目前看到 AMD 的 CPU offload 软件表现不佳，这可能在相同场景中对性能产生负面影响。
 
 关键是：真实世界的多轮数据集测试了更多 SOTA 推理引擎功能，能够在所有芯片上捕获更细致和可靠的性能数据。
 
diff --git a/packages/app/content/blog/zh/kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72.mdx b/packages/app/content/blog/zh/kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
new file mode 100644
index 000000000..1e0a4651a
--- /dev/null
+++ b/packages/app/content/blog/zh/kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72.mdx
@@ -0,0 +1,70 @@
+---
+title: 'Kimi K3 的 AgentX 结果：MI355X ATOM 在部分曲线上胜过 GB300 NVL72'
+subtitle: 'AMD 的厂商引擎在每美元性能前沿上赢下了实打实的一段，而 Hopper 连服务 K3 都很吃力'
+seoDescription: '在 Kimi K3 的 AgentX 运行中，端到端延迟 40 到 60 秒的区间内，MI355X ATOM 在每美元性能上胜过 GB300 NVL72 vLLM；而 Hopper 表现吃力，因为 K3 需要定制的 SM90 kernel 以及调优过的 TP32、EP32 形状。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - kimi
+  - mi355x
+  - gb300
+  - nvl72
+  - h200
+  - amd
+  - nvidia
+---
+
+在端到端延迟 40 到 60 秒的那段曲线上，MI355X ATOM 的每美元性能甚至胜过 GB300 NVL72 vLLM。在一个 2.8 万亿参数的模型上、面对 NVIDIA 的机架级系统，这是 AMD 实打实赢来的结果。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/27a3f80e-7959-4651-90eb-f498989908a3_1934x1284.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+## 上游这个前提
+
+AMD 用 [ATOM](/zh/glossary/atom) 快速推进 K3 性能是件好事。与之相伴的建议是：优先把这些改进上游到 vLLM。ATOM 目前是 AMD 性能最好的引擎，但对使用上游开源服务栈的客户来说，vLLM 才是更有参考价值的对比对象；只在厂商运行时下成立的曲线，并不能代表这些客户实际能部署的性能。
+
+## Hopper 基本无法参与
+
+Hopper 很难承接 Kimi K3 的 AgentX 工作负载。一方面 Kimi 确实是个庞大的模型，另一方面 vLLM 维护者和 NVIDIA 一直没有把精力放在为 K3 优化 Hopper 上。
+
+具体原因是明确的，而非笼统的：Hopper（SM90）需要为 K3 定制调优 kernel，并针对高[交互性](/zh/glossary/interactivity)服务调优 TP32 和 EP32 的 shape。这两者目前都不存在，因此真正的约束与其说是硬件，不如说是没有人在做这件事。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ce2b58fc-b7b7-46c2-ae05-baa14419d233_2048x1459.png"
+  caption="来源：SemiAnalysis InferenceX"
+/>
+
+这提醒我们：基准测试图表上的代际差距，有一部分其实是工程投入上的代际差距。Hopper 多年来能很好地服务 DeepSeek R1，是因为那个模型持续获得 kernel 层面的投入；K3 没有得到同等对待，曲线也就如实反映了这一差别。
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 Kimi K3 上，AMD 有胜过 NVIDIA 的地方吗？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "有，在部分曲线上。在端到端延迟 40 到 60 秒的区间内，MI355X ATOM 的每美元性能甚至胜过 GB300 NVL72 vLLM。前提是 ATOM 属于 AMD 的厂商引擎而非上游开源软件栈，因此对多数客户而言 vLLM 仍是更有参考价值的对比对象。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Hopper 在 Kimi K3 上为什么吃力？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Kimi K3 体量巨大，而 vLLM 维护者和 NVIDIA 都没有把精力放在为它优化 Hopper 上。要在 SM90 上以高交互性服务 K3，需要定制调优的 kernel 以及调优过的 TP32、EP32 形状，而这些工作尚未完成。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/kimi-k3-the-manos-the-mythos-the.mdx b/packages/app/content/blog/zh/kimi-k3-the-manos-the-mythos-the.mdx
index b6f96556d..2dac3c1c3 100644
--- a/packages/app/content/blog/zh/kimi-k3-the-manos-the-mythos-the.mdx
+++ b/packages/app/content/blog/zh/kimi-k3-the-manos-the-mythos-the.mdx
@@ -681,12 +681,12 @@ QB 的做法是：在当前批次的截断阈值和路由结果下，求出一
 
 在 InferenceX 上，我们直接基于内部录制的 Claude Code 轨迹（trace）来评测 Kimi K3 的服务性能。我们回放其中达到稳态的一小时轨迹。每轮的输入 token 中位数为 142k，输出 token 中位数为 444，每个会话的轮数中位数为 65。每轮输出 token 偏短，是 agentic harness 类负载的典型特征——agent 频繁调用工具，连编辑操作也是工具调用。
 
-相比我们此前的 8k1k/1k1k 基准测试，这套基准是一次大的跃进，因为它真实反映了现实中的 agentic 使用场景。从系统角度看，它也更贴近生产系统，能够体现包括 prefix cache 和 KV 向 DRAM 卸载在内的 KV cache 行为。
+相比我们此前的 8k1k/1k1k 基准测试，这套基准是一次大的跃进，因为它真实反映了现实中的 agentic 使用场景。从系统角度看，它也更贴近生产系统，能够体现包括 prefix cache 和 KV 向 DRAM offload 在内的 KV cache 行为。
 
 <Figure
   src="/images/kimi-k3-the-manos-the-mythos-the/inferencex-agentic-trace-distributions.png"
   alt="agentic 轨迹数据集的分布直方图：每轮输入 token、每轮输出 token、每次请求未命中缓存的输入 token，以及每次会话的轮数"
-  caption="来源：[InferenceX](https://inferencex.semianalysis.com/datasets)"
+  caption="来源：[InferenceX](https://inferencex.semianalysis.com/zh/agentx)"
 />
 
 对 Kimi K3 来说，Day 0 的 bringup 比 DSv4 更轻松，因为权重发布前的文档和准备工作更充分。合适的镜像和投机解码模型与权重同时发布。
@@ -766,7 +766,7 @@ _本文后续内容——B300 并发扫描、AMD MI355X 与 MI455X 的 bringup
       "name": "InferenceX 如何评测 Kimi K3？",
       "acceptedAnswer": {
         "@type": "Answer",
-        "text": "使用内部录制的 Claude Code 真实轨迹，而非固定的 8k1k/1k1k 合成负载。我们回放一小时达到稳态的轨迹，每轮输入 token 中位数 142k、输出 token 中位数 444，每个会话轮数中位数 65。每轮输出偏短是 agentic harness 的典型特征，因为 agent 频繁调用工具。回放真实轨迹能够复现生产环境中的 KV cache 行为，包括 prefix cache 命中和 KV 向 DRAM 的卸载。"
+        "text": "使用内部录制的 Claude Code 真实轨迹，而非固定的 8k1k/1k1k 合成负载。我们回放一小时达到稳态的轨迹，每轮输入 token 中位数 142k、输出 token 中位数 444，每个会话轮数中位数 65。每轮输出偏短是 agentic harness 的典型特征，因为 agent 频繁调用工具。回放真实轨迹能够复现生产环境中的 KV cache 行为，包括 prefix cache 命中和 KV 向 DRAM offload。"
       }
     },
     {
diff --git a/packages/app/content/blog/zh/minimax-m3-agentx-b200-b300-vs-rack-scale.mdx b/packages/app/content/blog/zh/minimax-m3-agentx-b200-b300-vs-rack-scale.mdx
new file mode 100644
index 000000000..37164b1da
--- /dev/null
+++ b/packages/app/content/blog/zh/minimax-m3-agentx-b200-b300-vs-rack-scale.mdx
@@ -0,0 +1,75 @@
+---
+title: 'MiniMax M3 的 AgentX 结果：B200 与 B300 为何胜过机架级的 GB200 NVL72 与 GB300 NVL72'
+subtitle: 'Dynamo router 成为瓶颈，没有提交使用上下文并行，而 AMD 完全没有用上 KV offload'
+seoDescription: '在 MiniMax M3 上，B200 与 B300 在按 TCO 归一化的吞吐量上胜过 GB200 与 GB300。Dynamo router 可能成为瓶颈，因为其工作量随在线前缀增长；尽管 p90 输入长度达 317k，仍没有提交使用上下文并行；AMD 的 Pareto 最优点则完全没有使用 KV offload。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - minimax
+  - b200
+  - b300
+  - gb200
+  - gb300
+  - nvl72
+  - dynamo
+  - amd
+  - rocm
+---
+
+在 MiniMax M3 上，按 TCO 归一化的吞吐量方面，B200 与 B300 完全胜过各自对应的机架级方案。对于一个由 NVIDIA 横扫的模型来说，这个内部结果颇令人意外，而原因并不在网络。
+
+在 AgentX 上，机架级的优势不那么明显，因为 Dynamo router 可能成为瓶颈：它的工作量随在线前缀的数量和长度增长，而不是随生成的 token 数增长。此外，这个模型上还没有针对宽专家并行、宽 DCP 或宽张量并行调优良好的 kernel。而 GB200 与 GB300 的 TCO 更高，在没有宽 EP 或宽 DCP 的情况下，这就直接表现为更差的每 TCO 性能。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/aec34e63-34fd-4c42-87a1-97e5c14661ed_2048x1302.png" />
+
+我们也预期 NVIDIA 会在该 SKU 的机架级方案上做进一步优化，并会在后续文章中重点介绍。
+
+## 没有人在使用上下文并行
+
+尽管 p90 输入序列长度达到 317k token，目前仍没有任何提交使用[上下文并行](/zh/glossary/context-parallelism)。这恰恰是该技术存在的意义所在，因此这种缺席需要解释。
+
+M3 只有 4 个 KV head，即使在 TP8 下 DCP 也最多只能到 2；而且 MSA indexer 还需要单独的上下文并行处理，相关 vLLM PR 已经提交。因此这个上限是架构层面的，而不是调优疏忽，它也限定了长上下文并行策略在这个模型上能带来多少收益。
+
+## AMD 的 offload 缺口
+
+NVIDIA 所有 Pareto 最优点在并发 20 以上都启用了 [KV offload](/zh/glossary/kv-cache-offload)。而 AMD 的 Pareto 最优点没有一个使用 DRAM KV offload；在其他模型上，AMD 使用 offload 的比例也低于 NVIDIA。
+
+原因很机械：AMD vLLM 上 CPU KV cache offload 所需的 GPU 到 CPU 传输效率极低，因为 hipMemcpyBatchAsync API 直到 ROCm 7.14 才提供。没有它，vLLM 原生的 simple CPU offloading 只能做串行 memcpy，而无法把多次拷贝合并成更大的消息。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2b40624a-223c-44db-a403-0be563b2af82_2048x1298.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=interactivity&i_best=0&i_hc=1&i_active=mi355x_vllm)"
+/>
+
+缺少一个批量拷贝 API，这种事不会出现在规格书上，在定长基准测试中也根本不会被触发。但在长时间的 agentic 会话中，它决定了整整一层内存是否可用。
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 MiniMax M3 上，B200 与 B300 为什么胜过 GB200 与 GB300？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在按 TCO 归一化的吞吐量上，AgentX 中机架级的优势被削弱了：Dynamo router 可能成为瓶颈，因为它的工作量随在线前缀的数量和长度增长，而不是随生成的 token 数增长。此外，该模型上还没有针对宽专家并行、宽 DCP 或宽张量并行调优良好的 kernel，缺少这些能力时，GB200 与 GB300 更高的 TCO 就赚不回来。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AMD 在 MiniMax M3 上为什么没有使用 KV offload？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在 AMD vLLM 上，CPU KV cache offload 所需的 GPU 到 CPU 传输效率极低，因为 hipMemcpyBatchAsync API 直到 ROCm 7.14 才提供。没有它，原生的 simple CPU offloading 只能做串行 memcpy 而无法合并成更大的消息；因此 AMD 没有任何 Pareto 最优点使用 DRAM offload，而 NVIDIA 在并发 20 以上的每个最优点都用了。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/minimax-m3-agentx-b300-trtllm-tp2-vs-the-field.mdx b/packages/app/content/blog/zh/minimax-m3-agentx-b300-trtllm-tp2-vs-the-field.mdx
new file mode 100644
index 000000000..0a7323dc5
--- /dev/null
+++ b/packages/app/content/blog/zh/minimax-m3-agentx-b300-trtllm-tp2-vs-the-field.mdx
@@ -0,0 +1,78 @@
+---
+title: 'MiniMax M3 的 AgentX 结果：B300 TRT-LLM TP2 拿下桂冠'
+subtitle: 'NVIDIA 横扫这个 432B 模型，而缺失的 DP attention 数据点说明了缓存局部性为何变成路由约束'
+seoDescription: 'MiniMax M3 432B 的 AgentX 结果。B300 TRT-LLM TP2 领跑前沿曲线；由于 KV cache 局部性变成路由约束，曲线上没有 DP attention 数据点；GB200 上 TP4/EP4/DPA 的吞吐量只有纯 TP4 的 0.60 倍，而 p90 TTFT 高出 3 倍以上。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - minimax
+  - b300
+  - b200
+  - gb200
+  - trtllm
+  - vllm
+  - nvidia
+---
+
+在 MiniMax M3 432B 上，NVIDIA 完胜所有对手。AMD 的软件性能非常糟糕，长上下文下尤其如此；我们认为根源在于 AMD 工程管理层只激励团队针对短上下文单轮工作负载做调优，而忽视了长上下文多轮工作负载。
+
+在 NVIDIA 阵营内部，B300 TRT-LLM 在 TP2 下拿下了 M3 的桂冠。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/5179bd8e-c147-46c8-aeb4-b347049b28a7_2048x1303.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft)"
+/>
+
+## 缺失的 DP attention 数据点
+
+M3 的前沿曲线上没有 [DP attention](/zh/glossary/dp-attention) 数据点，而这种缺席本身就说明了问题：在 M3 上，数据并行注意力并不划算，因为 KV cache 局部性会变成路由约束。
+
+具体数字让机制变得清晰。以并发 40 的 GB200 为例，TP4/EP4/DPA 的吞吐量只有纯 TP4 的 0.60 倍，而 p90 TTFT 高出 3 倍以上；并发 32 时缓存命中率仅 28.8%，理论值则是 96.0%。每个 DP rank 独占缓存池的四分之一，一个 30 万 token 的会话一旦落到错误的 rank 上，就要全部重算。
+
+28.8% 实测与 96.0% 理论之间的这道鸿沟，正是数据并行下[前缀复用](/zh/glossary/prefix-cache-hit-rate)的全部症结：缓存本身足够大，只是会话没有被发到持有它的那个 rank 上。
+
+M3 的前沿曲线上也没有出现任何带专家并行的 decode 配置，可能是因为并发还不足以让所有专家的负载均衡起来。
+
+## vLLM 与 TRT-LLM 的对比
+
+值得一提的是，从吞吐量与 p90 交互性的关系看，vLLM 的表现与 TRT-LLM 非常接近；而在吞吐量与 p90 TTFT 的关系上，vLLM 表现更好。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/729199d9-66c0-42b5-b1b8-f228503c1585_2048x1289.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=MiniMax-M3&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_active=b200_trt%2Cb200_vllm%2Cb300_trt%2Cb300_vllm&i_hc=1&i_label=0)"
+/>
+
+因此更倾向于哪一个，取决于你的产品是按哪条坐标轴来定义的。这也再次说明：应当把交互性和首 token 延迟放在一起读，而不是挑一张更符合预设结论的图。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/68f7ad68-3886-4980-9b01-390582728e47_2048x1292.png" />
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 MiniMax M3 上哪种配置最快？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "B300 TRT-LLM 在 TP2 下拿下 M3 桂冠。NVIDIA 在该模型上整体领先，而 AMD 的软件性能表现较差，长上下文场景尤其明显。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "M3 前沿曲线上为什么没有数据并行注意力的数据点？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在 M3 上 DP attention 并不划算，因为 KV cache 局部性会变成路由约束。在并发 40 的 GB200 上，TP4/EP4/DPA 的吞吐量只有纯 TP4 的 0.60 倍，p90 TTFT 高出 3 倍以上；并发 32 时实测缓存命中率为 28.8%，而理论上限为 96.0%，原因是每个 DP rank 只拥有缓存池的四分之一，会话落到错误的 rank 上就要全部重算。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/openai-jalapeno-better-than-nvidia.mdx b/packages/app/content/blog/zh/openai-jalapeno-better-than-nvidia.mdx
new file mode 100644
index 000000000..5daf684a3
--- /dev/null
+++ b/packages/app/content/blog/zh/openai-jalapeno-better-than-nvidia.mdx
@@ -0,0 +1,383 @@
+---
+title: 'OpenAI Jalapeño：比 NVIDIA Blackwell 更强'
+subtitle: 'OpenAI 自研 ASIC 对比 Rubin：Jalapeño 的 TCO、每兆瓦吞吐量，以及那些够辣的细节'
+seoDescription: 'OpenAI 首代自研推理 ASIC，与 Broadcom 合作打造，并在 InferenceX 上完成基准测试。文中对比它与 Blackwell、Vera Rubin 的每兆瓦性能，介绍 HBM4 与 MXFP4 规格、Gluon 软件栈、OpenAI 为何不做 prefill-decode 分离，以及 Katsu、Vindaloo 与 Chana 组成的机架架构。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - inference
+  - benchmark
+  - gpu
+  - nvidia
+  - openai
+  - asic
+  - rubin
+---
+
+过去两年，OpenAI 一直在低调打造一款名为「Jalapeño」的推理芯片，如今在 Hot Chips 上正式发布。此前关于它成功流片的传闻已经流传了一段时间，现在细节终于公开。OpenAI 邀请我们实地查看这颗芯片、走进他们的实验室确认它的真实性，并用我们的 [InferenceX](https://inferencex.semianalysis.com/) 套件对它做了[基准测试](https://openai.com/index/jalapeno-first-results/)。
+
+今年 6 月，[OpenAI 公布了这项芯片计划](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/)，与 Broadcom 合作，从零开始、专为 LLM 推理设计。[设计工作始于 2024 年年中](https://newsletter.semianalysis.com/p/openai-chip-team-is-now-serious)，从组建团队到量产流片只用了约 16 个月，这在 ASIC 开发中是极快的节奏。
+
+通常第一代芯片并不具备竞争力，但 OpenAI 打破了这一规律：在多个主流开源模型上，它击败了我们测试过的每一款 NVIDIA、AMD 和 Google 芯片，处于业界领先水平。OpenAI 靠的是极致的软硬件协同设计。令人意外的是，他们并没有针对推理的某个特定环节做过度特化，而是专注于做一颗在所有场景下都有高性能的通用芯片。
+
+本文将深入介绍 Jalapeño 的架构细节、软件细节，以及它在 InferenceX 上的性能结果。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c028f756-4748-4583-91b2-5c4a63893cb5_2048x1153.png"
+  caption="来源：[OpenAI](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/)"
+/>
+
+## 一颗通用的推理芯片
+
+外界普遍认为 OpenAI 的芯片是为自家模型特化的，但这是错的：OpenAI 做的是一颗面向 AI 推理的通用芯片。
+
+这个时间线堪称疯狂，说明「用 AI 加速芯片设计」的说法确有其事。抛开进度不谈，OpenAI 投入了大量资金、做出了务实的设计取舍，团队本身也很能打，因此这个结果并不令人意外。
+
+仅看规格，它就已经是有力的竞争者：
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/e346a4e5-76cb-4fd9-be95-00321116b605_1846x510.png"
+  caption="来源：SemiAnalysis"
+/>
+
+而采用 HBM4，使它足以与 NVIDIA 和 AMD 的旗舰 GPU 相提并论：
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/327f95f7-ab6b-44ae-9b8f-17c1ff5d296b_1712x1300.png"
+  caption="来源：OpenAI"
+/>
+
+关于这颗芯片的媒体报道，大多沿用了 OpenAI 随口提到的几句话，认为它会以其他芯片做不到的方式为自家模型做优化。这种说法是错的。Jalapeño 是一颗通用推理芯片，能够运行各类模型和各类工作负载，也包括我们的 InferenceX 基准测试——我们正是与 OpenAI 工程师一起在实验室里跑的。他们还开玩笑似地展示了在这颗芯片上运行 Doom，而移植工作只用 Codex 提示词就完成了。
+
+下面是我们的核心 perf/W 结果，看的是每兆瓦全量供电所能产出的 token 吞吐量。**Jalapeño 碾压了其他所有芯片**。而且这一切都是在不使用多 token 预测（MTP）的前提下取得的，图中其他芯片则都是各自 SKU 上开启 MTP 的最佳配置。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/19a7f45a-df8e-436e-ba04-df5d8610f3da_2048x1330.png"
+  caption="来源：SemiAnalysis"
+/>
+
+在几乎所有场景下，Jalapeño 的 perf/W 都胜过 Blackwell，而且并没有针对曲线上的某个点做专门调优。它不仅在低延迟场景表现出色，在高吞吐场景同样如此。更对等的比较是与单 token 预测的结果对比，此时它把所有竞争对手都甩在身后。在低并发场景下，Jalapeño 展现出惊人的交互性：在 DeepSeek R1 上，并发为 1 时超过 700 tok/s/user。
+
+更惊人的是，这一切都是在单 token 预测（STP）下完成的，没有投机解码，也没有 prefill 与 decode 分离。除 DeepSeek R1 外，我们还看到了其他一些模型，包括 Kimi-K2.5 和 GPT-OSS，后者约为 1,400 tok/s/user。对所有这些模型，我们确认 Jalapeño 的 GSM8k 评估结果与 NVIDIA 芯片相当。
+
+有几点需要说明。第一，所有数据均由 OpenAI 提供。我们在实验室现场验证了 InferenceX 的运行过程，但没有跑完整套 [InferenceX](https://inferencex.semianalysis.com/) 基准测试，也没有看到 [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) 结果。AgentX 是我们更倾向用于比较芯片性能的套件，因为其数据集具备长上下文和多轮特征，能反映真实生产流程中的缓存行为。在 8k1k 上表现良好的框架，在 AgentX 上可能明显变差，因为真实生产负载会给 router、前缀缓存机制、缓存管理、offload 基础设施等组件施加压力，而单轮 8k1k 根本考察不到这些。相关内容可参见我们的 AgentX 文章。
+
+[AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) — 自 2025 年 11 月的 Claude Code 拐点以来，长上下文、多轮的 agentic 工作负载快速增长，如今已经主导生产推理的流量。2026 年 4 月，OpenAI 的企业级 agentic 支出超过了 ChatGPT 支出。
+
+第二，我们认为拿它和 Blackwell 比并不完整，也不够公平。Jalapeño 真正的对手是同样使用 HBM4 的 Rubin 这类芯片。Vera Rubin 系统现在已经开始向客户发货，而 OpenAI 手上除了工程样片之外还需要一段时间才能有更多东西。
+
+因此，性能比较的对象应该是 Rubin 而不是 Blackwell；从某种意义上说，我们本就预期 Jalapeño 这样的定制芯片会胜过 Blackwell。Vera Rubin NVL72 的 perf/MW 是 GB200 NVL72 的 5.4 倍，[我们在分析 NVIDIA 上月与 CoreWeave 联合发布时所给出性能数据的文章中已有说明](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference)。下文我们会把 Jalapeño 与 Vera Rubin 7 月的性能数据做对比。
+
+[Vera Rubin NVL72 vs GB200 NVL72? Inference TCO & Architecture Analysis](https://newsletter.semianalysis.com/p/vera-rubin-nvl72-vs-gb200-nvl72-inference) — Vera Rubin NVL72 是 NVIDIA 机架级 Oberon 架构的第二代产品，其推理性能提升来自极致的协同设计。工程样片的早期结果令人鼓舞：运行 DeepSeek R1 的 Vera Rubin NVL72 相比 GB200 NVL72 目前有 5.4 倍的每兆瓦性能和 5 倍的每美元性能，而与 2025 年 GB200 NVL72 早期 bring-up 阶段相比差距更大。Vera Rubin 目前仍处于早期 bring-up 阶段，我们预计差距还会继续拉大。随着软件成熟，Rubin 的推理性能会持续提升，这与我们此前在 Blackwell 上观察到的规律一致。
+
+第三，被测试的模型并不处于开源前沿。NVIDIA 和 AMD 已经用 AgentX 公布了 DeepSeek V4 Pro、Kimi K3 等更大模型的结果。模型越大、发布越新，在一颗新芯片上完成 bring-up 就越复杂。话虽如此，OpenAI 在 Jalapeño 上跑通的这些模型也算不上小。
+
+## 性能分析
+
+OpenAI 是按 perf/W 来做设计的。原因很简单：OpenAI 当前的瓶颈是数据中心电力，而不是预算或场地，因此每兆瓦 token 数至关重要。在 Computex 2026 上，黄仁勋表示 perf/W、可靠性和长使用寿命是未来 GPU 的核心特性。用他的原话说：「如果你有 1 吉瓦电力，那么每瓦吞吐量就是营收。」他还提到，仅仅因为芯片更便宜就选择错误的架构是说不通的。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/69dc13f6-ffb1-4d56-8df7-b75b57280382_1980x1254.png"
+  caption="来源：Computex 2026 主题演讲"
+/>
+
+NVIDIA 在 Hot Chips 2026 的 Vera 演讲中展示同一张营收图时也强调了这一点：「今天的数据中心是电力受限的。」电力很重要，而且直接驱动营收。
+
+运营方无法简单地获取更多兆瓦，因为增加 GPU 与增加电网容量的时间尺度完全不同。数据中心的电力上限受制于并网接入、基础设施、制冷能力以及 UPS 与备用发电设计。电网审批的延误一再落后于硬件与建设进度，这催生了对 BtM（表后）电力容量的需求：直接建在数据中心现场的燃气轮机和自备发电机。这部分容量位于电网计费表之后，而不是从公共电网取电，因此运营方无需等待并网和电力扩容就能给园区供电——这正是 xAI 的 Colossus 2 高度依赖 BtM、而实际并网进度远远滞后的原因。更多内容参见我们的[能源模型](https://semianalysis.com/energy-model/)。
+
+正如我们在 X 上写过的，由于瓦特就是焦耳每秒，tok/s/MW 可以约化为每焦耳 token 数。这使 tok/s/MW 能够代表一套系统把能量转化为 token 的效率与能力。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c1a1da73-425c-4800-b396-01c7d2151be9_1200x1432.png"
+  caption="来源：[SemiAnalysis](https://x.com/SemiAnalysis_/status/2091208991271420076)"
+/>
+
+在这一点上，即便与 Rubin 相比，Jalapeño 也胜出。OpenAI 的 Jalapeño 在 STP 下的每兆瓦输出 token 吞吐量，超过了 [NVIDIA 与 CoreWeave 在 7 月公布的 Vera Rubin MTP 结果，也远超 GB200 在 2025 年的 MTP 结果](https://www.coreweave.com/blog/nvidia-vera-rubin-nvl72-on-coreweave-10x-more-tokens-per-megawatt-than-blackwell)。正如我们在 Vera Rubin 文章中所说，之所以拿 VR 与 2025 年的 GB200 结果对比，是因为二者处于类似的早期 bring-up 阶段，这样可以保持软件成熟度这一变量不变。沿用同样的逻辑，我们把 Vera Rubin 2026 年 7 月的最新结果、GB200 2025 年的结果与今天的 Jalapeño 结果放在一起比较。这是很合理的对比：这些是目前公开的最佳 Rubin 数据，而 OpenAI 的流片时间还在 Rubin 之后。OpenAI 和 Rubin 都还不成熟，性能都会继续提升。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/6e8f9fd2-ec7f-45fa-80b9-dc132e32c661_2048x1450.png"
+  caption="来源：OpenAI、SemiAnalysis"
+/>
+
+在 perf/TCO 上，Vera Rubin 与 Jalapeño 势均力敌，每美元产出的输出 token 数几乎相同。但如前所述，**Jalapeño 的结果是在不使用投机解码的前提下取得的**，而 Vera Rubin 的结果用了投机解码。投机解码可以把单 token 成本降低约 3 到 5 倍。等 Jalapeño 用上投机解码之后，它的 token 服务成本还会更低。当然，这部分 TCO 优势来自用 Broadcom 相对更低（尽管依然很高）的毛利替换了 NVIDIA 的高毛利，但这并不是全部原因。举例来说，Meta 和 Microsoft 的 AI ASIC 项目投入时间更久却始终没能真正跑起来，这说明成本只是方程式的一部分。Jalapeño 的完整 TCO 拆解见 [SemiAnalysis AI Cloud TCO 模型](https://semianalysis.com/ai-cloud-tco-model/)。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a6c58331-c05a-4823-9c4b-aae90705bb53_2048x1485.png"
+  caption="来源：OpenAI、SemiAnalysis"
+/>
+
+在架构上，OpenAI 选择不把 prefill 与 decode（PD）分离到不同的芯片池。draft 模型与主模型共用同一批芯片和同一张网络，这种设计理念用一部分理论效率换取了运维上的实用性。理由在于工作负载的构成会随时间变化：随着我们走过三个模型时代（[知识、推理和 agentic，见我们近期的文章](https://newsletter.semianalysis.com/p/are-open-models-catching-up)），输入 token、缓存写入、缓存读取与输出 token 之间的比例已经发生了显著变化。因此，事先固定异构的 prefill 硅片与 decode 硅片比例，长期来看会带来低效。OpenAI 在这套架构中选择了同构资源池，并努力让芯片在所有场景下都表现良好。
+
+事实也确实如此。在 Kimi K2.5（Cursor Composer 2.5 基于该模型）上，Jalapeño 达到接近 700 tok/s/user，在 100 tok/s/user 处更是次优芯片的 9 倍以上。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/3088f64c-9bee-46d9-a445-6818248ba573_2048x1366.png"
+  caption="来源：OpenAI、SemiAnalysis"
+/>
+
+在 GPT-OSS 上又是一边倒。Jalapeño 在等交互性下的每兆瓦吞吐量接近 GB200 最高吞吐点的两倍，是 GB200 并发为 1 时的 50 倍以上。Jalapeño 并发较高的那些点使用了 EP8。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9f245dd7-c6b7-48a3-9215-3976633bb77f_2048x1366.png"
+  caption="来源：OpenAI、SemiAnalysis"
+/>
+
+这些结果令人印象深刻！不过我们还是要挑刺：它们只是 8k1k，是一个明显更容易调优的工作负载，而且目前还没有 [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) 运行结果。正如我们在 [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) 文章中所说，多轮长上下文工作负载会给服务栈的更多环节施加压力，例如 router 和前缀缓存。要在 agentic 工作负载上取得优异表现，还需要大量优化。更多内容请见 [AgentX](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) 文章。
+
+[AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing?](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat) — 自 2025 年 11 月的 Claude Code 拐点以来，长上下文、多轮的 agentic 工作负载快速增长，如今已经主导生产推理的流量。2026 年 4 月，OpenAI 的企业级 agentic 支出超过了 ChatGPT 支出。
+
+## 深入规格与架构
+
+上述所有结果都是在 Jalapeño 的 A0 步进上取得的，距离项目启动仅 9 个月。而 B0 步进已经在 fab 中流片，其优化相比早期的 A0 硅片带来约 25% 的每瓦性能提升。具体来说，B0 步进在采用台积电 N3P 工艺、单颗满掩模尺寸的计算 die 上提供 13.4 PFLOPs 的 MXFP4 算力。作为对比，尺寸相近、同一工艺节点的单颗 Rubin 计算 die 提供 17.5 PFLOPs 的稠密 NVFP4 算力。
+
+考虑到 Jalapeño 的 TDP 只有 700W，而 Rubin 每颗计算 die 为 900 至 1,150W，这个数字就相当可观了。由于 Jalapeño 面向推理而非训练，OpenAI 不需要把 TDP 推高来榨取 FLOPs，这可以理解；但无论如何，上述数据说明 Jalapeño 的峰值理论算力是拿得出手的。
+
+与其他加速器直接对比时，Jalapeño 拥有最高的每瓦 HBM 带宽和最高的每瓦 FLOPs，与 1,800W 的 Rubin Max-Q 配置相当：
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/c5795c45-d23b-403c-8bc4-f13d5ba6fb45_2048x394.png"
+  caption="来源：SemiAnalysis"
+/>
+
+片外 I/O 由一颗 N3E I/O chiplet 提供，具备 32 条 800G SerDes 通道用于计算网络，其中 24 条（600GB/s）用于机架内的本地 scale-up，8 条（200GB/s）用于全局 scale-up，即 2,048 颗 XPU 的多机架域。系统 I/O 使用 PCIe Gen 5 连接 x86 主机 CPU。
+
+Jalapeño 将搭载 HBM4，使它成为继 NVIDIA 和 AMD 之后相对较早的采用者，甚至领先于已经成熟的 TPU 和 Trainium 项目。既然 Jalapeño 的核心架构原则之一是把 HBM 带宽用到极致，那么选择次优的 HBM 就与这一目标背道而驰。最终它实现了每封装 15.4TB/s 的显存带宽，超过所有在售的 HBM3E 加速器。15.4TB/s 说明其 HBM4 达到了 10Gbps 的引脚速率，略高于 NVIDIA 在 Rubin 上跑出的 9.6Gbps。这批 HBM 很可能由三星提供。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/0d7d181a-2bb9-45d4-bd67-2f7e525ce6d1_1426x376.png"
+  caption="来源：OpenAI"
+/>
+
+OpenAI 在 2025 年 11 月完成了 Jalapeño 的流片，更准确地说，这是 CoWoS 设计的流片，而不只是顶层 die 的硅片流片。在那次流片之后的 9 个月内、且只有 3 个月的实际硅片 bring-up 时间里，OpenAI 就用 Jalapeño 拿出了非常好的结果。考虑到团队的软件栈是从零起步，这就更令人印象深刻了。
+
+与之相对，Rubin 的 CoWoS 流片在 2025 年 10 月完成，早了一个月，但我们目前看到的早期结果只有 CoreWeave 的工程样片。NVIDIA 并没有像 OpenAI 这样让我们测试并发布基准结果，这说明他们的芯片软件仍不成熟。考虑到 OpenAI 在自家硅片上 bring-up 新模型的速度，CUDA 护城河有可能已经不复存在。
+
+它们都还远未优化到位，但总体上可以看到 Jalapeño 交出了更好的数据。我们并不认为 NVIDIA 的硬件更差，更多是 Jalapeño 的软件 bring-up 推进得比 NVIDIA 更快。这体现了软硬件协同设计的威力，而这正是一支能打的前沿实验室 ASIC 团队相比更成熟的商用芯片厂商最可能胜出的地方。反直觉的是，从零开始或许也帮了 OpenAI：他们可以做出干净利落的架构决策，而不必顾虑向后兼容或旧版软件。
+
+虽然 OpenAI 目前只有 Jalapeño 的工程样片，但量产计划将在 2027 年逐步爬坡，大部分产出目前排在明年年底。[关于出货量与 ASP 的更多细节，请见 SemiAnalysis 加速器模型](https://semianalysis.com/accelerator-hbm-model/)。
+
+可以说，OpenAI Jalapeno 是一款真正会大批量出货的 ASIC。
+
+与 Rubin 的时间线相比，Jalapeño 的推进速度快得惊人。如前所示，尽管 Rubin 起步更早，Jalapeño 的结果依然胜出。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/bb59cacc-1f16-4c6a-98a5-0f0db74dada7_2048x411.png"
+  caption="来源：SemiAnalysis"
+/>
+
+## Jalapeno 架构
+
+接下来深入架构。这颗芯片的矩阵引擎采用 MXFP 数值格式和权重驻留（weight stationary）的脉动阵列，与 TPU 类似。但与 TPU 直接对比时，它支持更小的形状与维度，这意味着不会像更大的脉动阵列那样，在遇到形状不规整的矩阵乘法时出现奇怪的性能悬崖。
+
+它还配有 64 位标量核心以及 FP32/INT32 向量核心。OpenAI 在托盘级别投入了冗余设计，并在核心与通道级别内置了良率收割机制。他们表示，AI 辅助设计在设计阶段使 SIMD 面积减少 8%、矩阵引擎面积减少 10%。虽然没有说明具体的工艺、电压、温度（PVT）条件，但他们也提到 AI 辅助生成的模块在时序和功耗上都优于最初版本。
+
+Jalapeño 的架构设计重点在于消除 KVCache 与权重的内存搬运，以及固定延迟和各类开销，从而即便在小 batch 或小形状下，也能比其他加速器更接近原始峰值算力与带宽。
+
+核心与 HBM 被划分为多个 slice，每个核心 slice 对自己那一份 HBM 具有低延迟的本地视图。slice 之间的同步走一张专用的高带宽集合通信网络。这种极简的内存层级结构，已经让 Jalapeño 相对 GPU 具备了很大的潜在优势：在 GPU 上，内存访问必须穿过复杂的内存系统，产生的高延迟只能靠更大的形状去摊薄或掩盖。
+
+之所以可行，是因为只要仔细安排权重和 KV 的放置，核心之间的同步就可以被限制在有限且已知的高带宽通信上，例如可以与计算重叠的张量并行通信。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/871631ee-8f0d-4fb6-869c-a6fb298e800e_2048x832.png"
+  caption="来源：OpenAI"
+/>
+
+此外还有一张通用 NoC，用于常规通信以及访问 scale-up 网络。总体而言，相比 NVIDIA 和 Google，OpenAI 依靠简化的 NoC 与内存子系统节省了大量功耗，并获得了可观的性能提升。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/1f93df86-2ea2-4904-8fd1-e0e5eff4b410_1638x854.png"
+  caption="来源：OpenAI"
+/>
+
+在核心层面，OpenAI 描述的是带 L1 缓存的乱序（OoO）核心。这与我们在其他加速器上看到的模式差异很大：后者普遍采用软件管理的 scratchpad，并搭配某种异步 DMA 支持。同样，这里的论点是它能让 Jalapeño 避免诸如 barrier 延迟这类固定开销；而在 GPU 等其他加速器上，这些开销必须靠给每个核心分配更多工作来掩盖或摊薄，也因此更难逼近原始的峰值带宽与算力。
+
+代价是 Jalapeño 因而更依赖良好的预取来保证内存请求及时到达，而这一点可预测性更差、也更难推理。不过，只要 Codex 处在一个好的 harness 中并能拿到详细 trace，为给定形状找到预取最优的 kernel 很可能几乎不需要人工介入。我们认为 OpenAI 正是这样做的，才得以如此迅速地跑通 DeepSeek R1、Kimi K2.5 和 GPT-OSS。
+
+这些核心还支持「小」矩阵维度。视其小到什么程度，这应当能让它在不同模型和 batch 维度上更通用，也更不易受矩阵维度对齐、padding 开销和 tiling 低效的影响。相比之下，TPU、Trainium 和 Etched 的芯片都采用很大的脉动阵列，往往需要较大 batch 或恰好可整除的模型维度，才能避免 tiling 带来的浪费。
+
+在 Jalapeño 上，OpenAI 专注于消除系统中的固定延迟，以便在 Pareto 曲线的各个区间都尽可能逼近 roofline 性能。理论上，这可能在多个运行点上带来相对 GPU 的优势：
+
+- 在低延迟 / 小 batch 推理上有高得多的性能上界，而 GPU 在这里受制于诸多固定开销，例如启动延迟、barrier 延迟和内存系统延迟
+- 在大 batch 或长上下文场景下，也有一定机会更接近硬件 roofline
+
+需要说明的是，即便理论上存在这样的性能上界，对真实 kernel 而言未必容易兑现。因此这套思路大致是：
+
+1. 面向所有工作负载形状，设计出尽可能高的性能上界
+2. 让 Codex 去做那些繁琐的工作，找出能够逼近该上界的 kernel
+
+从 OpenAI 团队在 Jalapeño 上跑通 InferenceX 工作负载的极快速度来看，我们对这条路线持乐观态度。
+
+如果 Jalapeño 成功，那将是一个强烈信号：业界对编程模型和完美通用编译器的执念，正在被前沿 AI 模型所瓦解。
+
+## 软件
+
+OpenAI 写 Jalapeño 的 kernel 就像写汇编。每个 kernel 都是手工调优的代码，有些长达约 3,000 行，并配有正确性检查和自研 sanitizer。早期的 kernel 工作以人在回路为主而非完全自动化，但随着内部一个规模更大的 Codex 版本投入使用，这一点发生了变化——OpenAI 计划把该版本推向企业客户。内部的服务引擎名为「Teacup」。有意思的是，**在用 InferenceX 对 DeepSeek 做基准测试之前，OpenAI 内部并没有 MLA kernel 的实现**。Codex 能在 OpenAI kernel 工程团队完全不介入的情况下这么快写出可用且高效的 kernel，足见其软件流水线的开发能力。
+
+OpenAI 用 Gluon 为 Jalapeño 编程。Gluon 是 OpenAI 的 kernel 编程语言，构建在 Triton 之上，保留了 Triton 的 SPMD（单程序多数据）编程模型，同时暴露出**更底层的编程抽象**。例如在 NVIDIA GPU 上，它提供映射到 PTX 指令的 API，包括 MMA 指令、TMA 指令、mbarrier 机制等等。Gluon 最独特的抽象是 **layout**。一般来说，layout 定义的是硬件资源（例如 warp 9 的第 5 个寄存器）与张量元素（例如第 6 行第 7 列的元素）之间的映射。Gluon 的 layout 抽象基于 [Linear Layouts](https://arxiv.org/abs/2505.23819)，这是 OpenAI 提出的一套 layout 代数，用数学方式形式化了 layout 的定义，并提供了在 layout 上做运算的工具。由此可以实现许多能力，例如可证明正确的 layout 转换和最优的内存 swizzling。
+
+就 Jalapeño 的编程模型而言，每个 Gluon 程序对应一个常驻线程。我们认为这暗示 Jalapeño 适合**常驻 kernel 编程模式**：每个程序处理多个 tile，且由程序员而非硬件调度器来分配工作。OpenAI 提到了 TensorInfo，这是一种显式编码 layout 的抽象，很可能就是为 Jalapeño 设计、并由 Linear Layouts 驱动的那套 layout。最后，每个核心都提供数据预取和解耦的乱序单元，例如用户可以对已预取的数据编写等待逻辑，由信号量控制解锁。
+
+命运的奇妙之处在于：像 GPT 5.6 Sol 这样目前跑在 NVIDIA GPU 上的 OpenAI 模型，被用来设计了一颗真正威胁 CUDA 护城河的芯片——NVIDIA 自己的 GPU 正在实时地帮助催生它潜在的接班者。
+
+把时间轴拉开看，还能看到 Jalapeño 的开发节奏：在不到两周的时间里，某些交互性区间的吞吐量提升超过 2 倍。Jalapeño 团队每次发来的压缩包里都藏着惊喜。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/b38d958b-6d53-4ebe-b92f-c2cb58de05c2_2048x1485.png"
+  caption="来源：OpenAI、SemiAnalysis"
+/>
+
+不只是 kernel 性能提升。在 8 天之内，Jalapeño 团队在此前 TP8 配置的基础上启用了 TP32，从单机扩展到完整的机架级配置来运行大模型。这样的开发速度确实令人印象深刻。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/a18d5611-f7c1-4e34-8919-99b1c683054f_2048x1485.png"
+  caption="来源：OpenAI、SemiAnalysis"
+/>
+
+为了在投入真实硬件运行之前验证性能，OpenAI 还有一个名为「chilisim」的模拟器，采用定宽 trace 总线，与实测硬件的误差在 5% 以内。A0 上的 tracing 能力有限，但在 B0 上有了大幅改进，很可能吸收了 A0 实际运行的反馈。工程师还演示了 Codex CLI 运行一个内部模型（代号「Raiku」或「5.3 Codex Spark」），TPOT 为 1.2ms。
+
+团队还展示了由 Codex 编写、直接在芯片上运行的演示：36 FPS 的 Doom、一个 FP32 流体力学模拟，以及一个名为「Liquid Light」的鼠标拖拽可视化。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/d410878f-5f19-40bd-ad2c-d92ad45d65a9_1252x1232.png"
+  caption="来源：SemiAnalysis"
+/>
+
+在模型侧，OpenAI 内部的 megakernel 方案（代号「gigakernel」）围绕单个 megakernel 构建，在设备上循环执行，以减少 CPU 开销和启动时间。团队还在进一步押注测试时计算策略，内部尤其关注如何连贯地使用 100 万次 rollout。
+
+## 分离还是不分离，这是个问题
+
+前面提到，OpenAI 在这些芯片上没有使用 prefill decode 分离。这让我们相当意外，因为 NVIDIA 和 AMD 的 GPU 即便在同构硬件上也能从 PDD 中获得显著收益。下面我们来看看 Jalapeño 团队为什么这样选。
+
+当工作负载固定不变时，prefill-decode 分离（PDD）看起来很有吸引力。prefill 与 decode 对硬件的压力方式不同，因此把两个阶段分别放到各自调优过的资源池中，可以在某个选定的输入输出比例下提升效率。但生产流量并不会停在那个比例上：输入与输出序列长度、并发、缓存命中率、投机接受率和延迟目标，都会在一天之内不断变化。
+
+一旦设备被划分为 prefill 池和 decode 池，prefill 需求过多时 decode 芯片就会闲置而请求排队；decode 需求过多时则反过来。运营方必须持续预测正确的比例、在两侧都预留冗余容量，并不断重新平衡一个理想比例始终在移动的系统。
+
+在统一的系统中，某些资源在特定阶段可能利用不足，但每台设备都随时可以服务下一条请求。而在分离式系统中，一颗芯片可能仅仅因为归属于错误的资源池就完全闲置：局部利用率看起来更好，全局利用率却可能很差。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/92d363ab-c391-4b9a-9f67-516dac22f534_1364x938.png"
+  caption="来源：SemiAnalysis"
+/>
+
+分离还会破坏局部性。prefill worker 产生的大量 KV cache 是 decode worker 立刻需要的，因此系统必须先把这份状态通过网络传过去，生成才能继续。这会带来额外的带宽消耗、同步、排队，以及又一个故障域。开销还会随输入序列长度上升，因为 KV cache 会变大。不过，避免搬运 KV 主要是一项功耗与延迟优化；愿意搬运一部分 KV，可以用一些功耗和单请求延迟换来更高的硬件利用率。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/5789e0ea-2d0d-4924-b338-f0559c7e7ee9_1928x1292.png" />
+
+可调配的统一集群能够在延迟敏感请求和吞吐导向的批处理之间转移容量，而固定的划分方式则会在流量构成变化时让硬件被闲置。此外，上下文长度会改变注意力与 FFN 之间的工作量平衡，使任何固定的硬件比例都只在其设计点附近才高效。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/9f3a37f8-0996-4dea-9631-02ade61d7fae_1358x938.png"
+  caption="来源：SemiAnalysis"
+/>
+
+同样的约束也适用于投机解码。draft 模型必须以极低延迟把候选 token 送给验证方。把两者拆到各自专用的资源池中，会把一个紧耦合的解码循环变成分布式协议，额外的通信与协调开销可能把起草省下的延迟又吃回去。让两个模型留在同一批设备和同一张低延迟网络上，才能保住让投机解码值得做的那份局部性。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/2eb19374-5a08-4d82-84a3-9875b3fc0f5b_1980x1220.png"
+  caption="来源：SemiAnalysis"
+/>
+
+不过，当需求足够大、足够稳定且可预测时，分离仍然可以胜出，尤其是当常规 GPU 需要较大的分阶段 batch 才能获得良好吞吐时。但它并不是免费的午餐。
+
+## 从日式到印式（由淡到辣）：Katsu、Vindaloo 与 Chana —— 这些咖喱菜如何组成一套机架系统
+
+Jalapeño 系统在机架层面由一个 CPU 主机机架和一个 ASIC 机架组成。主机机架容纳 16 个名为「Katsu」的主机 CPU 托盘，每个对应右侧一个名为「Vindaloo」的 ASIC 托盘。每台主机配备两颗 Turin 级 AMD EPYC CPU 和 1.5TB DRAM，每机架配 2 个 E1.S 和 2 个 M.2 SSD。每个托盘还配有 400G（2x200G）前端网络。每个 Katsu 托盘通过 8 条外部 PCIe DAC 线缆与对应的 Vindaloo 托盘相连，这些线缆在机架前部横向走线。系统级设计由 Celestica 合作完成。
+
+ASIC 机架由 16 个 Vindaloo 托盘和 8 个名为「Chana」的 scale-up 交换托盘（6 个用于本地域、2 个用于全局域）组成。每个 Vindaloo 托盘装有 8 颗 Jalapeño ASIC，每机架合计 128 颗。ASIC 通过铜缆背板连接到各个 Chana 交换托盘，与 NVIDIA 的 Oberon 类似。scale-up 拓扑分为机架内 128 颗 ASIC 的本地域，以及最多连接 16 个机架、2,048 颗 ASIC 的全局域。带宽与拓扑细节见下文。
+
+边车式主机机架的供电约为 50kW 规划容量（生产环境 31kW），ASIC 机架为 130kW，两机架系统合计约 160kW，基本相当于一个双倍宽度的 GB300 机架的功耗。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/596ca531-3bde-44fa-8d1b-34d04b8cc6b8_1304x1382.png"
+  caption="来源：SemiAnalysis、OpenAI"
+/>
+
+OpenAI 可以在单个 scale-up 网络内连接多达 2,048 颗 Jalapeño XPU。该 scale-up 网络由两个域组成：本地域通过背板连接机架内全部 128 颗 XPU；全局域采用铜缆与光互连混合方案，跨 16 个机架连接 2,048 颗 XPU。每个机架有 8 个 Chana 交换托盘，中间 6 个用于本地域，各配一颗 102.4T 的 Tomahawk 6 交换 ASIC；位于本地交换托盘上下方的 2 个用于全局域，我们认为它们可能各由 2 颗 102.4T Tomahawk 6 组成，单托盘可达 204.8T。
+
+在本地域中，128 颗 Jalapeño 芯片每颗具备 4.8Tb/s 的单向带宽，以 all-to-all 方式连接到 6 颗 102.4Tb/s 的 Tomahawk 6 ASIC。这相当于每颗 XPU 需要 48 对差分（DP）公母连接器，每机架用于本地 scale-up 的无源铜缆合计 6,144 DP。
+
+在全局域中，16 个机架共 2,048 颗 XPU 通过铜缆背板、204.8T TH6 电交换、1.6T 光模块和光路交换的组合连接在一起。每颗 XPU 的全局链路具备 1.6Tb/s 单向带宽，对应 XPU 与全局交换之间背板上每颗 XPU 16 对差分连接器。每个全局交换托盘（含 2 颗 ASIC）的出向带宽在背板与前面板光口之间分配。
+
+把本地域与全局域合并计算，每机架每颗 XPU 的背板连接器数量为 64 DP，全机架用于无源铜缆的连接器合计 8,192 DP。
+
+全局域采用 rail-only 架构，跨全局域共 8 条 rail。我们认为 OpenAI 通过安装在每个机架中的光路交换机（OCS）来疏导全局域的光链路。对每颗 XPU 而言，1.6Tb/s 的全局带宽先经铜缆背板到达全局交换托盘，再通过前面板的 1.6T 光模块出口，进入无源光交换后离开机架。由此把 scale-up world size 扩展到 16 个机架、每机架 128 颗、合计 2,048 颗 XPU。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/ae9b26b8-471a-43b2-a60c-cd99ce2b9b07_3480x1342.png"
+  caption="来源：SemiAnalysis"
+/>
+
+由于 scale-up 网络只占系统总成本约 10%，这种灵活性为未来 10 到 20 万亿参数的模型、或 200 万到 400 万 token 的上下文窗口买到了宝贵的可选性。在部署方面，OpenAI 正与 neocloud 合作，并在与数据中心伙伴一起采集可靠性数据直到 1 月，同时优化从到货到上架的部署时间。
+
+## 下一步
+
+接下来我们谈谈 Jalapeño 的未来，它的第一个生产 token 即将到来。下一个目标是 100MW，而障碍主要在硬件侧：能造出多少、数据中心能部署和运营得多好、监控与韧性如何处理等等。软件已经得到验证，而有了内部模型，任何软件上的先发优势都很容易被追上。付费部分我们会讨论这对 NVIDIA、AMD、Cerebras，以及未来几年与 OpenAI 签约的其他芯片公司意味着什么。
+
+[我们在加速器模型中还覆盖了下一代芯片的产量、出货单位与时间线](https://semianalysis.com/accelerator-hbm-model/)。
+
+## 对 NVIDIA、AMD 和 Cerebras 的直接影响
+
+<Blur>
+
+Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua. Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat. Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur. Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
+
+</Blur>
+
+---
+
+_本文完整版发布在我们的 Substack 上。[订阅 SemiAnalysis](https://newsletter.semianalysis.com/subscribe) 阅读完整文章。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "什么是 OpenAI Jalapeño？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "Jalapeño 是 OpenAI 首款自研推理 ASIC，与 Broadcom 合作、从零开始专为 LLM 推理设计，在 Hot Chips 上正式发布。设计工作始于 2024 年年中，CoWoS 设计于 2025 年 11 月流片，从组建团队到流片约 16 个月。目前在 fab 中的 B0 步进采用台积电 N3P 工艺，在单颗满掩模尺寸的计算 die 上提供 13.4 PFLOPs 的 MXFP4 算力，TDP 为 700W，并通过 HBM4 实现每封装 15.4 TB/s 的显存带宽。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Jalapeño 与 NVIDIA Blackwell、Vera Rubin 相比如何？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在每兆瓦全量供电的 token 吞吐量上，Jalapeño 在几乎所有场景中都胜过 Blackwell；其单 token 预测（STP）结果还超过了 NVIDIA 与 CoreWeave 在 2026 年 7 月公布的 Vera Rubin MTP 数据。在每美元 TCO 性能上，Jalapeño 与 Vera Rubin 大致势均力敌。与 Blackwell 比较的参考价值较弱，因为 Jalapeño 真正的对手是 Rubin 这类 HBM4 产品；而且 Jalapeño 取得这些结果时并未使用投机解码，NVIDIA 的对应配置则使用了。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "OpenAI 为什么没有采用 prefill-decode 分离？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "当工作负载构成固定时，分离很有吸引力，但生产流量并非如此：输入与输出长度、并发、缓存命中率、投机接受率和延迟目标都会在一天之内变化。一旦把芯片划分为 prefill 池和 decode 池，一颗芯片可能仅因归属错误的资源池而完全闲置，于是局部利用率提升、全局利用率反而下降。统一资源池还能保住投机解码所依赖的局部性：一旦把 draft 模型与验证方分开，紧耦合的解码循环就会变成分布式协议。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Jalapeño 的机架是如何组成的？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "该系统由两个机架组成。主机机架容纳 16 个名为 Katsu 的 CPU 托盘，每个配备两颗 Turin 级 AMD EPYC CPU 和 1.5TB DRAM，通过外部 PCIe DAC 线缆连接到 ASIC 机架。ASIC 机架容纳 16 个 Vindaloo 托盘，每个装有 8 颗 Jalapeño ASIC，每机架合计 128 颗，另有 8 个基于 Tomahawk 6 的 Chana scale-up 交换托盘。scale-up 分为通过铜缆背板连接的 128 颗 ASIC 本地域，以及跨 16 个机架、连接 2,048 颗 ASIC 的全局域。两机架系统功耗约 160kW。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "Jalapeño 上运行的是什么软件栈？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "OpenAI 使用自研 kernel 语言 Gluon 为该芯片编程。Gluon 构建在 Triton 之上，保留 Triton 的 SPMD 模型，同时暴露更底层的抽象。它最具特色的能力是 layout 抽象，基于 OpenAI 提出的 layout 代数 Linear Layouts，可实现可证明正确的 layout 转换和最优的内存 swizzling。内部服务引擎名为 Teacup，模拟器 chilisim 与实测硬件的误差在 5% 以内，而大量 kernel 代码由内部规模更大的 Codex 版本编写。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "这些 Jalapeño 结果有哪些需要注意的前提？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "所有数据均由 OpenAI 提供。SemiAnalysis 在实验室现场验证了 InferenceX 的运行过程，但没有跑完整套基准测试，目前也没有 AgentX 结果。公布的运行均为 8k1k 单轮场景，比长上下文多轮流量更容易调优：agentic 工作负载会给 router、前缀缓存、缓存管理和 offload 基础设施施加压力，而单轮 8k1k 完全考察不到这些。此外，所测试的模型也不是最新的前沿开放权重模型。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/qwen3-5-397b-agentx-b300-fp4-vs-h100.mdx b/packages/app/content/blog/zh/qwen3-5-397b-agentx-b300-fp4-vs-h100.mdx
new file mode 100644
index 000000000..08107acff
--- /dev/null
+++ b/packages/app/content/blog/zh/qwen3-5-397b-agentx-b300-fp4-vs-h100.mdx
@@ -0,0 +1,66 @@
+---
+title: 'Qwen3.5 397B 的 AgentX 结果：B300 FP4 的每美元性能是 H100 的 12 倍'
+subtitle: '在长上下文 agentic 工作负载上，四年硬件迭代加上 4 位格式能换来什么'
+seoDescription: '在 Qwen3.5 397B 的 AgentX 运行中，B300 FP4 的每美元性能是 H100 的 12 倍。该对比跨越四代硬件与两种数值格式，工作负载为截断到 256k 的 agentic trace 回放。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - qwen
+  - b300
+  - h100
+  - h200
+  - fp4
+  - fp8
+  - sglang
+  - nvidia
+---
+
+在 Qwen3.5 上，B300 FP4 的[每美元性能](/zh/glossary/performance-per-dollar)是 H100 的 12 倍。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/cc2cf6c7-aa58-4a76-a0a2-ee0457d9243c_2048x1298.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=interactivity&i_hc=1&i_best=0&i_prec=fp4%2Cfp8&i_active=b300_sglang%2Ch100_sglang%2Ch200_sglang)"
+/>
+
+## 这个数字代表什么，又不代表什么
+
+这里同时有两个变量在变，读数时都要考虑。第一是四年的硬件迭代；第二是数值格式：该对比跨越 [FP4](/zh/glossary/fp4) 与 [FP8](/zh/glossary/fp8)，并没有固定精度。这样比较反映的是运营方今天在各款硬件上实际能部署的方案，是诚实的口径，但它并不是只针对硅片本身的受控实验。
+
+工作负载同样重要。这不是 8k1k。Qwen3.5 的原生最大上下文长度为 262k token，因此 AgentX 对它回放的是截断到 256k 的 agentic 数据集，这意味着频繁的上下文压缩和大量前缀复用。在长上下文多轮回放上测得的比值，与在互相独立的短提示词上测得的同一比值，是两种不同的结论，而前者更接近真实的生产编码流量。
+
+## 请按你自己的运行点来读
+
+12 倍的总体比值是一个标题，而不是部署方案。真正承载决策的是曲线：先确定你的产品需要的[交互性](/zh/glossary/interactivity)，在该点读取每套系统的表现，并在同一点核对首 token 延迟，然后再下结论。
+
+这也是各模型的 AgentX 图表默认展示完整前沿、而不是单个数字的原因。较老的硬件可能在低并发下于每美元性能上仍有竞争力，却在高吞吐区间完全掉队；单一比值无法同时表达这条曲线的两端。
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 Qwen3.5 397B 上，B300 比 H100 好多少？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在 AgentX 的 Qwen3.5 运行中，B300 FP4 的每美元性能是 H100 的 12 倍。该对比同时跨越硬件代际差异和数值格式差异，因为它覆盖 FP4 与 FP8，并未固定精度。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "这个 12 倍是在定长序列工作负载上测得的吗？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "不是。Qwen3.5 的原生最大上下文长度为 262k token，因此 AgentX 对它回放的是截断到 256k 的 agentic trace 数据集。该工作负载包含频繁的上下文压缩和大量前缀复用，与互相独立的短提示词是完全不同的测量。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/qwen3-5-397b-agentx-nvidia-vs-amd-sglang.mdx b/packages/app/content/blog/zh/qwen3-5-397b-agentx-nvidia-vs-amd-sglang.mdx
new file mode 100644
index 000000000..20f07bfcf
--- /dev/null
+++ b/packages/app/content/blog/zh/qwen3-5-397b-agentx-nvidia-vs-amd-sglang.mdx
@@ -0,0 +1,67 @@
+---
+title: 'MI355X 对比 GB300 NVL72 推理性能：Qwen3.5 SGLang 上 20 倍差距'
+subtitle: 'GatedDeltaNet、262k 原生上下文，以及同一引擎上 AMD 完全缺席的竞争'
+seoDescription: 'Qwen3.5 397B 的 AgentX 结果。在 SGLang 对 SGLang 的比较中，NVIDIA 在 90 tok/s/user 下领先 20 倍以上，AMD 完全没有竞争力；NVIDIA 的 SGLang 提交在 p90 TTFT 上也远低于 TRT-LLM。'
+date: '2026-08-25'
+publishDate: '2026-08-25'
+tags:
+  - agentx
+  - agentic
+  - benchmark
+  - inference
+  - qwen
+  - sglang
+  - trtllm
+  - nvidia
+  - amd
+---
+
+Qwen3.5 397B 每隔几层就用 GatedDeltaNet 替代常规注意力。GatedDeltaNet 由 MIT 与 NVIDIA Research 提出，其状态存储需求在理论上是常数，而常规注意力是线性增长的，因此相比同等规模的稠密注意力模型，它的存储需求更低。与 Nemotron 那样惨败的端到端模型训练研究不同，NVIDIA Research 在 GDN、LatentMoE 这类被前沿模型采用的基础研究上做得非常出色。
+
+在看结果之前先说明一点测试方法：该模型原生最大上下文长度为 262k token，因此 AgentX 对它回放的是[截断到 256k 的数据集](https://inferencex.semianalysis.com/agentx/cc-traces-weka-062126-256k)。这模拟的是在上下文上限较小的模型上频繁触顶、频繁压缩上下文的工作负载，也就是用户实际使用这类模型的方式。
+
+## 差距
+
+在 SGLang 对 SGLang 的比较中，Qwen3.5 397B 是 NVIDIA 的强势领地：90 tok/s/user 时性能领先 20 倍以上。在 Qwen3.5 SGLang 上，AMD 目前完全没有竞争力。
+
+<Figure
+  src="https://substack-post-media.s3.amazonaws.com/public/images/519a7cd0-ac51-4a5c-9c45-ba7c41f6d058_2048x1300.png"
+  caption="来源：[SemiAnalysis InferenceX](https://inferencex.semianalysis.com/inference?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces&i_xmode=ttft&i_best=0&i_hc=1)"
+/>
+
+固定引擎这一变量，正是这组对比值得一提的原因。跨引擎的结果总有争论空间，因为不同软件栈的成熟度不同，投入优化的人力也不同；而这里两边都是 SGLang，差距依然是 20 倍。
+
+## NVIDIA 在用 TTFT 换交互性
+
+在 NVIDIA 阵营内部，我们再次看到以牺牲首 token 延迟为代价过度优化[交互性](/zh/glossary/interactivity)的现象，TRT-LLM 尤其明显。上图中所有 NVIDIA SGLang 提交的 p90 TTFT 都远低于 TRT-LLM。
+
+这是真实的取舍而非缺陷，但读者必须看得到它。一套让用户先等更久、之后流式输出更快的方案，在交互性坐标轴上会更好看，在首 token 的[服务级目标](/zh/glossary/service-level-objective)下则更差；只有两者一起读，才知道哪一种适合你的产品。
+
+<Figure src="https://substack-post-media.s3.amazonaws.com/public/images/c92da639-dcff-4e39-b9b5-49742677e93c_2048x1289.png" />
+
+---
+
+_本文结果只是 AgentX 1.0 的一个切面。完整分析、回放方法，以及该基准测试推动的 70 多个上游 PR，请见 [AgentX - InferenceXv3：CUDA 护城河在 agentic 推理中还站得住吗？](/zh/blog/agentx-inferencexv3-does-cuda-moat)。每个数据点都可以在[免费仪表板](https://inferencex.semianalysis.com/)上自行探索。_
+
+<JsonLd>{`{
+  "@context": "https://schema.org",
+  "@type": "FAQPage",
+  "mainEntity": [
+    {
+      "@type": "Question",
+      "name": "在 Qwen3.5 397B 上，NVIDIA 比 AMD 快多少？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "在 SGLang 对 SGLang 的比较中，NVIDIA 在 90 tok/s/user 下领先 20 倍以上；在 Qwen3.5 SGLang 上，AMD 目前完全没有竞争力。两边都固定使用同一引擎，正是这组对比有意义的前提。"
+      }
+    },
+    {
+      "@type": "Question",
+      "name": "AgentX 为什么对 Qwen3.5 回放截断后的数据集？",
+      "acceptedAnswer": {
+        "@type": "Answer",
+        "text": "该模型原生最大上下文长度为 262k token，因此 AgentX 使用截断到 256k 的上下文数据集。这模拟的是在上下文上限较小的模型上频繁触顶、频繁压缩上下文的工作负载，也就是用户实际使用这类模型的方式。"
+      }
+    }
+  ]
+}`}</JsonLd>
diff --git a/packages/app/content/blog/zh/ultra-high-interactivity-on-nvidia.mdx b/packages/app/content/blog/zh/ultra-high-interactivity-on-nvidia.mdx
index 041ee64b3..45cdc7a95 100644
--- a/packages/app/content/blog/zh/ultra-high-interactivity-on-nvidia.mdx
+++ b/packages/app/content/blog/zh/ultra-high-interactivity-on-nvidia.mdx
@@ -3,6 +3,7 @@ title: '在 NVIDIA GPU 上实现超高交互性？TileRT 登陆 InferenceX'
 subtitle: '运行在 NVIDIA GPU 上的 TileRT 软件能否与 Cerebras、Groq LPU、SambaNova 竞争？批大小为 1、分离式引擎、高吞吐量预填充引擎、高交互性解码引擎'
 date: '2026-08-10'
 publishDate: '2026-08-10'
+modifiedDate: '2026-08-18'
 tags:
   - benchmark
   - gpu
@@ -13,6 +14,8 @@ tags:
   - tilert
   - vllm
   - glm5
+  - agentx
+  - agentic
 ---
 
 _本文最初于 2026 年 8 月 10 日发布在 [SemiAnalysis 通讯](https://newsletter.semianalysis.com/p/ultra-high-interactivity-on-nvidia)。_
@@ -235,9 +238,9 @@ TileRT 继承了 ASIC 厂商最大的弱点。静态提前编译意味着模型
 
 AI 编码智能体能在已知模板内加速调优，但新颖的变换仍需专家判断。此外，单体式持久化 kernel 也削弱了传统逐 kernel 性能剖析时间线的作用，使自动化反馈闭环更加困难。
 
-## TileRT 与 InferenceX 的下一步
+## TileRT 的下一步
 
-我们正在积极推进，将 TileRT 的基准测试从 InferenceX 的单轮 8k/1k 场景扩展到我们新的智能体编码基准——AgentX。该场景回放真实的 Claude Code 与 Codex 轨迹，包含长上下文、多轮请求、真实的子智能体活动与动态工具调用延迟。其输入长度中位数为 140k tokens，理论缓存命中率中位数屋顶线可达 99.2%。
+InferenceX 的长上下文多轮编码场景 AgentX 现已在具备对应数据的模型与服务栈上正式上线。它回放从真实 Claude Code 与 Codex 会话衍生出的隐私保护工作负载形态，包含长上下文、多轮请求、真实的 subagent 活动与动态工具调用延迟。其输入长度中位数为 140k tokens，理论缓存命中率中位数屋顶线可达 99.2%。对 TileRT 而言，下一步仍是将目前的单轮 8k/1k 基准测试扩展到 AgentX。
 
 <Figure
   src="/images/ultra-high-interactivity-on-nvidia/agentx-benchmark-characteristics.png"
@@ -245,7 +248,7 @@ AI 编码智能体能在已知模板内加速调优，但新颖的变换仍需
   caption="来源：SemiAnalysis"
 />
 
-该工作负载将检验 TileRT 与 vLLM 的整套系统，而不仅是解码速度，涵盖增量 KV 传输、前缀缓存复用、缓存保留与卸载、路由与调度。关键问题在于：TileRT 能否在轮次之间只传输新增上下文，同时保持其超高交互性优势。
+该工作负载将检验 TileRT 与 vLLM 的整套系统，而不仅是解码速度，涵盖增量 KV 传输、前缀缓存复用、缓存保留与 offload、路由与调度。关键问题在于：TileRT 能否在轮次之间只传输新增上下文，同时保持其超高交互性优势。
 
 <Figure
   src="/images/ultra-high-interactivity-on-nvidia/multi-turn-thinking.png"
diff --git a/packages/app/content/blog/zh/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx b/packages/app/content/blog/zh/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx
index 02187d766..6385c0653 100644
--- a/packages/app/content/blog/zh/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx
+++ b/packages/app/content/blog/zh/vera-rubin-nvl72-vs-gb200-nvl72-inference.mdx
@@ -223,7 +223,7 @@ CoreWeave 性能结果的第二个细节是他们使用的是 DeepSeek R1 671B
 
 有趣的是，CoreWeave 选择 DeepSeek R1 671B 在理论上反而更有利于 Blackwell 基线，而非 Rubin。Rubin 的主要优势在于更大的 HBM 容量、更大的 CPU DRAM 容量和更高的 HBM 带宽，这意味着 Rubin 更适合 Fable 5、Gemini Pro、Kimi K3、Qwen3.8 2.4T 这类多万亿参数模型。
 
-第三点值得注意的是，CoreWeave 只使用了单轮 8k/1k 输入/输出 token。理论上，Agentic Coding 这类多轮长上下文工作负载在 Rubin 上应有更好表现（得益于其更大的 HBM 容量和带宽），但简单的单轮基准测试无法体现这一点。[我们即将推出的 AgentX 基准测试场景由 Weka、LMCache、vLLM/SGLang 社区、Nvidia、AMD 及社区众多伙伴合作打造，将提供真实的 agentic 工作负载来评测推理性能。](https://inferencex.semianalysis.com/datasets/cc-traces-weka-062126)我们鼓励所有人采用这一推理基准测试！
+第三点值得注意的是，CoreWeave 只使用了单轮 8k/1k 输入/输出 token。理论上，Agentic Coding 这类多轮长上下文工作负载在 Rubin 上应有更好表现（得益于其更大的 HBM 容量和带宽），但简单的单轮基准测试无法体现这一点。[我们即将推出的 AgentX 基准测试场景由 Weka、LMCache、vLLM/SGLang 社区、Nvidia、AMD 及社区众多伙伴合作打造，将提供真实的 agentic 工作负载来评测推理性能。](https://inferencex.semianalysis.com/zh/agentx/cc-traces-weka-062126)我们鼓励所有人采用这一推理基准测试！
 
 最后，我们注意到 CoreWeave 的测试是在一台没有 scale-out 网络的预量产机架上完成的。具体来说，CoreWeave 使用的是 Dell 工程样机（Engineering Sample，ES）机架。我们仍然认为这些结果有价值：它们启用了宽 EP 和 PD 分离，用到了 NVL72 的 scale-up 背板，证明其工作良好。这块背板在 GB200 NVL72 Oberon 爬坡期间曾面临诸多可靠性挑战，[我们在加速器模型中有过记录。](https://semianalysis.com/accelerator-hbm-model/)
 
diff --git a/packages/app/content/models/deepseek-r1.mdx b/packages/app/content/models/deepseek-r1.mdx
new file mode 100644
index 000000000..c42fbccc2
--- /dev/null
+++ b/packages/app/content/models/deepseek-r1.mdx
@@ -0,0 +1,81 @@
+---
+title: 'DeepSeek R1 0528'
+developer: 'DeepSeek'
+releaseDate: 'May 28, 2025'
+description: "DeepSeek's 671B-parameter / 37B-active MoE reasoning model with Multi-head Latent Attention, released under the MIT license."
+---
+
+## Overview
+
+DeepSeek-R1-0528 is a minor version upgrade of DeepSeek-AI's open-weights reasoning model DeepSeek-R1: per the official model card, "the DeepSeek R1 model has undergone a minor version upgrade, with the current version being DeepSeek-R1-0528," which "significantly improved its depth of reasoning and inference capabilities by leveraging increased computational resources and introducing algorithmic optimization mechanisms during post-training," with overall performance described as approaching leading models such as O3 and Gemini 2.5 Pro ([DeepSeek-R1-0528 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)). DeepSeek's own release note, dated May 28, 2025, summarizes the update as improved benchmark performance, enhanced front-end capabilities, reduced hallucinations, and support for JSON output and function calling, with open-source weights on Hugging Face ([DeepSeek-R1-0528 Release](https://api-docs.deepseek.com/news/news250528/)).
+
+Positioning is reasoning-first, with strong secondary emphasis on code and agentic tool use. The card attributes the gains to deeper thinking: AIME 2025 accuracy rose from 70% to 87.5%, with average reasoning length increasing from about 12K tokens per question in the previous version to about 23K tokens in R1-0528; the release also cites a reduced hallucination rate, enhanced function-calling support, and a better "vibe coding" experience ([DeepSeek-R1-0528 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)). Usage recommendations changed relative to earlier R1: a system prompt is now supported, and it is no longer necessary to prepend `<think>\n` to force thinking mode; the official web/app temperature is 0.6 ([DeepSeek-R1-0528 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+
+The underlying technique lineage is DeepSeek-R1's reinforcement-learning recipe: the R1 paper argues that LLM reasoning ability can be incentivized through pure RL without human-labeled reasoning trajectories, producing emergent self-reflection, verification and dynamic strategy adaptation ([DeepSeek-R1 paper, arXiv:2501.12948](https://arxiv.org/abs/2501.12948)). R1 (671B total / 37B active MoE) is trained on top of DeepSeek-V3-Base ([DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1)), whose architecture adopts Multi-head Latent Attention (MLA) and DeepSeekMoE with an auxiliary-loss-free load-balancing strategy and a multi-token-prediction objective usable for speculative decoding ([DeepSeek-V3 model card](https://huggingface.co/deepseek-ai/DeepSeek-V3)). Both the code repository and model use are under the MIT License, and the R1 series supports commercial use and distillation ([DeepSeek-R1-0528 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)). DeepSeek also distilled R1-0528's chain-of-thought into Qwen3-8B Base to produce DeepSeek-R1-0528-Qwen3-8B ([DeepSeek-R1-0528 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+
+## Architecture
+
+- Total parameters: 671B; activated per token: 37B (DeepSeek-R1 series table) ([DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1)); the Hugging Face repo metadata lists 685B params, which for DeepSeek-V3 is explained as 671B main weights + 14B MTP module weights ([DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1), [DeepSeek-V3 model card](https://huggingface.co/deepseek-ai/DeepSeek-V3))
+- Architecture family: Mixture-of-Experts; `architectures: ["DeepseekV3ForCausalLM"]`, `model_type: deepseek_v3` ([DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1), [config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- Base model: DeepSeek-R1 and R1-Zero are trained based on DeepSeek-V3-Base ([DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1))
+- Attention: Multi-head Latent Attention (MLA), adopted from the DeepSeek-V3/V2 architecture ([DeepSeek-V3 model card](https://huggingface.co/deepseek-ai/DeepSeek-V3)); the R1-0528 config exposes MLA ranks `kv_lora_rank: 512`, `q_lora_rank: 1536`, `qk_nope_head_dim: 128`, `qk_rope_head_dim: 64`, `v_head_dim: 128` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json)). Note: the R1-0528 card itself does not name the attention mechanism ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md))
+- Layers: `num_hidden_layers: 61`, with `first_k_dense_replace: 3` (first three layers dense rather than MoE) and `num_nextn_predict_layers: 1` (MTP block) ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- Attention heads: `num_attention_heads: 128`, `num_key_value_heads: 128`, `hidden_size: 7168`, dense `intermediate_size: 18432` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- MoE experts: `n_routed_experts: 256`, `n_shared_experts: 1`, `num_experts_per_tok: 8`, `moe_intermediate_size: 2048`, `moe_layer_freq: 1`, grouped routing with `n_group: 8` / `topk_group: 4`, `scoring_func: sigmoid`, `topk_method: noaux_tc`, `routed_scaling_factor: 2.5` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- Load balancing: auxiliary-loss-free load-balancing strategy (DeepSeek-V3 lineage; `noaux_tc` in config) ([DeepSeek-V3 model card](https://huggingface.co/deepseek-ai/DeepSeek-V3), [config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- Context window: the DeepSeek-R1 series table lists 128K ([DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1)); the R1-0528 config sets `max_position_embeddings: 163840` with YaRN rope scaling (`factor: 40` over `original_max_position_embeddings: 4096`, `rope_theta: 10000`) ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- Precision / quantization: `torch_dtype: bfloat16` with an FP8 quantization config (`quant_method: fp8`, `fmt: e4m3`, dynamic activation scheme, `weight_block_size: [128,128]`); repo tensor types for R1 are listed as BF16, F8_E4M3, F32 ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json), [DeepSeek-R1 model card](https://huggingface.co/deepseek-ai/DeepSeek-R1))
+- Vocabulary: `vocab_size: 129280`, `hidden_act: silu`, `rms_norm_eps: 1e-06`, `tie_word_embeddings: false` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- Multi-token prediction: MTP training objective from the V3 lineage, also usable for speculative decoding ([DeepSeek-V3 model card](https://huggingface.co/deepseek-ai/DeepSeek-V3)); `num_nextn_predict_layers: 1` in the R1-0528 config ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/config.json))
+- Reasoning budget: maximum generation length used in DeepSeek's own evals is 64K tokens; average AIME reasoning length ~23K tokens per question (up from ~12K in R1) ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md))
+- Distilled sibling: DeepSeek-R1-0528-Qwen3-8B has an architecture identical to Qwen3-8B but shares R1-0528's tokenizer configuration ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md))
+- License: MIT (`license: mit` in the card front matter); commercial use and distillation supported ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md))
+
+## Official vendor eval scores
+
+All values are DeepSeek-reported for DeepSeek-R1-0528 on its official Hugging Face model card. Eval settings: max generation length 64K, temperature 0.6, top-p 0.95, 16 samples per query for pass@1; SWE-Verified uses the Agentless framework; HLE is evaluated on text-only prompts; GPT-4.1 plays the user role in Tau-bench ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+
+| Benchmark                        | Score                                                                           | Nemotron3 Ultra | % better | Source                                                                                  |
+| -------------------------------- | ------------------------------------------------------------------------------- | --------------- | -------- | --------------------------------------------------------------------------------------- |
+| MMLU-Pro (EM)                    | 85.0                                                                            | 86.8            | -2.1%    | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| MMLU-Redux (EM)                  | 93.4                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| GPQA-Diamond (Pass@1)            | 81.0                                                                            | 87.0 (no tools) | -6.9%    | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| Humanity's Last Exam (Pass@1)    | 17.7                                                                            | 26.7 (no tools) | -33.7%   | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| SimpleQA (Correct)               | 27.8                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| FRAMES (Acc.)                    | 83.0                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| LiveCodeBench 2408–2505 (Pass@1) | 73.3                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| Codeforces-Div1 (Rating)         | 1930                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| SWE-bench Verified (Resolved)    | 57.6                                                                            | 70.7            | -18.5%   | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| Aider-Polyglot (Acc.)            | 71.6                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| AIME 2024 (Pass@1)               | 91.4                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| AIME 2025 (Pass@1)               | 87.5                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| HMMT 2025 (Pass@1)               | 79.4                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| CNMO 2024 (Pass@1)               | 86.9                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| BFCL_v3_MultiTurn (Acc)          | 37.0                                                                            | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| Tau-Bench (Pass@1)               | 53.5 (Airline) / 63.9 (Retail)                                                  | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| Terminal-Bench                   | n.a. — not reported by DeepSeek for R1-0528                                     | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| tau2-bench                       | n.a. — DeepSeek reports the original Tau-Bench (Airline/Retail), not tau2-bench | n.a.            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+| BrowseComp                       | n.a. — not reported by DeepSeek for R1-0528                                     | 44.4            | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md) |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+
+## Benchmark explanations
+
+- **MMLU-Pro** — a harder, cleaned multi-subject multiple-choice exam suite measuring broad academic and professional knowledge and reasoning, scored as exact match ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **MMLU-Redux** — a relabeled/error-corrected version of MMLU, measuring the same broad knowledge with cleaner ground truth ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **GPQA-Diamond** — the hardest subset of graduate-level physics, chemistry and biology questions written to resist simple web lookup; measures expert scientific reasoning ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **Humanity's Last Exam (HLE)** — an extremely difficult expert-written exam across many disciplines used as a frontier-difficulty ceiling; DeepSeek evaluated text-only prompts ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **SimpleQA** — short factual questions testing parametric factual recall and calibration; measures the share answered correctly ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **FRAMES** — multi-document retrieval-and-reasoning questions requiring synthesis of several sources into one factual answer ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **LiveCodeBench (2408–2505)** — competitive-programming problems drawn from contests in a fixed recent window to limit contamination; measures pass@1 code correctness ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **Codeforces-Div1** — an Elo-style rating estimated from Division-1 competitive-programming contest performance rather than a percentage ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **SWE-bench Verified** — human-validated real GitHub issues where the model must generate a patch passing the repo's tests; evaluated here with the Agentless framework ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **Aider-Polyglot** — multi-language code-editing tasks in the Aider harness, measuring the ability to apply correct edits across programming languages ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **AIME 2024 / AIME 2025** — American Invitational Mathematics Examination problem sets: hard short-answer competition math, a core reasoning-depth test ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **HMMT 2025** — the Harvard-MIT Math Tournament 2025 problem set, olympiad-style short-answer math ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **CNMO 2024** — the 2024 Chinese National High School Mathematics Olympiad, measuring olympiad math in a Chinese-language setting ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **BFCL_v3_MultiTurn** — the Berkeley Function-Calling Leaderboard v3 multi-turn split, measuring correct tool/function invocation across a dialogue ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **Tau-Bench (Airline / Retail)** — simulated customer-service agent tasks requiring tool use and policy adherence over multi-turn dialogue; GPT-4.1 played the user ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **Terminal-Bench** — multi-step shell/terminal agent tasks; not reported by DeepSeek for R1-0528 ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **tau2-bench** — the successor to Tau-Bench with additional domains and a dual-control user simulator; not reported by DeepSeek for R1-0528 ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
+- **BrowseComp** — hard-to-find information-seeking questions requiring persistent web browsing; not reported by DeepSeek for R1-0528 ([model card](https://huggingface.co/deepseek-ai/DeepSeek-R1-0528/raw/main/README.md)).
diff --git a/packages/app/content/models/deepseek-v4.mdx b/packages/app/content/models/deepseek-v4.mdx
new file mode 100644
index 000000000..c4bcc7075
--- /dev/null
+++ b/packages/app/content/models/deepseek-v4.mdx
@@ -0,0 +1,84 @@
+---
+title: 'DeepSeek V4 Pro'
+developer: 'DeepSeek'
+releaseDate: 'April 24, 2026 (preview); August 13, 2026 (V4-Pro-0813 GA)'
+description: "DeepSeek's 1.6T-parameter / 49B-active MoE flagship with hybrid CSA+HCA sparse attention, a 1M-token context window, and MIT-licensed open weights."
+---
+
+## Overview
+
+DeepSeek-V4-Pro is the flagship open-weights Mixture-of-Experts (MoE) language model of the DeepSeek-V4 series from DeepSeek-AI, with 1.6T total parameters and 49B activated per token and a native 1M-token context length ([DeepSeek-V4-Pro model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)). It was announced as "DeepSeek-V4 Preview … officially live & open-sourced" alongside the smaller DeepSeek-V4-Flash (284B total / 13B active) on DeepSeek's own API news page dated April 24, 2026, which also states that 1M context is the default across all official DeepSeek services and that both models support dual Thinking / Non-Thinking modes ([DeepSeek V4 Preview Release](https://api-docs.deepseek.com/news/news260424)). The accompanying technical report, "DeepSeek-V4: Towards Highly Efficient Million-Token …", was submitted to arXiv on 26 Apr 2026 and presents the two MoE models as a preview of the V4 series ([arXiv:2606.19348](https://arxiv.org/abs/2606.19348)). The weights are released under the MIT License ([DeepSeek-V4-Pro model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+
+Positioning is explicitly agentic and coding-first: DeepSeek describes V4-Pro as having enhanced agentic capabilities, as open-source SOTA on agentic-coding benchmarks, as leading all current open models while trailing only Gemini-3.1-Pro, and as being used for in-house agentic coding at DeepSeek, with integrations for Claude Code, OpenClaw and OpenCode ([DeepSeek V4 Preview Release](https://api-docs.deepseek.com/news/news260424)). The maximum reasoning-effort mode, DeepSeek-V4-Pro-Max, is described in the technical report as redefining the state of the art for open models ([arXiv:2606.19348](https://arxiv.org/abs/2606.19348)).
+
+The headline techniques are efficiency-oriented. V4-Pro uses a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), Manifold-Constrained Hyper-Connections (mHC) in place of conventional residual connections, and the Muon optimizer, pre-trained on more than 32T tokens; in the 1M-token setting it needs 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 ([DeepSeek-V4-Pro model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [arXiv:2606.19348](https://arxiv.org/abs/2606.19348)). A production checkpoint, DeepSeek-V4-Pro-0813, superseded the preview: it is built on the preview structure with a DSpark speculative decoding module attached and reports 1.7T params, with "greatly enhanced agentic capabilities" ([DeepSeek-V4-Pro-0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813)); DeepSeek's changelog dated 2026-08-13 states the GA release rolled out on app, web and API with low/high/max reasoning effort and native OpenAI Responses API support ([DeepSeek-V4-Pro GA Release](https://api-docs.deepseek.com/news/news260813/), [DeepSeek API Change Log](https://api-docs.deepseek.com/updates/)).
+
+## Architecture
+
+- Total parameters: 1.6T (preview checkpoint) ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)); the GA card lists 1.7T params for DeepSeek-V4-Pro-0813 ([0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813))
+- Activated parameters per token: 49B ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [arXiv:2606.19348](https://arxiv.org/abs/2606.19348))
+- Architecture family: Mixture-of-Experts decoder LM, `DeepseekV4ForCausalLM`, `model_type: deepseek_v4` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json))
+- Layers: `num_hidden_layers: 61`, plus `num_nextn_predict_layers: 1` (MTP block) ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json)); the 61-layer stack is described as layers 0–1 HCA and layers 2–60 alternating CSA/HCA, with the final MTP block using sliding-window attention only ([Hugging Face DeepSeek-V4 blog](https://huggingface.co/blog/deepseekv4))
+- Attention: hybrid architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [arXiv:2606.19348](https://arxiv.org/abs/2606.19348))
+- CSA details: compresses KV 4× along the sequence, softmax-gated pooling with learned positional bias, an FP4 "lightning indexer" selecting top-k compressed blocks, plus a sliding-window branch for recent tokens ([Hugging Face DeepSeek-V4 blog](https://huggingface.co/blog/deepseekv4))
+- HCA details: compresses KV 128×, dense (no sparse selection) attention over compressed blocks with a sliding-window recency branch ([Hugging Face DeepSeek-V4 blog](https://huggingface.co/blog/deepseekv4)); the config's `compress_ratios` list alternates 128 and 4 across layers, with `index_topk: 1024`, `index_n_heads: 64`, `index_head_dim: 128`, `sliding_window: 128` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json))
+- Attention shape: `num_attention_heads: 128`, `num_key_value_heads: 1`, `head_dim: 512`, `q_lora_rank: 1536`, `o_lora_rank: 1024`, `qk_rope_head_dim: 64`, `hidden_size: 7168` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json))
+- MoE experts: `n_routed_experts: 384`, `n_shared_experts: 1`, `num_experts_per_tok: 6`, `moe_intermediate_size: 3072`, routing `scoring_func: sqrtsoftplus` with `topk_method: noaux_tc` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json)); the model card itself does not state expert counts ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro))
+- Residual connections: replaced by Manifold-Constrained Hyper-Connections (mHC) ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [arXiv:2606.19348](https://arxiv.org/abs/2606.19348)); config exposes `hc_mult: 4`, `hc_sinkhorn_iters: 20`, `hc_eps: 1e-06`, `num_hash_layers: 3` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json))
+- Context window: 1M tokens; `max_position_embeddings: 1048576` with YaRN rope scaling factor 16 over `original_max_position_embeddings: 65536` ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json))
+- Precision / quantization: distributed as FP4 + FP8 Mixed — MoE expert parameters in FP4, most other parameters in FP8 (Base checkpoints are FP8 Mixed); HF tensor types listed as BF16, I64, F32, F8_E4M3, I8 ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)). Config shows `quant_method: fp8`, `fmt: e4m3`, `scale_fmt: ue8m0`, `weight_block_size: [128,128]`, and the 0813 config adds `expert_dtype: fp4` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json), [0813 config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/raw/main/config.json))
+- KV/FLOP efficiency: at 1M context, 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [arXiv:2606.19348](https://arxiv.org/abs/2606.19348))
+- Optimizer and pre-training data: Muon optimizer; more than 32T diverse, high-quality tokens ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [arXiv:2606.19348](https://arxiv.org/abs/2606.19348))
+- Vocabulary: `vocab_size: 129280`, `hidden_act: silu`, `swiglu_limit: 10.0` ([config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/raw/main/config.json))
+- Speculative decoding (GA checkpoint only): DSpark module attached, `dspark_target_layer_ids: [58,59,60]`, `dspark_block_size: 5`, `dspark_markov_rank: 512` ([0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813), [0813 config.json](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813/raw/main/config.json))
+- License: MIT ([model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813))
+- Modalities: text in / text out; the preview API announcement lists no image/audio input for Pro (a separate experimental `deepseek-v4-flash-vision-exp` accepts images) ([DeepSeek API docs](https://api-docs.deepseek.com/))
+
+## Official vendor eval scores
+
+All rows below are vendor-reported. Unless noted, values are DeepSeek-V4-Pro-**Max** (maximum reasoning effort) from the official Hugging Face model card.
+
+| Benchmark                          | Score                                                                                                                     | Nemotron3 Ultra   | % better | Source                                                                                                                                          |
+| ---------------------------------- | ------------------------------------------------------------------------------------------------------------------------- | ----------------- | -------- | ----------------------------------------------------------------------------------------------------------------------------------------------- |
+| MMLU-Pro (EM)                      | 87.5                                                                                                                      | 86.8              | +0.8%    | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| GPQA Diamond (Pass@1)              | 90.1                                                                                                                      | 87.0 (no tools)   | +3.6%    | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| HLE (Humanity's Last Exam, Pass@1) | 37.7 (preview Max); 42.7 without tools / 60.0 with tools for V4-Pro-0813                                                  | 26.7 (no tools)   | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro); [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) |
+| HLE with tools (Pass@1)            | 48.2 (preview Max)                                                                                                        | 37.4 (with tools) | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| LiveCodeBench (Pass@1)             | 93.5                                                                                                                      | 89.0 (v6)         | +5.1%    | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| Codeforces (Rating)                | 3206                                                                                                                      | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| HMMT 2026 Feb (Pass@1)             | 95.2                                                                                                                      | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| IMOAnswerBench (Pass@1)            | 89.8                                                                                                                      | 88.6 (no tools)   | +1.4%    | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| SWE-bench Verified (Resolved)      | 80.6                                                                                                                      | 70.7              | +14.0%   | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| SWE-bench Pro (Resolved)           | 55.4                                                                                                                      | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| SWE-bench Multilingual (Resolved)  | 76.2                                                                                                                      | 67.7              | +12.6%   | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| Terminal-Bench 2.0 (Acc)           | 67.9                                                                                                                      | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| Terminal-Bench 2.1 (V4-Pro-0813)   | 87.9                                                                                                                      | 56.4              | +55.9%   | [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813)                                                                      |
+| BrowseComp (Pass@1)                | 83.4                                                                                                                      | 44.4              | +87.8%   | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| Toolathlon (Pass@1)                | 51.8 (preview Max); Toolathlon-Verified 74.1 for V4-Pro-0813                                                              | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro); [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) |
+| MRCR 1M (MMR)                      | 83.5                                                                                                                      | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| DeepSWE (V4-Pro-0813)              | 62.7                                                                                                                      | n.a.              | —        | [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813)                                                                      |
+| AIME 2024 / AIME 2025              | n.a. — not reported by DeepSeek for V4-Pro; math reasoning is reported via HMMT 2026 Feb, IMOAnswerBench and Apex instead | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)                                                                             |
+| tau2-bench                         | n.a. — not present in DeepSeek's V4-Pro or V4-Pro-0813 result tables                                                      | n.a.              | —        | [HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813) |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+Additional vendor-reported values on the same cards, for reference: Apex 38.3, Apex Shortlist 90.2, SimpleQA-Verified 57.9, Chinese-SimpleQA 84.4, CorpusQA 1M 62.0, MCPAtlas Public 73.6, GDPval-AA Elo 1554 ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)); and for V4-Pro-0813: NL2Repo 61.5, Cybergym 83.3, Agents' Last Exam 25.7, AutomationBench (Public) 31.8, plus internal DSBench-FullStack 71.1 and DSBench-Hard 67.2 ([0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813)).
+
+## Benchmark explanations
+
+- **MMLU-Pro** — a harder, cleaned multiple-choice exam suite across many academic and professional subjects, measuring broad knowledge and reasoning; scored as exact match ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **GPQA Diamond** — the hardest subset of graduate-level, "Google-proof" physics, chemistry and biology questions written by domain PhDs; measures expert scientific reasoning ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **HLE (Humanity's Last Exam)** — an extremely difficult, expert-authored exam spanning many fields, used as a frontier-difficulty knowledge and reasoning ceiling; DeepSeek also reports a tool-augmented variant ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **LiveCodeBench** — competitive-programming problems collected continuously from live contests to limit contamination; measures code generation correctness at pass@1 ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **Codeforces** — an Elo-style rating estimated from competitive-programming contest performance rather than a percentage accuracy ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **HMMT 2026 Feb** — the February 2026 Harvard-MIT Math Tournament problem set, a hard olympiad-style short-answer math benchmark ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **IMOAnswerBench** — International Mathematical Olympiad-style problems graded on final answers, measuring olympiad math ability ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **SWE-bench Verified** — human-validated real GitHub issues where the model must produce a patch that passes the repository's tests; measures agentic software-engineering ability ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **SWE-bench Pro** — a harder, more realistic variant of the SWE-bench issue-resolution task ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **SWE-bench Multilingual** — the same issue-resolution setup extended beyond Python to repositories in multiple programming languages ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **Terminal-Bench 2.0 / 2.1** — multi-step shell and systems tasks executed in a real terminal environment, measuring end-to-end agentic tool use ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813)).
+- **BrowseComp** — hard-to-find information-seeking questions that require an agent to browse the web persistently to locate the answer ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **Toolathlon / Toolathlon-Verified** — broad tool-calling agent tasks across many tools and APIs; the Verified split is a human-checked subset ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro), [0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813)).
+- **MRCR 1M** — multi-round coreference resolution at one-million-token context, measuring long-context retrieval and tracking ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **DeepSWE** — a software-engineering agent evaluation reported by DeepSeek for the 0813 production checkpoint ([0813 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813)).
+- **AIME** — the American Invitational Mathematics Examination, a short-answer competition math benchmark; not reported by DeepSeek for V4-Pro ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
+- **tau2-bench** — a customer-service-style tool-use and policy-following agent benchmark; not reported by DeepSeek for V4-Pro ([HF model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)).
diff --git a/packages/app/content/models/glm-5-1.mdx b/packages/app/content/models/glm-5-1.mdx
new file mode 100644
index 000000000..34d434052
--- /dev/null
+++ b/packages/app/content/models/glm-5-1.mdx
@@ -0,0 +1,88 @@
+---
+title: 'GLM-5 / GLM-5.1'
+developer: 'Z.ai (Zhipu AI)'
+releaseDate: 'February 11, 2026 (GLM-5); April 7, 2026 (GLM-5.1)'
+description: "Z.ai's 744B-parameter / 40B-active MoE family with DeepSeek Sparse Attention (DSA), trained on 28.5T tokens and released under the MIT license."
+---
+
+## Overview
+
+GLM-5 is the flagship open-weights large language model of Zhipu AI (operating internationally as Z.ai, Hugging Face org `zai-org`), launched "targeting complex systems engineering and long-horizon agentic tasks" ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)). The GLM-5 weight repository was created on Hugging Face on 2026-02-11 ([Hugging Face model API for zai-org/GLM-5](https://huggingface.co/api/models/zai-org/GLM-5)), and secondary coverage dates the public release to February 11, 2026 ([NYU Shanghai RITS](https://rits.shanghai.nyu.edu/ai/glm-5-zhipu-ai-ships-a-744b-open-weight-frontier-model/)). Compared with GLM-4.5, GLM-5 scales from 355B parameters (32B active) to 744B parameters (40B active) and raises pre-training data from 23T to 28.5T tokens ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5); [Z.ai GLM-5 blog](https://z.ai/blog/glm-5)). Weights are distributed under the MIT license ([GLM-5 README](https://huggingface.co/zai-org/GLM-5/raw/main/README.md); [Hugging Face model API](https://huggingface.co/api/models/zai-org/GLM-5)), on Hugging Face and ModelScope, with API access on api.z.ai and BigModel.cn ([Z.ai GLM-5 blog](https://z.ai/blog/glm-5)).
+
+GLM-5.1 is the follow-up point release on the same architecture, described by Z.ai as "our next-generation flagship model for agentic engineering, with significantly stronger coding capabilities than its predecessor," reaching state-of-the-art performance on SWE-Bench Pro and leading GLM-5 by a wide margin on NL2Repo and Terminal-Bench 2.0 ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)). Its distinguishing claim is long-horizon durability: where earlier models "exhaust their repertoire early," GLM-5.1 "sustains optimization over hundreds of rounds and thousands of tool calls" ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)). The GLM-5.1 repository was created 2026-04-03 ([Hugging Face model API for zai-org/GLM-5.1](https://huggingface.co/api/models/zai-org/GLM-5.1)) and third-party documentation gives the public release date as April 7, 2026 ([Clore.ai GLM-5.1 guide](https://docs.clore.ai/guides/language-models/glm-5-1)); it is also MIT-licensed ([GLM-5.1 README](https://huggingface.co/zai-org/GLM-5.1/raw/main/README.md); [Z.ai GLM-5.1 blog](https://z.ai/blog/glm-5.1)). Both models are served through the Z.ai API with a 200K context window and 128K maximum output ([Z.ai docs: GLM-5](https://docs.z.ai/guides/llm/glm-5); [Z.ai docs: GLM-5.1](https://docs.z.ai/guides/llm/glm-5.1)).
+
+Notable techniques: GLM-5 integrates DeepSeek Sparse Attention (DSA), "largely reducing deployment cost while preserving long-context capacity" ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)), and the arXiv technical report _GLM-5: from Vibe Coding to Agentic Engineering_ (arXiv:2602.15763, submitted 17 Feb 2026) states that GLM-5 "adopts DSA" to cut training and inference cost while maintaining long-context fidelity, adds an asynchronous RL infrastructure that decouples generation from training, and proposes novel asynchronous agent RL algorithms ([arXiv:2602.15763](https://arxiv.org/abs/2602.15763)). That infrastructure is named **slime**, "a novel asynchronous RL infrastructure that substantially improves training throughput and efficiency, enabling more fine-grained post-training iterations" ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)). GLM-5.1 cites the same technical report as GLM-5, consistent with an architecture-sharing point release ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)). Beyond NVIDIA, Z.ai reports support on Huawei Ascend, Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame and Hygon accelerators ([Z.ai GLM-5 blog](https://z.ai/blog/glm-5)).
+
+## Architecture
+
+Values below come from the published `config.json` of each repository; GLM-5 and GLM-5.1 configs are byte-for-byte identical apart from the `transformers_version` field, confirming a shared architecture.
+
+- **Total parameters (vendor figure):** 744B ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5); [Z.ai docs: GLM-5](https://docs.z.ai/guides/llm/glm-5))
+- **Total parameters (checkpoint tensor count):** 753,864,139,008 (~754B) for both GLM-5 and GLM-5.1 ([GLM-5 HF API](https://huggingface.co/api/models/zai-org/GLM-5); [GLM-5.1 HF API](https://huggingface.co/api/models/zai-org/GLM-5.1))
+- **Active parameters per token:** 40B ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5); [Z.ai docs: GLM-5](https://docs.z.ai/guides/llm/glm-5))
+- **Model class / architecture string:** `GlmMoeDsaForCausalLM`, `model_type: glm_moe_dsa` ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json); [GLM-5.1 config.json](https://huggingface.co/zai-org/GLM-5.1/raw/main/config.json))
+- **Layers:** 78 hidden layers, with the first 3 layers dense (`first_k_dense_replace: 3`) and MoE in every subsequent layer (`moe_layer_freq: 1`) ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json))
+- **Hidden size / dense FFN:** hidden_size 6144, intermediate_size 12288 ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json))
+- **MoE experts:** 256 routed experts + 1 shared expert, 8 experts activated per token, expert FFN width 2048, sigmoid scoring with `noaux_tc` top-k routing and routed scaling factor 2.5 ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json); [GLM-5.1 config.json](https://huggingface.co/zai-org/GLM-5.1/raw/main/config.json))
+- **Attention:** DeepSeek Sparse Attention (DSA) ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5); [arXiv:2602.15763](https://arxiv.org/abs/2602.15763)); MLA-style low-rank projections with `q_lora_rank: 2048`, `kv_lora_rank: 512`, 64 attention heads, `qk_head_dim: 256` (192 nope + 64 rope), `v_head_dim: 256` ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json))
+- **DSA indexer:** 32 indexer heads, `index_head_dim: 128`, `index_topk: 2048` selected tokens per query, interleaved indexer RoPE ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json)); in DSA a lightweight "lightning indexer" picks the top-k relevant tokens, cutting core attention from O(L²) to O(Lk) ([IndexCache paper, arXiv:2603.12201](https://arxiv.org/abs/2603.12201))
+- **Multi-token prediction (MTP):** `num_nextn_predict_layers: 1`, used for speculative decoding (vLLM launch uses `--speculative-config.method mtp`, SGLang uses EAGLE) ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json); [GLM-5 model card](https://huggingface.co/zai-org/GLM-5))
+- **Context window:** `max_position_embeddings: 202752` in the checkpoint ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json)); Z.ai API lists a 200K context window and 128K max output for both GLM-5 and GLM-5.1 ([Z.ai docs: GLM-5](https://docs.z.ai/guides/llm/glm-5); [Z.ai docs: GLM-5.1](https://docs.z.ai/guides/llm/glm-5.1))
+- **RoPE:** default RoPE with theta 1,000,000, interleaved ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json))
+- **Precision:** BF16 checkpoint (`dtype: bfloat16`), plus an FP8 sibling repository `zai-org/GLM-5.1-FP8` listing `F8_E4M3` tensors at 754B parameters for cheaper serving ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json); [zai-org/GLM-5.1-FP8](https://huggingface.co/zai-org/GLM-5.1-FP8)); GLM-5.1 tensors are listed as BF16/F32 ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1))
+- **Vocabulary:** 154,880 tokens, untied embeddings ([GLM-5 config.json](https://huggingface.co/zai-org/GLM-5/raw/main/config.json))
+- **Pre-training data:** 28.5T tokens (up from 23T for GLM-4.5) ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5))
+- **License:** MIT ([GLM-5 README](https://huggingface.co/zai-org/GLM-5/raw/main/README.md); [GLM-5.1 README](https://huggingface.co/zai-org/GLM-5.1/raw/main/README.md))
+- **Serving stacks:** vLLM v0.19.0+, SGLang v0.5.10+, KTransformers v0.5.3+, Transformers, xLLM v0.8.0+ ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5); [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1))
+
+## Official vendor eval scores
+
+Vendor-reported only (Hugging Face model cards and Z.ai blog). GLM-5 column values are from the GLM-5 card/blog; GLM-5.1 values from the GLM-5.1 card/blog.
+
+| Benchmark                                                | Score                                                | Nemotron3 Ultra | % better | Source                                                                                                                 |
+| -------------------------------------------------------- | ---------------------------------------------------- | --------------- | -------- | ---------------------------------------------------------------------------------------------------------------------- |
+| SWE-bench Verified (GLM-5)                               | 77.8                                                 | 70.7            | +10.0%   | [GLM-5 model card](https://huggingface.co/zai-org/GLM-5) / [Z.ai GLM-5 blog](https://z.ai/blog/glm-5)                  |
+| SWE-bench Multilingual (GLM-5)                           | 73.3                                                 | 67.7            | +8.3%    | [GLM-5 model card](https://huggingface.co/zai-org/GLM-5)                                                               |
+| SWE-Bench Pro (GLM-5.1 / GLM-5)                          | 58.4 / 55.1                                          | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| NL2Repo (GLM-5.1 / GLM-5)                                | 42.7 / 35.9                                          | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| Terminal-Bench 2.0, Terminus-2 (GLM-5.1 / GLM-5)         | 63.5 / 56.2 (60.7 on verified set)                   | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1); [GLM-5 model card](https://huggingface.co/zai-org/GLM-5) |
+| Terminal-Bench 2.0, best self-reported harness (GLM-5.1) | 69.0 (Claude Code)                                   | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| τ²-Bench (GLM-5)                                         | 89.7                                                 | n.a.            | —        | [GLM-5 model card](https://huggingface.co/zai-org/GLM-5)                                                               |
+| τ³-Bench (GLM-5.1 / GLM-5)                               | 70.6 / 69.2                                          | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| BrowseComp (GLM-5.1 / GLM-5)                             | 68.0 / 62.0; with context management 79.3 / 75.9     | 44.4            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| BrowseComp-Zh (GLM-5)                                    | 72.7                                                 | n.a.            | —        | [GLM-5 model card](https://huggingface.co/zai-org/GLM-5)                                                               |
+| AIME 2026 (GLM-5.1 / GLM-5)                              | 95.3 / 95.4 (GLM-5 card reports 92.7 on AIME 2026 I) | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1); [GLM-5 model card](https://huggingface.co/zai-org/GLM-5) |
+| HMMT Nov. 2025 (GLM-5.1 / GLM-5)                         | 94.0 / 96.9                                          | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| HMMT Feb. 2026 (GLM-5.1 / GLM-5)                         | 82.6 / 82.8                                          | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| GPQA-Diamond (GLM-5.1 / GLM-5)                           | 86.2 / 86.0                                          | 87.0 (no tools) | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| IMOAnswerBench (GLM-5.1 / GLM-5)                         | 83.8 / 82.5                                          | 88.6 (no tools) | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| Humanity's Last Exam (GLM-5.1 / GLM-5)                   | 31.0 / 30.5 text-only; 52.3 / 50.4 with tools        | 26.7 (no tools) | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| CyberGym (GLM-5.1 / GLM-5)                               | 68.7 / 48.3 (GLM-5 card reports 43.2)                | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1); [GLM-5 model card](https://huggingface.co/zai-org/GLM-5) |
+| MCP-Atlas, public 500-task set (GLM-5.1 / GLM-5)         | 71.8 / 69.2 (GLM-5 card reports 67.8)                | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1); [GLM-5 model card](https://huggingface.co/zai-org/GLM-5) |
+| Tool-Decathlon (GLM-5.1 / GLM-5)                         | 40.7 / 38.0                                          | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| Vending Bench 2 (GLM-5.1 / GLM-5)                        | $5,634.41 / $4,432.12                                | n.a.            | —        | [GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)                                                           |
+| MMLU-Pro                                                 | n.a. (not reported on any fetched vendor page)       | 86.8            | —        | —                                                                                                                      |
+| LiveCodeBench                                            | n.a. (not reported on any fetched vendor page)       | 89.0 (v6)       | —        | —                                                                                                                      |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+Notes on evaluation harnesses (vendor-stated): SWE-bench runs use OpenHands with a 200K context window; Terminal-Bench 2.0 Claude Code scores are averaged over 5 runs; CyberGym is single-run Pass@1 over 1,507 tasks; Vending Bench 2 runs were conducted independently by Andon Labs ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)). GLM-5's τ²-Bench airline domain uses the fixes from the Claude Opus 4.5 system card ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)). Where the GLM-5 and GLM-5.1 cards disagree on a GLM-5 number (CyberGym, MCP-Atlas, AIME, Tool-Decathlon), both figures are shown above.
+
+## Benchmark explanations
+
+- **SWE-bench Verified** — human-validated GitHub issues that the model must resolve by patching a real repository; measures end-to-end software bug-fixing ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **SWE-bench Multilingual** — the same issue-resolution task across repositories in multiple programming languages, testing generality beyond Python ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **SWE-Bench Pro** — a harder, contamination-resistant issue-resolution suite (hosted by Scale AI) run here with the OpenHands agent under a 400K/200K context ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)).
+- **NL2Repo** — repository generation: build a working codebase from a natural-language specification ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)).
+- **Terminal-Bench 2.0** — real-world terminal tasks executed in a sandboxed shell under CPU/RAM limits, scored by task completion ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1); [GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **τ²-Bench / τ³-Bench** — multi-turn tool-use and customer-service style agent dialogues in Retail, Telecom and Airline domains, testing policy compliance with a simulated user ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **BrowseComp / BrowseComp-Zh** — hard web-browsing research questions requiring many search-and-read hops; the Zh variant is the Chinese-language version, with and without context-management strategies ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **AIME 2026** — American Invitational Mathematics Examination problems, short-answer competition math ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)).
+- **HMMT (Nov. 2025 / Feb. 2026)** — Harvard-MIT Mathematics Tournament problem sets, competition-level math reasoning ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)).
+- **IMOAnswerBench** — International-Mathematical-Olympiad-style problems graded on final answers ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)).
+- **GPQA-Diamond** — the hardest subset of graduate-level, Google-proof physics/chemistry/biology multiple-choice questions ([GLM-5.1 model card](https://huggingface.co/zai-org/GLM-5.1)).
+- **Humanity's Last Exam (HLE)** — a very broad expert-written exam of frontier knowledge and reasoning; reported text-only by default and separately with tool access ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **CyberGym** — starts from white-box source code and tests whether the model can identify and validate vulnerabilities by triggering faults ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)); evaluated as single-run Pass@1 over 1,507 tasks ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **MCP-Atlas** — 500-task public subset of tool-calling tasks over Model Context Protocol servers, LLM-judged ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5)).
+- **Tool-Decathlon** — a ten-domain battery of tool-use/agent tasks run through an official evaluation service ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **Vending Bench 2** — a simulated vending-machine business run over a one-year horizon, scored by final account balance; measures long-term planning and resource management ([Z.ai GLM-5 blog](https://z.ai/blog/glm-5)).
+- **VectorDBBench / KernelBench (GLM-5.1 case studies)** — building a high-QPS approximate-nearest-neighbour database in Rust (GLM-5.1: 21.5k QPS over 600+ iterations and 6,000+ tool calls) and generating faster GPU kernels from PyTorch references (GLM-5.1: 3.6× geometric-mean speedup over 50 problems) ([Z.ai GLM-5.1 blog](https://z.ai/blog/glm-5.1)).
diff --git a/packages/app/content/models/glm-5-2.mdx b/packages/app/content/models/glm-5-2.mdx
new file mode 100644
index 000000000..f7b0d24b8
--- /dev/null
+++ b/packages/app/content/models/glm-5-2.mdx
@@ -0,0 +1,102 @@
+---
+title: 'GLM-5.2 / GLM-5.3'
+developer: 'Z.ai (Zhipu AI)'
+releaseDate: 'June 16, 2026 (GLM-5.2); August 14, 2026 (GLM-5.3)'
+description: "Z.ai's 744B-class MoE line extending sparse attention with IndexShare for 1M-token contexts, with GLM-5.3 adding a major agentic post-training pass."
+---
+
+## Overview
+
+GLM-5.2 is Z.ai's (Zhipu AI's) flagship long-horizon model in the GLM-5 family, introduced as "our latest flagship model for long-horizon tasks," marking "a substantial leap in long-horizon task capability over its predecessor GLM-5.1" and, for the first time in the family, delivering it on "a solid 1M-token context" ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)). The weight repository was created on Hugging Face on 2026-06-16 ([Hugging Face model API for zai-org/GLM-5.2](https://huggingface.co/api/models/zai-org/GLM-5.2)), matching Z.ai's own launch announcement of "GLM-5.2: Frontier Intelligence, Open Weights" with "MIT-licensed open weights" and "same API pricing as GLM-5.1" ([Z.ai announcement](https://x.com/Zai_org/status/2066938937344495629)). Its headline features are the solid 1M context, "advanced coding with flexible effort" (multiple thinking-effort levels), the **IndexShare** architecture change, an improved MTP layer for speculative decoding, and "Pure Open: an MIT open-source license — no regional limits, technical access without borders" ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2); [GLM-5.2 README](https://huggingface.co/zai-org/GLM-5.2/raw/main/README.md)). Z.ai positions it as the highest-ranked open-source model across three long-horizon coding benchmarks (FrontierSWE, PostTrainBench, SWE-Marathon) and the strongest open-source model on standard coding benchmarks ([Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2)).
+
+GLM-5.3, announced on August 14, 2026 ([Yicai Global](https://www.yicaiglobal.com/news/chinas-zai-unveils-glm-53-says-it-leads-open-source-ai-models-in-key-benchmarks)), is a post-training-only release: "It uses the same base model as GLM-5.2 — every gain comes from post-training" ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [Z.ai docs: GLM-5.3](https://docs.z.ai/guides/llm/glm-5.3)). Z.ai frames it as "Frontier Coding with Emergent Cyber Capabilities": a ~50% coding gain over GLM-5.2 on the in-house Z.ai Code Bench, claimed open-source SOTA on Terminal-Bench 3.0 and Agents' Last Exam (CLI), and unexpectedly fast growth in vulnerability discovery and exploitation ability ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [Z.ai docs: GLM-5.3](https://docs.z.ai/guides/llm/glm-5.3)). Working with security teams in China, Z.ai reports the model surfaced 2,436 vulnerabilities across 269 projects, including 1,097 medium-to-high severity issues, with an average latent lifetime of 26.6 years and the oldest flaw introduced in 1981 ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)). GLM-5.3 is text-only, always reasoning, with a 1M-token context, 128K max output, and three effort levels (`low`, `high`, `max`, default `max`); disabling thinking is no longer supported ([Z.ai docs: GLM-5.3](https://docs.z.ai/guides/llm/glm-5.3)).
+
+Licensing/availability differs between the two. GLM-5.2 ships as MIT open weights on Hugging Face and ModelScope ([Hugging Face model API](https://huggingface.co/api/models/zai-org/GLM-5.2); [Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2)). GLM-5.3 weights are not yet public at the time of writing: Z.ai states "The model weights of GLM-5.3 will be publicly available soon in two weeks," with access initially via the Z.ai API, the GLM Coding Plan, and ZCode ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [Z.ai docs: GLM-5.3](https://docs.z.ai/guides/llm/glm-5.3)). Notable techniques across the pair: IndexShare for cheap long-context sparse attention, SAO (with compaction) for RL on long-horizon tasks, and **slime**, Z.ai's open-source asynchronous RL post-training framework with Megatron for training and SGLang for rollout — GLM-5.3 reports >2.3× end-to-end RL throughput gains from system-level optimizations and training-rollout logprob agreement controlled to the 1e-7 level ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)). Independent evaluation from Artificial Analysis placed GLM-5.3 at 60 on its Intelligence Index on August 18, 2026, level with Kimi K3 ([Unite.AI](https://www.unite.ai/glm-5-3-scores-60-on-artificial-analysis-intelligence-index-matching-kimi-k3/)).
+
+## Architecture
+
+GLM-5.3 shares GLM-5.2's base model exactly ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)), so the GLM-5.2 checkpoint config is the authoritative architecture source for both.
+
+- **Total parameters (checkpoint tensor count):** 753,329,940,480 (~753B) ([Hugging Face model API for zai-org/GLM-5.2](https://huggingface.co/api/models/zai-org/GLM-5.2)); the family's vendor-stated headline figure is 744B total / 40B active, stated for GLM-5 on the same architecture lineage ([GLM-5 model card](https://huggingface.co/zai-org/GLM-5); [Z.ai docs: GLM-5](https://docs.z.ai/guides/llm/glm-5))
+- **Active parameters per token:** n.a. for GLM-5.2/5.3 specifically — not stated on any fetched vendor page ([Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2); [GLM-5.2 README](https://huggingface.co/zai-org/GLM-5.2/raw/main/README.md))
+- **Model class:** `GlmMoeDsaForCausalLM`, `model_type: glm_moe_dsa` ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json))
+- **Layers:** 78 hidden layers; `mlp_layer_types` shows the first 3 layers dense and the remaining 75 sparse (MoE) ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json))
+- **Hidden size / dense FFN:** hidden_size 6144, intermediate_size 12288 ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json))
+- **MoE experts:** 256 routed experts + 1 shared expert, 8 activated per token, expert FFN width 2048, sigmoid scoring with `noaux_tc` routing, routed scaling factor 2.5 ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json))
+- **Attention:** DSA-based sparse attention with MLA-style low-rank projections — 64 attention heads, `q_lora_rank: 2048`, `kv_lora_rank: 512`, `qk_head_dim: 256` (192 nope + 64 rope), `v_head_dim: 256`, `head_dim: 192` ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json)); DSA's lightning indexer selects top-k tokens per query, reducing core attention from O(L²) to O(Lk) while the indexer itself stays O(L²) ([IndexCache paper, arXiv:2603.12201](https://arxiv.org/abs/2603.12201))
+- **IndexShare (novel component):** "reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length" ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)); the indexer sits at the first of each group of 4 layers and its top-k indices are reused for all 4, removing indexer dot-product and top-k work in 3/4 of layers; trained with IndexShare from mid-training at 128K sequence length ([Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2))
+- **IndexShare in config:** `index_topk_freq: 4`, `index_skip_topk_offset: 3`, `index_share_for_mtp_iteration: true`, and an `indexer_types` list alternating one `full` layer followed by three `shared` layers across all 78 layers ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json))
+- **DSA indexer sizing:** 32 indexer heads, `index_head_dim: 128`, `index_topk: 2048`, interleaved indexer RoPE ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json))
+- **Multi-token prediction (MTP):** `num_nextn_predict_layers: 1`; the MTP layer was improved for speculative decoding, raising acceptance length by up to 20% — ablation: baseline 4.56 → +IndexShare+KV Share 5.10 → +rejection sampling 5.29 → +end-to-end TV loss 5.47; MTP step parameters are shared as in GLM-5.1, and in multi-step MTP the indexer runs on the first step with its top-k indices reused for later steps ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json); [Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2))
+- **Context window:** `max_position_embeddings: 1048576` (1M) in the checkpoint ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json)); Z.ai API lists 1M context / 128K max output for GLM-5.2 and for GLM-5.3 ([Z.ai docs: GLM-5.2](https://docs.z.ai/guides/llm/glm-5.2); [Z.ai docs: GLM-5.3](https://docs.z.ai/guides/llm/glm-5.3)). GLM-5.1's context was 200K, so 5.2 is a 5× extension ([Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2))
+- **RoPE:** default RoPE with theta raised to 8,000,000 (vs 1,000,000 in GLM-5/5.1), interleaved ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json); [GLM-5.1 config.json](https://huggingface.co/zai-org/GLM-5.1/raw/main/config.json))
+- **Precision:** BF16 checkpoint (`dtype: bfloat16`); FP8 KV-cache is used in Z.ai's large-scale agentic RL training infrastructure ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json); [Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2))
+- **Vocabulary:** 154,880 tokens, untied embeddings ([GLM-5.2 config.json](https://huggingface.co/zai-org/GLM-5.2/raw/main/config.json))
+- **Effort control:** GLM-5.2 exposes High and Max thinking effort ([Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2)); GLM-5.3 exposes `low`/`high`/`max` with reasoning always enabled ([Z.ai docs: GLM-5.3](https://docs.z.ai/guides/llm/glm-5.3))
+- **License:** MIT for GLM-5.2 ([Hugging Face model API](https://huggingface.co/api/models/zai-org/GLM-5.2); [GLM-5.2 README](https://huggingface.co/zai-org/GLM-5.2/raw/main/README.md)); GLM-5.3 weights pending release ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3))
+- **Serving stacks:** SGLang v0.5.13.post1+, vLLM v0.23.0+, Transformers v0.5.12+, KTransformers v0.5.12+, Unsloth v0.1.47-beta+, plus vLLM-Ascend/xLLM/SGLang on Ascend NPU ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)); inference optimization builds on LayerSplit, with a throughput advantage that grows with context length ([Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2))
+
+## Official vendor eval scores
+
+Vendor-reported only. GLM-5.3 figures come from the Z.ai GLM-5.3 blog table (Z.ai ran the comparisons itself); GLM-5.2 figures from the GLM-5.2 Hugging Face card / Z.ai blog.
+
+| Benchmark                              | Score                                                                                              | Nemotron3 Ultra   | % better | Source                                                                                                       |
+| -------------------------------------- | -------------------------------------------------------------------------------------------------- | ----------------- | -------- | ------------------------------------------------------------------------------------------------------------ |
+| Terminal-Bench 3.0                     | GLM-5.3: 28.3 · GLM-5.2: 4.6                                                                       | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| Terminal-Bench 2.1                     | GLM-5.3: 88.2 · GLM-5.2: 81.0 (Terminus-2), 82.7 (best harness, Claude Code)                       | 56.4              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| DeepSWE v1.1                           | GLM-5.3: 66.9 · GLM-5.2: 46.2                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| SWE-bench Pro                          | GLM-5.2: 62.1 (GLM-5.1: 58.4) — not reported for GLM-5.3                                           | n.a.              | —        | [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)                                                 |
+| SWE-Marathon v1.1                      | GLM-5.3: 42.5 · GLM-5.2: 19.4 (GLM-5.2 card reports 13.0 for SWE-Marathon)                         | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| FrontierSWE                            | GLM-5.3: 78.1 · GLM-5.2: 67.5 (GLM-5.2 card reports 74.4 dominance as of 2026-06-16)               | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| NL2Repo                                | GLM-5.3: 58.0 · GLM-5.2: 48.9                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| ProgramBench (Almost Solved)           | GLM-5.3: 19.0 · GLM-5.2: 9.5 (GLM-5.2 card reports 63.7 on ProgramBench)                           | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| PostTrainBench                         | GLM-5.3: 39.8 · GLM-5.2: 31.7 (GLM-5.2 card reports 34.3)                                          | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| Agents' Last Exam (ALE-CLI)            | GLM-5.3: 28.5 · GLM-5.2: 23.8                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| Toolathlon Verified                    | GLM-5.3: 73.0 · GLM-5.2: 59.9                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| AutomationBench v1.0.6                 | GLM-5.3: 48.2 · GLM-5.2: 26.2                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| CyberGym                               | GLM-5.3: 84.5 · GLM-5.2: 77.2                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| ExploitBench                           | GLM-5.3: 54.4 · GLM-5.2: 24.4                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| ExploitGym (2h / 6h tasks completed)   | GLM-5.3: 105 / 130 · GLM-5.2: 29 / 39                                                              | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| Humanity's Last Exam (with tools)      | GLM-5.3: 62.5 · GLM-5.2: 54.7 (GLM-5.2 text-only: 40.5)                                            | 37.4 (with tools) | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| GDPval-AA v2 (Elo)                     | GLM-5.3: 1769 · GLM-5.2: 1508                                                                      | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| AIME 2026                              | GLM-5.2: 99.2 — not reported for GLM-5.3                                                           | n.a.              | —        | [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)                                                 |
+| HMMT Nov. 2025 / Feb. 2026             | GLM-5.2: 94.4 / 92.5 — not reported for GLM-5.3                                                    | n.a.              | —        | [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)                                                 |
+| GPQA-Diamond                           | GLM-5.2: 91.2 — not reported for GLM-5.3                                                           | 87.0 (no tools)   | —        | [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)                                                 |
+| IMOAnswerBench                         | GLM-5.2: 91.0 — not reported for GLM-5.3                                                           | 88.6 (no tools)   | —        | [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)                                                 |
+| CritPt                                 | GLM-5.2: 20.9 (blog) / 16.7 (model card)                                                           | 3.1 (no tools)    | —        | [Z.ai GLM-5.2 blog](https://z.ai/blog/glm-5.2); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2) |
+| MCP-Atlas (public set)                 | GLM-5.2: 76.8 — not reported for GLM-5.3                                                           | n.a.              | —        | [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)                                                 |
+| Tool-Decathlon                         | GLM-5.2: 48.2 — not reported for GLM-5.3                                                           | n.a.              | —        | [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)                                                 |
+| Z.ai Code Bench (private, High effort) | GLM-5.3: 31.4% at ~50K output tokens/task (Max effort: 34.5% at ~75K); GLM-5.2: 23.4% at 96K (Max) | n.a.              | —        | [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)                                                               |
+| SWE-bench Verified                     | n.a. — not reported for GLM-5.2 or GLM-5.3 on any fetched vendor page                              | 70.7              | —        | —                                                                                                            |
+| τ²-bench / τ³-Bench                    | n.a. — not reported for GLM-5.2 or GLM-5.3 on any fetched vendor page                              | n.a.              | —        | —                                                                                                            |
+| BrowseComp                             | n.a. — not reported for GLM-5.2 or GLM-5.3 on any fetched vendor page                              | 44.4              | —        | —                                                                                                            |
+| LiveCodeBench                          | n.a. — not reported on any fetched vendor page                                                     | 89.0 (v6)         | —        | —                                                                                                            |
+| MMLU-Pro                               | n.a. — not reported on any fetched vendor page                                                     | 86.8              | —        | —                                                                                                            |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+Vendor harness notes: GLM-5.2 reasoning evals use temperature 1.0 / top_p 0.95, 163,840-token generation limit, GPT-5.5 (medium) as judge, text-only subsets by default; SWE-Bench Pro uses OpenHands under a 400K context; DeepSWE uses the official pier framework with mini-swe-agent; Terminal-Bench 2.1 Claude Code scores are averaged over 5 runs; FrontierSWE was evaluated by Proximal, PostTrainBench by PostTrainBench, and SWE-Marathon by Abundant AI, all at 1M context and max effort ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)). GLM-5.3's HLE-with-tools run used a 300K context with context management and GPT-5.6-luna (medium) as judge ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+
+## Benchmark explanations
+
+- **Terminal-Bench 2.1 / 3.0** — real-world command-line tasks solved autonomously in a sandboxed terminal; 3.0 is the harder successor generation, described by Z.ai's docs as measuring complex task completion in a real terminal environment ([Z.ai docs: GLM-5.3](https://docs.z.ai/guides/llm/glm-5.3); [GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **DeepSWE v1.1** — long-horizon software-engineering task suite run with the official pier framework and mini-swe-agent harness, each task in an isolated no-internet container ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **SWE-bench Pro** — contamination-resistant GitHub issue resolution requiring repo-level patches, run with OpenHands under a 400K context ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **SWE-Marathon** — very long-horizon software engineering evaluated by Abundant AI at 1M context, testing sustained multi-hour work ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **FrontierSWE** — frontier-difficulty software-engineering tasks scored as a "dominance" rate, evaluated by Proximal at 1M context and max effort ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **NL2Repo** — generate an entire working repository from a natural-language specification, with rule- and LLM-based anti-cheat checks ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **ProgramBench** — 200 hard programming instances run through Claude Code with up to 2,000 turns and 6-hour timeouts; "Almost Solved" is the stricter partial-credit tier ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2); [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **PostTrainBench** — machine-learning post-training research tasks (run the experiment, improve a model) evaluated by the benchmark authors at 1M context ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **Agents' Last Exam (ALE-CLI)** — real-world professional agent scenarios emphasizing cross-tool collaboration and long-horizon execution, scored pass@1 in a CLI harness ([Yicai Global](https://www.yicaiglobal.com/news/chinas-zai-unveils-glm-53-says-it-leads-open-source-ai-models-in-key-benchmarks); [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **Toolathlon Verified** — verified multi-tool "chores" requiring correct orchestration of many external tools ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **AutomationBench v1.0.6** — end-to-end automation of real workflow tasks by an agent ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **CyberGym** — starts from white-box source code and tests whether the model can identify and validate vulnerabilities by triggering faults ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **ExploitBench** — requires deeper reasoning about real vulnerabilities and producing working exploitation, i.e. further up the exploitation chain than CyberGym ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **ExploitGym** — counts how many exploitation tasks a model completes under time-normalized budgets (2-hour and 6-hour), normalized by per-model throughput ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **Humanity's Last Exam (HLE)** — broad expert-written exam of frontier knowledge and reasoning; the "with tools" variant permits search/tool use ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2); [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **GDPval-AA v2** — Elo-style evaluation of real-world, economically valuable knowledge work spanning 44 occupations ([Yicai Global](https://www.yicaiglobal.com/news/chinas-zai-unveils-glm-53-says-it-leads-open-source-ai-models-in-key-benchmarks); [Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
+- **AIME 2026 / HMMT / IMOAnswerBench** — competition mathematics (American Invitational Mathematics Examination, Harvard-MIT Mathematics Tournament, IMO-style problems) graded on exact final answers ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **GPQA-Diamond** — hardest subset of graduate-level, Google-proof science multiple-choice questions ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **CritPt** — frontier physics-reasoning research problems, reported alongside HLE in the reasoning block ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **MCP-Atlas** — 500-task public subset of Model Context Protocol tool-calling tasks, LLM-judged (Gemini 3.0 Pro as judge) ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **Tool-Decathlon** — ten-domain agentic tool-use battery run through the official evaluation service with 128K max tokens ([GLM-5.2 model card](https://huggingface.co/zai-org/GLM-5.2)).
+- **Z.ai Code Bench** — Z.ai's private benchmark placing coding agents in complex local development environments, scoring end-to-end task completion and checklist accuracy plus output-token efficiency ([Z.ai GLM-5.3 blog](https://z.ai/blog/glm-5.3)).
diff --git a/packages/app/content/models/gptoss-120b.mdx b/packages/app/content/models/gptoss-120b.mdx
new file mode 100644
index 000000000..f702266d1
--- /dev/null
+++ b/packages/app/content/models/gptoss-120b.mdx
@@ -0,0 +1,69 @@
+---
+title: 'gpt-oss-120b'
+developer: 'OpenAI'
+releaseDate: 'August 5, 2025'
+description: "OpenAI's 117B-parameter / 5.1B-active open-weight MoE with alternating dense and sliding-window attention and native MXFP4 quantization, under Apache 2.0."
+---
+
+## Overview
+
+gpt-oss-120b is an open-weight reasoning language model released by OpenAI on August 5, 2025, alongside the smaller gpt-oss-20b, both distributed under the Apache 2.0 license ([OpenAI, "Introducing gpt-oss"](https://openai.com/index/introducing-gpt-oss/)). The Hugging Face model card confirms the Apache 2.0 license and lists the model at 117B total parameters with 5.1B active parameters ([Hugging Face: openai/gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b)). OpenAI's accompanying model card describes both models as open-weight reasoning models trained with large-scale distillation and reinforcement learning, and states that weights, inference implementations, tool environments, and tokenizers are all released under Apache 2.0 ([arXiv:2508.10925 abstract](https://arxiv.org/abs/2508.10925)).
+
+Positioning: OpenAI reports that gpt-oss-120b reaches near-parity with OpenAI o4-mini on core reasoning benchmarks, outperforms o3-mini and matches or exceeds o4-mini on Codeforces competition coding, MMLU, HLE, and TauBench tool calling, and beats o4-mini on AIME 2024/2025 competition mathematics and HealthBench ([OpenAI, "Introducing gpt-oss"](https://openai.com/index/introducing-gpt-oss/)). The models are optimized for deep research browsing, Python tool use, and developer-provided functions, and use a rendered chat format designed for clear instruction following and role delineation ([arXiv:2508.10925 abstract](https://arxiv.org/abs/2508.10925)).
+
+Notable techniques: the model is an autoregressive mixture-of-experts transformer that alternates banded-window and fully dense attention, uses Grouped Query Attention with learned attention-sink biases in the softmax denominator, RoPE positional embeddings with YaRN context extension, and gated SwiGLU MoE activations ([OpenAI gpt-oss model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)). Its MoE weights — more than 90% of all parameters — are natively quantized to MXFP4 at 4.25 bits per parameter, which lets the 120b model fit on a single 80 GB GPU such as an NVIDIA H100 or AMD MI300X ([OpenAI gpt-oss model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [Hugging Face model card](https://huggingface.co/openai/gpt-oss-120b)).
+
+## Architecture
+
+- Model type: autoregressive Mixture-of-Experts transformer; HF architecture class `GptOssForCausalLM`, model type `gpt_oss` ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Total parameters: 116.83B (≈117B as listed on the HF card) ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [Hugging Face](https://huggingface.co/openai/gpt-oss-120b))
+- Active parameters per token: 5.13B (≈5.1B) ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [Hugging Face](https://huggingface.co/openai/gpt-oss-120b))
+- Parameter breakdown: MLP 114.71B, attention 0.96B, embed + unembed 1.16B; checkpoint size 60.8 GiB ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf))
+- Layers: 36 (`num_hidden_layers: 36`) ([OpenAI blog](https://openai.com/index/introducing-gpt-oss/); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Residual stream / hidden size: 2,880 ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Experts: 128 MoE experts per block, top-4 selected per token (softmax-weighted over selected experts via a linear router); `num_local_experts: 128`, `num_experts_per_tok: 4` ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- MoE activation: gated SwiGLU (SwiGLU limit 7.0 in config) ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Attention pattern: alternates banded-window (sliding) and fully dense attention; config lists a strict `sliding_attention` / `full_attention` alternation across all 36 layers ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Sliding-window bandwidth: 128 tokens ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Attention heads: 64 query heads per layer, head dim 64, 8 key-value heads — Grouped Query Attention (group size 8) ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json); [OpenAI blog](https://openai.com/index/introducing-gpt-oss/))
+- Attention sinks: each attention head has a learned bias in the softmax denominator (off-by-one attention), letting heads attend to nothing ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf))
+- Positional embeddings: RoPE with YaRN scaling (`rope_theta: 150000`, scaling factor 32.0, original max positions 4096) ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Context window: 131,072 tokens for dense layers (128k native) ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json); [OpenAI blog](https://openai.com/index/introducing-gpt-oss/))
+- Vocabulary size: 201,088 ([config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json))
+- Quantization: MXFP4 on MoE weights at 4.25 bits/parameter (>90% of parameters); config `quant_method: mxfp4` with self-attention, router, embeddings and lm_head left unquantized; tensor types BF16 · U8 ([model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); [config.json](https://huggingface.co/openai/gpt-oss-120b/raw/main/config.json); [Hugging Face](https://huggingface.co/openai/gpt-oss-120b))
+- Hardware footprint: runs within 80 GB of memory / a single 80GB GPU ([OpenAI blog](https://openai.com/index/introducing-gpt-oss/); [Hugging Face](https://huggingface.co/openai/gpt-oss-120b))
+
+## Official vendor eval scores
+
+All values are OpenAI-reported for gpt-oss-120b. Scores vary by reasoning effort (low / medium / high); the table reports high reasoning effort, with medium shown where the HF card also lists it.
+
+| Benchmark                               | Score                                                                | Nemotron3 Ultra   | % better | Source                                                                                                                                                                                              |
+| --------------------------------------- | -------------------------------------------------------------------- | ----------------- | -------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| AIME 2024 (no tools, high reasoning)    | 95.8% (low 56.3%, medium 80.4%)                                      | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| AIME 2024 (with tools, high)            | 96.6%                                                                | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| AIME 2025 (no tools, high)              | 92.5% (low 50.4%, medium 80.0%)                                      | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| AIME 2025 (with tools, high)            | 97.9%                                                                | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| GPQA Diamond (no tools, high)           | 80.1% (medium 73.1%)                                                 | 87.0 (no tools)   | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); medium also on [Hugging Face card](https://huggingface.co/openai/gpt-oss-120b) |
+| GPQA Diamond (with tools, high)         | 80.9% (medium 73.5%)                                                 | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); medium also on [Hugging Face card](https://huggingface.co/openai/gpt-oss-120b) |
+| MMLU (high)                             | 90.0% (low 85.9%, medium 88.0%)                                      | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| Humanity's Last Exam (no tools, high)   | 14.9%                                                                | 26.7 (no tools)   | -44.2%   | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| Humanity's Last Exam (with tools, high) | 19.0%                                                                | 37.4 (with tools) | -49.2%   | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| SWE-bench Verified (high)               | 62.4% (low 47.9%, medium 52.6%)                                      | 70.7              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf); medium also on [Hugging Face card](https://huggingface.co/openai/gpt-oss-120b) |
+| Codeforces Elo (no tools, high)         | 2,463 Elo (low 1,595; medium 2,205)                                  | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| Codeforces Elo (with tools, high)       | 2,622 Elo (low 1,653; medium 2,365)                                  | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| τ-Bench Retail (high)                   | 67.8% (low 49.4%, medium 62.0%)                                      | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| HealthBench (high)                      | 57.6% (Hard 30.0%, Consensus 89.9%)                                  | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+| BFCL / IFEval                           | n.a. — not reported in the OpenAI model card or HF card fetched here | n.a.              | —        | [OpenAI model card PDF](https://cdn.openai.com/pdf/419b6906-9da6-406c-a19d-1bb078ac7637/oai_gpt-oss_model_card.pdf)                                                                                 |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+
+## Benchmark explanations
+
+- **AIME 2024 / AIME 2025** — American Invitational Mathematics Examination problem sets; measures competition-level mathematical reasoning, scored as accuracy with and without tool access ([OpenAI blog](https://openai.com/index/introducing-gpt-oss/)).
+- **GPQA Diamond** — the hardest subset of graduate-level, Google-proof science multiple-choice questions (biology, chemistry, physics); measures expert-level scientific reasoning ([Hugging Face card links the GPQA dataset](https://huggingface.co/openai/gpt-oss-120b)).
+- **MMLU** — Massive Multitask Language Understanding: broad multiple-choice knowledge and general problem solving across academic and professional subjects ([OpenAI blog](https://openai.com/index/introducing-gpt-oss/)).
+- **Humanity's Last Exam (HLE)** — extremely difficult expert-written questions spanning many disciplines; measures frontier general problem solving ([OpenAI blog](https://openai.com/index/introducing-gpt-oss/)).
+- **SWE-bench Verified** — human-validated real GitHub issues; measures whether a model can produce patches that resolve software bugs in actual repositories ([Hugging Face card links the SWE-bench Verified dataset](https://huggingface.co/openai/gpt-oss-120b)).
+- **Codeforces (Elo)** — competitive programming rating derived from Codeforces contest problems; measures algorithmic coding skill on a rating scale rather than a percentage ([OpenAI blog](https://openai.com/index/introducing-gpt-oss/)).
+- **τ-Bench Retail (Tau-Bench)** — agentic tool-calling evaluation in a retail customer-service environment; measures multi-turn function calling and policy-following ([OpenAI blog describes TauBench as a tool-calling / agentic evaluation suite](https://openai.com/index/introducing-gpt-oss/)).
+- **HealthBench / Hard / Consensus** — health-conversation evaluations scoring the quality and safety of medical responses, with a harder subset and a physician-consensus variant ([OpenAI blog](https://openai.com/index/introducing-gpt-oss/)).
diff --git a/packages/app/content/models/kimi-k26.mdx b/packages/app/content/models/kimi-k26.mdx
new file mode 100644
index 000000000..8b2b51dbd
--- /dev/null
+++ b/packages/app/content/models/kimi-k26.mdx
@@ -0,0 +1,94 @@
+---
+title: 'Kimi K2.5 / K2.6 / K2.7-Code'
+developer: 'Moonshot AI'
+releaseDate: 'January 27, 2026 (K2.5); April 20, 2026 (K2.6); June 12, 2026 (K2.7-Code)'
+description: "Moonshot AI's 1T-parameter / 32B-active MoE family sharing a DeepSeek-V3-style MLA backbone with native INT4 quantization and a 256K context window."
+---
+
+## Overview
+
+The Kimi K2.5 → K2.6 → K2.7-Code line is Moonshot AI's 1-trillion-parameter open-weights MoE family, all three sharing an identical architecture: 1T total parameters, 32B activated per token, 61 layers, MLA attention, 384 routed experts with 8 selected per token plus 1 shared expert, a 256K context window and a 400M-parameter MoonViT vision encoder ([Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6); [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)). Moonshot states this explicitly on the later cards — "Kimi-K2.6 has the same architecture as Kimi-K2.5, and the deployment method can be directly reused" and "Kimi-K2.7-Code has the same architecture as Kimi-K2.5/Kimi-K2.6" ([Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6); [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)). The text backbone is DeepSeek-style: K2.7-Code's `config.json` declares the text architecture as `DeepseekV3ForCausalLM` with `topk_method: noaux_tc`, sigmoid expert scoring, YaRN RoPE scaling and MLA-style `q_lora_rank`/`kv_lora_rank` projections ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json)).
+
+Release timeline, per Moonshot's own channels: **Kimi K2.5** was published 2026-01-27 and is an open-source native multimodal agentic model built by continual pretraining on ~15 trillion mixed visual and text tokens on top of Kimi-K2-Base, adding instant and thinking modes and the Agent Swarm execution scheme ([Kimi Research & Tech Blog](https://www.kimi.ai/blog/); [Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)); its card also logs a 2026.1.29 changelog removing the default system prompt and fixing the `<|media_begin|>` token ([Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)). **Kimi K2.6** was published 2026-04-20 and pushes long-horizon coding, coding-driven design, proactive autonomous execution and swarm orchestration scaling to 300 sub-agents executing 4,000 coordinated steps ([Kimi Research & Tech Blog](https://www.kimi.ai/blog/); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)). **Kimi K2.7-Code** was announced on Moonshot's official X account on June 12, 2026 as "our latest coding model … now released and open-sourced," with +21.8% on Kimi Code Bench v2, +11.0% on Program Bench and +31.5% on MLS Bench Lite versus K2.6, 30% lower reasoning-token usage, and availability via the Kimi API and Kimi Code ([@Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2065377579130142937)).
+
+Positioning shifts across the family from general multimodal agent to coding specialist. K2.5 is framed around native multimodality, coding from visual specifications, and an agent swarm that decomposes tasks into parallel domain-specific sub-agents ([Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)). K2.7-Code is explicitly "a coding-focused agentic model built upon Kimi K2.6," strengthening end-to-end completion of complex software-engineering workflows while cutting thinking tokens ~30% versus K2.6; it forces thinking and `preserve_thinking` on (they cannot be disabled), unlike K2.6 where preserve-thinking is off by default and an instant (non-thinking) mode exists ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)). All three ship under the Modified MIT License for both code repository and weights ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)). Notable techniques include native INT4 quantization ("the same native int4 quantization method as Kimi-K2-Thinking"), interleaved thinking with multi-step tool calls, and Kimi Code CLI as the recommended agent framework; supported inference engines are vLLM, SGLang and KTransformers with `transformers >=4.57.1, <5.0.0` ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)). Moonshot's API platform lists dedicated pricing pages for `kimi-k2.7-code` and `kimi-k2.6`, and notes that `kimi-k2.5` and the `moonshot-v1` series are no longer available to newly registered users following the K3 launch ([Kimi API pricing docs](https://platform.kimi.ai/docs/pricing); [Kimi API model list](https://platform.kimi.ai/docs/models)).
+
+## Architecture
+
+Shared by K2.5, K2.6 and K2.7-Code unless noted.
+
+- **Architecture family:** Mixture-of-Experts (MoE), native multimodal (image + video input, text output) ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5))
+- **Total parameters:** 1T; Hugging Face reports the K2.7-Code safetensors model size as 1T params ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code))
+- **Activated parameters per token:** 32B ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6); [Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5))
+- **Layers:** 61 including 1 dense layer ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)); `num_hidden_layers: 61`, `first_k_dense_replace: 1` ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Attention mechanism:** MLA (Multi-head Latent Attention) ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)); config: `q_lora_rank: 1536`, `kv_lora_rank: 512`, `qk_nope_head_dim: 128`, `qk_rope_head_dim: 64`, `v_head_dim: 128` ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Attention hidden dimension / heads:** 7168 hidden, 64 attention heads ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)); `hidden_size: 7168`, `num_attention_heads: 64`, `num_key_value_heads: 64` ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Experts:** 384 routed experts, 8 selected per token, 1 shared expert ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)); `n_routed_experts: 384`, `num_experts_per_tok: 8`, `n_shared_experts: 1`, `moe_layer_freq: 1` ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **MoE dimensions / routing:** MoE hidden dimension per expert 2048 ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)); `moe_intermediate_size: 2048`, dense `intermediate_size: 18432`, sigmoid expert scoring, `topk_method: noaux_tc`, `norm_topk_prob: true`, `routed_scaling_factor: 2.827`, `aux_loss_alpha: 0.001` — the DeepSeek-V3-style MoE recipe ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Text backbone class:** `DeepseekV3ForCausalLM` inside `KimiK25ForConditionalGeneration` (`model_type: kimi_k25`, text `model_type: kimi_k2`) ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Activation function:** SwiGLU per the model cards; `hidden_act: silu` in config ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Context window:** 256K per the cards; `max_position_embeddings: 262144` ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json)). Evaluations were run at a 262,144-token context length ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code))
+- **Positional encoding:** YaRN RoPE scaling, `rope_theta: 50000.0`, factor 64.0, `original_max_position_embeddings: 4096`, beta_fast 32.0 / beta_slow 1.0 ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Vocabulary:** 160K per the cards; `vocab_size: 163840` ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Vision encoder:** MoonViT, 400M parameters ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)); vision config: 27 layers, 16 heads, hidden size 1152, intermediate 4304, `patch_size: 14`, `patchmerger` projector, `video_attn_type: spatial_temporal` ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Precision / quantization:** native INT4 — same method as Kimi-K2-Thinking ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)); config declares `compressed-tensors` `pack-quantized`, `num_bits: 4`, `group_size: 32`, symmetric, with attention, shared experts, dense MLP, `lm_head`, vision tower and mm projector excluded; published tensor types are BF16/F32/I32 ([Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json); [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code))
+- **Novel components (KDA, LatentMoE, SiTU-GLU):** n.a. — these are Kimi K3 components and are not present in the K2.x cards or config, which specify MLA + SwiGLU + DeepSeek-style MoE instead ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.7-Code config.json](https://huggingface.co/moonshotai/Kimi-K2.7-Code/raw/main/config.json))
+- **Thinking behavior:** K2.7-Code forces thinking and `preserve_thinking` to True and does not support instant mode; K2.6/K2.5 support instant vs thinking modes with `preserve_thinking` off by default ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6))
+- **Recommended sampling:** temperature 1.0, top-p 0.95 in thinking mode (K2.6 instant mode: 0.6) ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code); [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6))
+- **Pretraining data (K2.5):** continual pretraining on ~15 trillion mixed visual and text tokens atop Kimi-K2-Base ([Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5))
+- **Token efficiency (K2.7-Code):** ~30% fewer thinking tokens than K2.6 ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code))
+- **Serving:** vLLM, SGLang, KTransformers; `transformers >=4.57.1, <5.0.0` ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code))
+
+## Official vendor eval scores
+
+Vendor-reported only. K2.7-Code's model card publishes six benchmarks (all with the K2.6 baseline alongside); the broader public suites (SWE-bench Verified, LiveCodeBench, AIME, HMMT, GPQA, MMLU-Pro) are reported by Moonshot for K2.6 and K2.5 but not re-run for K2.7-Code.
+
+| Benchmark                                             | Score                                                            | Nemotron3 Ultra | % better | Source                                                                                                                                   |
+| ----------------------------------------------------- | ---------------------------------------------------------------- | --------------- | -------- | ---------------------------------------------------------------------------------------------------------------------------------------- |
+| Kimi Code Bench v2 (in-house)                         | K2.7-Code **62.0**; K2.6 50.9                                    | n.a.            | —        | [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)                                                            |
+| Program Bench                                         | K2.7-Code **53.6**; K2.6 48.3                                    | n.a.            | —        | [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)                                                            |
+| MLS Bench Lite                                        | K2.7-Code **35.1**; K2.6 26.7                                    | n.a.            | —        | [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)                                                            |
+| Kimi Claw 24/7 Bench (in-house, long-horizon agentic) | K2.7-Code **46.9**; K2.6 42.9                                    | n.a.            | —        | [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)                                                            |
+| MCP Atlas                                             | K2.7-Code **76.0**; K2.6 69.4                                    | n.a.            | —        | [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)                                                            |
+| MCP Mark Verified                                     | K2.7-Code **81.1**; K2.6 72.8                                    | n.a.            | —        | [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)                                                            |
+| SWE-Bench Verified                                    | K2.7-Code n.a.; K2.6 **80.2**; K2.5 76.8                         | 70.7            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| SWE-Bench Pro                                         | K2.7-Code n.a.; K2.6 **58.6**; K2.5 50.7                         | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| SWE-Bench Multilingual                                | K2.7-Code n.a.; K2.6 **76.7**; K2.5 73.0                         | 67.7            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| Terminal-Bench 2.0 (Terminus-2)                       | K2.7-Code n.a.; K2.6 **66.7**; K2.5 50.8                         | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| LiveCodeBench (v6)                                    | K2.7-Code n.a.; K2.6 **89.6**; K2.5 85.0                         | 89.0 (v6)       | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| BrowseComp                                            | K2.7-Code n.a.; K2.6 **83.2** (86.3 with Agent Swarm); K2.5 74.9 | 44.4            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| HLE-Full (with tools)                                 | K2.7-Code n.a.; K2.6 **54.0**; K2.5 50.2                         | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| HLE-Full (no tools)                                   | K2.7-Code n.a.; K2.6 **34.7**; K2.5 30.1                         | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| AIME 2026                                             | K2.7-Code n.a.; K2.6 **96.4**; (K2.5 on AIME 2025: 96.1)         | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6); [Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5) |
+| HMMT 2026 (Feb)                                       | K2.7-Code n.a.; K2.6 **92.7**; (K2.5 on HMMT 2025 Feb: 95.4)     | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6); [Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5) |
+| GPQA-Diamond                                          | K2.7-Code n.a.; K2.6 **90.5**; K2.5 87.6                         | 87.0 (no tools) | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| MMLU-Pro                                              | K2.7-Code n.a.; K2.6 n.a.; K2.5 **87.1**                         | 86.8            | —        | [Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)                                                                      |
+| MMMU-Pro                                              | K2.7-Code n.a.; K2.6 **79.4**; K2.5 78.5                         | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+| tau2-bench                                            | n.a. — not reported by Moonshot for K2.5, K2.6 or K2.7-Code      | n.a.            | —        | [Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)                                                                      |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+Vendor-reported deltas for K2.7-Code versus K2.6, stated in Moonshot's launch post: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite, and 30% lower reasoning-token usage ([@Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2065377579130142937)).
+
+Evaluation conditions: K2.7-Code and K2.6 were tested with thinking mode enabled via Kimi Code CLI at temperature 1.0, top-p 0.95 and 262,144-token context; MCP-Atlas and MCPMark-Verified used a 100 tool-call/step budget with 32k max tokens per step, averaged over 3 runs ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)). K2.6's SWE-Bench family results used an in-house harness adapted from SWE-agent, averaged over 10 runs ([Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)).
+
+## Benchmark explanations
+
+- **Kimi Code Bench v2** — Moonshot's in-house coding-agent benchmark of realistic software-engineering tasks across 10+ mainstream languages and a full production stack (backend, infra, performance, systems, security, frontend, ML/data) ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **Program Bench** — 200 tasks asking an agent to recreate a program's behavior from only a compiled binary plus documentation (no source, decompilation or internet), judged against 248,000+ fuzz-generated behavioral tests ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **MLS Bench Lite** — official 30-task subset of MLS-Bench testing whether an agent can invent generalizable, scalable ML methods (LLM pretraining/post-training, robotics, world models, CV, RL, ML systems), with 5 hours per task ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **Kimi Claw 24/7 Bench** — Moonshot's in-house long-horizon agentic benchmark for persistent, multi-day coworking tasks: 17 professional scenarios and 610 evaluation points, executed through the OpenClaw harness ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **MCP Atlas** — realistic tool-use tasks served through scalable MCP servers, measuring multi-step tool orchestration ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **MCP Mark Verified** — human-verified MCPMark: MCP tool use across five real server environments (Notion, GitHub, Filesystem, Postgres, Playwright) ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **SWE-Bench Verified** — human-validated real GitHub issues where the model must produce a patch that passes hidden tests; measures repository-level bug fixing.
+- **SWE-Bench Pro** — harder, contamination-resistant SWE-bench variant with larger, more complex real-world tasks.
+- **SWE-Bench Multilingual** — SWE-bench-style issue resolution across repositories in multiple programming languages.
+- **Terminal-Bench 2.0** — terminal/command-line agent tasks run under the default Terminus-2 agent framework; measures end-to-end shell task completion ([Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)).
+- **LiveCodeBench (v6)** — competitive-programming problems drawn from recent contests to limit training contamination; measures algorithmic code generation.
+- **BrowseComp** — hard web-search/browsing agent benchmark for locating obscure facts online over many tool calls; K2.6 also reports an Agent Swarm variant ([Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)).
+- **HLE-Full** — Humanity's Last Exam full set (text and image): extremely hard expert-written questions, reported with and without search/code tools ([Kimi-K2.6 model card](https://huggingface.co/moonshotai/Kimi-K2.6)).
+- **AIME 2026 / AIME 2025** — American Invitational Mathematics Examination: olympiad-qualifier math problems, averaged over many runs ([Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)).
+- **HMMT 2026 (Feb) / HMMT 2025 (Feb)** — Harvard-MIT Mathematics Tournament February contest: harder competition math than AIME.
+- **GPQA-Diamond** — hardest subset of graduate-level, Google-proof science multiple-choice questions in physics, chemistry and biology.
+- **MMLU-Pro** — harder, reasoning-heavier successor to MMLU for broad multiple-choice knowledge.
+- **MMMU-Pro** — massive multi-discipline multimodal understanding, "Pro" version: college-level reasoning over images.
+- **tau2-bench** — customer-service-style tool-use agent benchmark with domain policies and simulated users (not reported by Moonshot for this family).
diff --git a/packages/app/content/models/kimi-k3.mdx b/packages/app/content/models/kimi-k3.mdx
new file mode 100644
index 000000000..14f29fd39
--- /dev/null
+++ b/packages/app/content/models/kimi-k3.mdx
@@ -0,0 +1,106 @@
+---
+title: 'Kimi K3'
+developer: 'Moonshot AI'
+releaseDate: 'July 16, 2026 (open weights by July 27, 2026)'
+description: "Moonshot AI's 2.8T-parameter / 104B-active MoE flagship built on Kimi Delta Attention and Gated MLA, with native vision and a 1M-token context window."
+---
+
+## Overview
+
+Kimi K3 is Moonshot AI's flagship open-weights model, described on its Hugging Face model card as "an open-weight, native multimodal agentic model and our most capable model to date" — a 2.8T-parameter Mixture-of-Experts model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision and a 1-million-token context window, billed as "the world's first open 3T-class model" ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)). Moonshot's own tech blog frames the release the same way: 2.8 trillion parameters, a 1M-token context window, native vision, Stable LatentMoE routing that effectively activates 16 of 896 experts, and roughly a 2.5× improvement in overall scaling efficiency versus Kimi K2 ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3)).
+
+Moonshot's official blog index dates the Kimi K3 post to 2026-07-16 ([Kimi Research & Tech Blog](https://www.kimi.ai/blog/)), and the launch post on Moonshot's official X account, published July 16, 2026, announced K3 as live on Kimi.com, Kimi Work, Kimi Code and the Kimi API, with "Open Weights by July 27, 2026" ([@Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2077830247446446450)). The tech blog confirms the same commitment: "the full Kimi K3 model weights will be released by July 27, 2026," with further architecture, training and evaluation detail promised in a forthcoming Kimi K3 technical report ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3)). The Hugging Face repository (`moonshotai/Kimi-K3`) carries model-card metadata `license: other` / `license_name: kimi-k3`, and the card states that "both the code repository and the model weights are released under the Kimi K3 License" — i.e. a custom Moonshot license, not MIT or Apache ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+
+Positioning is explicitly agentic and coding-first. The card lists long-horizon coding ("sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools — from GPU kernel optimization and compiler development to vision-in-the-loop game dev, CAD, and even chip design") and agentic knowledge work (deep research, interactive visualizations, dashboards, motion design and video editing) as headline capabilities, and states that K3 "always has thinking enabled" with a `reasoning_effort` field supporting `low`, `high` and `max` (default `max`), plus a preserved-thinking-history requirement for multi-turn and tool-call usage ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)). Notable techniques disclosed by Moonshot include KDA and AttnRes, Stable LatentMoE with Quantile Balancing routing, Per-Head Muon optimization, the Sigmoid Tanh Unit (SiTU) activation, and Gated MLA, alongside native MXFP4 quantization-aware training applied from the SFT stage onward ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3); [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)). Serving is recommended on vLLM, SGLang and TokenSpeed, with the hosted API exposed as model id `kimi-k3` at platform.kimi.ai ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); vendor API pricing on the tech blog is $3.00/MTok cache-miss input, $0.30/MTok cache-hit input and $15.00/MTok output ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3)).
+
+## Architecture
+
+- **Architecture family:** Mixture-of-Experts (MoE), autoregressive, native multimodal (text + image) ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md))
+- **Total parameters:** 2.8T ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md))
+- **Activated parameters per token:** 104B ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md))
+- **Layers:** 93 total, of which 1 dense layer ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); `num_hidden_layers: 93`, `first_k_dense_replace: 1` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Attention-layer composition:** 69 KDA layers + 24 Gated MLA layers ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)). `config.json` lists `full_attn_layers` at layers 4, 8, 12, … 88, 92, 93 — i.e. full (Gated MLA) attention roughly every 4th layer, with the remaining 69 indices in `kda_layers` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Attention mechanism:** KDA (Kimi Delta Attention, hybrid linear attention) & Gated MLA; MLA configured with `mla_use_nope: true`, `mla_use_output_gate: true`, `q_lora_rank: 1536`, `kv_lora_rank: 512`, `qk_nope_head_dim: 128`, `qk_rope_head_dim: 64`, `v_head_dim: 128` ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md); [config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **KDA configuration:** 96 heads, `head_dim: 128`, `short_conv_kernel_size: 4`, `gate_lower_bound: -5.0`, `use_full_rank_gate: true` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Attention Residuals (AttnRes):** `attn_res_block_size: 12` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json)); described by Moonshot as selectively retrieving representations across model depth rather than accumulating them uniformly, and as delivering ~25% higher training efficiency at &lt;2% additional cost ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3); [@Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2077830247446446450))
+- **Attention hidden dimension / heads:** 7168 hidden, 96 attention heads ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); `hidden_size: 7168`, `num_attention_heads: 96`, `num_key_value_heads: 96` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Experts:** 896 routed experts, 16 selected per token, 2 shared experts ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); `num_experts: 896`, `num_experts_per_token: 16`, `num_shared_experts: 2`, `moe_layer_freq: 1` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **LatentMoE dimensions:** Latent MoE dimension 3584, MoE hidden dimension per expert 3072 ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); `routed_expert_hidden_size: 3584`, `moe_intermediate_size: 3072`, `latent_moe_use_norm: true` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Routing:** Stable LatentMoE framework; sigmoid router with `noaux_tc` top-k method and renormalization (`moe_router_activation_func: sigmoid`, `topk_method: noaux_tc`, `moe_renormalize: true`) ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json)); Moonshot adds Quantile Balancing, which derives expert allocation from router-score quantiles and removes heuristic updates and the balancing hyperparameter ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3))
+- **Activation function:** SiTU-GLU per the model card ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); `hidden_act: "situ"` with `activation_situ_beta: 4.0` and `activation_situ_linear_beta: 25.0` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json)). The tech blog names the component Sigmoid Tanh Unit (SiTU) ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3))
+- **Dense FFN intermediate size:** 33792 ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Context window:** 1,048,576 tokens ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); `max_position_embeddings: 1048576` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Vocabulary:** 160K per the card; `vocab_size: 163840` in config ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md); [config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Vision encoder:** MoonViT-V2, 401M parameters ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); vision tower config: 27 layers, 12 heads, `vt_hidden_size: 1024`, `vt_intermediate_size: 4096`, `patch_size: 14`, `mm_projector_type: patchmergerv2` ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Precision / quantization:** MXFP4 weights with MXFP8 activations, via quantization-aware training from the SFT stage onward ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)); config declares `format: mxfp4-pack-quantized`, `num_bits: 4`, `group_size: 32`, `quant_method: compressed-tensors`, with attention, shared experts, dense MLP, `lm_head`, vision tower and mm projector excluded from quantization ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Model class / implementation:** `KimiK3ForConditionalGeneration` over a `KimiLinearForCausalLM` text backbone (`model_type: kimi_linear`) ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Multi-token prediction:** none configured (`num_nextn_predict_layers: 0`) ([config.json](https://huggingface.co/moonshotai/Kimi-K3/raw/main/config.json))
+- **Optimizer technique:** Per-Head Muon, extending Muon by optimizing attention heads independently ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3))
+- **Decoding efficiency claim:** KDA enables up to 6.3× faster decoding in million-token contexts ([@Kimi_Moonshot](https://x.com/Kimi_Moonshot/status/2077830247446446450))
+- **Recommended deployment scale:** supernode configurations with 64 or more accelerators; vLLM, SGLang and TokenSpeed recipes ([Kimi K3 Tech Blog](https://www.kimi.com/blog/kimi-k3); [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md))
+- **Modality:** Text + Image input per the card's model summary; the introduction also states K3 "understands text, images, and video within the same model" ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md))
+
+## Official vendor eval scores
+
+All values are the Kimi K3 (max reasoning effort, temperature 1.0) column of Moonshot's own model-card evaluation table ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+
+| Benchmark                                                             | Score                                                            | Nemotron3 Ultra | % better | Source                                                                              |
+| --------------------------------------------------------------------- | ---------------------------------------------------------------- | --------------- | -------- | ----------------------------------------------------------------------------------- |
+| GPQA Diamond                                                          | 93.5                                                             | 87.0 (no tools) | +7.5%    | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| HLE-Full (no tools / with tools)                                      | 43.5 / 56.0                                                      | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| AA-LCR (long-context reasoning)                                       | 74.7                                                             | 65.4            | +14.2%   | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| CritPt                                                                | 23.4                                                             | 3.1 (no tools)  | +654.8%  | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| Terminal-Bench 2.1 (Kimi Code harness)                                | 88.3                                                             | 56.4            | +56.6%   | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| DeepSWE (Kimi Code harness; 67.3 with mini-SWE-agent)                 | 67.5                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| ProgramBench                                                          | 77.8                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| FrontierSWE                                                           | 81.2                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| SWE-Marathon                                                          | 42.0                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| Kimi Code Bench 2.0 (in-house)                                        | 72.9 (73.7 with Claude Code harness)                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| MLS-Bench-Lite                                                        | 48.3                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| SciCode                                                               | 58.7                                                             | 44.6 (subtask)  | +31.6%   | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| BrowseComp (300K-token context compaction; 90.4 with full 1M context) | 91.2                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| DeepSearchQA (F1)                                                     | 95.0                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| MCPMark-Verified                                                      | 94.5                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| MCP-Atlas (500-task public subset)                                    | 84.2                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| OSWorld-Verified                                                      | 84.8                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| τ³-Banking                                                            | 33.4                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| GDPval-AA v2 (Elo)                                                    | 1686                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| MMMU-Pro (no tools / with Python)                                     | 81.6 / 83.4                                                      | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| Video-MME (w. subtitles)                                              | 90.0                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| OmniDocBench                                                          | 91.1                                                             | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| SWE-bench Verified                                                    | n.a. — not reported in Moonshot's K3 model-card table            | 70.7            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| LiveCodeBench                                                         | n.a. — not reported for K3                                       | 89.0 (v6)       | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| AIME / HMMT / MMLU-Pro                                                | n.a. — not reported for K3                                       | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+| tau2-bench                                                            | n.a. — not reported for K3 (Moonshot reports τ³-Banking instead) | n.a.            | —        | [Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md) |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+
+## Benchmark explanations
+
+- **GPQA Diamond** — hardest subset of graduate-level, Google-proof multiple-choice science questions in physics, chemistry and biology; measures expert-level scientific reasoning.
+- **HLE-Full** — Humanity's Last Exam, full set (text and image questions): very hard expert-written questions across many disciplines; Moonshot reports scores without and with tool augmentation ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **AA-LCR** — Artificial Analysis Long Context Reasoning: reasoning over very long inputs; K3's figure is cited from Artificial Analysis as of July 23, 2026 ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **CritPt** — hard research-level physics reasoning benchmark; K3's figure is cited from Artificial Analysis as of July 23, 2026 ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **Terminal-Bench 2.1** — end-to-end terminal/command-line agent tasks: can the model drive a shell to complete real system and engineering jobs.
+- **DeepSWE** — agentic software-engineering task suite (v1.1 tasks) measuring repository-level issue resolution with an agent harness ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **ProgramBench** — asks a coding agent to recreate a program's behavior from only a compiled binary plus documentation, judged against fuzz-generated behavioral tests ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **FrontierSWE** — frontier-difficulty software-engineering tasks scored by a dominance metric recomputed with the official evaluation script ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **SWE-Marathon** — very long-horizon software-engineering tasks, including GPU tasks with performance gates and anti-cheat validators ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **Kimi Code Bench 2.0** — Moonshot's in-house coding-agent benchmark spanning realistic engineering tasks across 10+ languages and a full production stack, including cybersecurity/safety tasks ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md); [Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **MLS-Bench-Lite** — 30-task official subset of MLS-Bench, testing whether an agent can invent generalizable, scalable ML methods within a 5-hour budget ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **SciCode** — scientific code-generation tasks derived from research problems; K3's figure is cited from Artificial Analysis ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **BrowseComp** — hard web-browsing/search agent benchmark: finding hard-to-locate facts online across many tool calls.
+- **DeepSearchQA (F1)** — deep research question answering with search tools, scored by F1 against reference answers.
+- **MCPMark-Verified** — human-verified edition of MCPMark, testing MCP tool use across Notion, GitHub, Filesystem, Postgres and Playwright servers ([Kimi-K2.7-Code model card](https://huggingface.co/moonshotai/Kimi-K2.7-Code)).
+- **MCP-Atlas** — realistic tool-use tasks over scalable MCP servers, run here on the 500-task public subset with a 100-turn limit ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **OSWorld-Verified** — computer-use agent benchmark: completing real desktop/OS tasks through GUI interaction.
+- **τ³-Banking** — tool-agent benchmark in a banking domain testing multi-turn policy-following and tool calls; K3's figure is cited from Artificial Analysis ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **GDPval-AA v2 (Elo)** — Artificial Analysis Elo rating for economically valuable, real-world professional work tasks ([Kimi-K3 model card](https://huggingface.co/moonshotai/Kimi-K3/blob/main/README.md)).
+- **MMMU-Pro** — massive multi-discipline multimodal understanding, "Pro" (harder) version: college-level reasoning over images.
+- **Video-MME** — video understanding across short, medium and long videos; reported here with subtitles.
+- **OmniDocBench** — document parsing/OCR accuracy, scored as (1 − normalized Levenshtein distance) × 100 ([Kimi-K2.5 model card](https://huggingface.co/moonshotai/Kimi-K2.5)).
+- **SWE-bench Verified** — human-validated subset of real GitHub issues requiring a patch that passes hidden tests (not reported for K3).
+- **LiveCodeBench** — contamination-resistant competitive-programming benchmark using recently published contest problems (not reported for K3).
+- **AIME / HMMT** — high-school olympiad-level competition math exams (not reported for K3).
+- **MMLU-Pro** — harder, reasoning-focused version of MMLU multiple-choice knowledge testing (not reported for K3).
+- **tau2-bench** — customer-service-style tool-use agent benchmark with domain policies (not reported for K3).
diff --git a/packages/app/content/models/llama-3-3-70b.mdx b/packages/app/content/models/llama-3-3-70b.mdx
new file mode 100644
index 000000000..69ccb27dc
--- /dev/null
+++ b/packages/app/content/models/llama-3-3-70b.mdx
@@ -0,0 +1,63 @@
+---
+title: 'Llama 3.3 70B Instruct'
+developer: 'Meta'
+releaseDate: 'December 6, 2024'
+description: "Meta's 70B dense instruction-tuned transformer with GQA and a 128K context window, released under the Llama 3.3 Community License."
+---
+
+## Overview
+
+Llama 3.3 70B Instruct is Meta's instruction-tuned, text-only multilingual large language model, released on December 6, 2024 ([Hugging Face: meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct); [Meta Llama 3.3 model card on GitHub](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)). It is an auto-regressive model built on an optimized transformer architecture with 70B parameters (the Hugging Face repo lists 71B params for the checkpoint), a 128k context window, and pretraining on more than 15 trillion tokens with a knowledge cutoff of December 2023 ([Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct)). It takes multilingual text as input and produces multilingual text and code, with official support for English, German, French, Italian, Portuguese, Hindi, Spanish, and Thai ([Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct)).
+
+Positioning: Meta describes Llama 3.3 70B as offering performance similar to Llama 3.1 405B at a fraction of the serving cost, released "to close out the year" 2024 ([Meta AI blog, "The future of AI: Built with Llama"](https://ai.meta.com/blog/future-of-ai-built-with-llama/)). Meta's own instruction-tuned comparison table supports this: Llama 3.3 70B Instruct matches Llama 3.1 70B on MMLU (86.0) while beating both 3.1 70B and, on several tasks, the 405B model — for example IFEval 92.1 vs. 88.6 for 405B and MATH 77.0 vs. 73.8 for 405B ([Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct)).
+
+Licensing and techniques: the model is distributed under the custom commercial Llama 3.3 Community License Agreement, which grants royalty-free use, reproduction, modification and distribution but requires "Built with Llama" attribution, prefixing derivative model names with "Llama", a NOTICE file attribution, compliance with the Acceptable Use Policy, and a separate Meta-granted license for licensees exceeding 700 million monthly active users as of the release date ([Hugging Face model card / license text](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct); [license file](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/LICENSE)). Architecturally, all Llama 3 versions use Grouped-Query Attention for improved inference scalability, and the released BF16 checkpoints can additionally be loaded in 8-bit or 4-bit via `bitsandbytes`/`transformers` ([Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct)).
+
+## Architecture
+
+- Model type: auto-regressive language model using an optimized transformer architecture; dense (no mixture-of-experts) ([Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct))
+- Total parameters: 70B per Meta's model information table (Hugging Face lists 71B params for the checkpoint); all parameters active per token since the model is dense ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct))
+- Active parameters: n.a. as a separately reported figure — Meta reports only the 70B dense parameter count ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md))
+- Layers: 80 (`num_hidden_layers: 80`; the Meta/HF cards do not state layer count — value from a mirrored copy of the official config.json, as the `meta-llama` repo is gated) ([config.json mirror](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/raw/main/config.json))
+- Hidden size: 8192 ([config.json mirror](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/raw/main/config.json))
+- Attention: Grouped-Query Attention (GQA), 64 query heads and 8 key-value heads (8 query heads per KV group) ([Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct); [config.json mirror](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/raw/main/config.json))
+- Experts / MoE: none — dense model, no expert routing reported ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md))
+- Context window: 128k tokens (`max_position_embeddings: 131072`) ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [config.json mirror](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/raw/main/config.json))
+- Positional embeddings: RoPE with `rope_theta: 500000.0` ([config.json mirror](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/raw/main/config.json))
+- Vocabulary size: 128,256 ([config.json mirror](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/raw/main/config.json))
+- Precision / quantization: released in BF16 (`torch_dtype: bfloat16`); checkpoints can be run in 8-bit or 4-bit with `bitsandbytes` (`load_in_8bit=True` / `load_in_4bit=True`) ([Hugging Face model card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct); [config.json mirror](https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/raw/main/config.json))
+- Training data: 15T+ pretraining tokens, knowledge cutoff December 2023 ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md))
+
+## Official vendor eval scores
+
+Meta-reported results for Llama 3.3 70B Instruct (English text, instruction-tuned evaluation table).
+
+| Benchmark                                             | Score                                              | Nemotron3 Ultra | % better | Source                                                                                                                                                                               |
+| ----------------------------------------------------- | -------------------------------------------------- | --------------- | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| MMLU (CoT, 0-shot, macro_avg/acc)                     | 86.0                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| MMLU Pro (CoT, 5-shot, macro_avg/acc)                 | 68.9                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| IFEval (steerability)                                 | 92.1                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| GPQA Diamond (CoT, 0-shot, acc)                       | 50.5                                               | 87.0 (no tools) | -42.0%   | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| HumanEval (0-shot, pass@1)                            | 88.4                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| MBPP EvalPlus (base, 0-shot, pass@1)                  | 87.6                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| MATH (CoT, 0-shot, sympy_intersection_score)          | 77.0                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| BFCL v2 (0-shot, overall_ast_summary/macro_avg/valid) | 77.3                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| MGSM (0-shot, exact match, multilingual)              | 91.1                                               | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md); [Hugging Face card](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| SWE-bench Verified                                    | n.a. — not reported in Meta's Llama 3.3 model card | 70.7            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)                                                                                |
+| Codeforces                                            | n.a. — not reported by Meta for this model         | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)                                                                                |
+| τ-bench (Tau-bench)                                   | n.a. — not reported by Meta for this model         | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)                                                                                |
+| AIME                                                  | n.a. — not reported by Meta for this model         | n.a.            | —        | [Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)                                                                                |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+
+## Benchmark explanations
+
+- **MMLU (CoT)** — Massive Multitask Language Understanding: multiple-choice questions across 57 academic and professional subjects, evaluated with chain-of-thought prompting; measures broad knowledge and reasoning ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
+- **MMLU Pro (CoT)** — a harder, more reasoning-intensive refresh of MMLU with more answer options and cleaner items; measures robust general reasoning ([Hugging Face card links the MMLU-Pro dataset](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct)).
+- **IFEval** — instruction-following evaluation with verifiable constraints (format, length, keywords); Meta classifies it under "Steerability" ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
+- **GPQA Diamond (CoT)** — hardest subset of graduate-level, Google-proof science questions; Meta classifies it under "Reasoning" and scores accuracy ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
+- **HumanEval** — Python function-completion problems scored pass@1 by unit tests; measures basic code generation correctness ([Meta model card, "Code" category](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
+- **MBPP EvalPlus (base)** — Mostly Basic Python Problems with the EvalPlus expanded test suite, pass@1; measures code correctness under stricter tests ([Meta model card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
+- **MATH (CoT)** — competition math word problems graded by symbolic equivalence (sympy); measures multi-step mathematical reasoning ([Meta model card, "Math" category](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
+- **BFCL v2** — Berkeley Function-Calling Leaderboard v2, scored on abstract-syntax-tree validity of generated function calls; measures tool-use/function-calling accuracy ([Meta model card, "Tool Use" category](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
+- **MGSM** — Multilingual Grade School Math, exact-match accuracy on GSM8K-style problems translated into multiple languages; measures multilingual arithmetic reasoning ([Meta model card, "Multilingual" category](https://github.com/meta-llama/llama-models/blob/main/models/llama3_3/MODEL_CARD.md)).
diff --git a/packages/app/content/models/minimax-m27.mdx b/packages/app/content/models/minimax-m27.mdx
new file mode 100644
index 000000000..6b8ebd50d
--- /dev/null
+++ b/packages/app/content/models/minimax-m27.mdx
@@ -0,0 +1,85 @@
+---
+title: 'MiniMax M2.5 / M2.7'
+developer: 'MiniMax'
+releaseDate: 'February 12, 2026 (M2.5); March 18, 2026 (M2.7)'
+description: "MiniMax's 229B-parameter / ~10B-active MoE family with GQA, FP8 block quantization, and a 192K context window."
+---
+
+## Overview
+
+MiniMax-M2.5 and MiniMax-M2.7 are the two most recent open-weights checkpoints of MiniMax's M2 series: sparse mixture-of-experts language models built, in the words of the series technical report, on the thesis that "mini activations can unleash maximum real-world intelligence," with a flagship configuration of 229.9B total parameters and only 9.8B activated per token ([The MiniMax-M2 Series, arXiv:2605.26494](https://arxiv.org/abs/2605.26494)). M2.5 was released on 2026-02-12 — the date MiniMax stamps on its launch announcement ([MiniMax M2.5 announcement](https://www.minimax.io/news/minimax-m25)), matching the Hugging Face repository creation timestamp of 2026-02-12 ([HF API metadata for MiniMax-M2.5](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.5)) and the "MiniMax-M2.5 Version Release Date: 2026-02-13" in its model license ([MiniMax-M2.5 LICENSE-MODEL](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/LICENSE-MODEL)). M2.7 followed on 2026-03-18 as a point release of the same series ([MiniMax M2.7 announcement](https://www.minimax.io/news/minimax-m27-en); [MiniMax API release notes](https://platform.minimax.io/docs/release-notes/models)), shipped in two serving variants, MiniMax-M2.7 and M2.7-highspeed ([MiniMax API release notes](https://platform.minimax.io/docs/release-notes/models)).
+
+Positioning is squarely agentic and coding/office-work oriented. MiniMax describes M2.5 as "extensively trained with reinforcement learning in hundreds of thousands of complex real-world environments" and "SOTA in coding, agentic tool use and search, office work, and a range of other economically valuable tasks," headlined by 80.2% SWE-Bench Verified, 51.3% Multi-SWE-Bench and 76.3% BrowseComp ([MiniMax-M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)). Cost and speed are central to the pitch: M2.5-Lightning sustains 100 tokens/second at $0.30/M input and $2.40/M output, standard M2.5 runs at 50 tokens/second for half that, and MiniMax claims "$1 to run the model continuously for an hour" at 100 tok/s ([MiniMax-M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)). M2.7 is pitched as "our first model deeply participating in its own evolution," able to build complex agent harnesses and complete elaborate productivity tasks using Agent Teams, complex Skills and dynamic tool search; MiniMax reports an internal M2.7 version autonomously optimizing a programming scaffold over 100+ rounds for a 30% performance improvement ([MiniMax-M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)).
+
+Architecturally the point release is a continuation, not a redesign: the series report treats the post-training pipeline as components that "co-evolve from M2 through M2.5 to the latest M2.7" on the same MoE base — a 62-layer decoder-only Transformer, 256 fine-grained experts with sigmoid gating, full multi-head attention with GQA, a 192K-token native context window and a Multi-Token Prediction module doubling as a speculative-decoding draft path ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)). Both released checkpoints report the same 229B model size on Hugging Face and identical `config.json` contents, including FP8 (`float8_e4m3fn`) block-wise quantization ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json); [M2.7 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/config.json)). Notable training-side techniques include the Forge agent-native RL framework (with a tree-structured sample-merging strategy MiniMax credits with ~40× training speedup) and the CISPO algorithm for MoE training stability ([MiniMax-M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)), plus windowed-FIFO scheduling and prefix-tree merging in the report's system description ([arXiv:2605.26494](https://arxiv.org/abs/2605.26494)). Licensing differs between the two: M2.5's HF metadata declares `license_name: modified-mit` pointing at the GitHub LICENSE ([HF API metadata for MiniMax-M2.5](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.5)) with a "MINIMAX MODEL LICENSE" file in-repo ([M2.5 LICENSE-MODEL](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/LICENSE-MODEL)), whereas M2.7 ships a "NON-COMMERCIAL LICENSE" that permits non-commercial use on MIT-style terms, requires "Built with MiniMax M2.7" attribution for commercial use, and prohibits any commercial use without prior written authorization from MiniMax ([M2.7 LICENSE](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/LICENSE)).
+
+## Architecture
+
+Shared by both checkpoints unless noted. M2.5 and M2.7 publish byte-identical architecture fields in `config.json`.
+
+- **Total parameters:** 229.9B for the series flagship per the technical report ([arXiv:2605.26494](https://arxiv.org/abs/2605.26494)); Hugging Face safetensors metadata reports 228,703,644,928 for M2.5 and 228,689,764,864 for M2.7, both surfaced as "229B params" on the cards ([HF API metadata M2.5](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.5); [HF API metadata M2.7](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.7); [M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)).
+- **Activated parameters:** 9.8B per token for the series flagship ([arXiv:2605.26494](https://arxiv.org/abs/2605.26494)); the report describes M2.7 as having "∼10B activated parameters" ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **Architecture class:** `MiniMaxM2ForCausalLM`, `model_type: minimax_m2` — i.e. M2.7 reuses the M2 model class ([M2.7 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/config.json); [M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)).
+- **Layers:** 62 (`num_hidden_layers: 62`), a decoder-only Transformer ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json); [MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **Hidden size:** 3,072; expert intermediate size 1,536; no shared expert (`shared_intermediate_size: 0`) ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)).
+- **Attention:** grouped-query attention with 48 query heads and 8 key-value heads, head dim 128, full (non-sliding) attention on every layer (`attn_type_list` is all 1s) ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)); the report confirms "full multi-head attention with GQA," 48 query / 8 KV heads ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **QK normalization:** `use_qk_norm: true`, `qk_norm_type: per_layer` ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json); [M2.7 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/config.json)).
+- **Positional encoding:** RoPE with `rope_theta: 5,000,000` and `rotary_dim: 64` (partial rotary over the 128-dim head) ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)); the report lists Rotary Position Embeddings ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **Experts:** 256 routed experts (`num_local_experts: 256`), 8 activated per token (`num_experts_per_tok: 8`), sigmoid scoring with learnable routing bias (`scoring_func: sigmoid`, `use_routing_bias: true`) ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)); the report describes "256 fine-grained experts with sigmoid gating" and expert-specific bias terms ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **Multi-token prediction:** `use_mtp: true`, `num_mtp_modules: 3`, `mtp_transformer_layers: 1` ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)). The report details a single MTP module (K=1) in pre-training expanded to three modules (K=3) during the continued-pretraining decay phase, MTP loss weight annealed 0.3 → 0.1, with the three modules used as speculative-decoding draft heads verified by the main model in one forward pass ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **Precision / quantization:** FP8 weights — `quant_method: fp8`, `fmt: float8_e4m3fn`, dynamic activation scaling, `weight_block_size: [128, 128]`, with `gate`, `e_score_correction_bias` and `lm_head` excluded from conversion; base dtype bfloat16 ([M2.7 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/raw/main/config.json)). HF lists tensor types F32 · BF16 · F8_E4M3 for both repos ([M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5); [M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)).
+- **Context window:** `max_position_embeddings: 196608` (192K) ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)); the report states a "192K-token native context window" reached via 8K → 32K → 192K context-extension stages ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **Vocabulary:** 200,064 tokens; `tie_word_embeddings: false`; activation `silu`; `rms_norm_eps: 1e-06` ([M2.5 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M2.5/raw/main/config.json)).
+- **Pre-training scale:** 29.2T tokens total (19.9T constant phase, 9.3T decay-phase budget) ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+- **Inference settings:** `temperature=1.0`, `top_p=0.95`, `top_k=40`, with per-model default system prompts; served through SGLang, vLLM, Transformers (plus KTransformers for M2.5, NVIDIA NIM for M2.7) ([M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5); [M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)).
+- **Multimodality:** text-in/text-out (`pipeline_tag: text-generation` for both repos) ([HF API metadata M2.7](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M2.7)); M2.7's MM Claw benchmark is described as an internal multi-modal office benchmark ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+
+## Official vendor eval scores
+
+Vendor-reported only. Unless noted, both columns come from Table 4 of MiniMax's own M2-series technical report, evaluated with thinking enabled and the interleaved-thinking protocol; AIME 2026, GPQA-Diamond, SciCode, IFBench, AA-LCR, HLE and MMLU-Pro follow the Artificial Analysis Index v4.0 protocol, no tools, pass@1 ([MiniMax-M2 series report HTML](https://arxiv.org/html/2605.26494v2)).
+
+| Benchmark                   | Score                              | Nemotron3 Ultra | % better | Source                                                                                                                                                                                             |
+| --------------------------- | ---------------------------------- | --------------- | -------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| SWE-Bench Verified          | M2.5: 80.2% · M2.7: n.a.           | 70.7            | —        | [MiniMax-M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) / [MiniMax M2.5 announcement](https://www.minimax.io/news/minimax-m25) (not reported for M2.7 in fetched MiniMax sources) |
+| SWE-bench Pro               | M2.5: 55.4 · M2.7: 56.2            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.7 56.22% also in [MiniMax M2.7 announcement](https://www.minimax.io/news/minimax-m27-en)                              |
+| SWE-bench Multilingual      | M2.5: 74.1 · M2.7: 76.5            | 67.7            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.7 76.5 also on the [M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)                                   |
+| Multi-SWE-bench             | M2.5: 51.3 · M2.7: 52.7            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.5 51.3% also in [MiniMax M2.5 announcement](https://www.minimax.io/news/minimax-m25)                                  |
+| Terminal-Bench 2.0          | M2.5: 51.7 · M2.7: 57.0            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.7 57.0% also in [MiniMax M2.7 announcement](https://www.minimax.io/news/minimax-m27-en)                               |
+| NL2Repo                     | M2.5: 26.6 · M2.7: 39.8            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2)                                                                                                                           |
+| MLE Bench Lite (medal rate) | M2.5: 51.5 · M2.7: 66.6            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.7 66.6% also on the [M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)                                  |
+| VIBE-Pro                    | M2.5: 54.2 · M2.7: 55.6            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2)                                                                                                                           |
+| BrowseComp                  | M2.5: 76.3 · M2.7: 77.8            | 44.4            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.5 76.3% (with context management) in [MiniMax M2.5 announcement](https://www.minimax.io/news/minimax-m25)             |
+| Wide Search                 | M2.5: 70.3 · M2.7: 75.2            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2)                                                                                                                           |
+| Toolathlon                  | M2.5: 38.3 · M2.7: 46.3            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.7 46.3% also in [MiniMax M2.7 announcement](https://www.minimax.io/news/minimax-m27-en)                               |
+| GDPval-AA                   | M2.5: 35.0 · M2.7: 50.0 (ELO 1495) | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); ELO 1495 from the [M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)                                       |
+| AIME 2026                   | M2.5: 87.2 · M2.7: 94.2            | n.a.            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2)                                                                                                                           |
+| GPQA-Diamond                | M2.5: 85.2 · M2.7: 89.8            | 87.0 (no tools) | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); M2.5 85.2 also in the [M2.5 model card appendix](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)                          |
+| MMLU-Pro                    | M2.5: 85.2 · M2.7: 81.8            | 86.8            | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2) (as printed; M2.7 is lower than M2.5 in the vendor table)                                                                 |
+| HLE (no tools)              | M2.5: 19.0 · M2.7: 28.0            | 26.7 (no tools) | —        | [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2); the M2.5 card appendix lists 19.4 for "HLE w/o tools" ([M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)) |
+| tau2-bench                  | n.a.                               | n.a.            | —        | Named among agent benchmarks in the report's attention-ablation section but no M2.5/M2.7 score published ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2))                         |
+| LiveCodeBench               | n.a.                               | 89.0 (v6)       | —        | Not reported for M2.5 or M2.7 in any fetched MiniMax source ([MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2))                                                             |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+Other vendor-reported figures: M2.5 scored 79.7 on SWE-Bench Verified under Droid scaffolding and 76.1 under OpenCode scaffolding, averaged 3.52M tokens and 22.8 minutes per SWE-Bench Verified task (vs 31.3 minutes for M2.1, a 37% speedup), and achieved a 59.0% average GDPval-MM win rate against mainstream models ([MiniMax M2.5 announcement](https://www.minimax.io/news/minimax-m25)). M2.5's card appendix additionally reports AIME25 86.3, SciCode 44.4, IFBench 70.0 and AA-LCR 69.5 ([M2.5 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)). For M2.7, MiniMax reports MM Claw 62.7%, 97% skill compliance across 40+ complex skills, MEWC v2 63.3, Finance Modeling Pro 57.0, RISE 64.3 and HyperTask 67.6 ([M2.7 model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7); [MiniMax-M2 series report, Table 4](https://arxiv.org/html/2605.26494v2)).
+
+## Benchmark explanations
+
+- **SWE-Bench Verified** — human-validated set of real GitHub issues; scores the share of issues whose generated patch passes the repository's tests ([MiniMax evaluated it with Claude Code scaffolding averaged over 4 runs](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)).
+- **SWE-bench Pro** — industry-grade repository repair, a harder successor to SWE-bench aimed at realistic enterprise codebases ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **SWE-bench Multilingual** — cross-language repository-level editing, testing coding competence beyond Python ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **Multi-SWE-bench** — multi-repo task transfer: issue resolution spanning multiple repositories/languages ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **Terminal-Bench 2.0** — terminal and system-operation tasks in a sandbox (8 vCPU / 16 GB, 2-hour timeout, Terminus-2 scaffold in MiniMax's setup), measuring shell-agent reliability ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **NL2Repo** — MiniMax's internal benchmark for synthesizing an entire repository from a natural-language specification ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **MLE Bench Lite** — autonomous machine-learning engineering across 22 Kaggle-style competitions; scored as test-set medal rate after 24-hour runs on a single A30 ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **VIBE-Pro** — MiniMax's internal end-to-end full-stack application development benchmark, verifying interaction logic and visual fidelity in containerized deployments ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **BrowseComp** — hard open-web search benchmark measuring an agent's ability to locate obscure facts and synthesize an answer ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **Wide Search** — broad-coverage web research benchmark requiring many parallel lookups aggregated into one answer, run in the WebExplorer agent framework ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **Toolathlon** — heterogeneous tool-use benchmark exercising many different external tools in one agent workflow ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **GDPval-AA** — the Artificial-Analysis-judged subset of OpenAI's GDPval suite of economically valuable office tasks, scored on deliverable quality ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **AIME 2026** — American Invitational Mathematics Examination problems; competition mathematics accuracy ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **GPQA-Diamond** — hardest split of graduate-level physics/chemistry/biology questions written to be Google-proof ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **MMLU-Pro** — broad multi-domain knowledge benchmark with harder, more distractor-rich multiple-choice items than MMLU ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **HLE (Humanity's Last Exam, no-tool subset)** — frontier-difficulty expert questions across disciplines, answered without tools ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **tau2-bench** — simulated-user tool-use benchmark with domain policies (retail/telecom splits) measuring policy-compliant multi-turn agent behavior; not scored for M2.5/M2.7 (n.a.) ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **LiveCodeBench** — rolling competitive-programming benchmark built from recent contest problems to resist training contamination; not reported by MiniMax for this family (n.a.).
+- **RISE / MM Claw / MEWC v2 / Finance Modeling Pro / HyperTask** — MiniMax-internal benchmarks for realistic multi-step browsing with a Playwright browser, multimodal office "claw" tasks, hard Microsoft Excel World Championship problems, expert-rubric-scored Excel financial modeling, and long-horizon multi-requirement "vibe coding" respectively ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
diff --git a/packages/app/content/models/minimax-m3.mdx b/packages/app/content/models/minimax-m3.mdx
new file mode 100644
index 000000000..9edceb30d
--- /dev/null
+++ b/packages/app/content/models/minimax-m3.mdx
@@ -0,0 +1,91 @@
+---
+title: 'MiniMax M3'
+developer: 'MiniMax'
+releaseDate: 'June 1, 2026'
+description: "MiniMax's ~428B-parameter / ~23B-active multimodal MoE with GQA attention, a 1M-token context window, and 7 multi-token-prediction modules."
+---
+
+## Overview
+
+MiniMax-M3 is the flagship open-weights model of the M-series from Shanghai-based AI company MiniMax, described on its Hugging Face model card as "a native multimodal model with 1M context" with "~428B parameters and ~23B activated parameters" ([MiniMaxAI/MiniMax-M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)). MiniMax's API release notes date the launch to "Jun. 1, 2026," describing M3 as "the latest M-series language model for agentic reasoning, tool use, coding, multimodal chat input, and long-context tasks" ([MiniMax API release notes](https://platform.minimax.io/docs/release-notes/models)); the Hugging Face repository itself was created on 2026-06-02 ([HF API metadata for MiniMaxAI/MiniMax-M3](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3)). The model is a mixture-of-experts vision-language model: it accepts text, image and video input and emits text, and the repository is published under `pipeline_tag: image-text-to-text` ([M3 README source](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/README.md)).
+
+Positioning is explicitly agentic and coding-first. MiniMax's launch blog markets M3 as delivering "coding & agentic frontier" capability with autonomous task decomposition, tool invocation, multi-step reasoning, autonomous browsing and long-range coding, and calls it the first open-weight model combining frontier coding capability, million-token context and native multimodality ([MiniMax M3 model page](https://www.minimax.io/models/text/m3)). The model card frames three headline capabilities: native multimodality ("mixed-modality training from the very first step"), context scaling through sparse attention, and "frontier-level performance across long-horizon agentic benchmarks, excelling in both coding and cowork" ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)). NVIDIA's hosted model card adds concrete target workloads: long-form video understanding up to 30 minutes, long-horizon coding tasks of 8+ hours, and agentic/tool-use workflows, with two switchable reasoning modes ([NVIDIA build.nvidia.com M3 model card](https://build.nvidia.com/minimaxai/minimax-m3/modelcard)); the card itself exposes three modes via the `thinking` parameter — `enabled`, `adaptive`, `disabled` ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)).
+
+The defining technical contribution is **MiniMax Sparse Attention (MSA)**, "a high-performance sparse attention operator designed for million-token contexts" that, compared with GQA, "dramatically reduces the attention compute and memory footprint while preserving model quality," yielding "9× prefill and 15× decode speedups compared to M2 at 1M context, reducing per-token compute to 1/20" ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)). The accompanying technical report describes MSA as blockwise sparse attention built on top of GQA, with a lightweight Index Branch that scores KV blocks and independently selects a Top-k subset per GQA group, a Main Branch performing exact block-sparse attention over selected blocks, exp-free Top-k selection, and a KV-outer sparse attention kernel; on a 109B-parameter native-multimodal testbed it reports 28.4× lower per-token attention compute at 1M context and 14.2× prefill / 7.6× decoding wall-clock speedups on H800 while performing "on par with GQA" ([MiniMax Sparse Attention, arXiv:2606.13392](https://arxiv.org/abs/2606.13392)). Licensing is **not** MIT: the repository ships the "MINIMAX COMMUNITY LICENSE," which grants rights for non-commercial purposes, requires prominent "Built with MiniMax M3" attribution for commercial use, requires a one-time notice to api@minimax.io below US$20M annual revenue, and requires separate prior written authorization above US$20M ([M3 LICENSE](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/LICENSE)); HF card metadata tags it `license: other`, `license_name: minimax-community` ([M3 README source](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/README.md)).
+
+## Architecture
+
+- **Total parameters:** ~428B per the model card ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)); Hugging Face safetensors metadata reports 427,040,140,160 parameters exactly ([HF API metadata](https://huggingface.co/api/models/MiniMaxAI/MiniMax-M3)).
+- **Activated parameters:** ~23B per token ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)).
+- **Architecture class / model type:** `MiniMaxM3SparseForConditionalGeneration`, `model_type: minimax_m3_vl`; the text tower is `MiniMaxM3SparseForCausalLM` ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Layers:** 60 (`num_hidden_layers: 60`) ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Hidden size:** 6,144; MoE expert intermediate size 3,072; dense-layer intermediate size 12,288; shared-expert intermediate size 3,072 ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Attention layout:** 64 query heads, 4 key-value heads (GQA), head dim 128 ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)); MSA is explicitly "blockwise sparse attention built upon GQA" ([arXiv:2606.13392](https://arxiv.org/abs/2606.13392)).
+- **QK normalization:** `use_qk_norm: true` with `qk_norm_type: per_head` ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Partial RoPE:** `rotary_dim: 64` with `partial_rotary_factor: 0.5` and `rope_theta: 5,000,000` — i.e. RoPE applied to half of each 128-dim head ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Sparse attention (MSA) configuration:** `use_sparse_attention: true`, `sparse_index_dim: 128`, `sparse_num_index_heads: 4`, `sparse_topk_blocks: 16`, `sparse_block_size: 128`, `sparse_score_type: max`, `sparse_local_block: 1`; sparse attention is disabled in the first 3 layers and enabled in layers 4–60 via the `sparse_attention_freq` list ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **MSA mechanism:** lightweight Index Branch scores KV blocks and selects a Top-k subset independently per GQA group (group-specific sparse retrieval); Main Branch runs exact block-sparse attention over only those blocks; exp-free Top-k selection; KV-outer sparse attention kernel for tensor-core utilization ([arXiv:2606.13392](https://arxiv.org/abs/2606.13392)). MiniMax states MSA partitions KV into blocks more precisely than DSA/MoBA-style approaches and is "more than 4× faster than Flash-Sparse-Attention and flash-moba" ([MiniMax M3 launch blog](https://www.minimax.io/blog/minimax-m3)).
+- **Experts:** 128 routed experts (`num_local_experts: 128`), top-4 routed per token (`num_experts_per_tok: 4`), plus 1 shared expert (`n_shared_experts: 1`); sigmoid routing scores with routing bias and `routed_scaling_factor: 2.0`; first 3 layers dense, layers 4–60 MoE per `moe_layer_freq` ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Activation / norms:** `hidden_act: swigluoai` with `swiglu_alpha: 1.702`, `swiglu_limit: 7.0`; RMSNorm with `use_gemma_norm: true`, `rms_norm_eps: 1e-06` ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Multi-token prediction:** `num_mtp_modules: 7`, `num_nextn_predict_layers: 1` ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Context window:** `max_position_embeddings: 1,048,576` (1M tokens) ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)); MiniMax markets "up to 1M tokens" with a guaranteed minimum of 512K ([MiniMax M3 model page](https://www.minimax.io/models/text/m3)).
+- **Vocabulary:** 200,064 tokens; image token index 200025, video token index 200026 ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Vision tower:** CLIP-style vision encoder, 32 layers, hidden size 1,280, 16 heads, patch size 14, image size 2016, 3D RoPE position embeddings, projection to 6,144 dims, patch-merge token compression (spatial merge 2, temporal patch 2) ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)).
+- **Precision:** `torch_dtype: bfloat16` ([M3 config.json](https://huggingface.co/MiniMaxAI/MiniMax-M3/raw/main/config.json)); the HF card lists tensor types BF16 · F32 ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)). MiniMax also publishes an MXFP8 checkpoint referenced in ATOM/MXFP4-MXFP8 usage guidance ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)).
+- **Recommended inference settings:** `temperature=1.0`, `top_p=0.95`; served via SGLang, vLLM, Transformers, KTransformers, unsloth, ATOM ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)).
+- **Reasoning modes:** `thinking` parameter with `enabled` / `adaptive` / `disabled` ([M3 model card](https://huggingface.co/MiniMaxAI/MiniMax-M3)).
+
+## Official vendor eval scores
+
+All rows below are MiniMax-reported scores taken from the official benchmark table published in the MiniMax-M3 Hugging Face repository (`figures/benchmark.jpeg`, referenced from the model card) or from MiniMax's own launch blog / model page.
+
+| Benchmark          | Score                  | Nemotron3 Ultra | % better | Source                                                                                                                                                                                                             |
+| ------------------ | ---------------------- | --------------- | -------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| SWE-Bench Verified | 80.5                   | 70.7            | +13.9%   | [MiniMax M3 vendor benchmark table (HF repo)](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                     |
+| SWE-Bench Pro      | 59.0                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg); also 59.0% in the [M3 launch blog](https://www.minimax.io/blog/minimax-m3)                   |
+| Terminal Bench 2.1 | 66.0                   | 56.4            | +17.0%   | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg); also 66.0% in the [M3 launch blog](https://www.minimax.io/blog/minimax-m3)                   |
+| SWE-fficiency      | 34.8                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg); also 34.8% in the [M3 launch blog](https://www.minimax.io/blog/minimax-m3)                   |
+| NL2Repo            | 42.1                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| LiveSQLBench       | 40.2                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| KernelBench Hard   | 28.8                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg); also 28.8% in the [M3 launch blog](https://www.minimax.io/blog/minimax-m3)                   |
+| PostTrainBench     | 37.1 (rank #3 overall) | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg); [MiniMax M3 model page](https://www.minimax.io/models/text/m3)                               |
+| BrowseComp         | 83.5                   | 44.4            | +88.1%   | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg); [MiniMax M3 model page](https://www.minimax.io/models/text/m3)                               |
+| MCP Atlas          | 74.2                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg); also 74.2% in the [M3 launch blog](https://www.minimax.io/blog/minimax-m3)                   |
+| Apex-Agents        | 27.7                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| OSWorld-Verified   | 75.2                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| MMMU-Pro           | 78.1                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| Video-MMMU         | 84.6                   | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| USAMO 2026         | 36 / 42                | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| IMO 2025           | 35 / 42                | n.a.            | —        | [MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)                                                                                               |
+| tau2-bench         | n.a.                   | n.a.            | —        | Not reported for M3 in any MiniMax source fetched ([vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg), [model page](https://www.minimax.io/models/text/m3)) |
+| LiveCodeBench      | n.a.                   | 89.0 (v6)       | —        | Not reported for M3 in any MiniMax source fetched ([vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg))                                                      |
+| AIME               | n.a.                   | n.a.            | —        | M3's reasoning block reports IMO 2025 and USAMO 2026 instead ([vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg))                                           |
+| GPQA Diamond       | n.a.                   | 87.0 (no tools) | —        | Not reported for M3 in any MiniMax source fetched ([vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg))                                                      |
+| MMLU-Pro           | n.a.                   | 86.8            | —        | Not reported for M3 in any MiniMax source fetched ([vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg))                                                      |
+| HLE                | n.a.                   | 26.7 (no tools) | —        | Not reported for M3 in any MiniMax source fetched ([vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg))                                                      |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+Additional vendor-reported figures from the same table (comparison context): the vendor chart also lists SWE Atlas-QnA 37.9, SWE Atlas-Test Writing 30.8, CL-bench 20.5, VIBE-V2 50.1, SVG-Bench 63.7, PaperBench 52.6, DRACO 73.2, GDPval rubrics 74.8, BankerToolBench 76.1, OfficeQA Pro 45.1, SpreadSheetBench-v1 89.4, YC-Bench 2.1M, LOCA-Bench (256k) 49.3, Claw-Eval 74.5, OmniDocBench 91.6 and VideoMME (w/ sub) 85.4 for M3 ([MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)).
+
+Vendor evaluation methodology notes (from the same chart): SWE-Bench Verified was run on internal infrastructure with Claude Code as scaffolding, default system prompt overridden, 4 runs averaged; Terminal-bench 2.1 used an 8C16G sandbox, 2-hour timeout, 128K max output tokens and Terminus 2 scaffolding; SWE-fficiency used the open-source dataset/workflow in a 1C2G sandbox with a 2-hour timeout ([MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)).
+
+## Benchmark explanations
+
+- **SWE-Bench Verified** — human-validated subset of real GitHub issues; measures whether the model's patch resolves the issue and passes the repo's tests ([MiniMax reports it as an agentic coding benchmark run with Claude Code scaffolding](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)).
+- **SWE-Bench Pro** — harder, industry-grade repository-repair benchmark for coding agents ([described as "industry-grade repository repair" in the MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **Terminal Bench 2.1** — terminal/system-operation tasks executed in a sandboxed shell, measuring command-line agent competence ([MiniMax describes Terminal-Bench as terminal and system-operation tasks](https://arxiv.org/html/2605.26494v2)).
+- **SWE-fficiency** — measures a coding agent's ability to make repository changes that improve software efficiency, using the open-source SWE-fficiency dataset and workflow ([evaluation methodology in MiniMax's M3 chart](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)).
+- **NL2Repo** — MiniMax's internal natural-language-to-repository synthesis benchmark: build a whole repo from a specification ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **LiveSQLBench** — text-to-SQL agent benchmark (600 questions across 22 PostgreSQL databases in MiniMax's setup), scoring correct query results ([M3 chart methodology](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)).
+- **KernelBench Hard** — hardest split of GPU-kernel generation tasks: write correct, fast CUDA-style kernels ([reported by MiniMax as a coding/agentic frontier result](https://www.minimax.io/blog/minimax-m3)).
+- **PostTrainBench** — autonomous post-training benchmark: given pretrain-only base models, the agent must synthesize data, train, evaluate and iterate without human help ([MiniMax M3 model page](https://www.minimax.io/models/text/m3)).
+- **BrowseComp** — hard web-browsing benchmark measuring whether an agent can find difficult-to-locate facts online ([used by MiniMax as a search/deep-research benchmark](https://arxiv.org/html/2605.26494v2)).
+- **MCP Atlas** — tool-use benchmark over Model Context Protocol tool servers, measuring multi-tool orchestration ([reported in MiniMax's M3 agentic results](https://www.minimax.io/blog/minimax-m3)).
+- **Apex-Agents** — Mercor's agentic evaluation suite of expert-designed real-work tasks ([listed in MiniMax's M3 cowork block and as a leaderboard on the M3 card](https://huggingface.co/MiniMaxAI/MiniMax-M3)).
+- **OSWorld-Verified** — computer-use benchmark: complete real desktop GUI tasks in a virtual OS environment ([reported as a GUI benchmark with task completion rate in MiniMax's M3 materials](https://www.minimax.io/blog/minimax-m3)).
+- **MMMU-Pro** — multimodal college-level multiple-choice reasoning across images and diagrams, a harder variant of MMMU ([listed as a multimodal benchmark on the M3 card](https://huggingface.co/MiniMaxAI/MiniMax-M3)).
+- **Video-MMMU / VideoMME** — long-form video understanding benchmarks; MiniMax reports Video-MME at 512 frames, measuring comprehension of extended video input ([M3 launch blog](https://www.minimax.io/blog/minimax-m3)).
+- **USAMO 2026 / IMO 2025** — olympiad proof-based mathematics contests, scored out of 42 points, measuring frontier mathematical reasoning ([MiniMax M3 vendor benchmark table](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/benchmark.jpeg)).
+- **tau2-bench** — customer-service style tool-use benchmark with simulated users and domain policies (retail/telecom splits); named among agent benchmarks in MiniMax's M-series report but not scored for M3 ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
+- **LiveCodeBench** — contamination-resistant competitive-programming benchmark using recently published contest problems; not reported by MiniMax for M3 (n.a.).
+- **AIME / GPQA Diamond / MMLU-Pro / HLE** — competition math, graduate-level science QA, broad multi-domain knowledge, and frontier-difficulty open knowledge respectively; MiniMax defines them this way in the M-series report but publishes no M3 numbers for them ([MiniMax-M2 series report](https://arxiv.org/html/2605.26494v2)).
diff --git a/packages/app/content/models/qwen-3-5.mdx b/packages/app/content/models/qwen-3-5.mdx
new file mode 100644
index 000000000..da34f1c27
--- /dev/null
+++ b/packages/app/content/models/qwen-3-5.mdx
@@ -0,0 +1,81 @@
+---
+title: 'Qwen3.5-397B-A17B'
+developer: 'Alibaba (Qwen)'
+releaseDate: 'February 2026'
+description: "Alibaba's 397B-parameter / 17B-active MoE with a 3:1 Gated DeltaNet to Gated Attention hybrid stack, 262K native context, and Apache 2.0 weights."
+---
+
+## Overview
+
+Qwen3.5-397B-A17B is the first open-weights model released in Alibaba's Qwen3.5 series, announced by the Qwen team as a **native vision-language foundation model** with 397B total parameters and 17B activated per forward pass ([Alibaba Cloud "Qwen3.5: Towards Native Multimodal Agents"](https://www.alibabacloud.com/blog/602894)). The official announcement blog is dated February 2026 ([Qwen blog](https://qwen.ai/blog?id=qwen3.5)), with third-party trackers placing the public release in mid-February 2026 — February 16, 2026 per [OpenRouter](https://openrouter.ai/qwen/qwen3.5-397b-a17b) and February 17, 2026 per [Artificial Analysis](https://artificialanalysis.ai/articles/qwen3-5-397b-a17b-everything-you-need-to-know). Weights are published on Hugging Face as a post-trained checkpoint compatible with Transformers, vLLM, SGLang and KTransformers, with a hosted counterpart (**Qwen3.5-Plus**) offering a 1M-token default context, built-in tools and adaptive tool use ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)). The repository ships the **Apache License, Version 2.0** ([LICENSE file on Hugging Face](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/blob/main/LICENSE)), also reported as Apache 2.0 by [Artificial Analysis](https://artificialanalysis.ai/articles/qwen3-5-397b-a17b-everything-you-need-to-know).
+
+Positioning is explicitly agentic and multimodal: the vendor frames the model as targeting "reasoning, coding, agent capabilities, and multimodal understanding" ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)), and the distribution description highlights language understanding, logical reasoning, code generation, agent-based tasks, image and video understanding, and GUI interaction, with "robust code-generation and agent capabilities" and strong generalization across agent tasks ([OpenRouter](https://openrouter.ai/qwen/qwen3.5-397b-a17b)). Qwen reports that early-fusion multimodal training achieves cross-generational parity with Qwen3 while outperforming the Qwen3-VL line on reasoning, coding, agents and visual understanding ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)). Independent evaluation placed it at 45 on the Artificial Analysis Intelligence Index — #3 among open-weights models at publication — with a GDPval-AA agentic ELO of 1,221 versus 860 for Qwen3 235B ([Artificial Analysis](https://artificialanalysis.ai/articles/qwen3-5-397b-a17b-everything-you-need-to-know)).
+
+On technique, the model unifies thinking and non-thinking behaviour in one checkpoint: it "operate[s] in thinking mode by default," emitting `<think>...</think>` content before the answer, and unlike Qwen3 it drops the `/think` and `/nothink` soft switches, so non-thinking responses are obtained via API parameters instead ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)); Artificial Analysis notes this reverses Qwen3's split instruct/thinking variants ([Artificial Analysis](https://artificialanalysis.ai/articles/qwen3-5-397b-a17b-everything-you-need-to-know)). The checkpoint is trained with **multi-token prediction (MTP)** across multiple steps ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)), listed by Qwen as an efficiency feature of the underlying Qwen3-Next architecture ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)), and serving stacks expose it directly as MTP speculative decoding, e.g. `{"method": "qwen3_5_mtp", "num_speculative_tokens": 3}` ([vLLM Ascend docs](https://docs.vllm.ai/projects/ascend/en/latest/tutorials/models/Qwen3.5-397B-A17B.html)). Training used a native FP8 pipeline over activations, MoE routing and GEMMs with BF16 preserved in sensitive layers (~50% activation-memory reduction, >10% speedup), plus a fully disaggregated asynchronous RL framework with FP8 end-to-end training, rollout router replay, speculative decoding and multi-turn rollout locking for a 3×–5× end-to-end speedup ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+
+## Architecture
+
+- **Total / active parameters:** 397B total, 17B activated ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894))
+- **Type:** Causal language model with vision encoder (native vision-language, early text-vision fusion) ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894))
+- **Base architecture:** Qwen3-Next — hybrid linear attention (Gated Delta Networks) + high-sparsity sparse MoE ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894))
+- **Layers:** 60 (`num_hidden_layers: 60`) ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [config.json](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/main/config.json))
+- **Hidden layout:** `15 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE))`; config shows three `linear_attention` layers per one `full_attention` layer, `full_attention_interval: 4` ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [config.json](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/main/config.json))
+- **Hidden dimension:** 4096 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B))
+- **Gated DeltaNet (linear attention):** 64 value heads, 16 QK heads, head dim 128, conv kernel dim 4 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [config.json](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/main/config.json))
+- **Gated Attention (full attention):** 32 Q heads, 2 KV heads, head dim 256, output gating enabled (`attn_output_gate: true`), no attention bias ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [config.json](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/main/config.json))
+- **Rotary position embedding dimension:** 64 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B))
+- **MoE experts:** 512 experts, 10 routed + 1 shared activated per token, expert intermediate dimension 1024 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [config.json](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/main/config.json))
+- **Vocabulary / embeddings:** token embedding and LM output 248,320 (padded); Qwen describes the vocabulary as expanded to 250k from 150k, giving 10–60% encoding efficiency gains ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894))
+- **Context window:** 262,144 tokens native (`max_position_embeddings: 262144`), extensible to 1,010,000 tokens; Qwen3.5-Plus defaults to 1M ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [config.json](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/main/config.json))
+- **Precision:** released weights `dtype: bfloat16` (tensor types BF16/F32); an FP8-quantized repo uses fine-grained FP8 with block size 128; training used a native FP8 pipeline with BF16 retained in sensitive layers ([config.json](https://huggingface.co/Qwen/Qwen3.5-397B-A17B/raw/main/config.json); [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [Qwen3.5-397B-A17B-FP8 card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8); [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894))
+- **Novel components:** MTP trained with multi-steps, exposed for speculative decoding in serving (`qwen3_5_mtp`) ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B); [vLLM Ascend docs](https://docs.vllm.ai/projects/ascend/en/latest/tutorials/models/Qwen3.5-397B-A17B.html))
+- **Multilingual coverage:** 201 languages and dialects (up from 119) ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894))
+- **Reported efficiency:** decoding throughput 8.6× / 19.0× that of Qwen3-Max at 32k / 256k context; 3.5× / 7.2× versus Qwen3-235B-A22B ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894))
+- **Thinking mode:** thinking on by default with `<think>...</think>` traces; no `/think` `/nothink` soft switch; recommended thinking sampling `temperature=0.6, top_p=0.95, top_k=20` ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B))
+
+## Official vendor eval scores
+
+All values below are Qwen-reported (thinking mode, post-trained checkpoint) as published on the Hugging Face model card and the official Qwen3.5 announcement blog.
+
+| Benchmark              | Score                                       | Nemotron3 Ultra | % better | Source                                                                                                                                    |
+| ---------------------- | ------------------------------------------- | --------------- | -------- | ----------------------------------------------------------------------------------------------------------------------------------------- |
+| SWE-bench Verified     | 76.4                                        | 70.7            | +8.1%    | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| SWE-bench Multilingual | 69.3                                        | 67.7            | +2.4%    | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| Terminal Bench 2       | 52.5                                        | n.a.            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| TAU2-Bench (τ²-bench)  | 86.7                                        | n.a.            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| BFCL-V4                | 72.9                                        | n.a.            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| BrowseComp             | 69.0 (context-folding) / 78.6 (discard-all) | 44.4            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| LiveCodeBench v6       | 83.6                                        | 89.0 (v6)       | -6.1%    | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| AIME 2026              | 91.3                                        | n.a.            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| HMMT Feb 25            | 94.8                                        | n.a.            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| HMMT Nov 25            | 92.7                                        | n.a.            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| GPQA (Diamond)         | 88.4                                        | 87.0 (no tools) | +1.6%    | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| MMLU-Pro               | 87.8                                        | 86.8            | +1.2%    | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| HLE                    | 28.7 (HLE-Verified: 37.6)                   | 26.7 (no tools) | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+| MMMU (multimodal)      | 85.0                                        | n.a.            | —        | [Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B) / [Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894) |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+Vendor scaffold caveats stated by Qwen: τ²-bench follows the official setup except the airline domain, which uses the Claude Opus 4.5 system-card fixes; search-agent benchmarks use a 256k simple context-folding strategy, and BrowseComp's 78.6 uses the discard-all strategy also used by DeepSeek-V3.2 and Kimi K2.5 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)).
+
+## Benchmark explanations
+
+- **SWE-bench Verified** — human-validated set of real GitHub issues; measures whether the model's agentic patch actually resolves the issue and passes tests ([SWE-bench Verified leaderboard reference on the model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)).
+- **SWE-bench Multilingual** — same issue-resolution task extended beyond Python to repositories in multiple programming languages ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+- **Terminal Bench 2** — end-to-end agentic terminal tasks: the model must drive a shell to complete real command-line workloads ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+- **TAU2-Bench (τ²-bench)** — multi-turn tool-using customer-service style agent tasks across domains (including airline), scoring task completion under policy constraints ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)).
+- **BFCL-V4** — Berkeley Function Calling Leaderboard v4; accuracy of function/tool call selection and argument construction ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+- **BrowseComp** — hard web-browsing research questions requiring an agent to search, navigate and synthesize evidence over many hops, here with context-management strategies ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)).
+- **LiveCodeBench v6** — contamination-controlled competitive-programming problems collected over time; measures code generation correctness ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+- **AIME 2026** — American Invitational Mathematics Examination 2026 problems; short-answer competition math reasoning ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+- **HMMT Feb 25 / Nov 25** — Harvard-MIT Mathematics Tournament problem sets from the February and November 2025 contests; olympiad-level math accuracy ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+- **GPQA (Diamond)** — graduate-level, Google-proof multiple-choice questions in biology, chemistry and physics written by domain PhDs ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)).
+- **MMLU-Pro** — harder, reasoning-heavy revision of MMLU with expanded answer options, testing broad academic and professional knowledge ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+- **HLE / HLE-Verified** — Humanity's Last Exam, extremely difficult expert-written cross-domain questions; HLE-Verified is a revised version with a component-wise verification protocol and error taxonomy ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)).
+- **MMMU** — massive multi-discipline multimodal understanding: college-level questions requiring joint image and text reasoning ([Alibaba Cloud blog](https://www.alibabacloud.com/blog/602894)).
+
+## Caveats
+
+- Neither the Qwen blog nor the model card states an exact calendar release day; only "February 2026" is official. Third-party dates disagree (Feb 15 / 16 / 17, 2026).
+- The model card does not carry an explicit license statement in prose; Apache 2.0 is confirmed from the repository LICENSE file.
+- Alibaba has not published a formal Qwen3.5 tech report; the architecture facts on this page come from the model card, its `config.json`, and the announcement blog.
+- Vendor benchmark numbers use Qwen's own scaffolds and context-management strategies; the two BrowseComp figures illustrate how sensitive agentic scores are to harness choice.
diff --git a/packages/app/content/models/qwen-3-8-flash-next.mdx b/packages/app/content/models/qwen-3-8-flash-next.mdx
new file mode 100644
index 000000000..1cb875b60
--- /dev/null
+++ b/packages/app/content/models/qwen-3-8-flash-next.mdx
@@ -0,0 +1,123 @@
+---
+title: 'Qwen3.8-Flash-Next'
+developer: 'Alibaba (Qwen)'
+releaseDate: 'August 2026'
+description: "Alibaba's open-weights preview of the Qwen4 architecture: 176B total across a 125B main model and a 51B n-gram embedding table, 6B active, with Qwen Sparse Attention, gated residuals and a Gated DeltaNet hybrid stack."
+---
+
+## Overview
+
+Qwen3.8-Flash-Next is an open-weights preview of the architecture underlying **Qwen4**, published by Alibaba's Qwen team in August 2026 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next); [Qwen blog](https://qwen.ai/blog?id=qwen3.8-flash-next)). Qwen frames the release as deliberately architectural rather than a scaling exercise: the card describes the design as emphasising "architectural innovation for sustainable AGI progress, departing from the scaling-focused approach of simply increasing parameters and context windows" ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)). The stated purpose of shipping it early is to give the developer community time to prepare for the full Qwen4 family ([TechNode](https://technode.com/2026/08/26/alibabas-qwen-to-open-source-qwen3-8-flash-next-previewing-qwen4-architecture/)).
+
+The parameter count needs care. The card leads with "125B with 6B activated", but that figure covers the main model only. A **51B n-gram embedding table** brings the total to **176B**, and a separate **4B MTP head** sits outside both numbers ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8)). Reading the headline alone understates the checkpoint by roughly a third, which matters for capacity planning even though the n-gram table is a lookup rather than a compute-bearing parameter block.
+
+Weights ship under the `qwen-community-1.0` license, in BF16 and in an FP8 checkpoint that uses fine-grained quantization with a block size of 128; Qwen states the FP8 build's "performance metrics are nearly identical to those of the original model" ([Qwen3.8-Flash-Next-FP8 card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8)). Thinking mode is enabled by default, emitting `<think>...</think>` blocks, and is configurable through `enable_thinking`, `preserve_thinking` and `reasoning_effort`; preserved thinking retains reasoning blocks across conversation history by default ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+
+On InferenceX the model is benchmarked on the [AgentX](/glossary/agentx) agentic-traces scenario rather than fixed sequence lengths, served by SGLang with native NEXTN [multi-token prediction](/glossary/multi-token-prediction). Hopper cannot run the NVFP4 checkpoint — there are no SM100 tensor cores on sm_90 — so H200 serves the FP8 weights while Blackwell parts serve `RadixArk/Qwen3.8-Flash-Next-NVFP4`.
+
+## Architecture
+
+- **Total / active parameters:** 176B total — a 125B main model plus a 51B n-gram embedding table — with 6B activated per forward pass, and a further 4B in the MTP head ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Layers:** 48 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Hidden layout:** `12 × (3 × (Gated DeltaNet → MoE) → 1 × (Qwen Sparse Attention → MoE))` — a 3:1 linear-to-sparse-attention hybrid ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Hidden dimension:** 2560 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Gated DeltaNet (linear attention):** 48 value heads, 16 QK heads, head dimension 128 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Qwen Sparse Attention (full attention):** 24 Q heads, 2 KV heads, head dimension 256, rotary position embedding dimension 64 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **QSA indexer:** MQA with 4 query heads and 1 shared key head, indexer head dimension 128, budget of "512 blocks or 2048 tokens" ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **MoE experts:** 512 experts, 10 routed + 1 shared activated per token, expert intermediate dimension 640 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **N-gram embedding:** 20,000,000 bigram/trigram embeddings applied at layer 2 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Gated residual:** 4 branches, bottleneck rank 320 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **MTP:** 1 layer, trained with multi-steps ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Vocabulary:** 248,320 (padded) ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Context window:** 262,144 tokens native, extensible up to 1,000,000 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **Precision:** BF16 weights, plus an FP8 checkpoint using fine-grained quantization at block size 128 ([Qwen3.8-Flash-Next-FP8 card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8))
+- **License:** `qwen-community-1.0` ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next))
+
+## The Qwen4 architecture innovations
+
+Four changes distinguish this checkpoint from Qwen3.5 and the Qwen3-Next line it descends from. Each one attacks a different cost in the serving loop, which is why the model is interesting on a benchmark that measures whole systems rather than kernels.
+
+### N-gram embedding: parameter scaling that is not MoE
+
+The headline idea is a second axis for adding parameters. Qwen describes the n-gram table as "a unique axis for parameter scaling that requires less computation" than mixture-of-experts, achieving "highly efficient parameter scaling for memory-constrained accelerators" by "indexing with short n-grams" ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+
+The mechanics matter for anyone sizing a deployment. Twenty million bigram and trigram embeddings are applied at layer 2, contributing 51B parameters — more than 40% of the checkpoint's non-MTP weight — while adding a table lookup rather than a matrix multiply. MoE buys capacity by activating more experts, which costs both memory bandwidth and all-to-all traffic when experts are sharded; an n-gram table buys capacity by growing a structure that is read, not computed. On an accelerator where HBM capacity binds before FLOPS do, those are very different trades.
+
+It also changes what "active parameters" means as a proxy for cost. Six billion active parameters implies a very cheap forward pass, and for the MoE and attention stack it is; the n-gram table is capacity you pay for in memory footprint rather than in arithmetic per token.
+
+### Qwen Sparse Attention: sparsity at the micro-block level
+
+QSA replaces the Gated Attention layers of Qwen3.5's hybrid with a sparse mechanism that "operates at the micro-block level" rather than per token, which the card says "cuts long-context latency significantly" ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+
+The selection machinery is small and explicit: an MQA indexer with 4 query heads and a single shared key head, head dimension 128, choosing under a fixed budget of 512 blocks or 2048 tokens. A bounded budget is the operationally important part. Attention cost stops growing with context length once the budget binds, so the per-token cost of a 200k-token session converges toward the cost of a much shorter one, and long-context [interactivity](/glossary/interactivity) stops degrading the way a dense-attention model's does.
+
+Sparsity of this kind is a systems claim as much as a modelling one. As InferenceX has repeatedly found with [sparse attention](/glossary/sparse-attention) stacks, theoretical sparsity says little about realised speed: index construction, irregular memory access, kernel fusion and precision support decide what actually arrives, which is why the benchmark pins engine and image versions alongside the model.
+
+### Gated residual: finer control across a widened residual stream
+
+The gated residual "modulates information flowing through widened residual streams via an element-wise, data-dependent read gate and per-branch scalar write gate", which Qwen says provides "finer-grained expressiveness across layers while preserving training stability and keeping inference overhead low" ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)). The configuration is 4 branches at bottleneck rank 320.
+
+The read gate is element-wise and data-dependent; the write gate is a single scalar per branch. That asymmetry is the design: expressive where it is cheap to be expressive, and deliberately coarse where per-element parameters would add inference cost for little return.
+
+### The hybrid ratio and MTP, carried forward
+
+Two elements are inherited rather than new, and both matter for how the model serves. The 3:1 interleave of Gated DeltaNet to full-attention layers is the same ratio Qwen3.5 uses, keeping most layers on [linear attention](/glossary/linear-attention) whose state does not grow with sequence length. And the model ships a 1-layer MTP head trained with multiple steps, exposed by SGLang as native NEXTN [speculative decoding](/glossary/speculative-decoding).
+
+MTP is the piece with the largest effect on measured throughput, and also the one most sensitive to how it is benchmarked. AgentX replays anonymized traces filled with synthetic tokens, so a speculator would accept an unrepresentative number of draft tokens if left to the content; runs therefore apply a fixed [acceptance length](/glossary/acceptance-length) measured separately, which for this model at MTP=3 was 3.24 with thinking off.
+
+### Training recipe
+
+Qwen also reports optimizer-level changes: Muon and AdamW applied to specific weight categories, and the elimination of traditional batch-size warmups in favour of starting directly at the target batch size, which "substantially reduces total optimizer steps" ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)). This has no bearing on inference performance, but it is part of what Qwen means by previewing Qwen4 rather than shipping a larger Qwen3.
+
+## Official vendor eval scores
+
+All values below are Qwen-reported for Qwen3.8-Flash-Next as published on the Hugging Face model card. The card's own comparison columns name Qwen3.8-27B, Qwen3.7-Plus and DeepSeek-V4-Flash; those are omitted here in favour of the cross-page Nemotron 3 Ultra reference.
+
+| Benchmark                    | Score     | Nemotron3 Ultra     | % better | Source                                                          |
+| ---------------------------- | --------- | ------------------- | -------- | --------------------------------------------------------------- |
+| GPQA Diamond                 | 91.7      | 87.0 (no tools)     | +5.4%    | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| LiveCodeBench v6             | 91.9      | 89.0 (v6)           | +3.3%    | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| SWE-bench Multilingual       | 81.0      | 67.7                | +19.6%   | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| SWE-bench Pro                | 62.5      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| IFBench                      | 81.3      | 81.7 (prompt loose) | -0.5%    | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| DeepSWE 1.1                  | 58.7      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| Toolathlon Verified          | 73.5      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| CoWorkBench                  | 73.9      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| JobBench                     | 55.7      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| Agents' Last Exam (Pass@1)   | 25.2      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| ClawEval-MM (Pass@3)         | 64.4      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| AndroidWorld                 | 84.5      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| OSWorld 2.0 (Binary/Partial) | 19.4/52.3 | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| Vision2Web                   | 64.0      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| ERQA                         | 72.3      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| LVBench                      | 76.6      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| RealWorldQA                  | 88.5      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| RecreationBench              | 49.9      | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| MathVision (Without/With CI) | 90.6/95.7 | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+| CharXiv RQ (Without/With CI) | 84.6/90.6 | n.a.                | —        | [HF model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) |
+
+Nemotron 3 Ultra figures are the vendor-reported numbers from the [NVIDIA Nemotron 3 Ultra model card](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16), shown as a common reference point across model pages. `n.a.` means NVIDIA does not report that benchmark, or reports a different variant of it that would not be a like-for-like comparison. **% better** is `(score − Nemotron) / Nemotron`, computed only where both sides are a plain number on the same metric; a rating scale, a range, or a prose cell shows `—` instead. Neither column is a head-to-head evaluation: each score is self-reported by its own vendor under its own harness.
+
+## Benchmark explanations
+
+- **GPQA Diamond** — graduate-level, Google-proof multiple-choice questions in biology, chemistry and physics written by domain PhDs ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **LiveCodeBench v6** — contamination-controlled competitive-programming problems collected over time; measures code generation correctness ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **SWE-bench Multilingual** — real GitHub issue resolution extended beyond Python to repositories in multiple programming languages ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **SWE-bench Pro** — a harder issue-resolution set than SWE-bench Verified, with longer and more involved patches ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **DeepSWE 1.1** — agentic software-engineering tasks scored on whether the agent's changes resolve the task end to end ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **Toolathlon Verified** — verified subset of multi-tool agentic tasks, scoring correct tool selection and sequencing ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **CoWorkBench / JobBench** — workplace-style agentic task suites measuring completion of realistic multi-step office and job workflows ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **Agents' Last Exam** — very hard agentic tasks designed to remain unsolved by current systems; reported Pass@1 ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **IFBench** — instruction-following accuracy under explicit constraints ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **AndroidWorld / OSWorld 2.0** — GUI-agent benchmarks driving a real Android environment and a real desktop OS respectively; OSWorld reports binary and partial credit ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **Vision2Web / ClawEval-MM / RecreationBench** — multimodal agentic and understanding suites over web pages and visual scenes ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **LVBench / RealWorldQA / ERQA** — long-video understanding, real-world visual question answering, and embodied reasoning question answering ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+- **MathVision / CharXiv RQ** — visual mathematics and chart-reasoning benchmarks, each reported with and without a code interpreter ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)).
+
+## Caveats
+
+- **The 125B figure is the main model, not the checkpoint.** Total parameters are 176B once the 51B n-gram embedding table is counted, with a further 4B in the MTP head ([Hugging Face model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next-FP8)).
+- **All eval scores are vendor-reported**, produced by Qwen under its own harness. The card's comparison columns are also Qwen's own selection of rivals.
+- **Precision differs by chip generation.** Hopper has no NVFP4 path, so H200 results use the FP8 checkpoint while Blackwell results use NVFP4. Compare across chips with that in mind.
+- **The acceptance length used for MTP is measured separately**, not derived from the replayed traces, because AgentX content is synthetic. The interim value at the first sweep was 3.24 at MTP=3, collected with thinking off.
+- **This is an architecture preview.** Qwen positions the checkpoint as an early look at Qwen4 rather than a finished flagship, and engine support was days old at first benchmark, so results should be expected to move as kernels mature.
diff --git a/packages/app/cypress/component/agentic-distribution.cy.tsx b/packages/app/cypress/component/agentic-distribution.cy.tsx
new file mode 100644
index 000000000..cf20d02bc
--- /dev/null
+++ b/packages/app/cypress/component/agentic-distribution.cy.tsx
@@ -0,0 +1,94 @@
+import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.shared-runtime';
+
+import { Distribution } from '@/components/inference/agentic-point/distribution';
+
+// Mounted outside the Next app shell; next-style-loader inserts the global
+// stylesheet before this anchor, so it must exist before the import below.
+const cssAnchor = document.createElement('noscript');
+cssAnchor.id = '__next_css__DO_NOT_USE__';
+document.head.append(cssAnchor);
+require('@/app/globals.css');
+
+/** Deterministic lognormal sample so bar counts and the fit are reproducible. */
+function lognormalSample(count: number, mu: number, sigma: number): number[] {
+  let seed = 987_654;
+  const uniform = () => {
+    seed = (seed * 1103515245 + 12345) % 2147483648;
+    return (seed + 1) / 2147483649;
+  };
+  return Array.from({ length: count }, () => {
+    const z = Math.sqrt(-2 * Math.log(uniform())) * Math.cos(2 * Math.PI * uniform());
+    return Math.round(Math.exp(mu + sigma * z));
+  });
+}
+
+const FIT_STROKE = '#a855f7';
+
+function mountDistribution(values: number[], pathname = '/inference/agentic/96255') {
+  cy.mount(
+    <PathnameContext.Provider value={pathname}>
+      <div style={{ width: 900, padding: 16 }}>
+        <Distribution values={values} unit="tokens" />
+      </div>
+    </PathnameContext.Provider>,
+  );
+}
+
+describe('Agentic ISL/OSL distribution', () => {
+  it('bins on a log axis without a fitted overlay', () => {
+    const values = lognormalSample(1500, 7.3, 0.9);
+    mountDistribution(values);
+
+    cy.contains('1,500 requests').should('be.visible');
+    cy.contains('log scale').should('be.visible');
+    // Bars only — no fitted overlay or stats readout.
+    cy.contains('lognormal fit').should('not.exist');
+    cy.contains('μ=').should('not.exist');
+    cy.get(`path[stroke="${FIT_STROKE}"]`).should('not.exist');
+    cy.get('rect[opacity="0.55"]').should('have.length.greaterThan', 10);
+
+    // Percentile guides remain, one per percentile.
+    cy.get(
+      'line[stroke="#3b82f6"], line[stroke="#22c55e"], line[stroke="#f59e0b"], line[stroke="#ef4444"]',
+    ).should('have.length', 8);
+  });
+
+  it('spreads mass across the axis instead of piling it into the first bins', () => {
+    // The point of the log axis: on a linear one this sample puts almost
+    // everything in the leftmost bins. Assert the tallest bar is not the first.
+    mountDistribution(lognormalSample(1500, 7.3, 0.9));
+    cy.get('rect[opacity="0.55"]').then(($bars) => {
+      const heights = [...$bars].map((bar) => Number(bar.getAttribute('height')));
+      const peakIndex = heights.indexOf(Math.max(...heights));
+      expect(peakIndex).to.be.greaterThan(1);
+      expect(peakIndex).to.be.lessThan(heights.length - 2);
+    });
+  });
+
+  it('reports zero-token requests that a log axis cannot place', () => {
+    const values = [...lognormalSample(400, 6, 0.8), 0, 0, 0];
+    mountDistribution(values);
+    cy.contains('400 requests').should('be.visible');
+    cy.contains('3 requests with 0 tokens excluded from the log axis').should('be.visible');
+  });
+
+  it('still draws the histogram when every request is the same length', () => {
+    mountDistribution(Array.from({ length: 50 }, () => 512));
+    cy.contains('50 requests').should('be.visible');
+    cy.get('rect[opacity="0.55"]').should('have.length.greaterThan', 0);
+  });
+
+  it('shows the empty placeholder when nothing can be plotted', () => {
+    mountDistribution([]);
+    cy.contains('No data').should('be.visible');
+    cy.get(`path[stroke="${FIT_STROKE}"]`).should('not.exist');
+  });
+
+  it('renders in Simplified Chinese on the /zh route', () => {
+    mountDistribution(lognormalSample(600, 7, 0.8), '/zh/inference/agentic/96255');
+    cy.contains('600 个请求').should('be.visible');
+    cy.contains('对数刻度').should('be.visible');
+    cy.contains('数值（tokens，对数刻度）').should('be.visible');
+    cy.contains('requests').should('not.exist');
+  });
+});
diff --git a/packages/app/cypress/component/chart-legend.cy.tsx b/packages/app/cypress/component/chart-legend.cy.tsx
index 15b9ab100..b14e78c4c 100644
--- a/packages/app/cypress/component/chart-legend.cy.tsx
+++ b/packages/app/cypress/component/chart-legend.cy.tsx
@@ -1,4 +1,4 @@
-import { useState } from 'react';
+import { Profiler, useState } from 'react';
 
 import LegendPointsDialog from '@/components/inference/ui/LegendPointsDialog';
 import { OffloadHaloLegendKey } from '@/components/inference/ui/OffloadHaloLegendKey';
@@ -81,25 +81,47 @@ describe('ChartLegend (sidebar variant)', () => {
     cy.get('.sidebar-legend label').should('have.length', 4);
   });
 
-  it('legend items have colored dots', () => {
-    cy.get('.sidebar-legend label').first().find('span').first().should('exist');
+  it('derives long unofficial labels without a nested update', () => {
+    const branch = 'qwen3.5-fp4-gb200-dynamo-sglang-agentic-mtp-pareto-refresh';
+    const onRender = cy.spy().as('legendRender');
+
+    cy.mount(
+      <Profiler id="long-unofficial-run-legend" onRender={onRender}>
+        <ChartLegend
+          legendItems={[
+            {
+              ...MOCK_ITEMS[0],
+              name: 'unofficial-run-32177976542',
+              label: `✕ ${branch}`,
+            },
+          ]}
+          isLegendExpanded={true}
+          onExpandedChange={() => {}}
+          variant="sidebar"
+        />
+      </Profiler>,
+    );
+
+    cy.get('[data-testid="chart-legend"]').should('contain.text', branch);
+    cy.get('@legendRender').should((renderSpy) => {
+      // Cypress loses the Sinon spy type when resolving an alias.
+      const profilerSpy = renderSpy as unknown as {
+        getCalls: () => { args: unknown[] }[];
+      };
+      const calls = profilerSpy.getCalls();
+      expect(
+        calls.map(({ args }) => args[1]),
+        'React render phases',
+      ).not.to.include('nested-update');
+    });
   });
 
-  it('search input filters legend items by hiding non-matches', () => {
-    cy.get('.sidebar-legend input[placeholder="Search..."]').should('exist');
-    cy.get('.sidebar-legend input[placeholder="Search..."]').clear().type('MI300');
-    // Non-matching items are hidden via overflow-hidden class, not removed from DOM
-    cy.get('.sidebar-legend li.overflow-hidden').should('have.length', 3);
-    cy.get('.sidebar-legend li:not(.overflow-hidden)').should('have.length', 1);
-    cy.get('.sidebar-legend li:not(.overflow-hidden)').should('contain.text', 'AMD MI300X');
+  it('legend items have colored dots', () => {
+    cy.get('.sidebar-legend label').first().find('span').first().should('exist');
   });
 
-  it('search clear button resets search', () => {
-    cy.get('.sidebar-legend input[placeholder="Search..."]').type('test');
-    cy.get('button[aria-label="Clear search"]').should('be.visible');
-    cy.get('button[aria-label="Clear search"]').click();
-    cy.get('.sidebar-legend input[placeholder="Search..."]').should('have.value', '');
-    cy.get('button[aria-label="Clear search"]').should('not.exist');
+  it('renders no search input (removed from the sidebar panel)', () => {
+    cy.get('.sidebar-legend input[type="text"]').should('not.exist');
   });
 
   it('clicking a legend item toggles its active state', () => {
@@ -115,13 +137,14 @@ describe('ChartLegend (sidebar variant)', () => {
     cy.contains('Reset filter').should('not.exist');
   });
 
-  it('expand/collapse button toggles legend state', () => {
-    cy.get('.sidebar-legend').should('have.class', 'bg-accent');
-    cy.get('.sidebar-legend button')
-      .filter(':contains("Collapse"), :contains("Expand")')
-      .first()
-      .click();
-    cy.get('.sidebar-legend').should('not.have.class', 'bg-accent');
+  it('close button hides the panel and the reopen button restores it', () => {
+    cy.get('.sidebar-legend').should('exist');
+    cy.get('[data-testid="legend-close-button"]').click();
+    cy.get('.sidebar-legend').should('not.exist');
+    cy.get('[data-testid="legend-open-button"]').should('be.visible');
+    cy.get('[data-testid="legend-open-button"]').click();
+    cy.get('.sidebar-legend').should('exist');
+    cy.get('[data-testid="legend-open-button"]').should('not.exist');
   });
 
   it('renders no points-table icon when items have no onShowPoints handler', () => {
@@ -232,7 +255,7 @@ function LegendWithPointsTable() {
             if (!open) setOpenSeries(null);
           }}
           title={isOverlay ? '✕ my-branch' : 'B300 (vLLM)'}
-          subtitle="DeepSeek V4 Pro · Agentic Traces"
+          subtitle="DeepSeek V4 Pro · Agentic"
           accentColor={isOverlay ? '#dc2626' : '#2b83ba'}
           rows={buildLegendPointsRows(isOverlay ? OVERLAY_POINTS : OFFICIAL_POINTS, isOverlay)}
           isOverlay={isOverlay}
@@ -258,7 +281,7 @@ describe('ChartLegend points-table icon + dialog', () => {
     cy.get('[data-testid="legend-points-dialog"]').should('contain.text', 'B300 (vLLM)');
     cy.get('[data-testid="legend-points-dialog"]').should(
       'contain.text',
-      'DeepSeek V4 Pro · Agentic Traces',
+      'DeepSeek V4 Pro · Agentic',
     );
     // Two rows, conc ascending, linked to the agentic detail pages
     cy.get('[data-testid="legend-points-row"]').should('have.length', 2);
diff --git a/packages/app/cypress/component/chart-selectors.cy.tsx b/packages/app/cypress/component/chart-selectors.cy.tsx
index fd3c16d27..490ef6546 100644
--- a/packages/app/cypress/component/chart-selectors.cy.tsx
+++ b/packages/app/cypress/component/chart-selectors.cy.tsx
@@ -1,12 +1,14 @@
 import { useState } from 'react';
+import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.shared-runtime';
 
 import {
   ModelSelector,
+  ScenarioSelector,
   SequenceSelector,
   PrecisionSelector,
 } from '@/components/ui/chart-selectors';
 import { TooltipProvider } from '@/components/ui/tooltip';
-import { Model } from '@/lib/data-mappings';
+import { Model, Sequence } from '@/lib/data-mappings';
 
 function ModelSelectorHarness() {
   const [value, setValue] = useState('DeepSeek-R1-0528');
@@ -42,6 +44,20 @@ function SequenceSelectorHarness() {
   );
 }
 
+function ScenarioSelectorHarness({ initial = Sequence.AgenticTraces }: { initial?: Sequence }) {
+  const [value, setValue] = useState<string>(initial);
+  return (
+    <TooltipProvider delayDuration={0}>
+      <ScenarioSelector
+        value={value}
+        onChange={setValue}
+        availableSequences={[Sequence.EightK_OneK, Sequence.AgenticTraces]}
+        data-testid="scenario-selector"
+      />
+    </TooltipProvider>
+  );
+}
+
 function PrecisionSelectorHarness() {
   const [value, setValue] = useState(['FP8']);
   return (
@@ -93,6 +109,22 @@ describe('Chart Selectors', () => {
         'be.visible',
       );
     });
+
+    it('localizes model category labels and reasons on Chinese routes', () => {
+      cy.mount(
+        <PathnameContext.Provider value="/zh/inference">
+          <ModelSelectorHarness />
+        </PathnameContext.Provider>,
+      );
+      cy.get('[data-testid="model-selector"]').click();
+      cy.contains('维护模式').should('be.visible');
+      cy.contains('已弃用').should('be.visible');
+      cy.contains('Maintenance Mode').should('not.exist');
+      cy.get('[data-testid="selector-category-maintenance-mode-info"]').trigger('pointermove', {
+        pointerType: 'mouse',
+      });
+      cy.contains('这些模型的相关性较低，因此以较低优先级更新。').should('be.visible');
+    });
   });
 
   describe('SequenceSelector', () => {
@@ -112,6 +144,85 @@ describe('Chart Selectors', () => {
     });
   });
 
+  describe('ScenarioSelector', () => {
+    it('labels the agentic scenario "Agentic"', () => {
+      cy.mount(<ScenarioSelectorHarness />);
+      cy.get('[data-testid="scenario-selector"]').should('have.text', 'Agentic');
+      cy.get('[data-testid="scenario-selector"]').click();
+      cy.contains('[role="option"]', 'Agentic').should('be.visible');
+      cy.contains('[role="option"]', 'Agentic Traces').should('not.exist');
+      // The lone agentic entry needs no "Agentic" heading above it.
+      cy.get('[data-slot="select-content"]')
+        .find('[data-slot="select-label"]')
+        .should('not.contain.text', 'Agentic');
+    });
+
+    it('explains the agentic workload in a tooltip that links to /agentx', () => {
+      cy.mount(<ScenarioSelectorHarness />);
+      cy.get('[data-testid="scenario-agentic-info"]').trigger('pointermove', {
+        pointerType: 'mouse',
+      });
+
+      cy.contains('Realistic Long Context Multi Turn Agentic Workload with Sub Agents.').should(
+        'be.visible',
+      );
+      cy.get('[data-testid="scenario-agentic-info-link"]')
+        .should('be.visible')
+        .and('have.attr', 'href', '/agentx');
+    });
+
+    it('renders nothing when the only scenario is fixed-sequence', () => {
+      cy.mount(
+        <TooltipProvider delayDuration={0}>
+          <div data-testid="selector-host">
+            <ScenarioSelector
+              value={Sequence.EightK_OneK}
+              onChange={() => {}}
+              availableSequences={[Sequence.EightK_OneK]}
+              data-testid="scenario-selector"
+            />
+          </div>
+        </TooltipProvider>,
+      );
+      cy.get('[data-testid="selector-host"]').should('exist');
+      cy.get('[data-testid="scenario-selector"]').should('not.exist');
+      cy.contains('Scenario').should('not.exist');
+    });
+
+    it('renders nothing when agentic is the only scenario', () => {
+      // A static "Scenario: Agentic" readout is as redundant as a one-option
+      // dropdown — the whole control disappears for single-scenario models.
+      cy.mount(
+        <TooltipProvider delayDuration={0}>
+          <div data-testid="selector-host">
+            <ScenarioSelector
+              value={Sequence.AgenticTraces}
+              onChange={() => {}}
+              availableSequences={[Sequence.AgenticTraces]}
+              data-testid="scenario-selector"
+            />
+          </div>
+        </TooltipProvider>,
+      );
+      cy.get('[data-testid="selector-host"]').should('exist');
+      cy.get('[data-testid="scenario-selector"]').should('not.exist');
+      cy.contains('Scenario').should('not.exist');
+      cy.get('[data-testid="scenario-static-value"]').should('not.exist');
+      cy.get('[data-testid="scenario-agentic-info"]').should('not.exist');
+    });
+
+    it('hides the agentic explainer on fixed-sequence scenarios', () => {
+      cy.mount(<ScenarioSelectorHarness initial={Sequence.EightK_OneK} />);
+      cy.get('[data-testid="scenario-selector"]').should('contain.text', '8K / 1K');
+      cy.get('[data-testid="scenario-agentic-info"]').should('not.exist');
+
+      // ...and appears as soon as the user picks the agentic scenario.
+      cy.get('[data-testid="scenario-selector"]').click();
+      cy.contains('[role="option"]', 'Agentic').click();
+      cy.get('[data-testid="scenario-agentic-info"]').should('exist');
+    });
+  });
+
   describe('PrecisionSelector', () => {
     beforeEach(() => {
       cy.mount(<PrecisionSelectorHarness />);
@@ -120,5 +231,23 @@ describe('Chart Selectors', () => {
     it('shows current selection', () => {
       cy.get('[data-testid="precision-multiselect"]').should('contain', 'FP8');
     });
+
+    it('renders nothing when only one precision is available', () => {
+      cy.mount(
+        <TooltipProvider>
+          <div data-testid="selector-host">
+            <PrecisionSelector
+              value={['FP8']}
+              onChange={() => {}}
+              availablePrecisions={['FP8']}
+              data-testid="precision-multiselect"
+            />
+          </div>
+        </TooltipProvider>,
+      );
+      cy.get('[data-testid="selector-host"]').should('exist');
+      cy.get('[data-testid="precision-multiselect"]').should('not.exist');
+      cy.contains('Precision').should('not.exist');
+    });
   });
 });
diff --git a/packages/app/cypress/component/dataset-list.cy.tsx b/packages/app/cypress/component/dataset-list.cy.tsx
index f7cfcb9a7..ce62d6fa6 100644
--- a/packages/app/cypress/component/dataset-list.cy.tsx
+++ b/packages/app/cypress/component/dataset-list.cy.tsx
@@ -3,6 +3,7 @@ import { AppRouterContext } from 'next/dist/shared/lib/app-router-context.shared
 
 import { DatasetList } from '@/components/datasets/dataset-list';
 import type { DatasetRecord } from '@/hooks/api/use-datasets';
+import { createMockRouter } from '../support/mock-router';
 
 const datasets: DatasetRecord[] = [
   {
@@ -43,17 +44,6 @@ const datasets: DatasetRecord[] = [
   },
 ];
 
-function createMockRouter() {
-  return {
-    push: cy.stub(),
-    replace: cy.stub(),
-    refresh: cy.stub(),
-    back: cy.stub(),
-    forward: cy.stub(),
-    prefetch: cy.stub().resolves(),
-  };
-}
-
 function mountList() {
   const queryClient = new QueryClient({ defaultOptions: { queries: { retry: false } } });
   cy.mount(
@@ -74,7 +64,7 @@ describe('DatasetList', () => {
     cy.contains('cc-traces-weka (256k)').should('be.visible');
     cy.contains('1,234').should('be.visible'); // conversation_count, localized
     cy.contains('82%').should('be.visible'); // cachedPct
-    cy.get('a[href="/datasets/cc-traces-weka-full"]').should('exist');
+    cy.get('a[href="/agentx/cc-traces-weka-full"]').should('exist');
   });
 
   it('shows the empty state when no datasets are ingested', () => {
diff --git a/packages/app/cypress/component/date-picker.cy.tsx b/packages/app/cypress/component/date-picker.cy.tsx
index 68465586b..631f03c50 100644
--- a/packages/app/cypress/component/date-picker.cy.tsx
+++ b/packages/app/cypress/component/date-picker.cy.tsx
@@ -1,4 +1,5 @@
 import { useState } from 'react';
+import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.shared-runtime';
 
 import { DatePicker } from '@/components/ui/date-picker';
 
@@ -16,9 +17,11 @@ const AVAILABLE_DATES = [
 function DatePickerWrapper({
   initialDate,
   availableDates = AVAILABLE_DATES,
+  useDefaults = false,
 }: {
   initialDate?: string;
   availableDates?: string[];
+  useDefaults?: boolean;
 }) {
   const [date, setDate] = useState<string | undefined>(initialDate);
   return (
@@ -28,7 +31,7 @@ function DatePickerWrapper({
       availableDates={availableDates}
       minDate={availableDates[0]}
       maxDate={availableDates.at(-1)}
-      placeholder="Select date"
+      placeholder={useDefaults ? undefined : 'Select date'}
     />
   );
 }
@@ -118,4 +121,23 @@ describe('DatePicker', () => {
     // Original date should still be shown
     cy.contains('Nov 15, 2025').should('be.visible');
   });
+
+  it('localizes the complete dialog chrome on Chinese routes', () => {
+    cy.mount(
+      <PathnameContext.Provider value="/zh/inference">
+        <DatePickerWrapper initialDate="2025-11-15" useDefaults />
+      </PathnameContext.Provider>,
+    );
+    cy.contains('运行日期：').click();
+    cy.contains('选择运行日期').should('be.visible');
+    cy.contains('选择运行日期以查看该次运行的性能数据。').should('be.visible');
+    cy.contains('button', '前往最新日期').should('be.visible');
+    cy.contains('button', '取消').should('be.visible');
+    cy.contains('button', '应用').should('be.visible');
+    cy.get('h3.font-semibold').should('contain', '2025年11月');
+    cy.get('[role="dialog"]')
+      .should('contain.text', '关闭')
+      .and('not.contain.text', 'Cancel')
+      .and('not.contain.text', 'Close');
+  });
 });
diff --git a/packages/app/cypress/component/date-range-picker.cy.tsx b/packages/app/cypress/component/date-range-picker.cy.tsx
index 9fd7c4ed2..61aeb7d86 100644
--- a/packages/app/cypress/component/date-range-picker.cy.tsx
+++ b/packages/app/cypress/component/date-range-picker.cy.tsx
@@ -1,13 +1,16 @@
 import { useState } from 'react';
+import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.shared-runtime';
 
 import { type DateRange, DateRangePicker } from '@/components/ui/date-range-picker';
 
 function DateRangePickerHarness({
   initialRange = { startDate: '', endDate: '' },
   availableDates,
+  useDefaults = false,
 }: {
   initialRange?: DateRange;
   availableDates?: string[];
+  useDefaults?: boolean;
 }) {
   const [range, setRange] = useState<DateRange>(initialRange);
   return (
@@ -16,7 +19,7 @@ function DateRangePickerHarness({
         dateRange={range}
         onChange={setRange}
         availableDates={availableDates}
-        placeholder="Select date range"
+        placeholder={useDefaults ? undefined : 'Select date range'}
       />
       <div data-testid="date-range-output">
         {range.startDate && range.endDate ? `${range.startDate} to ${range.endDate}` : 'no range'}
@@ -86,4 +89,19 @@ describe('DateRangePicker', () => {
     cy.get('[role="dialog"]').should('not.exist');
     cy.get('[data-testid="date-range-output"]').should('contain', '2026-01-01 to 2026-01-01');
   });
+
+  it('localizes status, quick actions, and dialog actions on Chinese routes', () => {
+    const dates = ['2025-12-01', '2026-01-01', '2026-02-01'];
+    cy.mount(
+      <PathnameContext.Provider value="/zh/inference">
+        <DateRangePickerHarness availableDates={dates} useDefaults />
+      </PathnameContext.Provider>,
+    );
+    cy.contains('选择日期范围').click();
+    cy.contains('请选择开始和结束日期以定义日期范围。').should('be.visible');
+    cy.contains('button', '最大范围').should('be.visible');
+    cy.contains('button', '取消').should('be.visible');
+    cy.contains('button', '应用').should('be.visible');
+    cy.get('[role="dialog"]').should('not.contain.text', 'Max Range');
+  });
 });
diff --git a/packages/app/cypress/component/faq-question-link.cy.tsx b/packages/app/cypress/component/faq-question-link.cy.tsx
new file mode 100644
index 000000000..df2502c9e
--- /dev/null
+++ b/packages/app/cypress/component/faq-question-link.cy.tsx
@@ -0,0 +1,55 @@
+import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.shared-runtime';
+
+import { FaqQuestionLink } from '@/components/about/faq-question-link';
+
+const ID = 'faq-normalized-interactivity';
+const QUESTION = 'What is the difference between normalized interactivity and interactivity?';
+
+function mountFaqQuestionLink(pathname = '/about') {
+  cy.mount(
+    <PathnameContext.Provider value={pathname}>
+      <div className="w-[760px] p-4 font-medium">
+        <FaqQuestionLink id={ID} question={QUESTION} />
+      </div>
+    </PathnameContext.Provider>,
+  );
+}
+
+describe('FaqQuestionLink', () => {
+  it('shows an icon-only control and copies the direct URL', () => {
+    cy.window().then((win) => {
+      cy.stub(win.navigator.clipboard, 'writeText').as('writeFaqLink').resolves();
+    });
+    mountFaqQuestionLink();
+
+    cy.get(`a[href="#${ID}"]`).should('have.text', QUESTION);
+    cy.get(`[data-testid="faq-copy-link-${ID}"]`)
+      .should('be.visible')
+      .and('have.text', '')
+      .and('have.attr', 'title', 'Copy link')
+      .and('have.attr', 'aria-label', `Copy link: ${QUESTION}`)
+      .find('svg.lucide-link')
+      .should('be.visible');
+    cy.get(`[data-testid="faq-copy-link-${ID}"]`)
+      .click()
+      .should('have.attr', 'title', 'Copied')
+      .and('have.attr', 'aria-label', `Copied: ${QUESTION}`)
+      .find('svg.lucide-check')
+      .should('be.visible');
+    cy.get('@writeFaqLink').should(
+      'have.been.calledOnceWith',
+      Cypress.sinon.match(new RegExp(`#${ID}$`, 'u')),
+    );
+  });
+
+  it('uses concise Chinese button copy on the Chinese route', () => {
+    mountFaqQuestionLink('/zh/about');
+
+    cy.get(`[data-testid="faq-copy-link-${ID}"]`)
+      .should('be.visible')
+      .and('have.attr', 'aria-label', `复制链接: ${QUESTION}`)
+      .and('have.attr', 'title', '复制链接')
+      .find('svg.lucide-link')
+      .should('be.visible');
+  });
+});
diff --git a/packages/app/cypress/component/footer.cy.tsx b/packages/app/cypress/component/footer.cy.tsx
index 6e7916779..4224d9f01 100644
--- a/packages/app/cypress/component/footer.cy.tsx
+++ b/packages/app/cypress/component/footer.cy.tsx
@@ -49,8 +49,13 @@ describe('Footer', () => {
     cy.get('[data-testid="footer-link-benchmarks"]')
       .should('have.attr', 'href')
       .and('include', 'github.com/SemiAnalysisAI/InferenceX');
-    cy.get('[data-testid="footer-link-frontend"]')
-      .should('have.attr', 'href')
+    cy.get('[data-testid="footer-link-agentx-harness"]')
+      .should('contain.text', 'AgentX Harness')
+      .and('have.attr', 'href')
+      .and('include', 'github.com/SemiAnalysisAI/agentx-harness');
+    cy.get('[data-testid="footer-link-visualization"]')
+      .should('contain.text', 'Visualization')
+      .and('have.attr', 'href')
       .and('include', 'github.com/SemiAnalysisAI/InferenceX-app');
   });
 
@@ -58,9 +63,9 @@ describe('Footer', () => {
     cy.get('[data-testid="footer-link-supporters"]')
       .should('contain.text', 'Supporters')
       .and('have.attr', 'href', '/quotes');
-    cy.get('[data-testid="footer-link-datasets"]')
-      .should('contain.text', 'Datasets')
-      .and('have.attr', 'href', '/datasets');
+    cy.get('[data-testid="footer-link-agentx"]')
+      .should('contain.text', 'AgentX')
+      .and('have.attr', 'href', '/agentx');
     cy.get('[data-testid="footer-link-articles"]')
       .should('contain.text', 'Articles')
       .and('have.attr', 'href', '/blog');
diff --git a/packages/app/cypress/component/gpu-graph.cy.tsx b/packages/app/cypress/component/gpu-graph.cy.tsx
index 0634ffff6..4589327fd 100644
--- a/packages/app/cypress/component/gpu-graph.cy.tsx
+++ b/packages/app/cypress/component/gpu-graph.cy.tsx
@@ -185,12 +185,12 @@ describe('GPUGraph', () => {
     cy.get('#test-gpu-agentic-decorations [data-testid="spec-decode-marker-key"]').should(
       'not.exist',
     );
-    cy.get('#test-gpu-agentic-decorations [data-testid="offload-halo-key"]').should('exist');
-    cy.get('#test-gpu-agentic-decorations [data-testid="agentic-optimization-note"]')
-      .should('contain.text', 'Inference optimizations enabled')
-      .find('button')
-      .focus();
-    cy.contains('Each configuration may use inference optimizations').should('be.visible');
+    // The KV-offload key and optimization note moved to the axis-metric info
+    // footer rendered by ChartDisplay, so the chart itself carries neither.
+    cy.get('#test-gpu-agentic-decorations [data-testid="offload-halo-key"]').should('not.exist');
+    cy.get('#test-gpu-agentic-decorations [data-testid="agentic-optimization-note"]').should(
+      'not.exist',
+    );
     cy.get('#test-gpu-agentic-decorations svg .dot-group').first().trigger('mouseenter');
     cy.get('[data-chart-tooltip]').should('contain.text', 'Speculative Decoding');
     cy.get('[data-chart-tooltip]').should('contain.text', 'MTP');
diff --git a/packages/app/cypress/component/header.cy.tsx b/packages/app/cypress/component/header.cy.tsx
index 8589264ab..6d85f6247 100644
--- a/packages/app/cypress/component/header.cy.tsx
+++ b/packages/app/cypress/component/header.cy.tsx
@@ -4,6 +4,7 @@ import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.share
 
 import { Header } from '@/components/header/header';
 import { ThemeProvider } from '@/components/ui/theme-provider';
+import { createMockRouter } from '../support/mock-router';
 
 // Mounted outside the Next app shell; next-style-loader inserts the global
 // stylesheet before this anchor, so it must exist before the import below.
@@ -21,17 +22,6 @@ const MIN_TOUCH_PX = 44;
 /** Tolerance for sub-pixel layout rounding. */
 const EPSILON = 0.5;
 
-function createMockRouter() {
-  return {
-    push: cy.stub(),
-    replace: cy.stub(),
-    refresh: cy.stub(),
-    back: cy.stub(),
-    forward: cy.stub(),
-    prefetch: cy.stub().resolves(),
-  };
-}
-
 function rectOf(selector: string) {
   return cy.get(selector).then(($el) => $el[0].getBoundingClientRect());
 }
@@ -80,6 +70,20 @@ describe('Header', () => {
     cy.wrap(mockRouter.push).should('have.been.calledTwice');
   });
 
+  it('retains resilient locale navigation outside Overview', () => {
+    cy.clock();
+    mountHeader('/inference');
+    cy.get('[data-testid="language-toggle"]')
+      .invoke('attr', 'href')
+      .then((href) => {
+        expect(href).to.include('/zh/inference');
+        cy.get('[data-testid="language-toggle"]').click();
+        cy.wrap(mockRouter.push).should('have.been.calledOnceWith', href);
+        cy.tick(250);
+        cy.wrap(mockRouter.push).should('have.been.calledTwice');
+      });
+  });
+
   it('makes a click on the tab already showing a no-op', () => {
     mountHeader('/overview');
     cy.get('[data-testid="nav-link-overview"]').click();
@@ -104,10 +108,50 @@ describe('Header', () => {
     cy.get('[data-testid="nav-link-compare"]').should('have.attr', 'href', '/compare');
   });
 
+  it('shows AgentX as a top-level nav link and highlights AgentX child pages', () => {
+    cy.get('[data-testid="nav-link-agentx"]')
+      .should('be.visible')
+      .and('have.attr', 'href', '/agentx')
+      .find('[data-nav-badge="agentx"]')
+      .should('have.text', 'NEW');
+
+    mountHeader('/agentx/claude-code-traces');
+    cy.get('[data-testid="nav-link-agentx"]').should('have.class', 'text-brand');
+  });
+
+  it('keeps AgentX in the Chinese navigation tree', () => {
+    mountHeader('/zh/agentx');
+    cy.get('[data-testid="nav-link-agentx"]')
+      .should('be.visible')
+      .and('contain.text', 'AgentX')
+      .and('have.attr', 'href', '/zh/agentx')
+      .and('have.class', 'text-brand');
+    cy.get('[data-testid="nav-link-agentx"]')
+      .find('[data-nav-badge="agentx"]')
+      .should('have.text', '新');
+  });
+
+  it('orders the nav with Home first and AgentX second', () => {
+    const expected = [
+      'Home',
+      'AgentX',
+      'Overview',
+      'Dashboard',
+      'Comparisons',
+      'Articles',
+      'About',
+    ];
+    cy.get('[data-testid^="nav-link-"]').then(($links) => {
+      // Strip the NEW badge so the comparison is against the label alone.
+      const labels = [...$links].map((link) => (link.textContent ?? '').replace('NEW', '').trim());
+      expect(labels).to.deep.equal(expected);
+    });
+  });
+
   it('keeps footer destinations out of the primary nav', () => {
     cy.get('[data-testid="nav-link-supporters"]').should('not.exist');
-    cy.get('[data-testid="nav-link-datasets"]').should('not.exist');
-    cy.get('[data-testid="nav-link-blog"]').should('not.exist');
+    // Telemetry moved to the footer; Articles took its place in the nav.
+    cy.get('[data-testid="nav-link-telemetry"]').should('not.exist');
   });
 
   it('shows the GitHub stars button linking to the correct repo', () => {
@@ -133,9 +177,14 @@ describe('Header', () => {
       cy.contains('a', 'Overview').should('be.visible').and('have.attr', 'href', '/overview');
       cy.contains('a', 'Dashboard').should('be.visible').and('have.attr', 'href', '/inference');
       cy.contains('a', 'Comparisons').should('be.visible').and('have.attr', 'href', '/compare');
+      cy.contains('a', 'AgentX')
+        .should('be.visible')
+        .and('have.attr', 'href', '/agentx')
+        .find('[data-nav-badge="agentx"]')
+        .should('have.text', 'NEW');
+      cy.contains('a', 'Articles').should('be.visible').and('have.attr', 'href', '/blog');
       cy.contains('a', 'Supporters').should('not.exist');
-      cy.contains('a', 'Datasets').should('not.exist');
-      cy.contains('a', 'Articles').should('not.exist');
+      cy.contains('a', 'Telemetry').should('not.exist');
     });
   });
 
@@ -149,6 +198,40 @@ describe('Header', () => {
     cy.wrap(mockRouter.push).should('have.been.calledTwice');
   });
 
+  it('uses the hamburger without horizontal overflow from 1009 through 1024 CSS pixels', () => {
+    [1009, 1012, 1020, 1024].forEach((width) => {
+      cy.viewport(width, 720);
+      cy.get('[data-testid="nav-link-dashboard"]').should('not.be.visible');
+      cy.get('[data-testid="mobile-menu-toggle"]').should('be.visible');
+      cy.document().then((doc) => {
+        expect(doc.documentElement.scrollWidth, `${width}px document scrollWidth`).to.be.at.most(
+          doc.documentElement.clientWidth,
+        );
+      });
+      cy.get('[data-testid="header"]').then(($header) => {
+        const header = $header[0];
+        expect(header.scrollWidth, `${width}px header scrollWidth`).to.be.at.most(
+          header.clientWidth,
+        );
+      });
+    });
+  });
+
+  it('keeps every primary link inside the header at the xl desktop breakpoint', () => {
+    cy.viewport(1280, 720);
+    cy.get('[data-testid="header"]').then(($header) => {
+      const header = $header[0];
+      const bounds = header.getBoundingClientRect();
+      expect(header.scrollWidth, 'header scrollWidth').to.be.at.most(header.clientWidth);
+
+      cy.get('[data-testid^="nav-link-"]:visible').each(($link) => {
+        const rect = $link[0].getBoundingClientRect();
+        expect(rect.left, `${$link.text()} left edge`).to.be.at.least(bounds.left - EPSILON);
+        expect(rect.right, `${$link.text()} right edge`).to.be.at.most(bounds.right + EPSILON);
+      });
+    });
+  });
+
   describe('at 320x700', () => {
     beforeEach(() => {
       cy.viewport(320, 700);
@@ -204,10 +287,12 @@ describe('Header', () => {
       cy.get('[data-testid="mobile-menu-toggle"]').click();
       cy.get('[data-testid="mobile-menu"]').should('be.visible');
       cy.get('[data-testid="mobile-menu"]').within(() => {
-        ['Home', 'Overview', 'Dashboard', 'Comparisons', 'About'].forEach((label) => {
-          cy.contains('a', label).should('be.visible');
-        });
-        ['Supporters', 'Datasets', 'Articles'].forEach((label) => {
+        ['Home', 'Overview', 'Dashboard', 'Comparisons', 'Articles', 'AgentX', 'About'].forEach(
+          (label) => {
+            cy.contains('a', label).should('be.visible');
+          },
+        );
+        ['Supporters', 'Telemetry'].forEach((label) => {
           cy.contains('a', label).should('not.exist');
         });
       });
diff --git a/packages/app/cypress/component/inference-chart-controls.cy.tsx b/packages/app/cypress/component/inference-chart-controls.cy.tsx
index f6db52302..3884388fa 100644
--- a/packages/app/cypress/component/inference-chart-controls.cy.tsx
+++ b/packages/app/cypress/component/inference-chart-controls.cy.tsx
@@ -26,6 +26,7 @@ describe('Inference ChartControls', () => {
 
   it('renders the Y-axis metric selector', () => {
     cy.get('[data-testid="yaxis-metric-selector"]').should('be.visible');
+    cy.get('[data-testid="cost-display-selector"]').should('not.exist');
   });
 
   it('Y-axis metric selector shows grouped options', () => {
@@ -41,6 +42,36 @@ describe('Inference ChartControls', () => {
     cy.get('@setSelectedYAxisMetric').should('have.been.calledOnce');
   });
 
+  it('lists and selects the schema-v2 derived axes in the Measured Energy group', () => {
+    const options = [
+      {
+        key: 'y_measuredJPerSuccessfulQuery',
+        label: 'Measured Joules per Successful Query',
+      },
+      {
+        key: 'y_measuredWhPerSuccessfulQuery',
+        label: 'Measured Watt-hours per Successful Query',
+      },
+      {
+        key: 'y_measuredPowerPercentTdp',
+        label: 'Measured Average Power as Percent of TDP',
+      },
+    ];
+
+    for (const option of options) {
+      cy.get('[data-testid="yaxis-metric-selector"]').click();
+      cy.contains('Measured Energy')
+        .parent()
+        .within(() => {
+          cy.contains('[role="option"]', option.label)
+            .scrollIntoView()
+            .should('be.visible')
+            .click();
+        });
+      cy.get('@setSelectedYAxisMetric').should('have.been.calledWith', option.key);
+    }
+  });
+
   it('hides the GPU comparison section when no GPUs are selected', () => {
     // Default mock: selectedGPUs = [] — GPU date range pickers should not render
     cy.contains('Comparison Date Range').should('not.exist');
@@ -54,6 +85,31 @@ describe('Inference ChartControls', () => {
   });
 });
 
+describe('Inference ChartControls cost metrics', () => {
+  beforeEach(() => {
+    mountWithProviders(<InferenceChartControls />, {
+      inference: { selectedYAxisMetric: 'y_costh' },
+    });
+  });
+
+  it('shows cost per million and tokens per dollar as separate Y-axis options', () => {
+    cy.get('[data-testid="yaxis-metric-selector"]').click();
+    cy.contains('[role="option"]', 'Cost per Million Total Tokens (Owning - Hyperscaler)').should(
+      'exist',
+    );
+    cy.contains('[role="option"]', 'Total Tokens per $1 USD (Owning - Hyperscaler)').should(
+      'exist',
+    );
+    cy.get('[data-testid="cost-display-selector"]').should('not.exist');
+  });
+
+  it('selects tokens per dollar through the Y-axis metric control', () => {
+    cy.get('[data-testid="yaxis-metric-selector"]').click();
+    cy.contains('[role="option"]', 'Total Tokens per $1 USD (Owning - Hyperscaler)').click();
+    cy.get('@setSelectedYAxisMetric').should('have.been.calledWith', 'y_tokensPerDollarH');
+  });
+});
+
 describe('Inference ChartControls with GPUs selected', () => {
   it('shows the date range picker when GPUs are selected', () => {
     mountWithProviders(<InferenceChartControls />, {
@@ -66,7 +122,7 @@ describe('Inference ChartControls with GPUs selected', () => {
     cy.contains('Comparison Date Range').should('be.visible');
   });
 
-  it('flags the date range when nothing has been picked to compare against', () => {
+  it('leaves the optional date range unflagged for a selected current config', () => {
     mountWithProviders(<InferenceChartControls />, {
       inference: {
         selectedGPUs: ['h100'],
@@ -75,7 +131,9 @@ describe('Inference ChartControls with GPUs selected', () => {
       },
     });
 
-    cy.contains('button', 'Select date range').should('have.class', 'animate-pulse');
+    cy.contains('button', 'Select date range')
+      .should('not.have.class', 'animate-pulse')
+      .and('not.have.class', 'border-red-500');
   });
 
   it('leaves the date range unflagged when exact comparison entries are pinned', () => {
diff --git a/packages/app/cypress/component/multi-date-picker.cy.tsx b/packages/app/cypress/component/multi-date-picker.cy.tsx
deleted file mode 100644
index 9f80ff427..000000000
--- a/packages/app/cypress/component/multi-date-picker.cy.tsx
+++ /dev/null
@@ -1,128 +0,0 @@
-import { useState } from 'react';
-
-import { MultiDatePicker } from '@/components/ui/multi-date-picker';
-
-const AVAILABLE_DATES = [
-  '2025-10-15',
-  '2025-11-02',
-  '2025-11-05',
-  '2025-11-10',
-  '2025-11-15',
-  '2025-11-20',
-  '2025-12-01',
-  '2025-12-05',
-  '2025-12-15',
-];
-
-function MultiDatePickerWrapper({
-  initialDates = [],
-  maxDates = 3,
-}: {
-  initialDates?: string[];
-  maxDates?: number;
-}) {
-  const [dates, setDates] = useState<string[]>(initialDates);
-  return (
-    <MultiDatePicker
-      dates={dates}
-      onChange={setDates}
-      maxDates={maxDates}
-      availableDates={AVAILABLE_DATES}
-      minDate={AVAILABLE_DATES[0]}
-      maxDate={AVAILABLE_DATES.at(-1)}
-      placeholder="Select dates"
-    />
-  );
-}
-
-describe('MultiDatePicker', () => {
-  it('renders placeholder when no dates selected', () => {
-    cy.mount(<MultiDatePickerWrapper />);
-    cy.contains('Select dates').should('be.visible');
-  });
-
-  it('shows formatted text for selected dates', () => {
-    // 1 date: shows the formatted date
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-15']} />);
-    cy.contains('Nov 15, 2025').should('be.visible');
-  });
-
-  it('shows "vs" text for 2 selected dates', () => {
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-10', '2025-11-20']} />);
-    cy.contains('Nov 10, 2025 vs Nov 20, 2025').should('be.visible');
-  });
-
-  it('shows count text for 3+ selected dates', () => {
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-01', '2025-11-10', '2025-11-20']} />);
-    cy.contains('3 dates selected').should('be.visible');
-  });
-
-  it('calendar opens on click', () => {
-    cy.mount(<MultiDatePickerWrapper />);
-    cy.contains('Select dates').click();
-    cy.contains('Select Comparison Dates').should('be.visible');
-  });
-
-  it('can select multiple dates', () => {
-    // Start with a date in November so calendar opens to November
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-02']} maxDates={3} />);
-    cy.contains('Nov 2, 2025').click();
-    // Calendar opens to November. Available within range: Nov 2, 5, 10, 15, 20.
-    cy.get('.grid-cols-7 button').contains(/^5$/u).click();
-    cy.get('.grid-cols-7 button').contains('10').click();
-    // Should see the selected dates as pills
-    cy.contains('Selected Dates:').should('be.visible');
-    cy.contains('Nov 2, 2025').should('be.visible');
-    cy.contains('Nov 5, 2025').should('be.visible');
-    cy.contains('Nov 10, 2025').should('be.visible');
-  });
-
-  it('cannot exceed maxDates', () => {
-    // Start with 1 date already selected in November, calendar opens to Nov
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-02']} maxDates={2} />);
-    cy.contains('Nov 2, 2025').click();
-    // Calendar opens to November (month of selected date). Select one more.
-    cy.get('.grid-cols-7 button').contains(/^5$/u).click();
-    // Now at maxDates=2, third date (10) should be disabled
-    cy.get('.grid-cols-7 button').contains('10').should('be.disabled');
-  });
-
-  it('selected dates show as removable pills', () => {
-    // Start with a date so calendar opens to its month
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-05']} maxDates={3} />);
-    cy.contains('Nov 5, 2025').click();
-    // Calendar opens to November. Add another date.
-    cy.get('.grid-cols-7 button').contains('10').click();
-    // Pill should have a remove button
-    cy.get('[aria-label="Remove Nov 10, 2025"]').should('be.visible');
-  });
-
-  it('remove button on pill deselects date', () => {
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-05']} maxDates={3} />);
-    cy.contains('Nov 5, 2025').click();
-    // Calendar opens to November. Select another date.
-    cy.get('.grid-cols-7 button').contains('10').click();
-    // Remove the first pill (Nov 5)
-    cy.get('[aria-label="Remove Nov 5, 2025"]').click();
-    // Only Nov 10 should remain as a pill in the selected dates section
-    cy.contains('Selected Dates:')
-      .parents('.rounded-md')
-      .first()
-      .within(() => {
-        cy.contains('Nov 5, 2025').should('not.exist');
-        cy.contains('Nov 10, 2025').should('be.visible');
-      });
-  });
-
-  it('Clear All removes all selected dates', () => {
-    cy.mount(<MultiDatePickerWrapper initialDates={['2025-11-05']} maxDates={3} />);
-    cy.contains('Nov 5, 2025').click();
-    // Calendar opens to November. Select another date.
-    cy.get('.grid-cols-7 button').contains('10').click();
-    cy.contains('Selected Dates:').should('be.visible');
-    // Click Clear All
-    cy.contains('Clear All').click();
-    // Selected Dates section should disappear (no pills)
-    cy.contains('Selected Dates:').should('not.exist');
-  });
-});
diff --git a/packages/app/cypress/component/multi-select.cy.tsx b/packages/app/cypress/component/multi-select.cy.tsx
index 76d1ffdfe..e20bf0442 100644
--- a/packages/app/cypress/component/multi-select.cy.tsx
+++ b/packages/app/cypress/component/multi-select.cy.tsx
@@ -1,4 +1,5 @@
 import { useState } from 'react';
+import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.shared-runtime';
 
 import { MultiSelect, type MultiSelectOption } from '@/components/ui/multi-select';
 
@@ -15,11 +16,13 @@ function MultiSelectWrapper({
   maxSelections,
   minSelections,
   searchable = true,
+  useDefaults = false,
 }: {
   initial?: string[];
   maxSelections?: number;
   minSelections?: number;
   searchable?: boolean;
+  useDefaults?: boolean;
 }) {
   const [value, setValue] = useState<string[]>(initial);
   return (
@@ -27,7 +30,7 @@ function MultiSelectWrapper({
       options={OPTIONS}
       value={value}
       onChange={setValue}
-      placeholder="Select Chips..."
+      placeholder={useDefaults ? undefined : 'Select Chips...'}
       maxSelections={maxSelections}
       minSelections={minSelections}
       searchable={searchable}
@@ -115,6 +118,23 @@ describe('MultiSelect', () => {
     cy.contains('Select Chips...').should('be.visible');
   });
 
+  it('localizes visible and accessible defaults on Chinese routes', () => {
+    cy.mount(
+      <PathnameContext.Provider value="/zh/inference">
+        <MultiSelectWrapper initial={['h100-sxm']} maxSelections={2} useDefaults />
+      </PathnameContext.Provider>,
+    );
+
+    cy.get('[aria-label="移除 NVIDIA H100 SXM"]').should('exist');
+    cy.get('[aria-label="清除所有选择"]').should('exist');
+    cy.get('[data-slot="select-trigger"]').click();
+    cy.get('input[placeholder="搜索..."]')
+      .should('have.attr', 'aria-label', '搜索选项')
+      .type('不存在');
+    cy.contains('没有结果').should('be.visible');
+    cy.contains('1 / 2 项已选择').should('be.visible');
+  });
+
   it('click outside closes dropdown', () => {
     cy.mount(
       <div>
diff --git a/packages/app/cypress/component/normalized-interactivity-help-link.cy.tsx b/packages/app/cypress/component/normalized-interactivity-help-link.cy.tsx
new file mode 100644
index 000000000..0774c8f81
--- /dev/null
+++ b/packages/app/cypress/component/normalized-interactivity-help-link.cy.tsx
@@ -0,0 +1,35 @@
+import { NormalizedInteractivityHelpLink } from '@/components/inference/ui/NormalizedInteractivityHelpLink';
+
+describe('NormalizedInteractivityHelpLink', () => {
+  it('renders a subtle icon-only link to the English FAQ', () => {
+    cy.mount(
+      <div className="relative h-10 w-80">
+        <NormalizedInteractivityHelpLink locale="en" />
+      </div>,
+    );
+
+    cy.get('[data-testid="normalized-interactivity-faq-link"]')
+      .should('be.visible')
+      .and('have.text', '')
+      .and('have.attr', 'href', '/about#faq-normalized-interactivity')
+      .and('have.attr', 'title', 'What does E2E Normalized Interactivity mean?')
+      .and('have.attr', 'aria-label', 'What does E2E Normalized Interactivity mean?')
+      .and('have.class', 'no-export')
+      .find('svg[aria-hidden="true"]')
+      .should('be.visible')
+      .and('have.attr', 'width', '24');
+  });
+
+  it('links the Chinese control to the Chinese FAQ', () => {
+    cy.mount(
+      <div className="relative h-10 w-80">
+        <NormalizedInteractivityHelpLink locale="zh" />
+      </div>,
+    );
+
+    cy.get('[data-testid="normalized-interactivity-faq-link"]')
+      .should('have.attr', 'href', '/zh/about#faq-normalized-interactivity')
+      .and('have.attr', 'title', '什么是端到端归一化交互性？')
+      .and('have.attr', 'aria-label', '什么是端到端归一化交互性？');
+  });
+});
diff --git a/packages/app/cypress/component/quick-filters-dialog.cy.tsx b/packages/app/cypress/component/quick-filters-dialog.cy.tsx
new file mode 100644
index 000000000..86cf90525
--- /dev/null
+++ b/packages/app/cypress/component/quick-filters-dialog.cy.tsx
@@ -0,0 +1,61 @@
+import { QuickFiltersDialog } from '@/components/inference/ui/QuickFiltersDialog';
+import type { QuickFilters } from '@/components/inference/types';
+import { Sequence } from '@/lib/data-mappings';
+import { mountWithProviders } from '../support/test-utils';
+
+const availableQuickFilters: QuickFilters = {
+  vendors: ['NVIDIA', 'AMD'],
+  frameworks: ['vllm', 'sglang'],
+  deployment: ['single-node', 'multi-node', 'disagg'],
+  spec: ['mtp', 'stp'],
+};
+
+describe('QuickFiltersDialog', () => {
+  it('shows every filter group for fixed-sequence charts', () => {
+    mountWithProviders(<QuickFiltersDialog open onOpenChange={cy.stub()} />, {
+      inference: { availableQuickFilters },
+    });
+
+    cy.get('[data-testid="quick-filters-dialog"]').should('be.visible');
+    cy.get('[data-testid="quick-filter-vendor-NVIDIA"]').should('exist');
+    cy.get('[data-testid="quick-filter-framework-vllm"]').should('exist');
+    cy.get('[data-testid="quick-filter-deployment-disagg"]').should('exist');
+    cy.get('[data-testid="quick-filter-spec-mtp"]').should('exist');
+  });
+
+  it('removes speculative decoding from agentic charts and toggles supported filters', () => {
+    mountWithProviders(<QuickFiltersDialog open onOpenChange={cy.stub()} />, {
+      inference: {
+        selectedSequence: Sequence.AgenticTraces,
+        availableQuickFilters,
+      },
+    });
+
+    cy.get('[data-testid="quick-filter-vendor-NVIDIA"]').click();
+    cy.get('@setQuickFilterVendors').should('have.been.calledWith', ['NVIDIA']);
+    cy.get('[data-testid="quick-filter-framework-vllm"]').should('exist');
+    cy.get('[data-testid="quick-filter-deployment-disagg"]').should('exist');
+    cy.get('[data-testid^="quick-filter-spec-"]').should('not.exist');
+    cy.contains('Spec Decoding').should('not.exist');
+  });
+
+  it('clears every filter category from one action', () => {
+    mountWithProviders(<QuickFiltersDialog open onOpenChange={cy.stub()} />, {
+      inference: {
+        quickFilters: {
+          vendors: ['AMD'],
+          frameworks: ['sglang'],
+          deployment: ['disagg'],
+          spec: ['mtp'],
+        },
+        availableQuickFilters,
+      },
+    });
+
+    cy.contains('button', 'Clear filters').click();
+    cy.get('@setQuickFilterVendors').should('have.been.calledWith', []);
+    cy.get('@setQuickFilterFrameworks').should('have.been.calledWith', []);
+    cy.get('@setQuickFilterDeployment').should('have.been.calledWith', []);
+    cy.get('@setQuickFilterSpec').should('have.been.calledWith', []);
+  });
+});
diff --git a/packages/app/cypress/component/quotes-content.cy.tsx b/packages/app/cypress/component/quotes-content.cy.tsx
index d32d13c5d..34101ab8e 100644
--- a/packages/app/cypress/component/quotes-content.cy.tsx
+++ b/packages/app/cypress/component/quotes-content.cy.tsx
@@ -32,6 +32,26 @@ describe('QuotesContent', () => {
     cy.get('button[aria-label^="Jump to"]').should('have.length', uniqueOrgsWithLogo.length);
   });
 
+  it('renders the Mooncake team quote with the KVCache.AI identity', () => {
+    cy.get('#quote-mooncake').within(() => {
+      cy.contains('Mooncake makes KV cache a first-class system resource').should('be.visible');
+      cy.contains('a', 'Mooncake Team')
+        .should('have.attr', 'href', 'https://kvcache.ai/')
+        .and('have.attr', 'target', '_blank');
+      cy.get('img[alt="Mooncake"]')
+        .should('have.attr', 'src', '/logos/kvcache-ai.svg')
+        .and(($logo) => {
+          expect(($logo[0] as HTMLImageElement).naturalWidth).to.be.greaterThan(0);
+        });
+    });
+
+    cy.mount(<QuotesContent locale="zh" />);
+    cy.get('#quote-mooncake').should(
+      'contain.text',
+      'Mooncake 将 KV cache 作为分离式 LLM 推理中的一等系统资源',
+    );
+  });
+
   it('scrolls to the matching quote when a logo is clicked', () => {
     const targetOrg = uniqueOrgsWithLogo[0];
 
diff --git a/packages/app/cypress/component/scatter-graph.cy.tsx b/packages/app/cypress/component/scatter-graph.cy.tsx
index 3ee91849d..504c69178 100644
--- a/packages/app/cypress/component/scatter-graph.cy.tsx
+++ b/packages/app/cypress/component/scatter-graph.cy.tsx
@@ -1,16 +1,20 @@
-import { useState } from 'react';
-import { GlobalFilterContext } from '@/components/GlobalFilterContext';
-import { InferenceContext } from '@/components/inference/InferenceContext';
-import { UnofficialRunContext } from '@/components/unofficial-run-provider';
+import { useReducer, useState } from 'react';
+import { GlobalFilterSelectionContext } from '@/components/GlobalFilterContext';
+import { InferenceContextsProvider } from '@/components/inference/InferenceContext';
+import {
+  overlaySelectionReducer,
+  UnofficialRunContext,
+} from '@/components/unofficial-run-provider';
 import ScatterGraph from '@/components/inference/ui/ScatterGraph';
 import ChartDisplay from '@/components/inference/ui/ChartDisplay';
 import { mountWithProviders } from '../support/test-utils';
+import { expandLegendAdvanced } from '../support/legend-advanced';
 import {
   createMockInferenceData,
   createMockChartDefinition,
   createMockHardwareConfig,
-  createMockGlobalFilterContext,
-  createMockInferenceContext,
+  createMockGlobalFilterContexts,
+  createMockInferenceContextValues,
   createMockUnofficialRunContext,
 } from '../support/mock-data';
 import { Model, Precision, Sequence } from '@/lib/data-mappings';
@@ -138,7 +142,7 @@ describe('ScatterGraph', () => {
         }),
       ),
     );
-    const baseInference = createMockInferenceContext();
+    const baseInference = createMockInferenceContextValues();
 
     function OfficialScopeHarness() {
       const [secondScope, setSecondScope] = useState(false);
@@ -155,7 +159,12 @@ describe('ScatterGraph', () => {
       };
 
       return (
-        <InferenceContext.Provider value={inference}>
+        <InferenceContextsProvider
+          data={inference}
+          filters={inference}
+          display={inference}
+          actions={inference}
+        >
           <button data-testid="change-official-scope" onClick={() => setSecondScope(true)}>
             Change official scope
           </button>
@@ -169,7 +178,7 @@ describe('ScatterGraph', () => {
               chartDefinition={chartDefinition}
             />
           </div>
-        </InferenceContext.Provider>
+        </InferenceContextsProvider>
       );
     }
 
@@ -204,7 +213,7 @@ describe('ScatterGraph', () => {
       chartType: 'interactivity',
       y_tpPerGpu_roofline: 'upper_left',
     });
-    const baseInference = createMockInferenceContext();
+    const baseInference = createMockInferenceContextValues();
     const baseUnofficial = createMockUnofficialRunContext();
 
     function DelayedOfficialScopeHarness() {
@@ -256,15 +265,22 @@ describe('ScatterGraph', () => {
         ...baseUnofficial,
         isUnofficialRun: true,
         activeOverlayHwTypes: activeOverlayKeys,
-        setActiveOverlayHwTypes: setActiveOverlayKeys,
+        setUnifiedOverlaySelection: (official: Set<string>, overlay: Set<string>) => {
+          setOfficialOverride(official);
+          setActiveOverlayKeys(overlay);
+        },
         allOverlayHwTypes: new Set(['h100_vllm', 'b200_vllm']),
         localOfficialOverride: officialOverride,
-        setLocalOfficialOverride: setOfficialOverride,
       };
 
       return (
         <UnofficialRunContext.Provider value={unofficial}>
-          <InferenceContext.Provider value={inference}>
+          <InferenceContextsProvider
+            data={inference}
+            filters={inference}
+            display={inference}
+            actions={inference}
+          >
             <button data-testid="change-delayed-chart-scope" onClick={() => setSecondScope(true)}>
               Change scope
             </button>
@@ -291,7 +307,7 @@ describe('ScatterGraph', () => {
                 overlayData={overlayData}
               />
             </div>
-          </InferenceContext.Provider>
+          </InferenceContextsProvider>
         </UnofficialRunContext.Provider>
       );
     }
@@ -346,6 +362,9 @@ describe('ScatterGraph', () => {
       chartType: 'interactivity',
       y_tpPerGpu_roofline: 'upper_left',
     });
+    const runId = 32177976542;
+    const runBranch = 'qwen3.5-fp4-gb200-dynamo-sglang-agentic-mtp-pareto-refresh';
+    const runUrl = `https://github.com/SemiAnalysisAI/InferenceX/actions/runs/${runId}`;
     const officialData = [
       createMockInferenceData({ hwKey: 'h100', x: 8, y: 240, precision: Precision.FP4 }),
       createMockInferenceData({ hwKey: 'h100', x: 16, y: 200, precision: Precision.FP4 }),
@@ -358,26 +377,26 @@ describe('ScatterGraph', () => {
           x: 8,
           y: 320,
           precision: Precision.FP4,
-          run_url: 'https://github.com/x/y/actions/runs/12345',
+          run_url: runUrl,
         }),
         createMockInferenceData({
           hwKey: 'b200_trt',
           x: 16,
           y: 280,
           precision: Precision.FP4,
-          run_url: 'https://github.com/x/y/actions/runs/12345',
+          run_url: runUrl,
         }),
         createMockInferenceData({
           hwKey: 'b200_trt',
           x: 32,
           y: 220,
           precision: Precision.FP4,
-          run_url: 'https://github.com/x/y/actions/runs/12345',
+          run_url: runUrl,
         }),
       ],
       hardwareConfig: hwConfig,
-      label: 'feature-branch',
-      runUrl: 'https://github.com/x/y/actions/runs/12345',
+      label: runBranch,
+      runUrl,
     };
 
     mountWithProviders(
@@ -403,15 +422,15 @@ describe('ScatterGraph', () => {
         unofficial: {
           activeOverlayHwTypes: new Set(['b200_trt']),
           allOverlayHwTypes: new Set(['b200_trt']),
-          runIndexByUrl: { 'https://github.com/x/y/actions/runs/12345': 0, '12345': 0 },
+          runIndexByUrl: { [runUrl]: 0, [String(runId)]: 0 },
           unofficialRunInfos: [
             {
-              id: 12345,
+              id: runId,
               name: 'CI run',
-              branch: 'feature-branch',
-              sha: 'abc123',
-              createdAt: '2026-05-01T00:00:00Z',
-              url: 'https://github.com/x/y/actions/runs/12345',
+              branch: runBranch,
+              sha: '7a4a06b',
+              createdAt: '2026-08-18T19:40:51Z',
+              url: runUrl,
               conclusion: 'success',
               status: 'completed',
               isNonMainBranch: true,
@@ -435,11 +454,104 @@ describe('ScatterGraph', () => {
     cy.get('#test-scatter-overlay-labels svg .line-label')
       .filter('[data-line-key]:not([data-line-key^="overlay-"])')
       .should('have.length.greaterThan', 0);
-    // Overlay label text is the run's branch name (matching the overlay legend),
-    // not the hw label.
+    // The exact branch that crashed the production page remains visible in the
+    // overlay line label and legend after ScatterGraph's render-time updates.
     cy.get('#test-scatter-overlay-labels svg .line-label[data-line-key^="overlay-"]')
       .find('text')
-      .should('contain.text', 'feature-branch');
+      .should('contain.text', runBranch);
+    cy.get(
+      '#test-scatter-overlay-labels svg .line-label[data-line-key^="overlay-"] .ll-gpu',
+    ).should('not.exist');
+    cy.get('#test-scatter-overlay-labels [data-testid="chart-legend"]').should(
+      'contain.text',
+      runBranch,
+    );
+  });
+
+  it('places precision between the GPU and engine in multi-precision labels', () => {
+    const runUrl = 'https://github.com/x/y/actions/runs/precision-order';
+    const chartDefinition = createMockChartDefinition({
+      chartType: 'interactivity',
+      y_tpPerGpu_roofline: 'upper_left',
+    });
+    const data = [Precision.FP4, Precision.FP8].flatMap((precision, precisionIndex) =>
+      [8, 16].map((x, index) =>
+        createMockInferenceData({
+          hwKey: 'b200_trt',
+          x,
+          y: 320 - precisionIndex * 20 - index * 40,
+          precision,
+        }),
+      ),
+    );
+    const overlayData = {
+      data: [Precision.FP4, Precision.FP8].flatMap((precision, precisionIndex) =>
+        [8, 16].map((x, index) =>
+          createMockInferenceData({
+            hwKey: 'h100_vllm',
+            x,
+            y: 260 - precisionIndex * 20 - index * 40,
+            precision,
+            run_url: runUrl,
+          }),
+        ),
+      ),
+      hardwareConfig: hwConfig,
+      label: '',
+      runUrl,
+    };
+
+    mountWithProviders(
+      <div style={{ width: 800, height: 600 }}>
+        <ScatterGraph
+          chartId="test-scatter-precision-order"
+          modelLabel="DeepSeek R1"
+          data={data}
+          xLabel="Concurrency"
+          yLabel="Throughput / Chip (tok/s)"
+          chartDefinition={chartDefinition}
+          overlayData={overlayData}
+        />
+      </div>,
+      {
+        inference: {
+          hardwareConfig: hwConfig,
+          activeHwTypes: new Set(['b200_trt']),
+          hwTypesWithData: new Set(['b200_trt']),
+          selectedPrecisions: [Precision.FP4, Precision.FP8],
+          showLineLabels: true,
+        },
+        unofficial: {
+          activeOverlayHwTypes: new Set(['h100_vllm']),
+          allOverlayHwTypes: new Set(['h100_vllm']),
+          runIndexByUrl: { [runUrl]: 0, 'precision-order': 0 },
+          // No run metadata: exercise the hardware-label fallback path.
+          unofficialRunInfos: [],
+        },
+      },
+    );
+
+    cy.get('#test-scatter-precision-order svg .line-label[data-hw-key="b200_trt"] .ll-text')
+      .should('have.length', 2)
+      .then(($labels) => {
+        expect($labels.toArray().map((label) => label.textContent)).to.have.members([
+          'B200 FP4 (TRTLLM)',
+          'B200 FP8 (TRTLLM)',
+        ]);
+        for (const label of $labels) {
+          expect(
+            [...label.querySelectorAll('tspan')].map((segment) => segment.className.baseVal),
+          ).to.deep.equal(['ll-gpu', 'll-precision', 'll-engine']);
+        }
+      });
+    cy.get('#test-scatter-precision-order svg .line-label[data-line-key^="overlay-"] .ll-text')
+      .should('have.length', 2)
+      .then(($labels) => {
+        expect($labels.toArray().map((label) => label.textContent)).to.have.members([
+          'H100 FP4 (vLLM)',
+          'H100 FP8 (vLLM)',
+        ]);
+      });
   });
 
   it('renders a line label for a singleton unofficial overlay series', () => {
@@ -548,21 +660,25 @@ describe('ScatterGraph', () => {
       createMockInferenceData({ hwKey: 'h100', x: 300, y: 190, precision: Precision.FP8 }),
       createMockInferenceData({ hwKey: 'h100', x: 340, y: 150, precision: Precision.FP8 }),
     ];
-    const baseInference = createMockInferenceContext();
+    const baseInference = createMockInferenceContextValues();
 
     function IngestedSingletonLabelHarness() {
       const [showLineLabels, setShowLineLabels] = useState(true);
+      const inference = {
+        ...baseInference,
+        hardwareConfig: hwConfig,
+        activeHwTypes: new Set(['b200_tilert_mtp', 'h100']),
+        hwTypesWithData: new Set(['b200_tilert_mtp', 'h100']),
+        selectedPrecisions: [Precision.FP8],
+        showLineLabels,
+        setShowLineLabels,
+      };
       return (
-        <InferenceContext.Provider
-          value={{
-            ...baseInference,
-            hardwareConfig: hwConfig,
-            activeHwTypes: new Set(['b200_tilert_mtp', 'h100']),
-            hwTypesWithData: new Set(['b200_tilert_mtp', 'h100']),
-            selectedPrecisions: [Precision.FP8],
-            showLineLabels,
-            setShowLineLabels,
-          }}
+        <InferenceContextsProvider
+          data={inference}
+          filters={inference}
+          display={inference}
+          actions={inference}
         >
           <div style={{ width: 800, height: 600 }}>
             <ScatterGraph
@@ -574,7 +690,7 @@ describe('ScatterGraph', () => {
               chartDefinition={interactivityChartDef}
             />
           </div>
-        </InferenceContext.Provider>
+        </InferenceContextsProvider>
       );
     }
 
@@ -585,7 +701,18 @@ describe('ScatterGraph', () => {
       .should('have.css', 'opacity', '1')
       .find('text')
       .should('have.text', 'B200 (TileRT, MTP)');
+    cy.get(
+      '#test-scatter-ingested-singleton-label svg .line-label[data-hw-key="b200_tilert_mtp"] .ll-gpu',
+    )
+      .should('have.text', 'B200')
+      .and('have.attr', 'font-weight', '700');
+    cy.get(
+      '#test-scatter-ingested-singleton-label svg .line-label[data-hw-key="b200_tilert_mtp"] .ll-engine',
+    )
+      .should('have.text', ' (TileRT, MTP)')
+      .and('have.attr', 'fill', '#d1d5db');
 
+    expandLegendAdvanced();
     cy.get('#scatter-line-labels').click();
     cy.get('#test-scatter-ingested-singleton-label svg .line-label').should('not.exist');
     cy.get('#scatter-line-labels').click();
@@ -676,6 +803,12 @@ describe('ScatterGraph', () => {
     cy.get('#test-scatter-m3-eagle svg .line-label[data-line-key^="overlay-"]')
       .find('text')
       .should('contain.text', 'EAGLE');
+    cy.get('#test-scatter-m3-eagle svg .line-label[data-line-key^="overlay-"] .ll-gpu')
+      .should('have.text', 'B200')
+      .and('have.attr', 'font-weight', '700');
+    cy.get('#test-scatter-m3-eagle svg .line-label[data-line-key^="overlay-"] .ll-engine')
+      .should('contain.text', 'EAGLE')
+      .and('have.attr', 'fill', '#d1d5db');
     // No label should show the generic MTP token for M3.
     cy.get('#test-scatter-m3-eagle svg .line-label text').should('not.contain.text', 'MTP');
   });
@@ -767,15 +900,84 @@ describe('ScatterGraph', () => {
     ).should('have.css', 'opacity', '1');
     // The exclusion resolver must be bypassed rather than resolving in favor of
     // either the official or overlay engine family.
-    cy.get('@setActiveOverlayHwTypes').should('not.have.been.called');
+    cy.get('@setUnifiedOverlaySelection').should('not.have.been.called');
 
     // An additional official engine can also be selected while the preview is
     // loaded; the production-only conflict toggle must not be consulted.
     cy.get('label[for="checkbox-h100_vllm"]').click();
     cy.get('@blockedComparisonToggle').should('not.have.been.called');
-    cy.get('@setLocalOfficialOverride').should((setOverride) => {
-      const selection = setOverride.lastCall.args[0] as Set<string>;
-      expect([...selection]).to.have.members(['b200_sglang', 'h100_vllm']);
+    cy.get('@setUnifiedOverlaySelection').should((setSelection) => {
+      const official = setSelection.lastCall.args[0] as Set<string>;
+      expect([...official]).to.have.members(['b200_sglang', 'h100_vllm']);
+    });
+  });
+
+  it('keeps the unofficial overlay active when soloing an official hardware series', () => {
+    const chartDefinition = createMockChartDefinition({
+      chartType: 'interactivity',
+      y_tpPerGpu_roofline: 'upper_left',
+    });
+    const officialData = ['b200_sglang', 'h100_vllm'].flatMap((hwKey, hwIndex) =>
+      [8, 16, 32].map((x, index) =>
+        createMockInferenceData({
+          hwKey,
+          x,
+          y: 320 - hwIndex * 20 - index * 40,
+          precision: Precision.FP4,
+        }),
+      ),
+    );
+    const runUrl = 'https://github.com/x/y/actions/runs/official-solo';
+    const overlayData = {
+      data: [8, 16, 32].map((x, index) =>
+        createMockInferenceData({
+          hwKey: 'h100_vllm',
+          x,
+          y: 260 - index * 40,
+          precision: Precision.FP4,
+          run_url: runUrl,
+        }),
+      ),
+      hardwareConfig: hwConfig,
+      label: 'official-solo',
+      runUrl,
+    };
+
+    mountWithProviders(
+      <div style={{ width: 800, height: 600 }}>
+        <ScatterGraph
+          chartId="test-scatter-official-solo"
+          modelLabel="DeepSeek V4 Pro"
+          data={officialData}
+          xLabel="Concurrency"
+          yLabel="Throughput / Chip (tok/s)"
+          chartDefinition={chartDefinition}
+          overlayData={overlayData}
+        />
+      </div>,
+      {
+        inference: {
+          hardwareConfig: hwConfig,
+          activeHwTypes: new Set(['b200_sglang', 'h100_vllm']),
+          hwTypesWithData: new Set(['b200_sglang', 'h100_vllm']),
+          selectedModel: Model.DeepSeek_V4_Pro,
+          selectedSequence: Sequence.AgenticTraces,
+          selectedPrecisions: [Precision.FP4],
+        },
+        unofficial: {
+          activeOverlayHwTypes: new Set(['h100_vllm']),
+          allOverlayHwTypes: new Set(['h100_vllm']),
+        },
+      },
+    );
+
+    cy.get('#test-scatter-official-solo svg .overlay-roofline-path').should('exist');
+    cy.get('label[for="checkbox-h100_vllm"]').click();
+    cy.get('@setUnifiedOverlaySelection').should((setSelection) => {
+      const official = setSelection.lastCall.args[0] as Set<string>;
+      const overlay = setSelection.lastCall.args[1] as Set<string>;
+      expect([...official]).to.deep.equal(['h100_vllm']);
+      expect([...overlay]).to.deep.equal(['h100_vllm']);
     });
   });
 
@@ -938,7 +1140,7 @@ describe('ScatterGraph', () => {
 });
 
 describe('ChartDisplay engine comparison guard', () => {
-  it('includes cost-clipped official and unofficial points in table mode', () => {
+  it('includes explicitly clipped official and unofficial points in table mode', () => {
     const chartDefinition = createMockChartDefinition({
       chartType: 'interactivity',
       x: 'median_intvty',
@@ -1044,8 +1246,8 @@ describe('ChartDisplay engine comparison guard', () => {
 
   it('keeps official table rows synchronized with legend state after a scope change', () => {
     const chartDefinition = createMockChartDefinition({ chartType: 'interactivity' });
-    const baseInference = createMockInferenceContext();
-    const baseGlobalFilters = createMockGlobalFilterContext();
+    const baseInference = createMockInferenceContextValues();
+    const baseGlobalFilters = createMockGlobalFilterContexts().selection;
 
     function OfficialRowsScopeHarness() {
       const [secondScope, setSecondScope] = useState(false);
@@ -1082,8 +1284,13 @@ describe('ChartDisplay engine comparison guard', () => {
       };
 
       return (
-        <GlobalFilterContext.Provider value={globalFilters}>
-          <InferenceContext.Provider value={inference}>
+        <GlobalFilterSelectionContext.Provider value={globalFilters}>
+          <InferenceContextsProvider
+            data={inference}
+            filters={inference}
+            display={inference}
+            actions={inference}
+          >
             <button data-testid="change-official-table-scope" onClick={() => setSecondScope(true)}>
               Change scope
             </button>
@@ -1094,8 +1301,8 @@ describe('ChartDisplay engine comparison guard', () => {
               Select vLLM
             </button>
             <ChartDisplay />
-          </InferenceContext.Provider>
-        </GlobalFilterContext.Provider>
+          </InferenceContextsProvider>
+        </GlobalFilterSelectionContext.Provider>
       );
     }
 
@@ -1111,7 +1318,62 @@ describe('ChartDisplay engine comparison guard', () => {
     cy.get('[data-testid="inference-results-table"] tbody tr').should('have.length', 1);
     cy.get('[data-testid="inference-results-table"] tbody').contains('vLLM').should('exist');
     cy.get('[data-testid="inference-results-table"] tbody').contains('SGLang').should('not.exist');
-    cy.get('@setLocalOfficialOverride').should('not.have.been.called');
+    cy.get('@setUnifiedOverlaySelection').should('not.have.been.called');
+  });
+
+  it('renders the table columns without the median interactivity or TTFT columns', () => {
+    // Mirror the real interactivity chart: x IS interactivity, which is what
+    // made the separate median column a duplicate.
+    const chartDefinition = createMockChartDefinition({
+      chartType: 'interactivity',
+      x: 'median_intvty',
+      x_label: 'Interactivity (tok/s/user)',
+    });
+    const row = createMockInferenceData({
+      hwKey: 'b200_sglang',
+      hw: 'Official SGLang',
+      model: Model.DeepSeek_V4_Pro,
+      precision: Precision.FP4,
+    });
+
+    mountWithProviders(<ChartDisplay />, {
+      inference: {
+        graphs: [
+          {
+            model: Model.DeepSeek_V4_Pro,
+            sequence: Sequence.AgenticTraces,
+            chartDefinition,
+            data: [row],
+          },
+        ],
+        selectedModel: Model.DeepSeek_V4_Pro,
+        selectedSequence: Sequence.AgenticTraces,
+        selectedXAxisMode: 'interactivity',
+        activeHwTypes: new Set(['b200_sglang']),
+        hwTypesWithData: new Set(['b200_sglang']),
+      },
+      globalFilters: {
+        selectedModel: Model.DeepSeek_V4_Pro,
+        selectedSequence: Sequence.AgenticTraces,
+        effectiveSequence: Sequence.AgenticTraces,
+      },
+      unofficial: {},
+    });
+
+    cy.get('[data-testid="inference-table-view-btn"]').click();
+    cy.get('[data-testid="inference-results-table"] thead th').then(($headers) => {
+      const headers = [...$headers].map((th) => (th.textContent ?? '').trim());
+      // Interactivity is already the x-axis column on the interactivity chart,
+      // so the median column duplicated it. Assert both that the column is gone
+      // and that exactly one interactivity column remains, so a rename cannot
+      // quietly reintroduce the duplicate.
+      expect(headers).to.not.include('Median Interactivity (tok/s)');
+      expect(headers.filter((h) => h.toLowerCase().includes('interactivity'))).to.have.length(1);
+      // Median TTFT was dropped too; unlike interactivity it is not duplicated
+      // by the x-axis column, so nothing else in the table should carry it.
+      expect(headers).to.not.include('Median TTFT (ms)');
+      expect(headers.filter((h) => h.toLowerCase().includes('ttft'))).to.have.length(0);
+    });
   });
 
   it('keeps same-hardware cross-engine AgentX STP rows out of table mode', () => {
@@ -1249,8 +1511,10 @@ describe('ChartDisplay engine comparison guard', () => {
     cy.get('[data-testid="inference-results-table"] tbody tr').should('have.length', 2);
     cy.get('[data-testid="inference-results-table"] tbody').contains('vLLM').should('exist');
     cy.get('[data-testid="inference-results-table"] tbody').contains('SGLang').should('exist');
-    // The reconciliation effect must not strip the run's hw types from the provider.
-    cy.get('@setActiveOverlayHwTypes').should('not.have.been.called');
+    cy.get('@reconcileOverlayScope').should((reconcile) => {
+      const scope = reconcile.lastCall.args[0] as { overlayHwTypes: Set<string> };
+      expect([...scope.overlayHwTypes]).to.have.members(['h100_vllm']);
+    });
   });
 
   it('keeps an explicitly empty official legend out of table mode', () => {
@@ -1320,15 +1584,24 @@ describe('ChartDisplay engine comparison guard', () => {
       status: 'completed',
       isNonMainBranch: true,
     };
-    const baseInference = createMockInferenceContext();
-    const baseGlobalFilters = createMockGlobalFilterContext();
+    const baseInference = createMockInferenceContextValues();
+    const baseGlobalFilters = createMockGlobalFilterContexts().selection;
     const baseUnofficial = createMockUnofficialRunContext();
 
     function OverlayScopeHarness() {
       const [secondScope, setSecondScope] = useState(false);
       const [secondScopeLoaded, setSecondScopeLoaded] = useState(false);
-      const [activeOverlayKeys, setActiveOverlayKeys] = useState(new Set(['h100_sglang']));
-      const [officialOverride, setOfficialOverride] = useState<Set<string> | null>(null);
+      const [selection, dispatchSelection] = useReducer(overlaySelectionReducer, {
+        availabilityKey: String(runInfo.id),
+        activeOverlayHwTypes: new Set(['h100_sglang']),
+        availableOverlayHwTypes: new Set(['h100_sglang', 'h200_sglang', 'b200_vllm']),
+        localOfficialOverride: null,
+        scopeKey: `${Model.DeepSeek_V4_Pro}|${Sequence.AgenticTraces}|${Precision.FP4}|${runInfo.url}|official:b200_sglang|overlay:h100_sglang,h200_sglang`,
+        scopeOverlayHwTypes: new Set(['h100_sglang', 'h200_sglang']),
+        scopeReady: true,
+        bestSelectionKey: '',
+        bestPerSku: false,
+      });
       const [, setRenderVersion] = useState(0);
       const model = secondScope ? Model.DeepSeek_R1 : Model.DeepSeek_V4_Pro;
       const officialKeys = secondScope ? ['h100_vllm', 'b200_sglang'] : ['b200_sglang'];
@@ -1387,17 +1660,25 @@ describe('ChartDisplay engine comparison guard', () => {
         unofficialRunInfos: [runInfo],
         runIndexByUrl: { [runInfo.url]: 0, [String(runInfo.id)]: 0 },
         getOverlayData: () => ({ data: overlayRows, hardwareConfig: hwConfig }),
-        activeOverlayHwTypes: activeOverlayKeys,
-        setActiveOverlayHwTypes: setActiveOverlayKeys,
-        allOverlayHwTypes: new Set(['h100_sglang', 'h200_sglang', 'b200_vllm']),
-        localOfficialOverride: officialOverride,
-        setLocalOfficialOverride: setOfficialOverride,
+        activeOverlayHwTypes: selection.activeOverlayHwTypes,
+        reconcileOverlayScope: (
+          input: Parameters<typeof baseUnofficial.reconcileOverlayScope>[0],
+        ) => dispatchSelection({ type: 'scope', input }),
+        setUnifiedOverlaySelection: (official: Set<string>, overlay: Set<string>) =>
+          dispatchSelection({ type: 'selection', official, overlay }),
+        allOverlayHwTypes: selection.availableOverlayHwTypes,
+        localOfficialOverride: selection.localOfficialOverride,
       };
 
       return (
-        <GlobalFilterContext.Provider value={globalFilters}>
+        <GlobalFilterSelectionContext.Provider value={globalFilters}>
           <UnofficialRunContext.Provider value={unofficial}>
-            <InferenceContext.Provider value={inference}>
+            <InferenceContextsProvider
+              data={inference}
+              filters={inference}
+              display={inference}
+              actions={inference}
+            >
               <button data-testid="change-overlay-scope" onClick={() => setSecondScope(true)}>
                 Change scope
               </button>
@@ -1406,7 +1687,13 @@ describe('ChartDisplay engine comparison guard', () => {
               </button>
               <button
                 data-testid="clear-overlay-scope"
-                onClick={() => setActiveOverlayKeys(new Set())}
+                onClick={() =>
+                  dispatchSelection({
+                    type: 'selection',
+                    official: selection.localOfficialOverride ?? new Set(officialKeys),
+                    overlay: new Set(),
+                  })
+                }
               >
                 Clear overlays
               </button>
@@ -1417,9 +1704,9 @@ describe('ChartDisplay engine comparison guard', () => {
                 Rerender
               </button>
               <ChartDisplay />
-            </InferenceContext.Provider>
+            </InferenceContextsProvider>
           </UnofficialRunContext.Provider>
-        </GlobalFilterContext.Provider>
+        </GlobalFilterSelectionContext.Provider>
       );
     }
 
diff --git a/packages/app/cypress/component/scatter-label-overlap.cy.tsx b/packages/app/cypress/component/scatter-label-overlap.cy.tsx
new file mode 100644
index 000000000..6740152fa
--- /dev/null
+++ b/packages/app/cypress/component/scatter-label-overlap.cy.tsx
@@ -0,0 +1,152 @@
+import ScatterGraph from '@/components/inference/ui/ScatterGraph';
+import { Precision } from '@/lib/data-mappings';
+
+import {
+  createMockChartDefinition,
+  createMockHardwareConfig,
+  createMockInferenceData,
+} from '../support/mock-data';
+import { mountWithProviders } from '../support/test-utils';
+
+const hwConfig = createMockHardwareConfig();
+
+const CHART_ID = 'test-scatter-label-overlap';
+
+interface Box {
+  left: number;
+  right: number;
+  top: number;
+  bottom: number;
+}
+
+/** Overlap depth of two boxes along each axis; negative means no overlap. */
+function overlap(a: Box, b: Box): { x: number; y: number } {
+  return {
+    x: Math.min(a.right, b.right) - Math.max(a.left, b.left),
+    y: Math.min(a.bottom, b.bottom) - Math.max(a.top, b.top),
+  };
+}
+
+/**
+ * Both axes must exceed this for a collision to count. The placement engine
+ * models a text block with fixed ascent/descent constants, so a glyph's real
+ * ink box can stick out by about a pixel without anything looking wrong.
+ */
+const TOLERANCE_PX = 1;
+
+/** A curve dense enough that its run-name pill has point labels to land on. */
+function curve(hwKey: string, ys: number[]) {
+  return ys.map((y, i) =>
+    createMockInferenceData({
+      hwKey,
+      x: 2 ** (i + 3),
+      y,
+      conc: 2 ** (i + 3),
+      precision: Precision.FP4,
+      // Widens the label in advanced mode (a disagg prefill/decode split),
+      // which is the shape that overlapped in practice.
+      disagg: true,
+      dp_attention: true,
+      prefill_tp: 8,
+      prefill_ep: 8,
+      prefill_dp_attention: true,
+      prefill_num_workers: 2,
+      decode_tp: 16,
+      decode_ep: 16,
+      decode_dp_attention: true,
+      decode_num_workers: 1,
+    }),
+  );
+}
+
+function mountChart(useAdvancedLabels: boolean) {
+  mountWithProviders(
+    <div style={{ width: 900, height: 600 }}>
+      <ScatterGraph
+        chartId={CHART_ID}
+        modelLabel="DeepSeek R1"
+        data={[
+          ...curve('h100', [240, 232, 210, 180, 150, 118]),
+          ...curve('b200_trt', [330, 315, 288, 250, 205, 160]),
+        ]}
+        xLabel="Concurrency"
+        yLabel="Throughput / Chip (tok/s)"
+        chartDefinition={createMockChartDefinition({
+          chartType: 'interactivity',
+          y_tpPerGpu_roofline: 'upper_left',
+        })}
+      />
+    </div>,
+    {
+      inference: {
+        hardwareConfig: hwConfig,
+        activeHwTypes: new Set(['h100', 'b200_trt']),
+        hwTypesWithData: new Set(['h100', 'b200_trt']),
+        selectedPrecisions: [Precision.FP4],
+        showLineLabels: true,
+        showPointLabels: true,
+        useAdvancedLabels,
+      },
+      unofficial: {},
+    },
+  );
+}
+
+/** Asserts no visible point label is drawn through a visible pill. */
+function expectNoPillOverlap() {
+  cy.get(`#${CHART_ID} svg .ll-bg`).should('exist');
+
+  cy.get(`#${CHART_ID} svg`).then(($svg) => {
+    const svg = $svg[0];
+
+    const pills = [...svg.querySelectorAll<SVGGElement>('.line-label, .parallelism-label')]
+      .filter((group) => group.style.opacity !== '0')
+      .map((group) => ({
+        label: group.textContent ?? '',
+        box: (group.querySelector('.ll-bg, .pl-bg') as SVGRectElement).getBoundingClientRect(),
+      }));
+
+    const points = [...svg.querySelectorAll<SVGTextElement>('.point-label')]
+      .filter((text) => {
+        const group = text.closest<SVGGElement>('.dot-group');
+        return text.style.opacity !== '0' && group?.style.opacity !== '0';
+      })
+      .map((text) => ({ label: text.textContent ?? '', box: text.getBoundingClientRect() }));
+
+    // Without these the assertion below would pass on an empty chart.
+    expect(pills.length, 'visible run-name pills').to.be.greaterThan(0);
+    expect(points.length, 'visible point labels').to.be.greaterThan(3);
+
+    const collisions: string[] = [];
+    for (const pill of pills) {
+      for (const point of points) {
+        const { x, y } = overlap(pill.box, point.box);
+        if (x > TOLERANCE_PX && y > TOLERANCE_PX) {
+          collisions.push(
+            `"${point.label}" overlaps pill "${pill.label}" by ${x.toFixed(1)}x${y.toFixed(1)}px`,
+          );
+        }
+      }
+    }
+
+    expect(collisions, collisions.join('; ')).to.have.length(0);
+  });
+}
+
+describe('Scatter label overlap', () => {
+  it('draws both label kinds (guards the overlap assertions below)', () => {
+    mountChart(false);
+    cy.get(`#${CHART_ID} svg .line-label`).should('have.length.greaterThan', 0);
+    cy.get(`#${CHART_ID} svg .point-label`).should('have.length.greaterThan', 3);
+  });
+
+  it('keeps point labels clear of the run-name pills', () => {
+    mountChart(false);
+    expectNoPillOverlap();
+  });
+
+  it('keeps the wider parallelism labels clear of the run-name pills', () => {
+    mountChart(true);
+    expectNoPillOverlap();
+  });
+});
diff --git a/packages/app/cypress/component/tab-nav.cy.tsx b/packages/app/cypress/component/tab-nav.cy.tsx
index 9d29cd7b8..24f86098f 100644
--- a/packages/app/cypress/component/tab-nav.cy.tsx
+++ b/packages/app/cypress/component/tab-nav.cy.tsx
@@ -4,17 +4,7 @@ import { PathnameContext } from 'next/dist/shared/lib/hooks-client-context.share
 import { TabNav } from '@/components/tab-nav';
 import { UnofficialRunContext } from '@/components/unofficial-run-provider';
 import { createMockUnofficialRunContext } from '../support/mock-data';
-
-function createMockRouter() {
-  return {
-    push: cy.stub(),
-    replace: cy.stub(),
-    refresh: cy.stub(),
-    back: cy.stub(),
-    forward: cy.stub(),
-    prefetch: cy.stub().resolves(),
-  };
-}
+import { createMockRouter } from '../support/mock-router';
 
 /**
  * Mount TabNav with the provided URL search string written into the window
diff --git a/packages/app/cypress/e2e/agentic-catalog.cy.ts b/packages/app/cypress/e2e/agentic-catalog.cy.ts
new file mode 100644
index 000000000..1ff5ada12
--- /dev/null
+++ b/packages/app/cypress/e2e/agentic-catalog.cy.ts
@@ -0,0 +1,74 @@
+import { unlockAgenticGate } from '../support/e2e';
+
+/**
+ * `/inference/agentic` is the telemetry catalog: a hero drawn from the AgentX
+ * telemetry tutorial, then one card per (model, SKU, engine, precision) config
+ * that has stored per-request telemetry.
+ */
+describe('AgentX telemetry catalog', () => {
+  beforeEach(() => {
+    cy.window().then((win) => {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+    });
+  });
+
+  it('leads with the telemetry hero and links to the full tutorial', () => {
+    cy.visit('/inference/agentic', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentic-catalog-hero"]').within(() => {
+      cy.get('h1').should('have.text', 'Exploring Agentic Workloads: Detailed Telemetry');
+      // Highlight counts come from the tutorial guide, which agentx-telemetry.test.ts
+      // pins to what the detail page actually renders.
+      cy.get('dl div').should('have.length', 4);
+      cy.contains('11').should('be.visible');
+      cy.get('[data-testid="agentic-catalog-tutorial-link"]')
+        .should('contain.text', 'Read the telemetry tutorial')
+        .and('have.attr', 'href', '/agentx/telemetry');
+    });
+  });
+
+  it('lists configurations grouped by model, each opening a point-detail page', () => {
+    cy.visit('/inference/agentic', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentic-catalog-summary"]')
+      .should('contain.text', 'AgentX runs with stored telemetry')
+      .and('contain.text', 'serving configurations across');
+
+    cy.get('[data-testid^="agentic-catalog-model-"]').should('have.length.greaterThan', 0);
+    cy.get('[data-testid^="agentic-catalog-card-"]')
+      .should('have.length.greaterThan', 0)
+      .first()
+      .should('have.attr', 'href')
+      .and('match', /^\/inference\/agentic\/\d+$/u);
+  });
+
+  it('reaches the catalog from the footer and highlights Dashboard in the header nav', () => {
+    cy.visit('/inference/agentic', { onBeforeLoad: unlockAgenticGate });
+
+    // Telemetry is a footer destination now, so the catalog sits under the
+    // Dashboard nav entry rather than owning one of its own.
+    cy.get('[data-testid="nav-link-telemetry"]').should('not.exist');
+    cy.get('[data-testid="nav-link-dashboard"]').should('have.class', 'text-brand');
+    cy.get('[data-testid="footer-link-telemetry"]').should(
+      'have.attr',
+      'href',
+      '/inference/agentic',
+    );
+  });
+
+  it('ships the Simplified Chinese catalog', () => {
+    cy.visit('/zh/inference/agentic', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentic-catalog-summary"]').should(
+      'contain.text',
+      '已存储遥测数据的 AgentX 运行',
+    );
+    cy.get('[data-testid="agentic-catalog-tutorial-link"]')
+      .should('contain.text', '阅读遥测数据教程')
+      .and('have.attr', 'href', '/zh/agentx/telemetry');
+    cy.get('[data-testid^="agentic-catalog-card-"]')
+      .first()
+      .should('have.attr', 'href')
+      .and('match', /^\/zh\/inference\/agentic\/\d+$/u);
+  });
+});
diff --git a/packages/app/cypress/e2e/agentic-detail-back-nav.cy.ts b/packages/app/cypress/e2e/agentic-detail-back-nav.cy.ts
new file mode 100644
index 000000000..9ce372500
--- /dev/null
+++ b/packages/app/cypress/e2e/agentic-detail-back-nav.cy.ts
@@ -0,0 +1,322 @@
+import { interceptDerivedAgenticMetrics } from '../support/e2e';
+
+// ---------------------------------------------------------------------------
+// Regression: returning from an agentic point-detail page reset the chart.
+//
+// Chart filters only ever reach an in-memory store — `url-state.ts` strips the
+// share params from the address bar after load — and the point-detail links are
+// plain `<a href>` (they must support open-in-new-tab), so following one is a
+// full-document navigation that destroys that store. Back therefore returned to
+// a bare `/inference` and the chart rebuilt from defaults: DeepSeek V4 Pro with
+// every legend series on, no matter what the reader had been looking at.
+//
+// The shared fixtures carry ZERO agentic_traces rows, so this spec injects its
+// own agentic coverage (same pattern as gpu-compare-agentic-detail.cy.ts).
+// ---------------------------------------------------------------------------
+
+const AGENTIC_DATE = '2026-06-12';
+const DEFAULT_MODEL_DB_KEY = 'dsv4'; // DeepSeek-V4-Pro, the dashboard default
+const TARGET_MODEL_DB_KEY = 'kimik3'; // Kimi-K3, the non-default the reader picks
+const TARGET_MODEL_LABEL = 'Kimi K3';
+
+const AGENTIC_HARDWARE = [
+  { hardware: 'b200', framework: 'vllm' },
+  { hardware: 'b300', framework: 'vllm' },
+  { hardware: 'h200', framework: 'vllm' },
+];
+
+/** The series the reader solos — must differ from the "all on" default. */
+const SOLO_HW_KEY = 'b300_vllm';
+const SOLO_HW_LABEL = 'B300 (vLLM)';
+
+const availabilityFor = (model: string) => [
+  ...AGENTIC_HARDWARE.map((g) => ({
+    model,
+    isl: null,
+    osl: null,
+    precision: 'fp4',
+    hardware: g.hardware,
+    framework: g.framework,
+    spec_method: 'none',
+    disagg: false,
+    benchmark_type: 'agentic_traces',
+    date: AGENTIC_DATE,
+  })),
+  // Fixed-seq rows alongside, so the scenario selector sees the "both exist"
+  // signal it needs to settle confidently on agentic.
+  ...AGENTIC_HARDWARE.map((g) => ({
+    model,
+    isl: 8192,
+    osl: 1024,
+    precision: 'fp4',
+    hardware: g.hardware,
+    framework: g.framework,
+    spec_method: 'none',
+    disagg: false,
+    benchmark_type: 'single_turn',
+    date: AGENTIC_DATE,
+  })),
+];
+
+const availability = [
+  ...availabilityFor(DEFAULT_MODEL_DB_KEY),
+  ...availabilityFor(TARGET_MODEL_DB_KEY),
+];
+
+const percentileLadder = (prefix: string, base: number): Record<string, number> => ({
+  [`median_${prefix}`]: base,
+  [`p75_${prefix}`]: base * 1.2,
+  [`p90_${prefix}`]: base * 1.5,
+  [`p95_${prefix}`]: base * 1.7,
+  [`p99_${prefix}`]: base * 2.2,
+  [`std_${prefix}`]: base * 0.3,
+});
+
+const agenticMetrics = (conc: number): Record<string, number> => {
+  const scale = conc / 16;
+  const itl = 0.011 * scale;
+  return {
+    ...percentileLadder('ttft', 0.4 * scale),
+    ...percentileLadder('tpot', 0.012 * scale),
+    ...percentileLadder('itl', itl),
+    ...percentileLadder('e2el', 8 * scale),
+    median_intvty: 1 / itl,
+    p75_intvty: 1 / (itl * 1.2),
+    p90_intvty: 1 / (itl * 1.5),
+    p99_intvty: 1 / (itl * 2.2),
+    std_intvty: (1 / itl) * 0.1,
+    tput_per_gpu: 950 / Math.sqrt(scale),
+    output_tput_per_gpu: 210,
+    input_tput_per_gpu: 740,
+    total_tput_tps: 7600 * conc * 0.05,
+  };
+};
+
+let benchIdCursor = 900100;
+const benchmarksFor = (model: string) =>
+  AGENTIC_HARDWARE.flatMap((g) =>
+    [16, 64, 128].map((conc) => ({
+      id: benchIdCursor++,
+      hardware: g.hardware,
+      framework: g.framework,
+      model,
+      precision: 'fp4',
+      spec_method: 'none',
+      disagg: false,
+      is_multinode: false,
+      prefill_tp: 8,
+      prefill_ep: 1,
+      prefill_dp_attention: false,
+      prefill_num_workers: 0,
+      decode_tp: 8,
+      decode_ep: 1,
+      decode_dp_attention: false,
+      decode_num_workers: 0,
+      num_prefill_gpu: 8,
+      num_decode_gpu: 8,
+      isl: null,
+      osl: null,
+      conc,
+      offload_mode: 'off',
+      benchmark_type: 'agentic_traces',
+      image: 'vllm/vllm-openai:v0.9.0',
+      metrics: agenticMetrics(conc),
+      workers: null,
+      date: AGENTIC_DATE,
+      run_url: null,
+    })),
+  );
+
+const benchmarks = [...benchmarksFor(DEFAULT_MODEL_DB_KEY), ...benchmarksFor(TARGET_MODEL_DB_KEY)];
+const tracedIds = new Set(benchmarks.map((b) => b.id));
+
+/** Minimal detail-page payloads — this spec only cares about its nav controls. */
+function interceptDetailPage(): void {
+  cy.intercept('GET', '/api/v1/trace-server-metrics*', { body: null });
+  cy.intercept('GET', '/api/v1/trace-histograms*', { body: {} });
+  cy.intercept('GET', '/api/v1/request-timeline*', { body: null });
+  cy.intercept('GET', '/api/v1/benchmark-siblings*', { body: null });
+  cy.intercept('GET', '/api/v1/agentic-aggregates*', { body: null });
+}
+
+function interceptChart(): void {
+  cy.intercept('GET', '/api/v1/availability', { body: availability }).as('availability');
+  cy.intercept('GET', '/api/v1/benchmarks*', { body: benchmarks }).as('benchmarks');
+  cy.intercept('GET', '/api/v1/trace-availability*', (request) => {
+    const ids = new URL(request.url).searchParams.get('ids')?.split(',') ?? [];
+    request.reply({
+      body: Object.fromEntries(
+        ids.filter((id) => tracedIds.has(Number(id))).map((id) => [id, true]),
+      ),
+    });
+  });
+  interceptDerivedAgenticMetrics();
+  interceptDetailPage();
+}
+
+function visitChart(path: string): void {
+  interceptChart();
+  cy.visit(path, {
+    onBeforeLoad(win) {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+    },
+  });
+  cy.get('[data-testid="inference-chart-display"]').should('exist');
+}
+
+/** Pick the non-default model through the dropdown, exactly as the report does. */
+function selectTargetModel(): void {
+  cy.get('[data-testid="model-selector"]').should('contain.text', 'DeepSeek V4 Pro');
+  cy.get('[data-testid="model-selector"]').click();
+  cy.contains('[role="option"]', TARGET_MODEL_LABEL).click();
+  cy.get('[data-testid="model-selector"]').should('contain.text', TARGET_MODEL_LABEL);
+}
+
+/** Solo one legend series so the restored state is distinguishable from "all on". */
+function soloLegendSeries(): void {
+  cy.get('[data-testid="chart-legend"] ul input[type="checkbox"]:checked').should(
+    'have.length.greaterThan',
+    1,
+  );
+  cy.contains('[data-testid="chart-legend"] label', SOLO_HW_LABEL).click();
+  cy.get('[data-testid="chart-legend"] ul input[type="checkbox"]:checked').should('have.length', 1);
+}
+
+/** Pin the tooltip on a visible point of the soloed series and open its detail page. */
+function openPointDetail(): void {
+  cy.get(`[data-testid="legend-points-${SOLO_HW_KEY}"]`).should('exist');
+  cy.get('[data-testid="inference-chart-display"] svg .dot-group')
+    .should('have.length.greaterThan', 0)
+    .then(($dots) => {
+      const target = [...$dots].find((node) => {
+        const datum = (node as unknown as { __data__?: { hwKey?: string; id?: number } }).__data__;
+        return datum?.hwKey === SOLO_HW_KEY && tracedIds.has(Number(datum?.id));
+      });
+      expect(target, `a traced ${SOLO_HW_KEY} point`).to.not.equal(undefined);
+      cy.wrap(target).find('.visible-shape').click({ force: true });
+    });
+  cy.get('[data-chart-tooltip]:visible [data-action="view-charts"]').should('be.visible').click();
+}
+
+/**
+ * The detail page's two nav controls, addressed by their visible label so this
+ * spec fails on the restoration assertion rather than on a missing hook when
+ * run against a build without the fix.
+ */
+const backControl = (locale: 'en' | 'zh') =>
+  cy.contains('button', locale === 'zh' ? '返回' : 'Back');
+const chartLink = (locale: 'en' | 'zh') =>
+  cy.contains('a', locale === 'zh' ? '推理图表' : 'Inference chart');
+
+/** The state the reader left behind, as it must come back. */
+function assertRestoredChart(): void {
+  cy.get('[data-testid="model-selector"]').should('contain.text', TARGET_MODEL_LABEL);
+  cy.get('[data-testid="chart-legend"] ul input[type="checkbox"]:checked')
+    .should('have.length', 1)
+    .closest('li')
+    .should('contain.text', SOLO_HW_LABEL);
+}
+
+describe('Agentic point detail — returning to the chart', () => {
+  it('carries the chart state into the detail link instead of dropping it', () => {
+    visitChart('/inference');
+    selectTargetModel();
+    soloLegendSeries();
+
+    cy.get('[data-testid="inference-chart-display"] svg .dot-group')
+      .should('have.length.greaterThan', 0)
+      .then(($dots) => {
+        const target = [...$dots].find((node) => {
+          const datum = (node as unknown as { __data__?: { hwKey?: string; id?: number } })
+            .__data__;
+          return datum?.hwKey === SOLO_HW_KEY && tracedIds.has(Number(datum?.id));
+        });
+        cy.wrap(target).find('.visible-shape').click({ force: true });
+      });
+
+    cy.get('[data-chart-tooltip]:visible [data-action="view-charts"]')
+      .should('be.visible')
+      .then(($link) => {
+        const href = $link.attr('href') ?? '';
+        expect(href, 'detail href').to.match(/^\/inference\/agentic\/\d+\?/u);
+        const params = new URLSearchParams(href.slice(href.indexOf('?')));
+        expect(params.get('g_model'), 'g_model carried into the detail link').to.equal('Kimi-K3');
+        expect(params.get('i_active'), 'legend selection carried').to.equal(SOLO_HW_KEY);
+      });
+  });
+
+  it('restores the model and legend when the reader clicks Back', () => {
+    visitChart('/inference');
+    selectTargetModel();
+    soloLegendSeries();
+    openPointDetail();
+
+    cy.location('pathname').should('match', /^\/inference\/agentic\/\d+$/u);
+    backControl('en').click();
+
+    // Picking Kimi K3 moved the chart's history entry onto its indexable
+    // subroute, so Back returns there — the path itself now pins the model.
+    cy.location('pathname').should('eq', '/inference/kimi-k3');
+    assertRestoredChart();
+  });
+
+  it('restores the model and legend through the "Inference chart" link', () => {
+    visitChart('/inference');
+    selectTargetModel();
+    soloLegendSeries();
+    openPointDetail();
+
+    chartLink('en')
+      .should('have.attr', 'href')
+      .and('match', /^\/inference\?/u);
+    chartLink('en').click();
+
+    cy.location('pathname').should('eq', '/inference');
+    assertRestoredChart();
+  });
+
+  it('keeps a Chinese reader on /zh for the whole round trip', () => {
+    visitChart('/zh/inference');
+    selectTargetModel();
+    soloLegendSeries();
+    openPointDetail();
+
+    cy.location('pathname').should('match', /^\/zh\/inference\/agentic\/\d+$/u);
+    chartLink('zh')
+      .should('have.attr', 'href')
+      .and('match', /^\/zh\/inference\?/u);
+    chartLink('zh').click();
+
+    cy.location('pathname').should('eq', '/zh/inference');
+    assertRestoredChart();
+  });
+
+  it('does not drop the unofficialruns overlay param on the round trip', () => {
+    visitChart('/inference?unofficialruns=987654321');
+    selectTargetModel();
+    soloLegendSeries();
+
+    cy.get('[data-testid="inference-chart-display"] svg .dot-group')
+      .should('have.length.greaterThan', 0)
+      .then(($dots) => {
+        const target = [...$dots].find((node) => {
+          const datum = (node as unknown as { __data__?: { hwKey?: string; id?: number } })
+            .__data__;
+          return datum?.hwKey === SOLO_HW_KEY && tracedIds.has(Number(datum?.id));
+        });
+        cy.wrap(target).find('.visible-shape').click({ force: true });
+      });
+
+    cy.get('[data-chart-tooltip]:visible [data-action="view-charts"]')
+      .should('be.visible')
+      .then(($link) => {
+        const href = $link.attr('href') ?? '';
+        const params = new URLSearchParams(href.slice(href.indexOf('?')));
+        expect(params.get('unofficialruns'), 'overlay run carried forward').to.equal('987654321');
+      });
+
+    cy.get('[data-chart-tooltip]:visible [data-action="view-charts"]').click();
+    cy.location('search').should('contain', 'unofficialruns=987654321');
+    chartLink('en').should('have.attr', 'href').and('contain', 'unofficialruns=987654321');
+  });
+});
diff --git a/packages/app/cypress/e2e/agentic-point-coach-mark.cy.ts b/packages/app/cypress/e2e/agentic-point-coach-mark.cy.ts
new file mode 100644
index 000000000..57c1b1dd0
--- /dev/null
+++ b/packages/app/cypress/e2e/agentic-point-coach-mark.cy.ts
@@ -0,0 +1,289 @@
+/**
+ * First-visit coach mark on the agentic inference chart.
+ *
+ * The affordance it teaches is invisible: clicking an agentic scatter point
+ * pins its tooltip, and the pinned tooltip carries a "View charts" link to the
+ * per-point server metrics and logs. The callout draws a pointer to a real,
+ * currently-visible point and goes away for good once the user either
+ * dismisses it or clicks a point.
+ */
+import {
+  interceptDerivedAgenticMetrics,
+  keepAgenticCoachMark,
+  unlockAgenticGate,
+} from '../support/e2e';
+import { plotBounds } from '@/lib/d3-chart/plot-bounds';
+import { OVERLAY_RUN_ID, interceptOverlayRun } from '../support/overlay-fixtures';
+
+// This spec owns coach-mark state, so it opts out of the global suppression in
+// cypress/support/e2e.ts rather than fighting it on every visit (that hook also
+// runs on cy.reload(), where onBeforeLoad cannot reach).
+keepAgenticCoachMark();
+
+const STORAGE_KEY = 'inferencex-agentic-point-coach-mark-dismissed';
+const COACH_MARK = '[data-testid="agentic-point-coach-mark"]';
+const AGENTIC_POINTS =
+  '[data-testid="scatter-graph"] .dot-group[data-benchmark-type="agentic_traces"]';
+// The pointer targets the marker inside the group, not the group's
+// label-inflated bounding box.
+const AGENTIC_MARKERS = `${AGENTIC_POINTS} .visible-shape`;
+
+/**
+ * Competing dashboard toasts land bottom-right, in the *overlay* slot rather
+ * than the coach mark's own slot — suppressed anyway so nothing else is
+ * animating while positions are measured.
+ */
+function freshCoachMarkState(win: Cypress.AUTWindow) {
+  unlockAgenticGate(win);
+  win.localStorage.removeItem(STORAGE_KEY);
+  win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+  win.sessionStorage.setItem('inferencex-reproducibility-nudge-shown', '1');
+  win.sessionStorage.setItem('inferencex-star-nudge-shown', '1');
+  win.localStorage.setItem('inferencex-filter-hint-nudge-dismissed', '1');
+}
+
+/**
+ * The callout only appears once there is a point on screen to point at, and at
+ * this viewport the chart starts below the fold — so bring it into view, the
+ * same thing a real first-time visitor does before they can see any of this.
+ */
+function visitAgenticChart(url = '/inference?i_seq=agentic-traces') {
+  interceptOverlayRun();
+  interceptDerivedAgenticMetrics();
+  cy.visit(url, { onBeforeLoad: freshCoachMarkState });
+  cy.get(AGENTIC_POINTS).should('have.length.greaterThan', 0);
+  cy.get('[data-testid="scatter-graph"]').scrollIntoView();
+}
+
+function centreOf(element: Element) {
+  const rect = element.getBoundingClientRect();
+  return { x: rect.left + rect.width / 2, y: rect.top + rect.height / 2 };
+}
+
+beforeEach(() => {
+  cy.clearAllLocalStorage();
+  cy.clearAllSessionStorage();
+  cy.intercept('GET', '/api/v1/trace-availability?ids=*', (request) => {
+    const ids = new URL(request.url).searchParams.get('ids')?.split(',').filter(Boolean) ?? [];
+    request.reply({
+      body: Object.fromEntries(ids.map((id) => [id, true])),
+    });
+  });
+});
+describe('Agentic point coach mark', () => {
+  it('points at a real, visible agentic point on the first visit', () => {
+    visitAgenticChart();
+
+    cy.get(COACH_MARK)
+      .should('be.visible')
+      .and('contain.text', 'Every point has a story')
+      .and('contain.text', 'Click any point to view server metrics and logs');
+
+    // Non-modal by construction: a focus trap or a scrim would block the very
+    // click the tip is asking for.
+    cy.get(COACH_MARK).should('have.attr', 'role', 'dialog');
+    cy.get(COACH_MARK).should('have.attr', 'aria-modal', 'false');
+    cy.get('[data-testid="agentic-point-coach-mark-layer"]').should(
+      'have.css',
+      'pointer-events',
+      'none',
+    );
+
+    // The callout must target an actual point, not empty plot area. The
+    // highlight ring sits exactly on the anchor (the pointer line deliberately
+    // stops short so its arrowhead doesn't cover the dot).
+    cy.get('[data-testid="agentic-point-coach-mark-pointer"]').should('exist');
+    cy.get('[data-testid="agentic-point-coach-mark-target"]').then(($ring) => {
+      const tipX = Number($ring.attr('cx'));
+      const tipY = Number($ring.attr('cy'));
+
+      cy.get(AGENTIC_MARKERS).then(($points) => {
+        const hit = [...$points].some((point) => {
+          const { x, y } = centreOf(point);
+          return Math.abs(x - tipX) < 1 && Math.abs(y - tipY) < 1;
+        });
+        expect(hit, 'pointer ends on an agentic point').to.eq(true);
+      });
+    });
+  });
+
+  it('waits for the chart to be on screen before pointing at anything', () => {
+    interceptOverlayRun();
+    interceptDerivedAgenticMetrics();
+    cy.visit('/inference?i_seq=agentic-traces', { onBeforeLoad: freshCoachMarkState });
+    cy.get(AGENTIC_POINTS).should('have.length.greaterThan', 0);
+
+    // Chart below the fold: no anchor, so no stranded arrow — and no dismissal
+    // burned either, since the tip has not been shown.
+    cy.window().its('scrollY').should('eq', 0);
+    cy.wait(2500);
+    cy.get(COACH_MARK).should('not.exist');
+    cy.window().then((win) => {
+      expect(win.localStorage.getItem(STORAGE_KEY)).to.eq(null);
+    });
+
+    cy.get('[data-testid="scatter-graph"]').scrollIntoView();
+    cy.get(COACH_MARK).should('be.visible');
+  });
+
+  it('targets the clipped plot area, not the axis gutters', () => {
+    // `plotBounds` reconstructs the clip region from the live chart DOM. If the
+    // skeleton it reads ever changes shape it returns null and the resolver
+    // silently falls back to the SVG box — which includes the 60px axis
+    // gutters, where a zoomed point is invisible. Assert against the real DOM
+    // so that fallback cannot go unnoticed.
+    visitAgenticChart();
+    cy.get(COACH_MARK).should('be.visible');
+
+    cy.get('[data-testid="scatter-graph"] [data-testid="d3-chart-svg"]').then(($svg) => {
+      const svg = $svg[0];
+      const bounds = plotBounds(svg);
+      expect(bounds, 'clip region resolves against the real chart').to.not.eq(null);
+
+      const box = svg.getBoundingClientRect();
+      expect(bounds!.left, 'left gutter excluded').to.be.greaterThan(box.left);
+      expect(bounds!.bottom, 'bottom gutter excluded').to.be.lessThan(box.bottom);
+
+      cy.get('[data-testid="agentic-point-coach-mark-target"]').then(($ring) => {
+        const cx = Number($ring.attr('cx'));
+        const cy = Number($ring.attr('cy'));
+        expect(cx).to.be.within(bounds!.left, bounds!.right);
+        expect(cy).to.be.within(bounds!.top, bounds!.bottom);
+      });
+    });
+  });
+
+  it('keeps the card fully on screen', () => {
+    visitAgenticChart();
+    cy.get(COACH_MARK).should('be.visible');
+
+    cy.window().then((win) => {
+      const rect = win.document.querySelector(COACH_MARK)!.getBoundingClientRect();
+      const { clientWidth, clientHeight } = win.document.documentElement;
+      expect(rect.left, 'left edge').to.be.at.least(0);
+      expect(rect.top, 'top edge').to.be.at.least(0);
+      expect(rect.right, 'right edge').to.be.at.most(clientWidth);
+      expect(rect.bottom, 'bottom edge').to.be.at.most(clientHeight);
+    });
+  });
+
+  it('follows its point when the chart is zoomed', () => {
+    visitAgenticChart();
+    cy.get('[data-testid="agentic-point-coach-mark-target"]').should('exist');
+
+    // Shift+wheel is the chart's zoom gesture (see the chart instructions).
+    cy.get('[data-testid="d3-chart-svg"]')
+      .first()
+      .trigger('wheel', { deltaY: -400, shiftKey: true, bubbles: true });
+
+    cy.get('[data-testid="agentic-point-coach-mark-target"]').should(($ring) => {
+      const tipX = Number($ring.attr('cx'));
+      const tipY = Number($ring.attr('cy'));
+      // Wherever the point ended up, the callout is still on top of one —
+      // no stranded arrow pointing at empty plot area.
+      const points = Cypress.$(AGENTIC_MARKERS)
+        .filter((_i, el) => getComputedStyle(el.parentElement!).opacity !== '0')
+        .toArray();
+      const hit = points.some((point) => {
+        const { x, y } = centreOf(point);
+        return Math.abs(x - tipX) < 1.5 && Math.abs(y - tipY) < 1.5;
+      });
+      expect(hit, 'callout still targets an agentic point after zoom').to.eq(true);
+    });
+  });
+
+  it('dismisses from the close button and stays gone after reload', () => {
+    visitAgenticChart();
+    cy.get(COACH_MARK).should('be.visible');
+
+    cy.get('[data-testid="agentic-point-coach-mark-dismiss"]').click();
+    cy.get(COACH_MARK).should('not.exist');
+    cy.window().then((win) => {
+      expect(win.localStorage.getItem(STORAGE_KEY)).to.eq('1');
+    });
+
+    interceptOverlayRun();
+    interceptDerivedAgenticMetrics();
+    cy.reload();
+    cy.get(AGENTIC_POINTS).should('have.length.greaterThan', 0);
+    cy.wait(1500);
+    cy.get(COACH_MARK).should('not.exist');
+  });
+
+  it('dismisses when the user actually clicks a point', () => {
+    visitAgenticChart();
+    cy.get(COACH_MARK).should('be.visible');
+
+    cy.get(AGENTIC_POINTS).first().find('.visible-shape').click({ force: true });
+
+    cy.get(COACH_MARK).should('not.exist');
+    // Clicking a point is engagement, so it persists the dismissal too.
+    cy.window().then((win) => {
+      expect(win.localStorage.getItem(STORAGE_KEY)).to.eq('1');
+    });
+    // …and the interaction it was teaching still happened.
+    cy.get('[data-chart-tooltip]:visible').should('have.length', 1);
+  });
+
+  it('is dismissable from the keyboard', () => {
+    visitAgenticChart();
+    cy.get(COACH_MARK).should('be.visible');
+
+    cy.get('body').type('{esc}');
+    cy.get(COACH_MARK).should('not.exist');
+    cy.window().then((win) => {
+      expect(win.localStorage.getItem(STORAGE_KEY)).to.eq('1');
+    });
+  });
+
+  it('never anchors to an unofficial-run overlay marker', () => {
+    // Overlay runs have no stored trace, so their tooltip offers no
+    // "View charts" link — anchoring there would teach a dead end.
+    visitAgenticChart(`/inference?unofficialrun=${OVERLAY_RUN_ID}&i_seq=agentic-traces`);
+    cy.wait('@unofficialRun');
+    cy.get('[data-testid="scatter-graph"] .unofficial-overlay-pt').should(
+      'have.length.greaterThan',
+      0,
+    );
+
+    cy.get(COACH_MARK).should('be.visible');
+    cy.get('[data-testid="agentic-point-coach-mark-target"]').then(($ring) => {
+      const tipX = Number($ring.attr('cx'));
+      const tipY = Number($ring.attr('cy'));
+      const onPoint = (element: Element) => {
+        const { x, y } = centreOf(element);
+        return Math.abs(x - tipX) < 1 && Math.abs(y - tipY) < 1;
+      };
+
+      cy.get('[data-testid="scatter-graph"] .unofficial-overlay-pt').then(($overlay) => {
+        expect([...$overlay].some(onPoint), 'pointer avoids overlay markers').to.eq(false);
+      });
+      cy.get(AGENTIC_MARKERS).then(($official) => {
+        expect([...$official].some(onPoint), 'pointer lands on an official point').to.eq(true);
+      });
+    });
+  });
+
+  it('does not appear on the agentic detail page it links to', () => {
+    cy.visit('/inference/agentic/206885', {
+      onBeforeLoad: freshCoachMarkState,
+      failOnStatusCode: false,
+    });
+    cy.wait(2000);
+    cy.get(COACH_MARK).should('not.exist');
+  });
+
+  it('renders Simplified Chinese copy on /zh', () => {
+    visitAgenticChart('/zh/inference?i_seq=agentic-traces');
+
+    cy.get(COACH_MARK)
+      .should('be.visible')
+      .and('contain.text', '每个数据点背后都有细节')
+      .and('contain.text', '点击任意数据点即可查看服务端指标与日志');
+    cy.get('[data-testid="agentic-point-coach-mark-dismiss"]').should(
+      'have.attr',
+      'aria-label',
+      '关闭提示',
+    );
+  });
+});
diff --git a/packages/app/cypress/e2e/agentic-point-time-series.cy.ts b/packages/app/cypress/e2e/agentic-point-time-series.cy.ts
index ebb7bba3c..d63920401 100644
--- a/packages/app/cypress/e2e/agentic-point-time-series.cy.ts
+++ b/packages/app/cypress/e2e/agentic-point-time-series.cy.ts
@@ -7,6 +7,7 @@ const timelineRequest = (
   overrides: Record<string, unknown> = {},
 ) => ({
   cid: 'conversation-1',
+  ri: 0,
   ti: index,
   wid: 'worker-1',
   ad: 0,
@@ -23,49 +24,344 @@ const timelineRequest = (
   ...overrides,
 });
 
+const requestChartPayload = (requests: ReturnType<typeof timelineRequest>[]) => {
+  const cids = [...new Set(requests.map((request) => request.cid))];
+  const phases = [...new Set(requests.map((request) => request.phase))];
+  return {
+    version: 602,
+    timelineVersion: 6,
+    startNs: 0,
+    endNs: 7_000_000_000,
+    durationS: 7,
+    cids,
+    phases,
+    requests: requests.map((request) => [
+      cids.indexOf(request.cid),
+      phases.indexOf(request.phase),
+      Math.round(request.start / 1_000),
+      Math.round(request.end / 1_000),
+      request.ttftMs,
+      request.tpotMs,
+      request.isl,
+      request.osl,
+      request.cancelled ? 1 : 0,
+    ]),
+  };
+};
+
+const benchmarkSiblings = {
+  sku: {
+    hardware: 'b300',
+    framework: 'vllm',
+    model: 'kimik3',
+    precision: 'fp4',
+    spec_method: 'mtp',
+    benchmark_type: 'agentic_traces',
+    github_run_id: 31893747354,
+    date: '2026-08-15',
+    dataset_slug: null,
+  },
+  siblings: [
+    {
+      id: 206885,
+      conc: 8,
+      offload_mode: 'off',
+      decode_tp: 8,
+      decode_ep: 1,
+      decode_pp: 1,
+      decode_dcp_size: 8,
+      decode_pcp_size: 1,
+      decode_dp_attention: false,
+      decode_num_workers: 0,
+      prefill_tp: 8,
+      prefill_ep: 1,
+      prefill_pp: 1,
+      prefill_dcp_size: 8,
+      prefill_pcp_size: 1,
+      prefill_dp_attention: false,
+      prefill_num_workers: 0,
+      num_prefill_gpu: 8,
+      num_decode_gpu: 8,
+      disagg: false,
+      is_multinode: false,
+      tput_per_gpu: 1.2,
+      total_requests: 395,
+      is_current: true,
+      has_trace: true,
+    },
+  ],
+};
+
 describe('Agentic point request metric time series', () => {
-  before(() => {
+  beforeEach(() => {
+    const requests = [
+      timelineRequest(0, 100, 10),
+      timelineRequest(1, 200, 20),
+      timelineRequest(2, 400, 25),
+      timelineRequest(3, 800, 40),
+      timelineRequest(4, 1600, 80, { ri: 1 }),
+      timelineRequest(5, 3200, 160, { phase: 'warmup' }),
+      timelineRequest(6, 6400, 320, { cancelled: true }),
+      timelineRequest(7, 0, 0, {
+        cid: 'conversation-1::sa:subagent_001_abcd',
+        credit: 1_100_000_000,
+        start: 1_100_000_000,
+        end: 1_900_000_000,
+        ttftMs: null,
+        tpotMs: null,
+        isl: null,
+        osl: null,
+      }),
+      timelineRequest(8, 0, 0, {
+        cid: 'conversation-1::sa:subagent_001_abcd:aux:011',
+        credit: 1_200_000_000,
+        start: 1_200_000_000,
+        end: 1_800_000_000,
+        ttftMs: null,
+        tpotMs: null,
+        isl: null,
+        osl: null,
+      }),
+    ];
     cy.intercept('GET', '/api/v1/trace-histograms*', { body: {} });
     cy.intercept('GET', '/api/v1/trace-server-metrics*', { body: null });
-    cy.intercept('GET', '/api/v1/benchmark-siblings*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/benchmark-siblings*', { body: benchmarkSiblings });
+    cy.intercept('GET', '/api/v1/request-chart-data*', {
+      body: requestChartPayload(requests),
+    }).as('requestChartData');
     cy.intercept('GET', '/api/v1/request-timeline*', {
       body: {
-        version: 3,
+        version: 6,
         startNs: 0,
         endNs: 7_000_000_000,
         durationS: 7,
-        requests: [
-          timelineRequest(0, 100, 10),
-          timelineRequest(1, 200, 20),
-          timelineRequest(2, 400, 25),
-          timelineRequest(3, 800, 40),
-          timelineRequest(4, 1600, 80),
-          timelineRequest(5, 3200, 160, { phase: 'warmup' }),
-          timelineRequest(6, 6400, 320, { cancelled: true }),
-          timelineRequest(7, 0, 0, {
-            cid: 'conversation-1::sa:subagent_001_abcd',
-            credit: 1_100_000_000,
-            start: 1_100_000_000,
-            end: 1_900_000_000,
-            ttftMs: null,
-            tpotMs: null,
-            isl: null,
-            osl: null,
-          }),
-          timelineRequest(8, 0, 0, {
-            cid: 'conversation-1::sa:subagent_001_abcd:aux:011',
-            credit: 1_200_000_000,
-            start: 1_200_000_000,
-            end: 1_800_000_000,
-            ttftMs: null,
-            tpotMs: null,
-            isl: null,
-            osl: null,
-          }),
-        ],
+        requests,
       },
     });
     cy.visit('/inference/agentic/206885', { onBeforeLoad: unlockAgenticGate });
+    // The segmented toggle is present in the server-rendered HTML. Waiting for
+    // this client-only query proves hydration has attached its click handlers,
+    // avoiding a Firefox race where an early click leaves `?view=` unchanged.
+    cy.wait('@requestChartData');
+  });
+
+  it('uses the shared topology label for the active agentic point', () => {
+    cy.contains('button', 'TP8/DCP8 • c=8').should('be.visible');
+  });
+
+  it('opens the stored server log and loads it incrementally', () => {
+    const longServerLogPath =
+      'agentic/conc_1152/aiperf_artifacts/logs/aiperf/2026-08-20/server.log';
+    const firstChunk = `INFO server ready\n${'trace line\n'.repeat(4_000)}`;
+    const searchMatchOffset = 70_027;
+    const searchJumpOffset = searchMatchOffset - 16 * 1024;
+    const routerJumpChunk = `${'context line\n'.repeat(1_260)}${'x'.repeat(
+      searchMatchOffset - searchJumpOffset - 'context line\n'.length * 1_260,
+    )}router ready for requests\n`;
+    cy.intercept(
+      { method: 'GET', pathname: '/api/v1/server-log-files' },
+      {
+        body: [longServerLogPath, 'results/benchmark.log', 'results/router.log'],
+      },
+    ).as('serverLogFiles');
+    cy.intercept({ method: 'GET', pathname: '/api/v1/server-log' }, (request) => {
+      const params = new URL(request.url).searchParams;
+      const offset = Number(params.get('offset') ?? 0);
+      const fileName = params.get('file') ?? longServerLogPath;
+      if (fileName === longServerLogPath && offset === 0) {
+        request.alias = 'serverLogInitialChunk';
+      } else if (fileName === longServerLogPath && offset === 31) {
+        request.alias = 'serverLogNextChunk';
+      } else if (fileName === 'results/router.log' && offset === searchJumpOffset) {
+        request.alias = 'serverLogJumpChunk';
+      } else if (fileName === 'results/router.log' && offset === 0) {
+        request.alias = 'serverLogRouterStart';
+      }
+      request.reply({
+        body:
+          fileName === 'results/router.log'
+            ? {
+                id: 206885,
+                fileName,
+                serverLog: offset === searchJumpOffset ? routerJumpChunk : 'INFO router ready\n',
+                offset,
+                nextOffset: null,
+              }
+            : offset === 0
+              ? {
+                  id: 206885,
+                  fileName,
+                  serverLog: `\u001B[32m${firstChunk}\u001B[0m`,
+                  offset: 0,
+                  nextOffset: 31,
+                }
+              : {
+                  id: 206885,
+                  fileName,
+                  serverLog: 'INFO benchmark complete\n',
+                  offset,
+                  nextOffset: null,
+                },
+      });
+    }).as('serverLogChunk');
+    cy.intercept({ method: 'GET', pathname: '/api/v1/server-log-search' }, (request) => {
+      expect(new URL(request.url).searchParams.get('q')).to.equal('router ready');
+      request.reply({
+        body: {
+          id: 206885,
+          query: 'router ready',
+          matches: [
+            {
+              fileName: 'results/router.log',
+              offset: searchMatchOffset,
+              before: 'INFO ',
+              match: 'router ready',
+              after: ' for requests',
+            },
+          ],
+          truncated: false,
+        },
+      });
+    }).as('serverLogSearch');
+    cy.intercept(
+      {
+        method: 'GET',
+        pathname: '/api/v1/server-log',
+        query: {
+          id: '206885',
+          file: longServerLogPath,
+          download: '1',
+        },
+      },
+      {
+        headers: {
+          'content-disposition': 'attachment; filename="server.log"',
+          'content-type': 'text/plain; charset=utf-8',
+        },
+        body: firstChunk,
+      },
+    ).as('serverLogDownload');
+    cy.viewport(480, 900);
+    cy.get('[data-testid="detail-view-logs"]').click();
+    cy.location('search').should('contain', 'view=logs');
+    cy.wait('@serverLogFiles');
+    cy.wait('@serverLogInitialChunk');
+
+    cy.get('[data-testid="agentic-server-log-viewer"]')
+      .should('contain.text', 'Log files')
+      .and('contain.text', longServerLogPath)
+      .then(($viewer) => {
+        cy.get('#agentic-log-file').then(($trigger) => {
+          const viewerBounds = $viewer[0].getBoundingClientRect();
+          const triggerBounds = $trigger[0].getBoundingClientRect();
+          expect(triggerBounds.left).to.be.at.least(viewerBounds.left);
+          expect(triggerBounds.right).to.be.at.most(viewerBounds.right);
+        });
+      });
+    cy.get('[data-testid="server-log-content"]')
+      .should('contain.text', 'INFO server ready')
+      .and('not.contain.text', '\u001B[32m');
+
+    cy.get('[data-testid="server-log-search"]').type('router ready');
+    cy.wait('@serverLogSearch');
+    cy.get('[data-testid="server-log-search-results"]')
+      .should('contain.text', '1 match')
+      .and('contain.text', 'results/router.log')
+      .and('contain.text', 'character 70,028')
+      .and('contain.text', 'INFO router ready for requests');
+
+    cy.get('[data-testid="go-to-server-log-match"]').click();
+    cy.wait('@serverLogJumpChunk').then(({ request }) => {
+      const params = new URL(request.url).searchParams;
+      expect(params.get('file')).to.equal('results/router.log');
+      expect(params.get('offset')).to.equal(String(searchJumpOffset));
+    });
+    cy.get('#agentic-log-file').should('contain.text', 'results/router.log');
+    cy.get('[data-testid="server-log-jump-highlight"]')
+      .should('be.visible')
+      .and('have.text', 'router ready');
+
+    cy.get('#agentic-log-file').click();
+    cy.contains('[role="option"]', longServerLogPath).click();
+    cy.get('[data-testid="server-log-content"]').should('contain.text', 'INFO server ready');
+    cy.get('[data-testid="download-selected-server-log"]')
+      .should('have.attr', 'href')
+      .and('include', 'download=1');
+    cy.get('[data-testid="download-selected-server-log"]').click();
+    cy.wait('@serverLogDownload');
+
+    cy.get('[data-testid="server-log-content"]').scrollTo('bottom');
+    cy.wait('@serverLogNextChunk');
+    cy.get('[data-testid="server-log-content"]')
+      .should('contain.text', 'INFO server ready')
+      .and('contain.text', 'INFO benchmark complete');
+    cy.contains('End of stored log').should('be.visible');
+
+    cy.get('#agentic-log-file').click();
+    cy.contains('[role="option"]', 'results/router.log').click();
+    cy.wait('@serverLogRouterStart');
+    cy.get('[data-testid="server-log-content"]').should('contain.text', 'INFO router ready');
+
+    cy.viewport(1280, 720);
+    cy.get('[data-testid="detail-view-point"]').click();
+  });
+
+  it('keeps loaded log text visible when a later chunk fails', () => {
+    const retainedLog = `INFO retained after failure\n${'trace line\n'.repeat(2_000)}`;
+    cy.intercept(
+      { method: 'GET', pathname: '/api/v1/server-log-files' },
+      { body: ['results/server.log', 'results/benchmark.log'] },
+    );
+    cy.intercept({ method: 'GET', pathname: '/api/v1/server-log' }, (request) => {
+      const params = new URL(request.url).searchParams;
+      const fileName = params.get('file');
+      const offset = Number(params.get('offset') ?? 0);
+      if (fileName !== 'results/benchmark.log') {
+        request.reply({
+          body: {
+            id: 206885,
+            fileName,
+            serverLog: 'INFO initial file\n',
+            offset,
+            nextOffset: null,
+          },
+        });
+        return;
+      }
+      request.alias = 'retainedLogChunk';
+      if (offset === 0) {
+        request.reply({
+          body: {
+            id: 206885,
+            fileName: 'results/benchmark.log',
+            serverLog: retainedLog,
+            offset: 0,
+            nextOffset: 28,
+          },
+        });
+      } else {
+        request.reply({ statusCode: 503, body: { error: 'temporary failure' } });
+      }
+    });
+
+    cy.contains('button', 'TP8/DCP8 • c=8').should('be.visible');
+    cy.get('[data-testid="detail-view-logs"]').click();
+    cy.location('search').should('contain', 'view=logs');
+    cy.get('#agentic-log-file').click();
+    cy.contains('[role="option"]', 'results/benchmark.log').click();
+    cy.wait('@retainedLogChunk');
+    cy.get('[data-testid="server-log-content"]').should(
+      'contain.text',
+      'INFO retained after failure',
+    );
+    cy.get('[data-testid="load-more-server-log"]').click();
+    cy.contains('The next chunk could not be loaded', { timeout: 15_000 }).should('be.visible');
+    cy.get('[data-testid="server-log-content"]').should(
+      'contain.text',
+      'INFO retained after failure',
+    );
+    cy.get('[data-testid="detail-view-point"]').click();
+    cy.location('search').should('not.contain', 'view=logs');
   });
 
   it('renders rolling P90 interactivity and TTFT by default using profiling requests only', () => {
@@ -181,14 +477,22 @@ describe('Agentic point request metric time series', () => {
     // coverage is now continuous across [0s, 7s]: idle 0ms (0.0%). A 1.00s value
     // here would mean the Gantt had regressed to label-based filtering.
     cy.get('[data-testid="timeline-total-idle-time"]').should('have.text', 'idle 0ms (0.0%)');
+    cy.get('[data-timeline-row-kind="parent"]')
+      .should('have.length', 2)
+      .then(($rows) => {
+        expect($rows.eq(0)).to.contain.text('conversation-1 · replay 1');
+        expect($rows.eq(1)).to.contain.text('conversation-1 · replay 2');
+      });
     cy.get('[data-timeline-row-kind="aux"]')
       .should('have.css', 'padding-left', '24px')
       .and('contain.text', 'aux 011 · parallel');
   });
 
   it('restores the request timeline view after browser Back from a dataset route', () => {
+    cy.get('[data-testid="detail-view-timeline"]').click();
+    cy.get('[data-testid="timeline-total-idle-time"]').should('be.visible');
     cy.window().then((win) => {
-      win.history.pushState({}, '', '/datasets/test-dataset/conversations/conversation-1');
+      win.history.pushState({}, '', '/agentx/test-dataset/conversations/conversation-1');
     });
     cy.go('back');
     cy.location('pathname').should('eq', '/inference/agentic/206885');
@@ -204,6 +508,82 @@ describe('Agentic point request metric time series', () => {
       cy.get('svg path[stroke="#ef4444"]').should('have.length', 1);
     });
   });
+
+  it('renders the log viewer in Simplified Chinese', () => {
+    cy.intercept('GET', '/api/v1/trace-server-metrics*', { body: null });
+    cy.intercept('GET', '/api/v1/benchmark-siblings*', { body: benchmarkSiblings });
+    cy.intercept(
+      { method: 'GET', pathname: '/api/v1/server-log-files' },
+      {
+        body: ['results/server.log'],
+      },
+    );
+    cy.intercept(
+      { method: 'GET', pathname: '/api/v1/server-log' },
+      {
+        body: {
+          id: 206885,
+          fileName: 'results/server.log',
+          serverLog: 'INFO server ready\n',
+          offset: 0,
+          nextOffset: null,
+        },
+      },
+    );
+    cy.visit('/zh/inference/agentic/206885?view=logs', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="detail-view-logs"]')
+      .should('have.attr', 'aria-selected', 'true')
+      .and('have.text', '日志');
+    cy.get('[data-testid="agentic-server-log-viewer"]')
+      .should('contain.text', '日志文件')
+      .and('contain.text', 'results/server.log')
+      .and('contain.text', '搜索所有日志文件')
+      .and('contain.text', '已到达日志末尾');
+  });
+});
+
+describe('Agentic request timeline virtualization', () => {
+  const requests = Array.from({ length: 2_000 }, (_, index) =>
+    timelineRequest(index, 10, 10, {
+      cid: `conversation-${String(index).padStart(4, '0')}`,
+      credit: index * 1_000_000_000,
+      start: index * 1_000_000_000,
+      end: (index + 1) * 1_000_000_000,
+    }),
+  );
+
+  beforeEach(() => {
+    cy.intercept('GET', '/api/v1/trace-server-metrics*', { body: null });
+    cy.intercept('GET', '/api/v1/benchmark-siblings*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/request-chart-data*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/request-timeline*', {
+      body: {
+        version: 5,
+        startNs: 0,
+        endNs: 2_000_000_000_000,
+        durationS: 2_000,
+        requests,
+      },
+    });
+    cy.visit('/inference/agentic/206885?view=timeline', { onBeforeLoad: unlockAgenticGate });
+  });
+
+  it('keeps row labels and SVG request bars bounded to the viewport', () => {
+    cy.get('[data-testid="request-timeline-scroll"]').should('be.visible');
+    cy.get('[data-timeline-row-kind]').should(($rows) => {
+      expect($rows.length).to.be.lessThan(40);
+    });
+    cy.get('[data-testid="request-timeline-svg"] rect').should(($bars) => {
+      expect($bars.length).to.be.lessThan(100);
+    });
+
+    cy.get('[data-testid="request-timeline-scroll"]').scrollTo('bottom');
+    cy.get('[data-timeline-row-kind]').should('contain.text', 'conversation-1999');
+    cy.get('[data-timeline-row-kind]').should(($rows) => {
+      expect($rows.length).to.be.lessThan(40);
+    });
+  });
 });
 
 const pointMeta = {
@@ -279,7 +659,12 @@ describe('Agentic point orchestrator metric sources', () => {
     );
     cy.intercept('GET', '/api/v1/trace-histograms*', { body: {} });
     cy.intercept('GET', '/api/v1/benchmark-siblings*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/request-chart-data*', { statusCode: 404 });
     cy.intercept('GET', '/api/v1/request-timeline*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/trace-server-metric-source*', (request) => {
+      const source = new URL(request.url).searchParams.get('source');
+      request.reply({ body: source === decode.source.id ? decode : prefill });
+    });
     cy.intercept('GET', '/api/v1/trace-server-metrics*', {
       body: {
         meta: pointMeta,
@@ -296,7 +681,7 @@ describe('Agentic point orchestrator metric sources', () => {
         prefixCacheHitsTps: prefill.prefixCacheHitsTps,
         hostKvCacheUsage: [],
         kvCacheUsageByEngine: [],
-        metricSources: [prefill, decode],
+        metricSources: [{ source: prefill.source }, { source: decode.source }],
       },
     });
     cy.visit('/inference/agentic/206885', { onBeforeLoad: unlockAgenticGate });
@@ -353,3 +738,64 @@ describe('Agentic point orchestrator metric sources', () => {
     cy.get('[data-testid="throughput-series-decode"]').should('have.attr', 'aria-pressed', 'false');
   });
 });
+
+const engineSeries = (engineLabel: string, value: number) => ({
+  engineLabel,
+  points: [
+    { t: 0, value },
+    { t: 1, value: value + 0.05 },
+  ],
+});
+
+describe('Agentic point per-engine KV overlay', () => {
+  beforeEach(() => {
+    cy.intercept('GET', '/api/v1/trace-histograms*', { body: {} });
+    cy.intercept('GET', '/api/v1/benchmark-siblings*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/request-chart-data*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/request-timeline*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/trace-server-metrics*', {
+      body: {
+        meta: pointMeta,
+        startNs: 0,
+        endNs: 2_000_000_000,
+        durationS: 2,
+        timeslicesCount: 2,
+        kvCacheUsage: [
+          { t: 0, value: 0.3 },
+          { t: 1, value: 0.35 },
+        ],
+        prefixCacheHitRate: [],
+        queueDepth: [],
+        promptTokensBySource: {},
+        prefillTps: [],
+        decodeTps: [],
+        prefixCacheHitsTps: [],
+        hostKvCacheUsage: [],
+        // Bare DP ranks plus a role-qualified engine, as the ETL emits for a
+        // disaggregated run where the decode worker reports no rank label.
+        kvCacheUsageByEngine: [
+          engineSeries('0', 0.2),
+          engineSeries('1', 0.4),
+          engineSeries('decode', 0.6),
+        ],
+        metricSources: [],
+      },
+    });
+    cy.visit('/inference/agentic/206885', { onBeforeLoad: unlockAgenticGate });
+  });
+
+  it('draws one legend entry per engine and leaves named engines unprefixed', () => {
+    cy.contains('svg', 'KV cache (%)')
+      .first()
+      .within(() => {
+        cy.contains('text', 'DP 0').should('be.visible');
+        cy.contains('text', 'DP 1').should('be.visible');
+        // Already self-describing — must NOT come out as "DP decode".
+        cy.contains('text', 'decode').should('be.visible');
+        cy.contains('text', 'DP decode').should('not.exist');
+        cy.contains('text', 'Avg').should('be.visible');
+        // One line per engine plus the average, and no duplicate chips.
+        cy.get('path[fill="none"]').should('have.length', 4);
+      });
+  });
+});
diff --git a/packages/app/cypress/e2e/agentic-quick-filters.cy.ts b/packages/app/cypress/e2e/agentic-quick-filters.cy.ts
new file mode 100644
index 000000000..9fa68cf49
--- /dev/null
+++ b/packages/app/cypress/e2e/agentic-quick-filters.cy.ts
@@ -0,0 +1,153 @@
+import { unlockAgenticGate } from '../support/e2e';
+
+const MODEL = 'dsv4';
+const DATE = '2026-06-12';
+
+const percentileLadder = (prefix: string, base: number): Record<string, number> => ({
+  [`median_${prefix}`]: base,
+  [`p75_${prefix}`]: base * 1.2,
+  [`p90_${prefix}`]: base * 1.5,
+  [`p95_${prefix}`]: base * 1.7,
+  [`p99_${prefix}`]: base * 2.2,
+  [`std_${prefix}`]: base * 0.3,
+});
+
+const metrics = (conc: number): Record<string, number> => {
+  const scale = conc / 16;
+  const itl = 0.011 * scale;
+  return {
+    ...percentileLadder('ttft', 0.4 * scale),
+    ...percentileLadder('tpot', 0.012 * scale),
+    ...percentileLadder('itl', itl),
+    ...percentileLadder('e2el', 8 * scale),
+    median_intvty: 1 / itl,
+    p75_intvty: 1 / (itl * 1.2),
+    p90_intvty: 1 / (itl * 1.5),
+    p99_intvty: 1 / (itl * 2.2),
+    std_intvty: (1 / itl) * 0.1,
+    tput_per_gpu: 950 / Math.sqrt(scale),
+    output_tput_per_gpu: 210,
+    input_tput_per_gpu: 740,
+    total_tput_tps: 380 * conc,
+  };
+};
+
+const configs = [
+  { hardware: 'b200', framework: 'vllm', disagg: false },
+  { hardware: 'mi300x', framework: 'sglang', disagg: true },
+];
+
+const availability = [
+  ...configs.map((config) => ({
+    model: MODEL,
+    isl: null,
+    osl: null,
+    precision: 'fp4',
+    hardware: config.hardware,
+    framework: config.framework,
+    spec_method: 'mtp',
+    disagg: config.disagg,
+    benchmark_type: 'agentic_traces',
+    date: DATE,
+  })),
+  ...configs.map((config) => ({
+    model: MODEL,
+    isl: 8192,
+    osl: 1024,
+    precision: 'fp4',
+    hardware: config.hardware,
+    framework: config.framework,
+    spec_method: 'none',
+    disagg: config.disagg,
+    benchmark_type: 'single_turn',
+    date: DATE,
+  })),
+];
+
+let benchmarkId = 980000;
+const benchmarks = configs.flatMap((config) =>
+  [16, 64, 128].map((conc) => ({
+    id: benchmarkId++,
+    ...config,
+    model: MODEL,
+    precision: 'fp4',
+    spec_method: 'mtp',
+    is_multinode: false,
+    prefill_tp: 8,
+    decode_tp: 8,
+    num_prefill_gpu: 8,
+    num_decode_gpu: 8,
+    isl: null,
+    osl: null,
+    conc,
+    offload_mode: 'off',
+    benchmark_type: 'agentic_traces',
+    image: `${config.framework}/server:test`,
+    metrics: metrics(conc),
+    workers: null,
+    date: DATE,
+    run_url: null,
+  })),
+);
+
+function visitAgenticQuickFilters() {
+  cy.intercept('GET', '/api/v1/availability', { body: availability }).as('availability');
+  cy.intercept('GET', '/api/v1/benchmarks*', { body: benchmarks }).as('benchmarks');
+  cy.visit('/inference?g_model=DeepSeek-V4-Pro&i_seq=agentic-traces&i_prec=fp4&i_spec=mtp', {
+    onBeforeLoad(win) {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+      unlockAgenticGate(win);
+    },
+  });
+  cy.wait(['@availability', '@benchmarks']);
+  cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
+}
+
+describe('Agentic Quick Filters', () => {
+  it('opens from the legend, omits spec decoding, and filters the rendered agentic points', () => {
+    visitAgenticQuickFilters();
+
+    // A stale speculative-decoding URL selection is ignored for agentic data.
+    cy.get('.dot-group[data-hw-key^="b200"]').should('exist');
+    cy.get('.dot-group[data-hw-key^="mi300x"]').should('exist');
+    cy.get('[data-testid="quick-filters-dialog"]').should('not.exist');
+    cy.get('[data-testid="scatter-quick-filters"]').should('contain.text', 'Quick Filters').click();
+    cy.get('[data-testid="quick-filters-dialog"]').should('be.visible');
+    cy.get('[data-testid="quick-filters-selected-count"]').should('not.exist');
+
+    cy.get('[data-testid^="quick-filter-spec-"]').should('not.exist');
+    cy.contains('Spec Decoding').should('not.exist');
+
+    cy.get('[data-testid="quick-filter-vendor-AMD"]').click();
+    cy.get('.dot-group[data-hw-key^="b200"]').should('not.exist');
+    cy.get('.dot-group[data-hw-key^="mi300x"]').should('exist');
+    cy.get('[data-testid="quick-filters-selected-count"]').should('contain.text', '1 selected');
+
+    // An incompatible cross-group selection produces zero points. The dialog
+    // must stay mounted so the reader can recover without reloading the page.
+    cy.get('[data-testid="quick-filter-framework-vllm"]').click();
+    cy.get('[data-testid="scatter-empty-quick-filters"]').should('exist');
+    cy.get('[data-testid="quick-filters-dialog"]').should('be.visible');
+
+    cy.contains('button', 'Clear filters').click();
+    cy.get('.dot-group[data-hw-key^="b200"]').should('exist');
+    cy.get('.dot-group[data-hw-key^="mi300x"]').should('exist');
+    cy.get('[data-testid="quick-filters-selected-count"]').should('not.exist');
+  });
+
+  it('clears Quick Filters from the legend reset action', () => {
+    visitAgenticQuickFilters();
+
+    cy.get('[data-testid="scatter-quick-filters"]').click();
+    cy.get('[data-testid="quick-filter-vendor-AMD"]').click();
+    cy.contains('button', 'Done').click();
+
+    cy.get('.dot-group[data-hw-key^="b200"]').should('not.exist');
+    cy.get('[data-testid="scatter-reset-filter"]').click();
+    cy.get('.dot-group[data-hw-key^="b200"]').should('exist');
+    cy.get('.dot-group[data-hw-key^="mi300x"]').should('exist');
+
+    cy.get('[data-testid="scatter-quick-filters"]').click();
+    cy.get('[data-testid="quick-filter-vendor-AMD"]').should('have.attr', 'aria-pressed', 'false');
+  });
+});
diff --git a/packages/app/cypress/e2e/agentx-optimizations.cy.ts b/packages/app/cypress/e2e/agentx-optimizations.cy.ts
new file mode 100644
index 000000000..acf4e0616
--- /dev/null
+++ b/packages/app/cypress/e2e/agentx-optimizations.cy.ts
@@ -0,0 +1,148 @@
+import { unlockAgenticGate } from '../support/e2e';
+
+const FRAMEWORK_SLUGS = [
+  'vllm',
+  'sglang',
+  'tensorrt-llm',
+  'atom',
+  'aiter',
+  'dynamo',
+  'lmcache',
+  'mooncake',
+];
+
+describe('AgentX optimizations', () => {
+  beforeEach(() => {
+    cy.intercept('GET', '/api/v1/datasets', { statusCode: 200, body: [] });
+  });
+
+  it('links from /agentx to the optimizations index and to each project', () => {
+    cy.visit('/agentx', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentx-optimizations-callout"]').within(() => {
+      cy.contains('h2', 'Optimizations for Agentic Workloads').should('be.visible');
+      cy.contains('50+ upstream pull requests').should('be.visible');
+      cy.get('[data-testid="agentx-optimizations-framework-link"]').should(
+        'have.length',
+        FRAMEWORK_SLUGS.length,
+      );
+      // Direct route into one project, for a reader who knows which engine they run.
+      cy.get('[data-testid="agentx-optimizations-framework-link"][data-framework="sglang"]').should(
+        'have.attr',
+        'href',
+        '/agentx/optimizations/sglang',
+      );
+      cy.get('[data-testid="agentx-optimizations-cta"]')
+        .should('contain.text', 'Read the optimizations')
+        .click();
+    });
+
+    cy.location('pathname').should('eq', '/agentx/optimizations');
+  });
+
+  it('publishes the index with the ecosystem primer, every project, and sourced PRs', () => {
+    cy.visit('/agentx/optimizations', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentx-optimizations-index"]').within(() => {
+      cy.get('h1').should('have.text', 'Optimizations for Agentic Workloads');
+      cy.get('[data-testid="agentx-optimizations-card"]').should(
+        'have.length',
+        FRAMEWORK_SLUGS.length,
+      );
+      cy.get('[data-testid="agentx-optimizations-section-ecosystem"]').should(
+        'contain.text',
+        'A brief introduction to the distributed inference ecosystem',
+      );
+      cy.get('[data-testid="agentx-optimizations-figure-servingStack"] img').should('be.visible');
+      // The day-zero section keeps its own upstream references.
+      cy.get('[data-testid="agentx-optimizations-section-other"]')
+        .find('[data-testid="agentx-optimizations-pr"]')
+        .first()
+        .should('have.attr', 'href')
+        .and('match', /^https:\/\/github\.com\/vllm-project\/vllm\/pull\/\d+$/u);
+      cy.get('[data-testid="agentx-optimizations-card"][data-framework="vllm"]').click();
+    });
+
+    cy.location('pathname').should('eq', '/agentx/optimizations/vllm');
+  });
+
+  it('renders a project page with its measurements, figures, and PR links', () => {
+    cy.visit('/agentx/optimizations/vllm', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentx-optimizations-article"]')
+      .should('have.attr', 'data-framework', 'vllm')
+      .within(() => {
+        cy.get('h1').should('have.text', 'vLLM');
+        cy.contains('prefix-cache hit rate above 95%').should('be.visible');
+        cy.contains('81.7% higher output throughput').should('be.visible');
+        cy.get('[data-testid="agentx-optimizations-figure-vllmSelectiveRetention"] img').should(
+          'be.visible',
+        );
+        cy.get('[data-testid="agentx-optimizations-pr"]')
+          .should('have.length.greaterThan', 20)
+          .first()
+          .should('have.attr', 'href', 'https://github.com/vllm-project/vllm/pull/43447')
+          .and('contain.text', 'vLLM #43447');
+        // Sibling projects stay one click away.
+        cy.get('[data-testid="agentx-optimizations-card"]').should(
+          'have.length',
+          FRAMEWORK_SLUGS.length - 1,
+        );
+      });
+  });
+
+  it('serves every project page and pairs each with its Chinese sibling', () => {
+    for (const slug of FRAMEWORK_SLUGS) {
+      cy.request(`/agentx/optimizations/${slug}`).its('status').should('eq', 200);
+      cy.request(`/zh/agentx/optimizations/${slug}`).its('status').should('eq', 200);
+    }
+  });
+
+  it('returns 404 for a project that does not exist', () => {
+    cy.request({ url: '/agentx/optimizations/not-a-framework', failOnStatusCode: false })
+      .its('status')
+      .should('eq', 404);
+  });
+
+  it('ships the Simplified Chinese index and project pages', () => {
+    cy.visit('/zh/agentx', { onBeforeLoad: unlockAgenticGate });
+    cy.get('[data-testid="agentx-optimizations-callout"]').within(() => {
+      cy.contains('h2', '面向智能体负载的优化').should('be.visible');
+      cy.get('[data-testid="agentx-optimizations-cta"]')
+        .should('contain.text', '查看优化详情')
+        .and('have.attr', 'href', '/zh/agentx/optimizations');
+    });
+
+    cy.visit('/zh/agentx/optimizations', { onBeforeLoad: unlockAgenticGate });
+    cy.get('[data-testid="agentx-optimizations-index"]').within(() => {
+      cy.get('h1').should('have.text', '面向智能体负载的优化');
+      cy.contains('分布式推理生态简介').should('be.visible');
+      cy.get('[data-testid="agentx-optimizations-card"]').should(
+        'have.length',
+        FRAMEWORK_SLUGS.length,
+      );
+      cy.get('[data-testid="agentx-optimizations-card"][data-framework="lmcache"]').should(
+        'have.attr',
+        'href',
+        '/zh/agentx/optimizations/lmcache',
+      );
+    });
+    cy.get('link[rel="alternate"][hreflang="en"]').should(
+      'have.attr',
+      'href',
+      'https://inferencex.semianalysis.com/agentx/optimizations',
+    );
+
+    cy.visit('/zh/agentx/optimizations/sglang', { onBeforeLoad: unlockAgenticGate });
+    cy.get('[data-testid="agentx-optimizations-article"]').within(() => {
+      // Product names stay English; the prose around them is translated.
+      cy.get('h1').should('have.text', 'SGLang');
+      cy.contains('HiCache offload').should('be.visible');
+      cy.contains('并发 384 下输出吞吐量提升 26.75%').should('be.visible');
+      cy.get('[data-testid="agentx-optimizations-pr"]')
+        .first()
+        .should('have.attr', 'href')
+        .and('match', /^https:\/\/github\.com\/sgl-project\/sglang\/pull\/\d+$/u);
+    });
+  });
+});
diff --git a/packages/app/cypress/e2e/agentx-telemetry.cy.ts b/packages/app/cypress/e2e/agentx-telemetry.cy.ts
new file mode 100644
index 000000000..f3d5fc7eb
--- /dev/null
+++ b/packages/app/cypress/e2e/agentx-telemetry.cy.ts
@@ -0,0 +1,157 @@
+import { keepTelemetryTutorial, unlockAgenticGate } from '../support/e2e';
+
+// This spec owns the telemetry-tutorial card's storage, so it opts out of the
+// global suppression seeded in support/e2e.ts (which also re-seeds on reload).
+keepTelemetryTutorial();
+
+const SECTION_IDS = [
+  'why-per-point-telemetry',
+  'one-curve-per-stack',
+  'point-detail-page',
+  'kv-offload',
+  'request-timeline',
+  'flamegraph',
+];
+
+const FIGURE_KEYS = [
+  'perPointOverview',
+  'chartPointTooltip',
+  'pointDetailFull',
+  'requestTimeline',
+  'flamegraph',
+];
+
+describe('AgentX telemetry tutorial — English page', () => {
+  beforeEach(() => {
+    cy.visit('/agentx/telemetry');
+  });
+
+  it('renders the article, every section, and every figure', () => {
+    cy.get('[data-testid="agentx-telemetry-article"]').should('be.visible');
+    cy.contains('h1', 'Exploring Agentic Workloads: Detailed Telemetry').should('be.visible');
+
+    for (const id of SECTION_IDS) {
+      cy.get(`[data-testid="agentx-telemetry-section-${id}"]`).should('exist');
+      // Every section id is also an anchor target for the on-this-page nav.
+      cy.get(`a[href="#${id}"]`).should('exist');
+    }
+
+    for (const key of FIGURE_KEYS) {
+      cy.get(`[data-testid="agentx-telemetry-figure-${key}"]`)
+        .should('exist')
+        // next/image lazy-loads below the fold, so nothing decodes until the
+        // figure is scrolled into view.
+        .scrollIntoView()
+        .find('img')
+        // naturalWidth is only non-zero once the file actually decodes, so this
+        // fails on a missing or broken asset rather than a missing tag.
+        .should(($img) => {
+          expect($img[0].naturalWidth, `decoded width of ${key}`).to.be.greaterThan(0);
+        });
+    }
+  });
+
+  it('states the eleven-chart claim the detail page has to back', () => {
+    cy.get('[data-testid="agentx-telemetry-article"]')
+      .should('contain.text', '11')
+      .and('contain.text', 'per-point telemetry charts');
+  });
+
+  it('links back to /agentx and out to the agentic results', () => {
+    cy.get('[data-testid="agentx-telemetry-results-cta"]')
+      .should('be.visible')
+      .and('have.attr', 'href', '/inference?i_seq=agentic-traces');
+
+    cy.contains('a', 'AgentX datasets').click();
+    cy.location('pathname').should('eq', '/agentx');
+  });
+});
+
+describe('AgentX telemetry tutorial — Chinese page', () => {
+  it('renders translated copy and keeps the same section anchors', () => {
+    cy.visit('/zh/agentx/telemetry');
+    cy.get('[data-testid="agentx-telemetry-article"]').should('be.visible');
+    cy.contains('h1', '通过详细遥测数据解析智能体负载').should('be.visible');
+    // Anchors are language-neutral so a deep link survives the locale switch.
+    for (const id of SECTION_IDS) {
+      cy.get(`[data-testid="agentx-telemetry-section-${id}"]`).should('exist');
+    }
+    cy.get('[data-testid="agentx-telemetry-results-cta"]').should(
+      'have.attr',
+      'href',
+      '/zh/inference?i_seq=agentic-traces',
+    );
+  });
+});
+
+describe('AgentX telemetry tutorial — entry point on /agentx', () => {
+  it('shows the callout and navigates to the tutorial', () => {
+    cy.visit('/agentx');
+    cy.get('[data-testid="agentx-telemetry-callout"]').should('be.visible');
+    cy.get('[data-testid="agentx-telemetry-cta"]').click();
+    cy.location('pathname').should('eq', '/agentx/telemetry');
+  });
+
+  it('shows the Chinese callout on /zh/agentx', () => {
+    cy.visit('/zh/agentx');
+    cy.get('[data-testid="agentx-telemetry-callout"]').should('be.visible');
+    cy.get('[data-testid="agentx-telemetry-cta"]').click();
+    cy.location('pathname').should('eq', '/zh/agentx/telemetry');
+  });
+});
+
+describe('AgentX telemetry tutorial — popup on the agentic point-detail page', () => {
+  beforeEach(() => {
+    // The detail page's own data is irrelevant here; stub it so the spec is
+    // about the nudge, not about whichever benchmark row happens to exist.
+    cy.intercept('GET', '/api/v1/trace-histograms*', { body: {} });
+    cy.intercept('GET', '/api/v1/trace-server-metrics*', { body: null });
+    cy.intercept('GET', '/api/v1/benchmark-siblings*', { statusCode: 404 });
+    cy.intercept('GET', '/api/v1/request-timeline*', { statusCode: 404 });
+    cy.visit('/inference/agentic/206885', {
+      onBeforeLoad: (win) => {
+        unlockAgenticGate(win);
+        win.localStorage.removeItem('inferencex-agentx-telemetry-tutorial-dismissed');
+      },
+    });
+  });
+
+  it('appears after the charts settle, without a backdrop over them', () => {
+    cy.get('[data-testid="telemetry-tutorial-modal"]', { timeout: 10_000 }).should('be.visible');
+    // A bottom-right card, not a centered dialog — the charts stay usable.
+    cy.get('[data-testid="telemetry-tutorial-modal"]').should('have.attr', 'aria-modal', 'false');
+    cy.get('[data-testid="telemetry-tutorial-modal"]').should(
+      'contain.text',
+      'New to these charts?',
+    );
+  });
+
+  it('is the only card in the corner — dashboard nudges stay off this route', () => {
+    // The detail page sits inside the (dashboard) route group, so the dashboard
+    // NudgeEngine would otherwise mount here too and stack its own bottom-right
+    // cards under the tutorial. `reproducibility` (1.5s timer, no conditions) is
+    // the one that always fires, so it is the canary for the carve-out in
+    // DashboardShell.
+    cy.get('[data-testid="telemetry-tutorial-modal"]', { timeout: 10_000 }).should('be.visible');
+    cy.get('[data-testid="reproducibility-nudge"]').should('not.exist');
+    cy.get('[data-testid="filter-hint-nudge"]').should('not.exist');
+  });
+
+  it('opens the tutorial from its primary action', () => {
+    cy.get('[data-testid="telemetry-tutorial-modal-action"]', { timeout: 10_000 }).click();
+    cy.location('pathname').should('eq', '/agentx/telemetry');
+  });
+
+  it('stays dismissed across a reload', () => {
+    cy.get('[data-testid="telemetry-tutorial-modal-dismiss"]', { timeout: 10_000 }).click();
+    cy.get('[data-testid="telemetry-tutorial-modal"]').should('not.exist');
+    cy.reload();
+    cy.get('[data-testid="detail-view-toggle"]').should('exist');
+    // A `not.exist` assertion passes on its FIRST check, so a timeout would not
+    // outlast the nudge's 2.5s trigger delay — the card could still appear
+    // afterwards and the test would have gone green. Burn the delay first, then
+    // assert absence, so this only passes if dismissal actually persisted.
+    cy.wait(3_500);
+    cy.get('[data-testid="telemetry-tutorial-modal"]').should('not.exist');
+  });
+});
diff --git a/packages/app/cypress/e2e/ai-chart.cy.ts b/packages/app/cypress/e2e/ai-chart.cy.ts
new file mode 100644
index 000000000..f9ba41bc8
--- /dev/null
+++ b/packages/app/cypress/e2e/ai-chart.cy.ts
@@ -0,0 +1,171 @@
+type FixtureRow = Record<string, any> & { metrics: Record<string, number> };
+
+function fixtureRow(
+  rows: FixtureRow[],
+  hardware: 'b200' | 'mi355x',
+  x: number,
+  throughput: number,
+  measuredPower?: number,
+): FixtureRow {
+  const template = rows.find(
+    (row) =>
+      row.hardware === hardware &&
+      row.framework === 'sglang' &&
+      row.spec_method === 'mtp' &&
+      row.disagg === false &&
+      row.isl === 8192 &&
+      row.osl === 1024,
+  );
+  if (!template) throw new Error(`Missing ${hardware} SGLang fixture row`);
+
+  const metrics: Record<string, number> = {
+    ...template.metrics,
+    median_intvty: x,
+    tput_per_gpu: throughput,
+    output_tput_per_gpu: throughput / 2,
+    input_tput_per_gpu: throughput / 2,
+  };
+  if (measuredPower === undefined) delete metrics.avg_power_w;
+  else metrics.avg_power_w = measuredPower;
+
+  return {
+    ...template,
+    conc: x,
+    metrics,
+  };
+}
+
+function interceptGeneration(spec: Record<string, unknown>, rows: FixtureRow[]): void {
+  cy.intercept('GET', '**/api/v1/benchmarks?*', rows).as('benchmarks');
+  cy.intercept('POST', 'https://api.openai.com/v1/chat/completions', (request) => {
+    const systemPrompt = request.body.messages?.[0]?.content ?? '';
+    const content = systemPrompt.includes('chart generation assistant')
+      ? JSON.stringify(spec)
+      : 'Deterministic test summary.';
+    request.reply({ choices: [{ message: { content } }] });
+  }).as('openAi');
+}
+
+function generateChart(): void {
+  cy.get('input[placeholder="OpenAI API Key"]').type('test-api-key', { log: false });
+  cy.get('textarea').type('Generate the requested chart');
+  cy.contains('button', 'Generate Chart').click();
+  cy.wait('@benchmarks');
+}
+
+function benchmarkSpec(overrides: Record<string, unknown>): Record<string, unknown> {
+  return {
+    chartType: 'bar',
+    dataSource: 'benchmarks',
+    model: 'DeepSeek-R1-0528',
+    sequence: '8k/1k',
+    hardwareKeys: ['b200', 'mi355x'],
+    precisions: [],
+    frameworks: ['sglang'],
+    disagg: false,
+    yAxisMetric: 'y_tpPerGpu',
+    yAxisLabel: 'Throughput/Chip',
+    targetInteractivity: 40,
+    sortOrder: 'desc',
+    radarMetrics: null,
+    topN: null,
+    topNDistinctGpus: true,
+    title: 'AI chart regression',
+    description: 'Deterministic fixture data.',
+    ...overrides,
+  };
+}
+
+describe('AI chart metric semantics', () => {
+  beforeEach(() => {
+    cy.visit('/ai-chart');
+  });
+
+  it('excludes unavailable zero cost when selecting the top configuration', () => {
+    cy.fixture<FixtureRow[]>('api/benchmarks.json').then((fixtureRows) => {
+      const rows = [
+        fixtureRow(fixtureRows, 'b200', 40, 0),
+        fixtureRow(fixtureRows, 'mi355x', 40, 10_000),
+      ];
+      interceptGeneration(
+        benchmarkSpec({
+          yAxisMetric: 'y_costh',
+          yAxisLabel: 'Cost per Million Total Tokens',
+          sortOrder: 'asc',
+          topN: 1,
+          topNDistinctGpus: false,
+        }),
+        rows,
+      );
+
+      generateChart();
+
+      cy.get('#ai-chart-bar-export')
+        .should('contain.text', 'MI355X')
+        .and('not.contain.text', 'B200');
+      cy.get('#ai-chart-bar-export .bar').should('have.length', 1);
+    });
+  });
+
+  it('renders lower measured power farther from the radar center', () => {
+    cy.fixture<FixtureRow[]>('api/benchmarks.json').then((fixtureRows) => {
+      const rows = [
+        fixtureRow(fixtureRows, 'b200', 40, 100, 600),
+        fixtureRow(fixtureRows, 'mi355x', 40, 100, 300),
+      ];
+      interceptGeneration(
+        benchmarkSpec({
+          chartType: 'radar',
+          yAxisMetric: 'y_measuredAvgPower',
+          yAxisLabel: 'Measured Average Power per Chip',
+          radarMetrics: ['y_measuredAvgPower'],
+        }),
+        rows,
+      );
+
+      generateChart();
+
+      cy.get('#ai-chart-radar-export .radar-dot')
+        .should('have.length', 2)
+        .then(($dots) => {
+          const b200Radius = Math.hypot(
+            Number($dots.eq(0).attr('cx')),
+            Number($dots.eq(0).attr('cy')),
+          );
+          const mi355xRadius = Math.hypot(
+            Number($dots.eq(1).attr('cx')),
+            Number($dots.eq(1).attr('cy')),
+          );
+          expect(mi355xRadius).to.be.greaterThan(b200Radius);
+        });
+    });
+  });
+
+  it('keeps a missing middle metric as a line gap instead of a zero-valued dot', () => {
+    cy.fixture<FixtureRow[]>('api/benchmarks.json').then((fixtureRows) => {
+      const rows = [
+        fixtureRow(fixtureRows, 'b200', 10, 100, 500),
+        fixtureRow(fixtureRows, 'b200', 20, 200),
+        fixtureRow(fixtureRows, 'b200', 30, 300, 400),
+      ];
+      interceptGeneration(
+        benchmarkSpec({
+          chartType: 'line',
+          hardwareKeys: ['b200'],
+          yAxisMetric: 'y_measuredAvgPower',
+          yAxisLabel: 'Measured Average Power per Chip',
+        }),
+        rows,
+      );
+
+      generateChart();
+
+      cy.get('#ai-chart-line-export .dot-group').should('have.length', 2);
+      cy.get('#ai-chart-line-export .line-path')
+        .invoke('attr', 'd')
+        .should((path) => {
+          expect(path?.match(/M/gu)).to.have.length(2);
+        });
+    });
+  });
+});
diff --git a/packages/app/cypress/e2e/api-documentation.cy.ts b/packages/app/cypress/e2e/api-documentation.cy.ts
index f96d70f81..f22f9e9b5 100644
--- a/packages/app/cypress/e2e/api-documentation.cy.ts
+++ b/packages/app/cypress/e2e/api-documentation.cy.ts
@@ -76,6 +76,6 @@ describe('API documentation', () => {
       .and('have.attr', 'hreflang', 'en');
     cy.get('[data-testid="footer-link-api"]')
       .should('have.attr', 'href', '/zh/api')
-      .and('have.text', 'API 参考文档');
+      .and('have.text', 'API 文档');
   });
 });
diff --git a/packages/app/cypress/e2e/blog.cy.ts b/packages/app/cypress/e2e/blog.cy.ts
index 0af6efc89..351f37fef 100644
--- a/packages/app/cypress/e2e/blog.cy.ts
+++ b/packages/app/cypress/e2e/blog.cy.ts
@@ -28,6 +28,9 @@ describe('Blog', () => {
 
   describe('Blog post page', () => {
     before(() => {
+      cy.intercept('GET', 'https://substack-post-media.s3.amazonaws.com/**', {
+        statusCode: 204,
+      });
       cy.visit('/blog/inferencemax-open-source-inference-benchmarking');
     });
 
diff --git a/packages/app/cypress/e2e/chart-overflow-continuation.cy.ts b/packages/app/cypress/e2e/chart-overflow-continuation.cy.ts
index 5b78f53de..572ddfd1e 100644
--- a/packages/app/cypress/e2e/chart-overflow-continuation.cy.ts
+++ b/packages/app/cypress/e2e/chart-overflow-continuation.cy.ts
@@ -85,6 +85,9 @@ const availability = [
 const visitOverflowChart = (withOverlay: boolean) => {
   cy.intercept('GET', '/api/v1/availability', { body: availability }).as('availability');
   cy.intercept('GET', '/api/v1/benchmarks*', { body: rows(null) }).as('benchmarks');
+  cy.intercept('GET', '/api/v1/workflow-info*', {
+    body: { runs: [], changelogs: [], configs: [] },
+  }).as('workflowInfo');
   if (withOverlay) {
     cy.intercept('GET', '/api/unofficial-run*', {
       body: {
diff --git a/packages/app/cypress/e2e/collectivex.cy.ts b/packages/app/cypress/e2e/collectivex.cy.ts
index a7cda6230..204e7a69d 100644
--- a/packages/app/cypress/e2e/collectivex.cy.ts
+++ b/packages/app/cypress/e2e/collectivex.cy.ts
@@ -43,6 +43,16 @@ const kvDataset = buildDataset({
   ],
   meta: { run_id: '162', generated_at: '2026-08-07T12:20:00Z', source_sha: 'e'.repeat(40) },
 });
+const kvComparisonDataset = buildDataset({
+  shards: [makeRawShard()],
+  kv: [{}],
+  meta: { run_id: '164', generated_at: '2026-08-09T12:20:00Z', source_sha: 'a'.repeat(40) },
+});
+const kvOnlyDataset = buildDataset({
+  shards: [],
+  kv: [{}],
+  meta: { run_id: '163', generated_at: '2026-08-08T12:20:00Z', source_sha: 'f'.repeat(40) },
+});
 const ADMIN_TOKEN_KEY = 'collectivex-admin-token';
 
 function installRuns(bodies: CollectiveXDataset[] = [dataset]) {
@@ -174,7 +184,7 @@ describe('CollectiveX neutral run view', () => {
       .and('not.contain.text', 'Payload rate is derived');
   });
 
-  it('allows both measured kernel modes to be selected together', () => {
+  it('selects normal and low-latency kernel modes by default', () => {
     const withLowLatency = buildDataset({
       shards: [makeRawShard(), makeRawShard({ mode: 'low-latency' }), makeRawShard({ ep: 16 })],
     });
@@ -185,13 +195,6 @@ describe('CollectiveX neutral run view', () => {
     cy.wait('@run');
 
     cy.get('[data-testid="collectivex-mode-toggle"]').should('be.visible');
-    cy.get('[data-testid="collectivex-mode-toggle"] button')
-      .contains('Normal')
-      .should('have.attr', 'aria-pressed', 'true');
-    cy.get('[data-testid="collectivex-main-chart"]').should('contain.text', 'deepep-v2');
-    cy.get('[data-testid="collectivex-explorer-chart"] .line-path').should('have.length', 1);
-
-    cy.get('[data-testid="collectivex-mode-toggle"]').contains('Low-latency').click();
     cy.get('[data-testid="collectivex-mode-toggle"] button[aria-pressed="true"]').should(
       'have.length',
       2,
@@ -214,6 +217,42 @@ describe('CollectiveX neutral run view', () => {
     cy.get('[data-testid="collectivex-explorer-chart"] .line-path').should('have.length', 2);
   });
 
+  it('normalizes runner-pool suffixes in SKU controls and chart labels', () => {
+    const withRunnerPoolSkus = buildDataset({
+      shards: [
+        makeRawShard({ sku: 'b200-nscale' }),
+        makeRawShard({ sku: 'h100-dgxc', mode: 'low-latency' }),
+      ],
+    });
+    installRuns([withRunnerPoolSkus]);
+    installRun(withRunnerPoolSkus);
+    cy.reload();
+    cy.wait('@runs');
+    cy.wait('@run');
+
+    cy.get('[data-testid="chart-legend"]')
+      .should('contain.text', 'B200')
+      .and('contain.text', 'H100')
+      .and('not.contain.text', 'B200-NSCALE')
+      .and('not.contain.text', 'H100-DGXC');
+    cy.get(`[data-testid="collectivex-run-row-${withRunnerPoolSkus.run.run_id}"]`)
+      .should('contain.text', 'B200')
+      .and('contain.text', 'H100')
+      .and('not.contain.text', 'B200-NSCALE')
+      .and('not.contain.text', 'H100-DGXC');
+    cy.get('[data-testid="collectivex-sku-select"]').click();
+    cy.get('[role="option"]').then(($options) => {
+      const labels = [...$options].map((option) => option.textContent?.trim());
+      expect(labels).to.include.members(['B200', 'H100']);
+      expect(labels).not.to.include.members(['B200-NSCALE', 'H100-DGXC']);
+    });
+    cy.contains('[role="option"]', 'B200').click();
+    cy.get('[data-testid="chart-legend"]')
+      .should('contain.text', 'B200')
+      .and('not.contain.text', 'H100');
+    cy.get('[data-testid="collectivex-explorer-chart"] .line-path').should('have.length', 1);
+  });
+
   it('selects the available phase when a partial run only measured prefill', () => {
     const prefill = buildDataset({ shards: [makeRawShard({ phase: 'prefill' })] });
     installRuns([prefill]);
@@ -333,14 +372,51 @@ describe('CollectiveX neutral run view', () => {
     cy.get('[data-testid="collectivex-explorer-chart"] .line-path').should('have.length', 1);
   });
 
-  it('keeps the chart on top and presents the matrix inventory', () => {
+  it('filters the run table by EP and KV suite without hiding mixed-suite runs', () => {
+    installRuns([kvOnlyDataset, kvDataset, dataset]);
+    installRun(kvOnlyDataset, 'kvOnlyRun');
+    cy.reload();
+    cy.wait('@runs');
+    cy.wait('@kvOnlyRun');
+
+    cy.get('[data-testid="collectivex-suite-filter"] [role="tab"][aria-selected="true"]').should(
+      'contain.text',
+      'All',
+    );
+    cy.get(`[data-testid="collectivex-run-row-${kvOnlyDataset.run.run_id}"]`).should('be.visible');
+    cy.get(`[data-testid="collectivex-run-row-${kvDataset.run.run_id}"]`).should('be.visible');
+    cy.get(`[data-testid="collectivex-run-row-${dataset.run.run_id}"]`).should('be.visible');
+
+    cy.get('[data-testid="collectivex-suite-filter"]').contains('[role="tab"]', 'EP').click();
+    cy.get(`[data-testid="collectivex-run-row-${kvOnlyDataset.run.run_id}"]`).should('not.exist');
+    cy.get(`[data-testid="collectivex-run-row-${kvDataset.run.run_id}"]`).should('be.visible');
+    cy.get(`[data-testid="collectivex-run-row-${dataset.run.run_id}"]`).should('be.visible');
+
+    cy.get('[data-testid="collectivex-suite-filter"]').contains('[role="tab"]', 'KV').click();
+    cy.get(`[data-testid="collectivex-run-row-${kvOnlyDataset.run.run_id}"]`).should('be.visible');
+    cy.get(`[data-testid="collectivex-run-row-${kvDataset.run.run_id}"]`).should('be.visible');
+    cy.get(`[data-testid="collectivex-run-row-${dataset.run.run_id}"]`).should('not.exist');
+  });
+
+  it('does not render the matrix case inventory', () => {
     cy.get('[data-testid="collectivex-main-chart"]').should('be.visible');
-    cy.get('[data-testid="collectivex-inventory"]')
-      .should('contain.text', 'Matrix case inventory')
-      .and('contain.text', `${dataset.coverage.length} cases`);
-    cy.get('[data-testid="collectivex-inventory-table"]')
-      .should('contain.text', 'H200-DGXC')
-      .and('contain.text', 'B300');
+    cy.get('[data-testid="collectivex-inventory"]').should('not.exist');
+    cy.get('[data-testid="collectivex-display"]').should(
+      'not.contain.text',
+      'Matrix case inventory',
+    );
+  });
+
+  it('localizes the suite filter on the Chinese route', () => {
+    cy.visit('/zh/collectivex');
+    cy.wait('@runs');
+    cy.wait('@run');
+
+    cy.get('[data-testid="collectivex-suite-filter"]')
+      .should('have.attr', 'aria-label', '按测试套件筛选 CollectiveX 运行')
+      .and('contain.text', '全部')
+      .and('contain.text', 'EP')
+      .and('contain.text', 'KV');
   });
 });
 
@@ -581,6 +657,44 @@ describe('CollectiveX kv-transfer card', () => {
     });
   });
 
+  it('plots the fixed frontier axes and keeps multi-run lines aligned with the legend', () => {
+    installRuns([kvComparisonDataset, kvDataset]);
+    installRun(kvComparisonDataset);
+    installRun(kvDataset, 'comparisonKvRun');
+    openCollectiveX();
+
+    cy.get(`[data-testid="collectivex-run-visible-${kvDataset.run.run_id}"]`).check();
+    cy.wait('@comparisonKvRun');
+    cy.get('[data-testid="collectivex-kv-xaxis-toggle"]').contains('button', 'Frontier').click();
+
+    cy.get('[data-testid="collectivex-kv-metric-toggle"]').should('not.exist');
+    cy.get('[data-testid="collectivex-kv-frontier-chart"]')
+      .should('contain.text', 'Aggregate pull bandwidth at p50 (GB/s, log)')
+      .and('contain.text', 'Burst p95 latency per in-flight request (ms, log)');
+    cy.get('[data-testid="collectivex-kv-frontier-chart"] .line-path')
+      .should('have.length', 2)
+      .then(($lines) => {
+        expect([...$lines].map((line) => line.getAttribute('stroke-dasharray'))).to.have.members([
+          'none',
+          '9 4',
+        ]);
+      });
+  });
+
+  it('localizes the frontier control and chart copy on the Chinese page', () => {
+    installRuns([kvDataset]);
+    installRun(kvDataset);
+    cy.visit('/zh/collectivex');
+    cy.wait('@runs');
+    cy.wait('@run');
+
+    cy.get('[data-testid="collectivex-kv-xaxis-toggle"]').contains('button', '帕累托前沿').click();
+    cy.get('[data-testid="collectivex-kv-frontier-chart"]')
+      .should('contain.text', 'p50 聚合 pull 带宽（GB/s，对数）')
+      .and('contain.text', '每个在途请求的突发 p95 延迟（ms，对数）')
+      .and('contain.text', '越靠右下越优');
+  });
+
   it('renders no kv card and no KV suite badge for an EP-only run', () => {
     installRuns();
     installRun();
diff --git a/packages/app/cypress/e2e/compare-per-dollar-table.cy.ts b/packages/app/cypress/e2e/compare-per-dollar-table.cy.ts
index e21a0baa5..f1134c682 100644
--- a/packages/app/cypress/e2e/compare-per-dollar-table.cy.ts
+++ b/packages/app/cypress/e2e/compare-per-dollar-table.cy.ts
@@ -1,3 +1,5 @@
+import { TCO_SOURCE_TITLE, TCO_SOURCE_URL } from '@semianalysisai/inferencex-constants';
+
 describe('Compare-per-dollar slug page — slimmed table + cross-link', () => {
   before(() => {
     cy.window().then((win) => {
@@ -59,6 +61,12 @@ describe('Compare-per-dollar slug page — slimmed table + cross-link', () => {
     cy.contains('Performance per Dollar').should('be.visible');
   });
 
+  it('attributes pricing to the shared TCO source', () => {
+    cy.get('[data-testid="compare-per-dollar-pricing"]')
+      .contains('a', TCO_SOURCE_TITLE)
+      .should('have.attr', 'href', TCO_SOURCE_URL);
+  });
+
   it('renders an indexable comparison PNG with descriptive alt text', () => {
     cy.get('[data-testid="compare-per-dollar-indexed-image"] img')
       .should('be.visible')
@@ -73,6 +81,15 @@ describe('Compare-per-dollar slug page — slimmed table + cross-link', () => {
   });
 });
 
+describe('Chinese compare-per-dollar table', () => {
+  it('localizes the per-dollar metric label', () => {
+    cy.visit('/zh/compare-per-dollar/deepseek-r1-gb200-vs-h100');
+    cy.get('[data-testid="compare-interpolated-table"] tbody')
+      .should('contain.text', '每百万 token 美元成本')
+      .and('not.contain.text', 'Dollar per Million Tokens');
+  });
+});
+
 describe('Compare slug page — cross-link to per-dollar view', () => {
   before(() => {
     cy.window().then((win) => {
diff --git a/packages/app/cypress/e2e/compare-precision.cy.ts b/packages/app/cypress/e2e/compare-precision.cy.ts
index 9b07b388d..411c62ced 100644
--- a/packages/app/cypress/e2e/compare-precision.cy.ts
+++ b/packages/app/cypress/e2e/compare-precision.cy.ts
@@ -5,14 +5,74 @@ describe('Compare precision index page', () => {
     });
   });
 
-  it('renders the /compare index with precision and spec-decode CTA links', () => {
+  it('leads the /compare index with AgentX results and keeps fixed-sequence tools', () => {
     cy.visit('/compare');
+    cy.get('[data-testid="compare-agentx-primary"]').within(() => {
+      cy.get('h1').should('have.text', 'Compare Realistic Agentic Inference Perf');
+      cy.get('[data-testid^="compare-agentx-model-"]').should('have.length', 6);
+      cy.get('[data-testid="compare-agentx-model-kimi-k3"]').should(
+        'have.attr',
+        'href',
+        '/inference/kimi-k3',
+      );
+      cy.get('[data-testid="compare-agentx-overview-link"]')
+        .should('contain.text', 'Overview')
+        .and('have.attr', 'href', '/overview');
+      cy.get('[data-testid="compare-agentx-dashboard-link"]')
+        .should('contain.text', 'Full dashboard')
+        .and('have.attr', 'href', '/inference/kimi-k3');
+      cy.get('[data-testid="compare-agentx-methodology-link"]').should('not.exist');
+    });
+    cy.get('[data-testid="compare-model-catalog"]')
+      .should('contain.text', 'AgentX and 8K→1K results')
+      .and('contain.text', 'Each card identifies its scenario');
     cy.get('[data-testid="compare-index-precision-link"]')
       .should('have.attr', 'href', '/compare-precision')
       .and('contain.text', 'Compare precisions');
     cy.get('[data-testid="compare-index-spec-decode-link"]')
       .should('have.attr', 'href', '/compare-spec-decode')
       .and('contain.text', 'Compare speculative decoding');
+    cy.get('#deepseek-v4 a[data-scenario="AgentX"]')
+      .first()
+      .should('contain.text', 'AgentX')
+      .and('have.attr', 'href')
+      .and('match', /\/agentic$/u);
+    cy.get('#deepseek-r1 a[data-scenario="8K→1K"]')
+      .first()
+      .should('contain.text', '8K→1K')
+      .and('have.attr', 'href')
+      .and('match', /\/8k-1k$/u);
+    // Vendor logos render beside each hardware label in the pair cards:
+    // NVIDIA uses the full-color green mark, AMD its monochrome brand mark.
+    cy.get('a[data-scenario] img[src="/logos/nvidia-color.svg"]').should('exist');
+    cy.get('a[data-scenario] img[src="/logos/amd.svg"]').should('exist');
+  });
+
+  it('ships the same AgentX-first hierarchy on the Simplified Chinese page', () => {
+    cy.visit('/zh/compare');
+    cy.get('[data-testid="compare-agentx-primary"]').within(() => {
+      cy.get('h1').should('have.text', '真实智能体工作负载下的推理性能对比');
+      cy.get('[data-testid^="compare-agentx-model-"]').should('have.length', 6);
+      cy.get('[data-testid="compare-agentx-model-deepseek-v4"]').should(
+        'have.attr',
+        'href',
+        '/zh/inference/deepseek-v4',
+      );
+      cy.get('[data-testid="compare-agentx-overview-link"]')
+        .should('contain.text', '总览')
+        .and('have.attr', 'href', '/zh/overview');
+      cy.get('[data-testid="compare-agentx-dashboard-link"]')
+        .should('contain.text', '查看完整仪表板')
+        .and('have.attr', 'href', '/zh/inference/kimi-k3');
+      cy.get('[data-testid="compare-agentx-methodology-link"]').should('not.exist');
+    });
+    cy.get('[data-testid="compare-model-catalog"]').should('contain.text', 'AgentX 与 8K→1K 结果');
+    cy.get('#deepseek-v4 a[data-scenario="AgentX"]')
+      .first()
+      .should('have.attr', 'href')
+      .and('match', /^\/zh\/compare\/.+\/agentic$/u);
+    cy.get('a[data-scenario] img[src="/logos/nvidia-color.svg"]').should('exist');
+    cy.get('a[data-scenario] img[src="/logos/amd.svg"]').should('exist');
   });
 
   it('renders the /compare-per-dollar index with precision and spec-decode CTA links', () => {
diff --git a/packages/app/cypress/e2e/compare-scenario-routes.cy.ts b/packages/app/cypress/e2e/compare-scenario-routes.cy.ts
new file mode 100644
index 000000000..994210bd4
--- /dev/null
+++ b/packages/app/cypress/e2e/compare-scenario-routes.cy.ts
@@ -0,0 +1,121 @@
+/**
+ * The workload is addressable as a path segment — `/compare/<slug>/agentic`,
+ * `/compare/<slug>/8k-1k` — not only as `?i_seq=`. The bare slug keeps
+ * rendering the pair's default workload.
+ */
+const AGENTX_SLUG = 'deepseek-v4-b200-vs-h200';
+const FIXED_SEQ_SLUG = 'deepseek-r1-h100-vs-h200';
+
+describe('Compare scenario routes', () => {
+  beforeEach(() => {
+    cy.window().then((win) => {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+    });
+  });
+
+  it('serves the agentic workload at its own path with the AgentX explainer', () => {
+    cy.visit(`/compare/${AGENTX_SLUG}/agentic`);
+
+    cy.get('[data-testid="compare-agentic-intro"]')
+      .should('contain.text', 'AgentX replays real coding-agent sessions')
+      .find('[data-testid="compare-agentic-intro-link"]')
+      .should('contain.text', 'Learn more about AgentX')
+      .and('have.attr', 'href', '/agentx');
+  });
+
+  it('serves the fixed-sequence workload at its own path, without the explainer', () => {
+    cy.visit(`/compare/${AGENTX_SLUG}/8k-1k`);
+
+    cy.get('[data-testid="compare-narrative"]').should('exist');
+    cy.get('[data-testid="compare-agentic-intro"]').should('not.exist');
+  });
+
+  it('lets the path segment override the pair default', () => {
+    // Which workload a bare slug defaults to is `pickPairDefaults`, unit-tested
+    // in compare-pair-defaults.test.ts against rows this fixture set does not
+    // carry. What belongs here is the routing contract: whatever the default
+    // is, a scenario segment replaces it.
+    cy.visit(`/compare/${AGENTX_SLUG}`);
+    cy.get('[data-testid="compare-narrative"]')
+      .invoke('text')
+      .then((bare) => {
+        cy.visit(`/compare/${AGENTX_SLUG}/8k-1k`);
+        cy.get('[data-testid="compare-narrative"]').should('contain.text', '8k/1k');
+        cy.get('[data-testid="compare-narrative"]')
+          .invoke('text')
+          .should((pinned) => {
+            expect(pinned, 'segment changes the rendered workload').to.not.equal(bare);
+          });
+      });
+  });
+
+  it('shows no AgentX explainer for a model with no AgentX data', () => {
+    cy.visit(`/compare/${FIXED_SEQ_SLUG}`);
+    cy.get('[data-testid="compare-agentic-intro"]').should('not.exist');
+  });
+
+  it('canonicalizes every scenario view onto the bare slug URL', () => {
+    // The segments are views of one comparison; without a shared canonical the
+    // default view would appear twice in the index.
+    cy.visit(`/compare/${AGENTX_SLUG}/8k-1k`);
+    cy.get('link[rel="canonical"]')
+      .should('have.attr', 'href')
+      .and('match', new RegExp(`/compare/${AGENTX_SLUG}$`, 'u'));
+  });
+
+  it('404s an unknown scenario segment instead of silently defaulting', () => {
+    cy.request({ url: `/compare/${AGENTX_SLUG}/bogus`, failOnStatusCode: false })
+      .its('status')
+      .should('eq', 404);
+  });
+
+  it('keeps the segment when redirecting a non-canonical slug', () => {
+    cy.request({
+      url: `/compare/deepseek-v4-h200-vs-b200/8k-1k`,
+      followRedirect: false,
+      failOnStatusCode: false,
+    }).then((res) => {
+      expect(res.status).to.eq(308);
+      expect(res.headers.location).to.eq(`/compare/${AGENTX_SLUG}/8k-1k`);
+    });
+  });
+
+  it('carries the scenario routes across the other compare families', () => {
+    for (const base of ['/compare-per-dollar', '/zh/compare']) {
+      cy.request({ url: `${base}/${AGENTX_SLUG}/agentic`, failOnStatusCode: false })
+        .its('status')
+        .should('eq', 200);
+    }
+  });
+
+  it('links catalog cards to scenario paths rather than query strings', () => {
+    for (const base of ['/compare', '/compare-per-dollar']) {
+      cy.visit(base);
+      cy.get('#deepseek-v4 a[data-scenario="AgentX"]')
+        .first()
+        .should('have.attr', 'href')
+        .and('match', new RegExp(`^${base}/.+/agentic$`, 'u'));
+      cy.get('#deepseek-r1 a[data-scenario="8K→1K"]')
+        .first()
+        .should('have.attr', 'href')
+        .and('match', new RegExp(`^${base}/.+/8k-1k$`, 'u'));
+    }
+  });
+
+  it('splits the per-dollar catalog by scenario the way /compare does', () => {
+    cy.visit('/compare-per-dollar');
+    cy.contains('Models with AgentX data open long-context').should('be.visible');
+    cy.get('a[data-scenario="AgentX"]').should('have.length.greaterThan', 0);
+    cy.get('a[data-scenario="8K→1K"]').should('have.length.greaterThan', 0);
+  });
+
+  it('ships the Chinese explainer on the Chinese scenario route', () => {
+    cy.visit(`/zh/compare/${AGENTX_SLUG}/agentic`);
+
+    cy.get('[data-testid="compare-agentic-intro"]')
+      .should('contain.text', 'AgentX 回放真实的编码 agent 会话')
+      .find('[data-testid="compare-agentic-intro-link"]')
+      .should('contain.text', '进一步了解 AgentX')
+      .and('have.attr', 'href', '/zh/agentx');
+  });
+});
diff --git a/packages/app/cypress/e2e/csv-export-overlay.cy.ts b/packages/app/cypress/e2e/csv-export-overlay.cy.ts
index b43295c38..0d0034c9e 100644
--- a/packages/app/cypress/e2e/csv-export-overlay.cy.ts
+++ b/packages/app/cypress/e2e/csv-export-overlay.cy.ts
@@ -35,9 +35,10 @@ function readCapturedCsv(): Cypress.Chainable<string> {
 describe('Inference CSV export with an unofficial-run overlay', () => {
   before(() => {
     interceptOverlayRun();
-    // The agentic default mode is E2E Normalized Interactivity (which suppresses overlays and
-    // fetches derived metrics) — stub the fetch, then switch to the
-    // Interactivity mode this suite's overlay assertions rely on.
+    // Agentic charts default to Interactivity, where the overlay renders. The
+    // derived-metrics fetch now happens only under E2E Normalized Interactivity,
+    // so the stub below is a guard against a stray request rather than a
+    // dependency of this suite.
     interceptDerivedAgenticMetrics();
     cy.visit(`/inference?unofficialrun=${OVERLAY_RUN_ID}&i_seq=agentic-traces&i_pctl=p90`, {
       onBeforeLoad(win) {
@@ -47,12 +48,11 @@ describe('Inference CSV export with an unofficial-run overlay', () => {
       },
     });
     cy.wait('@unofficialRun');
-    // Interactivity is nested under the Advanced menu on agentic charts.
-    cy.get('[data-testid="x-axis-mode-advanced"]').click();
-    cy.get('[data-testid="x-axis-mode-interactivity"]').click();
-    cy.get('[data-testid="x-axis-mode-advanced"]')
-      .should('have.attr', 'data-state', 'active')
-      .and('contain.text', 'Interactivity');
+    // Every x-axis metric is a top-level tab on agentic charts, and Interactivity
+    // is the default — clicked anyway so the suite does not depend on that.
+    cy.get('[data-testid="x-axis-mode-interactivity"]')
+      .click()
+      .should('have.attr', 'data-state', 'active');
     cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should('exist');
   });
 
diff --git a/packages/app/cypress/e2e/custom-user-values.cy.ts b/packages/app/cypress/e2e/custom-user-values.cy.ts
index a45db8f9c..e5b7a6c34 100644
--- a/packages/app/cypress/e2e/custom-user-values.cy.ts
+++ b/packages/app/cypress/e2e/custom-user-values.cy.ts
@@ -24,7 +24,7 @@ describe('Custom User Values', () => {
   describe('Custom Chip Costs', () => {
     it('renders the custom costs input section when custom cost metric is selected', () => {
       selectCustomCostMetric();
-      cy.get('[data-testid="custom-costs-section"]').should('be.visible');
+      cy.get('[data-testid="custom-costs-section"]').scrollIntoView().should('be.visible');
       cy.get('[data-testid="custom-costs-section"]').should('contain.text', 'Custom Chip Costs');
     });
 
@@ -98,7 +98,7 @@ describe('Custom User Values', () => {
   describe('Custom Chip Powers', () => {
     it('renders the custom powers input section when custom power metric is selected', () => {
       selectCustomPowerMetric();
-      cy.get('[data-testid="custom-powers-section"]').should('be.visible');
+      cy.get('[data-testid="custom-powers-section"]').scrollIntoView().should('be.visible');
       cy.get('[data-testid="custom-powers-section"]').should('contain.text', 'Custom Chip Powers');
     });
 
diff --git a/packages/app/cypress/e2e/datasets-distributions.cy.ts b/packages/app/cypress/e2e/datasets-distributions.cy.ts
index 0d2a7789a..a5759ce90 100644
--- a/packages/app/cypress/e2e/datasets-distributions.cy.ts
+++ b/packages/app/cypress/e2e/datasets-distributions.cy.ts
@@ -19,6 +19,20 @@ const distribution = (values: {
   },
 });
 
+const LAYOUT_DATASET = {
+  id: 'layout-dataset',
+  slug: 'layout-dataset',
+  label: 'Layout dataset',
+  variant: 'full',
+  description: null,
+  hf_url: null,
+  license: 'apache-2.0',
+  conversation_count: 0,
+  summary: {},
+  chart_data: {},
+  ingested_at: '2026-06-23T00:00:00Z',
+};
+
 describe('Dataset distribution percentiles', () => {
   before(() => {
     cy.intercept('GET', '/api/v1/datasets/test-dataset', {
@@ -87,7 +101,7 @@ describe('Dataset distribution percentiles', () => {
     cy.intercept('GET', '/api/v1/datasets/test-dataset/conversations*', {
       body: { total: 0, items: [] },
     });
-    cy.visit('/datasets/test-dataset', { onBeforeLoad: unlockAgenticGate });
+    cy.visit('/agentx/test-dataset', { onBeforeLoad: unlockAgenticGate });
   });
 
   it('shows P50/P75/P90/P95 for ISL, OSL, and uncached input', () => {
@@ -133,3 +147,28 @@ describe('Dataset distribution percentiles', () => {
     }
   });
 });
+
+describe('Dataset detail loading stability', () => {
+  for (const path of ['/agentx/layout-dataset', '/zh/agentx/layout-dataset']) {
+    it(`keeps the footer below the viewport while ${path} loads`, () => {
+      cy.intercept('GET', '/api/v1/datasets/layout-dataset', {
+        delay: 1500,
+        body: LAYOUT_DATASET,
+      }).as('dataset');
+      cy.intercept('GET', '/api/v1/datasets/layout-dataset/conversations*', {
+        body: { total: 0, items: [] },
+      });
+
+      cy.visit(path, { onBeforeLoad: unlockAgenticGate });
+      cy.get('[data-testid="dataset-detail-loading"]').should('be.visible');
+      cy.get('[data-testid="footer"]').then(($footer) => {
+        cy.window().then((win) => {
+          expect($footer[0].getBoundingClientRect().top).to.be.at.least(win.innerHeight);
+        });
+      });
+
+      cy.wait('@dataset');
+      cy.contains('h1', 'Layout dataset').should('be.visible');
+    });
+  }
+});
diff --git a/packages/app/cypress/e2e/datasets-flamegraph-time.cy.ts b/packages/app/cypress/e2e/datasets-flamegraph-time.cy.ts
index bdb1adfc0..5a4e8df22 100644
--- a/packages/app/cypress/e2e/datasets-flamegraph-time.cy.ts
+++ b/packages/app/cypress/e2e/datasets-flamegraph-time.cy.ts
@@ -95,7 +95,7 @@ describe('Dataset conversation flamegraph timing', () => {
         },
       },
     });
-    cy.visit('/datasets/test-dataset/conversations/conversation-1', {
+    cy.visit('/agentx/test-dataset/conversations/conversation-1', {
       onBeforeLoad: unlockAgenticGate,
     });
   });
diff --git a/packages/app/cypress/e2e/datasets-methodology.cy.ts b/packages/app/cypress/e2e/datasets-methodology.cy.ts
new file mode 100644
index 000000000..e9fccbb20
--- /dev/null
+++ b/packages/app/cypress/e2e/datasets-methodology.cy.ts
@@ -0,0 +1,165 @@
+import { unlockAgenticGate } from '../support/e2e';
+
+const DATASET = {
+  id: 'agentx-v1-full',
+  slug: 'agentx-v1-full',
+  label: 'AgentX v1.0 (full)',
+  variant: 'full',
+  description: 'The complete AgentX v1.0 replay set.',
+  hf_url: null,
+  license: 'apache-2.0',
+  conversation_count: 393,
+  summary: {
+    medianRequestsPerConversation: 20,
+    meanRequestsPerConversation: 24,
+    mainTurns: 7800,
+    subagentGroups: 1300,
+    cachedPct: 0.84,
+    totalIn: 56_000_000,
+    totalOut: 175_000,
+  },
+  ingested_at: '2026-06-21T00:00:00Z',
+};
+
+describe('AgentX dataset methodology', () => {
+  beforeEach(() => {
+    cy.intercept('GET', '/api/v1/datasets', { statusCode: 200, body: [DATASET] });
+  });
+
+  it('explains the source, replay sequence, controls, and interpretation in English', () => {
+    cy.visit('/agentx', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentx-methodology"]').within(() => {
+      cy.get('h1').should('have.text', 'AgentX Benchmark Datasets');
+      cy.get('[data-testid="agentx-methodology-step"]').should('have.length', 4);
+      cy.get('[data-testid="agentx-methodology-step"] h3').then(($headings) => {
+        expect([...$headings].map((heading) => heading.textContent)).to.deep.equal([
+          'Capture',
+          'Transform',
+          'Reconstruct',
+          'Replay and measure',
+        ]);
+      });
+      cy.contains('393 Claude Code sessions').should('be.visible');
+      cy.contains('Claude Code 2.1.139 or newer').should('be.visible');
+      cy.contains('SPEED-Bench').should('be.visible');
+      cy.contains('Its synthetic payloads do not support model-quality evaluation.').should(
+        'be.visible',
+      );
+      cy.get('[data-testid="agentx-methodology-cta"]')
+        .should('contain.text', 'Read the full methodology')
+        .and('have.attr', 'href', '/agentx/methodology');
+      cy.get('[data-testid="agentx-results-cta"]')
+        .should('contain.text', 'View AgentX Performance Results')
+        .and('have.attr', 'href', '/inference');
+    });
+  });
+
+  it('ships the same methodology on the Simplified Chinese page', () => {
+    cy.visit('/zh/agentx', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentx-methodology"]').within(() => {
+      cy.get('h1').should('have.text', 'AgentX 基准测试数据集');
+      cy.get('[data-testid="agentx-methodology-step"]').should('have.length', 4);
+      cy.contains('AgentX 如何构建一次回放').should('be.visible');
+      cy.contains('回放控制').should('be.visible');
+      cy.contains('参与者自愿提供的 Claude Code 会话').should('be.visible');
+      cy.contains('会话数').should('be.visible');
+      cy.contains('单请求 input token 数中位数').should('be.visible');
+      cy.contains('单请求 output token 数中位数').should('be.visible');
+      cy.contains('对外报告的指标仅覆盖一小时 profiling 窗口').should('be.visible');
+      cy.contains('合成 payload 不能用于模型质量评估').should('be.visible');
+      cy.contains('对于没有标准 DRAM 配置的服务器，可用 DRAM 上限为 3 TB').should('be.visible');
+      cy.get('[data-testid="agentx-methodology-cta"]')
+        .should('contain.text', '查看完整测试方法')
+        .and('have.attr', 'href', '/zh/agentx/methodology');
+      cy.get('[data-testid="agentx-results-cta"]')
+        .should('contain.text', '查看 AgentX 性能结果')
+        .and('have.attr', 'href', '/zh/inference');
+    });
+
+    cy.get('a[href="/zh/agentx/agentx-v1-full"]').within(() => {
+      cy.contains('会话数').should('be.visible');
+      cy.contains('单会话请求数中位数').should('be.visible');
+      cy.contains('单会话平均请求数').should('be.visible');
+      cy.contains('main agent 轮次').should('be.visible');
+      cy.contains('subagent 组').should('be.visible');
+      cy.contains('cached input 占比').should('be.visible');
+      cy.contains('input token 总数').should('be.visible');
+      cy.contains('output token 总数').should('be.visible');
+    });
+  });
+
+  it('publishes the full English methodology with sourced figures and locale pairing', () => {
+    cy.visit('/agentx/methodology', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentx-methodology-article"]').within(() => {
+      cy.get('h1').should('have.text', 'AgentX Methodology');
+      cy.contains('393 sessions built on June 21, 2026').should('be.visible');
+      cy.contains('directed acyclic graph (DAG)').should('be.visible');
+      cy.contains('following one-hour profiling window').should('be.visible');
+      cy.contains('capped at 3 TB').should('be.visible');
+      cy.get('figure[data-testid^="agentx-methodology-figure-"]').should('have.length', 21);
+      cy.get('[data-testid="agentx-methodology-figure-corpus"]')
+        .should('contain.text', 'View full-resolution image')
+        .parent('a')
+        .should('have.attr', 'href', '/images/agentx-methodology/corpus-scale.png')
+        .and('have.attr', 'target', '_blank');
+      cy.get('[data-testid="agentx-methodology-figure-corpus"] img')
+        .invoke('attr', 'src')
+        .should('include', 'q=100');
+      [
+        'requestDistributions256k',
+        'subagentDistributions256k',
+        'replayJoined',
+        'replayFlatspawn',
+        'replaySidecars',
+      ].forEach((figure) => {
+        cy.get(`[data-testid="agentx-methodology-figure-${figure}"] img`)
+          .should('be.visible')
+          .invoke('attr', 'alt')
+          .should('not.be.empty');
+      });
+      cy.get('a[href="https://arxiv.org/abs/2604.09557"]').should('exist');
+      cy.contains('mostly vibe coded').should('not.exist');
+      cy.contains('Distillation is bad').should('not.exist');
+    });
+
+    cy.get('[data-testid="language-toggle"]').should('have.attr', 'href', '/zh/agentx/methodology');
+    cy.get('link[rel="alternate"][hreflang="en"]').should('exist');
+    cy.get('link[rel="alternate"][hreflang="zh-CN"]').should('exist');
+  });
+
+  it('publishes the natural Simplified Chinese methodology sibling', () => {
+    cy.visit('/zh/agentx/methodology', { onBeforeLoad: unlockAgenticGate });
+
+    cy.get('[data-testid="agentx-methodology-article"]').within(() => {
+      cy.get('h1').should('have.text', 'AgentX 方法论');
+      cy.contains('2026 年 6 月 21 日构建').should('be.visible');
+      cy.contains('有向无环图（DAG）').should('be.visible');
+      cy.contains('上限为 3 TB').should('be.visible');
+      cy.get('figure[data-testid^="agentx-methodology-figure-"]').should('have.length', 21);
+      cy.get('[data-testid="agentx-methodology-figure-corpus"]')
+        .should('contain.text', '查看原始分辨率图片')
+        .parent('a')
+        .should('have.attr', 'href', '/images/agentx-methodology/corpus-scale.png');
+    });
+
+    cy.get('[data-testid="language-toggle"]').should('have.attr', 'href', '/agentx/methodology');
+  });
+
+  it('permanently redirects legacy dataset routes without dropping path or query', () => {
+    cy.request({
+      url: '/datasets/test-set/conversations/abc?turn=3',
+      followRedirect: false,
+    }).then((response) => {
+      expect(response.status).to.eq(308);
+      expect(response.headers.location).to.eq('/agentx/test-set/conversations/abc?turn=3');
+    });
+
+    cy.request({ url: '/zh/datasets', followRedirect: false }).then((response) => {
+      expect(response.status).to.eq(308);
+      expect(response.headers.location).to.eq('/zh/agentx');
+    });
+  });
+});
diff --git a/packages/app/cypress/e2e/dropdown-switching.cy.ts b/packages/app/cypress/e2e/dropdown-switching.cy.ts
index 4f6f32058..1af072681 100644
--- a/packages/app/cypress/e2e/dropdown-switching.cy.ts
+++ b/packages/app/cypress/e2e/dropdown-switching.cy.ts
@@ -25,6 +25,11 @@ describe('Dropdown one-click switching', () => {
   });
 
   it('only one MultiSelect content panel is open at a time when switching dropdowns', () => {
+    // The default model is FP4-only in the fixtures, which hides the Precision
+    // control — switch to a multi-precision model so both dropdowns exist.
+    cy.visit('/inference?g_model=DeepSeek-R1-0528');
+    cy.get('[data-testid="inference-chart-display"]').should('exist');
+
     cy.get('[data-testid="model-selector"]').click();
     cy.get('[data-slot="select-content"]').should('have.length', 1);
 
@@ -44,6 +49,21 @@ describe('Dropdown one-click switching', () => {
     cy.get('[data-slot="select-content"]').should('not.exist');
   });
 
+  it('marks the featured AgentX models with a NEW pill in the dropdown', () => {
+    cy.get('[data-testid="model-selector"]').click();
+
+    // A featured AgentX model carries the pill… (MiniMax M3 rather than the
+    // Kimi K3 default because the availability fixtures don't ship kimik3 rows)
+    cy.contains('[role="option"]', 'MiniMax M3 428B')
+      .find('[data-new-badge="model-option"]')
+      .should('be.visible')
+      .and('have.text', 'NEW');
+    // …while non-featured models render without one.
+    cy.contains('[role="option"]', 'DeepSeek R1 0528 671B')
+      .find('[data-new-badge="model-option"]')
+      .should('not.exist');
+  });
+
   it('separates maintenance-mode models from deprecated models', () => {
     cy.get('[data-testid="model-selector"]').click();
 
diff --git a/packages/app/cypress/e2e/fixed-sequence-logs.cy.ts b/packages/app/cypress/e2e/fixed-sequence-logs.cy.ts
new file mode 100644
index 000000000..590505536
--- /dev/null
+++ b/packages/app/cypress/e2e/fixed-sequence-logs.cy.ts
@@ -0,0 +1,55 @@
+import { unlockAgenticGate } from '../support/e2e';
+
+describe('Fixed-sequence benchmark logs', () => {
+  beforeEach(() => {
+    cy.intercept('GET', '/api/v1/server-log-files*', {
+      body: ['watchtower-teal-cn03_decode_w0.out', 'results/benchmark.log'],
+    }).as('logFiles');
+    cy.intercept({ method: 'GET', pathname: '/api/v1/server-log' }, (request) => {
+      const params = new URL(request.url).searchParams;
+      const fileName = params.get('file') ?? 'watchtower-teal-cn03_decode_w0.out';
+      request.reply({
+        body: {
+          id: 96255,
+          fileName,
+          serverLog: fileName.endsWith('.out')
+            ? 'INFO decode worker ready\n'
+            : 'benchmark complete\n',
+          offset: 0,
+          nextOffset: null,
+        },
+      });
+    }).as('logContent');
+  });
+
+  it('opens all stored .out and .log files on the fixed-sequence detail route', () => {
+    cy.visit('/inference/logs/96255', { onBeforeLoad: unlockAgenticGate });
+    cy.wait('@logFiles');
+    cy.wait('@logContent');
+
+    cy.contains('h1', 'Benchmark logs').should('be.visible');
+    cy.get('[data-testid="agentic-server-log-viewer"]')
+      .should('have.attr', 'data-log-context', 'fixed-sequence')
+      .and('contain.text', 'watchtower-teal-cn03_decode_w0.out');
+    cy.get('[data-testid="server-log-content"]').should('contain.text', 'INFO decode worker ready');
+
+    cy.get('#agentic-log-file').click();
+    cy.contains('[role="option"]', 'results/benchmark.log').click();
+    cy.get('[data-testid="server-log-content"]').should('contain.text', 'benchmark complete');
+    cy.get('[data-testid="download-selected-server-log"]')
+      .should('have.attr', 'href')
+      .and('include', 'id=96255')
+      .and('include', 'download=1');
+  });
+
+  it('provides the same viewer on the Simplified Chinese route', () => {
+    cy.visit('/zh/inference/logs/96255', { onBeforeLoad: unlockAgenticGate });
+    cy.wait('@logFiles');
+    cy.wait('@logContent');
+
+    cy.contains('h1', '基准测试日志').should('be.visible');
+    cy.get('[data-testid="agentic-server-log-viewer"]')
+      .should('contain.text', '日志文件')
+      .and('contain.text', '搜索所有日志文件');
+  });
+});
diff --git a/packages/app/cypress/e2e/fleet-lifecycle.cy.ts b/packages/app/cypress/e2e/fleet-lifecycle.cy.ts
new file mode 100644
index 000000000..30635edc2
--- /dev/null
+++ b/packages/app/cypress/e2e/fleet-lifecycle.cy.ts
@@ -0,0 +1,995 @@
+import {
+  availability as agenticAvailability,
+  b300Rows as agenticB300Rows,
+} from '../support/overlay-fixtures';
+import { unlockAgenticGate } from '../support/e2e';
+
+type CsvCaptureWindow = Cypress.AUTWindow & {
+  __capturedLifecycleCsvBlob?: Blob;
+};
+
+function captureLifecycleCsvDownloads(win: Cypress.AUTWindow): void {
+  const captureWindow = win as CsvCaptureWindow;
+  win.URL.createObjectURL = (object: Blob | MediaSource) => {
+    if (object instanceof win.Blob) captureWindow.__capturedLifecycleCsvBlob = object;
+    return 'blob:calculator-lifecycle-csv-test';
+  };
+  win.HTMLAnchorElement.prototype.click = () => {};
+}
+
+function readCapturedLifecycleCsv(): Cypress.Chainable<string> {
+  return cy.window().then((win) => {
+    const blob = (win as CsvCaptureWindow).__capturedLifecycleCsvBlob;
+    expect(blob, 'captured lifecycle CSV Blob').to.be.instanceOf(win.Blob);
+    return blob!.text();
+  });
+}
+
+// Fleet Lifecycle section on its own /fleet page. The distinguishing behaviours, all
+// worth locking down:
+//  - it reads the FULL run history, so a row's operating point can come from an
+//    earlier date than the caption's run date — and must say which;
+//  - clamped (never-measured-at-target) reads are excluded rather than clamped,
+//    so chips can legitimately be absent and must be explained, not dropped;
+//  - the token price defaults to the cheapest visible chip's break-even;
+//  - agentic traces have no ISL/OSL, so history is keyed on benchmark_type
+//    instead, and their cached input tokens bill at a discount.
+
+/**
+ * This figure's export menu button. Scoped: the bar chart above has its own, and
+ * `export-button` is not unique across the page.
+ */
+const exportMenu = () =>
+  cy.get('[data-testid="calculator-lifecycle-figure"]').find('[data-testid="export-button"]');
+
+/** Switch the section to the table tab. Idempotent — clicking the active tab is a no-op. */
+const showTable = () => cy.get('[data-testid="calculator-lifecycle-table-view-btn"]').click();
+
+/** Switch the section back to the chart tab. */
+const showChart = () => cy.get('[data-testid="calculator-lifecycle-chart-view-btn"]').click();
+
+/**
+ * Put the section back on the chart tab if it has one rendered.
+ *
+ * `testIsolation` is false, so all of these tests share one page and one tab
+ * state. Without this, a test that reads the table leaves the next one looking
+ * for a chart that is not mounted — an ordering coupling that fails in whichever
+ * order the file happens to be written in. Tolerant of the tab not existing yet:
+ * when the power budget is cleared there is no figure at all.
+ */
+const resetToChart = () =>
+  cy.get('body').then(($body) => {
+    if ($body.find('[data-testid="calculator-lifecycle-chart-view-btn"]').length > 0) showChart();
+  });
+
+/**
+ * A named cell of the first lifecycle table row. Switches to the table tab first:
+ * the section opens on the chart, and these assertions are about the numbers.
+ *
+ * Resolved by header text rather than by position on purpose — these were once
+ * bare indices, and inserting a column silently re-pointed nine assertions at
+ * their neighbours, each of which then failed on a value that was correct.
+ */
+const firstRowCell = (header: string) => {
+  showTable();
+  return cy.get('[data-testid="calculator-lifecycle-table"] thead th').then(($ths) => {
+    const index = $ths.toArray().findIndex((th) => (th.textContent ?? '').includes(header));
+    expect(index, `a column headed "${header}"`).to.be.greaterThan(-1);
+    return cy
+      .get('[data-testid="calculator-lifecycle-table"] tbody tr')
+      .first()
+      .find('td')
+      .eq(index);
+  });
+};
+
+/** The time-axis tick labels, as one string — changes when the x domain moves. */
+const xAxisTicks = () =>
+  cy.get('[data-testid="calculator-lifecycle-chart-svg"] .x-axis .tick text').invoke('text');
+
+/** The rect that owns hover for the whole plot area. */
+const plotOverlay = () =>
+  cy.get('[data-testid="calculator-lifecycle-chart-svg"] .proximity-overlay');
+
+/** Move the cursor to a fraction of the way across the plot. */
+const hoverPlot = (fraction: number) =>
+  plotOverlay().then(($rect) => {
+    const width = $rect[0]!.getBoundingClientRect().width;
+    plotOverlay().trigger('mousemove', width * fraction, 40);
+  });
+
+/** Click at a fraction of the way across the plot. */
+const clickPlot = (fraction: number) =>
+  plotOverlay().then(($rect) => {
+    const width = $rect[0]!.getBoundingClientRect().width;
+    plotOverlay().click(width * fraction, 40);
+  });
+
+/**
+ * Plot x of the last step dot. Steps are matched to hover positions by proximity,
+ * so this is the x a reader points at to ask about a measured config.
+ */
+const stepX = () =>
+  cy
+    .get('[data-testid="calculator-lifecycle-chart-svg"] .dot-group')
+    .last()
+    .invoke('attr', 'transform')
+    .then((transform) => {
+      const x = Number(/translate\((?<x>[-\d.]+)/u.exec(String(transform))?.groups?.x);
+      expect(x, 'a numeric dot x').to.be.a('number');
+      return x;
+    });
+
+/** This chart's tooltip: a portal on <body>, keyed by chart id. */
+const readout = () => cy.get('[data-chart-tooltip="fleet-lifecycle"]');
+
+/**
+ * A table money cell as a number. Cells are compact ($1.2M, -$430k), so comparing
+ * their text lexically is meaningless — these have to be parsed to be compared.
+ */
+const money = (text: string): number => {
+  const match = /^(?<sign>-?)\$(?<value>[\d.]+)(?<suffix>[kMB])?/u.exec(text.trim());
+  if (!match?.groups) throw new Error(`not a money cell: ${text}`);
+  const scale = { k: 1e3, M: 1e6, B: 1e9 }[match.groups.suffix ?? ''] ?? 1;
+  return Number(match.groups.value) * scale * (match.groups.sign ? -1 : 1);
+};
+
+/** All text in the chart SVG, including the axis labels. */
+const chartText = () => cy.get('[data-testid="calculator-lifecycle-chart-svg"]').invoke('text');
+
+/** Vertex count of the first plotted line — a sampled rollout curve has many. */
+const lineVertices = () =>
+  cy
+    .get('[data-testid="calculator-lifecycle-chart-svg"] path.line-path')
+    .first()
+    .invoke('attr', 'd')
+    .then((d) => (d ?? '').split('L').length);
+
+describe('Fleet — Fleet Lifecycle', () => {
+  before(() => {
+    cy.window().then((win) => {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+      // The reproducibility nudge is on a 1.5s timer, and showing a toast
+      // re-renders the chart tree — which rewrites the portal tooltip's inline
+      // styles and wipes an open hover readout. That is app-wide behaviour, not
+      // something this section can fix, so suppress the nudge rather than race it.
+      win.sessionStorage.setItem('inferencex-reproducibility-nudge-shown', '1');
+    });
+    cy.visit('/fleet');
+    // Readiness: the lifecycle section only mounts once run data has loaded.
+    cy.get('[data-testid="calculator-lifecycle-section"]').should('exist');
+  });
+
+  beforeEach(resetToChart);
+
+  it('defaults the power budget to 10 MW and renders the lifecycle', () => {
+    cy.get('[data-testid="calculator-lifecycle-section"]')
+      .should('be.visible')
+      .and('contain.text', 'Fleet Lifecycle');
+    cy.get('[data-testid="calc-fleet-mw-input"]').should('have.value', '10');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-empty"]').should('not.exist');
+  });
+
+  it('the default MW budget drives the lifecycle table', () => {
+    // The figure opens on the chart; the table is the other tab.
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('not.exist');
+    showTable();
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-table"]').within(() => {
+      cy.contains('th', 'Config Now').should('exist');
+      cy.contains('th', 'First Run').should('exist');
+      cy.contains('th', 'Latest Best').should('exist');
+      cy.contains('th', 'Improvements').should('exist');
+      // Absorbed from the old Fleet Projection table: the physical sizing the
+      // economics rest on, on the same row as the economics.
+      cy.contains('th', 'Chips').should('exist');
+      cy.contains('th', 'Concurrent Users now').should('exist');
+      cy.contains('th', 'Margin $/day').should('exist');
+      cy.contains('th', 'Payback').should('exist');
+      cy.contains('th', 'Cumulative Margin').should('exist');
+      cy.get('tbody tr').should('have.length.greaterThan', 0);
+    });
+    cy.get('[data-testid="calculator-lifecycle-empty"]').should('not.exist');
+    // No search box over one row per chip — every one of them is named in the
+    // legend and on the chart already.
+    cy.get('[data-testid="calculator-lifecycle-table"] input[type="text"]').should('not.exist');
+    // No cached-input assumption on a fixed sequence: those runs record no cache
+    // hits at all, so the control would be a knob that moves nothing.
+    cy.get('[data-testid="calc-lifecycle-cache-input"]').should('not.exist');
+    // Both prices are seeded as a pair, output at 4x input — roughly where the
+    // major vendors price. Asserted here, on a page nothing has edited yet: the
+    // spec shares one page across tests, so a later assertion would be reading
+    // whatever ratio an earlier test happened to leave behind.
+    cy.get('[data-testid="calc-lifecycle-price-input"]')
+      .invoke('val')
+      .then((input) => {
+        expect(Number(input)).to.be.greaterThan(0);
+        cy.get('[data-testid="calc-lifecycle-output-price-input"]')
+          .invoke('val')
+          .should((output) => {
+            expect(Number(output)).to.be.closeTo(Number(input) * 4, Number(input) * 0.01);
+          });
+      });
+  });
+
+  it('quotes both token prices in the caption, and tracks the fields', () => {
+    // Every margin plotted is linear in these two, and on the margin metrics they
+    // are seeded from break-even rather than typed — so without them in the
+    // caption two screenshots of different scenarios are indistinguishable.
+    cy.get('[data-testid="calc-lifecycle-price-input"]').clear().type('12');
+    cy.get('[data-testid="calc-lifecycle-output-price-input"]').clear().type('48');
+    // A whole-dollar price reads as `$12`, not `$12.0000` — the fields pad for
+    // typing, the caption is prose. Cents survive; trailing zeros do not.
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('not.contain', '$12.0000');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').within(() => {
+      cy.contains('$12 in / $48 out per M tok').should('be.visible');
+      // Beside the MW figure, not instead of it.
+      cy.contains('MW').should('be.visible');
+    });
+
+    // It is the figure's caption, so it survives the tab switch with the rest.
+    showTable();
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should(
+      'contain',
+      '$12 in / $48 out per M tok',
+    );
+    showChart();
+
+    cy.get('[data-testid="calc-lifecycle-price-input"]').clear().type('7.25');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').within(() => {
+      cy.contains('$7.25 in / $48 out per M tok').should('be.visible');
+    });
+
+    cy.get('[data-testid="calc-lifecycle-price-input"]').clear().type('7');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').within(() => {
+      cy.contains('$7 in / $48 out per M tok').should('be.visible');
+      cy.contains('$12 in').should('not.exist');
+    });
+
+    // `testIsolation: false`, so hand the prices back to break-even before
+    // leaving — later tests read the seeded values and a stale 7 makes them fail
+    // for a reason that has nothing to do with what they assert. The caption
+    // following the reset is the last thing worth checking anyway.
+    cy.get('[data-testid="calc-lifecycle-price-reset"]').click();
+    cy.get('[data-testid="calc-lifecycle-price-input"]')
+      .invoke('val')
+      .then((seeded) => {
+        // The field pads to four decimals; the caption drops trailing zeros, so
+        // the expectation has to apply the same trim rather than echo the field.
+        const trimmed = String(seeded).includes('.')
+          ? String(seeded).replace(/\.?0+$/u, '')
+          : String(seeded);
+        cy.get('[data-testid="calculator-lifecycle-figure"]').should('contain', `$${trimmed} in /`);
+      });
+  });
+
+  it('puts the chart and the table on their own tabs, with a header on both', () => {
+    // The header is the figure's own caption, so it survives the switch — a reader
+    // on the table tab still sees what model, scenario and target they are reading.
+    cy.get('[data-testid="calculator-lifecycle-figure"]').within(() => {
+      cy.contains('h2', 'Fleet Lifecycle').should('be.visible');
+      cy.contains('tok/s/user').should('be.visible');
+    });
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('not.exist');
+
+    showTable();
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"]').should('not.exist');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').within(() => {
+      cy.contains('h2', 'Fleet Lifecycle').should('be.visible');
+    });
+
+    showChart();
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"]').should('be.visible');
+  });
+
+  it('offers a download, and disables the PNG where a PNG is meaningless', () => {
+    exportMenu().should('be.visible').click();
+    cy.get('[data-testid="export-csv-button"]').should('be.visible');
+    // On the chart a PNG is a real artefact.
+    cy.get('[data-testid="export-png-button"]').should('not.have.attr', 'aria-disabled', 'true');
+    cy.get('body').type('{esc}');
+
+    showTable();
+    exportMenu().click();
+    // A PNG of a paginated HTML table is not; the CSV is the export for that view.
+    cy.get('[data-testid="export-png-button"]').should('have.attr', 'aria-disabled', 'true');
+    cy.get('[data-testid="export-csv-button"]').should('be.visible');
+    cy.get('body').type('{esc}');
+  });
+
+  it('every row is traceable to the dated run it came from', () => {
+    // The caption's run date no longer describes these numbers, so each row
+    // carries its own date and links its run.
+    firstRowCell('First Run')
+      .invoke('text')
+      .should('match', /\d{4}-\d{2}-\d{2}/u);
+    firstRowCell('Latest Best')
+      .invoke('text')
+      .should('match', /\d{4}-\d{2}-\d{2}/u);
+  });
+
+  it('defaults the token price to a break-even figure', () => {
+    cy.get('[data-testid="calc-lifecycle-price-input"]')
+      .invoke('val')
+      .should('match', /^\d+(?:\.\d+)?$/u);
+    cy.get('[data-testid="calc-lifecycle-price-input"]')
+      .invoke('val')
+      .then((val) => expect(Number(val)).to.be.greaterThan(0));
+    // Break-even means the cheapest chip earns nothing: some row is at/near zero
+    // margin, and none of the visible fleet is comfortably profitable there.
+    showTable();
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('contain.text', 'Never');
+  });
+
+  it('raising the price above break-even turns margin positive and produces a payback', () => {
+    cy.get('[data-testid="calc-lifecycle-price-input"]').clear();
+    cy.get('[data-testid="calc-lifecycle-price-input"]').type('50');
+    // Margin $/day column: positive, so no leading minus.
+    firstRowCell('Margin $/day').invoke('text').should('match', /^\$/u);
+    firstRowCell('Payback')
+      .invoke('text')
+      .should('match', /\d+(?:\.\d+)? mo/u);
+  });
+
+  it('editing the price offers a reset back to break-even', () => {
+    cy.get('[data-testid="calc-lifecycle-price-reset"]').should('be.visible').click();
+    cy.get('[data-testid="calc-lifecycle-price-input"]')
+      .invoke('val')
+      .then((val) => expect(Number(val)).to.be.lessThan(50));
+  });
+
+  it('bills the two streams apart — the output price alone moves margin', () => {
+    // On 8k/1k output is a ninth of the tokens, so a blended single price could
+    // not respond to this at all.
+    cy.get('[data-testid="calc-lifecycle-price-input"]').clear().type('10');
+    cy.get('[data-testid="calc-lifecycle-output-price-input"]').clear().type('10');
+    firstRowCell('Margin $/day')
+      .invoke('text')
+      .then((before) => {
+        cy.get('[data-testid="calc-lifecycle-output-price-input"]').clear().type('500');
+        firstRowCell('Margin $/day')
+          .invoke('text')
+          .should((after) => {
+            expect(money(after)).to.be.greaterThan(money(before));
+          });
+      });
+  });
+
+  it('resets both prices together, keeping the ratio the user set', () => {
+    // Break-even with two prices is a line, so the reset has to pick a point on
+    // it: the one that preserves the ratio currently in the fields.
+    cy.get('[data-testid="calc-lifecycle-price-input"]').clear().type('10');
+    cy.get('[data-testid="calc-lifecycle-output-price-input"]').clear().type('30');
+    cy.get('[data-testid="calc-lifecycle-price-reset"]').click();
+    cy.get('[data-testid="calc-lifecycle-price-input"]')
+      .invoke('val')
+      .then((input) => {
+        expect(Number(input)).to.be.greaterThan(0);
+        cy.get('[data-testid="calc-lifecycle-output-price-input"]')
+          .invoke('val')
+          .should((output) => {
+            // 3x in, 3x out — not the 4x default, and not one field left behind.
+            expect(Number(output)).to.be.closeTo(Number(input) * 3, Number(input) * 0.01);
+          });
+      });
+  });
+
+  it('renders the lifecycle chart with one line per visible chip', () => {
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] path.line-path').should(
+      'have.length.greaterThan',
+      0,
+    );
+    // The break-even rule is what makes the sign of the margin readable.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should('exist');
+  });
+
+  it('names each line with its chip at the right edge', () => {
+    // One label per line, so a series is identifiable without the legend.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] path.line-path').then((lines) => {
+      cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-series-label').should(
+        'have.length',
+        lines.length,
+      );
+    });
+    // The label text is a chip name, matching the table's first column.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-series-label')
+      .first()
+      .invoke('text')
+      .should('match', /\S/u)
+      .then((label) => {
+        showTable();
+        cy.get('[data-testid="calculator-lifecycle-table"]').should('contain', String(label));
+      });
+  });
+
+  it('reads every plotted chip at the hovered date, not just one point', () => {
+    // Hovering anywhere in the plot has to answer "what is each chip doing here?".
+    // Per-point tooltips could not: comparing two chips at one date meant hovering
+    // twice and holding the first number in your head.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] path.line-path').then((lines) => {
+      // Late in the window, where every chip has been measured — a chip first
+      // measured mid-window has no line, and so no row, before that.
+      hoverPlot(0.9);
+      cy.get('[data-testid="calculator-lifecycle-chart-svg"] .ruler-group').should(
+        'have.css',
+        'display',
+        'block',
+      );
+      cy.get('[data-testid="calculator-lifecycle-chart-svg"] .vertical-ruler').should('exist');
+      // One row per line, each with a money figure, under the hovered date.
+      // Not `:visible` — Cypress treats a fixed-position element as hidden when
+      // something else answers elementFromPoint at its centre, and an unfrozen
+      // readout sets pointer-events: none, so the plot answers instead.
+      readout()
+        .should('have.css', 'display', 'block')
+        .and('contain.text', '$')
+        .invoke('text')
+        .should('match', /\d{2} \w{3} \d{4}/u);
+      readout().find('tbody tr').should('have.length', lines.length);
+    });
+    // Earlier than the newest chip's first measurement, that chip has no number
+    // to give, and the readout says nothing rather than inventing one.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] path.line-path').then((lines) => {
+      hoverPlot(0.05);
+      readout()
+        .find('tbody tr')
+        .should('have.length.greaterThan', 0)
+        .and('have.length.lessThan', lines.length);
+    });
+  });
+
+  it('freezes the readout on click, and releases it on the next click', () => {
+    hoverPlot(0.4);
+    readout()
+      .invoke('text')
+      .then((hovered) => {
+        // The date is what identifies a readout; the rest of the text legitimately
+        // changes on freeze, since freezing adds any step's config detail.
+        const frozenDate = /\d{2} \w{3} \d{4}/u.exec(hovered)?.[0];
+        expect(frozenDate, 'a date in the hover readout').to.be.a('string');
+        clickPlot(0.4);
+        readout()
+          .should('have.css', 'pointer-events', 'auto')
+          .and('contain.text', String(frozenDate));
+        // Frozen means frozen: moving the cursor elsewhere must not re-read.
+        hoverPlot(0.7);
+        readout().should('contain.text', String(frozenDate));
+        // The same gesture releases it, so the reader is never stuck with a
+        // stale readout pinned over the lines.
+        clickPlot(0.7);
+        readout().should('have.css', 'display', 'none');
+      });
+  });
+
+  it('keeps config detail out of the hover readout, showing it only once frozen', () => {
+    // Hovering is for scanning: the popup keeps one shape wherever the cursor is.
+    // A step block appearing as the cursor crosses a dot reflows the rows under
+    // the reader and buries the comparison they came for.
+    stepX().then((x) => {
+      plotOverlay().trigger('mousemove', x, 40);
+      readout().should('have.css', 'display', 'block').and('not.contain.text', 'Config');
+      plotOverlay().click(x, 40);
+      readout().should('contain.text', 'Config');
+      // Leave nothing frozen for the next test.
+      plotOverlay().click(x, 40);
+    });
+  });
+
+  it('links the run behind any step, not just the first and last', () => {
+    // The table links only the opening and closing sweeps. Intermediate rungs are
+    // exactly where an anomalous run that was never purged would sit, so freezing
+    // one has to expose its run — otherwise that rung is auditable nowhere.
+    // The hover grid never lands exactly on a step's instant, so steps match by
+    // proximity; this asserts that freezing works from a dot's own x.
+    stepX().then((x) => {
+      plotOverlay().click(x, 40);
+    });
+    readout()
+      .find('a')
+      .should('have.length.greaterThan', 0)
+      .first()
+      .should('have.attr', 'href')
+      .and('match', /^https?:\/\//u);
+    // The pin is left in place: the next test's zoom dismisses it, and the
+    // tooltip is a portal outside the SVG so it does not disturb what follows.
+  });
+
+  it('zooms the time axis, as its own instructions promise', () => {
+    xAxisTicks().then((before) => {
+      cy.get('[data-testid="calculator-lifecycle-chart-svg"]').trigger('wheel', {
+        deltaY: -400,
+        shiftKey: true,
+        clientX: 700,
+        clientY: 300,
+        bubbles: true,
+      });
+      // The axis has to actually move: the chart advertises shift+scroll zoom in
+      // its caption, and for a while it advertised it without wiring it up.
+      xAxisTicks().should('not.equal', before);
+      // Double-click resets, as the same caption promises. Also leaves the chart
+      // unzoomed for the tests that follow — they share this page.
+      cy.get('[data-testid="calculator-lifecycle-chart-svg"]').dblclick();
+      xAxisTicks().should('equal', before);
+      // d3's reset is a ~750ms transition whose trailing events repaint the axes
+      // from the scales captured when it started. Let it finish before the next
+      // test changes the metric, or that repaint lands on top of the new axis.
+      cy.wait(900);
+    });
+    // The chip labels survive the zoom, one per line, still inside the SVG.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] path.line-path').then((lines) => {
+      cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-series-label').should(
+        'have.length',
+        lines.length,
+      );
+    });
+  });
+
+  it('switches the y axis between margin and revenue', () => {
+    chartText().should('contain', 'Margin ($/day)');
+    // Break-even is meaningful on a margin axis, so the rule is drawn.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should('exist');
+
+    cy.get('[data-testid="calc-lifecycle-metric-revenue"]').click();
+    chartText().should('contain', 'Revenue ($/day)');
+    // Zero is not break-even on a revenue axis — each chip breaks even at its own
+    // cost line — so the rule must not be drawn claiming otherwise.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should(
+      'not.exist',
+    );
+    // Revenue drops the cost term, so nothing plotted can be negative.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .y-axis').should(
+      'not.contain.text',
+      '-$',
+    );
+
+    // Cumulative revenue is a running total in $, not a rate in $/day. Zero is
+    // where every chip starts rather than a threshold anything crosses, so the
+    // break-even rule stays off here too. (The axis does carry one negative tick:
+    // the domain is padded 5% below its floor, which on a tens-of-millions span
+    // rounds to a visible one. That is scale padding, not plotted data.)
+    cy.get('[data-testid="calc-lifecycle-metric-cumulative-revenue"]').click();
+    chartText().should('contain', 'Cumulative Revenue ($)');
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should(
+      'not.exist',
+    );
+
+    cy.get('[data-testid="calc-lifecycle-metric-margin"]').click();
+    chartText().should('contain', 'Margin ($/day)');
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should('exist');
+
+    // Margin per MW is the same quantity in the unit a power-constrained plan is
+    // written in, so it is a positive rescale: zero is still break-even and the
+    // rule stays. The axis ticks must differ from the $/day ones captured just
+    // above — the fleet is several MW, so a metric that was relabelled but never
+    // divided would reproduce them exactly. Comparing against any other metric's
+    // axis would pass without the division ever happening.
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .y-axis')
+      .invoke('text')
+      .then((perDay) => {
+        cy.get('[data-testid="calc-lifecycle-metric-margin-per-mw"]').click();
+        chartText().should('contain', 'Margin ($/MW/day)');
+        cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should(
+          'exist',
+        );
+        cy.get('[data-testid="calculator-lifecycle-chart-svg"] .y-axis')
+          .invoke('text')
+          .should((perMw) => {
+            expect(perMw, 'the per-MW axis is not the $/day axis relabelled').to.not.equal(perDay);
+          });
+      });
+
+    cy.get('[data-testid="calc-lifecycle-metric-margin"]').click();
+    chartText().should('contain', 'Margin ($/day)');
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should('exist');
+  });
+
+  it('a longer horizon accumulates more of whatever the daily margin is', () => {
+    // Direction, not just difference: a regression that moved this the wrong way
+    // would satisfy a bare `not.equal` while inverting the finding. The direction
+    // is not "up" — the price defaults to the cheapest chip's break-even, so the
+    // first row is typically losing money, and more months of a negative rate is a
+    // deeper hole. Reading the sign off the $/day column keeps the assertion true
+    // of the physics rather than of one fixture's profitability.
+    cy.get('[data-testid="calc-lifecycle-horizon-input"]').clear();
+    cy.get('[data-testid="calc-lifecycle-horizon-input"]').type('24');
+    firstRowCell('Margin $/day')
+      .invoke('text')
+      .then((perDay) => {
+        const sign = Math.sign(money(perDay));
+        expect(sign, 'a non-zero daily margin to accumulate').to.not.equal(0);
+        firstRowCell('Cumulative Margin')
+          .invoke('text')
+          .then((short) => {
+            cy.get('[data-testid="calc-lifecycle-horizon-input"]').clear();
+            cy.get('[data-testid="calc-lifecycle-horizon-input"]').type('96');
+            firstRowCell('Cumulative Margin')
+              .invoke('text')
+              .should((long) => {
+                const delta = (money(long) - money(short)) * sign;
+                expect(delta, 'cumulative margin moves with the daily rate').to.be.greaterThan(0);
+              });
+          });
+      });
+  });
+
+  it('tracks measured config improvements as steps, not one flat plateau', () => {
+    showTable();
+    // The point of the section: a chip's revenue follows its optimisation
+    // history, so at least one chip must show more than zero improvements.
+    cy.get('[data-testid="calculator-lifecycle-table"] tbody tr').then(($rows) => {
+      const improvements = [...$rows].map((row) =>
+        Number.parseInt(row.querySelectorAll('td')[4]!.textContent ?? '0', 10),
+      );
+      expect(Math.max(...improvements)).to.be.greaterThan(0);
+    });
+    // Gain over the opening config is reported as a multiple.
+    firstRowCell('Gain')
+      .invoke('text')
+      .should('match', /^\d+(?:\.\d+)?×$/u);
+  });
+
+  it('rolls each config out over the ramp instead of stepping instantly', () => {
+    // Reads both views, so it switches deliberately: `firstRowCell` moves to the
+    // table, and the vertex counts are only meaningful on the chart.
+    cy.get('[data-testid="calc-lifecycle-ramp-input"]')
+      .should('have.value', '0.5')
+      .and('have.attr', 'step', '0.25');
+    lineVertices().then((curved) => {
+      // Rollouts are sampled into curves, so the line cannot be a bare staircase.
+      expect(curved).to.be.greaterThan(20);
+      // A longer ramp reaches each config's numbers later, so it earns less.
+      firstRowCell('Cumulative Margin')
+        .invoke('text')
+        .then((short) => {
+          cy.get('[data-testid="calc-lifecycle-ramp-input"]').clear();
+          cy.get('[data-testid="calc-lifecycle-ramp-input"]').type('18');
+          // "Earns less" is the claim, so assert less — not merely different.
+          firstRowCell('Cumulative Margin')
+            .invoke('text')
+            .should((long) => expect(money(long)).to.be.lessThan(money(short)));
+          // Ramp 0 means configs take effect instantly — a pure staircase, which
+          // needs far fewer vertices than the sampled curves.
+          cy.get('[data-testid="calc-lifecycle-ramp-input"]').clear();
+          cy.get('[data-testid="calc-lifecycle-ramp-input"]').type('0');
+          showChart();
+          lineVertices().should((stepped) => expect(stepped).to.be.lessThan(curved));
+        });
+    });
+  });
+
+  it('a shorter MTBI lowers availability', () => {
+    cy.get('[data-testid="calc-lifecycle-mtbi-input"]').clear();
+    cy.get('[data-testid="calc-lifecycle-mtbi-input"]').type('60');
+    firstRowCell('Availability')
+      .invoke('text')
+      .then((high) => {
+        cy.get('[data-testid="calc-lifecycle-mtbi-input"]').clear();
+        cy.get('[data-testid="calc-lifecycle-mtbi-input"]').type('2');
+        firstRowCell('Availability')
+          .invoke('text')
+          .should((low) => {
+            expect(Number.parseFloat(low)).to.be.lessThan(Number.parseFloat(high));
+          });
+      });
+  });
+
+  it('states the hybrid basis and the unofficial-run exclusion', () => {
+    // Today's TCO rates on a dated operating point, and no overlay support.
+    cy.get('[data-testid="calculator-lifecycle-section"]')
+      .should('contain.text', 'Power and $/chip/hr are today')
+      .and('contain.text', 'Unofficial runs loaded via a run link are not shown here');
+    cy.get('[data-testid="calculator-lifecycle-section"]').should(
+      'contain.text',
+      'SemiAnalysis Datacenter Industry Model',
+    );
+  });
+
+  it('draws one row per chip, pooling its software configs', () => {
+    showTable();
+    // The history holds many hwKeys per chip (b200_sglang, b200_trtllm, …) and
+    // they are one piece of silicon, so the chip must not repeat down the table.
+    cy.get('[data-testid="calculator-lifecycle-table"] tbody tr').then(($rows) => {
+      const chips = [...$rows].map((row) => row.querySelectorAll('td')[0]!.textContent?.trim());
+      expect(chips.length).to.be.greaterThan(0);
+      expect(new Set(chips).size).to.equal(chips.length);
+    });
+    // And each row names the config it ended up on, since that changes along the line.
+    firstRowCell('Config Now').invoke('text').should('not.be.empty');
+  });
+
+  it('follows legend visibility', () => {
+    // The legend lives in the controls card, outside the figure, so it stays
+    // reachable from the table view and even when nothing is plottable.
+    showTable();
+    // Legend entries are configs and clicking one isolates it. Configs sit one
+    // level below the lines now, so isolating a single config leaves at most one
+    // chip — and none at all when that config was never measured at the target,
+    // which is a legitimate state rather than a broken chart. The section-level
+    // assertion covers both: the table shows at most one row, or is absent
+    // entirely in favour of the explanatory message.
+    cy.get('[data-testid="calculator-lifecycle-table"] tbody tr').then(($rows) => {
+      const fullCount = $rows.length;
+      expect(fullCount).to.be.greaterThan(1);
+      cy.get('[data-testid="fleet-legend"] ul label').first().click();
+      cy.get('[data-testid="calculator-lifecycle-section"]').should(($section) => {
+        expect(
+          $section.find('[data-testid="calculator-lifecycle-table"] tbody tr').length,
+        ).to.be.at.most(1);
+      });
+      cy.get('[data-testid="fleet-legend"] ul label').first().click();
+      cy.get('[data-testid="calculator-lifecycle-table"] tbody tr').should(
+        'have.length',
+        fullCount,
+      );
+    });
+  });
+
+  it('explains chips that were never measured at an extreme target instead of dropping them', () => {
+    // Pushing the target to the top of the range leaves chips outside their
+    // measured interactivity; the honest answer is to say so, with the range.
+    cy.get('[data-testid="fleet-controls"] input[type="range"]')
+      .invoke('attr', 'max')
+      .then((max) => {
+        cy.get('[data-testid="fleet-controls"] input[type="number"]').clear();
+        cy.get('[data-testid="fleet-controls"] input[type="number"]').type(`${max}{enter}`);
+      });
+    cy.get('[data-testid="calculator-lifecycle-section"]').should(
+      'contain.text',
+      'Not measured at this interactivity',
+    );
+    cy.get('[data-testid="calculator-lifecycle-unmeasured"]').should('contain.text', 'measured');
+  });
+
+  it('names the budget, not the slider, when a budget is too small to size a fleet', () => {
+    // The two ways to have nothing to plot need different fixes, so they must not
+    // share a message: this one is the budget, and blaming the interactivity
+    // slider would send the reader to the wrong control.
+    //
+    // The previous test leaves the target at the top of the range, where nothing
+    // is measured — which is the *other* cause, and would make `noneMeasured`
+    // the correct answer. So put the target back in a measured range first, or
+    // this asserts nothing about the budget at all.
+    cy.get('[data-testid="fleet-controls"] input[type="number"]').clear();
+    cy.get('[data-testid="fleet-controls"] input[type="number"]').type('35{enter}');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('be.visible');
+
+    cy.get('[data-testid="calc-fleet-mw-input"]').clear();
+    cy.get('[data-testid="calc-fleet-mw-input"]').type('0.0001');
+    cy.get('[data-testid="calculator-lifecycle-none"]')
+      .should('be.visible')
+      .and('contain.text', 'too small to power a single chip');
+  });
+
+  it('clearing the MW budget restores the empty state, and the input survives it', () => {
+    cy.get('[data-testid="calc-fleet-mw-input"]').clear();
+    cy.get('[data-testid="calculator-lifecycle-empty"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('not.exist');
+    // The one control that must never vanish with the body it gates.
+    cy.get('[data-testid="calc-fleet-mw-input"]').should('be.visible');
+  });
+});
+
+describe('Fleet — self-contained lifecycle regressions', () => {
+  before(() => {
+    const rows = agenticB300Rows(null);
+    cy.intercept('GET', '/api/v1/availability', { body: agenticAvailability });
+    cy.intercept('GET', '/api/v1/benchmarks/history*', { body: rows }).as('legacyPriceHistory');
+    cy.intercept('GET', '/api/v1/benchmarks*', { body: rows }).as('legacyPriceBenchmarks');
+    cy.visit('/fleet?g_model=DeepSeek-V4-Pro&i_seq=agentic-traces&i_prec=fp4&c_mw=10&c_price=7', {
+      onBeforeLoad(win) {
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+        win.sessionStorage.setItem('inferencex-reproducibility-nudge-shown', '1');
+        unlockAgenticGate(win);
+      },
+    });
+    cy.wait('@legacyPriceBenchmarks');
+    cy.wait('@legacyPriceHistory');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('be.visible');
+  });
+
+  it('preserves c_price and derives a non-zero output price', () => {
+    cy.get('[data-testid="calc-lifecycle-price-input"]').should('have.value', '7');
+    cy.get('[data-testid="calc-lifecycle-output-price-input"]')
+      .invoke('val')
+      .should((output) => {
+        expect(Number(output)).to.equal(28);
+      });
+  });
+
+  it('plots revenue per MW as a real normalized y-axis metric', () => {
+    cy.get('[data-testid="calc-lifecycle-metric-revenue"]').click();
+    cy.get('[data-testid="calculator-lifecycle-chart-svg"] .y-axis')
+      .invoke('text')
+      .then((revenuePerDay) => {
+        cy.get('[data-testid="calc-lifecycle-metric-revenue-per-mw"]').click();
+        chartText().should('contain', 'Revenue ($/MW/day)');
+        cy.get('[data-testid="calculator-lifecycle-chart-svg"] .lifecycle-zero-rule').should(
+          'not.exist',
+        );
+        cy.get('[data-testid="calculator-lifecycle-chart-svg"] .y-axis')
+          .invoke('text')
+          .should((revenuePerMw) => {
+            expect(revenuePerMw, 'Revenue/MW is not Revenue/day relabelled').to.not.equal(
+              revenuePerDay,
+            );
+          });
+      });
+    cy.get('[data-testid="calc-lifecycle-metric-margin"]').click();
+  });
+
+  it('exports the physical fleet sizing alongside its economics', () => {
+    showTable();
+    cy.window().then(captureLifecycleCsvDownloads);
+    exportMenu().click();
+    cy.get('[data-testid="export-csv-button"]').click();
+
+    readCapturedLifecycleCsv().then((csv) => {
+      const header = csv.split('\n').find((line) => line.startsWith('Chip,'));
+      expect(header, 'lifecycle CSV header').to.equal(
+        'Chip,Config Now,First Run,Latest Best,Improvements,Gain,Chips,tok/s/MW now,Concurrent Users now,Revenue $/day,Cost $/day,Margin $/day,Payback,Cumulative Margin,Availability',
+      );
+    });
+  });
+});
+
+describe('Fleet — Fleet Lifecycle with agentic traces', () => {
+  // Two run dates so the staircase has a step in it, and a measured cache hit
+  // rate on every row so the cached-input discount has something to apply to.
+  // The shared history fixture carries ZERO agentic rows, so this spec serves
+  // its own — the pattern gpu-compare-agentic-detail.cy.ts uses.
+  const LATER_DATE = '2026-07-26';
+  const historyRows = () => {
+    const day1 = agenticB300Rows(null).map((row) => ({
+      ...row,
+      metrics: { ...row.metrics, server_gpu_cache_hit_rate: 0.9 } as Record<string, number>,
+    }));
+    // A later, faster sweep: the improvement the section exists to show.
+    const day2 = day1.map((row) => ({
+      ...row,
+      id: row.id + 5000,
+      date: LATER_DATE,
+      metrics: {
+        ...row.metrics,
+        tput_per_gpu: row.metrics.tput_per_gpu * 1.4,
+        output_tput_per_gpu: row.metrics.output_tput_per_gpu * 1.4,
+        input_tput_per_gpu: row.metrics.input_tput_per_gpu * 1.4,
+      },
+    }));
+    const b300 = [...day1, ...day2];
+    return [...b300, ...b300.map((row) => ({ ...row, id: row.id + 20000, hardware: 'b200' }))];
+  };
+
+  before(() => {
+    // Agentic rows have null ISL/OSL, which the default fixture model has none
+    // of — so serve them the same way the main calculator spec does.
+    const b300Rows = agenticB300Rows(null);
+    const b200Rows = b300Rows.map((row) => ({ ...row, hardware: 'b200' }));
+    cy.intercept('GET', '/api/v1/availability', {
+      body: [
+        ...agenticAvailability,
+        ...agenticAvailability.map((row) => ({ ...row, hardware: 'b200' })),
+        ...agenticAvailability.map((row) => ({ ...row, date: LATER_DATE })),
+      ],
+    });
+    cy.intercept('GET', '/api/v1/benchmarks/history*', { body: historyRows() }).as(
+      'agenticHistory',
+    );
+    cy.intercept('GET', '/api/v1/benchmarks*', { body: [...b300Rows, ...b200Rows] }).as(
+      'agenticBenchmarks',
+    );
+    cy.visit('/fleet?g_model=DeepSeek-V4-Pro&i_seq=agentic-traces&i_prec=fp4', {
+      onBeforeLoad(win) {
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+        unlockAgenticGate(win);
+      },
+    });
+    cy.wait('@agenticBenchmarks');
+    // The agentic fixture exposes a single scenario, so the scenario
+    // control disappears entirely — no dropdown, no static readout.
+    cy.get('[data-testid="scenario-static-value"]').should('not.exist');
+    cy.get('[data-testid="calc-fleet-mw-input"]').should('have.value', '10');
+    cy.wait('@agenticHistory');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('be.visible');
+    showTable();
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('be.visible');
+  });
+
+  it('projects a lifecycle from history that has no ISL/OSL to key on', () => {
+    // The section used to refuse this outright. The endpoint keys agentic
+    // history on benchmark_type instead, so there is a real projection here.
+    cy.get('[data-testid="calculator-lifecycle-unsupported"]').should('not.exist');
+    cy.get('[data-testid="calculator-lifecycle-table"] tbody tr').should(
+      'have.length.greaterThan',
+      0,
+    );
+  });
+
+  it('names the cached tier in the caption, since most input bills at it', () => {
+    // On agentic the input price beside it is not what most tokens actually cost:
+    // the measured hit rate is ~0.9 here, so quoting the input price alone would
+    // overstate what a reader thinks the fleet charges.
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('contain', ', cached 10%');
+    cy.get('[data-testid="calc-lifecycle-cache-input"]').clear().type('35');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('contain', ', cached 35%');
+    cy.get('[data-testid="calc-lifecycle-cache-input"]').clear().type('10');
+  });
+
+  it('raises the break-even price it seeds, because fewer tokens are billable', () => {
+    cy.get('[data-testid="calc-lifecycle-cache-input"]').should('have.value', '10');
+    // While the price is still auto-seeded, the discount does NOT move margin —
+    // break-even re-solves to keep the cheapest fleet at zero either way. What it
+    // moves is the price you need to charge, which is the economically
+    // interesting statement: discounting 90% of a 133:1 input mix means asking
+    // materially more per token to stand still.
+    cy.get('[data-testid="calc-lifecycle-price-input"]')
+      .invoke('val')
+      .then((discountedPrice) => {
+        cy.get('[data-testid="calc-lifecycle-cache-input"]').clear().type('100');
+        cy.get('[data-testid="calc-lifecycle-price-input"]')
+          .invoke('val')
+          .should((fullPrice) => {
+            expect(Number(discountedPrice)).to.be.greaterThan(Number(fullPrice));
+          });
+      });
+  });
+
+  it('lowers margin at a fixed price, where the discount has nowhere to hide', () => {
+    // Take the price over so it stops re-seeding; now the cached-token discount
+    // lands where a reader would expect it — on the margin.
+    cy.get('[data-testid="calc-lifecycle-cache-input"]').clear().type('100');
+    cy.get('[data-testid="calc-lifecycle-price-input"]').clear().type('500');
+    firstRowCell('Margin $/day')
+      .invoke('text')
+      .then((full) => {
+        cy.get('[data-testid="calc-lifecycle-cache-input"]').clear().type('10');
+        firstRowCell('Margin $/day')
+          .invoke('text')
+          .should((discounted) => {
+            expect(money(discounted)).to.be.lessThan(money(full));
+          });
+      });
+  });
+});
+
+describe('Fleet — Fleet Lifecycle in Chinese', () => {
+  before(() => {
+    cy.visit('/zh/fleet', {
+      onBeforeLoad(win) {
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+      },
+    });
+    // Readiness: the lifecycle section only mounts once run data has loaded.
+    cy.get('[data-testid="calculator-lifecycle-section"]').should('exist');
+  });
+
+  it('translates the section, including the table headers and notes', () => {
+    cy.get('[data-testid="calculator-lifecycle-section"]')
+      .should('contain.text', '集群生命周期')
+      .and('contain.text', '设施功率 (MW)');
+    cy.get('[data-testid="calc-fleet-mw-input"]').should('have.value', '10');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('be.visible');
+    // The caption's price line is translated too, units and all.
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('contain', '每百万 token');
+    cy.get('[data-testid="calculator-lifecycle-figure"]').should('contain', '输入 $');
+    // The table headers this test checks are on the other tab now.
+    showTable();
+    cy.get('[data-testid="calculator-lifecycle-table"]').should('be.visible');
+    cy.get('[data-testid="calculator-lifecycle-section"]')
+      .should('contain.text', 'Token 价格')
+      .and('contain.text', '每 MW 收入')
+      .and('contain.text', '当前配置')
+      .and('contain.text', '首次运行')
+      .and('contain.text', '累计利润')
+      .and('contain.text', '来源：');
+    // Nothing from the English table leaks through.
+    cy.get('[data-testid="calculator-lifecycle-section"]')
+      .should('not.contain.text', 'Enter a facility power budget')
+      .and('not.contain.text', 'Cumulative Margin')
+      .and('not.contain.text', 'First Run');
+  });
+});
diff --git a/packages/app/cypress/e2e/gpu-compare-agentic-detail.cy.ts b/packages/app/cypress/e2e/gpu-compare-agentic-detail.cy.ts
index d7ffe31fa..d643a2d44 100644
--- a/packages/app/cypress/e2e/gpu-compare-agentic-detail.cy.ts
+++ b/packages/app/cypress/e2e/gpu-compare-agentic-detail.cy.ts
@@ -1,4 +1,5 @@
 import { interceptDerivedAgenticMetrics, unlockAgenticGate } from '../support/e2e';
+import { agenticMetrics } from '../support/agentic-fixtures';
 
 // ---------------------------------------------------------------------------
 // Spec-scoped fixture helpers
@@ -51,37 +52,6 @@ const agenticAvailability = [
   })),
 ];
 
-// Minimal per-metric percentile ladder matching what the chart expects for
-// agentic rows (median/p75/p90/p95/p99 + std for each family).
-const percentileLadder = (prefix: string, base: number): Record<string, number> => ({
-  [`median_${prefix}`]: base,
-  [`p75_${prefix}`]: base * 1.2,
-  [`p90_${prefix}`]: base * 1.5,
-  [`p95_${prefix}`]: base * 1.7,
-  [`p99_${prefix}`]: base * 2.2,
-  [`std_${prefix}`]: base * 0.3,
-});
-
-const agenticMetrics = (conc: number): Record<string, number> => {
-  const scale = conc / 16;
-  const itl = 0.011 * scale;
-  return {
-    ...percentileLadder('ttft', 0.4 * scale),
-    ...percentileLadder('tpot', 0.012 * scale),
-    ...percentileLadder('itl', itl),
-    ...percentileLadder('e2el', 8 * scale),
-    median_intvty: 1 / itl,
-    p75_intvty: 1 / (itl * 1.2),
-    p90_intvty: 1 / (itl * 1.5),
-    p99_intvty: 1 / (itl * 2.2),
-    std_intvty: (1 / itl) * 0.1,
-    tput_per_gpu: 950 / Math.sqrt(scale),
-    output_tput_per_gpu: 210,
-    input_tput_per_gpu: 740,
-    total_tput_tps: 7600 * conc * 0.05,
-  };
-};
-
 // IDs must be unique numbers — the GPU graph uses them as D3 data keys and
 // trace-availability is keyed on them.
 let benchIdCursor = 800100;
@@ -147,6 +117,14 @@ describe('GPU comparison agentic point detail', () => {
       );
       request.reply({ body: result });
     });
+    cy.intercept('GET', '/api/v1/log-availability*', (request) => {
+      const ids = new URL(request.url).searchParams.get('ids')?.split(',') ?? [];
+      request.reply({
+        body: Object.fromEntries(
+          ids.filter((id) => agenticIds.has(Number(id))).map((id) => [id, true]),
+        ),
+      });
+    }).as('gpuLogAvailability');
     // The agentic default x-axis mode (E2E Normalized Interactivity) fetches derived metrics on
     // mount; without values every point drops out of the (remapped) data set.
     interceptDerivedAgenticMetrics();
@@ -160,6 +138,10 @@ describe('GPU comparison agentic point detail', () => {
 
     cy.get('[data-testid="gpu-multiselect"] [role="combobox"]').click({ force: true });
     cy.get('[role="option"]').first().click();
+    cy.contains('Comparison Date Range').should('be.visible');
+    cy.contains('button', 'Select date range')
+      .should('not.have.class', 'animate-pulse')
+      .and('not.have.class', 'border-red-500');
     cy.contains('button', 'Select date range').click();
     cy.get('body').then(($body) => {
       if ($body.text().includes('View anyway')) {
@@ -172,6 +154,7 @@ describe('GPU comparison agentic point detail', () => {
 
     cy.get('[data-testid="gpu-graph"]').first().should('be.visible');
     cy.wait('@gpuTraceAvailability');
+    cy.wait('@gpuLogAvailability');
     cy.wait(100);
     cy.get('[data-testid="gpu-graph"]')
       .first()
@@ -199,7 +182,26 @@ describe('GPU comparison agentic point detail', () => {
       .then(($link) => {
         expect($link).to.match('a');
         expect($link).not.to.have.attr('target');
-        expect($link.attr('href')).to.match(/^\/inference\/agentic\/\d+$/u);
+        const href = $link.attr('href') ?? '';
+        expect(href).to.match(/^\/inference\/agentic\/\d+\?/u);
+        // The chart state rides along so the detail page can link back to the
+        // view the reader left — see agentic-detail-back-nav.cy.ts. Only
+        // non-default values are carried, so `g_model` (DeepSeek-V4-Pro IS the
+        // default) is absent while the scenario and GPU selection are not.
+        const params = new URLSearchParams(href.slice(href.indexOf('?')));
+        expect(params.get('i_seq')).to.equal('agentic-traces');
+        expect(params.get('i_gpus')).to.be.a('string').and.not.equal('');
+      });
+    cy.get('[data-chart-tooltip]:visible [data-action="view-logs"]')
+      .should('be.visible')
+      .then(($link) => {
+        expect($link).to.match('a');
+        const href = $link.attr('href') ?? '';
+        expect(href).to.match(/^\/inference\/agentic\/\d+\?/u);
+        const params = new URLSearchParams(href.slice(href.indexOf('?')));
+        expect(params.get('view')).to.equal('logs');
+        expect(params.get('i_seq')).to.equal('agentic-traces');
+        expect(params.get('i_gpus')).to.be.a('string').and.not.equal('');
       });
     cy.location('pathname').should('eq', '/inference');
   });
diff --git a/packages/app/cypress/e2e/gradient-labels.cy.ts b/packages/app/cypress/e2e/gradient-labels.cy.ts
index 05df38a46..5fdce43df 100644
--- a/packages/app/cypress/e2e/gradient-labels.cy.ts
+++ b/packages/app/cypress/e2e/gradient-labels.cy.ts
@@ -1,3 +1,5 @@
+import { expandLegendAdvanced } from '../support/legend-advanced';
+
 describe('Gradient Labels Toggle', () => {
   before(() => {
     cy.visit('/inference', {
@@ -8,6 +10,8 @@ describe('Gradient Labels Toggle', () => {
     // Wait for chart to load
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
     cy.get('.sidebar-legend').first().should('be.visible');
+    // Parallelism Labels lives in the collapsed-by-default Advanced drawer.
+    expandLegendAdvanced();
   });
 
   it('Gradient Labels toggle exists in the legend', () => {
@@ -94,6 +98,7 @@ describe('Gradient Labels Toggle', () => {
       },
     });
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
+    expandLegendAdvanced();
     cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'checked');
   });
 });
diff --git a/packages/app/cypress/e2e/historical-trends.cy.ts b/packages/app/cypress/e2e/historical-trends.cy.ts
index 55b0e2749..51c9d5b21 100644
--- a/packages/app/cypress/e2e/historical-trends.cy.ts
+++ b/packages/app/cypress/e2e/historical-trends.cy.ts
@@ -94,7 +94,12 @@ describe('Historical Trends — Content & Interactions', () => {
     cy.get('body').type('{esc}');
   });
 
-  it('precision multi-select is present', () => {
+  it('precision multi-select is present only for multi-precision models', () => {
+    // The default model is FP4-only in the fixtures, so the control is hidden;
+    // a multi-precision model brings it back.
+    cy.get('[data-testid="precision-multiselect"]').should('not.exist');
+    cy.visit('/historical?g_model=DeepSeek-R1-0528');
+    cy.get('[data-testid="historical-trends-display"]').should('be.visible');
     cy.get('[data-testid="precision-multiselect"]').should('be.visible');
   });
 
diff --git a/packages/app/cypress/e2e/inference-chart.cy.ts b/packages/app/cypress/e2e/inference-chart.cy.ts
index 35ca6922e..33bcc96c9 100644
--- a/packages/app/cypress/e2e/inference-chart.cy.ts
+++ b/packages/app/cypress/e2e/inference-chart.cy.ts
@@ -1,3 +1,6 @@
+import { unlockAgenticGate } from '../support/e2e';
+import { interceptOverlayRun, OVERLAY_RUN_ID } from '../support/overlay-fixtures';
+
 describe('Inference Chart', () => {
   before(() => {
     cy.window().then((win) => {
@@ -58,7 +61,9 @@ describe('Inference Chart', () => {
   });
 
   it('renders quick filters and toggles a vendor pill', () => {
-    cy.get('[data-testid="quick-filters"]').should('exist');
+    cy.get('[data-testid="quick-filters-dialog"]').should('not.exist');
+    cy.get('[data-testid="scatter-quick-filters"]').click();
+    cy.get('[data-testid="quick-filters-dialog"]').should('be.visible');
     cy.get('[data-testid="quick-filter-deployment-single-node"]').should('contain', 'Single-node');
     cy.get('[data-testid="quick-filter-deployment-multi-node"]').should('contain', 'Multi-node');
     cy.get('[data-testid="quick-filter-deployment-disagg"]').should('contain', 'Disaggregated');
@@ -69,4 +74,132 @@ describe('Inference Chart', () => {
       .click()
       .should('have.attr', 'aria-pressed', 'false');
   });
+
+  it('plots OpenRouter-priced token revenue for official and unofficial runs', () => {
+    cy.intercept('GET', 'https://openrouter.ai/api/v1/models', {
+      statusCode: 200,
+      body: {
+        data: [
+          {
+            id: 'deepseek/deepseek-v4-pro-0813',
+            pricing: {
+              prompt: '0.000001122',
+              completion: '0.000003366',
+            },
+          },
+        ],
+      },
+    }).as('openRouterPricing');
+    interceptOverlayRun();
+    cy.visit(
+      `/inference?unofficialrun=${OVERLAY_RUN_ID}&i_seq=agentic-traces&i_pctl=p90&i_metric=y_tokenRevenuePerGpuHour&i_revenue=openrouter`,
+      {
+        onBeforeLoad(win) {
+          win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+          unlockAgenticGate(win);
+        },
+      },
+    );
+    cy.wait('@unofficialRun');
+    cy.wait('@openRouterPricing');
+
+    cy.get('[data-testid="yaxis-metric-selector"]').should(
+      'contain.text',
+      'Token Revenue per GPU Hour',
+    );
+    cy.get('[data-testid="token-revenue-price-source"]').should(
+      'contain.text',
+      'OpenRouter current pricing',
+    );
+    cy.get('[data-testid="openrouter-price-summary"]')
+      .should('contain.text', 'Input $1.122/M tok')
+      .and('contain.text', 'Output $3.366/M tok');
+    cy.get('[data-testid="chart-figure"]')
+      .first()
+      .find('[data-testid="token-revenue-subtitle-prices"]')
+      .should('have.text', 'Input $1.122/M tok · Output $3.366/M tok')
+      .then(($prices) => {
+        const subtitle = $prices.parent().text();
+        expect(subtitle.indexOf($prices.text())).to.be.lessThan(subtitle.indexOf('Updated:'));
+      });
+    cy.get('[data-testid="openrouter-pricing-link"]').should(
+      'have.attr',
+      'href',
+      'https://openrouter.ai/deepseek/deepseek-v4-pro-0813',
+    );
+    cy.get('[data-testid="chart-figure"]')
+      .first()
+      .find('h2')
+      .should('contain.text', 'Token Revenue per GPU Hour at OpenRouter Pricing');
+    cy.get('[data-testid="inference-chart-display"] svg .dot-group').should(
+      'have.length.greaterThan',
+      0,
+    );
+    cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(
+      'have.length.greaterThan',
+      0,
+    );
+    cy.get('[data-testid^="axis-metric-row-y-"]').first().click();
+    cy.get('[data-testid^="axis-metric-body-y-"]')
+      .first()
+      .should('contain.text', 'OpenRouter')
+      .and('contain.text', '$/GPU/hr =');
+  });
+
+  it('ships OpenRouter-priced token revenue in Chinese', () => {
+    cy.viewport(390, 844);
+    cy.intercept('GET', 'https://openrouter.ai/api/v1/models', {
+      statusCode: 200,
+      body: {
+        data: [
+          {
+            id: 'deepseek/deepseek-v4-pro-0813',
+            pricing: {
+              prompt: '0.000001122',
+              completion: '0.000003366',
+            },
+          },
+        ],
+      },
+    }).as('openRouterPricingZh');
+    interceptOverlayRun();
+    cy.visit(
+      `/zh/inference?unofficialrun=${OVERLAY_RUN_ID}&i_seq=agentic-traces&i_pctl=p90&i_metric=y_tokenRevenuePerGpuHour&i_revenue=openrouter`,
+      { onBeforeLoad: unlockAgenticGate },
+    );
+    cy.wait('@unofficialRun');
+    cy.wait('@openRouterPricingZh');
+    cy.get('[data-testid="yaxis-metric-selector"]').should(
+      'contain.text',
+      '每 GPU 小时 token 收入',
+    );
+    cy.get('[data-testid="token-revenue-price-source"]').should(
+      'contain.text',
+      'OpenRouter 当前价格',
+    );
+    cy.get('[data-testid="openrouter-price-summary"]')
+      .should('contain.text', '输入 $1.122/百万 token')
+      .and('contain.text', '输出 $3.366/百万 token');
+    cy.get('[data-testid="chart-figure"]')
+      .first()
+      .find('[data-testid="token-revenue-subtitle-prices"]')
+      .should('have.text', '输入 $1.122/百万 token · 输出 $3.366/百万 token')
+      .then(($prices) => {
+        const subtitle = $prices.parent().text();
+        expect(subtitle.indexOf($prices.text())).to.be.lessThan(subtitle.indexOf('更新时间：'));
+      });
+    cy.get('[data-testid="chart-figure"]')
+      .first()
+      .find('h2')
+      .should('contain.text', '按 OpenRouter 价格计算的每 GPU 小时 token 收入');
+  });
+
+  it('surfaces the error instead of an endless skeleton when availability fails', () => {
+    cy.intercept('GET', '/api/v1/availability*', { statusCode: 500, body: {} }).as(
+      'availabilityFailure',
+    );
+    cy.visit('/inference');
+    cy.wait('@availabilityFailure');
+    cy.contains('h2', 'Something went wrong!', { timeout: 10000 }).should('be.visible');
+  });
 });
diff --git a/packages/app/cypress/e2e/inference-replay.cy.ts b/packages/app/cypress/e2e/inference-replay.cy.ts
index 8984b4e37..f44ce302f 100644
--- a/packages/app/cypress/e2e/inference-replay.cy.ts
+++ b/packages/app/cypress/e2e/inference-replay.cy.ts
@@ -1,3 +1,5 @@
+import { expandLegendAdvanced } from '../support/legend-advanced';
+
 const openReplayDialog = () => {
   cy.get('[data-testid="chart-figure"]')
     .first()
@@ -120,23 +122,32 @@ describe('Inference Replay', () => {
   it('re-renders the replay frame when a parent-chart toggle changes', () => {
     cy.get('body').then(($body) => {
       if ($body.find('[data-testid="replay-panel-chart-0"]').length === 0) return;
+      // The log-scale switch lives in the collapsed-by-default Advanced drawer,
+      // so expand it on the parent chart's legend (not the replay panel's).
+      cy.get('[data-testid="legend-advanced-toggle"]').then(($toggles) => {
+        const parentAdvanced = [...$toggles].find(
+          (toggle) => !toggle.closest('[data-testid^="replay-panel-chart-"]'),
+        );
+        if (!parentAdvanced) throw new Error('Parent chart Advanced toggle is missing');
+        if (parentAdvanced.getAttribute('aria-expanded') !== 'true') {
+          cy.wrap(parentAdvanced).click({ force: true });
+        }
+      });
       // Capture the SVG path data for the first roofline as a stable signature.
       cy.get('[data-testid="replay-panel-chart-0"] svg path.roofline-path')
         .first()
         .invoke('attr', 'd')
         .then((beforeD) => {
-          // Toggle the log-scale setting in the underlying inference context —
-          // the replay panel shares state with the parent chart, so the chart
-          // re-renders without us touching the replay UI.
-          cy.window().then((win) => {
-            const url = new URL(win.location.href);
-            const cur = url.searchParams.get('i_log') === '1';
-            url.searchParams.set('i_log', cur ? '0' : '1');
-            win.history.replaceState(null, '', url.toString());
-            // Dispatch a popstate so InferenceContext picks up the change.
-            win.dispatchEvent(new win.PopStateEvent('popstate'));
+          // Toggle the control on the parent chart rather than mutating the address
+          // bar. Dashboard URL state is intentionally snapshotted on load, so a
+          // synthetic popstate is not a supported control update.
+          cy.get('[data-testid="scatter-log-scale"]').then(($toggles) => {
+            const parentToggle = [...$toggles].find(
+              (toggle) => !toggle.closest('[data-testid^="replay-panel-chart-"]'),
+            );
+            if (!parentToggle) throw new Error('Parent chart log-scale toggle is missing');
+            cy.wrap(parentToggle).click({ force: true });
           });
-          cy.wait(400);
           cy.get('[data-testid="replay-panel-chart-0"] svg path.roofline-path')
             .first()
             .invoke('attr', 'd')
@@ -153,6 +164,7 @@ describe('Inference Replay', () => {
       // Enable line labels inside the replay panel (scoped — the parent chart
       // renders the same control behind the dialog).
       cy.get('[data-testid="replay-panel-chart-0"]').within(() => {
+        expandLegendAdvanced();
         cy.get('[data-testid="scatter-line-labels"]').then(($el) => {
           if ($el.attr('data-state') !== 'checked') cy.wrap($el).click();
         });
diff --git a/packages/app/cypress/e2e/jalapeno-preview-notice.cy.ts b/packages/app/cypress/e2e/jalapeno-preview-notice.cy.ts
new file mode 100644
index 000000000..73f99eac3
--- /dev/null
+++ b/packages/app/cypress/e2e/jalapeno-preview-notice.cy.ts
@@ -0,0 +1,131 @@
+const JALAPENO_QUERY =
+  'g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_outputTputPerGpu';
+const JALAPENO_NOTICE = '[data-testid="jalapeno-official-preview-notice"]';
+const VERA_RUBIN_NOTICE = '[data-testid="vera-rubin-official-preview-notice"]';
+
+describe('official preview notices', () => {
+  beforeEach(() => {
+    // Keep this spec independent of Neon response time. The client appends the
+    // supplemental Jalapeño and July Rubin snapshots to these successful empty
+    // responses, which is exactly the path under test here.
+    cy.intercept('GET', '**/api/v1/availability*', { body: [] });
+    cy.intercept('GET', '**/api/v1/benchmarks/history*', { body: [] });
+    cy.intercept('GET', '**/api/v1/benchmarks?*', { body: [] });
+  });
+
+  it('renders both graph notices and tracks Jalapeño legend visibility', () => {
+    cy.visit(`/inference?${JALAPENO_QUERY}`);
+
+    cy.get(JALAPENO_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'InferenceX Official Preview')
+      .and('contain.text', 'Results may change as validation and publication continue.')
+      .and('not.contain.text', '\u2014');
+    cy.get(VERA_RUBIN_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'Vera Rubin (July)')
+      .and('not.contain.text', '\u2014');
+
+    cy.get(
+      '[data-testid="chart-legend"] [role="button"][aria-label^="Hide"][aria-label*="Jalapeño"]',
+    )
+      .first()
+      .click({ force: true });
+    cy.get(JALAPENO_NOTICE).should('not.exist');
+    cy.get(VERA_RUBIN_NOTICE).should('be.visible');
+  });
+
+  it('keeps the July Vera Rubin notice off metrics where its curve is unsupported', () => {
+    cy.visit('/inference?g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_tpPerGpu');
+
+    cy.get(JALAPENO_NOTICE, { timeout: 20000 }).should('be.visible');
+    cy.get(VERA_RUBIN_NOTICE, { timeout: 20000 }).should('not.exist');
+  });
+
+  it('appears with Jalapeño calculator results', () => {
+    cy.visit(`/calculator?${JALAPENO_QUERY}`);
+
+    cy.get(JALAPENO_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'InferenceX Official Preview');
+
+    cy.get('[data-testid="calculator-controls"]').within(() => {
+      cy.get('#calc-cost-type').click();
+    });
+    cy.get('[role="option"]').contains('Output Tokens').click();
+    cy.get('[data-testid="calculator-controls"] input[type="number"]')
+      .first()
+      .clear()
+      .type('100')
+      .blur();
+    cy.get(VERA_RUBIN_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'Vera Rubin (July) results are an official preview');
+  });
+
+  it('appears with Jalapeño historical trend lines', () => {
+    cy.visit(`/historical?${JALAPENO_QUERY}`);
+
+    // The historical view defaults to 35 tok/s/user, below Jalapeño's measured
+    // range, and correctly does not extrapolate. Move into its measured range
+    // so a Jalapeño trend line, and therefore the preview notice, appears.
+    cy.get('[data-testid="historical-trends-display"] input[type="number"]', {
+      timeout: 20000,
+    })
+      .clear()
+      .type('100')
+      .blur();
+    cy.get(JALAPENO_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'InferenceX Official Preview');
+    cy.get(VERA_RUBIN_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'Vera Rubin (July) results are an official preview');
+  });
+
+  it('keeps the July Vera Rubin notice off unsupported historical metrics', () => {
+    cy.visit('/historical?g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_tpPerGpu');
+
+    cy.get('[data-testid="historical-trends-display"] input[type="number"]', {
+      timeout: 20000,
+    })
+      .clear()
+      .type('100')
+      .blur();
+    cy.get(JALAPENO_NOTICE, { timeout: 20000 }).should('be.visible');
+    cy.get(VERA_RUBIN_NOTICE).should('not.exist');
+  });
+
+  it('shows the Vera Rubin preview when its output-token fleet curve is visible', () => {
+    cy.visit('/fleet?g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4');
+
+    cy.get('[data-testid="fleet-cost-type-selector"]').click();
+    cy.get('[role="option"]').contains('Output Tokens').click();
+    cy.get('[data-testid="fleet-legend"]', { timeout: 20000 }).should('contain.text', 'Vera Rubin');
+    cy.get('[data-testid="fleet-controls"] input[type="number"]').clear().type('100').blur();
+    cy.get('[data-testid="calc-fleet-mw-input"]').should('have.value', '10');
+    cy.get(VERA_RUBIN_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'Vera Rubin (July) results are an official preview')
+      .and('not.contain.text', '\u2014');
+  });
+
+  it('uses the Chinese disclaimer on /zh and stays absent without Jalapeño results', () => {
+    cy.visit(`/zh/inference?${JALAPENO_QUERY}`);
+    cy.get(JALAPENO_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'InferenceX 官方预览')
+      .and('contain.text', '随着验证和发布工作的推进，结果可能会调整。');
+    cy.get(VERA_RUBIN_NOTICE, { timeout: 20000 })
+      .should('be.visible')
+      .and('contain.text', 'Vera Rubin (July)')
+      .and('contain.text', '随着验证和发布工作的推进，结果可能会调整。')
+      .and('not.contain.text', '\u2014');
+
+    cy.visit(
+      '/zh/inference?g_model=DeepSeek-V4-Pro&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_outputTputPerGpu',
+    );
+    cy.get(JALAPENO_NOTICE).should('not.exist');
+    cy.get(VERA_RUBIN_NOTICE).should('not.exist');
+  });
+});
diff --git a/packages/app/cypress/e2e/landing-performance.cy.ts b/packages/app/cypress/e2e/landing-performance.cy.ts
index ed580a3d4..c114a7d32 100644
--- a/packages/app/cypress/e2e/landing-performance.cy.ts
+++ b/packages/app/cypress/e2e/landing-performance.cy.ts
@@ -47,7 +47,8 @@ describe('Landing page performance', () => {
     cy.viewport(412, 823);
     cy.request('/')
       .its('body')
-      .should('contain', 'See more supporters')
+      .should('contain', 'See full quotes')
+      .and('contain', 'href="/quotes#quote-minimax"')
       .and('contain', 'data-testid="launch-banner"');
 
     cy.intercept('GET', '**/_next/static/**/*.js', (request) => {
@@ -58,18 +59,20 @@ describe('Landing page performance', () => {
 
     cy.visit('/', {
       onBeforeLoad(win) {
-        win.localStorage.removeItem('inferencex-kimi-k3-modal-dismissed');
-        win.localStorage.removeItem('inferencex-kimi-k3-banner-dismissed');
+        win.localStorage.removeItem('inferencex-openai-rubin-banner-dismissed');
         observeLayoutShifts(win);
       },
     });
 
     cy.get('[data-testid="launch-banner"]').should('be.visible');
-    cy.get('[data-testid="intro-section"]').should('contain.text', 'See more supporters');
-    cy.get('[data-testid="quote-carousel-more-row"]')
+    cy.get('[data-testid="intro-section"]').should(
+      'contain.text',
+      'See full quotes & more supporters',
+    );
+    cy.get('[data-testid="supporters-more-row"]')
       .should('have.class', 'justify-end')
       .find('a')
-      .should('have.text', 'See more supporters →');
+      .should('have.text', 'See full quotes & more supporters →');
     expectLowCls();
   });
 
@@ -77,14 +80,17 @@ describe('Landing page performance', () => {
     cy.viewport(412, 823);
     cy.visit('/', {
       onBeforeLoad(win) {
-        win.localStorage.setItem('inferencex-kimi-k3-banner-dismissed', '1');
+        win.localStorage.setItem('inferencex-openai-rubin-banner-dismissed', '1');
         observeLayoutShifts(win);
       },
     });
 
     cy.get('html').should('have.attr', 'data-landing-banner-dismissed');
     cy.get('[data-testid="launch-banner"]').should('not.exist');
-    cy.get('[data-testid="intro-section"]').should('contain.text', 'See more supporters');
+    cy.get('[data-testid="intro-section"]').should(
+      'contain.text',
+      'See full quotes & more supporters',
+    );
     expectLowCls();
   });
 
@@ -100,11 +106,14 @@ describe('Landing page performance', () => {
       );
       expect(resourceNames.some((name) => name.includes('/minecraft-click.mp3'))).to.eq(false);
       expect(resourceNames.some((name) => name.includes('/Monocraft-'))).to.eq(false);
-      // The carousel only renders the active quote's logo, and it always starts on
-      // the first (MiniMax) quote, so a mobile load fetches at most that one logo —
-      // never the full supporter set.
-      const carouselLogos = new Set(resourceNames.filter((name) => name.includes('/logos/')));
-      expect(carouselLogos.size).to.be.lessThan(3);
+      // The landing AgentX ledger has five lazy model marks. A mobile viewport may
+      // fetch any visible subset, but the text-only supporter strip must not pull
+      // in its former logo set.
+      const landingLogos = new Set(resourceNames.filter((name) => name.includes('/logos/')));
+      expect(landingLogos.size).to.be.lessThan(6);
+      expect(
+        [...landingLogos].filter((name) => !new URL(name).pathname.endsWith('-color.svg')),
+      ).to.deep.eq([]);
       expect(
         resourceNames.some(
           (name) => name.includes('/brand/logo-color.webp') && name.includes('w=128'),
@@ -113,7 +122,7 @@ describe('Landing page performance', () => {
     });
   });
 
-  it('preloads only the default font and initially visible supporter logo', () => {
+  it('preloads only the default font and no supporter logos', () => {
     cy.request('/').then((response) => {
       // Next emits resource preloads as a `Link` response header (when `/` renders
       // dynamically) and/or as inlined <link rel="preload"> tags in the document
@@ -150,10 +159,9 @@ describe('Landing page performance', () => {
       }
 
       expect([...fonts]).to.have.length(1);
-      expect([...logos]).to.have.length(1);
-      // MiniMax is pinned first in the carousel order, so it's the initially
-      // visible supporter the server renders (index 0) and preloads.
-      expect([...logos][0]).to.eq('/logos/minimax.svg');
+      // The supporters band no longer renders a quote block, so no supporter
+      // logo is visible on the landing page and none should be preloaded.
+      expect([...logos]).to.have.length(0);
     });
   });
 
diff --git a/packages/app/cypress/e2e/line-labels.cy.ts b/packages/app/cypress/e2e/line-labels.cy.ts
index 345ac17f1..d516fdb6f 100644
--- a/packages/app/cypress/e2e/line-labels.cy.ts
+++ b/packages/app/cypress/e2e/line-labels.cy.ts
@@ -1,3 +1,5 @@
+import { expandLegendAdvanced } from '../support/legend-advanced';
+
 describe('Line Labels Toggle', () => {
   before(() => {
     cy.visit('/inference', {
@@ -8,6 +10,8 @@ describe('Line Labels Toggle', () => {
     // Wait for chart to load
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
     cy.get('.sidebar-legend').first().should('be.visible');
+    // Label switches live in the collapsed-by-default Advanced drawer.
+    expandLegendAdvanced();
   });
 
   it('Line Labels toggle exists in the legend', () => {
@@ -57,7 +61,7 @@ describe('Line Labels Toggle', () => {
     });
     cy.get('[data-testid="scatter-graph"] svg g.line-label').should('have.length.greaterThan', 0);
 
-    cy.get('[data-testid="scatter-graph"] svg').then(($svg) => {
+    cy.get('[data-testid="scatter-graph"] [data-testid="d3-chart-svg"]').then(($svg) => {
       const svg = $svg[0];
       const dots = svg.querySelectorAll('.dot-group');
       const labels = svg.querySelectorAll('g.line-label');
@@ -88,7 +92,7 @@ describe('Line Labels Toggle', () => {
 
     // The chart requires Shift for wheel zoom (so bare scroll doesn't hijack
     // the page). Dispatch a few shift+wheel events over the plot to zoom in.
-    cy.get('[data-testid="scatter-graph"] svg').then(($svg) => {
+    cy.get('[data-testid="scatter-graph"] [data-testid="d3-chart-svg"]').then(($svg) => {
       const svg = $svg[0];
       const r = svg.getBoundingClientRect();
       for (let i = 0; i < 3; i++) {
@@ -106,7 +110,7 @@ describe('Line Labels Toggle', () => {
     });
     cy.wait(300);
 
-    cy.get('[data-testid="scatter-graph"] svg').then(($svg) => {
+    cy.get('[data-testid="scatter-graph"] [data-testid="d3-chart-svg"]').then(($svg) => {
       const svg = $svg[0];
       const dots = svg.querySelectorAll('.dot-group');
       const labels = svg.querySelectorAll('g.line-label');
@@ -156,6 +160,7 @@ describe('Line Labels Toggle', () => {
       },
     });
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
+    expandLegendAdvanced();
     cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'checked');
 
     // Labels should be rendered
@@ -169,6 +174,7 @@ describe('Line Labels Toggle', () => {
       },
     });
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
+    expandLegendAdvanced();
     cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'unchecked');
 
     // Labels should not be rendered
@@ -182,6 +188,7 @@ describe('Line Labels Toggle', () => {
       },
     });
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
+    expandLegendAdvanced();
     cy.get('#scatter-point-labels').should('have.attr', 'data-state', 'unchecked');
   });
 
@@ -192,6 +199,7 @@ describe('Line Labels Toggle', () => {
       },
     });
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
+    expandLegendAdvanced();
     cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'checked');
     // Labels toggle is auto-enabled by the URL hydration so the advanced
     // (parallelism) point labels actually render.
@@ -205,11 +213,12 @@ describe('Line Labels Toggle', () => {
       },
     });
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
+    expandLegendAdvanced();
     cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'checked');
     cy.get('#scatter-point-labels').should('have.attr', 'data-state', 'unchecked');
   });
 
-  it('appends the precision to each line label when multiple precisions are selected', () => {
+  it('places the precision between the GPU and engine when multiple precisions are selected', () => {
     // Pair the FP4+FP8 selection with a model that has both precisions in the
     // fixtures. The default model (DeepSeek-V4-Pro) only has FP4, so
     // `effectivePrecisions` would drop FP8 and the test couldn't observe the
@@ -222,8 +231,9 @@ describe('Line Labels Toggle', () => {
     cy.get('[data-testid="scatter-graph"]').should('be.visible');
 
     // With both FP4 and FP8 shown, each curve is its own line and the label
-    // must carry the precision so the two curves of the same hardware are
-    // distinguishable (e.g. "B200 (vLLM) FP8" vs "B200 (vLLM) FP4").
+    // must carry the precision between the GPU and engine so the two curves of
+    // the same hardware are distinguishable (e.g. "B200 FP8 (vLLM)" vs
+    // "B200 FP4 (vLLM)").
     cy.get('[data-testid="scatter-graph"] svg g.line-label .ll-text')
       .should('have.length.greaterThan', 0)
       .then(($texts) => {
@@ -237,6 +247,11 @@ describe('Line Labels Toggle', () => {
           labels.some((t) => /\bFP4\b/u.test(t)),
           'an FP4 line label exists',
         ).to.equal(true);
+        for (const label of labels.filter((t) => /\bFP(?:4|8)\b/u.test(t) && /\(.+\)/u.test(t))) {
+          expect(label, 'precision precedes the engine suffix').to.match(
+            /^.+\sFP(?:4|8)\s\(.+\)$/u,
+          );
+        }
       });
   });
 });
diff --git a/packages/app/cypress/e2e/model-architecture.cy.ts b/packages/app/cypress/e2e/model-architecture.cy.ts
index a1dddb435..fb3b3c452 100644
--- a/packages/app/cypress/e2e/model-architecture.cy.ts
+++ b/packages/app/cypress/e2e/model-architecture.cy.ts
@@ -1,38 +1,42 @@
-describe('Model Architecture Diagram', () => {
-  before(() => {
-    // Use desktop viewport to ensure all UI elements are visible
-    cy.viewport(1280, 800);
-    // Pin to DeepSeek R1 — it has a rich architecture diagram (MoE + dense blocks)
-    // that the tests below exercise. The app's default model is DeepSeek V4 Pro,
-    // which has no architecture entry, so we select R1 explicitly via URL.
-    cy.visit('/inference?g_model=DeepSeek-R1-0528', {
-      onBeforeLoad(win) {
-        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
-      },
-    });
-    // Wait for the page to load
-    cy.get('[data-testid="inference-chart-display"]').should('be.visible');
-  });
-
-  it('architecture toggle renders for DeepSeek R1 with MoE badges', () => {
-    cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
-    cy.get('[data-testid="model-architecture-toggle"]').should(
-      'contain.text',
-      'Model Architecture',
-    );
-    cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
-    cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MLA');
-    cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '671B');
-  });
-
-  it('clicking toggle expands and renders the SVG diagram', () => {
-    cy.get('[data-testid="model-architecture-toggle"]').click();
-    cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+/**
+ * The architecture diagram lives on the /model/[slug] deep-dive pages, always
+ * expanded (`variant="inline"`). The dashboard renders a link row to those
+ * pages where the collapsible drawer used to be — covered by the final
+ * describe block.
+ */
+
+/** Visit a model deep-dive page and wait for the inline diagram to render. */
+function visitModelPage(slug: string) {
+  cy.viewport(1280, 800);
+  cy.visit(`/model/${slug}`, {
+    onBeforeLoad(win) {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+    },
   });
+  cy.get('[data-testid="model-architecture-inline"]').should('be.visible');
+  cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+}
 
+describe('Model Architecture Diagram', () => {
   describe('Collapsible Transformer Blocks (MoE model - DeepSeek R1)', () => {
     before(() => {
-      // SVG is already visible from previous test (testIsolation: false)
+      // DeepSeek R1 has a rich architecture diagram (MoE + dense blocks)
+      // that the tests below exercise.
+      visitModelPage('deepseek-r1');
+    });
+
+    it('inline architecture header renders with MoE badges', () => {
+      cy.get('[data-testid="model-architecture-inline"]').should(
+        'contain.text',
+        'Model Architecture',
+      );
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MLA');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', '671B');
+    });
+
+    it('renders the SVG diagram without any toggle click', () => {
+      cy.get('[data-testid="model-architecture-toggle"]').should('not.exist');
       cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
     });
 
@@ -103,22 +107,7 @@ describe('Model Architecture Diagram', () => {
 
   describe('Collapsible Transformer Block (Dense model - Llama 3.3 70B)', () => {
     before(() => {
-      // Switch model and open architecture
-      // Clear any stale Radix scroll lock from prior Select interactions
-      cy.document().then((doc) => {
-        delete doc.body.dataset.scrollLocked;
-        doc.body.style.removeProperty('pointer-events');
-      });
-      cy.get('[role="combobox"]').filter(':visible').first().click();
-      cy.get('[role="option"]').contains('Llama 3.3').click();
-
-      // Only click toggle if SVG is not already visible (previous describe may have left it open)
-      cy.get('body').then(($body) => {
-        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
-          cy.get('[data-testid="model-architecture-toggle"]').click();
-        }
-      });
-      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+      visitModelPage('llama-3-3-70b');
     });
 
     it('shows single transformer block without dense sub-block', () => {
@@ -140,34 +129,20 @@ describe('Model Architecture Diagram', () => {
     });
 
     it('shows Dense badge and GQA badge', () => {
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'Dense');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'GQA');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'Dense');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'GQA');
     });
   });
 
   describe('Collapsible Transformer Blocks (MoE model - Kimi K2.5)', () => {
     before(() => {
-      // Clear any stale Radix scroll lock from prior Select interactions
-      cy.document().then((doc) => {
-        delete doc.body.dataset.scrollLocked;
-        doc.body.style.removeProperty('pointer-events');
-      });
-      cy.get('[role="combobox"]').filter(':visible').first().click();
-      cy.get('[role="option"]').contains('Kimi K2.5/2.6').click();
-
-      cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
-      cy.get('body').then(($body) => {
-        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
-          cy.get('[data-testid="model-architecture-toggle"]').click();
-        }
-      });
-      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+      visitModelPage('kimi-k26');
     });
 
     it('shows MoE and MLA badges for Kimi K2.5', () => {
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MLA');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '1.0T');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MLA');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', '1.0T');
     });
 
     it('shows both dense and MoE transformer blocks', () => {
@@ -195,27 +170,13 @@ describe('Model Architecture Diagram', () => {
 
   describe('Collapsible Transformer Blocks (MoE model - MiniMax M2.5)', () => {
     before(() => {
-      // Clear any stale Radix scroll lock from prior Select interactions
-      cy.document().then((doc) => {
-        delete doc.body.dataset.scrollLocked;
-        doc.body.style.removeProperty('pointer-events');
-      });
-      cy.get('[role="combobox"]').filter(':visible').first().click();
-      cy.get('[role="option"]').contains('MiniMax M2.5/2.7').click();
-
-      cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
-      cy.get('body').then(($body) => {
-        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
-          cy.get('[data-testid="model-architecture-toggle"]').click();
-        }
-      });
-      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+      visitModelPage('minimax-m27');
     });
 
     it('shows MoE and GQA badges for MiniMax M2.5', () => {
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'GQA');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '230B');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'GQA');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', '230B');
     });
 
     it('shows single MoE transformer block without dense sub-block', () => {
@@ -243,27 +204,13 @@ describe('Model Architecture Diagram', () => {
 
   describe('Collapsible Transformer Blocks (MoE model - MiniMax M3)', () => {
     before(() => {
-      // Clear any stale Radix scroll lock from prior Select interactions
-      cy.document().then((doc) => {
-        delete doc.body.dataset.scrollLocked;
-        doc.body.style.removeProperty('pointer-events');
-      });
-      cy.get('[role="combobox"]').filter(':visible').first().click();
-      cy.get('[role="option"]').contains('MiniMax M3').click();
-
-      cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
-      cy.get('body').then(($body) => {
-        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
-          cy.get('[data-testid="model-architecture-toggle"]').click();
-        }
-      });
-      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+      visitModelPage('minimax-m3');
     });
 
     it('shows MoE and GQA badges for MiniMax M3', () => {
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'GQA');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '428B');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'GQA');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', '428B');
     });
 
     it('GQA attention is NOT expandable (sparse attention rendered as a static block)', () => {
@@ -280,27 +227,13 @@ describe('Model Architecture Diagram', () => {
 
   describe('Alternating Attention Blocks (MoE model - gpt-oss 120B)', () => {
     before(() => {
-      // Clear any stale Radix scroll lock from prior Select interactions
-      cy.document().then((doc) => {
-        delete doc.body.dataset.scrollLocked;
-        doc.body.style.removeProperty('pointer-events');
-      });
-      cy.get('[role="combobox"]').filter(':visible').first().click();
-      cy.get('[role="option"]').contains('gpt-oss').click();
-
-      cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
-      cy.get('body').then(($body) => {
-        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
-          cy.get('[data-testid="model-architecture-toggle"]').click();
-        }
-      });
-      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+      visitModelPage('gptoss-120b');
     });
 
     it('shows MoE and Sink/Full GQA badges for gpt-oss', () => {
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'Sink/Full GQA');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '120B');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'Sink/Full GQA');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', '120B');
     });
 
     it('shows two separate transformer blocks (no single expand-transformer)', () => {
@@ -356,27 +289,13 @@ describe('Model Architecture Diagram', () => {
 
   describe('Hybrid Attention Blocks (MoE model - DeepSeek V4 Pro)', () => {
     before(() => {
-      // Clear any stale Radix scroll lock from prior Select interactions
-      cy.document().then((doc) => {
-        delete doc.body.dataset.scrollLocked;
-        doc.body.style.removeProperty('pointer-events');
-      });
-      cy.get('[role="combobox"]').filter(':visible').first().click();
-      cy.get('[role="option"]').contains('DeepSeek V4 Pro').click();
-
-      cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
-      cy.get('body').then(($body) => {
-        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
-          cy.get('[data-testid="model-architecture-toggle"]').click();
-        }
-      });
-      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+      visitModelPage('deepseek-v4');
     });
 
     it('shows MoE and Hybrid badges for DeepSeek V4 Pro', () => {
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'Hybrid');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '1.6T');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'Hybrid');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', '1.6T');
     });
 
     it('shows two separate hybrid (CSA/HCA) blocks with an alternating indicator', () => {
@@ -446,48 +365,16 @@ describe('Model Architecture Diagram', () => {
 
   describe('Hybrid Attention Blocks (MoE model - Kimi K3)', () => {
     before(() => {
-      // The model dropdown only lists models that have availability rows, and the
-      // shared fixtures carry none for kimik3 (nothing ingested yet). Append one
-      // so K3 is selectable, then re-visit with the intercept in place. This block
-      // runs last in the spec, so no earlier block sees the patched availability.
-      cy.fixture('api/availability.json').then((rows: unknown[]) => {
-        cy.intercept('GET', '/api/v1/availability', {
-          body: [
-            ...rows,
-            {
-              model: 'kimik3',
-              isl: 8192,
-              osl: 1024,
-              precision: 'fp4',
-              hardware: 'mi355x',
-              framework: 'vllm',
-              spec_method: 'none',
-              disagg: false,
-              date: '2026-07-18',
-            },
-          ],
-        }).as('availability');
-      });
-      cy.visit('/inference?g_model=Kimi-K3', {
-        onBeforeLoad(win) {
-          win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
-        },
-      });
-      cy.get('[data-testid="inference-chart-display"]').should('be.visible');
-
-      cy.get('[data-testid="model-architecture-toggle"]').should('be.visible');
-      cy.get('body').then(($body) => {
-        if ($body.find('[data-testid="model-architecture-svg"]:visible').length === 0) {
-          cy.get('[data-testid="model-architecture-toggle"]').click();
-        }
-      });
-      cy.get('[data-testid="model-architecture-svg"]').should('be.visible');
+      // The /model page renders the diagram regardless of benchmark
+      // availability, so no fixture patching is needed (unlike the old
+      // dashboard drawer, which required the model to be selectable).
+      visitModelPage('kimi-k3');
     });
 
     it('shows MoE and Hybrid badges for Kimi K3', () => {
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'MoE');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', 'Hybrid');
-      cy.get('[data-testid="model-architecture-toggle"]').should('contain.text', '2.8T');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'MoE');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', 'Hybrid');
+      cy.get('[data-testid="model-architecture-inline"]').should('contain.text', '2.8T');
     });
 
     it('renders the KDA and gated-MLA layer categories as two alternating blocks', () => {
@@ -557,4 +444,107 @@ describe('Model Architecture Diagram', () => {
       cy.contains('Released by Moonshot AI').should('be.visible');
     });
   });
+
+  describe('Model index page (/model)', () => {
+    before(() => {
+      cy.viewport(1280, 800);
+      cy.visit('/model', {
+        onBeforeLoad(win) {
+          win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+        },
+      });
+    });
+
+    it('lists every model page with architecture badges', () => {
+      cy.get('h1').should('contain.text', 'Model Architectures');
+      cy.get('[data-testid="model-index-list"] a').should('have.length.at.least', 11);
+      cy.get('[data-testid="model-index-link-kimi-k3"]')
+        .should('contain.text', 'Kimi K3')
+        .and('contain.text', 'MoE')
+        .and('contain.text', 'Hybrid')
+        .and('contain.text', '2.8T')
+        .and('have.attr', 'href', '/model/kimi-k3');
+      // Models without a MODEL_ARCHITECTURES entry still get a card (no badges).
+      cy.get('[data-testid="model-index-link-glm-5-2"]').should('contain.text', 'GLM-5.2');
+    });
+
+    it('navigates to a model deep-dive page', () => {
+      cy.get('[data-testid="model-index-link-deepseek-r1"]').click();
+      cy.url().should('include', '/model/deepseek-r1');
+      cy.get('[data-testid="model-architecture-inline"]').should('be.visible');
+    });
+
+    it('is linked from the footer', () => {
+      cy.get('[data-testid="footer-link-model-architectures"]')
+        .should('have.attr', 'href', '/model')
+        .and('contain.text', 'Model Architectures');
+    });
+  });
+
+  describe('Embedded dashboard changelog starts collapsed', () => {
+    before(() => {
+      cy.viewport(1280, 800);
+      // deepseek-r1 has fixture benchmark data, so the changelog renders.
+      cy.visit('/model/deepseek-r1', {
+        onBeforeLoad(win) {
+          win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+        },
+      });
+    });
+
+    it('renders the changelog header collapsed inside the embed', () => {
+      cy.contains('button', 'Config Changelog', { timeout: 20000 })
+        .should('be.visible')
+        .and('have.attr', 'aria-expanded', 'false');
+    });
+
+    it('can still be expanded manually', () => {
+      cy.contains('button', 'Config Changelog').click();
+      cy.contains('button', 'Config Changelog').should('have.attr', 'aria-expanded', 'true');
+    });
+  });
+
+  describe('Dashboard architecture icon link (replaces the banner row)', () => {
+    before(() => {
+      cy.viewport(1280, 800);
+      cy.visit('/inference?g_model=DeepSeek-R1-0528', {
+        onBeforeLoad(win) {
+          win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+        },
+      });
+      cy.get('[data-testid="inference-chart-display"]').should('be.visible');
+    });
+
+    it('renders an icon link beside the model selector instead of a banner row', () => {
+      cy.get('[data-testid="model-architecture-toggle"]').should('not.exist');
+      cy.get('[data-testid="model-architecture-link"]').should('be.visible');
+      cy.get('[data-testid="model-architecture-link"]').should(
+        'have.attr',
+        'aria-label',
+        'Learn more about the DeepSeek R1 0528 671B architecture',
+      );
+      cy.get('[data-testid="model-architecture-link"]')
+        .should('have.attr', 'href')
+        .and('equal', '/model/deepseek-r1');
+      // The former banner copy and badges now live in the tooltip; Radix
+      // tooltips open on keyboard focus, which is more reliable in Cypress
+      // than synthetic hover.
+      cy.get('[data-testid="model-architecture-link"]').focus();
+      cy.get('[data-testid="model-architecture-tooltip"]')
+        .should('be.visible')
+        .and('contain.text', 'Learn more about the DeepSeek R1 0528 671B architecture')
+        .and('contain.text', 'MoE')
+        .and('contain.text', 'MLA')
+        .and('contain.text', '671B');
+      // Dismiss the tooltip so it doesn't cover the link for the next test.
+      cy.get('body').type('{esc}');
+    });
+
+    it('navigates to the model deep-dive page', () => {
+      cy.get('[data-testid="model-architecture-link"]').click();
+      cy.url().should('include', '/model/deepseek-r1');
+      cy.get('[data-testid="model-architecture-inline"]').should('be.visible');
+      cy.get('[data-testid="model-page-dashboard"]').should('exist');
+    });
+  });
 });
diff --git a/packages/app/cypress/e2e/model-route-slugs.cy.ts b/packages/app/cypress/e2e/model-route-slugs.cy.ts
new file mode 100644
index 000000000..1c20c43c9
--- /dev/null
+++ b/packages/app/cypress/e2e/model-route-slugs.cy.ts
@@ -0,0 +1,82 @@
+/**
+ * Per-model SEO routes: /historical/<model> and /calculator/<model> (and
+ * their /zh siblings) render the tab seeded to the routed model. Switching
+ * models on the page rewrites the pathname in place via history.replaceState
+ * — no full page reload and no App Router remount.
+ */
+const visitOptions = {
+  onBeforeLoad(win: Cypress.AUTWindow) {
+    win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+  },
+};
+
+type MarkerWindow = Cypress.AUTWindow & { __modelRouteNoReloadMarker?: boolean };
+
+const unlockPointerEvents = () => {
+  cy.document().then((doc) => {
+    delete doc.body.dataset.scrollLocked;
+    doc.body.style.removeProperty('pointer-events');
+  });
+};
+
+describe('Per-model dashboard routes', () => {
+  it('seeds /historical/<slug> to the routed model and rewrites the URL on switch without a reload', () => {
+    cy.visit('/historical/minimax-m3', visitOptions);
+    cy.get('[data-testid="historical-trends-display"]').should('be.visible');
+    cy.get('[data-testid="model-selector"]').should('contain.text', 'MiniMax');
+
+    // A full page load (or App Router remount via hard navigation) would wipe
+    // this marker; the in-place history.replaceState rewrite must keep it.
+    cy.window().then((win) => {
+      (win as MarkerWindow).__modelRouteNoReloadMarker = true;
+    });
+    unlockPointerEvents();
+    cy.get('[data-testid="model-selector"]').click();
+    cy.get('[role="option"]').not('[aria-selected="true"]').first().click();
+
+    cy.location('pathname').should('not.eq', '/historical/minimax-m3');
+    cy.location('pathname').should('match', /^\/historical\/[a-z0-9-]+$/u);
+    cy.window().its('__modelRouteNoReloadMarker').should('eq', true);
+  });
+
+  it('moves from the bare tab path to a slugged path when a non-default model is chosen', () => {
+    cy.visit('/historical', visitOptions);
+    cy.get('[data-testid="historical-trends-display"]').should('be.visible');
+    unlockPointerEvents();
+    cy.get('[data-testid="model-selector"]').click();
+    cy.get('[role="option"]').not('[aria-selected="true"]').first().click();
+    cy.location('pathname').should('match', /^\/historical\/[a-z0-9-]+$/u);
+  });
+
+  it('redirects alias slugs to the canonical model slug, keeping share-link params', () => {
+    cy.visit('/calculator/kimi?i_seq=8k%2F1k', visitOptions);
+    cy.location('pathname').should('eq', '/calculator/kimi-k26');
+    cy.location('search').should('contain', 'i_seq=8k%2F1k');
+  });
+
+  it('language toggle follows the in-place rewritten per-model path', () => {
+    cy.visit('/historical/minimax-m3', visitOptions);
+    cy.get('[data-testid="historical-trends-display"]').should('be.visible');
+    unlockPointerEvents();
+    cy.get('[data-testid="model-selector"]').click();
+    cy.get('[role="option"]').not('[aria-selected="true"]').first().click();
+    cy.location('pathname')
+      .should('match', /^\/historical\/[a-z0-9-]+$/u)
+      .then((rewritten) => {
+        cy.get('[data-testid="language-toggle"]').click();
+        cy.location('pathname').should('eq', `/zh${rewritten}`);
+      });
+  });
+
+  it('returns 404 for unknown model slugs', () => {
+    cy.request({ url: '/historical/not-a-model', failOnStatusCode: false })
+      .its('status')
+      .should('eq', 404);
+  });
+
+  it('renders the /zh sibling with a model-specific Chinese intro', () => {
+    cy.visit('/zh/historical/minimax-m3', visitOptions);
+    cy.get('[data-testid="zh-tab-intro"]').should('be.visible').and('contain.text', 'MiniMax M3');
+    cy.get('[data-testid="historical-trends-display"]').should('be.visible');
+  });
+});
diff --git a/packages/app/cypress/e2e/navigation.cy.ts b/packages/app/cypress/e2e/navigation.cy.ts
index 12fc661da..9ca19924a 100644
--- a/packages/app/cypress/e2e/navigation.cy.ts
+++ b/packages/app/cypress/e2e/navigation.cy.ts
@@ -46,18 +46,18 @@ describe('First-load navigation', () => {
     cy.visit('/', {
       onBeforeLoad(win) {
         win.localStorage.removeItem('inferencex-starred');
-        win.localStorage.removeItem('inferencex-star-modal-dismissed');
-        win.localStorage.removeItem('inferencex-kimi-k3-modal-dismissed');
-        win.localStorage.removeItem('inferencex-kimi-k3-banner-dismissed');
+        // Snoozed, not cleared: the star modal is the eligible landing nudge
+        // on first load, and its corner card would sit over the footer links
+        // these specs click.
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+        win.localStorage.removeItem('inferencex-openai-rubin-banner-dismissed');
       },
     });
 
-    // Banner (inline) and overlay modal coexist in independent slots.
-    cy.get('[data-testid="launch-modal"]').should('be.visible');
     cy.get('body').should('not.have.attr', 'data-scroll-locked');
   });
 
-  it('navigates to articles from the footer while the launch modal is visible', () => {
+  it('navigates to articles from the footer', () => {
     cy.get('[data-testid="footer-link-articles"]').scrollIntoView().click();
     cy.location('pathname').should('eq', '/blog');
   });
@@ -77,17 +77,50 @@ describe('First-load navigation', () => {
     cy.location('pathname').should('eq', '/compare');
   });
 
-  it('navigates to overview and the full dashboard from the landing CTAs', () => {
-    cy.get('[data-testid="landing-overview-link"]')
-      .should('have.attr', 'href', '/overview')
-      .click();
-    cy.location('pathname').should('eq', '/overview');
+  it('navigates to AgentX from the header with one click', () => {
+    cy.get('[data-testid="nav-link-agentx"]')
+      .should('have.attr', 'href', '/agentx')
+      .find('[data-nav-badge="agentx"]')
+      .should('be.visible')
+      .and('have.text', 'NEW');
+    cy.get('[data-testid="nav-link-agentx"]').click();
+    cy.location('pathname').should('eq', '/agentx');
+  });
 
-    cy.visit('/');
-    cy.get('[data-testid="landing-full-dashboard-link"]')
-      .should('have.attr', 'href', '/inference')
-      .click();
-    cy.location('pathname').should('eq', '/inference');
+  it('leads the landing page with the AgentX hero and its two CTAs', () => {
+    cy.get('[data-testid="compare-agentx-primary"]').within(() => {
+      // The hero owns /compare's h1; on the landing page it is a section heading.
+      cy.get('h2').should('have.text', 'Compare Realistic Agentic Inference Perf');
+      cy.get('h1').should('not.exist');
+      cy.get('[data-testid="compare-agentx-overview-link"]')
+        .should('contain.text', 'Overview')
+        .and('have.attr', 'href', '/overview');
+      cy.get('[data-testid="compare-agentx-dashboard-link"]')
+        .should('contain.text', 'Full dashboard')
+        .and('have.attr', 'href', '/inference/kimi-k3');
+      cy.get('[data-testid="compare-agentx-methodology-link"]').should('not.exist');
+      cy.get('[data-testid^="compare-agentx-model-"]').should('have.length', 6);
+      // Editorial order, not alphabetical — see FEATURED_AGENTX_MODEL_SLUGS.
+      cy.get('[data-testid^="compare-agentx-model-"]').then(($rows) => {
+        const slugs = [...$rows].map((row) =>
+          (row.dataset.testid ?? '').replace('compare-agentx-model-', ''),
+        );
+        expect(slugs).to.deep.equal([
+          'kimi-k3',
+          'deepseek-v4',
+          'glm-5-2',
+          'minimax-m3',
+          'qwen-3-5',
+          'qwen-3-8-flash-next',
+        ]);
+      });
+      // Every featured ledger row carries the NEW pill.
+      cy.get('[data-testid^="compare-agentx-model-"] [data-new-badge="agentx-ledger"]')
+        .should('have.length', 6)
+        .each(($badge) => expect($badge.text()).to.equal('NEW'));
+    });
+    cy.get('[data-testid="compare-agentx-overview-link"]').click();
+    cy.location('pathname').should('eq', '/overview');
   });
 
   it('navigates to submissions from the landing CTA', () => {
diff --git a/packages/app/cypress/e2e/nudge-system.cy.ts b/packages/app/cypress/e2e/nudge-system.cy.ts
index a6d4fe031..530d8a775 100644
--- a/packages/app/cypress/e2e/nudge-system.cy.ts
+++ b/packages/app/cypress/e2e/nudge-system.cy.ts
@@ -14,8 +14,7 @@ function clearAllNudgeStorage(win: Cypress.AUTWindow) {
   const keys = [
     'inferencex-starred',
     'inferencex-star-modal-dismissed',
-    'inferencex-kimi-k3-modal-dismissed',
-    'inferencex-kimi-k3-banner-dismissed',
+    'inferencex-openai-rubin-banner-dismissed',
     'inferencex-reproducibility-nudge-shown',
     'inferencex-star-nudge-shown',
     'inferencex-export-nudge-shown',
@@ -43,90 +42,86 @@ beforeEach(() => {
 // ---------------------------------------------------------------------------
 
 describe('Landing nudges — modals', () => {
-  it('shows launch modal and banner simultaneously on fresh first load', () => {
+  it('shows the launch banner on fresh first load', () => {
     cy.visit('/', {
       onBeforeLoad: clearAllNudgeStorage,
     });
-    // Banner (inline) and modal (overlay) occupy independent slots
-    cy.get('[data-testid="launch-banner"]').should('be.visible');
-    cy.get('[data-testid="launch-modal"]')
+    cy.get('[data-testid="launch-banner"]')
       .should('be.visible')
-      .and('match', 'div[role="dialog"][aria-modal="false"]');
-    // Only one overlay at a time — star modal should not appear
-    cy.get('[data-testid="github-star-modal"]').should('not.exist');
+      .and('contain.text', "OpenAI's Latest In House Chip verus Rubin NVL72")
+      .and(
+        'contain.text',
+        'Compare Jalapeño (Teacup) with Vera Rubin (July) NVL72 on DeepSeek R1 at 8K / 1K.',
+      )
+      .and('contain.text', 'View results');
+    // Banner + header-nav badges, plus the six AgentX hero ledger rows — the
+    // shared pill must render at the same fixed size everywhere it appears.
+    cy.get('[data-new-badge]')
+      .should('have.length', 8)
+      .then(($badges) => {
+        const sizes = [...$badges].map((badge) => {
+          const rect = badge.getBoundingClientRect();
+          return { width: rect.width, height: rect.height };
+        });
+        for (const size of sizes) {
+          expect(size.width).to.eq(sizes[0].width);
+          expect(size.height).to.eq(sizes[0].height);
+        }
+        expect(sizes[0]).to.deep.eq({ width: 32, height: 16 });
+
+        for (const badge of $badges) {
+          const label = badge.querySelector('[data-new-badge-label]');
+          expect(label, 'badge label').not.to.eq(null);
+
+          const badgeRect = badge.getBoundingClientRect();
+          const labelRect = label!.getBoundingClientRect();
+          const horizontalOffset =
+            labelRect.left + labelRect.width / 2 - (badgeRect.left + badgeRect.width / 2);
+          const verticalOffset =
+            labelRect.top + labelRect.height / 2 - (badgeRect.top + badgeRect.height / 2);
+
+          expect(horizontalOffset).to.be.closeTo(0, 0.1);
+          expect(verticalOffset).to.eq(0);
+
+          // The label box can sit dead centre while the glyphs themselves spill
+          // out of it, so measure the rendered text and not just its container.
+          const range = badge.ownerDocument.createRange();
+          range.selectNodeContents(label!);
+          const inkRect = range.getBoundingClientRect();
+          const inkOffset =
+            inkRect.left + inkRect.width / 2 - (badgeRect.left + badgeRect.width / 2);
+
+          expect(inkRect.left, 'label ink stays inside the pill').to.be.at.least(badgeRect.left);
+          expect(inkRect.right, 'label ink stays inside the pill').to.be.at.most(badgeRect.right);
+          expect(inkOffset, 'label ink is centred').to.be.closeTo(0, 0.5);
+        }
+      });
   });
 
-  it('dismissing launch modal persists — not shown on reload', () => {
-    cy.visit('/', {
+  it('localizes the Rubin comparison banner title in Chinese', () => {
+    cy.visit('/zh', {
       onBeforeLoad: clearAllNudgeStorage,
     });
-    cy.get('[data-testid="launch-modal"]').should('be.visible');
-    cy.get('[data-testid="launch-modal-dismiss"]').click();
-    cy.get('[data-testid="launch-modal"]').should('not.exist');
-
-    cy.reload();
-    cy.get('[data-testid="launch-modal"]').should('not.exist');
+    cy.get('[data-testid="launch-banner"]')
+      .should('be.visible')
+      .and('contain.text', 'OpenAI 最新自研芯片对比 Rubin NVL72')
+      .and(
+        'contain.text',
+        '对比 Jalapeño (Teacup) 与 Vera Rubin (July) NVL72 在 DeepSeek R1 8K / 1K 工作负载下的表现。',
+      )
+      .and('contain.text', '查看结果');
   });
 
-  it('launch modal Explore action persists dismissal in localStorage', () => {
+  it('does not float a duplicate GitHub star card over the footer', () => {
+    // The persistent star CTA lives in the footer grid (footer-star-cta) and
+    // the header; the old immediate star modal duplicated it and covered the
+    // footer, so it must stay gone.
     cy.visit('/', {
       onBeforeLoad: clearAllNudgeStorage,
     });
-    cy.get('[data-testid="launch-modal"]').should('be.visible');
-
-    // The action writes localStorage synchronously before navigation. Check
-    // the storage value before the navigation completes; combined with the
-    // "Maybe Later" persists-across-reload test, this covers the explore
-    // path without needing to stub window.location.
-    cy.get('[data-testid="launch-modal-action"]').click();
-    cy.window().then((win) => {
-      expect(win.localStorage.getItem('inferencex-kimi-k3-modal-dismissed')).to.eq('1');
-    });
-  });
-
-  it('shows star modal when launch modal was previously dismissed', () => {
-    cy.visit('/', {
-      onBeforeLoad(win) {
-        clearAllNudgeStorage(win);
-        win.localStorage.setItem('inferencex-kimi-k3-modal-dismissed', '1');
-      },
-    });
-    cy.get('[data-testid="launch-modal"]').should('not.exist');
-    cy.get('[data-testid="github-star-modal"]').should('be.visible');
-  });
-
-  it('star modal dismiss uses timed strategy — re-shows after expiry', () => {
-    cy.visit('/', {
-      onBeforeLoad(win) {
-        clearAllNudgeStorage(win);
-        win.localStorage.setItem('inferencex-kimi-k3-modal-dismissed', '1');
-      },
-    });
-    cy.get('[data-testid="github-star-modal"]').should('be.visible');
-    cy.get('[data-testid="github-star-modal-dismiss"]').click();
-    cy.get('[data-testid="github-star-modal"]').should('not.exist');
-
-    cy.window().then((win) => {
-      const value = win.localStorage.getItem('inferencex-star-modal-dismissed');
-      expect(value).to.not.equal(null);
-      expect(Number(value)).to.be.greaterThan(0);
-    });
-  });
-
-  it('starring permanently suppresses both star modal and star nudge', () => {
-    cy.visit('/', {
-      onBeforeLoad(win) {
-        clearAllNudgeStorage(win);
-        win.localStorage.setItem('inferencex-kimi-k3-modal-dismissed', '1');
-      },
-    });
-    cy.get('[data-testid="github-star-modal"]').should('be.visible');
-    cy.get('[data-testid="github-star-modal-action"]').click();
+    cy.get('[data-testid="launch-banner"]').should('be.visible');
     cy.get('[data-testid="github-star-modal"]').should('not.exist');
-
-    cy.window().then((win) => {
-      expect(win.localStorage.getItem('inferencex-starred')).to.eq('1');
-    });
+    cy.get('[data-testid="footer-star-cta"]').should('exist');
   });
 });
 
@@ -137,16 +132,8 @@ describe('Landing nudges — modals', () => {
 describe('Landing nudges — banner', () => {
   it('shows launch banner on landing page', () => {
     cy.visit('/', {
-      onBeforeLoad(win) {
-        clearAllNudgeStorage(win);
-        // Dismiss modals so the banner (highest priority at 60) is the active nudge.
-        // Actually the banner has priority 60 > launch modal 50, so it should show first.
-        // But the engine only shows one nudge at a time; the banner wins because of priority.
-      },
+      onBeforeLoad: clearAllNudgeStorage,
     });
-    // The banner has the highest priority (60), so it should appear.
-    // However, NudgeEngine only shows one nudge at a time.
-    // With immediate triggers and priority 60 > 50 > 40, the banner wins.
     cy.get('[data-testid="launch-banner"]').should('be.visible');
   });
 
@@ -178,7 +165,7 @@ describe('Landing nudges — banner', () => {
     cy.get('[data-testid="launch-banner"]').should('be.visible');
     cy.window().then((win) => {
       // Only the X button should persist a dismissal — show alone must not.
-      expect(win.localStorage.getItem('inferencex-kimi-k3-banner-dismissed')).to.eq(null);
+      expect(win.localStorage.getItem('inferencex-openai-rubin-banner-dismissed')).to.eq(null);
     });
   });
 
@@ -189,11 +176,16 @@ describe('Landing nudges — banner', () => {
     cy.get('[data-testid="launch-banner"]').should('be.visible');
     cy.get('[data-testid="launch-banner"]').click();
     cy.location('pathname', { timeout: 10000 }).should('eq', '/inference');
+    cy.location('search')
+      .should('include', 'g_model=DeepSeek-R1-0528')
+      .and('include', 'i_seq=8k%2F1k')
+      .and('include', 'i_prec=fp4')
+      .and('include', 'i_metric=y_outputTputPerMw');
 
     // Body click must not write the dismissal key — the banner should still
     // render on a fresh visit to landing.
     cy.window().then((win) => {
-      expect(win.localStorage.getItem('inferencex-kimi-k3-banner-dismissed')).to.eq(null);
+      expect(win.localStorage.getItem('inferencex-openai-rubin-banner-dismissed')).to.eq(null);
     });
 
     cy.visit('/');
@@ -261,6 +253,24 @@ describe('Dashboard nudges — filter-hint toast', () => {
     cy.get('[data-testid="filter-hint-nudge"]').should('not.exist');
   });
 
+  it('arms the filter hint before a client transition into inference', () => {
+    cy.visit('/evaluation', { onBeforeLoad: suppressCompetingDashboardToasts });
+    cy.get('[data-testid="filter-hint-nudge"]').should('not.exist');
+
+    cy.get('[data-testid="tab-trigger-inference"]').click();
+    cy.location('pathname').should('eq', '/inference');
+    cy.get('[data-testid="filter-hint-nudge"]', { timeout: 3000 }).should('be.visible');
+  });
+
+  it('re-evaluates the localized filter hint after a Chinese tab transition', () => {
+    cy.visit('/zh/evaluation', { onBeforeLoad: suppressCompetingDashboardToasts });
+    cy.get('[data-testid="tab-trigger-inference"]').click();
+    cy.location('pathname').should('eq', '/zh/inference');
+    cy.get('[data-testid="filter-hint-nudge"]', { timeout: 3000 })
+      .should('be.visible')
+      .and('contain.text', '图表太拥挤？');
+  });
+
   it('dismissal persists to localStorage and the nudge stays gone after reload', () => {
     cy.visit('/inference', { onBeforeLoad: suppressCompetingDashboardToasts });
     cy.get('[data-testid="filter-hint-nudge"]', { timeout: 5000 }).should('be.visible');
@@ -344,7 +354,6 @@ describe('Nudge scope isolation', () => {
     cy.visit('/inference', {
       onBeforeLoad: clearAllNudgeStorage,
     });
-    cy.get('[data-testid="launch-modal"]').should('not.exist');
     cy.get('[data-testid="github-star-modal"]').should('not.exist');
     cy.get('[data-testid="launch-banner"]').should('not.exist');
   });
@@ -354,8 +363,7 @@ describe('Nudge scope isolation', () => {
       onBeforeLoad(win) {
         clearAllNudgeStorage(win);
         // Dismiss all landing nudges so nothing blocks visibility checks
-        win.localStorage.setItem('inferencex-kimi-k3-modal-dismissed', '1');
-        win.localStorage.setItem('inferencex-kimi-k3-banner-dismissed', '1');
+        win.localStorage.setItem('inferencex-openai-rubin-banner-dismissed', '1');
         win.localStorage.setItem('inferencex-starred', '1');
       },
     });
diff --git a/packages/app/cypress/e2e/overlay-legend-remove.cy.ts b/packages/app/cypress/e2e/overlay-legend-remove.cy.ts
index d52df3965..07c555f80 100644
--- a/packages/app/cypress/e2e/overlay-legend-remove.cy.ts
+++ b/packages/app/cypress/e2e/overlay-legend-remove.cy.ts
@@ -4,7 +4,7 @@
  *
  * Regression: with an overlay active, the chart reads official visibility from
  * `localOfficialOverride` (the unified overlay-mode selection), but the legend
- * X routed straight to InferenceContext's `removeHwType`, which mutates
+ * X routed straight to the inference action domain's `removeHwType`, which mutates
  * `activeHwTypes` — a set the chart ignores in overlay mode. The click
  * appeared to do nothing. The legend toggle already had the overlay-aware
  * split (`unifiedToggle`); the X now shares it (`handleRemoveHwType`).
@@ -22,9 +22,10 @@ describe('Official legend X works while an unofficial overlay is loaded', () =>
     // Use distinct hardware so the engine-comparison exclusion policy does
     // not resolve the official and unofficial rows as one competing family.
     interceptOverlayRun({ overlayHardware: 'h100' });
-    // The agentic default mode is E2E Normalized Interactivity (which suppresses overlays and
-    // fetches derived metrics) — stub the fetch, then switch to Interactivity
-    // where the overlay renders.
+    // Agentic charts default to Interactivity, where the overlay renders. The
+    // derived-metrics fetch now happens only under E2E Normalized Interactivity,
+    // so the stub below is a guard against a stray request rather than a
+    // dependency of this suite.
     interceptDerivedAgenticMetrics();
     cy.visit(`/inference?unofficialrun=${OVERLAY_RUN_ID}&i_seq=agentic-traces&i_pctl=p90`, {
       onBeforeLoad(win) {
@@ -33,9 +34,11 @@ describe('Official legend X works while an unofficial overlay is loaded', () =>
       },
     });
     cy.wait('@unofficialRun');
-    // Interactivity is nested under the Advanced menu on agentic charts.
-    cy.get('[data-testid="x-axis-mode-advanced"]').click();
-    cy.get('[data-testid="x-axis-mode-interactivity"]').click();
+    // Every x-axis metric is a top-level tab on agentic charts, and Interactivity
+    // is the default — clicked anyway so the suite does not depend on that.
+    cy.get('[data-testid="x-axis-mode-interactivity"]')
+      .click()
+      .should('have.attr', 'data-state', 'active');
     cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
     cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(
       'have.length',
@@ -43,6 +46,16 @@ describe('Official legend X works while an unofficial overlay is loaded', () =>
     );
   });
 
+  // Cypress clears intercepts between tests, so the derived-metrics stub is
+  // re-registered per test rather than once in `before`. Until #736 the
+  // agentic default was E2E Normalized Interactivity, so the fetch happened
+  // during `before` while the stub was still alive and React Query held the
+  // result for the rest of the spec. The default no longer fetches, so any
+  // later switch into that mode issues a fresh request.
+  beforeEach(() => {
+    interceptDerivedAgenticMetrics();
+  });
+
   it('shows official points and an official legend entry initially', () => {
     cy.get('[data-testid="inference-chart-display"] svg .dot-group').should(($dots) => {
       expect(countVisible($dots), 'visible official points').to.be.greaterThan(0);
@@ -69,19 +82,50 @@ describe('Official legend X works while an unofficial overlay is loaded', () =>
     cy.get('[data-testid="inference-chart-display"] svg .dot-group').should(($dots) => {
       expect(countVisible($dots), 'visible official points after remove').to.eq(0);
     });
-    // The overlay series is untouched. Optimal Only still hides its trace-less
-    // points because unofficial rows cannot join the canonical frontier.
+    // The overlay series is untouched — hiding an official SKU must not disturb
+    // it. All five overlay points stay visible under Optimal Only: they are
+    // non-dominated on the interactivity axes, and since #736 lacking persisted
+    // traces no longer excludes them from the frontier.
     cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(($pts) => {
-      expect(countVisible($pts), 'visible overlay X markers').to.eq(0);
+      expect(countVisible($pts), 'visible overlay X markers').to.eq(REAL_CONFIGS.length);
     });
     // Inactive row: the hover affordance flips to the "+" restore indicator
     // (explicit "clicking the name brings it back"), and the Hide X is gone.
     cy.get('[data-testid="chart-legend"] [title^="Show B300"]').should('exist');
+    // Best per SKU lives in the Quick Filters dialog now.
+    cy.get('[data-testid="scatter-quick-filters"]').click();
+    cy.get('[data-testid="quick-filter-best-per-sku"]').should(
+      'have.attr',
+      'data-state',
+      'unchecked',
+    );
+    cy.contains('button', 'Done').click();
+    cy.get('[data-testid="quick-filters-dialog"]').should('not.exist');
     cy.get(
       '[data-testid="chart-legend"] [role="button"][aria-label^="Hide"][aria-label*="B300"]',
     ).should('not.exist');
   });
 
+  it('keeps the official SKU hidden when chart metrics change', () => {
+    cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
+    cy.contains('[role="option"]', 'Cost per Million Total Tokens (Owning - Hyperscaler)').click({
+      force: true,
+    });
+
+    cy.get('[data-testid="chart-legend"] [title^="Show B300"]').should('exist');
+    cy.get('[data-testid="inference-chart-display"] svg .dot-group').should(($dots) => {
+      expect(countVisible($dots), 'visible official points after Y-axis change').to.eq(0);
+    });
+
+    cy.get('[data-testid="x-axis-mode-ttft"]').click().should('have.attr', 'data-state', 'active');
+    cy.get('[data-testid="chart-legend"] [title^="Show B300"]').should('exist');
+    cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(($pts) => {
+      expect(countVisible($pts), 'visible overlay points after metric changes').to.be.greaterThan(
+        0,
+      );
+    });
+  });
+
   it('re-activating the SKU from the legend restores the official points', () => {
     cy.get('[data-testid="chart-legend"]').contains('B300').click();
     cy.get('[data-testid="inference-chart-display"] svg .dot-group').should(($dots) => {
diff --git a/packages/app/cypress/e2e/overlay-optimal-only.cy.ts b/packages/app/cypress/e2e/overlay-optimal-only.cy.ts
index 520567806..f9f388c4d 100644
--- a/packages/app/cypress/e2e/overlay-optimal-only.cy.ts
+++ b/packages/app/cypress/e2e/overlay-optimal-only.cy.ts
@@ -1,22 +1,31 @@
 /**
- * Unofficial runs do not have persisted request traces, so they cannot join
- * the canonical E2E Normalized Interactivity frontier. Optimal Only must hide
- * them on every AgentX axis; Show All remains the explicit way to inspect them.
+ * Optimal Only must filter overlay (unofficial-run) points by the same rule it
+ * applies to official ones: the Pareto frontier of the *selected* axes.
+ *
+ * Until #736 the rule was different — unofficial runs have no persisted request
+ * traces, so they could never join the canonical E2E Normalized Interactivity
+ * frontier and Optimal Only hid every one of them. That gating is gone, so an
+ * overlay point now survives exactly when it is non-dominated on the axes on
+ * screen, and Show All remains the way to inspect the rest.
  */
 import { interceptDerivedAgenticMetrics, unlockAgenticGate } from '../support/e2e';
 import {
   countVisible,
+  DOMINATED_CONFIG,
   interceptOverlayRun,
   OVERLAY_RUN_ID,
   REAL_CONFIGS,
 } from '../support/overlay-fixtures';
 
-describe('Overlay points follow canonical Optimal Only policy (agentic interactivity)', () => {
+// The five real configs are all non-dominated on the interactivity axes, so a
+// sixth, deliberately dominated point is what gives Optimal Only something to
+// remove. Without it this suite would assert the same count either way and pass
+// no matter what the filter did.
+const OVERLAY_CONFIGS = [...REAL_CONFIGS, DOMINATED_CONFIG];
+
+describe('Overlay points follow Optimal Only on the selected axes', () => {
   before(() => {
-    interceptOverlayRun();
-    // The agentic default mode is E2E Normalized Interactivity (which suppresses overlays and
-    // fetches derived metrics) — stub the fetch, then switch to the
-    // Interactivity mode this suite is about.
+    interceptOverlayRun({ overlayConfigs: OVERLAY_CONFIGS });
     interceptDerivedAgenticMetrics();
     cy.visit(`/inference?unofficialrun=${OVERLAY_RUN_ID}&i_seq=agentic-traces&i_pctl=p90`, {
       onBeforeLoad(win) {
@@ -25,30 +34,37 @@ describe('Overlay points follow canonical Optimal Only policy (agentic interacti
       },
     });
     cy.wait('@unofficialRun');
-    // Interactivity is nested under the Advanced menu on agentic charts.
-    cy.get('[data-testid="x-axis-mode-advanced"]').click();
-    cy.get('[data-testid="x-axis-mode-interactivity"]').click();
+    // Every x-axis metric is a top-level tab on agentic charts, and Interactivity
+    // is the default — clicked anyway so the suite does not depend on that.
+    cy.get('[data-testid="x-axis-mode-interactivity"]')
+      .click()
+      .should('have.attr', 'data-state', 'active');
     cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
-    cy.get('[data-testid="x-axis-mode-advanced"]')
-      .should('have.attr', 'data-state', 'active')
-      .and('contain.text', 'Interactivity');
+    // All six are rendered; visibility is what Optimal Only changes.
     cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(
       'have.length',
-      REAL_CONFIGS.length,
+      OVERLAY_CONFIGS.length,
     );
   });
 
-  it('hides trace-less overlay points in the default Optimal Only view', () => {
+  // Cypress clears intercepts between tests, so the derived-metrics stub is
+  // re-registered per test rather than once in `before`. Until #736 the agentic
+  // default was E2E Normalized Interactivity, so the fetch happened during
+  // `before` while the stub was still alive and React Query held the result for
+  // the rest of the spec. The default no longer fetches.
+  beforeEach(() => {
+    interceptDerivedAgenticMetrics();
+  });
+
+  it('drops only the dominated overlay point in the default Optimal Only view', () => {
     cy.get('#scatter-hide-non-optimal').should('have.attr', 'data-state', 'checked');
-    // The deterministic derived-metric stub puts all five official rows on the
-    // canonical frontier.
     cy.get('[data-testid="inference-chart-display"] svg .dot-group').should(($dots) => {
       expect(countVisible($dots), 'visible official points').to.eq(REAL_CONFIGS.length);
     });
-    // Overlay rows have no persisted trace ids and therefore no canonical
-    // frontier membership.
+    // Five of six survive: every real config is non-dominated on these axes, and
+    // the overlay's trace-less rows are no longer excluded for lacking traces.
     cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(($pts) => {
-      expect(countVisible($pts), 'visible overlay X markers').to.eq(0);
+      expect(countVisible($pts), 'visible overlay X markers').to.eq(REAL_CONFIGS.length);
     });
   });
 
@@ -56,15 +72,15 @@ describe('Overlay points follow canonical Optimal Only policy (agentic interacti
     cy.get('#scatter-hide-non-optimal').click();
     cy.get('#scatter-hide-non-optimal').should('have.attr', 'data-state', 'unchecked');
     cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(($pts) => {
-      expect(countVisible($pts), 'visible overlay X markers').to.eq(REAL_CONFIGS.length);
+      expect(countVisible($pts), 'visible overlay X markers').to.eq(OVERLAY_CONFIGS.length);
     });
   });
 
-  it('re-hides trace-less overlay points when Optimal Only is re-enabled', () => {
+  it('re-drops the dominated overlay point when Optimal Only is re-enabled', () => {
     cy.get('#scatter-hide-non-optimal').click();
     cy.get('#scatter-hide-non-optimal').should('have.attr', 'data-state', 'checked');
     cy.get('[data-testid="inference-chart-display"] svg .unofficial-overlay-pt').should(($pts) => {
-      expect(countVisible($pts), 'visible overlay X markers').to.eq(0);
+      expect(countVisible($pts), 'visible overlay X markers').to.eq(REAL_CONFIGS.length);
     });
   });
 });
diff --git a/packages/app/cypress/e2e/overview-a11y.cy.ts b/packages/app/cypress/e2e/overview-a11y.cy.ts
new file mode 100644
index 000000000..64ff36457
--- /dev/null
+++ b/packages/app/cypress/e2e/overview-a11y.cy.ts
@@ -0,0 +1,47 @@
+import type { Result } from 'axe-core';
+
+const OVERVIEW_SURFACE = '[data-testid="overview-page"]';
+
+const OVERVIEW_ROUTES = [
+  { name: 'English hardware comparison', path: '/overview' },
+  { name: 'Chinese hardware comparison', path: '/zh/overview' },
+  { name: 'historical comparison', path: '/overview?compare=30d' },
+] as const;
+
+const OVERVIEW_THEMES = ['light', 'dark'] as const;
+
+const WCAG_AA_RULES = {
+  runOnly: {
+    type: 'tag' as const,
+    values: ['wcag2a', 'wcag2aa', 'wcag21a', 'wcag21aa'],
+  },
+};
+
+function reportViolations(violations: Result[]): void {
+  const summary = violations
+    .map(
+      ({ help, id, nodes }) =>
+        `${id}: ${help}\n${nodes.map(({ target }) => `  ${target.join(' ')}`).join('\n')}`,
+    )
+    .join('\n');
+  throw new Error(`Overview accessibility violations:\n${summary}`);
+}
+
+describe('Overview accessibility', () => {
+  for (const { name, path } of OVERVIEW_ROUTES) {
+    for (const theme of OVERVIEW_THEMES) {
+      it(`has no WCAG 2.0/2.1 A or AA violations in the ${name} view using the ${theme} theme`, () => {
+        cy.visit(path, {
+          onBeforeLoad(win) {
+            win.localStorage.setItem('theme', theme);
+          },
+        });
+        cy.get('html').should(theme === 'dark' ? 'have.class' : 'not.have.class', 'dark');
+        cy.get(OVERVIEW_SURFACE).should('be.visible');
+
+        cy.injectAxe();
+        cy.checkA11y(OVERVIEW_SURFACE, WCAG_AA_RULES, reportViolations);
+      });
+    }
+  }
+});
diff --git a/packages/app/cypress/e2e/overview.cy.ts b/packages/app/cypress/e2e/overview.cy.ts
index 6ff5c79eb..875e9dd08 100644
--- a/packages/app/cypress/e2e/overview.cy.ts
+++ b/packages/app/cypress/e2e/overview.cy.ts
@@ -1,10 +1,13 @@
+import { TCO_SOURCE_TITLE, TCO_SOURCE_URL } from '@semianalysisai/inferencex-constants';
+
 // Order mirrors DEFAULT_MODELS (MODEL_CONFIG insertion order), which fixes the
-// matrix row order.
+// row order within each scenario group — AgentX rows lead the matrix, 8K/1K
+// rows follow.
 const MODEL_LABELS = [
-  'DeepSeek V4 Pro 1.6T',
+  'DeepSeek V4 Pro 0813 1.6T',
   'Kimi K3 2.8T',
   'MiniMax M3 428B',
-  'GLM5.2',
+  'GLM5.2/GLM5.3',
   'Qwen3.5 397B',
 ];
 
@@ -19,24 +22,52 @@ const PLATFORM_HEADERS = [
 ];
 
 const SINGLE_TURN = 'single_turn_8k1k';
-/** Five models, three of them with both a single-turn and an AgentX row. */
-const MATRIX_ROWS = 8;
+/** Six models: three with both a single-turn and an AgentX row, and three
+ *  curated AgentX-only (Kimi K3, GLM 5.2, Qwen3.8-Flash-Next). */
+const MATRIX_ROWS = 9;
 const AGENTX = 'agentx';
 const AGENTX_LABEL = 'Long Context Multi-Turn Realistic Agentic Scenario (AgentX)';
-const AGENTX_LABEL_ZH = '长上下文多轮真实智能体场景（AgentX）';
-
-const PAGE_TITLE = 'Inference Cost per Million Tokens';
-const PAGE_TITLE_ZH = '推理每百万 token 成本';
-const SOURCE_NOTE = 'Source: InferenceX & SemiAnalysis Market July 2026 AI Cloud TCO Model';
-const SOURCE_LINK_TEXT = 'SemiAnalysis Market July 2026 AI Cloud TCO Model';
-const SOURCE_NOTE_ZH = '来源：InferenceX 与 SemiAnalysis Market July 2026 AI Cloud TCO Model';
-const SOURCE_HREF = 'https://semianalysis.com/ai-cloud-tco-model/';
+const AGENTX_LABEL_ZH = '长上下文、多轮交互的真实智能体场景（AgentX）';
+/** Shared by both locales: the scenario is named after its acronym. */
+const AGENTX_SHORT = 'AgentX';
+
+const PAGE_TITLE = 'Agentic Inference Costs';
+const PAGE_TITLE_ZH = '智能体推理成本';
+const SOURCE_NOTE = `Source: InferenceX & ${TCO_SOURCE_TITLE}`;
+const SOURCE_NOTE_ZH = `来源：InferenceX 与 ${TCO_SOURCE_TITLE}`;
 const SCOPE_METRIC = 'Hyperscaler cost';
 const SCOPE_DIRECTION = '↓ Lower is better';
 const SCOPE_LINE = `${SCOPE_METRIC} · ${SCOPE_DIRECTION} · ${SOURCE_NOTE}`;
-const SCOPE_METRIC_ZH = '超大规模云（hyperscaler）成本';
+const SCOPE_METRIC_ZH = '超大规模云厂商成本';
 const SCOPE_DIRECTION_ZH = '↓ 越低越好';
 const SCOPE_LINE_ZH = `${SCOPE_METRIC_ZH} · ${SCOPE_DIRECTION_ZH} · ${SOURCE_NOTE_ZH}`;
+const OVERVIEW_TIERS = [30, 50, 75, 100, 150, 200] as const;
+
+function selectOverviewTier(tier: (typeof OVERVIEW_TIERS)[number]) {
+  const targetIndex = OVERVIEW_TIERS.indexOf(tier);
+  // Unlike clicks, synthetic input events are not replayed by React, so a
+  // dispatch that lands before hydration finishes is silently dropped. The
+  // `.should` callback retries the whole dispatch until React has committed
+  // the selection, which only happens once the slider is interactive.
+  return cy.get<HTMLInputElement>('[data-testid="overview-tier-slider"]').should(([slider]) => {
+    const view = slider.ownerDocument.defaultView;
+    if (view === null) throw new Error('Slider has no window');
+    const tracker = (slider as { _valueTracker?: { setValue: (next: string) => void } })
+      ._valueTracker;
+    // React installs the value tracker when hydration commits the input.
+    expect(tracker, 'tier slider is hydrated').to.not.eq(undefined);
+    const nativeValueSetter = Object.getOwnPropertyDescriptor(
+      view.HTMLInputElement.prototype,
+      'value',
+    )?.set;
+    if (nativeValueSetter === undefined) throw new Error('Range input has no native value setter');
+    nativeValueSetter.call(slider, String(targetIndex));
+    // Desync the tracker so React always registers the dispatched input.
+    tracker?.setValue('');
+    slider.dispatchEvent(new view.Event('input', { bubbles: true }));
+    expect(slider.dataset.tier, 'slider committed tier').to.eq(String(tier));
+  });
+}
 
 function expectNoHorizontalOverflow() {
   cy.document().then((doc) => {
@@ -61,6 +92,52 @@ function expectNoHorizontalScroller(testId: string) {
   });
 }
 
+function stubFullscreenApi(win: Window) {
+  let fullscreenElement: Element | null = null;
+  const document = win.document;
+  const EventConstructor = document.defaultView?.Event ?? Event;
+  const ElementConstructor = document.defaultView?.Element ?? Element;
+  Object.defineProperty(document, 'fullscreenEnabled', {
+    configurable: true,
+    value: true,
+  });
+  Object.defineProperty(document, 'fullscreenElement', {
+    configurable: true,
+    get: () => fullscreenElement,
+  });
+  Object.defineProperty(document, 'exitFullscreen', {
+    configurable: true,
+    value: () => {
+      fullscreenElement = null;
+      document.dispatchEvent(new EventConstructor('fullscreenchange'));
+      return Promise.resolve();
+    },
+  });
+  Object.defineProperty(ElementConstructor.prototype, 'requestFullscreen', {
+    configurable: true,
+    value() {
+      fullscreenElement = document.querySelector('[data-testid="overview-presentation-surface"]');
+      document.dispatchEvent(new EventConstructor('fullscreenchange'));
+      return Promise.resolve();
+    },
+  });
+}
+
+/**
+ * The row header shows the acronym so the 22%-wide model column stays one line,
+ * and keeps the full scenario name for assistive tech and the hover title. Both
+ * strings are in the cell, so assert the two layers rather than their
+ * concatenation. Scenarios already named by a short label render one node.
+ */
+function expectAgentxScenario(fullLabel: string) {
+  cy.get('[data-testid="overview-model-scenario"]')
+    .should('have.attr', 'title', fullLabel)
+    .within(() => {
+      cy.get('.sr-only').should('have.text', fullLabel);
+      cy.get('[aria-hidden="true"]').should('have.text', AGENTX_SHORT);
+    });
+}
+
 /** Visible dates and snapshot framing must be gone; evidence stays in labels. */
 function expectNoVisibleDatesOrSnapshot() {
   cy.get('[data-testid="overview-pair-evidence-date"]').should('not.exist');
@@ -140,14 +217,14 @@ describe('Overview page', () => {
         'preserved';
     });
 
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '75').click();
+    selectOverviewTier(75);
     cy.wait('@overviewJson');
-    cy.location('search', { timeout: 15_000 }).should('eq', '?tier=75');
+    cy.location('search').should('eq', '?tier=75');
     cy.window().its('__overviewNavigationSentinel').should('eq', 'preserved');
 
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '50').click();
+    selectOverviewTier(50);
     cy.location('search').should('eq', '');
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '75').click();
+    selectOverviewTier(75);
     cy.location('search').should('eq', '?tier=75');
     cy.then(() => {
       expect(jsonRequests, 'one request; both visited selections are cached').to.equal(1);
@@ -158,17 +235,17 @@ describe('Overview page', () => {
       .find('[data-overview-engine-scope="all"]')
       .click();
     cy.wait('@overviewJson');
-    cy.location('search', { timeout: 15_000 }).should('eq', '?tier=75&engine=all');
+    cy.location('search').should('eq', '?tier=75&engine=all');
     cy.window().its('__overviewNavigationSentinel').should('eq', 'preserved');
 
     cy.get('[data-overview-comparison="30d"]').click();
     cy.wait('@overviewJson');
-    cy.location('search', { timeout: 15_000 }).should('eq', '?tier=75&engine=all&compare=30d');
+    cy.location('search').should('eq', '?tier=75&engine=all&compare=30d');
     cy.window().its('__overviewNavigationSentinel').should('eq', 'preserved');
 
     cy.go('back');
     cy.get('[data-overview-comparison="hardware"]').should('have.attr', 'aria-current', 'true');
-    cy.location('search', { timeout: 15_000 }).should('eq', '?tier=75&engine=all');
+    cy.location('search').should('eq', '?tier=75&engine=all');
     cy.then(() => {
       expect(rscRequests, 'selector and popstate RSC requests').to.equal(0);
     });
@@ -178,7 +255,7 @@ describe('Overview page', () => {
     cy.viewport(1280, 900);
     cy.visit('/overview');
     cy.intercept('GET', '**/api/v1/overview*').as('overviewJson');
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '75').click();
+    selectOverviewTier(75);
     cy.wait('@overviewJson');
     cy.location('search').should('eq', '?tier=75');
 
@@ -211,16 +288,19 @@ describe('Overview page', () => {
       });
     }).as('overviewJson');
 
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '75').click();
+    selectOverviewTier(75);
     cy.get('[data-testid="overview-engine-scope-switcher"]')
       .find('[data-overview-engine-scope="all"]')
       .click();
 
-    cy.location('search', { timeout: 15_000 }).should('eq', '?tier=75&engine=all');
+    cy.location('search').should('eq', '?tier=75&engine=all');
     // The rendered state, not just the URL: the losing response must not win.
-    cy.get('[data-testid="overview-tier-switcher"] [aria-current="page"]', {
+    // Keeps its explicit timeout because it waits on the real API through
+    // `request.continue` plus the delay above — unlike the URL assertions
+    // around it, which the click resolves synchronously.
+    cy.get('[data-testid="overview-tier-slider"]', {
       timeout: 15_000,
-    }).should('have.text', '75');
+    }).should('have.attr', 'data-tier', '75');
     cy.get('[data-overview-engine-scope="all"]').should('have.attr', 'aria-current', 'true');
     cy.location('search').should('eq', '?tier=75&engine=all');
   });
@@ -234,15 +314,79 @@ describe('Overview page', () => {
       });
     }).as('overviewJson');
 
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '75').click();
+    selectOverviewTier(75);
     cy.get('[data-testid="overview-page"] [aria-busy="true"]').should('exist');
     cy.wait('@overviewJson');
-    cy.get('[data-testid="overview-page"] [aria-busy="true"]', { timeout: 15_000 }).should(
-      'not.exist',
+    cy.get('[data-testid="overview-page"] [aria-busy="true"]').should('not.exist');
+  });
+
+  it('keeps the last matrix and history entry when a popstate request fails', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview');
+    // A client-derived reference switch proves hydration and the popstate
+    // listener are established without warming the tier payload under test.
+    cy.get('[data-testid="overview-reference-select"]').click();
+    cy.get('[data-overview-reference="b300"]').click();
+    cy.location('search').should('eq', '?ref=b300');
+    cy.intercept('GET', '**/api/v1/overview?tier=75', { statusCode: 500 }).as(
+      'overviewPopstateFailure',
     );
+    cy.window().then((win) => {
+      (win as Window & { __overviewNavigationSentinel?: string }).__overviewNavigationSentinel =
+        'preserved';
+      win.History.prototype.pushState.call(
+        win.history,
+        win.history.state,
+        '',
+        '/overview?tier=75&ref=b300',
+      );
+      cy.wrap(win.history.length).as('overviewHistoryLength');
+    });
+
+    cy.go('back');
+    cy.location('search').should('eq', '?ref=b300');
+    cy.go('forward');
+    cy.wait('@overviewPopstateFailure');
+    cy.get('[data-testid="overview-navigation-error"]')
+      .should('have.attr', 'role', 'alert')
+      .and(
+        'have.text',
+        'Could not load the selected comparison. Showing the last successfully loaded data.',
+      );
+    cy.get('[data-testid="overview-tier-slider"]').should('have.attr', 'data-tier', '50');
+    cy.get('[data-testid="overview-page"] [aria-busy="true"]').should('not.exist');
+    cy.location('search').should('eq', '?tier=75&ref=b300');
+    cy.get<number>('@overviewHistoryLength').then((expectedHistoryLength) => {
+      cy.window().then((after) => {
+        expect(after.history.length).to.equal(expectedHistoryLength);
+        expect(
+          (after as Window & { __overviewNavigationSentinel?: string })
+            .__overviewNavigationSentinel,
+        ).to.equal('preserved');
+      });
+    });
+  });
+
+  it('shows a localized empty state after a selector returns no models', () => {
+    cy.viewport(1280, 900);
+    cy.request('/api/v1/overview?tier=75').then(({ body }) => {
+      cy.intercept('GET', '**/api/v1/overview?tier=75', {
+        statusCode: 200,
+        body: { ...body, models: [] },
+      }).as('emptyOverview');
+      cy.visit('/zh/overview');
+
+      selectOverviewTier(75);
+      cy.wait('@emptyOverview');
+      cy.get('[data-testid="overview-empty-state"]')
+        .should('have.attr', 'role', 'status')
+        .and('have.text', '没有符合当前筛选条件的总览结果。');
+      cy.get('[data-testid="overview-desktop-matrix"]').should('not.exist');
+      cy.get('[data-testid="overview-mobile-list"]').should('not.exist');
+    });
   });
 
-  it('rewrites one history entry when the overview request fails', () => {
+  it('keeps one history entry when the overview request fails', () => {
     cy.viewport(1280, 900);
     cy.visit('/inference');
     cy.visit('/overview');
@@ -250,19 +394,27 @@ describe('Overview page', () => {
 
     cy.window().then((win) => {
       const before = win.history.length;
-      cy.get('[data-testid="overview-tier-switcher"]').contains('a', '75').click();
+      selectOverviewTier(75);
       cy.wait('@overviewJsonFailure');
-      cy.location('search', { timeout: 15_000 }).should('eq', '?tier=75');
+      cy.location('search').should('eq', '?tier=75');
       // A plain `.should` would be satisfied by the transient extra entry.
       cy.window().then((after) => {
         expect(after.history.length - before, 'one entry for one selection').to.equal(1);
       });
+      // The unsuccessful stop stays retryable: activating it again re-requests
+      // without stacking another history entry for the byte-identical href.
+      selectOverviewTier(75);
+      cy.wait('@overviewJsonFailure');
+      cy.location('search').should('eq', '?tier=75');
+      cy.window().then((after) => {
+        expect(after.history.length - before, 'retry adds no history entry').to.equal(1);
+      });
     });
 
     cy.go('back');
-    cy.location('search', { timeout: 15_000 }).should('eq', '');
+    cy.location('search').should('eq', '');
     cy.go('back');
-    cy.location('pathname', { timeout: 15_000 }).should('eq', '/inference');
+    cy.location('pathname').should('eq', '/inference');
   });
 
   it('warms a hovered option and derives the reference without a request', () => {
@@ -273,22 +425,27 @@ describe('Overview page', () => {
       jsonRequests += 1;
     }).as('overviewJson');
 
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '100').trigger('pointerover');
+    cy.get('[data-testid="overview-engine-scope-switcher"]')
+      .find('[data-overview-engine-scope="all"]')
+      .trigger('pointerover');
     cy.wait('@overviewJson');
     cy.location('search').should('eq', '');
     cy.then(() => {
       expect(jsonRequests, 'hover warms exactly one response').to.equal(1);
     });
 
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '100').click();
-    cy.location('search').should('eq', '?tier=100');
+    cy.get('[data-testid="overview-engine-scope-switcher"]')
+      .find('[data-overview-engine-scope="all"]')
+      .click();
+    cy.location('search').should('eq', '?engine=all');
+    cy.get('[data-overview-engine-scope="all"]').should('have.attr', 'aria-current', 'true');
     cy.then(() => {
       expect(jsonRequests, 'the click reuses the warmed response').to.equal(1);
     });
 
     cy.get('[data-testid="overview-reference-select"]').click();
     cy.get('[data-overview-reference="b300"]').click();
-    cy.location('search', { timeout: 15_000 }).should('eq', '?tier=100&ref=b300');
+    cy.location('search').should('eq', '?engine=all&ref=b300');
     cy.get('[data-overview-comparison="hardware"]').should('contain.text', 'vs B300');
     cy.then(() => {
       expect(jsonRequests, 'a reference change is derived, not fetched').to.equal(1);
@@ -300,6 +457,8 @@ describe('Overview page', () => {
     cy.visit('/overview');
 
     cy.get('[data-overview-comparison="30d"]').click();
+    // Explicit timeout retained: this waits on a real /api/v1/overview response,
+    // not a client-side URL change.
     cy.get('[data-overview-comparison="30d"]', { timeout: 15_000 }).should(
       'have.attr',
       'aria-current',
@@ -321,7 +480,8 @@ describe('Overview page', () => {
     desktopModel('gpt-oss-120b')
       .find('[data-testid="overview-model-category-badge"]')
       .should('have.attr', 'data-category', 'deprecated')
-      .and('contain.text', 'Deprecated');
+      .and('contain.text', 'Deprecated')
+      .and('contain.text', 'Model is no longer actively benchmarked.');
     desktopModel('DeepSeek-R1-0528')
       .find('[data-testid="overview-model-category-badge"]')
       .should('have.attr', 'data-category', 'maintenance');
@@ -335,7 +495,7 @@ describe('Overview page', () => {
       );
     });
 
-    cy.get('[data-testid="overview-tier-switcher"]').contains('a', '75').click();
+    selectOverviewTier(75);
     cy.location('search').should('eq', '?tier=75&models=all');
     cy.get('[data-testid="overview-desktop-model"][data-model="gpt-oss-120b"]').should('exist');
 
@@ -350,13 +510,16 @@ describe('Overview page', () => {
     cy.viewport(1280, 900);
     cy.visit('/zh/overview?models=all');
 
-    cy.get('[data-testid="overview-model-scope-toggle"]').should(
-      'contain.text',
-      '隐藏已弃用与维护模式模型',
-    );
+    // The chip carries the short label; the full sentence stays on the
+    // accessible name and hover title.
+    cy.get('[data-testid="overview-model-scope-toggle"]')
+      .should('contain.text', '隐藏停用模型')
+      .find('[data-overview-model-scope="default"]')
+      .should('have.attr', 'aria-label', '隐藏已弃用与维护模式模型');
     desktopModel('gpt-oss-120b')
       .find('[data-testid="overview-model-category-badge"]')
-      .should('contain.text', '已弃用');
+      .should('contain.text', '已弃用')
+      .and('contain.text', '目前已不再对该模型进行持续基准测试。');
   });
 
   it('uses a selectable hardware reference and preserves it across overview controls', () => {
@@ -379,25 +542,46 @@ describe('Overview page', () => {
       platform('b200').find('[data-testid="overview-cost-delta"]').should('exist');
     });
 
-    cy.get('[data-testid="overview-tier-switcher"]')
-      .contains('a', '100')
-      .should('have.attr', 'href', '/overview?tier=100&ref=b300');
+    selectOverviewTier(100);
+    cy.location('search').should('eq', '?tier=100&ref=b300');
     cy.get('[data-testid="overview-engine-scope-switcher"]')
       .find('[data-overview-engine-scope="all"]')
-      .should('have.attr', 'href', '/overview?engine=all&ref=b300');
+      .should('have.attr', 'href', '/overview?tier=100&engine=all&ref=b300');
     cy.get('[data-overview-comparison="30d"]').should(
       'have.attr',
       'href',
-      '/overview?ref=b300&compare=30d',
+      '/overview?tier=100&ref=b300&compare=30d',
     );
     cy.get('[data-testid="language-toggle"]')
-      .should('have.attr', 'href', '/zh/overview?ref=b300')
+      .should('have.attr', 'href', '/zh/overview?tier=100&ref=b300')
       .click();
     cy.location('pathname').should('eq', '/zh/overview');
-    cy.location('search').should('eq', '?ref=b300');
+    cy.location('search').should('eq', '?tier=100&ref=b300');
     cy.get('[data-overview-comparison="hardware"]').should('contain.text', '对比 B300');
   });
 
+  it('sends one RSC request when a slow language switch preserves overview state', () => {
+    let zhOverviewRscRequests = 0;
+
+    cy.intercept('GET', '/zh/overview*', (request) => {
+      if (request.query._rsc === undefined && request.headers.rsc !== '1') return;
+
+      zhOverviewRscRequests += 1;
+      request.continue((response) => {
+        response.setDelay(600);
+      });
+    });
+
+    cy.visit('/overview?tier=100&ref=b300');
+    cy.get('[data-testid="language-toggle"]')
+      .should('have.attr', 'href', '/zh/overview?tier=100&ref=b300')
+      .click();
+
+    cy.location('pathname').should('eq', '/zh/overview');
+    cy.location('search').should('eq', '?tier=100&ref=b300');
+    cy.then(() => expect(zhOverviewRscRequests).to.equal(1));
+  });
+
   it('uses rack SKU labels when GB200 or GB300 is the comparison reference', () => {
     cy.viewport(1280, 900);
     cy.visit('/overview?ref=gb200');
@@ -410,8 +594,11 @@ describe('Overview page', () => {
     cy.viewport(1280, 900);
     cy.visit('/overview');
 
-    cy.get('[data-testid="overview-page"]')
-      .children('[data-testid="overview-comparison-switcher"]')
+    // The tabs sit inside the surface handed to the Fullscreen API so that
+    // presenting keeps them with the matrix, which costs them their old spot as
+    // a direct child of the page.
+    cy.get('[data-testid="overview-presentation-surface"]')
+      .find('[data-testid="overview-comparison-switcher"]')
       .should('have.length', 1)
       .and('have.class', 'justify-center');
     cy.get('[data-testid="overview-comparison-switcher"]')
@@ -440,17 +627,16 @@ describe('Overview page', () => {
     cy.get('[data-testid="overview-desktop-matrix"] thead').should('contain.text', 'B200');
     cy.get('[data-testid="overview-desktop-matrix"] thead').should('not.contain.text', 'Reference');
 
-    cy.get('[data-testid="overview-tier-switcher"]')
-      .contains('a', '100')
-      .should('have.attr', 'href', '/overview?tier=100&compare=30d');
+    selectOverviewTier(100);
+    cy.location('search').should('eq', '?tier=100&compare=30d');
     cy.get('[data-testid="overview-engine-scope-switcher"]')
       .find('[data-overview-engine-scope="all"]')
-      .should('have.attr', 'href', '/overview?engine=all&compare=30d');
+      .should('have.attr', 'href', '/overview?tier=100&engine=all&compare=30d');
     cy.get('[data-testid="language-toggle"]')
-      .should('have.attr', 'href', '/zh/overview?compare=30d')
+      .should('have.attr', 'href', '/zh/overview?tier=100&compare=30d')
       .click();
     cy.location('pathname').should('eq', '/zh/overview');
-    cy.location('search').should('eq', '?compare=30d');
+    cy.location('search').should('eq', '?tier=100&compare=30d');
     cy.get('[data-testid="overview-comparison-switcher"]')
       .should('have.attr', 'aria-label', '对比方式')
       .within(() => {
@@ -459,8 +645,109 @@ describe('Overview page', () => {
           .should('have.attr', 'aria-current', 'true')
           .and('have.text', '对比 1 个月前');
       });
-    cy.contains('当前成本及其相对 30–60 天前最近一次有效平台结果的变化。').should('exist');
-    cy.contains('缺少有效 30 天对比的平台仅显示当前成本。').should('exist');
+    // The prose is editorial; the stable contract is that both history notes
+    // render the selected window while the caption also renders its upper bound.
+    cy.get('[data-testid="overview-methodology"] p')
+      .eq(0)
+      .should('be.visible')
+      .and('contain.text', '30')
+      .should('contain.text', '60');
+    cy.get('[data-testid="overview-methodology"] p')
+      .eq(1)
+      .should('be.visible')
+      .and('contain.text', '30');
+  });
+
+  it('keeps client-only presentation intent across selectors without forcing fullscreen on load', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview?tier=75&present=1&utm_source=deck#matrix', {
+      onBeforeLoad: stubFullscreenApi,
+    });
+
+    cy.get('[data-testid="overview-presentation-surface"]').should(
+      'have.attr',
+      'data-presenting',
+      'false',
+    );
+    cy.get('[data-testid="language-toggle"]').should(
+      'have.attr',
+      'href',
+      '/zh/overview?tier=75&present=1&utm_source=deck',
+    );
+
+    cy.get('[data-testid="overview-present-toggle"]').click();
+    cy.get('[data-testid="overview-presentation-surface"]').should(
+      'have.attr',
+      'data-presenting',
+      'true',
+    );
+    cy.location('search').should('eq', '?tier=75&present=1&utm_source=deck');
+    cy.location('hash').should('eq', '#matrix');
+
+    selectOverviewTier(100);
+    cy.location('search').should('eq', '?tier=100&present=1&utm_source=deck');
+    cy.get('[data-testid="overview-presentation-surface"]').should(
+      'have.attr',
+      'data-presenting',
+      'true',
+    );
+
+    cy.get('[data-overview-comparison="30d"]').click();
+    cy.location('search').should('eq', '?tier=100&compare=30d&present=1&utm_source=deck');
+    cy.get('[data-testid="overview-presentation-surface"]').should(
+      'have.attr',
+      'data-presenting',
+      'true',
+    );
+
+    cy.get('[data-testid="overview-present-toggle"]').click();
+    cy.get('[data-testid="overview-presentation-surface"]').should(
+      'have.attr',
+      'data-presenting',
+      'false',
+    );
+    cy.location('search').should('eq', '?tier=100&compare=30d&utm_source=deck');
+    cy.location('hash').should('eq', '#matrix');
+    cy.get('[data-testid="language-toggle"]').should(
+      'have.attr',
+      'href',
+      '/zh/overview?tier=100&compare=30d&utm_source=deck',
+    );
+  });
+
+  it('reasserts presentation intent when browser history changes during fullscreen', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview?tier=75', { onBeforeLoad: stubFullscreenApi });
+
+    // A selector push creates a real same-document history entry.
+    selectOverviewTier(100);
+    cy.location('search').should('eq', '?tier=100');
+    cy.get('[data-testid="overview-present-toggle"]').click();
+    cy.location('search').should('eq', '?tier=100&present=1');
+
+    cy.window().then((win) => win.history.back());
+    cy.location('search').should('eq', '?tier=75&present=1');
+    cy.get('[data-testid="overview-presentation-surface"]').should(
+      'have.attr',
+      'data-presenting',
+      'true',
+    );
+  });
+
+  it('leaves comparison paging to a focused presentation select option', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview?compare=30d', { onBeforeLoad: stubFullscreenApi });
+    cy.get('[data-testid="overview-present-toggle"]').click();
+
+    cy.get('[data-testid="overview-history-window-select"]').click();
+    cy.get('[role="option"]').first().focus().trigger('keydown', {
+      key: 'ArrowRight',
+      code: 'ArrowRight',
+      bubbles: true,
+    });
+
+    cy.location('search').should('eq', '?compare=30d&present=1');
+    cy.get('[data-overview-comparison="30d"]').should('have.attr', 'aria-current', 'true');
   });
 
   it('switches the history window through the embedded selector', () => {
@@ -469,7 +756,7 @@ describe('Overview page', () => {
 
     cy.get('[data-testid="overview-history-window-select"]').click();
     cy.get('[data-overview-window="7d"]').click();
-    cy.location('search', { timeout: 15_000 }).should('eq', '?compare=7d');
+    cy.location('search').should('eq', '?compare=7d');
     cy.get('[data-overview-comparison="7d"]').should('have.attr', 'aria-current', 'true');
     cy.contains(
       'Current cost and change versus the latest validated platform result 7–14 days earlier.',
@@ -477,7 +764,7 @@ describe('Overview page', () => {
 
     cy.get('[data-testid="overview-history-window-select"]').click();
     cy.get('[data-overview-window="90d"]').click();
-    cy.location('search', { timeout: 15_000 }).should('eq', '?compare=90d');
+    cy.location('search').should('eq', '?compare=90d');
     cy.contains(
       'Current cost and change versus the latest validated platform result 90–180 days earlier.',
     ).should('exist');
@@ -573,8 +860,11 @@ describe('Overview page', () => {
 
         cy.get('[data-testid="inference-chart-display"]').should('exist');
         cy.get('[data-testid="model-selector"]').click();
-        cy.contains('[role="option"]', 'DeepSeek V4 Pro 1.6T').click();
-        cy.get('[data-testid="model-selector"]').should('contain.text', 'DeepSeek V4 Pro 1.6T');
+        cy.contains('[role="option"]', 'DeepSeek V4 Pro 0813 1.6T').click();
+        cy.get('[data-testid="model-selector"]').should(
+          'contain.text',
+          'DeepSeek V4 Pro 0813 1.6T',
+        );
         cy.get('[data-testid="inference-chart-display"]').should(
           'not.contain.text',
           'No data available',
@@ -598,6 +888,47 @@ describe('Overview page', () => {
       });
   });
 
+  it('keeps a benchmark result usable without comparison dates', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview');
+
+    desktopModel('Qwen-3.5-397B-A17B', SINGLE_TURN)
+      .find(
+        '[data-testid="overview-pair-value"][data-hardware="mi355x"] [data-testid="overview-cost-evidence-link"]',
+      )
+      .then(($link) => {
+        const href = String($link.attr('href'));
+        expect(href).not.to.contain('i_dates=');
+        cy.visit(href);
+
+        cy.contains('Comparison Date Range').should('be.visible');
+        cy.contains('button', 'Select date range')
+          .should('not.have.class', 'animate-pulse')
+          .and('not.have.class', 'border-red-500');
+        cy.get('[data-testid="scatter-graph"]')
+          .should('be.visible')
+          .find('svg .dot-group')
+          .should('have.length.greaterThan', 0);
+        cy.contains('Select a date range or add a run to view chip comparison').should('not.exist');
+      });
+  });
+
+  it('does not require comparison dates when opening AgentX details', () => {
+    cy.viewport(1280, 900);
+    cy.visit('/overview');
+
+    desktopModel('Qwen-3.5-397B-A17B', AGENTX)
+      .contains('a', 'View details')
+      .then(($link) => {
+        const href = String($link.attr('href'));
+        expect(href).to.contain('i_seq=agentic-traces');
+        expect(href).not.to.contain('i_dates=');
+        cy.visit(href);
+
+        cy.contains('Comparison Date Range').should('not.exist');
+      });
+  });
+
   it('keeps the historical comparison complete and non-scrolling across desktop, tablet and phone', () => {
     for (const width of [320, 390, 768, 1024, 1279, 1280, 1440]) {
       cy.viewport(width, 900);
@@ -626,6 +957,48 @@ describe('Overview page', () => {
     }
   });
 
+  // Clicking the control, not just building its href: the matrix reads from a
+  // client data cache, and a cache keyed without the row params moves the
+  // address bar while leaving every row on screen.
+  it('narrows the matrix when the row filter is clicked in either comparison mode', () => {
+    cy.viewport(1280, 900);
+
+    for (const [href, attribute, key, sentence] of [
+      ['/overview?compare=30d', 'data-overview-row-scope', 'rows=changed', 'no 30-day change'],
+      [
+        '/overview',
+        'data-overview-hardware-row-scope',
+        'hwrows=priced',
+        'no result on any platform',
+      ],
+    ] as const) {
+      cy.visit(href);
+      cy.get('[data-testid="overview-desktop-model"]').should('have.length', MATRIX_ROWS);
+
+      cy.get(`a[${attribute}]`)
+        .should('contain.text', 'Hide ')
+        .invoke('text')
+        .then((label) => {
+          const hidden = Number(/\d+/.exec(label)?.[0]);
+          expect(hidden, `hidden row count in "${label}"`).to.be.greaterThan(0);
+
+          cy.get(`a[${attribute}]`).click();
+          cy.location('search').should('contain', key);
+          cy.get('[data-testid="overview-desktop-model"]').should(
+            'have.length',
+            MATRIX_ROWS - hidden,
+          );
+          // The action label names the click, so it flips once the scope
+          // lands; the chip keeps the short label and the counted sentence
+          // rides the accessible name.
+          cy.get(`a[${attribute}]`)
+            .should('contain.text', 'Show all rows')
+            .and('contain.text', String(hidden))
+            .and('have.attr', 'aria-label', `Show ${hidden} rows with ${sentence}`);
+        });
+    }
+  });
+
   it('defaults to community engine scope and switches with canonical links preserving tier and locale', () => {
     cy.viewport(1280, 900);
     cy.visit('/overview');
@@ -655,7 +1028,7 @@ describe('Overview page', () => {
     });
 
     desktopModel('GLM-5.2').within(() => {
-      cy.get('[data-testid="overview-model-scenario"]').should('have.text', AGENTX_LABEL);
+      expectAgentxScenario(AGENTX_LABEL);
       cy.get('[data-testid="overview-pair-missing"]').should('have.length', 5);
     });
     cy.get(
@@ -663,10 +1036,7 @@ describe('Overview page', () => {
     ).click();
     cy.location('search').should('eq', '?engine=all');
     desktopModel('GLM-5.2').find('[data-testid="overview-pair-missing"]').should('have.length', 5);
-    cy.get('[data-testid="overview-tier-switcher"]')
-      .contains('a', '100')
-      .should('have.attr', 'href', '/overview?tier=100&engine=all')
-      .click();
+    selectOverviewTier(100);
     cy.location('search').should('eq', '?tier=100&engine=all');
 
     cy.get('[data-testid="overview-engine-scope-switcher"]')
@@ -720,12 +1090,12 @@ describe('Overview page', () => {
           .should(
             'have.attr',
             'title',
-            'Estimated from validated benchmark runs. Open raw source dashboard for Jul 18: DeepSeek V4 Pro 1.6T · B200 · SGLang · FP4 · MTP',
+            'Estimated from validated benchmark runs. Open raw source dashboard for Jul 18: DeepSeek V4 Pro 0813 1.6T · B200 · SGLang · FP4 · MTP',
           )
           .and(
             'have.attr',
             'aria-label',
-            'Approximately $0.059. Estimated from validated benchmark runs. Open raw source dashboard for Jul 18: DeepSeek V4 Pro 1.6T · B200 · SGLang · FP4 · MTP',
+            'Approximately $0.059. Estimated from validated benchmark runs. Open raw source dashboard for Jul 18: DeepSeek V4 Pro 0813 1.6T · B200 · SGLang · FP4 · MTP',
           );
         cy.get('[data-testid="overview-pair-missing"]').should('not.exist');
       });
@@ -837,8 +1207,8 @@ describe('Overview page', () => {
       .and('not.contain.text', '8K→1K');
     // The TCO model behind every $/GPU/hr is cited under the metric.
     cy.get('[data-testid="overview-source-link"]')
-      .should('have.text', SOURCE_LINK_TEXT)
-      .and('have.attr', 'href', SOURCE_HREF)
+      .should('have.text', TCO_SOURCE_TITLE)
+      .and('have.attr', 'href', TCO_SOURCE_URL)
       .and('have.attr', 'target', '_blank')
       .and('have.attr', 'rel', 'noopener noreferrer');
     // Opening off-site is signalled by the shared external-link glyph.
@@ -895,9 +1265,9 @@ describe('Overview page', () => {
       cy.get('[data-testid="overview-desktop-matrix"]').should('contain.text', label);
     }
     for (const model of ['Kimi-K3', 'GLM-5.2']) {
-      desktopModel(model)
-        .find('[data-testid="overview-model-scenario"]')
-        .should('have.text', AGENTX_LABEL);
+      desktopModel(model).within(() => {
+        expectAgentxScenario(AGENTX_LABEL);
+      });
     }
     for (const model of ['DeepSeek-V4-Pro', 'MiniMax-M3', 'Qwen-3.5-397B-A17B']) {
       desktopModel(model, SINGLE_TURN)
@@ -914,14 +1284,15 @@ describe('Overview page', () => {
       'have.length',
       2,
     );
-    // Single-turn first, AgentX directly below it, both under the same label.
+    // The AgentX row sits in the leading AgentX group; the single-turn row
+    // follows in the 8K/1K group below it, both under the same label.
     cy.get('[data-testid="overview-desktop-model"][data-model="DeepSeek-V4-Pro"]').then(($rows) => {
-      expect([...$rows].map((row) => row.dataset.scenario)).to.deep.equal([SINGLE_TURN, AGENTX]);
+      expect([...$rows].map((row) => row.dataset.scenario)).to.deep.equal([AGENTX, SINGLE_TURN]);
     });
 
     desktopModel('DeepSeek-V4-Pro', AGENTX).within(() => {
-      cy.get('[data-testid="overview-model-scenario"]').should('have.text', AGENTX_LABEL);
-      cy.contains('DeepSeek V4 Pro 1.6T').should('exist');
+      expectAgentxScenario(AGENTX_LABEL);
+      cy.contains('DeepSeek V4 Pro 0813 1.6T').should('exist');
       // Priced from the AgentX rows alone — the single-turn sweep never leaks in.
       cy.get(
         '[data-testid="overview-pair-value"][data-hardware="b200"] [data-testid="overview-cost-evidence-link"]',
@@ -939,14 +1310,14 @@ describe('Overview page', () => {
       cy.contains('a', 'View details').should(
         'have.attr',
         'aria-label',
-        `View details: DeepSeek V4 Pro 1.6T · ${AGENTX_LABEL}`,
+        `View details: DeepSeek V4 Pro 0813 1.6T · ${AGENTX_LABEL}`,
       );
     });
     desktopModel('DeepSeek-V4-Pro', SINGLE_TURN).within(() => {
       cy.contains('a', 'View details').should(
         'have.attr',
         'aria-label',
-        'View details: DeepSeek V4 Pro 1.6T · 8K/1K',
+        'View details: DeepSeek V4 Pro 0813 1.6T · 8K/1K',
       );
     });
   });
@@ -1085,7 +1456,7 @@ describe('Overview page', () => {
       .should('not.match', /∞\s*%/);
   });
 
-  it('re-renders the whole matrix at the service level the URL names, via plain links', () => {
+  it('re-renders the whole matrix from the six-stop SLO slider', () => {
     cy.viewport(1280, 900);
     cy.visit('/overview');
 
@@ -1095,22 +1466,42 @@ describe('Overview page', () => {
       .and('contain.text', 'SLO');
     cy.get('body').should('not.contain.text', 'Service level');
     cy.get('[data-testid="overview-tier-switcher"]').within(() => {
-      cy.get('[aria-current="page"]').should('have.text', '50');
-      // 30 / 75 / 100 / 150 / 200 link out; the active 50 is inert text.
-      cy.get('a').should('have.length', 5);
-      cy.contains('a', '30').should('have.attr', 'href', '/overview?tier=30');
-      cy.contains('a', '150').should('have.attr', 'href', '/overview?tier=150');
-      cy.contains('a', '200').should('have.attr', 'href', '/overview?tier=200');
-      cy.contains('a', '100').should('have.attr', 'href', '/overview?tier=100').click();
+      cy.get('a, button').should('not.exist');
+      cy.get('[data-tier-option]').should('have.length', 6);
+      cy.get('[data-tier-option="50"]').should('have.attr', 'data-selected', 'true');
+      cy.get('[data-testid="overview-tier-ridge"]').should('have.length', 6);
+      cy.get('[data-tier-ridge="30"]').should('have.attr', 'data-state', 'filled');
+      cy.get('[data-tier-ridge="50"]').should('have.attr', 'data-state', 'selected');
+      cy.get('[data-tier-ridge="75"]').should('have.attr', 'data-state', 'unfilled');
+      cy.get('[data-testid="overview-tier-slider"]')
+        .should('have.attr', 'min', '0')
+        .and('have.attr', 'max', '5')
+        .and('have.attr', 'step', '1')
+        .and('have.value', '1')
+        .and('have.attr', 'data-tier', '50')
+        .and('have.attr', 'aria-valuetext', '50 tok/s/user');
     });
 
+    selectOverviewTier(100);
+
     cy.location('search').should('eq', '?tier=100');
     // The metric line never repeats the tier — the switcher states it.
     cy.get('[data-testid="overview-scope"]').should('have.text', SCOPE_LINE);
     cy.get('[data-testid="overview-tier-switcher"]').within(() => {
-      cy.get('[aria-current="page"]').should('have.text', '100');
-      cy.contains('a', '50').should('have.attr', 'href', '/overview');
+      cy.get('[data-tier-option="100"]').should('have.attr', 'data-selected', 'true');
+      cy.get('[data-tier-ridge="75"]').should('have.attr', 'data-state', 'filled');
+      cy.get('[data-tier-ridge="100"]').should('have.attr', 'data-state', 'selected');
+      cy.get('[data-tier-ridge="150"]').should('have.attr', 'data-state', 'unfilled');
+      cy.get('[data-testid="overview-tier-slider"]')
+        .should('have.attr', 'type', 'range')
+        .should('have.value', '3')
+        .and('have.attr', 'data-tier', '100')
+        .and('have.attr', 'aria-valuetext', '100 tok/s/user');
     });
+    selectOverviewTier(75);
+    cy.location('search').should('eq', '?tier=75');
+    selectOverviewTier(100);
+    cy.location('search').should('eq', '?tier=100');
 
     desktopModel('Qwen-3.5-397B-A17B', SINGLE_TURN).within(() => {
       platform('b200').should('contain.text', '$0.124').and('contain.text', 'FP8');
@@ -1374,13 +1765,13 @@ describe('Overview page', () => {
       .should('have.attr', 'href', '/zh/overview')
       .and('contain.text', '总览');
     cy.contains('h1', PAGE_TITLE_ZH).should('exist');
-    cy.contains('按各模型标注的场景，基于各平台最佳观测服务包络线计算每百万总 token 成本').should(
-      'exist',
-    );
+    cy.contains(
+      '根据各模型标注的测试场景，采用各平台实测的最优服务包络线，计算每百万总 token 成本',
+    ).should('exist');
     cy.get('[data-testid="overview-scope"]').should('have.text', SCOPE_LINE_ZH);
     cy.get('[data-testid="overview-source-link"]')
-      .should('have.text', SOURCE_LINK_TEXT)
-      .and('have.attr', 'href', SOURCE_HREF);
+      .should('have.text', TCO_SOURCE_TITLE)
+      .and('have.attr', 'href', TCO_SOURCE_URL);
     cy.get('body').should('not.contain.text', '一眼对比');
     cy.contains('— = 无结果。∞ = 缺少 B200 基线。').should('exist');
     cy.get('[data-testid="overview-methodology"]').children('p').should('have.length', 2);
@@ -1398,11 +1789,11 @@ describe('Overview page', () => {
       )
       .as('estimatedB200')
       .invoke('attr', 'title')
-      .should('include', '根据已验证的基准运行结果估算。')
-      .and('include', '原始数据仪表板：DeepSeek V4 Pro 1.6T · B200 · SGLang · FP4 · MTP');
+      .should('include', '根据已验证的基准测试结果估算。')
+      .and('include', '原始数据仪表板：DeepSeek V4 Pro 0813 1.6T · B200 · SGLang · FP4 · MTP');
     cy.get('@estimatedB200')
       .invoke('attr', 'aria-label')
-      .should('include', '约 $0.059。根据已验证的基准运行结果估算。');
+      .should('include', '约 $0.059。根据已验证的基准测试结果估算。');
     cy.get('@estimatedB200').should('have.text', '$0.059');
     cy.get('@estimatedB200')
       .should('have.attr', 'href')
@@ -1435,19 +1826,19 @@ describe('Overview page', () => {
         .and('have.attr', 'title', '缺少可比较的 B200 基线');
     });
     desktopModel('GLM-5.2').within(() => {
-      cy.get('[data-testid="overview-model-scenario"]').should('have.text', AGENTX_LABEL_ZH);
+      expectAgentxScenario(AGENTX_LABEL_ZH);
       cy.get('[data-testid="overview-pair-missing"]').should('have.length', 5);
       platform('b300')
         .find('[data-testid="overview-pair-missing"]')
         .should('contain.text', '该场景暂无数据');
     });
     desktopModel('DeepSeek-V4-Pro', AGENTX).within(() => {
-      cy.get('[data-testid="overview-model-scenario"]').should('have.text', AGENTX_LABEL_ZH);
+      expectAgentxScenario(AGENTX_LABEL_ZH);
       cy.get(
         '[data-testid="overview-pair-value"][data-hardware="b200"] [data-testid="overview-cost-evidence-link"]',
       ).should('have.text', '$0.064');
     });
-    cy.contains('若某款芯片不支持 FP4 推测解码，则采用次优的可用配置。').should('exist');
+    cy.contains('如果某款芯片没有可用的 FP4 投机解码配置，则改用次优配置。').should('exist');
 
     cy.visit('/zh/overview?tier=100');
     cy.get('[data-testid="overview-scope"]').should('have.text', SCOPE_LINE_ZH);
@@ -1456,7 +1847,11 @@ describe('Overview page', () => {
       .and('contain.text', 'SLO');
     cy.get('body').should('not.contain.text', '服务档位');
     cy.get('[data-testid="overview-tier-switcher"]').within(() => {
-      cy.contains('a', '50').should('have.attr', 'href', '/zh/overview');
+      cy.get('[data-testid="overview-tier-slider"]')
+        .should('have.attr', 'data-tier', '100')
+        // tok/s/user is a protected unit that stays untranslated on the zh route.
+        .and('have.attr', 'aria-valuetext', '100 tok/s/user');
+      cy.get('[data-tier-option]').should('have.length', 6);
     });
   });
 });
diff --git a/packages/app/cypress/e2e/resident-sequence-lengths.cy.ts b/packages/app/cypress/e2e/resident-sequence-lengths.cy.ts
new file mode 100644
index 000000000..7f6e6b93c
--- /dev/null
+++ b/packages/app/cypress/e2e/resident-sequence-lengths.cy.ts
@@ -0,0 +1,63 @@
+import { buildTokenLengthSketch } from '@semianalysisai/inferencex-constants';
+
+import { interceptDerivedAgenticMetrics, unlockAgenticGate } from '../support/e2e';
+import { interceptOverlayRun, OVERLAY_RUN_ID } from '../support/overlay-fixtures';
+
+function visitAgenticChart(extraQuery = '') {
+  cy.visit(`/inference?g_model=DeepSeek-V4-Pro&i_seq=agentic-traces&i_prec=fp4${extraQuery}`, {
+    onBeforeLoad(win) {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+      unlockAgenticGate(win);
+    },
+  });
+}
+
+describe('resident sequence-length subtitle', () => {
+  it('summarizes every resident official point without following legend selection', () => {
+    interceptOverlayRun();
+    interceptDerivedAgenticMetrics();
+    cy.intercept('GET', '/api/v1/resident-sequence-lengths*', (request) => {
+      const ids = new URL(request.url).searchParams.get('ids')?.split(',').filter(Boolean) ?? [];
+      expect(ids.length).to.be.greaterThan(1);
+      request.reply({
+        body: {
+          isl: buildTokenLengthSketch(Array.from({ length: 12_345 }, () => 8_192)),
+          osl: buildTokenLengthSketch(Array.from({ length: 12_345 }, () => 1_024)),
+          coveredPoints: ids.length,
+          requestedPoints: ids.length,
+        },
+      });
+    }).as('residentSequenceLengths');
+
+    visitAgenticChart();
+    cy.wait('@residentSequenceLengths');
+    cy.get('[data-testid="resident-sequence-lengths"]')
+      .should('be.visible')
+      .and('contain.text', 'Completed requests across all resident points (n=12,345)')
+      .and('contain.text', 'ISL p50 8.2k')
+      .and('contain.text', 'p95 8.2k')
+      .and('contain.text', 'OSL p50 1k');
+
+    // Legend visibility is downstream of the resident set. Hiding a series
+    // must not remove the subtitle or issue a narrower stats request.
+    cy.get('.sidebar-legend button').first().click({ force: true });
+    cy.get('[data-testid="resident-sequence-lengths"]').should('be.visible');
+    cy.get('@residentSequenceLengths.all').should('have.length', 1);
+  });
+
+  it('does not label official-only stats as covering an unofficial overlay', () => {
+    let requestCount = 0;
+    interceptOverlayRun();
+    interceptDerivedAgenticMetrics();
+    cy.intercept('GET', '/api/v1/resident-sequence-lengths*', (request) => {
+      requestCount += 1;
+      request.reply({ body: {} });
+    });
+
+    visitAgenticChart(`&unofficialrun=${OVERLAY_RUN_ID}`);
+    cy.wait('@unofficialRun');
+    cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
+    cy.get('[data-testid="resident-sequence-lengths"]').should('not.exist');
+    cy.then(() => expect(requestCount).to.equal(0));
+  });
+});
diff --git a/packages/app/cypress/e2e/sanity.cy.ts b/packages/app/cypress/e2e/sanity.cy.ts
index a18e11cc3..89366e088 100644
--- a/packages/app/cypress/e2e/sanity.cy.ts
+++ b/packages/app/cypress/e2e/sanity.cy.ts
@@ -58,11 +58,78 @@ describe('Page Load & Navigation', () => {
       );
     });
   });
+
+  it('opens and preserves a direct link to an FAQ answer', () => {
+    cy.visit('/about#faq-normalized-interactivity', {
+      onBeforeLoad(win) {
+        cy.stub(win.navigator.clipboard, 'writeText').as('writeFaqLink').resolves();
+      },
+    });
+
+    cy.get('#faq-normalized-interactivity')
+      .should('be.visible')
+      .within(() => {
+        cy.contains(
+          'a[href="#faq-normalized-interactivity"]',
+          'What is the difference between E2E Normalized Interactivity and Interactivity?',
+        ).should('be.visible');
+        cy.contains('The normalized value penalizes slow TTFT').should('be.visible');
+        cy.get('[data-testid="faq-copy-link-faq-normalized-interactivity"]')
+          .should('be.visible')
+          .and('have.text', '')
+          .and('have.attr', 'title', 'Copy link')
+          .find('svg.lucide-link')
+          .should('be.visible');
+        cy.get('[data-testid="faq-copy-link-faq-normalized-interactivity"]')
+          .click()
+          .should('have.attr', 'title', 'Copied')
+          .find('svg.lucide-check')
+          .should('be.visible');
+      });
+    cy.get('@writeFaqLink').should(
+      'have.been.calledOnceWith',
+      `${Cypress.config('baseUrl')}/about#faq-normalized-interactivity`,
+    );
+    cy.location('hash').should('eq', '#faq-normalized-interactivity');
+  });
+
+  it('shows a copy-link button for every FAQ question', () => {
+    cy.visit('/about');
+
+    cy.get('[data-testid^="faq-copy-link-"]')
+      .should('have.length', 15)
+      .each(($button) => {
+        cy.wrap($button)
+          .should('be.visible')
+          .and('have.text', '')
+          .and('have.attr', 'title', 'Copy link')
+          .find('svg.lucide-link')
+          .should('be.visible');
+      });
+  });
 });
 
 // Toggle visibility, click behavior, and aria-label are covered by
 // cypress/component/mode-toggle.cy.tsx. Only the reload-persistence test
 // requires a full page load (true e2e concern).
+describe('Splash text', () => {
+  it('announces AgentX on the landing page in both light and dark mode', () => {
+    cy.visit('/', {
+      onBeforeLoad(win) {
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+        win.localStorage.setItem('theme', 'light');
+      },
+    });
+    cy.get('html').should('not.have.class', 'dark');
+    cy.get('[data-testid="splash-text"]').should('be.visible').and('have.text', 'AgentX is here!!');
+
+    // Same splash after switching themes — it is no longer minecraft-only.
+    cy.get('[data-testid="theme-toggle"]').click();
+    cy.get('html').should('have.class', 'dark');
+    cy.get('[data-testid="splash-text"]').should('be.visible').and('have.text', 'AgentX is here!!');
+  });
+});
+
 describe('Theme Toggle', () => {
   it('theme persists across page reload (localStorage)', () => {
     cy.window().then((win) => {
diff --git a/packages/app/cypress/e2e/speed-overlay.cy.ts b/packages/app/cypress/e2e/speed-overlay.cy.ts
deleted file mode 100644
index 1b5c8ef1f..000000000
--- a/packages/app/cypress/e2e/speed-overlay.cy.ts
+++ /dev/null
@@ -1,204 +0,0 @@
-const advancedToggle = () =>
-  cy.contains('figure', 'vs. Interactivity').find('[data-testid="legend-advanced-toggle"]');
-
-describe('Bus / Race Car Speed Overlay', () => {
-  before(() => {
-    cy.visit('/inference', {
-      onBeforeLoad(win) {
-        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
-      },
-    });
-    cy.get('[data-testid="scatter-graph"]').should('be.visible');
-    cy.get('.sidebar-legend').first().should('be.visible');
-  });
-
-  it('toggle is hidden in the collapsed Advanced drawer and is off by default', () => {
-    advancedToggle().should('be.visible').and('have.attr', 'aria-expanded', 'false');
-    cy.get('#scatter-speed-overlay').should('not.exist');
-
-    advancedToggle().click();
-    advancedToggle().should('have.attr', 'aria-expanded', 'true');
-    cy.get('#scatter-speed-overlay').should('exist');
-    cy.get('label[for="scatter-speed-overlay"]').should('contain.text', 'Bus / Race Car');
-    cy.get('#scatter-speed-overlay').should('have.attr', 'data-state', 'unchecked');
-    cy.get('[data-testid="speed-overlay-bus"]').should('not.exist');
-    cy.get('[data-testid="speed-overlay-car"]').should('not.exist');
-  });
-
-  it('toggling on renders bus and car images on the chart', () => {
-    cy.get('#scatter-speed-overlay').click();
-    cy.get('#scatter-speed-overlay').should('have.attr', 'data-state', 'checked');
-    cy.get('[data-testid="speed-overlay-bus"]').should('exist');
-    cy.get('[data-testid="speed-overlay-car"]').should('exist');
-  });
-
-  it('default tput-vs-interactivity chart places bus top-left and car bottom-right', () => {
-    cy.get('[data-testid="speed-overlay-bus"]').should('have.attr', 'data-corner', 'top-left');
-    cy.get('[data-testid="speed-overlay-car"]').should('have.attr', 'data-corner', 'bottom-right');
-  });
-
-  it('switching to a cost metric flips the bus to bottom-left, car to top-right', () => {
-    cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
-    cy.get('[role="option"]')
-      .contains('Cost per Million Total Tokens (Owning - Hyperscaler)')
-      .click({ force: true });
-    cy.get('[data-testid="speed-overlay-bus"]').should('have.attr', 'data-corner', 'bottom-left');
-    cy.get('[data-testid="speed-overlay-car"]').should('have.attr', 'data-corner', 'top-right');
-  });
-
-  it('switching to an input-token throughput (TTFT X) places bus top-right, car bottom-left', () => {
-    cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
-    cy.get('[role="option"]').contains('Input Token Throughput per Chip').click({ force: true });
-    cy.get('[data-testid="speed-overlay-bus"]').should('have.attr', 'data-corner', 'top-right');
-    cy.get('[data-testid="speed-overlay-car"]').should('have.attr', 'data-corner', 'bottom-left');
-  });
-
-  it('toggling off removes the overlay images cleanly', () => {
-    cy.get('#scatter-speed-overlay').click();
-    cy.get('#scatter-speed-overlay').should('have.attr', 'data-state', 'unchecked');
-    cy.get('[data-testid="speed-overlay-bus"]').should('not.exist');
-    cy.get('[data-testid="speed-overlay-car"]').should('not.exist');
-  });
-
-  it('URL param i_speed=1 enables the overlay on first load', () => {
-    cy.visit('/inference?i_speed=1', {
-      onBeforeLoad(win) {
-        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
-      },
-    });
-    cy.get('[data-testid="scatter-graph"]').should('be.visible');
-    advancedToggle().click();
-    cy.get('#scatter-speed-overlay').should('have.attr', 'data-state', 'checked');
-    cy.get('[data-testid="speed-overlay-bus"]').should('exist');
-    cy.get('[data-testid="speed-overlay-car"]').should('exist');
-  });
-});
-
-describe('Donkey / Elytra Minecraft Overlay', () => {
-  before(() => {
-    cy.visit('/inference', {
-      onBeforeLoad(win) {
-        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
-      },
-    });
-    cy.get('[data-testid="scatter-graph"]').should('be.visible');
-    cy.get('.sidebar-legend').first().should('be.visible');
-  });
-
-  it('toggle is hidden in the Advanced drawer and is independent of the bus/car toggle', () => {
-    advancedToggle().should('have.attr', 'aria-expanded', 'false');
-    cy.get('#scatter-minecraft-overlay').should('not.exist');
-
-    advancedToggle().click();
-    cy.get('#scatter-minecraft-overlay').should('exist');
-    cy.get('label[for="scatter-minecraft-overlay"]').should('contain.text', 'Donkey / Elytra');
-    cy.get('#scatter-minecraft-overlay').should('have.attr', 'data-state', 'unchecked');
-    cy.get('[data-testid="speed-overlay-minecraft-slow"]').should('not.exist');
-    cy.get('[data-testid="speed-overlay-minecraft-fast"]').should('not.exist');
-  });
-
-  it('toggling on renders the donkey-with-chest and elytra images', () => {
-    cy.get('#scatter-minecraft-overlay').click();
-    cy.get('#scatter-minecraft-overlay').should('have.attr', 'data-state', 'checked');
-    cy.get('[data-testid="speed-overlay-minecraft-slow"]')
-      .should('exist')
-      .and('have.attr', 'href')
-      .and('include', 'donkey-chest.png');
-    cy.get('[data-testid="speed-overlay-minecraft-fast"]')
-      .should('exist')
-      .and('have.attr', 'href')
-      .and('include', 'elytra.png');
-  });
-
-  it('donkey sits in the same corner as the bus would (batch-friendly side)', () => {
-    cy.get('[data-testid="speed-overlay-minecraft-slow"]').should(
-      'have.attr',
-      'data-corner',
-      'top-left',
-    );
-    cy.get('[data-testid="speed-overlay-minecraft-fast"]').should(
-      'have.attr',
-      'data-corner',
-      'bottom-right',
-    );
-  });
-
-  it('enabling both toggles stacks the second pair inward from the first', () => {
-    cy.get('#scatter-speed-overlay').click();
-    cy.get('#scatter-speed-overlay').should('have.attr', 'data-state', 'checked');
-    cy.get('[data-testid="speed-overlay-bus"]').should('exist');
-    cy.get('[data-testid="speed-overlay-minecraft-slow"]').should('exist');
-    // Both pairs anchor to the same corner — the donkey sits inward (larger x)
-    // when the slow corner is on the left.
-    cy.get('[data-testid="speed-overlay-bus"]').then(($bus) => {
-      const busX = Number($bus.attr('x'));
-      cy.get('[data-testid="speed-overlay-minecraft-slow"]').then(($donkey) => {
-        const donkeyX = Number($donkey.attr('x'));
-        expect(donkeyX).to.be.greaterThan(busX);
-      });
-    });
-  });
-
-  it('URL param i_mc=1 enables the minecraft overlay on first load', () => {
-    cy.visit('/inference?i_mc=1', {
-      onBeforeLoad(win) {
-        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
-      },
-    });
-    cy.get('[data-testid="scatter-graph"]').should('be.visible');
-    advancedToggle().click();
-    cy.get('#scatter-minecraft-overlay').should('have.attr', 'data-state', 'checked');
-    cy.get('[data-testid="speed-overlay-minecraft-slow"]').should('exist');
-    cy.get('[data-testid="speed-overlay-minecraft-fast"]').should('exist');
-    // Bus/car toggle is independent — should remain off and not render.
-    cy.get('#scatter-speed-overlay').should('have.attr', 'data-state', 'unchecked');
-    cy.get('[data-testid="speed-overlay-bus"]').should('not.exist');
-  });
-});
-
-describe('Y-Axis Metric Search', () => {
-  before(() => {
-    cy.visit('/inference', {
-      onBeforeLoad(win) {
-        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
-      },
-    });
-    cy.get('[data-testid="scatter-graph"]').should('be.visible');
-  });
-
-  it('renders a search input inside the Y-axis selector dropdown', () => {
-    cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
-    // Other selectors (legend, MultiSelect) reuse the "Search..." placeholder, so
-    // scope every lookup to the open Y-axis dropdown's content.
-    cy.get('[data-slot="select-content"]')
-      .find('input[placeholder="Search..."]')
-      .should('be.visible');
-    cy.get('body').click(0, 0);
-  });
-
-  it('typing in the search filters the option list across groups', () => {
-    cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
-    cy.get('[data-slot="select-content"]')
-      .find('input[placeholder="Search..."]')
-      .type('input token throughput');
-    cy.get('[data-slot="select-content"]').find('[role="option"]').should('have.length', 2);
-    cy.get('[data-slot="select-content"]')
-      .find('[role="option"]')
-      .first()
-      .should('contain.text', 'Input Token Throughput');
-    cy.get('body').click(0, 0);
-  });
-
-  it('selecting a filtered option applies the metric and closes the dropdown', () => {
-    cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
-    cy.get('[data-slot="select-content"]')
-      .find('input[placeholder="Search..."]')
-      .type('cost per million total');
-    cy.get('[data-slot="select-content"]')
-      .find('[role="option"]')
-      .contains('3 Year Rental')
-      .click({ force: true });
-    cy.get('[data-slot="select-content"]').should('not.exist');
-    cy.get('[data-testid="yaxis-metric-selector"]').should('contain.text', '3 Year Rental');
-  });
-});
diff --git a/packages/app/cypress/e2e/throughput-calculator.cy.ts b/packages/app/cypress/e2e/throughput-calculator.cy.ts
index 70fc19cc7..55a85f9de 100644
--- a/packages/app/cypress/e2e/throughput-calculator.cy.ts
+++ b/packages/app/cypress/e2e/throughput-calculator.cy.ts
@@ -1,3 +1,5 @@
+import { TCO_SOURCE_TITLE, TCO_SOURCE_URL } from '@semianalysisai/inferencex-constants';
+
 import {
   availability as agenticAvailability,
   b300Rows as agenticB300Rows,
@@ -78,9 +80,12 @@ describe('TCO Calculator', () => {
       });
     });
 
-    it('renders Precision multi-selector', () => {
+    it('hides the Precision selector for a single-precision model', () => {
+      // DeepSeek-V4-Pro is FP4-only in the fixtures — with one precision there
+      // is nothing to choose, so the control is hidden entirely.
+      cy.get('[data-testid="calc-cost-selector"]').should('exist');
       cy.get('[data-testid="calculator-controls"]').within(() => {
-        cy.contains('Precision').should('exist');
+        cy.contains('Precision').should('not.exist');
       });
     });
 
@@ -217,6 +222,9 @@ describe('TCO Calculator', () => {
     it('shows TCO badges when cost metric is selected', () => {
       cy.get('[data-testid="calculator-cost-badges"]').should('contain.text', 'TCO $/chip/hr');
       cy.get('[data-testid="calculator-cost-badges"]').should('contain.text', '$');
+      cy.get('[data-testid="calculator-chart-section"]')
+        .contains('a', TCO_SOURCE_TITLE)
+        .should('have.attr', 'href', TCO_SOURCE_URL);
     });
 
     it('displays chart title that updates when metric changes', () => {
@@ -517,10 +525,13 @@ describe('TCO Calculator', () => {
         'contain.text',
         'Disaggregated inference configurations',
       );
-      cy.get('[data-testid="calculator-chart-section"]').should(
-        'contain.text',
-        'throughput per decode chip',
-      );
+      cy.get('[data-testid="calculator-chart-section"]')
+        // The direction is the point of the warning, not the phrasing: on these
+        // token types a disaggregated config reads high. Asserting the old wording
+        // alone let the note be reworded into something that no longer said which
+        // way it was wrong.
+        .should('contain.text', 'per prefill chip and per decode chip')
+        .and('contain.text', 'reads faster per chip than it is');
     });
 
     // A disagg config's input/output cost is attributed to only its prefill or
@@ -541,7 +552,10 @@ describe('TCO Calculator', () => {
         cy.get('body').type('{esc}');
         cy.get('[data-testid="calculator-disagg-cost-note"]')
           .should('be.visible')
-          .and('contain.text', 'cost per decode chip');
+          .and('contain.text', 'per prefill chip and per decode chip')
+          // Cheaper, specifically — the flattering direction is what a reader
+          // comparing costs needs told.
+          .and('contain.text', 'reads cheaper than it is');
       }
 
       cy.get('[data-testid="calc-cost-type-selector"]').click();
@@ -552,10 +566,13 @@ describe('TCO Calculator', () => {
 
     it('shows disaggregated throughput disclaimer for power metric', () => {
       cy.get('[data-testid="calculator-metric-power"]').click();
-      cy.get('[data-testid="calculator-chart-section"]').should(
-        'contain.text',
-        'throughput per decode chip',
-      );
+      cy.get('[data-testid="calculator-chart-section"]')
+        // The direction is the point of the warning, not the phrasing: on these
+        // token types a disaggregated config reads high. Asserting the old wording
+        // alone let the note be reworded into something that no longer said which
+        // way it was wrong.
+        .should('contain.text', 'per prefill chip and per decode chip')
+        .and('contain.text', 'reads faster per chip than it is');
     });
   });
 
@@ -578,7 +595,7 @@ describe('TCO Calculator', () => {
     // open straight to the right model without a flash of the default. See #430.
     it('?g_model= seeds the model selector before client hydration', () => {
       cy.request('/calculator?g_model=DeepSeek-V4-Pro').then((response) => {
-        expect(response.body).to.contain('DeepSeek V4 Pro 1.6T');
+        expect(response.body).to.contain('DeepSeek V4 Pro 0813 1.6T');
         expect(response.body).not.to.contain('DeepSeek R1 0528 671B');
       });
     });
@@ -588,7 +605,10 @@ describe('TCO Calculator', () => {
         win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
       });
       cy.visit('/calculator?g_model=DeepSeek-V4-Pro');
-      cy.get('[data-testid="calc-model-selector"]').should('contain.text', 'DeepSeek V4 Pro 1.6T');
+      cy.get('[data-testid="calc-model-selector"]').should(
+        'contain.text',
+        'DeepSeek V4 Pro 0813 1.6T',
+      );
     });
   });
 
@@ -596,6 +616,21 @@ describe('TCO Calculator', () => {
   // DeepSeek V4 agentic trace calculations
   // ---------------------------------------------------------------------------
 
+  describe('availability failure degradation', () => {
+    it('keeps usable benchmark results visible when availability metadata fails', () => {
+      cy.intercept('GET', '/api/v1/availability*', {
+        statusCode: 503,
+        body: { error: 'temporarily unavailable' },
+      }).as('availabilityFailure');
+
+      cy.visit('/calculator');
+      cy.wait('@availabilityFailure');
+      cy.get('[data-testid="calculator-bar-chart"] svg .bar').should('have.length.greaterThan', 0);
+      cy.get('[data-testid="calculator-controls"]').should('be.visible');
+      cy.contains('Error loading calculator data').should('not.exist');
+    });
+  });
+
   describe('DeepSeek V4 agentic calculations', () => {
     beforeEach(() => {
       const b300Rows = agenticB300Rows(null);
@@ -624,7 +659,10 @@ describe('TCO Calculator', () => {
     });
 
     it('renders throughput and cost calculations from null-ISL/OSL agentic rows', () => {
-      cy.get('[data-testid="calc-sequence-selector"]').should('contain.text', 'Agentic Traces');
+      // The agentic fixture exposes a single scenario, so the scenario
+      // control disappears entirely — no dropdown, no static readout.
+      cy.get('[data-testid="scenario-static-value"]').should('not.exist');
+      cy.get('[data-testid="calc-sequence-selector"]').should('not.exist');
       cy.get('[data-testid="calc-percentile-selector"]').should('contain.text', 'p90');
       cy.get('[data-testid="calculator-no-data"]').should('not.exist');
       cy.get('[data-testid="calculator-bar-chart"] svg .bar').should('have.length', 2);
@@ -681,7 +719,10 @@ describe('TCO Calculator', () => {
       });
       cy.wait('@agenticBenchmarks');
 
-      cy.get('[data-testid="calc-sequence-selector"]').should('contain.text', 'Agentic Traces');
+      // Single-scenario fixture → the scenario control disappears entirely;
+      // the gate is locked, so no percentile selector either.
+      cy.get('[data-testid="scenario-static-value"]').should('not.exist');
+      cy.get('[data-testid="calc-sequence-selector"]').should('not.exist');
       cy.get('[data-testid="calc-percentile-selector"]').should('not.exist');
       cy.get('[data-testid="calculator-chart-section"] h2')
         .first()
@@ -694,10 +735,12 @@ describe('TCO Calculator', () => {
   });
 
   // ---------------------------------------------------------------------------
-  // Fleet planner (MW projection + cost-target inverse lookup)
+  // Cost target: the inverse of the interactivity slider. The MW-budget input
+  // lives here now that the Fleet Lifecycle section moved to its own /fleet
+  // page — see fleet-lifecycle.cy.ts. Both pages share the `c_mw` URL param.
   // ---------------------------------------------------------------------------
 
-  describe('fleet planner', () => {
+  describe('cost target', () => {
     before(() => {
       cy.window().then((win) => {
         win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
@@ -706,42 +749,37 @@ describe('TCO Calculator', () => {
       cy.get('[data-testid="calculator-bar-chart"] svg .bar').should('have.length.greaterThan', 0);
     });
 
-    it('renders both fleet planner cards with empty states', () => {
-      cy.get('[data-testid="calculator-fleet-section"]').should('be.visible');
-      cy.get('[data-testid="calculator-fleet-section"]').should('contain.text', 'Fleet Projection');
-      cy.get('[data-testid="calculator-fleet-empty"]').should('be.visible');
+    it('sits directly under the interactivity slider, since it inverts it', () => {
+      // Reading order is the point: the slider fixes a speed and the cost falls
+      // out; this fixes the cost and the speed falls out. Anything between them
+      // makes the pair read as unrelated.
       cy.get('[data-testid="calculator-costcap-section"]').should('be.visible');
       cy.get('[data-testid="calculator-costcap-empty"]').should('be.visible');
-    });
-
-    it('entering a MW budget renders the fleet projection table', () => {
-      cy.get('[data-testid="calc-fleet-mw-input"]').type('10');
-      cy.get('[data-testid="calculator-fleet-table"]').should('be.visible');
-      cy.get('[data-testid="calculator-fleet-table"]').within(() => {
-        cy.contains('th', 'Chips').should('exist');
-        cy.contains('th', 'Concurrent Users').should('exist');
-        cy.contains('th', 'Fleet $/mo').should('exist');
-        cy.get('tbody tr').should('have.length.greaterThan', 0);
-        // GPU counts are whole, comma-grouped numbers
-        cy.get('tbody tr')
-          .first()
-          .find('td')
-          .eq(1)
-          .invoke('text')
-          .should('match', /^[\d,]+$/u);
+      cy.get('[data-testid="calculator-controls"]').then(($controls) => {
+        cy.get('[data-testid="calculator-costcap-section"]').then(($costcap) => {
+          const order = $controls[0]!.compareDocumentPosition($costcap[0]!);
+          // eslint-disable-next-line no-bitwise
+          expect(order & Node.DOCUMENT_POSITION_FOLLOWING).to.be.greaterThan(0);
+        });
+      });
+      cy.get('[data-testid="calculator-chart-section"]').then(($chart) => {
+        cy.get('[data-testid="calculator-costcap-section"]').then(($costcap) => {
+          const order = $chart[0]!.compareDocumentPosition($costcap[0]!);
+          // eslint-disable-next-line no-bitwise
+          expect(order & Node.DOCUMENT_POSITION_PRECEDING).to.be.greaterThan(0);
+        });
       });
-      cy.get('[data-testid="calculator-fleet-empty"]').should('not.exist');
     });
 
-    it('fleet table shows the utilization and facility-power assumptions', () => {
-      cy.get('[data-testid="calculator-fleet-section"]').should(
-        'contain.text',
-        'Assumes 100% utilization',
-      );
-      cy.get('[data-testid="calculator-fleet-section"]').should(
-        'contain.text',
-        'SemiAnalysis Datacenter Industry Model',
-      );
+    it('owns the MW budget input now that the lifecycle section moved to /fleet', () => {
+      cy.get('[data-testid="calculator-costcap-section"]')
+        .find('[data-testid="calc-costcap-mw-input"]')
+        .should('have.value', '10');
+      cy.get('[data-testid="calculator-lifecycle-section"]').should('not.exist');
+      // A pointer to the new page replaces the section.
+      cy.get('[data-testid="calculator-fleet-pointer-link"]')
+        .should('have.attr', 'href')
+        .and('match', /\/fleet$/);
     });
 
     it('entering a generous cost target renders reachable interactivity per GPU', () => {
@@ -766,23 +804,6 @@ describe('TCO Calculator', () => {
       cy.get('[data-testid="calculator-costcap-table"]').should('contain.text', 'Not reachable');
     });
 
-    it('a budget too small for one GPU shows a dedicated message, not the enter-a-value prompt', () => {
-      cy.get('[data-testid="calc-fleet-mw-input"]').clear();
-      cy.get('[data-testid="calc-fleet-mw-input"]').type('0.0001');
-      cy.get('[data-testid="calculator-fleet-empty"]')
-        .should('be.visible')
-        .and('contain.text', 'too small to power a single chip');
-      cy.get('[data-testid="calculator-fleet-table"]').should('not.exist');
-    });
-
-    it('clearing the MW input restores the empty state', () => {
-      cy.get('[data-testid="calc-fleet-mw-input"]').clear();
-      cy.get('[data-testid="calculator-fleet-empty"]')
-        .should('be.visible')
-        .and('contain.text', 'Enter a facility power budget');
-      cy.get('[data-testid="calculator-fleet-table"]').should('not.exist');
-    });
-
     it('cost-cap table follows legend visibility (soloing a GPU filters its rows)', () => {
       cy.get('[data-testid="calc-costcap-input"]').clear();
       cy.get('[data-testid="calc-costcap-input"]').type('100');
@@ -802,4 +823,43 @@ describe('TCO Calculator', () => {
       });
     });
   });
+
+  describe('folding sections away', () => {
+    before(() => {
+      cy.window().then((win) => {
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+      });
+      cy.visit('/calculator');
+      cy.get('[data-testid="calculator-bar-chart"] svg .bar').should('have.length.greaterThan', 0);
+    });
+
+    it('folds the chart away, keeping its title as the handle', () => {
+      // Folded means unmounted, not hidden: these bodies hold D3 charts, and a
+      // reader who folded one should not still be paying to render it.
+      cy.get('[data-testid="calculator-chart-collapse"]').click();
+      cy.get('[data-testid="calculator-figure"]').should('not.exist');
+      // The title only appears in the header once folded — while open the chart's
+      // own caption carries it, and two copies would be worse than none.
+      cy.get('[data-testid="calculator-chart-section"] h2')
+        .should('have.length', 1)
+        .and('contain.text', 'Total Token Throughput per Chip');
+
+      cy.get('[data-testid="calculator-chart-collapse"]').click();
+      cy.get('[data-testid="calculator-figure"]').should('be.visible');
+      cy.get('[data-testid="calculator-bar-chart"] svg .bar').should('have.length.greaterThan', 0);
+    });
+
+    it('folds the cost-target section away, keeping its heading as the handle', () => {
+      cy.get('[data-testid="calc-costcap-input"]').should('be.visible');
+      cy.get('[data-testid="calculator-costcap-collapse"]').click();
+      cy.get('[data-testid="calc-costcap-input"]').should('not.exist');
+      // The heading stays put, so the folded section still says what it is.
+      cy.get('[data-testid="calculator-costcap-section"]').should(
+        'contain.text',
+        'Interactivity Within a Cost Target',
+      );
+      cy.get('[data-testid="calculator-costcap-collapse"]').click();
+      cy.get('[data-testid="calc-costcap-input"]').should('be.visible');
+    });
+  });
 });
diff --git a/packages/app/cypress/e2e/tokens-per-currency.cy.ts b/packages/app/cypress/e2e/tokens-per-currency.cy.ts
new file mode 100644
index 000000000..b7a4e74d8
--- /dev/null
+++ b/packages/app/cypress/e2e/tokens-per-currency.cy.ts
@@ -0,0 +1,51 @@
+/**
+ * The dashboard now opens on tokens purchasable per $1 USD, and the same
+ * quantities are also available priced in yuan. Fixed-sequence Quick Filters
+ * remain available from the chart legend.
+ */
+describe('Tokens per currency and agentic controls', () => {
+  beforeEach(() => {
+    cy.window().then((win) => {
+      win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+    });
+  });
+
+  it('defaults the y-axis to total tokens per $1 USD', () => {
+    cy.visit('/inference');
+    cy.get('[data-testid="yaxis-metric-selector"]').should(
+      'contain.text',
+      'Total Tokens per $1 USD (Owning - Neocloud Giant)',
+    );
+    cy.get('[data-testid="scatter-graph"]')
+      .first()
+      .find('svg .dot-group')
+      .should('have.length.greaterThan', 0);
+  });
+
+  it('still honors an explicit ?i_metric=, so shared links are unaffected', () => {
+    cy.visit('/inference?i_metric=y_tpPerGpu');
+    cy.get('[data-testid="yaxis-metric-selector"]').should(
+      'contain.text',
+      'Token Throughput per Chip',
+    );
+  });
+
+  it('offers the same quantities priced in yuan', () => {
+    cy.visit('/inference');
+    cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
+    cy.get('[role="option"]')
+      .contains('Total Tokens per ¥1 RMB (Owning - Hyperscaler)')
+      .click({ force: true });
+    cy.get('[data-testid="scatter-graph"]')
+      .first()
+      .find('svg .dot-group')
+      .should('have.length.greaterThan', 0);
+  });
+
+  it('keeps Quick Filters for a fixed-sequence scenario', () => {
+    cy.visit('/inference?i_seq=8k%2F1k');
+    cy.get('[data-testid="scatter-quick-filters"]').click();
+    cy.get('[data-testid="quick-filters-dialog"]').should('be.visible');
+    cy.get('[data-testid="quick-filter-spec-mtp"]').should('exist');
+  });
+});
diff --git a/packages/app/cypress/e2e/ttft-x-axis-toggle.cy.ts b/packages/app/cypress/e2e/ttft-x-axis-toggle.cy.ts
index f62aed384..3dd584e8d 100644
--- a/packages/app/cypress/e2e/ttft-x-axis-toggle.cy.ts
+++ b/packages/app/cypress/e2e/ttft-x-axis-toggle.cy.ts
@@ -1,53 +1,26 @@
 import { interceptDerivedAgenticMetrics, unlockAgenticGate } from '../support/e2e';
+import { agenticMetrics } from '../support/agentic-fixtures';
+import { expandLegendAdvanced } from '../support/legend-advanced';
 
 // This spec exercises the agentic x-axis modes, which only exist when the
-// selected model resolves to the Agentic Traces scenario. The default e2e
-// fixtures (cypress/fixtures/api/*.json) have NO agentic rows for any model, and
-// the app default scenario is 8K/1K, so bare /inference always resolves to a
-// fixed-seq scenario. We therefore inject agentic availability + benchmark rows
-// for the default model VIA SPEC-SCOPED INTERCEPTS (not the shared fixtures) and
-// visit with an explicit ?i_seq=agentic-traces so this test — and only this
-// test — sees the agentic view.
+// selected model resolves to the Agentic scenario. The default e2e
+// fixtures (cypress/fixtures/api/*.json) have NO agentic rows for any model, so
+// bare /inference always resolves to a fixed-seq scenario there. We therefore
+// inject agentic availability + benchmark rows for the default model VIA
+// SPEC-SCOPED INTERCEPTS (not the shared fixtures) so this spec — and only this
+// spec — sees the agentic view. Most cases still pass ?i_seq=agentic-traces
+// explicitly; the "Default scenario" block below covers the implicit path,
+// where availability opens the agentic scenario on its own.
 const DEFAULT_MODEL_DB_KEY = 'dsv4'; // DeepSeek-V4-Pro is the default model
 const AGENTIC_DATE = '2026-06-12';
 
-// Percentile ladder for one metric family (median/p75/p90/p95/p99/std).
-const percentileLadder = (prefix: string, base: number): Record<string, number> => ({
-  [`median_${prefix}`]: base,
-  [`p75_${prefix}`]: base * 1.2,
-  [`p90_${prefix}`]: base * 1.5,
-  [`p95_${prefix}`]: base * 1.7,
-  [`p99_${prefix}`]: base * 2.2,
-  [`std_${prefix}`]: base * 0.3,
-});
-
-const agenticMetrics = (conc: number): Record<string, number> => {
-  const scale = conc / 16;
-  const itl = 0.011 * scale;
-  return {
-    ...percentileLadder('ttft', 0.4 * scale),
-    ...percentileLadder('tpot', 0.012 * scale),
-    ...percentileLadder('itl', itl),
-    ...percentileLadder('e2el', 8 * scale),
-    median_intvty: 1 / itl,
-    p75_intvty: 1 / (itl * 1.2),
-    p90_intvty: 1 / (itl * 1.5),
-    p99_intvty: 1 / (itl * 2.2),
-    std_intvty: (1 / itl) * 0.1,
-    tput_per_gpu: 950 / Math.sqrt(scale),
-    output_tput_per_gpu: 210,
-    input_tput_per_gpu: 740,
-    total_tput_tps: 7600 * conc * 0.05,
-  };
-};
-
 const agenticGpus = [
   { hardware: 'b200', framework: 'vllm', disagg: false },
   { hardware: 'b300', framework: 'vllm', disagg: false },
 ];
 
-// Availability: default model has BOTH agentic and fixed-seq. The agentic view
-// is selected explicitly via ?i_seq=agentic-traces (the app default is 8K/1K).
+// Availability: default model has BOTH agentic and fixed-seq, so the scenario
+// the chart lands on is a real choice rather than the only option.
 const agenticAvailability = [
   ...agenticGpus.map((g) => ({
     model: DEFAULT_MODEL_DB_KEY,
@@ -118,31 +91,42 @@ const interceptAgenticData = () => {
   cy.intercept('GET', '/api/v1/benchmarks*', { body: agenticBenchmarks }).as('benchmarks');
 };
 
+// Same rows re-keyed to another model to prove the default follows availability
+// rather than a hard-coded model registry.
+const OTHER_MODEL_DB_KEY = 'dsr1';
+const otherModelAvailability = agenticAvailability.map((row) => ({
+  ...row,
+  model: OTHER_MODEL_DB_KEY,
+}));
+const otherModelBenchmarks = agenticBenchmarks.map((row, index) => ({
+  ...row,
+  id: 920000 + index,
+  model: OTHER_MODEL_DB_KEY,
+}));
+
 const interceptFixedSequenceData = () => {
   cy.intercept('GET', '/api/v1/availability', { body: agenticAvailability }).as('availability');
   cy.intercept('GET', '/api/v1/benchmarks*', { body: fixedSequenceBenchmarks }).as('benchmarks');
 };
 
 /**
- * On agentic charts, Interactivity / E2E Latency / TTFT live inside the
- * "Advanced" menu rather than as top-level tabs, so they must be revealed
- * before they can be clicked. The menu closes on select, so the post-select
- * assertion targets the trigger, which carries the active state and the
- * selected metric's label.
+ * Every x-axis metric is a top-level tab, on agentic charts as well as fixed
+ * sequences — #736 removed the "Advanced" popover that used to hide
+ * Interactivity / E2E Latency / TTFT behind a trigger. The tab itself now
+ * carries both the selected state and the metric's label.
  */
-function selectAdvancedXAxisMode(mode: 'interactivity' | 'e2e' | 'ttft', label: string) {
-  cy.get('[data-testid="x-axis-mode-advanced"]').click();
+function selectXAxisMode(mode: 'interactivity' | 'e2e' | 'ttft', label: string) {
   cy.get(`[data-testid="x-axis-mode-${mode}"]`).click();
-  cy.get('[data-testid="x-axis-mode-advanced"]')
-    .should('have.attr', 'data-state', 'active')
+  cy.get(`[data-testid="x-axis-mode-${mode}"]`)
+    .should('have.attr', 'aria-selected', 'true')
     .and('contain.text', label);
 }
 
 describe('X-Axis Mode Toggle (inference chart)', () => {
   before(() => {
     interceptAgenticData();
-    // The agentic default mode is E2E Normalized Interactivity, which fetches derived metrics
-    // on first render — stub them before the visit.
+    // Agentic defaults to Interactivity, which needs no derived metrics. The stub
+    // covers the first switch into E2E Normalized Interactivity further down.
     interceptDerivedAgenticMetrics();
     cy.visit('/inference?i_seq=agentic-traces', {
       onBeforeLoad(win) {
@@ -154,39 +138,69 @@ describe('X-Axis Mode Toggle (inference chart)', () => {
     cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
   });
 
-  it('shows E2E Normalized Interactivity by default for the agentic view, as the leftmost option', () => {
-    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic Traces');
-    // The three per-request latency modes are nested under Advanced on agentic.
-    cy.get('[data-testid="x-axis-mode-ttft"]').should('not.exist');
-    cy.get('[data-testid="x-axis-mode-e2e"]').should('not.exist');
-    cy.get('[data-testid="x-axis-mode-interactivity"]').should('not.exist');
-    cy.get('[data-testid="x-axis-mode-advanced"]')
-      .should('be.visible')
-      .and('have.attr', 'data-state', 'inactive');
-    cy.get('[data-testid="x-axis-mode-e2e-normalized-interactivity"]')
-      .should('be.visible')
-      .and('have.attr', 'aria-selected', 'true');
-    // E2E Normalized Interactivity leads the mode list for agentic.
+  // Cypress clears intercepts between tests, so the derived-metrics stub is
+  // re-registered per test rather than once in `before`. Until #736 the
+  // agentic default was E2E Normalized Interactivity, so the fetch happened
+  // during `before` while the stub was still alive and React Query held the
+  // result for the rest of the spec. The default no longer fetches, so any
+  // later switch into that mode issues a fresh request.
+  beforeEach(() => {
+    interceptDerivedAgenticMetrics();
+  });
+
+  it('defaults the agentic view to Interactivity, with all four modes as flat tabs', () => {
+    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic');
+    // #736 made every latency mode a top-level tab and moved the default off E2E
+    // Normalized Interactivity, which still leads the strip without being selected.
+    cy.get('[data-testid="x-axis-mode-advanced"]').should('not.exist');
+    for (const mode of ['e2e-normalized-interactivity', 'interactivity', 'e2e', 'ttft']) {
+      cy.get(`[data-testid="x-axis-mode-${mode}"]`).should('be.visible');
+    }
+    cy.get('[data-testid="x-axis-mode-buttons"] [role="tab"]').should('have.length', 4);
     cy.get('[data-testid="x-axis-mode-buttons"] [role="tab"]')
       .first()
-      .should('have.attr', 'data-testid', 'x-axis-mode-e2e-normalized-interactivity');
-    cy.get('[data-testid="chart-figure"] h2').should(
-      'contain.text',
-      'P90 E2E Normalized Interactivity',
+      .should('have.attr', 'data-testid', 'x-axis-mode-e2e-normalized-interactivity')
+      .and('have.attr', 'aria-selected', 'false');
+    cy.get('[data-testid="x-axis-mode-interactivity"]').should(
+      'have.attr',
+      'aria-selected',
+      'true',
     );
+    cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Interactivity');
     cy.get('[data-testid="chart-figure"] svg').should(
       'contain.text',
-      'P90 E2E Normalized Interactivity (tok/s/user)',
+      'P90 Interactivity (tok/s/user)',
     );
   });
 
-  it('switches to Interactivity and updates the heading', () => {
-    selectAdvancedXAxisMode('interactivity', 'Interactivity');
+  it('offers unobtrusive FAQ help beside E2E Normalized Interactivity', () => {
+    cy.get('[data-testid="normalized-interactivity-faq-link"]')
+      .should('be.visible')
+      .and('have.attr', 'href', '/about#faq-normalized-interactivity')
+      .and('have.attr', 'title', 'What does E2E Normalized Interactivity mean?')
+      .and('have.class', 'no-export')
+      .find('svg[aria-hidden="true"]')
+      .should('be.visible');
+
+    // The metric tabs sit outside the chart capture root, and no-export is a
+    // second guard if this control moves into that tree in the future.
+    cy.get('#chart-0 [data-testid="normalized-interactivity-faq-link"]').should('not.exist');
+  });
+
+  it('switches to E2E Normalized Interactivity and updates the heading', () => {
+    // The first entry into this mode fetches the trace-derived metrics, which the
+    // suite's intercept stubs; the default no longer fetches them on load.
+    cy.get('[data-testid="x-axis-mode-e2e-normalized-interactivity"]').click();
+    cy.get('[data-testid="chart-figure"] h2').should(
+      'contain.text',
+      'P90 E2E Normalized Interactivity',
+    );
+    selectXAxisMode('interactivity', 'Interactivity');
     cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Interactivity');
   });
 
-  it('explains the offload halo in the legend and distinguishes it from plain points', () => {
-    cy.get('#chart-0 [data-testid="offload-halo-key"]')
+  it('explains the offload halo in the info footer and distinguishes it from plain points', () => {
+    cy.get('[data-testid="axis-metric-footer-chart-0"] [data-testid="offload-halo-key"]')
       .should('be.visible')
       .and('contain.text', 'KV offload ON');
     cy.get('#chart-0 .offload-halo').should('have.length.at.least', 1);
@@ -197,30 +211,19 @@ describe('X-Axis Mode Toggle (inference chart)', () => {
     });
   });
 
-  it('keeps the offload halo explanation in the PNG export clone', () => {
-    cy.window().then((win) => {
-      const exportContainer = win.document.querySelector('#chart-0-export');
-      expect(exportContainer).not.to.equal(null);
-      const state = { seen: false };
-      const observer = new win.MutationObserver(() => {
-        if (exportContainer?.querySelector('[data-testid="offload-halo-key"]')) {
-          state.seen = true;
-          observer.disconnect();
-        }
-      });
-      observer.observe(exportContainer!, { childList: true, subtree: true });
-      (win as typeof win & { __offloadHaloExportState: typeof state }).__offloadHaloExportState =
-        state;
-    });
-
-    cy.get('[data-testid="chart-figure"]').first().find('[data-testid="export-button"]').click();
-    cy.get('[data-testid="export-png-button"]').click();
-    cy.window().its('__offloadHaloExportState.seen').should('eq', true);
+  it('keeps the offload halo explanation out of the chart capture tree', () => {
+    // The key lives in the axis-metric info footer (a `no-export` sibling of
+    // the chart), so the PNG export clone — built from #chart-0 — carries the
+    // halo decoration itself but not the textual key.
+    cy.get('[data-testid="axis-metric-footer-chart-0"] [data-testid="offload-halo-key"]').should(
+      'be.visible',
+    );
+    cy.get('#chart-0 [data-testid="offload-halo-key"]').should('not.exist');
   });
 
   it('shows the selected percentile in the Interactivity axis label', () => {
     // Explicitly select the mode — do not rely on the agentic default mode.
-    selectAdvancedXAxisMode('interactivity', 'Interactivity');
+    selectXAxisMode('interactivity', 'Interactivity');
     // Agentic plots percentile fields (p90_intvty), so the axis label carries it.
     cy.get('[data-testid="chart-figure"] svg').should(
       'contain.text',
@@ -228,10 +231,14 @@ describe('X-Axis Mode Toggle (inference chart)', () => {
     );
   });
 
-  it('defaults to parallelism labels without line labels for the agentic view', () => {
+  it('defaults to parallelism, point, and line labels for the agentic view', () => {
+    // Line labels name the curve and point labels name the point, so the
+    // agentic view turns on both — it differs from fixed-seq only in the
+    // parallelism and point labels.
+    expandLegendAdvanced();
     cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'checked');
     cy.get('#scatter-point-labels').should('have.attr', 'data-state', 'checked');
-    cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'unchecked');
+    cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'checked');
   });
 
   it('honors explicit label URL overrides for the agentic view', () => {
@@ -239,25 +246,26 @@ describe('X-Axis Mode Toggle (inference chart)', () => {
     // Fresh page load → fresh React Query cache → the default E2E Normalized Interactivity
     // mode refetches derived metrics.
     interceptDerivedAgenticMetrics();
-    cy.visit('/inference?i_seq=agentic-traces&i_label=0&i_advlabel=0&i_linelabel=1', {
+    cy.visit('/inference?i_seq=agentic-traces&i_label=0&i_advlabel=0&i_linelabel=0', {
       onBeforeLoad(win) {
         win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
         unlockAgenticGate(win);
       },
     });
-    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic Traces');
+    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic');
+    expandLegendAdvanced();
     cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'unchecked');
     cy.get('#scatter-point-labels').should('have.attr', 'data-state', 'unchecked');
-    cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'checked');
+    cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'unchecked');
   });
 
   it('switches the x-axis to TTFT and updates the heading', () => {
-    selectAdvancedXAxisMode('ttft', 'TTFT');
+    selectXAxisMode('ttft', 'TTFT');
     cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Time To First Token');
   });
 
   it('switches the x-axis to E2E Latency and updates the heading', () => {
-    selectAdvancedXAxisMode('e2e', 'E2E Latency');
+    selectXAxisMode('e2e', 'E2E Latency');
     cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'End-to-end Latency');
     cy.get('[data-testid="chart-figure"] svg').should('contain.text', 'P90 End-to-end Latency (s)');
   });
@@ -298,19 +306,12 @@ describe('X-Axis Mode Toggle (inference chart)', () => {
     );
   });
 
-  // Documents the intended pairing of the Advanced menu with manual tab
-  // activation: focus may move to the lone remaining tab without the x-axis
-  // snapping back to it. The load-bearing guard for manual activation is the
-  // 8K/1K test below, where focus-activation is directly observable; this one
-  // states the agentic-side expectation.
-  it('keeps the Advanced selection when the remaining tab is focused', () => {
-    selectAdvancedXAxisMode('ttft', 'TTFT');
-
-    // Let the popover finish closing first: Radix restores focus to its own
-    // trigger on close, which would otherwise win the race against the focus
-    // below and mask the revert this test is guarding against.
-    cy.get('[data-testid="x-axis-mode-advanced-menu"]').should('not.exist');
-    cy.get('[data-testid="x-axis-mode-advanced"]').should('have.focus');
+  // Tabs are manually activated: moving focus along the strip must not change
+  // the x-axis. The load-bearing guard is the 8K/1K test below, where the same
+  // behaviour is asserted on the fixed-sequence strip; this one states the
+  // agentic-side expectation, across the agentic-only mode.
+  it('keeps the selected mode when another tab is focused', () => {
+    selectXAxisMode('ttft', 'TTFT');
 
     cy.get('[data-testid="x-axis-mode-e2e-normalized-interactivity"]').focus();
     cy.get('[data-testid="x-axis-mode-e2e-normalized-interactivity"]').should('have.focus');
@@ -320,14 +321,12 @@ describe('X-Axis Mode Toggle (inference chart)', () => {
       'aria-selected',
       'false',
     );
-    cy.get('[data-testid="x-axis-mode-advanced"]')
-      .should('have.attr', 'data-state', 'active')
-      .and('contain.text', 'TTFT');
+    cy.get('[data-testid="x-axis-mode-ttft"]').should('have.attr', 'aria-selected', 'true');
     cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Time To First Token');
   });
 
   it('switches back to Interactivity', () => {
-    selectAdvancedXAxisMode('interactivity', 'Interactivity');
+    selectXAxisMode('interactivity', 'Interactivity');
     cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Interactivity');
     cy.get('[data-testid="chart-figure"] svg').should(
       'contain.text',
@@ -338,7 +337,7 @@ describe('X-Axis Mode Toggle (inference chart)', () => {
   it('follows the percentile selector in the Interactivity axis label', () => {
     // Select p75 here rather than inheriting it from another test — the axis
     // label must track the selector on its own.
-    selectAdvancedXAxisMode('interactivity', 'Interactivity');
+    selectXAxisMode('interactivity', 'Interactivity');
     cy.get('[data-testid="percentile-selector"]').click();
     cy.contains('[role="option"]', 'p75').click();
     cy.get('[data-testid="chart-figure"] svg').should(
@@ -352,24 +351,28 @@ describe('X-axis mode URL param', () => {
   // Regression: the reconcile effect used to run before availability resolved
   // the sequence. It recorded the fixed-seq placeholder kind, then treated the
   // switch to agentic as a user-driven kind change and clobbered the
-  // URL-restored mode with the agentic default (E2E Normalized Interactivity).
+  // URL-restored mode with the agentic default.
+  //
+  // The restored mode must differ from that default or the test cannot fail:
+  // #736 moved the default to Interactivity, so restoring Interactivity now
+  // proves nothing. TTFT is a mode the snap would visibly overwrite.
   it('keeps a URL-restored mode through the agentic sequence resolving', () => {
     interceptAgenticData();
     interceptDerivedAgenticMetrics();
-    cy.visit('/inference?i_seq=agentic-traces&i_xmode=interactivity', {
+    cy.visit('/inference?i_seq=agentic-traces&i_xmode=ttft', {
       onBeforeLoad(win) {
         win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
         unlockAgenticGate(win);
       },
     });
-    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic Traces');
-    cy.get('[data-testid="x-axis-mode-advanced"]')
-      .should('have.attr', 'data-state', 'active')
-      .and('contain.text', 'Interactivity');
+    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic');
+    cy.get('[data-testid="x-axis-mode-ttft"]')
+      .should('have.attr', 'aria-selected', 'true')
+      .and('contain.text', 'TTFT');
     // Assert on the rendered chart too: the clobber happened one tick after
     // the buttons first painted, so a button-only check could pass too early.
-    cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Interactivity');
-    cy.get('[data-testid="x-axis-mode-e2e-normalized-interactivity"]').should(
+    cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Time To First Token');
+    cy.get('[data-testid="x-axis-mode-interactivity"]').should(
       'have.attr',
       'aria-selected',
       'false',
@@ -388,28 +391,61 @@ describe('X-axis mode URL param', () => {
       },
     });
 
-    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic Traces');
+    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic');
     cy.get('[data-testid="percentile-selector"]').should('not.exist');
     cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'P90');
   });
 });
 
 describe('Default scenario', () => {
-  it('bare /inference defaults to 8K / 1K even when the model has agentic data', () => {
-    // Availability contains BOTH agentic and fixed-seq rows for the default
-    // model; the default selection must still resolve to 8K/1K, not agentic.
-    interceptFixedSequenceData();
+  it('bare /inference opens on the Agentic scenario when the model has corresponding data', () => {
+    // Availability contains BOTH agentic and fixed-seq rows for DeepSeek-V4-Pro,
+    // so the untouched 8K/1K selection must not win.
+    interceptAgenticData();
+    interceptDerivedAgenticMetrics();
     cy.visit('/inference', {
       onBeforeLoad(win) {
         win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
       },
     });
+    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic');
+    // The explainer sits beside the trigger, linking out to the dataset page.
+    cy.get('[data-testid="scenario-agentic-info"]').should('exist');
+    cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
+    cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'P90');
+  });
+
+  it('keeps 8K / 1K when the link asks for it explicitly', () => {
+    // An explicit selection always beats the availability-driven default.
+    interceptFixedSequenceData();
+    cy.visit('/inference?i_seq=8k%2F1k', {
+      onBeforeLoad(win) {
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+      },
+    });
     cy.get('[data-testid="scenario-selector"]').should('contain.text', '8K / 1K');
+    cy.get('[data-testid="scenario-agentic-info"]').should('not.exist');
     cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
     // Fixed-seq plots the mean field — no percentile prefix on the axis label.
     cy.get('[data-testid="chart-figure"] svg').should('contain.text', 'Interactivity (tok/s/user)');
     cy.get('[data-testid="chart-figure"] svg').should('not.contain.text', 'P90 Interactivity');
   });
+
+  it('opens the Agentic scenario for another model with corresponding data', () => {
+    // DeepSeek-R1 is intentionally outside the original hard-coded model list;
+    // its agentic availability must still make the Agentic scenario the default.
+    cy.intercept('GET', '/api/v1/availability', { body: otherModelAvailability }).as(
+      'availability',
+    );
+    cy.intercept('GET', '/api/v1/benchmarks*', { body: otherModelBenchmarks }).as('benchmarks');
+    cy.visit('/inference?g_model=DeepSeek-R1-0528', {
+      onBeforeLoad(win) {
+        win.localStorage.setItem('inferencex-star-modal-dismissed', String(Date.now()));
+      },
+    });
+    cy.get('[data-testid="scenario-selector"]').should('contain.text', 'Agentic');
+    cy.get('[data-testid="scenario-agentic-info"]').should('exist');
+  });
 });
 
 describe('Label defaults for fixed-sequence scenarios', () => {
@@ -421,6 +457,7 @@ describe('Label defaults for fixed-sequence scenarios', () => {
       },
     });
     cy.get('[data-testid="scenario-selector"]').should('contain.text', '8K / 1K');
+    expandLegendAdvanced();
     cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'unchecked');
     cy.get('#scatter-point-labels').should('have.attr', 'data-state', 'unchecked');
     cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'checked');
@@ -454,8 +491,9 @@ describe('Label defaults for fixed-sequence scenarios', () => {
     cy.get('[data-testid="x-axis-mode-ttft"]').should('have.attr', 'aria-selected', 'true');
   });
 
-  // Only agentic nests the latency modes: E2E Normalized Interactivity is
-  // agentic-only, so collapsing them here would leave an empty tab strip.
+  // The strip is flat everywhere since #736; what is still specific to a fixed
+  // sequence is that E2E Normalized Interactivity is agentic-only, so this strip
+  // carries three tabs rather than four.
   it('keeps the flat x-axis strip with no Advanced menu', () => {
     interceptFixedSequenceData();
     cy.visit('/inference?i_seq=8k%2F1k', {
@@ -481,6 +519,7 @@ describe('Label defaults for fixed-sequence scenarios', () => {
       },
     });
     cy.get('[data-testid="scenario-selector"]').should('contain.text', '8K / 1K');
+    expandLegendAdvanced();
     cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'checked');
     cy.get('#scatter-point-labels').should('have.attr', 'data-state', 'checked');
     cy.get('#scatter-line-labels').should('have.attr', 'data-state', 'unchecked');
@@ -568,9 +607,19 @@ describe('X-Axis Mode Toggle — overlay path (finding #8 regression guard)', ()
     cy.get('[data-testid="chart-figure"]').should('have.length.at.least', 1);
   });
 
+  // Cypress clears intercepts between tests, so the derived-metrics stub is
+  // re-registered per test rather than once in `before`. Until #736 the
+  // agentic default was E2E Normalized Interactivity, so the fetch happened
+  // during `before` while the stub was still alive and React Query held the
+  // result for the rest of the spec. The default no longer fetches, so any
+  // later switch into that mode issues a fresh request.
+  beforeEach(() => {
+    interceptDerivedAgenticMetrics();
+  });
+
   it('shows overlay (unofficial-run) watermark SVG when an overlay is loaded', () => {
     // Explicitly select Interactivity — do not rely on the agentic default mode.
-    selectAdvancedXAxisMode('interactivity', 'Interactivity');
+    selectXAxisMode('interactivity', 'Interactivity');
     // The unofficial-run pattern watermark appears when isUnofficialRun is true.
     cy.get('[data-testid="inference-chart-display"] svg pattern[id^="unofficial-pattern-"]').should(
       'exist',
@@ -580,13 +629,13 @@ describe('X-Axis Mode Toggle — overlay path (finding #8 regression guard)', ()
       'contain.text',
       'P90 Interactivity (tok/s/user)',
     );
-    cy.get('#chart-0 [data-testid="offload-halo-key"]')
+    cy.get('[data-testid="axis-metric-footer-chart-0"] [data-testid="offload-halo-key"]')
       .should('be.visible')
       .and('contain.text', 'KV offload ON');
   });
 
   it('switches to ttft x-axis mode and renders SVG with overlay points', () => {
-    selectAdvancedXAxisMode('ttft', 'TTFT');
+    selectXAxisMode('ttft', 'TTFT');
     cy.get('[data-testid="chart-figure"] h2').should('contain.text', 'Time To First Token');
     // Overlay points render as triangles or circles inside the chart SVG.
     cy.get('[data-testid="inference-chart-display"] svg').should('exist');
diff --git a/packages/app/cypress/e2e/unofficial-watermark.cy.ts b/packages/app/cypress/e2e/unofficial-watermark.cy.ts
index 043dc516f..944d25024 100644
--- a/packages/app/cypress/e2e/unofficial-watermark.cy.ts
+++ b/packages/app/cypress/e2e/unofficial-watermark.cy.ts
@@ -1,3 +1,5 @@
+import { expandLegendAdvanced } from '../support/legend-advanced';
+
 describe('Unofficial-run watermark', () => {
   before(() => {
     cy.fixture('api/benchmarks.json').then((benchmarks) => {
@@ -20,6 +22,10 @@ describe('Unofficial-run watermark', () => {
           benchmarks: benchmarks.map((row: Record<string, unknown>) => ({
             ...row,
             is_multinode: true,
+            prefill_tp: 8,
+            prefill_ep: 1,
+            decode_tp: 8,
+            decode_ep: 1,
             metrics: {
               ...(row.metrics as Record<string, unknown> | undefined),
               kv_offloading: 'dram',
@@ -27,6 +33,8 @@ describe('Unofficial-run watermark', () => {
               kv_p2p_transfer: 'nixl',
               router_name: 'sglang-router',
               router_version: '0.3.2',
+              dcp_size: 8,
+              pcp_size: 4,
               server_gpu_cache_hit_rate: 0.875,
             },
             run_url: runUrl,
@@ -79,9 +87,18 @@ describe('Unofficial-run watermark', () => {
         expect(tooltip).to.contain.text('KV Offload Engine: HiCache');
         expect(tooltip).to.contain.text('KV Transfer Engine: NIXL');
         expect(tooltip).to.contain.text('Router: SGLang Router 0.3.2');
+        expect(tooltip).to.contain.text('Decode Context Parallelism (DCP): 8');
+        expect(tooltip).to.contain.text('Prefill Context Parallelism (PCP): 4');
         expect(tooltip).to.contain.text('Chip Cache Hit Rate: 87.5%');
       });
 
+    expandLegendAdvanced();
+    cy.get('#scatter-parallelism-labels').click();
+    cy.get('[data-testid="scatter-graph"] .unofficial-overlay-pt')
+      .first()
+      .find('.overlay-label')
+      .should('contain.text', 'TP8/DCP8/PCP4');
+
     cy.get('[data-testid="scatter-graph"]').first().scrollIntoView();
     cy.screenshot('unofficial-watermark', { capture: 'viewport' });
 
diff --git a/packages/app/cypress/e2e/url-params.cy.ts b/packages/app/cypress/e2e/url-params.cy.ts
index 6c8272181..6f160bcec 100644
--- a/packages/app/cypress/e2e/url-params.cy.ts
+++ b/packages/app/cypress/e2e/url-params.cy.ts
@@ -3,6 +3,7 @@
  * update the visible output (selector text, SVG axis labels).
  * Merged from url-params.cy.ts + chart-filter-effects.cy.ts + high-contrast.cy.ts.
  */
+import { expandLegendAdvanced } from '../support/legend-advanced';
 const visitWithDismissedModal = (path: string) => {
   cy.visit(path, {
     onBeforeLoad(win) {
@@ -38,10 +39,134 @@ describe('URL Parameter Persistence', () => {
   });
 
   describe('Inference legend', () => {
-    it('i_legend=0 collapses the sidebar legend on load', () => {
+    it('i_legend=0 hides the sidebar legend on load and the reopen button restores it', () => {
       visitWithDismissedModal('/inference?i_legend=0');
+      cy.get('[data-testid="legend-open-button"]').first().should('be.visible');
+      cy.get('.sidebar-legend').should('not.exist');
+
+      cy.get('[data-testid="legend-open-button"]').first().click();
+      cy.get('.sidebar-legend').first().should('be.visible');
+      cy.get('[data-testid="legend-open-button"]').should('not.exist');
+    });
+
+    // The address bar is deliberately left clean after load (see url-state.ts)
+    // — filter changes like closing the legend only reach the in-memory share
+    // state, so this asserts the UI transition rather than location.search.
+    it('legend close button hides the panel and the reopen button restores it', () => {
+      visitWithDismissedModal('/inference');
+      cy.get('.sidebar-legend').first().should('be.visible');
+
+      cy.get('[data-testid="legend-close-button"]').first().click();
+      cy.get('.sidebar-legend').should('not.exist');
+      cy.get('[data-testid="legend-open-button"]').first().should('be.visible');
+
+      cy.get('[data-testid="legend-open-button"]').first().click();
       cy.get('.sidebar-legend').first().should('be.visible');
-      cy.get('.sidebar-legend').first().should('not.have.class', 'bg-accent');
+      cy.get('[data-testid="legend-open-button"]').should('not.exist');
+    });
+
+    it('preserves a legend subset when chart metrics change', () => {
+      visitWithDismissedModal('/inference');
+
+      cy.get('[data-testid="chart-legend"] input[type="checkbox"]:checked').should(
+        'have.length.greaterThan',
+        1,
+      );
+      cy.get('[data-testid="chart-legend"] [role="button"][aria-label^="Hide "]')
+        .first()
+        .closest('li')
+        .find('input[type="checkbox"]')
+        .invoke('attr', 'id')
+        .then((hiddenInputId) => {
+          expect(hiddenInputId).to.be.a('string');
+          expect(hiddenInputId).to.have.length.greaterThan(0);
+          const selector = `#${CSS.escape(hiddenInputId!)}`;
+
+          cy.get(selector).parent().find('[role="button"][aria-label^="Hide "]').click();
+          cy.get(selector).should('not.be.checked');
+
+          cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
+          cy.contains('[role="option"]', 'All-in Provisioned Joules per Total Token').click({
+            force: true,
+          });
+          cy.get(selector).should('not.be.checked');
+
+          cy.get('[data-testid="x-axis-mode-ttft"]').click();
+          cy.get('[data-testid="x-axis-mode-ttft"]').should('have.attr', 'aria-selected', 'true');
+          cy.get(selector).should('not.be.checked');
+        });
+    });
+
+    it('refreshes the automatic Best per SKU selection when the metric changes', () => {
+      visitWithDismissedModal('/inference');
+
+      // Best per SKU lives in the Quick Filters dialog now.
+      cy.get('[data-testid="scatter-quick-filters"]').click();
+      cy.get('[data-testid="quick-filter-best-per-sku"]').should(
+        'have.attr',
+        'data-state',
+        'checked',
+      );
+      cy.contains('button', 'Done').click();
+      cy.get('[data-testid="quick-filters-dialog"]').should('not.exist');
+
+      cy.get('[data-testid="chart-legend"] ul input[type="checkbox"]:checked')
+        .then(($inputs) => [...$inputs].map((input) => input.id).toSorted())
+        .then((before) => {
+          cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
+          cy.contains(
+            '[role="option"]',
+            'Cost per Million Total Tokens (Owning - Hyperscaler)',
+          ).click({ force: true });
+
+          cy.get('[data-testid="scatter-quick-filters"]').click();
+          cy.get('[data-testid="quick-filter-best-per-sku"]').should(
+            'have.attr',
+            'data-state',
+            'checked',
+          );
+          cy.contains('button', 'Done').click();
+          cy.get('[data-testid="quick-filters-dialog"]').should('not.exist');
+          cy.get('[data-testid="x-axis-mode-ttft"]').click();
+          cy.get('[data-testid="x-axis-mode-ttft"]').should('have.attr', 'aria-selected', 'true');
+          cy.get('[data-testid="chart-legend"] ul input[type="checkbox"]:checked').then(
+            ($inputs) => {
+              const after = [...$inputs].map((input) => input.id).toSorted();
+              expect(after, 'metric-specific Best per SKU winners').not.to.deep.equal(before);
+            },
+          );
+        });
+    });
+  });
+
+  describe('Provider remount state', () => {
+    it('preserves filters across a standalone dashboard route', () => {
+      visitWithDismissedModal('/inference');
+      cy.get('[data-testid="model-selector"]').click();
+      cy.contains('[role="option"]', 'Qwen3.5 397B').click();
+      cy.get('[data-testid="model-selector"]').should('contain.text', 'Qwen3.5 397B');
+      // Navigate immediately to cover pending writes inside the debounce window.
+      cy.get('[data-testid="tab-trigger-gpu-specs"]').click();
+      cy.url().should('include', '/gpu-specs');
+      cy.get('[data-testid="tab-trigger-inference"]').click();
+
+      cy.get('[data-testid="model-selector"]').should('contain.text', 'Qwen3.5 397B');
+    });
+
+    it('preserves the newest pending model across a retained-provider tab switch', () => {
+      visitWithDismissedModal('/inference');
+      cy.get('[data-testid="model-selector"]').click();
+      cy.contains('[role="option"]', 'MiniMax M3 428B').click();
+      cy.get('[data-testid="model-selector"]').should('contain.text', 'MiniMax M3 428B');
+      cy.wait(200);
+
+      cy.get('[data-testid="model-selector"]').click();
+      cy.contains('[role="option"]', 'Qwen3.5 397B').click();
+      cy.get('[data-testid="tab-trigger-evaluation"]').click();
+      cy.location('pathname').should('eq', '/evaluation');
+      cy.get('[data-testid="tab-trigger-inference"]').click();
+      cy.location('pathname').should('eq', '/inference');
+      cy.get('[data-testid="model-selector"]').should('contain.text', 'Qwen3.5 397B');
     });
   });
 
@@ -63,10 +188,12 @@ describe('URL Parameter Persistence', () => {
     it('changing Y-axis metric via dropdown updates SVG axis label', () => {
       visitWithDismissedModal('/inference');
 
+      // The dashboard opens on the tokens-per-dollar default; this asserts the
+      // starting label before switching, not that throughput is the default.
       cy.get('[data-testid="scatter-graph"]')
         .first()
         .find('svg text[transform="rotate(-90)"]')
-        .should('contain.text', 'Throughput');
+        .should('contain.text', 'Total Tokens per $1 USD');
 
       cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
       cy.contains('[role="option"]', 'Cost per Million Total Tokens (Owning - Hyperscaler)').click({
@@ -79,6 +206,28 @@ describe('URL Parameter Persistence', () => {
         .should('have.text', 'Cost per Million Total Tokens ($)');
     });
 
+    it('tokens-per-dollar URL metric is independent from cost per million', () => {
+      visitWithDismissedModal('/inference?i_metric=y_tokensPerDollarH');
+
+      cy.get('[data-testid="yaxis-metric-selector"]').should(
+        'contain.text',
+        'Total Tokens per $1 USD (Owning - Hyperscaler)',
+      );
+      cy.get('[data-testid="scatter-graph"]')
+        .first()
+        .find('svg text[transform="rotate(-90)"]')
+        .should('have.text', 'Total Tokens per $1 USD (tok/$)');
+    });
+
+    it('keeps the legacy i_metric=y alias on raw throughput', () => {
+      visitWithDismissedModal('/inference?i_metric=y');
+
+      cy.get('[data-testid="yaxis-metric-selector"]').should(
+        'contain.text',
+        'Token Throughput per Chip',
+      );
+    });
+
     it('selecting a Y-axis metric updates the displayed value', () => {
       visitWithDismissedModal('/inference');
       cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
@@ -121,6 +270,62 @@ describe('URL Parameter Persistence', () => {
         .find('svg text[transform="rotate(-90)"]')
         .should('contain.text', 'Token Throughput per All in Utility MW');
     });
+
+    it('keeps tooltip rulers aligned after a zoomed metric switch', () => {
+      visitWithDismissedModal('/inference?i_metric=y_tpPerGpu');
+      expandLegendAdvanced();
+      cy.get('#scatter-log-scale').first().click();
+
+      cy.get('[data-testid="scatter-graph"] [data-testid="d3-chart-svg"]')
+        .first()
+        .then(($svg) => {
+          const svg = $svg[0] as unknown as SVGSVGElement & { __zoom?: { k: number } };
+          const bounds = svg.getBoundingClientRect();
+          for (let i = 0; i < 2; i += 1) {
+            svg.dispatchEvent(
+              new WheelEvent('wheel', {
+                deltaY: -240,
+                clientX: bounds.x + bounds.width / 2,
+                clientY: bounds.y + bounds.height / 2,
+                shiftKey: true,
+                bubbles: true,
+                cancelable: true,
+              }),
+            );
+          }
+          expect(svg.__zoom?.k, 'active zoom scale').to.be.greaterThan(1);
+        });
+
+      cy.get('[data-testid="yaxis-metric-selector"]').click({ force: true });
+      cy.contains('[role="option"]', 'Input Token Throughput per Chip').click({ force: true });
+      cy.get('[data-testid="yaxis-metric-selector"]').should(
+        'contain.text',
+        'Input Token Throughput per Chip',
+      );
+
+      cy.get('[data-testid="scatter-graph"]')
+        .first()
+        .within(() => {
+          cy.get<SVGGElement>('.dot-group')
+            .first()
+            .then(($point) => {
+              const match = $point
+                .attr('transform')
+                ?.match(/translate\((?<x>[^,]+),(?<y>[^)]+)\)/u);
+              expect(match, 'point transform').not.to.equal(null);
+              const pointX = Number(match?.groups?.x);
+              const pointY = Number(match?.groups?.y);
+
+              cy.wrap($point).trigger('mouseenter', { force: true });
+              cy.get('.vertical-ruler')
+                .invoke('attr', 'x1')
+                .then((x) => expect(Number(x)).to.be.closeTo(pointX, 1));
+              cy.get('.horizontal-ruler')
+                .invoke('attr', 'y1')
+                .then((y) => expect(Number(y)).to.be.closeTo(pointY, 1));
+            });
+        });
+    });
   });
 
   describe('Reliability date range', () => {
@@ -196,7 +401,9 @@ describe('URL Parameter Persistence', () => {
     });
 
     it('/inference with invalid ?i_prec=junk falls back to the default', () => {
-      visitWithErrorSpy('/inference?i_prec=junk');
+      // Pair with a multi-precision model — the FP4-only default model hides
+      // the precision selector entirely, leaving nothing to assert against.
+      visitWithErrorSpy('/inference?g_model=DeepSeek-R1-0528&i_prec=junk');
       cy.get('[data-testid="precision-multiselect"]').invoke('text').should('not.contain', 'junk');
       assertNoHydrationMismatch();
     });
@@ -239,24 +446,28 @@ describe('URL Parameter Persistence', () => {
     it('inference loads with high contrast off by default', () => {
       visitWithDismissedModal('/inference');
       cy.get('[data-testid="scatter-graph"]').should('exist');
+      expandLegendAdvanced();
       cy.get('#scatter-high-contrast').first().should('have.attr', 'data-state', 'unchecked');
     });
 
     it('i_hc=0 disables high contrast on load', () => {
       visitWithDismissedModal('/inference?i_hc=0');
       cy.get('[data-testid="scatter-graph"]').should('exist');
+      expandLegendAdvanced();
       cy.get('#scatter-high-contrast').first().should('have.attr', 'data-state', 'unchecked');
     });
 
     it('i_hc=1 applies high contrast on load', () => {
       visitWithDismissedModal('/inference?i_hc=1');
       cy.get('[data-testid="scatter-graph"]').should('exist');
+      expandLegendAdvanced();
       cy.get('#scatter-high-contrast').first().should('have.attr', 'data-state', 'checked');
     });
 
     it('multiple high contrast params can coexist in URL', () => {
       visitWithDismissedModal('/inference?i_hc=1&r_hc=1&e_hc=1');
       cy.get('[data-testid="scatter-graph"]').should('exist');
+      expandLegendAdvanced();
       cy.get('#scatter-high-contrast').first().should('have.attr', 'data-state', 'checked');
     });
 
@@ -274,8 +485,8 @@ describe('URL Parameter Persistence', () => {
     });
 
     it('historical trends tab shares the inference high-contrast default (off)', () => {
-      // Historical reads highContrast from the same InferenceContext as the
-      // scatter chart, so it inherits the default-off behavior.
+      // Historical reads highContrast from the shared inference display domain,
+      // so it inherits the default-off behavior.
       visitWithDismissedModal('/historical');
       cy.get('[data-testid="historical-trends-display"]').should('exist');
       cy.get('#historical-high-contrast').first().should('have.attr', 'data-state', 'unchecked');
@@ -292,6 +503,7 @@ describe('URL Parameter Persistence', () => {
     it('a bare /inference link with neither param renders high contrast AND parallelism labels off', () => {
       visitWithDismissedModal('/inference');
       cy.get('[data-testid="scatter-graph"]').should('exist');
+      expandLegendAdvanced();
       cy.get('#scatter-high-contrast').first().should('have.attr', 'data-state', 'unchecked');
       cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'unchecked');
     });
@@ -299,6 +511,7 @@ describe('URL Parameter Persistence', () => {
     it('i_hc=1&i_advlabel=1 enables both high contrast and parallelism labels on load', () => {
       visitWithDismissedModal('/inference?i_hc=1&i_advlabel=1');
       cy.get('[data-testid="scatter-graph"]').should('exist');
+      expandLegendAdvanced();
       cy.get('#scatter-high-contrast').first().should('have.attr', 'data-state', 'checked');
       cy.get('#scatter-parallelism-labels').should('have.attr', 'data-state', 'checked');
     });
diff --git a/packages/app/cypress/e2e/yaxis-metrics-render.cy.ts b/packages/app/cypress/e2e/yaxis-metrics-render.cy.ts
index c2a9a7bec..98a66f46a 100644
--- a/packages/app/cypress/e2e/yaxis-metrics-render.cy.ts
+++ b/packages/app/cypress/e2e/yaxis-metrics-render.cy.ts
@@ -21,7 +21,26 @@ describe('Y-Axis Metrics All Render Data', () => {
     'Cost per Million Input Tokens (Owning - Hyperscaler)',
     'Cost per Million Input Tokens (Owning - Neocloud Giant)',
     'Cost per Million Input Tokens (3 Year Rental)',
+    'Total Tokens per $1 USD (Owning - Hyperscaler)',
+    'Total Tokens per $1 USD (Owning - Neocloud Giant)',
+    'Total Tokens per $1 USD (3 Year Rental)',
+    'Output Tokens per $1 USD (Owning - Hyperscaler)',
+    'Output Tokens per $1 USD (Owning - Neocloud Giant)',
+    'Output Tokens per $1 USD (3 Year Rental)',
+    'Input Tokens per $1 USD (Owning - Hyperscaler)',
+    'Input Tokens per $1 USD (Owning - Neocloud Giant)',
+    'Input Tokens per $1 USD (3 Year Rental)',
+    'Total Tokens per ¥1 RMB (Owning - Hyperscaler)',
+    'Total Tokens per ¥1 RMB (Owning - Neocloud Giant)',
+    'Total Tokens per ¥1 RMB (3 Year Rental)',
+    'Output Tokens per ¥1 RMB (Owning - Hyperscaler)',
+    'Output Tokens per ¥1 RMB (Owning - Neocloud Giant)',
+    'Output Tokens per ¥1 RMB (3 Year Rental)',
+    'Input Tokens per ¥1 RMB (Owning - Hyperscaler)',
+    'Input Tokens per ¥1 RMB (Owning - Neocloud Giant)',
+    'Input Tokens per ¥1 RMB (3 Year Rental)',
     'Cost per Million Total Tokens (Custom User Values)',
+    'Total Tokens per $1 USD (Custom User Values)',
     'Token Throughput per All in Utility MW (Custom User Values)',
     'All-in Provisioned Joules per Total Token',
     'All-in Provisioned Joules per Output Token',
diff --git a/packages/app/cypress/e2e/zh-pages.cy.ts b/packages/app/cypress/e2e/zh-pages.cy.ts
index 2ae1441db..7e114d37c 100644
--- a/packages/app/cypress/e2e/zh-pages.cy.ts
+++ b/packages/app/cypress/e2e/zh-pages.cy.ts
@@ -5,17 +5,13 @@ describe('Chinese (/zh) pages', () => {
     });
 
     it('renders the Chinese landing content', () => {
-      cy.contains('h2', '探索 InferenceX').should('exist');
-      cy.contains('快速对比').should('exist');
-    });
-
-    it('links to the Chinese overview and full dashboard', () => {
-      cy.get('[data-testid="landing-overview-link"]')
-        .should('have.attr', 'href', '/zh/overview')
-        .and('have.text', '总览');
-      cy.get('[data-testid="landing-full-dashboard-link"]')
-        .should('have.attr', 'href', '/zh/inference')
-        .and('have.text', '完整仪表板');
+      cy.get('[data-testid="intro-section"]').should('contain.text', '智能体推理基准测试');
+      cy.get('[data-testid="splash-text"]').should('have.text', 'AgentX 来了！！');
+      // Quick Comparisons is hidden behind SHOW_QUICK_COMPARISONS in
+      // landing-page.tsx; the card and its Chinese strings still exist in the
+      // source, so assert it is not rendered rather than dropping the check.
+      cy.get('[data-testid="landing-quick-comparisons"]').should('not.exist');
+      cy.contains('快速对比').should('not.exist');
     });
 
     it('sets hreflang alternates to the English homepage', () => {
@@ -27,14 +23,40 @@ describe('Chinese (/zh) pages', () => {
       cy.get('[data-testid="language-toggle"]').should('have.attr', 'href', '/');
     });
 
+    it('header links to the Chinese AgentX page', () => {
+      cy.get('[data-testid="nav-link-agentx"]')
+        .should('contain.text', 'AgentX')
+        .and('have.attr', 'href', '/zh/agentx');
+    });
+
+    it('renders the AgentX hero on the Chinese landing page', () => {
+      cy.get('[data-testid="compare-agentx-primary"]').within(() => {
+        cy.get('h2').should('have.text', '真实智能体工作负载下的推理性能对比');
+        cy.get('[data-testid="compare-agentx-overview-link"]')
+          .should('contain.text', '总览')
+          .and('have.attr', 'href', '/zh/overview');
+        cy.get('[data-testid="compare-agentx-methodology-link"]').should('not.exist');
+        // Ledger NEW pills localize to 新 on the Chinese landing page.
+        cy.get('[data-testid^="compare-agentx-model-"] [data-new-badge="agentx-ledger"]')
+          .should('have.length', 6)
+          .each(($badge) => expect($badge.text()).to.equal('新'));
+      });
+    });
+
     it('footer renders in Chinese with zh-internal links', () => {
-      cy.get('[data-testid="footer-brand-description"]').should('contain.text', '开源推理基准测试');
+      cy.get('[data-testid="footer-brand-description"]').should(
+        'contain.text',
+        'InferenceX 持续开展开源的 agentic 推理基准测试',
+      );
       cy.get('[data-testid="footer-link-supporters"]')
-        .should('contain.text', '支持者')
+        .should('contain.text', '业界评价')
         .and('have.attr', 'href', '/zh/quotes');
-      cy.get('[data-testid="footer-link-datasets"]')
-        .should('contain.text', '数据集')
-        .and('have.attr', 'href', '/zh/datasets');
+      cy.get('[data-testid="footer-link-agentx"]')
+        .should('contain.text', 'AgentX')
+        .and('have.attr', 'href', '/zh/agentx');
+      cy.get('[data-testid="footer-link-about"]')
+        .should('have.text', '关于 SemiAnalysis')
+        .and('have.attr', 'href', 'https://semianalysis.com/about/');
       cy.get('[data-testid="footer-link-articles"]')
         .should('contain.text', '文章')
         .and('have.attr', 'href', '/zh/blog');
@@ -43,9 +65,7 @@ describe('Chinese (/zh) pages', () => {
         'href',
         '/zh/land-acknowledgement',
       );
-      cy.get('[data-testid="footer-link-zh"]')
-        .should('contain.text', 'English')
-        .and('have.attr', 'href', '/');
+      cy.get('[data-testid="footer-link-zh"]').should('not.exist');
     });
   });
 
@@ -56,7 +76,8 @@ describe('Chinese (/zh) pages', () => {
 
     it('renders the Chinese SEO intro above the chart', () => {
       cy.get('[data-testid="zh-tab-intro"]').within(() => {
-        cy.contains('h1', 'AI 推理基准测试').should('exist');
+        cy.contains('h1', '智能体推理基准测试').should('exist');
+        cy.contains('长上下文、多轮').should('exist');
       });
     });
 
@@ -68,6 +89,28 @@ describe('Chinese (/zh) pages', () => {
     });
   });
 
+  describe('zh FAQ', () => {
+    it('uses the same direct-link anchor as the English FAQ', () => {
+      cy.visit('/zh/about#faq-normalized-interactivity');
+
+      cy.get('#faq-normalized-interactivity')
+        .should('be.visible')
+        .within(() => {
+          cy.contains(
+            'a[href="#faq-normalized-interactivity"]',
+            '端到端归一化交互性与交互性有何区别？',
+          ).should('be.visible');
+          cy.contains('TTFT 越长，归一化指标越低').should('be.visible');
+          cy.get('[data-testid="faq-copy-link-faq-normalized-interactivity"]')
+            .should('be.visible')
+            .and('have.attr', 'title', '复制链接')
+            .find('svg.lucide-link')
+            .should('be.visible');
+        });
+      cy.location('hash').should('eq', '#faq-normalized-interactivity');
+    });
+  });
+
   describe('zh blog', () => {
     before(() => {
       cy.visit('/zh/blog');
diff --git a/packages/app/cypress/fixtures/api/_manifest.json b/packages/app/cypress/fixtures/api/_manifest.json
new file mode 100644
index 000000000..089c83483
--- /dev/null
+++ b/packages/app/cypress/fixtures/api/_manifest.json
@@ -0,0 +1,97 @@
+{
+  "schemaVersion": 1,
+  "generatedAt": "2026-08-20T08:25:30.857Z",
+  "fixtures": {
+    "agentic-catalog": {
+      "bytes": 10027,
+      "capturedAt": null,
+      "sha256": "b4d7da73dbb839a8c9444af4bf90cdcd2e25c53970214937ecc122c86c13ae58",
+      "source": null,
+      "topLevel": "array"
+    },
+    "availability": {
+      "bytes": 147942,
+      "capturedAt": null,
+      "sha256": "ffcf2ed4d741a592c50a4720d288acff176202b737942edb048ac46f0a9409ba",
+      "source": null,
+      "topLevel": "array"
+    },
+    "benchmarks-history": {
+      "bytes": 1859995,
+      "capturedAt": null,
+      "sha256": "9e7efc17d018d25acd13095ac0c33ed3c20aaf0d86c225792c0a73bde8e12d15",
+      "source": null,
+      "topLevel": "array"
+    },
+    "benchmarks": {
+      "bytes": 775884,
+      "capturedAt": null,
+      "sha256": "f46823a76cb63248705329f327e1494df02b4dfa4df0898b7ee084c75c20bd52",
+      "source": null,
+      "topLevel": "array"
+    },
+    "collectivex-latest": {
+      "bytes": 63278,
+      "capturedAt": "2026-08-20T08:25:30.857Z",
+      "sha256": "66dd9287f7b5c1d33fa2dc76f69622b25dc42d89fcd62644b01c8edd7d6290b3",
+      "source": "synthetic://collectivex",
+      "topLevel": "object"
+    },
+    "collectivex-run-159": {
+      "bytes": 20374,
+      "capturedAt": "2026-08-20T08:25:30.857Z",
+      "sha256": "54e33df39b51158ed3fb0956dd70e10e2c3540845dd62ab7c64fabad175d2259",
+      "source": "synthetic://collectivex",
+      "topLevel": "object"
+    },
+    "collectivex-runs": {
+      "bytes": 687,
+      "capturedAt": "2026-08-20T08:25:30.857Z",
+      "sha256": "9f3d47af2543fc56b0afec7b0c577ba08fcfc2d6c91b07d98163e721e50146c8",
+      "source": "synthetic://collectivex",
+      "topLevel": "object"
+    },
+    "evaluations": {
+      "bytes": 474940,
+      "capturedAt": null,
+      "sha256": "edd2740ddbf1104c763114018fa081e96c39d5edfa3f2e2424a7ce4b724a06c9",
+      "source": null,
+      "topLevel": "array"
+    },
+    "overview-history-rows": {
+      "bytes": 5002,
+      "capturedAt": null,
+      "sha256": "73caa715550d07044dcad6f14dd6661fb28a203883470ad47f272df1eb3c4c82",
+      "source": null,
+      "topLevel": "object"
+    },
+    "overview-rows": {
+      "bytes": 22152,
+      "capturedAt": null,
+      "sha256": "ccdc4c23fc6b5ab6f664e19db2ec31efe2b34318976baec06e9219987e240922",
+      "source": null,
+      "topLevel": "object"
+    },
+    "reliability": {
+      "bytes": 530238,
+      "capturedAt": null,
+      "sha256": "cf36cdc2446c75bd191a22f027823bce6ec1380937b5387f274a773dbc19ee66",
+      "source": null,
+      "topLevel": "array"
+    },
+    "submissions": {
+      "bytes": 130122,
+      "capturedAt": null,
+      "sha256": "cd3d37a878b708eab504b0a6bca6d2af408101d88bee47503f0a656d6c60f1f9",
+      "source": null,
+      "topLevel": "object"
+    },
+    "workflow-info": {
+      "bytes": 4941,
+      "capturedAt": null,
+      "sha256": "f28d2220e5246ad70113e38e49394bc7a0bfe9afba8cedf08b90e39042e53357",
+      "source": null,
+      "topLevel": "object"
+    }
+  }
+}
diff --git a/packages/app/cypress/fixtures/api/agentic-catalog.json b/packages/app/cypress/fixtures/api/agentic-catalog.json
new file mode 100644
index 000000000..d0f1536f8
--- /dev/null
+++ b/packages/app/cypress/fixtures/api/agentic-catalog.json
@@ -0,0 +1,519 @@
+[
+  {
+    "id": 440056,
+    "model": "dsv4",
+    "hardware": "b200",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 41,
+    "minConc": 1,
+    "maxConc": 68,
+    "latestDate": "2026-08-19"
+  },
+  {
+    "id": 439967,
+    "model": "dsv4",
+    "hardware": "b200",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 38,
+    "minConc": 1,
+    "maxConc": 196,
+    "latestDate": "2026-08-17"
+  },
+  {
+    "id": 439778,
+    "model": "dsv4",
+    "hardware": "b300",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 41,
+    "minConc": 1,
+    "maxConc": 196,
+    "latestDate": "2026-08-13"
+  },
+  {
+    "id": 439526,
+    "model": "dsv4",
+    "hardware": "b300",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 43,
+    "minConc": 1,
+    "maxConc": 576,
+    "latestDate": "2026-08-11"
+  },
+  {
+    "id": 439970,
+    "model": "dsv4",
+    "hardware": "gb200",
+    "framework": "dynamo-vllm",
+    "precision": "fp4",
+    "points": 39,
+    "minConc": 1,
+    "maxConc": 576,
+    "latestDate": "2026-08-18"
+  },
+  {
+    "id": 439331,
+    "model": "dsv4",
+    "hardware": "gb300",
+    "framework": "dynamo-sglang",
+    "precision": "fp4",
+    "points": 17,
+    "minConc": 1,
+    "maxConc": 1536,
+    "latestDate": "2026-08-10"
+  },
+  {
+    "id": 439874,
+    "model": "dsv4",
+    "hardware": "gb300",
+    "framework": "dynamo-vllm",
+    "precision": "fp4",
+    "points": 6,
+    "minConc": 1,
+    "maxConc": 1152,
+    "latestDate": "2026-08-14"
+  },
+  {
+    "id": 439550,
+    "model": "dsv4",
+    "hardware": "h200",
+    "framework": "dynamo-sglang",
+    "precision": "fp8",
+    "points": 5,
+    "minConc": 1,
+    "maxConc": 16,
+    "latestDate": "2026-08-11"
+  },
+  {
+    "id": 439512,
+    "model": "dsv4",
+    "hardware": "mi355x",
+    "framework": "mori-sglang",
+    "precision": "fp4",
+    "points": 8,
+    "minConc": 1,
+    "maxConc": 128,
+    "latestDate": "2026-08-11"
+  },
+  {
+    "id": 439840,
+    "model": "dsv4",
+    "hardware": "mi355x",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 8,
+    "minConc": 1,
+    "maxConc": 48,
+    "latestDate": "2026-08-14"
+  },
+  {
+    "id": 440031,
+    "model": "dsv4",
+    "hardware": "mi355x",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 18,
+    "minConc": 1,
+    "maxConc": 64,
+    "latestDate": "2026-08-19"
+  },
+  {
+    "id": 439990,
+    "model": "glm5.2",
+    "hardware": "b200",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 10,
+    "minConc": 1,
+    "maxConc": 16,
+    "latestDate": "2026-08-18"
+  },
+  {
+    "id": 439987,
+    "model": "glm5.2",
+    "hardware": "b300",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 10,
+    "minConc": 1,
+    "maxConc": 16,
+    "latestDate": "2026-08-18"
+  },
+  {
+    "id": 440079,
+    "model": "glm5.2",
+    "hardware": "gb200",
+    "framework": "dynamo-sglang",
+    "precision": "fp4",
+    "points": 14,
+    "minConc": 1,
+    "maxConc": 128,
+    "latestDate": "2026-08-19"
+  },
+  {
+    "id": 439925,
+    "model": "glm5.2",
+    "hardware": "gb300",
+    "framework": "dynamo-sglang",
+    "precision": "fp4",
+    "points": 7,
+    "minConc": 2,
+    "maxConc": 192,
+    "latestDate": "2026-08-16"
+  },
+  {
+    "id": 439294,
+    "model": "glm5.2",
+    "hardware": "h200",
+    "framework": "dynamo-sglang",
+    "precision": "fp8",
+    "points": 3,
+    "minConc": 8,
+    "maxConc": 16,
+    "latestDate": "2026-08-08"
+  },
+  {
+    "id": 439298,
+    "model": "glm5.2",
+    "hardware": "mi325x",
+    "framework": "sglang",
+    "precision": "fp8",
+    "points": 7,
+    "minConc": 1,
+    "maxConc": 8,
+    "latestDate": "2026-08-08"
+  },
+  {
+    "id": 440067,
+    "model": "glm5.2",
+    "hardware": "mi355x",
+    "framework": "atom",
+    "precision": "fp4",
+    "points": 7,
+    "minConc": 1,
+    "maxConc": 10,
+    "latestDate": "2026-08-19"
+  },
+  {
+    "id": 439530,
+    "model": "glm5.2",
+    "hardware": "mi355x",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 10,
+    "minConc": 1,
+    "maxConc": 16,
+    "latestDate": "2026-08-11"
+  },
+  {
+    "id": 440070,
+    "model": "kimik3",
+    "hardware": "b200",
+    "framework": "dynamo-vllm",
+    "precision": "fp4",
+    "points": 15,
+    "minConc": 1,
+    "maxConc": 96,
+    "latestDate": "2026-08-19"
+  },
+  {
+    "id": 439937,
+    "model": "kimik3",
+    "hardware": "b300",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 20,
+    "minConc": 1,
+    "maxConc": 70,
+    "latestDate": "2026-08-16"
+  },
+  {
+    "id": 440103,
+    "model": "kimik3",
+    "hardware": "gb200",
+    "framework": "dynamo-vllm",
+    "precision": "fp4",
+    "points": 31,
+    "minConc": 1,
+    "maxConc": 384,
+    "latestDate": "2026-08-18"
+  },
+  {
+    "id": 440000,
+    "model": "kimik3",
+    "hardware": "gb300",
+    "framework": "dynamo-vllm",
+    "precision": "fp4",
+    "points": 7,
+    "minConc": 1,
+    "maxConc": 74,
+    "latestDate": "2026-08-18"
+  },
+  {
+    "id": 438883,
+    "model": "kimik3",
+    "hardware": "h200",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 35,
+    "minConc": 1,
+    "maxConc": 32,
+    "latestDate": "2026-08-07"
+  },
+  {
+    "id": 439660,
+    "model": "kimik3",
+    "hardware": "mi355x",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 4,
+    "minConc": 1,
+    "maxConc": 10,
+    "latestDate": "2026-08-12"
+  },
+  {
+    "id": 440146,
+    "model": "minimaxm3",
+    "hardware": "b200",
+    "framework": "trt",
+    "precision": "fp4",
+    "points": 22,
+    "minConc": 1,
+    "maxConc": 45,
+    "latestDate": "2026-08-20"
+  },
+  {
+    "id": 439916,
+    "model": "minimaxm3",
+    "hardware": "b200",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 26,
+    "minConc": 1,
+    "maxConc": 40,
+    "latestDate": "2026-08-16"
+  },
+  {
+    "id": 440040,
+    "model": "minimaxm3",
+    "hardware": "b300",
+    "framework": "trt",
+    "precision": "fp4",
+    "points": 11,
+    "minConc": 1,
+    "maxConc": 30,
+    "latestDate": "2026-08-19"
+  },
+  {
+    "id": 439907,
+    "model": "minimaxm3",
+    "hardware": "b300",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 24,
+    "minConc": 1,
+    "maxConc": 75,
+    "latestDate": "2026-08-16"
+  },
+  {
+    "id": 439888,
+    "model": "minimaxm3",
+    "hardware": "gb200",
+    "framework": "dynamo-vllm",
+    "precision": "fp4",
+    "points": 17,
+    "minConc": 1,
+    "maxConc": 48,
+    "latestDate": "2026-08-15"
+  },
+  {
+    "id": 439584,
+    "model": "minimaxm3",
+    "hardware": "h100",
+    "framework": "vllm",
+    "precision": "fp8",
+    "points": 7,
+    "minConc": 1,
+    "maxConc": 8,
+    "latestDate": "2026-08-12"
+  },
+  {
+    "id": 439641,
+    "model": "minimaxm3",
+    "hardware": "h200",
+    "framework": "vllm",
+    "precision": "fp8",
+    "points": 8,
+    "minConc": 1,
+    "maxConc": 14,
+    "latestDate": "2026-08-12"
+  },
+  {
+    "id": 439832,
+    "model": "minimaxm3",
+    "hardware": "mi300x",
+    "framework": "vllm",
+    "precision": "fp8",
+    "points": 6,
+    "minConc": 2,
+    "maxConc": 16,
+    "latestDate": "2026-08-13"
+  },
+  {
+    "id": 439824,
+    "model": "minimaxm3",
+    "hardware": "mi325x",
+    "framework": "vllm",
+    "precision": "fp8",
+    "points": 9,
+    "minConc": 1,
+    "maxConc": 18,
+    "latestDate": "2026-08-13"
+  },
+  {
+    "id": 439556,
+    "model": "minimaxm3",
+    "hardware": "mi355x",
+    "framework": "vllm",
+    "precision": "fp4",
+    "points": 17,
+    "minConc": 1,
+    "maxConc": 40,
+    "latestDate": "2026-08-12"
+  },
+  {
+    "id": 438230,
+    "model": "qwen3.5",
+    "hardware": "b200",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 32,
+    "minConc": 1,
+    "maxConc": 72,
+    "latestDate": "2026-08-03"
+  },
+  {
+    "id": 438930,
+    "model": "qwen3.5",
+    "hardware": "b200",
+    "framework": "sglang",
+    "precision": "fp8",
+    "points": 32,
+    "minConc": 1,
+    "maxConc": 76,
+    "latestDate": "2026-08-07"
+  },
+  {
+    "id": 438219,
+    "model": "qwen3.5",
+    "hardware": "b300",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 28,
+    "minConc": 1,
+    "maxConc": 72,
+    "latestDate": "2026-08-03"
+  },
+  {
+    "id": 438967,
+    "model": "qwen3.5",
+    "hardware": "b300",
+    "framework": "sglang",
+    "precision": "fp8",
+    "points": 34,
+    "minConc": 1,
+    "maxConc": 72,
+    "latestDate": "2026-08-07"
+  },
+  {
+    "id": 440086,
+    "model": "qwen3.5",
+    "hardware": "gb200",
+    "framework": "dynamo-sglang",
+    "precision": "fp4",
+    "points": 33,
+    "minConc": 1,
+    "maxConc": 72,
+    "latestDate": "2026-08-19"
+  },
+  {
+    "id": 438856,
+    "model": "qwen3.5",
+    "hardware": "gb300",
+    "framework": "dynamo-sglang",
+    "precision": "fp4",
+    "points": 14,
+    "minConc": 1,
+    "maxConc": 128,
+    "latestDate": "2026-08-07"
+  },
+  {
+    "id": 439945,
+    "model": "qwen3.5",
+    "hardware": "gb300",
+    "framework": "dynamo-trt",
+    "precision": "fp4",
+    "points": 6,
+    "minConc": 7,
+    "maxConc": 704,
+    "latestDate": "2026-08-16"
+  },
+  {
+    "id": 438985,
+    "model": "qwen3.5",
+    "hardware": "h100",
+    "framework": "sglang",
+    "precision": "fp8",
+    "points": 9,
+    "minConc": 1,
+    "maxConc": 16,
+    "latestDate": "2026-08-07"
+  },
+  {
+    "id": 439978,
+    "model": "qwen3.5",
+    "hardware": "h200",
+    "framework": "sglang",
+    "precision": "fp8",
+    "points": 16,
+    "minConc": 2,
+    "maxConc": 24,
+    "latestDate": "2026-08-18"
+  },
+  {
+    "id": 439573,
+    "model": "qwen3.5",
+    "hardware": "mi300x",
+    "framework": "sglang",
+    "precision": "fp8",
+    "points": 10,
+    "minConc": 4,
+    "maxConc": 32,
+    "latestDate": "2026-08-12"
+  },
+  {
+    "id": 439628,
+    "model": "qwen3.5",
+    "hardware": "mi325x",
+    "framework": "sglang",
+    "precision": "fp8",
+    "points": 32,
+    "minConc": 1,
+    "maxConc": 80,
+    "latestDate": "2026-08-12"
+  },
+  {
+    "id": 440140,
+    "model": "qwen3.5",
+    "hardware": "mi355x",
+    "framework": "sglang",
+    "precision": "fp4",
+    "points": 31,
+    "minConc": 1,
+    "maxConc": 40,
+    "latestDate": "2026-08-20"
+  }
+]
diff --git a/packages/app/cypress/fixtures/api/collectivex-latest.json b/packages/app/cypress/fixtures/api/collectivex-latest.json
index 42f0c5e76..c0ae5a471 100644
--- a/packages/app/cypress/fixtures/api/collectivex-latest.json
+++ b/packages/app/cypress/fixtures/api/collectivex-latest.json
@@ -1 +1 @@
-{"version":1,"run":{"run_id":"160","run_attempt":1,"generated_at":"2026-07-08T12:20:00Z","conclusion":"success","source_sha":"cccccccccccccccccccccccccccccccccccccccc","requested_cases":5,"terminal_cases":4,"measured_cases":3,"unsupported_cases":1,"failed_cases":0,"requested_points":50,"terminal_points":40,"measured_points":30,"covered_skus":["b200-dgxc","b300","h200-dgxc","mi355x"]},"coverage":[{"case_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-bf16","label":"h200-dgxc · deepep-v2 · normal · decode · EP8 · bf16","disposition":"runnable","sku":"h200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null},{"case_id":"mi355x-mori-deepseek-v3-normal-decode-ep16-uniform-bf16","label":"mi355x · mori · normal · decode · EP16 · bf16","disposition":"runnable","sku":"mi355x","backend":"mori","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":16,"nodes":2,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","topology_class":"mi355x-xgmi-rdma"},"points":[{"tokens_per_rank":1,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":4096,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":8192,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null},{"case_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","label":"h200-dgxc · deepep-v2 · normal · decode · EP8 · fp8","disposition":"runnable","sku":"h200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"fp8","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null},{"case_id":"b300-deepep-v2-deepseek-v3-normal-decode-ep16-uniform-bf16","label":"b300 · deepep-v2 · normal · decode · EP16 · bf16","disposition":"unsupported","sku":"b300","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":16,"nodes":2,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":"rdma","topology_class":"b300-nvlink-rdma"},"points":[{"tokens_per_rank":1,"global_tokens":16,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":2,"global_tokens":32,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":4,"global_tokens":64,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":8,"global_tokens":128,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":16,"global_tokens":256,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":32,"global_tokens":512,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":64,"global_tokens":1024,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":128,"global_tokens":2048,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":256,"global_tokens":4096,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":512,"global_tokens":8192,"terminal_status":"unsupported","reason":"backend-platform-unsupported"}],"outcome":"unsupported","reason":"backend-platform-unsupported","detail":"unsupported by the selected backend/platform"},{"case_id":"b200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-bf16","label":"b200-dgxc · deepep-v2 · normal · decode · EP8 · bf16","disposition":"runnable","sku":"b200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"b200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"pending","reason":"pending"}],"outcome":"pending","reason":"pending","detail":null}],"series":[{"series_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-bf16","phase":"decode","mode":"normal","precision":"bf16","backend":"deepep-v2","system":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island","sku":"h200-dgxc","vendor":"nvidia"},"points":[{"tokens_per_rank":1,"global_tokens":8,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.90407673860912,"p90":111.02229327648992,"p95":107.05721137375814,"p99":99.92006394884093},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":200.3978666666667,"p90":185.55358024691355,"p95":178.92666666666665,"p99":166.9982222222222},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.69257142857144,"p90":113.60423280423281,"p95":109.5469387755102,"p99":102.24380952380953},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.50975895765473,"p90":98.62014718301363,"p95":95.09799906933456,"p99":88.75813246471228},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.61722488038276,"p90":110.75668970405812,"p95":106.8010936431989,"p99":99.68102073365232},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":198.74168595041323,"p90":184.02007958371595,"p95":177.4479338842975,"p99":165.61807162534438},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.3803765903308,"p90":113.31516350956554,"p95":109.2681933842239,"p99":101.983647158609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.39423861171366,"p90":98.51318389973487,"p95":94.99485590331577,"p99":88.6618655097614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.33174224343675,"p90":110.49235392910809,"p95":106.54619843163995,"p99":99.4431185361973},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":197.11265573770493,"p90":182.51171827565267,"p95":175.9934426229508,"p99":164.26054644808744},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.06976649746193,"p90":113.02756157172401,"p95":108.99086294416244,"p99":101.72480541455162},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.27896858071506,"p90":98.40645238955098,"p95":94.8919362327813,"p99":88.5658071505959},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.04761904761905,"p90":110.22927689594357,"p95":106.29251700680271,"p99":99.2063492063492},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":195.51011382113822,"p90":181.02788316772057,"p95":174.56260162601623,"p99":162.9250948509485},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.76072911392406,"p90":112.74141584622596,"p95":108.71493670886076,"p99":101.46727426160338},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.16394805194805,"p90":98.29995189995189,"p95":94.78923933209646,"p99":88.4699567099567},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.76484560570071,"p90":109.96744963490808,"p95":106.04004071937563,"p99":98.97070467141727},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":193.9334193548387,"p90":179.5679808841099,"p95":173.1548387096774,"p99":161.61118279569894},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.45325252525252,"p90":112.45671530115975,"p95":108.44040404040403,"p99":101.21104377104378},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.04917621621622,"p90":98.19368168168167,"p95":94.68676447876449,"p99":88.37431351351351},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.48341232227487,"p90":109.70686326136563,"p95":105.78876100203114,"p99":98.73617693522907},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":192.38195199999998,"p90":178.13143703703705,"p95":171.7696,"p99":160.31829333333334},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.1473249370277,"p90":112.17344901576638,"p95":108.16725440806044,"p99":100.95610411418976},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.93465226781858,"p90":98.0876409887209,"p95":94.58451095340943,"p99":88.2788768898488},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.2033096926714,"p90":109.44750897469572,"p95":105.5386693684566,"p99":98.50275807722618},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":190.8551111111111,"p90":176.717695473251,"p95":170.4063492063492,"p99":159.04592592592593},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.84293467336684,"p90":111.89160617904336,"p95":107.89547738693467,"p99":100.70244556113903},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.82037540453075,"p90":97.98182907826921,"p95":94.48247803975958,"p99":88.1836461704423},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.9245283018868,"p90":109.18937805730259,"p95":105.2897574123989,"p99":98.27044025157234},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":189.3523149606299,"p90":175.32621755613883,"p95":169.06456692913383,"p99":157.79359580052494},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.5400701754386,"p90":111.61117608836908,"p95":107.62506265664159,"p99":100.45005847953217},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.70634482758621,"p90":97.87624521072797,"p95":94.38066502463055,"p99":88.08862068965517},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.64705882352942,"p90":108.93246187363833,"p95":105.04201680672269,"p99":98.0392156862745},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":187.873,"p90":173.95648148148146,"p95":167.74374999999998,"p99":156.56083333333333},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.23872,"p90":111.33214814814815,"p95":107.356,"p99":100.19893333333334},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.5925597416577,"p90":97.77088864968306,"p95":94.27907119790866,"p99":87.99379978471475},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.37089201877934,"p90":108.67675186924012,"p95":104.79543930248154,"p99":97.80907668231613},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":186.41662015503874,"p90":172.60798162503588,"p95":166.44341085271316,"p99":155.3471834625323},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.93887281795512,"p90":111.05451186847694,"p95":107.08827930174563,"p99":99.94906068162926},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.47901935483871,"p90":97.66575866188768,"p95":94.17769585253455,"p99":87.89918279569893},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]},{"series_id":"mi355x-mori-deepseek-v3-normal-decode-ep16-uniform-bf16","phase":"decode","mode":"normal","precision":"bf16","backend":"mori","system":{"ep_size":16,"nodes":2,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","topology_class":"mi355x-xgmi-rdma","sku":"mi355x","vendor":"amd"},"points":[{"tokens_per_rank":1,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.95203836930456,"p90":55.51114663824496,"p95":53.52860568687907,"p99":49.96003197442047},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":100.19893333333334,"p90":92.77679012345678,"p95":89.46333333333332,"p99":83.4991111111111},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":61.34628571428572,"p90":56.802116402116404,"p95":54.7734693877551,"p99":51.121904761904766},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.25487947882736,"p90":49.310073591506814,"p95":47.54899953466728,"p99":44.37906623235614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.80861244019138,"p90":55.37834485202906,"p95":53.40054682159945,"p99":49.84051036682616},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":99.37084297520661,"p90":92.01003979185798,"p95":88.72396694214875,"p99":82.80903581267219},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":61.1901882951654,"p90":56.65758175478277,"p95":54.63409669211195,"p99":50.9918235793045},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.19711930585683,"p90":49.256591949867435,"p95":47.49742795165788,"p99":44.3309327548807},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.665871121718375,"p90":55.246176964554046,"p95":53.273099215819975,"p99":49.72155926809865},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":98.55632786885246,"p90":91.25585913782633,"p95":87.9967213114754,"p99":82.13027322404372},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":61.034883248730964,"p90":56.513780785862004,"p95":54.49543147208122,"p99":50.86240270727581},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.13948429035753,"p90":49.20322619477549,"p95":47.44596811639065,"p99":44.28290357529795},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.523809523809526,"p90":55.114638447971785,"p95":53.146258503401356,"p99":49.6031746031746},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":97.75505691056911,"p90":90.51394158386029,"p95":87.28130081300812,"p99":81.46254742547426},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.88036455696203,"p90":56.37070792311298,"p95":54.35746835443038,"p99":50.73363713080169},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.08197402597403,"p90":49.149975949975946,"p95":47.39461966604823,"p99":44.23497835497835},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.38242280285036,"p90":54.98372481745404,"p95":53.02002035968781,"p99":49.48535233570863},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":96.96670967741935,"p90":89.78399044205494,"p95":86.5774193548387,"p99":80.80559139784947},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.72662626262626,"p90":56.22835765057987,"p95":54.22020202020202,"p99":50.60552188552189},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.02458810810811,"p90":49.096840840840834,"p95":47.343382239382244,"p99":44.18715675675676},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.24170616113744,"p90":54.85343163068281,"p95":52.89438050101557,"p99":49.368088467614534},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":96.19097599999999,"p90":89.06571851851852,"p95":85.8848,"p99":80.15914666666667},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.57366246851385,"p90":56.08672450788319,"p95":54.08362720403022,"p99":50.47805205709488},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.96732613390929,"p90":49.04382049436045,"p95":47.29225547670472,"p99":44.1394384449244},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.1016548463357,"p90":54.72375448734786,"p95":52.7693346842283,"p99":49.25137903861309},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":95.42755555555556,"p90":88.3588477366255,"p95":85.2031746031746,"p99":79.52296296296296},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.42146733668342,"p90":55.94580308952168,"p95":53.947738693467336,"p99":50.351222780569515},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.910187702265375,"p90":48.99091453913461,"p95":47.24123901987979,"p99":44.09182308522115},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":58.9622641509434,"p90":54.594689028651295,"p95":52.64487870619945,"p99":49.13522012578617},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":94.67615748031496,"p90":87.66310877806941,"p95":84.53228346456692,"p99":78.89679790026247},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.2700350877193,"p90":55.80558804418454,"p95":53.812531328320794,"p99":50.22502923976609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.853172413793104,"p90":48.938122605363986,"p95":47.19033251231527,"p99":44.044310344827586},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":58.82352941176471,"p90":54.466230936819166,"p95":52.52100840336134,"p99":49.01960784313725},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":93.9365,"p90":86.97824074074073,"p95":83.87187499999999,"p99":78.28041666666667},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.11936,"p90":55.666074074074075,"p95":53.678,"p99":50.09946666666667},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.79627987082885,"p90":48.88544432484153,"p95":47.13953559895433,"p99":43.996899892357376},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":8192,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":58.68544600938967,"p90":54.33837593462006,"p95":52.39771965124077,"p99":48.904538341158066},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":93.20831007751937,"p90":86.30399081251794,"p95":83.22170542635658,"p99":77.67359173126616},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.96943640897756,"p90":55.52725593423847,"p95":53.544139650872815,"p99":49.97453034081463},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.739509677419356,"p90":48.83287933094384,"p95":47.08884792626728,"p99":43.94959139784947},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]},{"series_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","phase":"decode","mode":"normal","precision":"fp8","backend":"deepep-v2","system":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island","sku":"h200-dgxc","vendor":"nvidia"},"points":[{"tokens_per_rank":1,"global_tokens":8,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.90407673860912,"p90":111.02229327648992,"p95":107.05721137375814,"p99":99.92006394884093},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":200.3978666666667,"p90":185.55358024691355,"p95":178.92666666666665,"p99":166.9982222222222},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.69257142857144,"p90":113.60423280423281,"p95":109.5469387755102,"p99":102.24380952380953},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.50975895765473,"p90":98.62014718301363,"p95":95.09799906933456,"p99":88.75813246471228},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.61722488038276,"p90":110.75668970405812,"p95":106.8010936431989,"p99":99.68102073365232},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":198.74168595041323,"p90":184.02007958371595,"p95":177.4479338842975,"p99":165.61807162534438},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.3803765903308,"p90":113.31516350956554,"p95":109.2681933842239,"p99":101.983647158609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.39423861171366,"p90":98.51318389973487,"p95":94.99485590331577,"p99":88.6618655097614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.33174224343675,"p90":110.49235392910809,"p95":106.54619843163995,"p99":99.4431185361973},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":197.11265573770493,"p90":182.51171827565267,"p95":175.9934426229508,"p99":164.26054644808744},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.06976649746193,"p90":113.02756157172401,"p95":108.99086294416244,"p99":101.72480541455162},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.27896858071506,"p90":98.40645238955098,"p95":94.8919362327813,"p99":88.5658071505959},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.04761904761905,"p90":110.22927689594357,"p95":106.29251700680271,"p99":99.2063492063492},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":195.51011382113822,"p90":181.02788316772057,"p95":174.56260162601623,"p99":162.9250948509485},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.76072911392406,"p90":112.74141584622596,"p95":108.71493670886076,"p99":101.46727426160338},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.16394805194805,"p90":98.29995189995189,"p95":94.78923933209646,"p99":88.4699567099567},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.76484560570071,"p90":109.96744963490808,"p95":106.04004071937563,"p99":98.97070467141727},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":193.9334193548387,"p90":179.5679808841099,"p95":173.1548387096774,"p99":161.61118279569894},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.45325252525252,"p90":112.45671530115975,"p95":108.44040404040403,"p99":101.21104377104378},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.04917621621622,"p90":98.19368168168167,"p95":94.68676447876449,"p99":88.37431351351351},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.48341232227487,"p90":109.70686326136563,"p95":105.78876100203114,"p99":98.73617693522907},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":192.38195199999998,"p90":178.13143703703705,"p95":171.7696,"p99":160.31829333333334},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.1473249370277,"p90":112.17344901576638,"p95":108.16725440806044,"p99":100.95610411418976},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.93465226781858,"p90":98.0876409887209,"p95":94.58451095340943,"p99":88.2788768898488},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.2033096926714,"p90":109.44750897469572,"p95":105.5386693684566,"p99":98.50275807722618},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":190.8551111111111,"p90":176.717695473251,"p95":170.4063492063492,"p99":159.04592592592593},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.84293467336684,"p90":111.89160617904336,"p95":107.89547738693467,"p99":100.70244556113903},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.82037540453075,"p90":97.98182907826921,"p95":94.48247803975958,"p99":88.1836461704423},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.9245283018868,"p90":109.18937805730259,"p95":105.2897574123989,"p99":98.27044025157234},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":189.3523149606299,"p90":175.32621755613883,"p95":169.06456692913383,"p99":157.79359580052494},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.5400701754386,"p90":111.61117608836908,"p95":107.62506265664159,"p99":100.45005847953217},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.70634482758621,"p90":97.87624521072797,"p95":94.38066502463055,"p99":88.08862068965517},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.64705882352942,"p90":108.93246187363833,"p95":105.04201680672269,"p99":98.0392156862745},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":187.873,"p90":173.95648148148146,"p95":167.74374999999998,"p99":156.56083333333333},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.23872,"p90":111.33214814814815,"p95":107.356,"p99":100.19893333333334},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.5925597416577,"p90":97.77088864968306,"p95":94.27907119790866,"p99":87.99379978471475},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.37089201877934,"p90":108.67675186924012,"p95":104.79543930248154,"p99":97.80907668231613},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":186.41662015503874,"p90":172.60798162503588,"p95":166.44341085271316,"p99":155.3471834625323},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.93887281795512,"p90":111.05451186847694,"p95":107.08827930174563,"p99":99.94906068162926},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.47901935483871,"p90":97.66575866188768,"p95":94.17769585253455,"p99":87.89918279569893},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]}]}
+{"version":1,"run":{"run_id":"160","run_attempt":1,"generated_at":"2026-07-08T12:20:00Z","conclusion":"success","source_sha":"cccccccccccccccccccccccccccccccccccccccc","requested_cases":5,"terminal_cases":4,"measured_cases":3,"unsupported_cases":1,"failed_cases":0,"requested_points":50,"terminal_points":40,"measured_points":30,"covered_skus":["b200-dgxc","b300","h200-dgxc","mi355x"],"kv_requested_cases":0,"kv_measured_cases":0},"coverage":[{"case_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-bf16","label":"h200-dgxc · deepep-v2 · normal · decode · EP8 · bf16","disposition":"runnable","sku":"h200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null},{"case_id":"mi355x-mori-deepseek-v3-normal-decode-ep16-uniform-bf16","label":"mi355x · mori · normal · decode · EP16 · bf16","disposition":"runnable","sku":"mi355x","backend":"mori","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":16,"nodes":2,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","topology_class":"mi355x-xgmi-rdma"},"points":[{"tokens_per_rank":1,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":4096,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":8192,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null},{"case_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","label":"h200-dgxc · deepep-v2 · normal · decode · EP8 · fp8","disposition":"runnable","sku":"h200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"fp8","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null},{"case_id":"b300-deepep-v2-deepseek-v3-normal-decode-ep16-uniform-bf16","label":"b300 · deepep-v2 · normal · decode · EP16 · bf16","disposition":"unsupported","sku":"b300","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":16,"nodes":2,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":"rdma","topology_class":"b300-nvlink-rdma"},"points":[{"tokens_per_rank":1,"global_tokens":16,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":2,"global_tokens":32,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":4,"global_tokens":64,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":8,"global_tokens":128,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":16,"global_tokens":256,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":32,"global_tokens":512,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":64,"global_tokens":1024,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":128,"global_tokens":2048,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":256,"global_tokens":4096,"terminal_status":"unsupported","reason":"backend-platform-unsupported"},{"tokens_per_rank":512,"global_tokens":8192,"terminal_status":"unsupported","reason":"backend-platform-unsupported"}],"outcome":"unsupported","reason":"backend-platform-unsupported","detail":"unsupported by the selected backend/platform"},{"case_id":"b200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-bf16","label":"b200-dgxc · deepep-v2 · normal · decode · EP8 · bf16","disposition":"runnable","sku":"b200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"bf16","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"b200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"pending","reason":"pending"},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"pending","reason":"pending"}],"outcome":"pending","reason":"pending","detail":null}],"series":[{"series_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-bf16","phase":"decode","mode":"normal","precision":"bf16","backend":"deepep-v2","system":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island","sku":"h200-dgxc","vendor":"nvidia"},"points":[{"tokens_per_rank":1,"global_tokens":8,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.90407673860912,"p90":111.02229327648992,"p95":107.05721137375814,"p99":99.92006394884093},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":200.3978666666667,"p90":185.55358024691355,"p95":178.92666666666665,"p99":166.9982222222222},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.69257142857144,"p90":113.60423280423281,"p95":109.5469387755102,"p99":102.24380952380953},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.50975895765473,"p90":98.62014718301363,"p95":95.09799906933456,"p99":88.75813246471228},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.61722488038276,"p90":110.75668970405812,"p95":106.8010936431989,"p99":99.68102073365232},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":198.74168595041323,"p90":184.02007958371595,"p95":177.4479338842975,"p99":165.61807162534438},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.3803765903308,"p90":113.31516350956554,"p95":109.2681933842239,"p99":101.983647158609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.39423861171366,"p90":98.51318389973487,"p95":94.99485590331577,"p99":88.6618655097614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.33174224343675,"p90":110.49235392910809,"p95":106.54619843163995,"p99":99.4431185361973},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":197.11265573770493,"p90":182.51171827565267,"p95":175.9934426229508,"p99":164.26054644808744},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.06976649746193,"p90":113.02756157172401,"p95":108.99086294416244,"p99":101.72480541455162},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.27896858071506,"p90":98.40645238955098,"p95":94.8919362327813,"p99":88.5658071505959},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.04761904761905,"p90":110.22927689594357,"p95":106.29251700680271,"p99":99.2063492063492},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":195.51011382113822,"p90":181.02788316772057,"p95":174.56260162601623,"p99":162.9250948509485},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.76072911392406,"p90":112.74141584622596,"p95":108.71493670886076,"p99":101.46727426160338},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.16394805194805,"p90":98.29995189995189,"p95":94.78923933209646,"p99":88.4699567099567},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.76484560570071,"p90":109.96744963490808,"p95":106.04004071937563,"p99":98.97070467141727},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":193.9334193548387,"p90":179.5679808841099,"p95":173.1548387096774,"p99":161.61118279569894},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.45325252525252,"p90":112.45671530115975,"p95":108.44040404040403,"p99":101.21104377104378},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.04917621621622,"p90":98.19368168168167,"p95":94.68676447876449,"p99":88.37431351351351},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.48341232227487,"p90":109.70686326136563,"p95":105.78876100203114,"p99":98.73617693522907},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":192.38195199999998,"p90":178.13143703703705,"p95":171.7696,"p99":160.31829333333334},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.1473249370277,"p90":112.17344901576638,"p95":108.16725440806044,"p99":100.95610411418976},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.93465226781858,"p90":98.0876409887209,"p95":94.58451095340943,"p99":88.2788768898488},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.2033096926714,"p90":109.44750897469572,"p95":105.5386693684566,"p99":98.50275807722618},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":190.8551111111111,"p90":176.717695473251,"p95":170.4063492063492,"p99":159.04592592592593},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.84293467336684,"p90":111.89160617904336,"p95":107.89547738693467,"p99":100.70244556113903},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.82037540453075,"p90":97.98182907826921,"p95":94.48247803975958,"p99":88.1836461704423},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.9245283018868,"p90":109.18937805730259,"p95":105.2897574123989,"p99":98.27044025157234},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":189.3523149606299,"p90":175.32621755613883,"p95":169.06456692913383,"p99":157.79359580052494},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.5400701754386,"p90":111.61117608836908,"p95":107.62506265664159,"p99":100.45005847953217},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.70634482758621,"p90":97.87624521072797,"p95":94.38066502463055,"p99":88.08862068965517},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.64705882352942,"p90":108.93246187363833,"p95":105.04201680672269,"p99":98.0392156862745},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":187.873,"p90":173.95648148148146,"p95":167.74374999999998,"p99":156.56083333333333},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.23872,"p90":111.33214814814815,"p95":107.356,"p99":100.19893333333334},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.5925597416577,"p90":97.77088864968306,"p95":94.27907119790866,"p99":87.99379978471475},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.37089201877934,"p90":108.67675186924012,"p95":104.79543930248154,"p99":97.80907668231613},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":186.41662015503874,"p90":172.60798162503588,"p95":166.44341085271316,"p99":155.3471834625323},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.93887281795512,"p90":111.05451186847694,"p95":107.08827930174563,"p99":99.94906068162926},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.47901935483871,"p90":97.66575866188768,"p95":94.17769585253455,"p99":87.89918279569893},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]},{"series_id":"mi355x-mori-deepseek-v3-normal-decode-ep16-uniform-bf16","phase":"decode","mode":"normal","precision":"bf16","backend":"mori","system":{"ep_size":16,"nodes":2,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"xgmi","scale_out_transport":"rdma","topology_class":"mi355x-xgmi-rdma","sku":"mi355x","vendor":"amd"},"points":[{"tokens_per_rank":1,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.95203836930456,"p90":55.51114663824496,"p95":53.52860568687907,"p99":49.96003197442047},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":100.19893333333334,"p90":92.77679012345678,"p95":89.46333333333332,"p99":83.4991111111111},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":61.34628571428572,"p90":56.802116402116404,"p95":54.7734693877551,"p99":51.121904761904766},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.25487947882736,"p90":49.310073591506814,"p95":47.54899953466728,"p99":44.37906623235614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.80861244019138,"p90":55.37834485202906,"p95":53.40054682159945,"p99":49.84051036682616},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":99.37084297520661,"p90":92.01003979185798,"p95":88.72396694214875,"p99":82.80903581267219},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":61.1901882951654,"p90":56.65758175478277,"p95":54.63409669211195,"p99":50.9918235793045},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.19711930585683,"p90":49.256591949867435,"p95":47.49742795165788,"p99":44.3309327548807},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.665871121718375,"p90":55.246176964554046,"p95":53.273099215819975,"p99":49.72155926809865},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":98.55632786885246,"p90":91.25585913782633,"p95":87.9967213114754,"p99":82.13027322404372},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":61.034883248730964,"p90":56.513780785862004,"p95":54.49543147208122,"p99":50.86240270727581},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.13948429035753,"p90":49.20322619477549,"p95":47.44596811639065,"p99":44.28290357529795},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.523809523809526,"p90":55.114638447971785,"p95":53.146258503401356,"p99":49.6031746031746},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":97.75505691056911,"p90":90.51394158386029,"p95":87.28130081300812,"p99":81.46254742547426},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.88036455696203,"p90":56.37070792311298,"p95":54.35746835443038,"p99":50.73363713080169},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.08197402597403,"p90":49.149975949975946,"p95":47.39461966604823,"p99":44.23497835497835},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.38242280285036,"p90":54.98372481745404,"p95":53.02002035968781,"p99":49.48535233570863},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":96.96670967741935,"p90":89.78399044205494,"p95":86.5774193548387,"p99":80.80559139784947},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.72662626262626,"p90":56.22835765057987,"p95":54.22020202020202,"p99":50.60552188552189},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":53.02458810810811,"p90":49.096840840840834,"p95":47.343382239382244,"p99":44.18715675675676},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.24170616113744,"p90":54.85343163068281,"p95":52.89438050101557,"p99":49.368088467614534},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":96.19097599999999,"p90":89.06571851851852,"p95":85.8848,"p99":80.15914666666667},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.57366246851385,"p90":56.08672450788319,"p95":54.08362720403022,"p99":50.47805205709488},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.96732613390929,"p90":49.04382049436045,"p95":47.29225547670472,"p99":44.1394384449244},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.1016548463357,"p90":54.72375448734786,"p95":52.7693346842283,"p99":49.25137903861309},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":95.42755555555556,"p90":88.3588477366255,"p95":85.2031746031746,"p99":79.52296296296296},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.42146733668342,"p90":55.94580308952168,"p95":53.947738693467336,"p99":50.351222780569515},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.910187702265375,"p90":48.99091453913461,"p95":47.24123901987979,"p99":44.09182308522115},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":58.9622641509434,"p90":54.594689028651295,"p95":52.64487870619945,"p99":49.13522012578617},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":94.67615748031496,"p90":87.66310877806941,"p95":84.53228346456692,"p99":78.89679790026247},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.2700350877193,"p90":55.80558804418454,"p95":53.812531328320794,"p99":50.22502923976609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.853172413793104,"p90":48.938122605363986,"p95":47.19033251231527,"p99":44.044310344827586},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":58.82352941176471,"p90":54.466230936819166,"p95":52.52100840336134,"p99":49.01960784313725},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":93.9365,"p90":86.97824074074073,"p95":83.87187499999999,"p99":78.28041666666667},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":60.11936,"p90":55.666074074074075,"p95":53.678,"p99":50.09946666666667},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.79627987082885,"p90":48.88544432484153,"p95":47.13953559895433,"p99":43.996899892357376},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":8192,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":58.68544600938967,"p90":54.33837593462006,"p95":52.39771965124077,"p99":48.904538341158066},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":93.20831007751937,"p90":86.30399081251794,"p95":83.22170542635658,"p99":77.67359173126616},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":59.96943640897756,"p90":55.52725593423847,"p95":53.544139650872815,"p99":49.97453034081463},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":52.739509677419356,"p90":48.83287933094384,"p95":47.08884792626728,"p99":43.94959139784947},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]},{"series_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","phase":"decode","mode":"normal","precision":"fp8","backend":"deepep-v2","system":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island","sku":"h200-dgxc","vendor":"nvidia"},"points":[{"tokens_per_rank":1,"global_tokens":8,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.90407673860912,"p90":111.02229327648992,"p95":107.05721137375814,"p99":99.92006394884093},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":200.3978666666667,"p90":185.55358024691355,"p95":178.92666666666665,"p99":166.9982222222222},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.69257142857144,"p90":113.60423280423281,"p95":109.5469387755102,"p99":102.24380952380953},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.50975895765473,"p90":98.62014718301363,"p95":95.09799906933456,"p99":88.75813246471228},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.61722488038276,"p90":110.75668970405812,"p95":106.8010936431989,"p99":99.68102073365232},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":198.74168595041323,"p90":184.02007958371595,"p95":177.4479338842975,"p99":165.61807162534438},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.3803765903308,"p90":113.31516350956554,"p95":109.2681933842239,"p99":101.983647158609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.39423861171366,"p90":98.51318389973487,"p95":94.99485590331577,"p99":88.6618655097614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.33174224343675,"p90":110.49235392910809,"p95":106.54619843163995,"p99":99.4431185361973},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":197.11265573770493,"p90":182.51171827565267,"p95":175.9934426229508,"p99":164.26054644808744},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.06976649746193,"p90":113.02756157172401,"p95":108.99086294416244,"p99":101.72480541455162},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.27896858071506,"p90":98.40645238955098,"p95":94.8919362327813,"p99":88.5658071505959},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.04761904761905,"p90":110.22927689594357,"p95":106.29251700680271,"p99":99.2063492063492},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":195.51011382113822,"p90":181.02788316772057,"p95":174.56260162601623,"p99":162.9250948509485},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.76072911392406,"p90":112.74141584622596,"p95":108.71493670886076,"p99":101.46727426160338},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.16394805194805,"p90":98.29995189995189,"p95":94.78923933209646,"p99":88.4699567099567},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.76484560570071,"p90":109.96744963490808,"p95":106.04004071937563,"p99":98.97070467141727},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":193.9334193548387,"p90":179.5679808841099,"p95":173.1548387096774,"p99":161.61118279569894},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.45325252525252,"p90":112.45671530115975,"p95":108.44040404040403,"p99":101.21104377104378},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.04917621621622,"p90":98.19368168168167,"p95":94.68676447876449,"p99":88.37431351351351},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.48341232227487,"p90":109.70686326136563,"p95":105.78876100203114,"p99":98.73617693522907},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":192.38195199999998,"p90":178.13143703703705,"p95":171.7696,"p99":160.31829333333334},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.1473249370277,"p90":112.17344901576638,"p95":108.16725440806044,"p99":100.95610411418976},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.93465226781858,"p90":98.0876409887209,"p95":94.58451095340943,"p99":88.2788768898488},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.2033096926714,"p90":109.44750897469572,"p95":105.5386693684566,"p99":98.50275807722618},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":190.8551111111111,"p90":176.717695473251,"p95":170.4063492063492,"p99":159.04592592592593},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.84293467336684,"p90":111.89160617904336,"p95":107.89547738693467,"p99":100.70244556113903},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.82037540453075,"p90":97.98182907826921,"p95":94.48247803975958,"p99":88.1836461704423},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.9245283018868,"p90":109.18937805730259,"p95":105.2897574123989,"p99":98.27044025157234},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":189.3523149606299,"p90":175.32621755613883,"p95":169.06456692913383,"p99":157.79359580052494},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.5400701754386,"p90":111.61117608836908,"p95":107.62506265664159,"p99":100.45005847953217},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.70634482758621,"p90":97.87624521072797,"p95":94.38066502463055,"p99":88.08862068965517},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.64705882352942,"p90":108.93246187363833,"p95":105.04201680672269,"p99":98.0392156862745},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":187.873,"p90":173.95648148148146,"p95":167.74374999999998,"p99":156.56083333333333},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.23872,"p90":111.33214814814815,"p95":107.356,"p99":100.19893333333334},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.5925597416577,"p90":97.77088864968306,"p95":94.27907119790866,"p99":87.99379978471475},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.37089201877934,"p90":108.67675186924012,"p95":104.79543930248154,"p99":97.80907668231613},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":186.41662015503874,"p90":172.60798162503588,"p95":166.44341085271316,"p99":155.3471834625323},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.93887281795512,"p90":111.05451186847694,"p95":107.08827930174563,"p99":99.94906068162926},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.47901935483871,"p90":97.66575866188768,"p95":94.17769585253455,"p99":87.89918279569893},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]}],"kv":[]}
diff --git a/packages/app/cypress/fixtures/api/collectivex-run-159.json b/packages/app/cypress/fixtures/api/collectivex-run-159.json
index 6cc60f58e..bc6b2fb26 100644
--- a/packages/app/cypress/fixtures/api/collectivex-run-159.json
+++ b/packages/app/cypress/fixtures/api/collectivex-run-159.json
@@ -1 +1 @@
-{"version":1,"run":{"run_id":"159","run_attempt":1,"generated_at":"2026-07-07T12:20:00Z","conclusion":"success","source_sha":"dddddddddddddddddddddddddddddddddddddddd","requested_cases":1,"terminal_cases":1,"measured_cases":1,"unsupported_cases":0,"failed_cases":0,"requested_points":10,"terminal_points":10,"measured_points":10,"covered_skus":["h200-dgxc"]},"coverage":[{"case_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","label":"h200-dgxc · deepep-v2 · normal · decode · EP8 · fp8","disposition":"runnable","sku":"h200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"fp8","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null}],"series":[{"series_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","phase":"decode","mode":"normal","precision":"fp8","backend":"deepep-v2","system":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island","sku":"h200-dgxc","vendor":"nvidia"},"points":[{"tokens_per_rank":1,"global_tokens":8,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.90407673860912,"p90":111.02229327648992,"p95":107.05721137375814,"p99":99.92006394884093},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":200.3978666666667,"p90":185.55358024691355,"p95":178.92666666666665,"p99":166.9982222222222},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.69257142857144,"p90":113.60423280423281,"p95":109.5469387755102,"p99":102.24380952380953},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.50975895765473,"p90":98.62014718301363,"p95":95.09799906933456,"p99":88.75813246471228},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.61722488038276,"p90":110.75668970405812,"p95":106.8010936431989,"p99":99.68102073365232},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":198.74168595041323,"p90":184.02007958371595,"p95":177.4479338842975,"p99":165.61807162534438},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.3803765903308,"p90":113.31516350956554,"p95":109.2681933842239,"p99":101.983647158609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.39423861171366,"p90":98.51318389973487,"p95":94.99485590331577,"p99":88.6618655097614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.33174224343675,"p90":110.49235392910809,"p95":106.54619843163995,"p99":99.4431185361973},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":197.11265573770493,"p90":182.51171827565267,"p95":175.9934426229508,"p99":164.26054644808744},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.06976649746193,"p90":113.02756157172401,"p95":108.99086294416244,"p99":101.72480541455162},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.27896858071506,"p90":98.40645238955098,"p95":94.8919362327813,"p99":88.5658071505959},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.04761904761905,"p90":110.22927689594357,"p95":106.29251700680271,"p99":99.2063492063492},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":195.51011382113822,"p90":181.02788316772057,"p95":174.56260162601623,"p99":162.9250948509485},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.76072911392406,"p90":112.74141584622596,"p95":108.71493670886076,"p99":101.46727426160338},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.16394805194805,"p90":98.29995189995189,"p95":94.78923933209646,"p99":88.4699567099567},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.76484560570071,"p90":109.96744963490808,"p95":106.04004071937563,"p99":98.97070467141727},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":193.9334193548387,"p90":179.5679808841099,"p95":173.1548387096774,"p99":161.61118279569894},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.45325252525252,"p90":112.45671530115975,"p95":108.44040404040403,"p99":101.21104377104378},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.04917621621622,"p90":98.19368168168167,"p95":94.68676447876449,"p99":88.37431351351351},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.48341232227487,"p90":109.70686326136563,"p95":105.78876100203114,"p99":98.73617693522907},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":192.38195199999998,"p90":178.13143703703705,"p95":171.7696,"p99":160.31829333333334},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.1473249370277,"p90":112.17344901576638,"p95":108.16725440806044,"p99":100.95610411418976},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.93465226781858,"p90":98.0876409887209,"p95":94.58451095340943,"p99":88.2788768898488},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.2033096926714,"p90":109.44750897469572,"p95":105.5386693684566,"p99":98.50275807722618},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":190.8551111111111,"p90":176.717695473251,"p95":170.4063492063492,"p99":159.04592592592593},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.84293467336684,"p90":111.89160617904336,"p95":107.89547738693467,"p99":100.70244556113903},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.82037540453075,"p90":97.98182907826921,"p95":94.48247803975958,"p99":88.1836461704423},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.9245283018868,"p90":109.18937805730259,"p95":105.2897574123989,"p99":98.27044025157234},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":189.3523149606299,"p90":175.32621755613883,"p95":169.06456692913383,"p99":157.79359580052494},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.5400701754386,"p90":111.61117608836908,"p95":107.62506265664159,"p99":100.45005847953217},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.70634482758621,"p90":97.87624521072797,"p95":94.38066502463055,"p99":88.08862068965517},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.64705882352942,"p90":108.93246187363833,"p95":105.04201680672269,"p99":98.0392156862745},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":187.873,"p90":173.95648148148146,"p95":167.74374999999998,"p99":156.56083333333333},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.23872,"p90":111.33214814814815,"p95":107.356,"p99":100.19893333333334},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.5925597416577,"p90":97.77088864968306,"p95":94.27907119790866,"p99":87.99379978471475},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.37089201877934,"p90":108.67675186924012,"p95":104.79543930248154,"p99":97.80907668231613},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":186.41662015503874,"p90":172.60798162503588,"p95":166.44341085271316,"p99":155.3471834625323},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.93887281795512,"p90":111.05451186847694,"p95":107.08827930174563,"p99":99.94906068162926},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.47901935483871,"p90":97.66575866188768,"p95":94.17769585253455,"p99":87.89918279569893},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]}]}
+{"version":1,"run":{"run_id":"159","run_attempt":1,"generated_at":"2026-07-07T12:20:00Z","conclusion":"success","source_sha":"dddddddddddddddddddddddddddddddddddddddd","requested_cases":1,"terminal_cases":1,"measured_cases":1,"unsupported_cases":0,"failed_cases":0,"requested_points":10,"terminal_points":10,"measured_points":10,"covered_skus":["h200-dgxc"],"kv_requested_cases":0,"kv_measured_cases":0},"coverage":[{"case_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","label":"h200-dgxc · deepep-v2 · normal · decode · EP8 · fp8","disposition":"runnable","sku":"h200-dgxc","backend":"deepep-v2","phase":"decode","mode":"normal","precision":"fp8","topology":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island"},"points":[{"tokens_per_rank":1,"global_tokens":8,"terminal_status":"measured","reason":null},{"tokens_per_rank":2,"global_tokens":16,"terminal_status":"measured","reason":null},{"tokens_per_rank":4,"global_tokens":32,"terminal_status":"measured","reason":null},{"tokens_per_rank":8,"global_tokens":64,"terminal_status":"measured","reason":null},{"tokens_per_rank":16,"global_tokens":128,"terminal_status":"measured","reason":null},{"tokens_per_rank":32,"global_tokens":256,"terminal_status":"measured","reason":null},{"tokens_per_rank":64,"global_tokens":512,"terminal_status":"measured","reason":null},{"tokens_per_rank":128,"global_tokens":1024,"terminal_status":"measured","reason":null},{"tokens_per_rank":256,"global_tokens":2048,"terminal_status":"measured","reason":null},{"tokens_per_rank":512,"global_tokens":4096,"terminal_status":"measured","reason":null}],"outcome":"success","reason":null,"detail":null}],"series":[{"series_id":"h200-dgxc-deepep-v2-deepseek-v3-normal-decode-ep8-uniform-fp8","phase":"decode","mode":"normal","precision":"fp8","backend":"deepep-v2","system":{"ep_size":8,"nodes":1,"gpus_per_node":8,"scale_up_domain":8,"scale_up_transport":"nvlink","scale_out_transport":null,"topology_class":"h200-nvlink-island","sku":"h200-dgxc","vendor":"nvidia"},"points":[{"tokens_per_rank":1,"global_tokens":8,"components":{"dispatch":{"latency_us":{"p50":417,"p90":450.36,"p95":467.04,"p99":500.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":922.6952134292566,"p90":854.3474198419042,"p95":823.8350119904076,"p99":768.9126778577139},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.90407673860912,"p90":111.02229327648992,"p95":107.05721137375814,"p99":99.92006394884093},"payload_bytes":400000000},"stage":{"latency_us":{"p50":120,"p90":129.60000000000002,"p95":134.4,"p99":144},"activation_data_rate_gbps_at_latency_percentile":{"p50":1603.1829333333335,"p90":1484.4286419753084,"p95":1431.4133333333332,"p99":1335.9857777777777},"payload_data_rate_gbps_at_latency_percentile":{"p50":200.3978666666667,"p90":185.55358024691355,"p95":178.92666666666665,"p99":166.9982222222222},"payload_bytes":192381952},"combine":{"latency_us":{"p50":392,"p90":423.36,"p95":439.04,"p99":470.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":981.5405714285715,"p90":908.8338624338625,"p95":876.3755102040816,"p99":817.9504761904763},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.69257142857144,"p90":113.60423280423281,"p95":109.5469387755102,"p99":102.24380952380953},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":921,"p90":994.6800000000001,"p95":1031.5200000000002,"p99":1105.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":835.5350792616721,"p90":773.6435919089556,"p95":746.0134636264928,"p99":696.27923271806},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.50975895765473,"p90":98.62014718301363,"p95":95.09799906933456,"p99":88.75813246471228},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":2,"global_tokens":16,"components":{"dispatch":{"latency_us":{"p50":418,"p90":451.44000000000005,"p95":468.16,"p99":501.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":920.4878086124403,"p90":852.3035264930002,"p95":821.8641148325358,"p99":767.0731738437003},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.61722488038276,"p90":110.75668970405812,"p95":106.8010936431989,"p99":99.68102073365232},"payload_bytes":400000000},"stage":{"latency_us":{"p50":121,"p90":130.68,"p95":135.52,"p99":145.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1589.9334876033058,"p90":1472.1606366697276,"p95":1419.58347107438,"p99":1324.944573002755},"payload_data_rate_gbps_at_latency_percentile":{"p50":198.74168595041323,"p90":184.02007958371595,"p95":177.4479338842975,"p99":165.61807162534438},"payload_bytes":192381952},"combine":{"latency_us":{"p50":393,"p90":424.44000000000005,"p95":440.16,"p99":471.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":979.0430127226464,"p90":906.5213080765243,"p95":874.1455470737912,"p99":815.869177268872},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.3803765903308,"p90":113.31516350956554,"p95":109.2681933842239,"p99":101.983647158609},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":922,"p90":995.7600000000001,"p95":1032.64,"p99":1106.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":834.6288590021693,"p90":772.8044990760825,"p95":745.2043383947938,"p99":695.5240491684744},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.39423861171366,"p90":98.51318389973487,"p95":94.99485590331577,"p99":88.6618655097614},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":4,"global_tokens":32,"components":{"dispatch":{"latency_us":{"p50":419,"p90":452.52000000000004,"p95":469.28000000000003,"p99":502.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":918.2909403341289,"p90":850.2693891982674,"p95":819.9026252983293,"p99":765.2424502784409},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.33174224343675,"p90":110.49235392910809,"p95":106.54619843163995,"p99":99.4431185361973},"payload_bytes":400000000},"stage":{"latency_us":{"p50":122,"p90":131.76000000000002,"p95":136.64000000000001,"p99":146.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1576.9012459016394,"p90":1460.0937462052214,"p95":1407.9475409836064,"p99":1314.0843715846995},"payload_data_rate_gbps_at_latency_percentile":{"p50":197.11265573770493,"p90":182.51171827565267,"p95":175.9934426229508,"p99":164.26054644808744},"payload_bytes":192381952},"combine":{"latency_us":{"p50":394,"p90":425.52000000000004,"p95":441.28000000000003,"p99":472.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":976.5581319796954,"p90":904.2204925737921,"p95":871.9269035532996,"p99":813.798443316413},"payload_data_rate_gbps_at_latency_percentile":{"p50":122.06976649746193,"p90":113.02756157172401,"p95":108.99086294416244,"p99":101.72480541455162},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":923,"p90":996.84,"p95":1033.76,"p99":1107.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":833.724602383532,"p90":771.9672244291962,"p95":744.3969664138677,"p99":694.7705019862767},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.27896858071506,"p90":98.40645238955098,"p95":94.8919362327813,"p99":88.5658071505959},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":8,"global_tokens":64,"components":{"dispatch":{"latency_us":{"p50":420,"p90":453.6,"p95":470.40000000000003,"p99":504},"activation_data_rate_gbps_at_latency_percentile":{"p50":916.1045333333334,"p90":848.244938271605,"p95":817.9504761904761,"p99":763.4204444444445},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.04761904761905,"p90":110.22927689594357,"p95":106.29251700680271,"p99":99.2063492063492},"payload_bytes":400000000},"stage":{"latency_us":{"p50":123,"p90":132.84,"p95":137.76000000000002,"p99":147.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1564.0809105691058,"p90":1448.2230653417646,"p95":1396.5008130081299,"p99":1303.400758807588},"payload_data_rate_gbps_at_latency_percentile":{"p50":195.51011382113822,"p90":181.02788316772057,"p95":174.56260162601623,"p99":162.9250948509485},"payload_bytes":192381952},"combine":{"latency_us":{"p50":395,"p90":426.6,"p95":442.40000000000003,"p99":474},"activation_data_rate_gbps_at_latency_percentile":{"p50":974.0858329113925,"p90":901.9313267698077,"p95":869.7194936708861,"p99":811.738194092827},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.76072911392406,"p90":112.74141584622596,"p95":108.71493670886076,"p99":101.46727426160338},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":924,"p90":997.9200000000001,"p95":1034.88,"p99":1108.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":832.822303030303,"p90":771.1317620650954,"p95":743.5913419913419,"p99":694.0185858585859},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.16394805194805,"p90":98.29995189995189,"p95":94.78923933209646,"p99":88.4699567099567},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":16,"global_tokens":128,"components":{"dispatch":{"latency_us":{"p50":421,"p90":454.68,"p95":471.52000000000004,"p99":505.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":913.928513064133,"p90":846.230104689012,"p95":816.0076009501188,"p99":761.607094220111},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.76484560570071,"p90":109.96744963490808,"p95":106.04004071937563,"p99":98.97070467141727},"payload_bytes":400000000},"stage":{"latency_us":{"p50":124,"p90":133.92000000000002,"p95":138.88000000000002,"p99":148.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1551.4673548387095,"p90":1436.543847072879,"p95":1385.2387096774191,"p99":1292.8894623655915},"payload_data_rate_gbps_at_latency_percentile":{"p50":193.9334193548387,"p90":179.5679808841099,"p95":173.1548387096774,"p99":161.61118279569894},"payload_bytes":192381952},"combine":{"latency_us":{"p50":396,"p90":427.68,"p95":443.52000000000004,"p99":475.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":971.6260202020202,"p90":899.653722409278,"p95":867.5232323232323,"p99":809.6883501683502},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.45325252525252,"p90":112.45671530115975,"p95":108.44040404040403,"p99":101.21104377104378},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":925,"p90":999.0000000000001,"p95":1036,"p99":1110},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.9219545945946,"p90":770.298106106106,"p95":742.7874594594595,"p99":693.2682954954955},"payload_data_rate_gbps_at_latency_percentile":{"p50":106.04917621621622,"p90":98.19368168168167,"p95":94.68676447876449,"p99":88.37431351351351},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":32,"global_tokens":256,"components":{"dispatch":{"latency_us":{"p50":422,"p90":455.76000000000005,"p95":472.64000000000004,"p99":506.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":911.7628056872038,"p90":844.2248200807443,"p95":814.0739336492891,"p99":759.8023380726698},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.48341232227487,"p90":109.70686326136563,"p95":105.78876100203114,"p99":98.73617693522907},"payload_bytes":400000000},"stage":{"latency_us":{"p50":125,"p90":135,"p95":140,"p99":150},"activation_data_rate_gbps_at_latency_percentile":{"p50":1539.0556159999999,"p90":1425.0514962962964,"p95":1374.1568,"p99":1282.5463466666668},"payload_data_rate_gbps_at_latency_percentile":{"p50":192.38195199999998,"p90":178.13143703703705,"p95":171.7696,"p99":160.31829333333334},"payload_bytes":192381952},"combine":{"latency_us":{"p50":397,"p90":428.76000000000005,"p95":444.64000000000004,"p99":476.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":969.1785994962216,"p90":897.387592126131,"p95":865.3380352644836,"p99":807.648832913518},"payload_data_rate_gbps_at_latency_percentile":{"p50":121.1473249370277,"p90":112.17344901576638,"p95":108.16725440806044,"p99":100.95610411418976},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":926,"p90":1000.08,"p95":1037.1200000000001,"p99":1111.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":831.0235507559396,"p90":769.466250699944,"p95":741.9853131749459,"p99":692.5196256299496},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.93465226781858,"p90":98.0876409887209,"p95":94.58451095340943,"p99":88.2788768898488},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":64,"global_tokens":512,"components":{"dispatch":{"latency_us":{"p50":423,"p90":456.84000000000003,"p95":473.76000000000005,"p99":507.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":909.6073380614658,"p90":842.2290167235793,"p95":812.1494089834514,"p99":758.0061150512215},"payload_data_rate_gbps_at_latency_percentile":{"p50":118.2033096926714,"p90":109.44750897469572,"p95":105.5386693684566,"p99":98.50275807722618},"payload_bytes":400000000},"stage":{"latency_us":{"p50":126,"p90":136.08,"p95":141.12,"p99":151.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":1526.840888888889,"p90":1413.741563786008,"p95":1363.2507936507936,"p99":1272.3674074074074},"payload_data_rate_gbps_at_latency_percentile":{"p50":190.8551111111111,"p90":176.717695473251,"p95":170.4063492063492,"p99":159.04592592592593},"payload_bytes":192381952},"combine":{"latency_us":{"p50":398,"p90":429.84000000000003,"p95":445.76000000000005,"p99":477.59999999999997},"activation_data_rate_gbps_at_latency_percentile":{"p50":966.7434773869347,"p90":895.1328494323469,"p95":863.1638190954774,"p99":805.6195644891122},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.84293467336684,"p90":111.89160617904336,"p95":107.89547738693467,"p99":100.70244556113903},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":927,"p90":1001.1600000000001,"p95":1038.24,"p99":1112.3999999999999},"activation_data_rate_gbps_at_latency_percentile":{"p50":830.1270852211435,"p90":768.6361900195773,"p95":741.1848975188781,"p99":691.7725710176197},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.82037540453075,"p90":97.98182907826921,"p95":94.48247803975958,"p99":88.1836461704423},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":128,"global_tokens":1024,"components":{"dispatch":{"latency_us":{"p50":424,"p90":457.92,"p95":474.88000000000005,"p99":508.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":907.4620377358491,"p90":840.2426275331936,"p95":810.2339622641508,"p99":756.2183647798744},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.9245283018868,"p90":109.18937805730259,"p95":105.2897574123989,"p99":98.27044025157234},"payload_bytes":400000000},"stage":{"latency_us":{"p50":127,"p90":137.16,"p95":142.24,"p99":152.4},"activation_data_rate_gbps_at_latency_percentile":{"p50":1514.8185196850393,"p90":1402.6097404491106,"p95":1352.5165354330707,"p99":1262.3487664041995},"payload_data_rate_gbps_at_latency_percentile":{"p50":189.3523149606299,"p90":175.32621755613883,"p95":169.06456692913383,"p99":157.79359580052494},"payload_bytes":192381952},"combine":{"latency_us":{"p50":399,"p90":430.92,"p95":446.88000000000005,"p99":478.79999999999995},"activation_data_rate_gbps_at_latency_percentile":{"p50":964.3205614035088,"p90":892.8894087069526,"p95":861.0005012531327,"p99":803.6004678362574},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.5400701754386,"p90":111.61117608836908,"p95":107.62506265664159,"p99":100.45005847953217},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":928,"p90":1002.24,"p95":1039.3600000000001,"p99":1113.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":829.232551724138,"p90":767.8079182630906,"p95":740.3862068965517,"p99":691.0271264367817},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.70634482758621,"p90":97.87624521072797,"p95":94.38066502463055,"p99":88.08862068965517},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":256,"global_tokens":2048,"components":{"dispatch":{"latency_us":{"p50":425,"p90":459.00000000000006,"p95":476.00000000000006,"p99":510},"activation_data_rate_gbps_at_latency_percentile":{"p50":905.3268329411765,"p90":838.2655860566448,"p95":808.3275294117645,"p99":754.4390274509803},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.64705882352942,"p90":108.93246187363833,"p95":105.04201680672269,"p99":98.0392156862745},"payload_bytes":400000000},"stage":{"latency_us":{"p50":128,"p90":138.24,"p95":143.36,"p99":153.6},"activation_data_rate_gbps_at_latency_percentile":{"p50":1502.984,"p90":1391.6518518518517,"p95":1341.9499999999998,"p99":1252.4866666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":187.873,"p90":173.95648148148146,"p95":167.74374999999998,"p99":156.56083333333333},"payload_bytes":192381952},"combine":{"latency_us":{"p50":400,"p90":432,"p95":448.00000000000006,"p99":480},"activation_data_rate_gbps_at_latency_percentile":{"p50":961.90976,"p90":890.6571851851852,"p95":858.848,"p99":801.5914666666667},"payload_data_rate_gbps_at_latency_percentile":{"p50":120.23872,"p90":111.33214814814815,"p95":107.356,"p99":100.19893333333334},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":929,"p90":1003.32,"p95":1040.48,"p99":1114.8},"activation_data_rate_gbps_at_latency_percentile":{"p50":828.3399440258343,"p90":766.9814296535502,"p95":739.5892357373519,"p99":690.2832866881952},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.5925597416577,"p90":97.77088864968306,"p95":94.27907119790866,"p99":87.99379978471475},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}},{"tokens_per_rank":512,"global_tokens":4096,"components":{"dispatch":{"latency_us":{"p50":426,"p90":460.08000000000004,"p95":477.12000000000006,"p99":511.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":903.2016525821597,"p90":836.2978264649626,"p95":806.4300469483567,"p99":752.6680438184663},"payload_data_rate_gbps_at_latency_percentile":{"p50":117.37089201877934,"p90":108.67675186924012,"p95":104.79543930248154,"p99":97.80907668231613},"payload_bytes":400000000},"stage":{"latency_us":{"p50":129,"p90":139.32000000000002,"p95":144.48000000000002,"p99":154.79999999999998},"activation_data_rate_gbps_at_latency_percentile":{"p50":1491.33296124031,"p90":1380.863853000287,"p95":1331.5472868217053,"p99":1242.7774677002585},"payload_data_rate_gbps_at_latency_percentile":{"p50":186.41662015503874,"p90":172.60798162503588,"p95":166.44341085271316,"p99":155.3471834625323},"payload_bytes":192381952},"combine":{"latency_us":{"p50":401,"p90":433.08000000000004,"p95":449.12000000000006,"p99":481.2},"activation_data_rate_gbps_at_latency_percentile":{"p50":959.5109825436409,"p90":888.4360949478155,"p95":856.706234413965,"p99":799.5924854530341},"payload_data_rate_gbps_at_latency_percentile":{"p50":119.93887281795512,"p90":111.05451186847694,"p95":107.08827930174563,"p99":99.94906068162926},"payload_bytes":384763904},"roundtrip":{"latency_us":{"p50":930,"p90":1004.4000000000001,"p95":1041.6000000000001,"p99":1116},"activation_data_rate_gbps_at_latency_percentile":{"p50":827.4492559139785,"p90":766.1567184388689,"p95":738.7939784946236,"p99":689.541046594982},"payload_data_rate_gbps_at_latency_percentile":{"p50":105.47901935483871,"p90":97.66575866188768,"p95":94.17769585253455,"p99":87.89918279569893},"payload_bytes":784763904}},"roundtrip_token_rate_at_latency_percentile":{"p50":8338218,"p90":9005275.440000001,"p95":9338804.16,"p99":10005861.6}}]}],"kv":[]}
diff --git a/packages/app/cypress/fixtures/api/collectivex-runs.json b/packages/app/cypress/fixtures/api/collectivex-runs.json
index efdc80210..6888fbfea 100644
--- a/packages/app/cypress/fixtures/api/collectivex-runs.json
+++ b/packages/app/cypress/fixtures/api/collectivex-runs.json
@@ -1 +1 @@
-{"version":1,"runs":[{"run_id":"160","run_attempt":1,"generated_at":"2026-07-08T12:20:00Z","conclusion":"success","covered_skus":["b200-dgxc","b300","h200-dgxc","mi355x"],"requested_cases":5,"measured_cases":3,"requested_points":50,"terminal_points":40,"terminal_counts":{"measured":3,"unsupported":1,"failed":0}},{"run_id":"159","run_attempt":1,"generated_at":"2026-07-07T12:20:00Z","conclusion":"success","covered_skus":["h200-dgxc"],"requested_cases":1,"measured_cases":1,"requested_points":10,"terminal_points":10,"terminal_counts":{"measured":1,"unsupported":0,"failed":0}}],"discovery_complete":true}
+{"version":1,"runs":[{"run_id":"160","run_attempt":1,"generated_at":"2026-07-08T12:20:00Z","conclusion":"success","covered_skus":["b200-dgxc","b300","h200-dgxc","mi355x"],"requested_cases":5,"measured_cases":3,"requested_points":50,"terminal_points":40,"terminal_counts":{"measured":3,"unsupported":1,"failed":0},"kv_cases":{"requested":0,"measured":0}},{"run_id":"159","run_attempt":1,"generated_at":"2026-07-07T12:20:00Z","conclusion":"success","covered_skus":["h200-dgxc"],"requested_cases":1,"measured_cases":1,"requested_points":10,"terminal_points":10,"terminal_counts":{"measured":1,"unsupported":0,"failed":0},"kv_cases":{"requested":0,"measured":0}}],"discovery_complete":true}
diff --git a/packages/app/cypress/support/agentic-fixtures.ts b/packages/app/cypress/support/agentic-fixtures.ts
new file mode 100644
index 000000000..66adff223
--- /dev/null
+++ b/packages/app/cypress/support/agentic-fixtures.ts
@@ -0,0 +1,30 @@
+export function percentileLadder(prefix: string, base: number): Record<string, number> {
+  return {
+    [`median_${prefix}`]: base,
+    [`p75_${prefix}`]: base * 1.2,
+    [`p90_${prefix}`]: base * 1.5,
+    [`p95_${prefix}`]: base * 1.7,
+    [`p99_${prefix}`]: base * 2.2,
+    [`std_${prefix}`]: base * 0.3,
+  };
+}
+
+export function agenticMetrics(concurrency: number): Record<string, number> {
+  const scale = concurrency / 16;
+  const itl = 0.011 * scale;
+  return {
+    ...percentileLadder('ttft', 0.4 * scale),
+    ...percentileLadder('tpot', 0.012 * scale),
+    ...percentileLadder('itl', itl),
+    ...percentileLadder('e2el', 8 * scale),
+    median_intvty: 1 / itl,
+    p75_intvty: 1 / (itl * 1.2),
+    p90_intvty: 1 / (itl * 1.5),
+    p99_intvty: 1 / (itl * 2.2),
+    std_intvty: (1 / itl) * 0.1,
+    tput_per_gpu: 950 / Math.sqrt(scale),
+    output_tput_per_gpu: 210,
+    input_tput_per_gpu: 740,
+    total_tput_tps: 7600 * concurrency * 0.05,
+  };
+}
diff --git a/packages/app/cypress/support/e2e.ts b/packages/app/cypress/support/e2e.ts
index b3c15b58b..a6678efa4 100644
--- a/packages/app/cypress/support/e2e.ts
+++ b/packages/app/cypress/support/e2e.ts
@@ -2,14 +2,54 @@
  * Global e2e setup. Loaded before every `cy.visit` via `supportFile` in
  * `cypress.config.ts`.
  *
- * Snoozes the feedback-modal nudge so it doesn't render its centered modal
- * + backdrop on top of the UI under test. Specs that want to exercise the
+ * Suppresses the centered feedback-modal on the dashboard so its backdrop
+ * doesn't sit on top of the UI under test. Specs that want to exercise the
  * feedback-modal flow can clear `inferencex-feedback-modal-snoozed` in their
- * own `onBeforeLoad`.
+ * own `onBeforeLoad`, which runs after this hook.
  */
+import 'cypress-axe';
+
+let suppressTelemetryTutorial = true;
+let suppressAgenticCoachMark = true;
+
+/**
+ * Opt the whole spec out of the telemetry-tutorial suppression.
+ *
+ * Clearing the key in a visit's `onBeforeLoad` is not enough: this hook also
+ * fires on `cy.reload()`, which takes no `onBeforeLoad`, so the key would be
+ * re-seeded behind the test and a "still dismissed after reload" assertion
+ * would pass even if dismissal never persisted anything. Call this at the top
+ * of any spec that owns telemetry-tutorial state.
+ *
+ * The card is a bottom-right modal on /inference/agentic/[id]. It has no
+ * backdrop, but it sits over the last chart in the grid, so agentic specs
+ * suppress it by default.
+ */
+export function keepTelemetryTutorial(): void {
+  suppressTelemetryTutorial = false;
+}
+
+/**
+ * Opt the whole spec out of the agentic point coach-mark suppression, for the
+ * same reason as `keepTelemetryTutorial` — the key is re-seeded on
+ * `cy.reload()`, so a per-visit `onBeforeLoad` clear cannot own its state.
+ *
+ * The callout is anchored to a point inside `[data-testid="scatter-graph"]`,
+ * so on the agentic view it sits over the plot area that other specs click.
+ */
+export function keepAgenticCoachMark(): void {
+  suppressAgenticCoachMark = false;
+}
+
 Cypress.on('window:before:load', (win) => {
   try {
     win.localStorage.setItem('inferencex-feedback-modal-snoozed', String(Date.now()));
+    if (suppressTelemetryTutorial) {
+      win.localStorage.setItem('inferencex-agentx-telemetry-tutorial-dismissed', '1');
+    }
+    if (suppressAgenticCoachMark) {
+      win.localStorage.setItem('inferencex-agentic-point-coach-mark-dismissed', '1');
+    }
   } catch {
     // localStorage unavailable — fine, the test will just see the modal.
   }
@@ -19,14 +59,14 @@ Cypress.on('window:before:load', (win) => {
  * Seed the shared feature-gate flag (the same localStorage key the ↑↑↓↓ konami
  * unlock writes — see use-feature-gate.ts).
  *
- * The agentic surfaces (the "Agentic Traces" scenario, /datasets,
- * /inference/agentic/[id], and the Datasets nav link) are now PUBLIC by default
+ * The agentic surfaces (the "Agentic" scenario, /agentx,
+ * /inference/agentic/[id], and the AgentX nav link) are now PUBLIC by default
  * — they no longer sit behind this gate — so agentic specs no longer need it.
  * The helper is retained as a harmless no-op for those specs (and still unlocks
  * the remaining hidden features: the "Hidden" tab dropdown and Measured Energy).
  *
  * Call from a spec's `cy.visit(..., { onBeforeLoad })`:
- *   cy.visit('/datasets/x', { onBeforeLoad: unlockAgenticGate });
+ *   cy.visit('/agentx/x', { onBeforeLoad: unlockAgenticGate });
  * or compose inside an existing hook: `unlockAgenticGate(win)`.
  */
 export function unlockAgenticGate(win: Window): void {
diff --git a/packages/app/cypress/support/legend-advanced.ts b/packages/app/cypress/support/legend-advanced.ts
new file mode 100644
index 000000000..ff0f48a48
--- /dev/null
+++ b/packages/app/cypress/support/legend-advanced.ts
@@ -0,0 +1,20 @@
+/**
+ * Expand the chart legend's collapsed-by-default Advanced drawer.
+ *
+ * Log Scale, Labels, High Contrast, Parallelism Labels, and Line Labels live
+ * inside the Advanced section, so tests must open the drawer before reading
+ * or clicking those switches. Idempotent — safe to call when the drawer is
+ * already expanded.
+ *
+ * `cy.get` respects `cy.within` scoping, so call this inside a `.within()`
+ * block to target a specific chart's legend (e.g. a replay panel).
+ */
+export const expandLegendAdvanced = (): void => {
+  cy.get('[data-testid="legend-advanced-toggle"]')
+    .first()
+    .then(($btn) => {
+      if ($btn.attr('aria-expanded') !== 'true') {
+        cy.wrap($btn).click();
+      }
+    });
+};
diff --git a/packages/app/cypress/support/mock-data.ts b/packages/app/cypress/support/mock-data.ts
index c66bf850c..b75ef5342 100644
--- a/packages/app/cypress/support/mock-data.ts
+++ b/packages/app/cypress/support/mock-data.ts
@@ -2,8 +2,11 @@ import type {
   AggDataEntry,
   ChartDefinition,
   HardwareConfig,
-  InferenceChartContextType,
+  InferenceActionsContextType,
   InferenceData,
+  InferenceDataContextType,
+  InferenceDisplayContextType,
+  InferenceFiltersContextType,
 } from '@/components/inference/types';
 import type {
   EvaluationChartContextType,
@@ -13,7 +16,13 @@ import type {
   ModelSuccessRateData,
   ReliabilityChartContextType,
 } from '@/components/reliability/types';
-import type { GlobalFilterContextType } from '@/components/GlobalFilterContext';
+import type {
+  GlobalFilterActionsContextType,
+  GlobalFilterAvailabilityContextType,
+  GlobalFilterRunContextType,
+  GlobalFilterSelectionContextType,
+  GlobalFilterWorkflowContextType,
+} from '@/components/GlobalFilterContext';
 import type { UnofficialRunContextType } from '@/components/unofficial-run-provider';
 import { computeToggle } from '@/hooks/useTogglableSet';
 import { Model, Sequence, Precision } from '@/lib/data-mappings';
@@ -153,12 +162,17 @@ export function createMockInferenceData(overrides?: Partial<InferenceData>): Inf
 }
 
 // ---------------------------------------------------------------------------
-// Inference context
+// Inference domain contexts
 // ---------------------------------------------------------------------------
 
-export function createMockInferenceContext(
-  overrides?: Partial<InferenceChartContextType>,
-): InferenceChartContextType {
+export type MockInferenceContextValues = InferenceDataContextType &
+  InferenceFiltersContextType &
+  InferenceDisplayContextType &
+  InferenceActionsContextType;
+
+export function createMockInferenceContextValues(
+  overrides?: Partial<MockInferenceContextValues>,
+): MockInferenceContextValues {
   const hwConfig = createMockHardwareConfig();
   return {
     activeHwTypes: new Set(['h100', 'b200', 'b200_trt', 'mi300x', 'h200']),
@@ -195,15 +209,23 @@ export function createMockInferenceContext(
     setSelectedPrecisions: namedStub('setSelectedPrecisions'),
     loading: false,
     error: null,
-    workflowInfo: null,
     selectedYAxisMetric: 'y_tpPerGpu',
     setSelectedYAxisMetric: namedStub('setSelectedYAxisMetric'),
+    tokenRevenuePriceSource: 'normalized',
+    tokenRevenuePricing: {
+      source: 'normalized',
+      inputPerMillion: 1,
+      outputPerMillion: 1,
+    },
+    openRouterModelId: 'deepseek/deepseek-r1-0528',
+    openRouterPricingLoading: false,
+    openRouterPricingError: null,
+    setTokenRevenuePriceSource: namedStub('setTokenRevenuePriceSource'),
     selectedPercentile: 'p90',
     setSelectedPercentile: namedStub('setSelectedPercentile'),
     selectedXAxisMetric: null,
     setSelectedXAxisMetric: namedStub('setSelectedXAxisMetric'),
     selectedE2eXAxisMetric: null,
-    setSelectedE2eXAxisMetric: namedStub('setSelectedE2eXAxisMetric'),
     selectedXAxisMode: 'interactivity' as const,
     setSelectedXAxisMode: namedStub('setSelectedXAxisMode'),
     scaleType: 'auto',
@@ -226,14 +248,12 @@ export function createMockInferenceContext(
     setLogScale: namedStub('setLogScale'),
     useAdvancedLabels: false,
     setUseAdvancedLabels: namedStub('setUseAdvancedLabels'),
+    showConcurrencyLabels: false,
+    setShowConcurrencyLabels: namedStub('setShowConcurrencyLabels'),
     showGradientLabels: false,
     setShowGradientLabels: namedStub('setShowGradientLabels'),
     showLineLabels: false,
     setShowLineLabels: namedStub('setShowLineLabels'),
-    showSpeedOverlay: false,
-    setShowSpeedOverlay: namedStub('setShowSpeedOverlay'),
-    showMinecraftOverlay: false,
-    setShowMinecraftOverlay: namedStub('setShowMinecraftOverlay'),
     selectedGPUs: [],
     setSelectedGPUs: namedStub('setSelectedGPUs'),
     availableGPUs: [
@@ -254,7 +274,7 @@ export function createMockInferenceContext(
     availableDates: ['2025-02-28', '2025-03-01'],
     dateRangeAvailableDates: ['2025-02-28', '2025-03-01'],
     isCheckingAvailableDates: false,
-    availableRuns: null,
+    availableRuns: {},
     selectedRunId: '12345678',
     setSelectedRunId: namedStub('setSelectedRunId'),
     availablePrecisions: [Precision.FP4, Precision.FP8],
@@ -414,14 +434,29 @@ export function createMockReliabilityContext(
   };
 }
 
-// ---------------------------------------------------------------------------
-// Global filter context
+// Global filter contexts
 // ---------------------------------------------------------------------------
 
-export function createMockGlobalFilterContext(
-  overrides?: Partial<GlobalFilterContextType>,
-): GlobalFilterContextType {
-  return {
+export type GlobalFilterContextOverrides = Partial<
+  GlobalFilterSelectionContextType &
+    GlobalFilterActionsContextType &
+    GlobalFilterRunContextType &
+    GlobalFilterAvailabilityContextType &
+    GlobalFilterWorkflowContextType
+>;
+
+export interface MockGlobalFilterContexts {
+  selection: GlobalFilterSelectionContextType;
+  actions: GlobalFilterActionsContextType;
+  run: GlobalFilterRunContextType;
+  availability: GlobalFilterAvailabilityContextType;
+  workflow: GlobalFilterWorkflowContextType;
+}
+
+export function createMockGlobalFilterContexts(
+  overrides: GlobalFilterContextOverrides = {},
+): MockGlobalFilterContexts {
+  const values = {
     selectedModel: Model.DeepSeek_R1,
     setSelectedModel: namedStub('setSelectedModel_global'),
     selectedSequence: Sequence.EightK_OneK,
@@ -429,8 +464,6 @@ export function createMockGlobalFilterContext(
     selectedPrecisions: [Precision.FP4],
     setSelectedPrecisions: namedStub('setSelectedPrecisions_global'),
     effectiveSequence: Sequence.EightK_OneK,
-    // Mocks represent a settled state: availability is known and the sequence is
-    // resolved. Tests exercising the pre-availability window override this.
     sequenceResolved: true,
     effectivePrecisions: [Precision.FP4],
     selectedRunDate: '2025-03-01',
@@ -444,12 +477,51 @@ export function createMockGlobalFilterContext(
     availableDates: ['2025-02-28', '2025-03-01'],
     effectiveRunDate: '2025-03-01',
     availabilityRows: undefined,
-    workflowInfo: null,
+    availabilitySettled: true,
+    availabilityError: null,
     availableRuns: {},
     workflowLoading: false,
     workflowError: null,
     ...overrides,
   };
+
+  return {
+    selection: {
+      selectedModel: values.selectedModel,
+      selectedSequence: values.selectedSequence,
+      selectedPrecisions: values.selectedPrecisions,
+      effectiveSequence: values.effectiveSequence,
+      sequenceResolved: values.sequenceResolved,
+      effectivePrecisions: values.effectivePrecisions,
+    },
+    actions: {
+      setSelectedModel: values.setSelectedModel,
+      setSelectedSequence: values.setSelectedSequence,
+      setSelectedPrecisions: values.setSelectedPrecisions,
+      setSelectedRunDate: values.setSelectedRunDate,
+      setSelectedRunId: values.setSelectedRunId,
+    },
+    run: {
+      selectedRunDate: values.selectedRunDate,
+      selectedRunDateRev: values.selectedRunDateRev,
+      selectedRunId: values.selectedRunId,
+      effectiveRunDate: values.effectiveRunDate,
+    },
+    availability: {
+      availableModels: values.availableModels,
+      availableSequences: values.availableSequences,
+      availablePrecisions: values.availablePrecisions,
+      availableDates: values.availableDates,
+      availabilityRows: values.availabilityRows,
+      availabilitySettled: values.availabilitySettled,
+      availabilityError: values.availabilityError,
+    },
+    workflow: {
+      availableRuns: values.availableRuns,
+      workflowLoading: values.workflowLoading,
+      workflowError: values.workflowError,
+    },
+  };
 }
 
 // ---------------------------------------------------------------------------
@@ -476,12 +548,11 @@ export function createMockUnofficialRunContext(
       .stub()
       .returns(null) as unknown as UnofficialRunContextType['getOverlayData'],
     activeOverlayHwTypes: new Set<string>(),
-    setActiveOverlayHwTypes: namedStub('setActiveOverlayHwTypes'),
     allOverlayHwTypes: new Set<string>(),
-    toggleOverlayHwType: namedStub('toggleOverlayHwType'),
-    resetOverlayHwTypes: namedStub('resetOverlayHwTypes'),
     localOfficialOverride: null,
-    setLocalOfficialOverride: namedStub('setLocalOfficialOverride'),
+    reconcileOverlayScope: namedStub('reconcileOverlayScope'),
+    setUnifiedOverlaySelection: namedStub('setUnifiedOverlaySelection'),
+    resetOverlaySelection: namedStub('resetOverlaySelection'),
     ...overrides,
   };
 }
diff --git a/packages/app/cypress/support/mock-router.ts b/packages/app/cypress/support/mock-router.ts
new file mode 100644
index 000000000..c7130ccef
--- /dev/null
+++ b/packages/app/cypress/support/mock-router.ts
@@ -0,0 +1,16 @@
+import type { ContextType } from 'react';
+import type { AppRouterContext } from 'next/dist/shared/lib/app-router-context.shared-runtime';
+
+export type MockAppRouter = NonNullable<ContextType<typeof AppRouterContext>>;
+
+export function createMockRouter(): MockAppRouter {
+  return {
+    push: cy.stub(),
+    replace: cy.stub(),
+    refresh: cy.stub(),
+    back: cy.stub(),
+    forward: cy.stub(),
+    prefetch: cy.stub().resolves(),
+    bfcacheId: '',
+  };
+}
diff --git a/packages/app/cypress/support/overlay-fixtures.ts b/packages/app/cypress/support/overlay-fixtures.ts
index 2f3268500..ffd006585 100644
--- a/packages/app/cypress/support/overlay-fixtures.ts
+++ b/packages/app/cypress/support/overlay-fixtures.ts
@@ -4,8 +4,11 @@
  * The benchmark values are the real numbers from GitHub run 29682242847
  * (GLM5.2 B300 agentic hicache, offload=on rows):
  *   conc, p90_intvty (tok/s/user), tput_per_gpu, p90_e2el (s)
- * C=4 is dominated on e2e by C=8 (12874 tok/s @ 33.1s vs 9415 @ 48.0s), which
- * makes the set a ready-made probe for the e2e-restricted frontier behaviors.
+ * C=4 is dominated on e2e by C=8 (12874 tok/s @ 33.1s vs 9415 @ 48.0s). That
+ * used to make the set a probe for the e2e-restricted frontier; #736 removed
+ * that gating, so on the interactivity axes all five are Pareto-optimal —
+ * throughput falls monotonically as interactivity rises. Use
+ * `DOMINATED_CONFIG` when a test needs a point Optimal Only will actually drop.
  */
 export const DEFAULT_MODEL_DB_KEY = 'dsv4';
 export const AGENTIC_DATE = '2026-07-19';
@@ -39,9 +42,21 @@ export const metricsFor = (intvty: number, tput: number, e2el: number): Record<s
   input_tput_per_gpu: tput * 0.7,
 });
 
+/**
+ * A sixth config that is dominated on the *interactivity* axes: C=8 beats it on
+ * both interactivity (68.5 > 60) and throughput (12874 > 5000). Since #736 the
+ * Pareto frontier is computed from the selected axes alone, so this is the point
+ * Optimal Only must drop — the e2e-dominated C=4 above no longer is one.
+ */
+export const DOMINATED_CONFIG: [number, number, number, number] = [6, 60, 5000, 40];
+
 let idCursor = 900000;
-export const b300Rows = (runUrl: string | null, hardware = 'b300') =>
-  REAL_CONFIGS.map(([conc, intvty, tput, e2el]) => ({
+export const b300Rows = (
+  runUrl: string | null,
+  hardware = 'b300',
+  configs: [number, number, number, number][] = REAL_CONFIGS,
+) =>
+  configs.map(([conc, intvty, tput, e2el]) => ({
     id: runUrl ? 0 : idCursor++,
     hardware,
     framework: 'sglang',
@@ -81,8 +96,18 @@ export const availability = [
   },
 ];
 
-/** Intercept availability + benchmarks + unofficial-run with the B300 fixture. */
-export const interceptOverlayRun = ({ overlayHardware = 'b300' } = {}) => {
+/**
+ * Intercept availability + benchmarks + unofficial-run with the B300 fixture.
+ * `overlayConfigs` overrides only the overlay run's rows, leaving the official
+ * rows as the standard five.
+ */
+export const interceptOverlayRun = ({
+  overlayHardware = 'b300',
+  overlayConfigs = REAL_CONFIGS,
+}: {
+  overlayHardware?: string;
+  overlayConfigs?: [number, number, number, number][];
+} = {}) => {
   cy.intercept('GET', '/api/v1/availability', { body: availability }).as('availability');
   cy.intercept('GET', '/api/v1/benchmarks*', { body: b300Rows(null) }).as('benchmarks');
   cy.intercept('GET', '/api/unofficial-run*', {
@@ -100,7 +125,7 @@ export const interceptOverlayRun = ({ overlayHardware = 'b300' } = {}) => {
           isNonMainBranch: true,
         },
       ],
-      benchmarks: b300Rows(OVERLAY_RUN_URL, overlayHardware),
+      benchmarks: b300Rows(OVERLAY_RUN_URL, overlayHardware, overlayConfigs),
       evaluations: [],
     },
   }).as('unofficialRun');
diff --git a/packages/app/cypress/support/test-utils.tsx b/packages/app/cypress/support/test-utils.tsx
index f7de968f1..ff7fe2784 100644
--- a/packages/app/cypress/support/test-utils.tsx
+++ b/packages/app/cypress/support/test-utils.tsx
@@ -1,35 +1,39 @@
 import React from 'react';
 import { QueryClient, QueryClientProvider } from '@tanstack/react-query';
 
-import { InferenceContext } from '@/components/inference/InferenceContext';
+import { InferenceContextsProvider } from '@/components/inference/InferenceContext';
 import { EvaluationContext } from '@/components/evaluation/EvaluationContext';
 import { ReliabilityContext } from '@/components/reliability/ReliabilityContext';
 import {
-  GlobalFilterContext,
-  type GlobalFilterContextType,
+  GlobalFilterActionsContext,
+  GlobalFilterAvailabilityContext,
+  GlobalFilterRunContext,
+  GlobalFilterSelectionContext,
+  GlobalFilterWorkflowContext,
 } from '@/components/GlobalFilterContext';
 import {
   UnofficialRunContext,
   type UnofficialRunContextType,
 } from '@/components/unofficial-run-provider';
 
-import type { InferenceChartContextType } from '@/components/inference/types';
 import type { EvaluationChartContextType } from '@/components/evaluation/types';
 import type { ReliabilityChartContextType } from '@/components/reliability/types';
 
 import {
-  createMockInferenceContext,
+  createMockInferenceContextValues,
   createMockEvaluationContext,
   createMockReliabilityContext,
-  createMockGlobalFilterContext,
+  createMockGlobalFilterContexts,
   createMockUnofficialRunContext,
+  type GlobalFilterContextOverrides,
+  type MockInferenceContextValues,
 } from './mock-data';
 
 export interface ProviderOverrides {
-  inference?: Partial<InferenceChartContextType>;
+  inference?: Partial<MockInferenceContextValues>;
   evaluation?: Partial<EvaluationChartContextType>;
   reliability?: Partial<ReliabilityChartContextType>;
-  globalFilters?: Partial<GlobalFilterContextType>;
+  globalFilters?: GlobalFilterContextOverrides;
   unofficial?: Partial<UnofficialRunContextType>;
 }
 
@@ -53,8 +57,12 @@ export function mountWithProviders(
   let tree = component;
 
   if (overrides.inference !== undefined) {
-    const value = createMockInferenceContext(overrides.inference);
-    tree = <InferenceContext.Provider value={value}>{tree}</InferenceContext.Provider>;
+    const value = createMockInferenceContextValues(overrides.inference);
+    tree = (
+      <InferenceContextsProvider data={value} filters={value} display={value} actions={value}>
+        {tree}
+      </InferenceContextsProvider>
+    );
   }
 
   if (overrides.evaluation !== undefined) {
@@ -68,8 +76,20 @@ export function mountWithProviders(
   }
 
   if (overrides.globalFilters !== undefined) {
-    const value = createMockGlobalFilterContext(overrides.globalFilters);
-    tree = <GlobalFilterContext.Provider value={value}>{tree}</GlobalFilterContext.Provider>;
+    const values = createMockGlobalFilterContexts(overrides.globalFilters);
+    tree = (
+      <GlobalFilterActionsContext.Provider value={values.actions}>
+        <GlobalFilterSelectionContext.Provider value={values.selection}>
+          <GlobalFilterRunContext.Provider value={values.run}>
+            <GlobalFilterAvailabilityContext.Provider value={values.availability}>
+              <GlobalFilterWorkflowContext.Provider value={values.workflow}>
+                {tree}
+              </GlobalFilterWorkflowContext.Provider>
+            </GlobalFilterAvailabilityContext.Provider>
+          </GlobalFilterRunContext.Provider>
+        </GlobalFilterSelectionContext.Provider>
+      </GlobalFilterActionsContext.Provider>
+    );
   }
 
   if (overrides.unofficial !== undefined) {
diff --git a/packages/app/next.config.ts b/packages/app/next.config.ts
index b8445156c..bd96a0437 100644
--- a/packages/app/next.config.ts
+++ b/packages/app/next.config.ts
@@ -12,6 +12,20 @@ const nextConfig: NextConfig = {
   allowedDevOrigins: allowedDevOriginsFromEnv(),
   transpilePackages: ['@semianalysisai/inferencex-constants'],
   serverExternalPackages: ['shiki'],
+  redirects() {
+    return Promise.resolve([
+      {
+        source: '/datasets/:path*',
+        destination: '/agentx/:path*',
+        permanent: true,
+      },
+      {
+        source: '/zh/datasets/:path*',
+        destination: '/zh/agentx/:path*',
+        permanent: true,
+      },
+    ]);
+  },
   experimental: {
     optimizePackageImports: ['lucide-react', 'd3', '@tanstack/react-query'],
     // NOTE: experimental.inlineCss was evaluated (2026-07) for the PageSpeed
@@ -26,6 +40,7 @@ const nextConfig: NextConfig = {
     ...(process.env.GITHUB_ACTIONS === 'true' && { turbopackFileSystemCacheForBuild: true }),
   },
   images: {
+    qualities: [75, 100],
     remotePatterns: [
       { hostname: 'placehold.co' },
       { hostname: 'substack-post-media.s3.amazonaws.com' },
diff --git a/packages/app/package.json b/packages/app/package.json
index 7ed615df2..579ec2852 100644
--- a/packages/app/package.json
+++ b/packages/app/package.json
@@ -5,7 +5,7 @@
   "description": "InferenceX - Inference performance benchmarking and visualization",
   "repository": {
     "type": "git",
-    "url": "https://github.com/SemiAnalysisAI/inferencemax-app"
+    "url": "https://github.com/SemiAnalysisAI/InferenceX-app"
   },
   "sideEffects": [
     "**/*.css",
@@ -20,7 +20,7 @@
     "test:e2e": "bun run test:e2e:quick",
     "test:e2e:quick": "bun run test:e2e:quick:component && bun run test:e2e:quick:integration",
     "test:e2e:quick:component": "cypress run --component --spec cypress/component/chart-buttons.cy.tsx,cypress/component/scatter-graph.cy.tsx",
-    "test:e2e:quick:integration": "cypress run --spec cypress/e2e/sanity.cy.ts,cypress/e2e/inference-chart.cy.ts,cypress/e2e/evaluation-chart.cy.ts,cypress/e2e/reliability-chart.cy.ts,cypress/e2e/zh-pages.cy.ts,cypress/e2e/csv-export-overlay.cy.ts,cypress/e2e/overlay-optimal-only.cy.ts",
+    "test:e2e:quick:integration": "cypress run --spec cypress/e2e/sanity.cy.ts,cypress/e2e/inference-chart.cy.ts,cypress/e2e/evaluation-chart.cy.ts,cypress/e2e/reliability-chart.cy.ts,cypress/e2e/zh-pages.cy.ts,cypress/e2e/csv-export-overlay.cy.ts,cypress/e2e/overlay-optimal-only.cy.ts,cypress/e2e/resident-sequence-lengths.cy.ts,cypress/e2e/overview-a11y.cy.ts",
     "test:e2e:full": "cypress run --component && cypress run",
     "test:e2e:component": "cypress run --component",
     "test:e2e:integration": "cypress run",
@@ -35,9 +35,8 @@
   "dependencies": {
     "@chenglou/pretext": "^0.0.8",
     "@jpinsonneau/html-to-image": "^1.11.13",
-    "@noble/ciphers": "^2.2.0",
-    "@posthog/nextjs-config": "^1.9.68",
-    "@radix-ui/react-accordion": "^1.2.20",
+    "@noble/ciphers": "^2.3.0",
+    "@posthog/nextjs-config": "^1.9.69",
     "@radix-ui/react-dialog": "^1.1.23",
     "@radix-ui/react-label": "^2.1.15",
     "@radix-ui/react-popover": "^1.1.23",
@@ -46,26 +45,26 @@
     "@radix-ui/react-switch": "^1.3.7",
     "@radix-ui/react-tabs": "^1.1.21",
     "@radix-ui/react-tooltip": "^1.2.16",
-    "@react-three/fiber": "^9.6.1",
+    "@react-three/fiber": "^9.7.0",
     "@semianalysisai/inferencex-constants": "workspace:*",
     "@semianalysisai/inferencex-db": "workspace:*",
-    "@shikijs/rehype": "^4.3.1",
+    "@shikijs/rehype": "^4.4.3",
     "@tailwindcss/typography": "^0.5.20",
     "@tanstack/react-query": "^5.101.4",
     "@vercel/analytics": "^2.0.1",
-    "@vercel/blob": "^2.6.1",
+    "@vercel/blob": "^2.8.0",
     "@vercel/speed-insights": "^2.0.0",
     "adm-zip": "^0.6.0",
     "class-variance-authority": "^0.7.1",
     "clsx": "^2.1.1",
     "d3": "^7.9.0",
-    "dompurify": "^3.4.12",
+    "dompurify": "^3.4.13",
     "gray-matter": "^4.0.3",
     "iwanthue": "^2.1.0",
     "katex": "^0.16.47",
-    "lucide-react": "^1.27.0",
+    "lucide-react": "^1.31.0",
     "mp4-muxer": "^5.2.2",
-    "next": "^16.2.12",
+    "next": "^16.3.1",
     "next-mdx-remote": "^6.0.0",
     "next-themes": "^0.4.6",
     "posthog-js": "^1.407.2",
@@ -74,7 +73,7 @@
     "rehype-katex": "^7.0.1",
     "remark-gfm": "^4.0.1",
     "remark-math": "^6.0.0",
-    "shiki": "^4.3.1",
+    "shiki": "^4.4.3",
     "tailwind-merge": "^3.6.0",
     "three": "^0.185.1"
   },
@@ -84,16 +83,17 @@
     "@types/adm-zip": "^0.5.8",
     "@types/d3": "^7.4.3",
     "@types/mdast": "^4.0.4",
-    "@types/node": "^26.1.1",
-    "@types/react": "^19.2.17",
-    "@types/react-dom": "^19.2.3",
-    "@types/three": "^0.185.1",
+    "@types/node": "^26.2.0",
+    "@types/react": "19.2.17",
+    "@types/react-dom": "19.2.3",
+    "@types/three": "^0.185.4",
     "@vitest/coverage-v8": "^4.1.10",
-    "cypress": "^15.19.0",
+    "axe-core": "^4.13.0",
+    "cypress": "15.19.0",
+    "cypress-axe": "^1.7.0",
     "cypress-split": "^1.25.0",
-    "esbuild": "^0.28.1",
-    "fake-indexeddb": "^6.2.5",
-    "jsdom": "^29.1.1",
+    "esbuild": "^0.28.2",
+    "jsdom": "^30.0.1",
     "rimraf": "^6.1.3",
     "tailwindcss": "^4.3.3",
     "tw-animate-css": "^1.4.0",
diff --git a/packages/app/public/decorative/bus.png b/packages/app/public/decorative/bus.png
deleted file mode 100644
index 7ec6619bf..000000000
Binary files a/packages/app/public/decorative/bus.png and /dev/null differ
diff --git a/packages/app/public/decorative/donkey-chest.png b/packages/app/public/decorative/donkey-chest.png
deleted file mode 100644
index 7697e9b91..000000000
Binary files a/packages/app/public/decorative/donkey-chest.png and /dev/null differ
diff --git a/packages/app/public/decorative/elytra.png b/packages/app/public/decorative/elytra.png
deleted file mode 100644
index 79d83fc3e..000000000
Binary files a/packages/app/public/decorative/elytra.png and /dev/null differ
diff --git a/packages/app/public/decorative/racing-car.png b/packages/app/public/decorative/racing-car.png
deleted file mode 100644
index b463385a5..000000000
Binary files a/packages/app/public/decorative/racing-car.png and /dev/null differ
diff --git a/packages/app/public/images/agentx-methodology/corpus-scale.png b/packages/app/public/images/agentx-methodology/corpus-scale.png
new file mode 100644
index 000000000..a526a6df4
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/corpus-scale.png differ
diff --git a/packages/app/public/images/agentx-methodology/engine-acceptance-controls.png b/packages/app/public/images/agentx-methodology/engine-acceptance-controls.png
new file mode 100644
index 000000000..0606f78b0
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/engine-acceptance-controls.png differ
diff --git a/packages/app/public/images/agentx-methodology/golden-acceptance-length.png b/packages/app/public/images/agentx-methodology/golden-acceptance-length.png
new file mode 100644
index 000000000..5e2893d73
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/golden-acceptance-length.png differ
diff --git a/packages/app/public/images/agentx-methodology/hash-token-ratio.png b/packages/app/public/images/agentx-methodology/hash-token-ratio.png
new file mode 100644
index 000000000..1730ca439
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/hash-token-ratio.png differ
diff --git a/packages/app/public/images/agentx-methodology/pareto-b200-minimax-m3.png b/packages/app/public/images/agentx-methodology/pareto-b200-minimax-m3.png
new file mode 100644
index 000000000..dd8158824
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/pareto-b200-minimax-m3.png differ
diff --git a/packages/app/public/images/agentx-methodology/profiling-cache-sources.png b/packages/app/public/images/agentx-methodology/profiling-cache-sources.png
new file mode 100644
index 000000000..d3a3c49e1
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/profiling-cache-sources.png differ
diff --git a/packages/app/public/images/agentx-methodology/replay-flatspawn.png b/packages/app/public/images/agentx-methodology/replay-flatspawn.png
new file mode 100644
index 000000000..19e99a110
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/replay-flatspawn.png differ
diff --git a/packages/app/public/images/agentx-methodology/replay-linear.png b/packages/app/public/images/agentx-methodology/replay-linear.png
new file mode 100644
index 000000000..75ab6d5aa
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/replay-linear.png differ
diff --git a/packages/app/public/images/agentx-methodology/replay-multi-subagent.png b/packages/app/public/images/agentx-methodology/replay-multi-subagent.png
new file mode 100644
index 000000000..5e8072111
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/replay-multi-subagent.png differ
diff --git a/packages/app/public/images/agentx-methodology/replay-single-subagent.png b/packages/app/public/images/agentx-methodology/replay-single-subagent.png
new file mode 100644
index 000000000..185068040
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/replay-single-subagent.png differ
diff --git a/packages/app/public/images/agentx-methodology/replay-subagent-sidecars.png b/packages/app/public/images/agentx-methodology/replay-subagent-sidecars.png
new file mode 100644
index 000000000..a60af7fed
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/replay-subagent-sidecars.png differ
diff --git a/packages/app/public/images/agentx-methodology/replay-two-subagents-joined.png b/packages/app/public/images/agentx-methodology/replay-two-subagents-joined.png
new file mode 100644
index 000000000..3e39e3de2
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/replay-two-subagents-joined.png differ
diff --git a/packages/app/public/images/agentx-methodology/replay-two-subagents.png b/packages/app/public/images/agentx-methodology/replay-two-subagents.png
new file mode 100644
index 000000000..54d03ed10
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/replay-two-subagents.png differ
diff --git a/packages/app/public/images/agentx-methodology/request-distributions-256k.png b/packages/app/public/images/agentx-methodology/request-distributions-256k.png
new file mode 100644
index 000000000..8f6e03660
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/request-distributions-256k.png differ
diff --git a/packages/app/public/images/agentx-methodology/request-distributions.png b/packages/app/public/images/agentx-methodology/request-distributions.png
new file mode 100644
index 000000000..640c7b68f
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/request-distributions.png differ
diff --git a/packages/app/public/images/agentx-methodology/request-queue-depth.png b/packages/app/public/images/agentx-methodology/request-queue-depth.png
new file mode 100644
index 000000000..724e4dba2
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/request-queue-depth.png differ
diff --git a/packages/app/public/images/agentx-methodology/subagent-distributions-256k.png b/packages/app/public/images/agentx-methodology/subagent-distributions-256k.png
new file mode 100644
index 000000000..e695991ca
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/subagent-distributions-256k.png differ
diff --git a/packages/app/public/images/agentx-methodology/subagent-distributions.png b/packages/app/public/images/agentx-methodology/subagent-distributions.png
new file mode 100644
index 000000000..01b72a3e3
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/subagent-distributions.png differ
diff --git a/packages/app/public/images/agentx-methodology/warmup-cache-sources.png b/packages/app/public/images/agentx-methodology/warmup-cache-sources.png
new file mode 100644
index 000000000..f1bdfd761
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/warmup-cache-sources.png differ
diff --git a/packages/app/public/images/agentx-methodology/warmup-snapshot.png b/packages/app/public/images/agentx-methodology/warmup-snapshot.png
new file mode 100644
index 000000000..d806e911f
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/warmup-snapshot.png differ
diff --git a/packages/app/public/images/agentx-methodology/weka-trace-record.png b/packages/app/public/images/agentx-methodology/weka-trace-record.png
new file mode 100644
index 000000000..43c8263c3
Binary files /dev/null and b/packages/app/public/images/agentx-methodology/weka-trace-record.png differ
diff --git a/packages/app/public/images/agentx-optimizations/aiter-64bit-offsets.png b/packages/app/public/images/agentx-optimizations/aiter-64bit-offsets.png
new file mode 100644
index 000000000..76923c99d
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/aiter-64bit-offsets.png differ
diff --git a/packages/app/public/images/agentx-optimizations/atom-prefix-promotion.png b/packages/app/public/images/agentx-optimizations/atom-prefix-promotion.png
new file mode 100644
index 000000000..371389de3
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/atom-prefix-promotion.png differ
diff --git a/packages/app/public/images/agentx-optimizations/dynamo-ownership-table.png b/packages/app/public/images/agentx-optimizations/dynamo-ownership-table.png
new file mode 100644
index 000000000..b3b39ad11
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/dynamo-ownership-table.png differ
diff --git a/packages/app/public/images/agentx-optimizations/lmcache-tiers.png b/packages/app/public/images/agentx-optimizations/lmcache-tiers.png
new file mode 100644
index 000000000..7bdb5d247
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/lmcache-tiers.png differ
diff --git a/packages/app/public/images/agentx-optimizations/serving-stack.png b/packages/app/public/images/agentx-optimizations/serving-stack.png
new file mode 100644
index 000000000..afc647a1b
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/serving-stack.png differ
diff --git a/packages/app/public/images/agentx-optimizations/sglang-hicache-asymmetry.png b/packages/app/public/images/agentx-optimizations/sglang-hicache-asymmetry.png
new file mode 100644
index 000000000..86a1159f9
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/sglang-hicache-asymmetry.png differ
diff --git a/packages/app/public/images/agentx-optimizations/sglang-radix-staging.png b/packages/app/public/images/agentx-optimizations/sglang-radix-staging.png
new file mode 100644
index 000000000..5e6674142
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/sglang-radix-staging.png differ
diff --git a/packages/app/public/images/agentx-optimizations/trtllm-pipelined-transfer.png b/packages/app/public/images/agentx-optimizations/trtllm-pipelined-transfer.png
new file mode 100644
index 000000000..67c531479
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/trtllm-pipelined-transfer.png differ
diff --git a/packages/app/public/images/agentx-optimizations/vllm-connector-layer.png b/packages/app/public/images/agentx-optimizations/vllm-connector-layer.png
new file mode 100644
index 000000000..a5d79fccf
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/vllm-connector-layer.png differ
diff --git a/packages/app/public/images/agentx-optimizations/vllm-selective-retention.png b/packages/app/public/images/agentx-optimizations/vllm-selective-retention.png
new file mode 100644
index 000000000..da734273f
Binary files /dev/null and b/packages/app/public/images/agentx-optimizations/vllm-selective-retention.png differ
diff --git a/packages/app/public/images/agentx-telemetry/chart-point-tooltip.png b/packages/app/public/images/agentx-telemetry/chart-point-tooltip.png
new file mode 100644
index 000000000..bac3efc25
Binary files /dev/null and b/packages/app/public/images/agentx-telemetry/chart-point-tooltip.png differ
diff --git a/packages/app/public/images/agentx-telemetry/flamegraph.png b/packages/app/public/images/agentx-telemetry/flamegraph.png
new file mode 100644
index 000000000..4864123aa
Binary files /dev/null and b/packages/app/public/images/agentx-telemetry/flamegraph.png differ
diff --git a/packages/app/public/images/agentx-telemetry/per-point-overview.png b/packages/app/public/images/agentx-telemetry/per-point-overview.png
new file mode 100644
index 000000000..49b205c8c
Binary files /dev/null and b/packages/app/public/images/agentx-telemetry/per-point-overview.png differ
diff --git a/packages/app/public/images/agentx-telemetry/point-detail-full.png b/packages/app/public/images/agentx-telemetry/point-detail-full.png
new file mode 100644
index 000000000..e4577b359
Binary files /dev/null and b/packages/app/public/images/agentx-telemetry/point-detail-full.png differ
diff --git a/packages/app/public/images/agentx-telemetry/request-timeline.png b/packages/app/public/images/agentx-telemetry/request-timeline.png
new file mode 100644
index 000000000..38c4b7e46
Binary files /dev/null and b/packages/app/public/images/agentx-telemetry/request-timeline.png differ
diff --git a/packages/app/public/images/brief-overview-of-agentic-workloads/context-accumulation.png b/packages/app/public/images/brief-overview-of-agentic-workloads/context-accumulation.png
new file mode 100644
index 000000000..ac2a8db38
Binary files /dev/null and b/packages/app/public/images/brief-overview-of-agentic-workloads/context-accumulation.png differ
diff --git a/packages/app/public/logos/deepseek-color.svg b/packages/app/public/logos/deepseek-color.svg
new file mode 100644
index 000000000..3fc230240
--- /dev/null
+++ b/packages/app/public/logos/deepseek-color.svg
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>DeepSeek</title><path d="M23.748 4.482c-.254-.124-.364.113-.512.234-.051.039-.094.09-.137.136-.372.397-.806.657-1.373.626-.829-.046-1.537.214-2.163.848-.133-.782-.575-1.248-1.247-1.548-.352-.156-.708-.311-.955-.65-.172-.241-.219-.51-.305-.774-.055-.16-.11-.323-.293-.35-.2-.031-.278.136-.356.276-.313.572-.434 1.202-.422 1.84.027 1.436.633 2.58 1.838 3.393.137.093.172.187.129.323-.082.28-.18.552-.266.833-.055.179-.137.217-.329.14a5.526 5.526 0 01-1.736-1.18c-.857-.828-1.631-1.742-2.597-2.458a11.365 11.365 0 00-.689-.471c-.985-.957.13-1.743.388-1.836.27-.098.093-.432-.779-.428-.872.004-1.67.295-2.687.684a3.055 3.055 0 01-.465.137 9.597 9.597 0 00-2.883-.102c-1.885.21-3.39 1.102-4.497 2.623C.082 8.606-.231 10.684.152 12.85c.403 2.284 1.569 4.175 3.36 5.653 1.858 1.533 3.997 2.284 6.438 2.14 1.482-.085 3.133-.284 4.994-1.86.47.234.962.327 1.78.397.63.059 1.236-.03 1.705-.128.735-.156.684-.837.419-.961-2.155-1.004-1.682-.595-2.113-.926 1.096-1.296 2.746-2.642 3.392-7.003.05-.347.007-.565 0-.845-.004-.17.035-.237.23-.256a4.173 4.173 0 001.545-.475c1.396-.763 1.96-2.015 2.093-3.517.02-.23-.004-.467-.247-.588zM11.581 18c-2.089-1.642-3.102-2.183-3.52-2.16-.392.024-.321.471-.235.763.09.288.207.486.371.739.114.167.192.416-.113.603-.673.416-1.842-.14-1.897-.167-1.361-.802-2.5-1.86-3.301-3.307-.774-1.393-1.224-2.887-1.298-4.482-.02-.386.093-.522.477-.592a4.696 4.696 0 011.529-.039c2.132.312 3.946 1.265 5.468 2.774.868.86 1.525 1.887 2.202 2.891.72 1.066 1.494 2.082 2.48 2.914.348.292.625.514.891.677-.802.09-2.14.11-3.054-.614zm1-6.44a.306.306 0 01.415-.287.302.302 0 01.2.288.306.306 0 01-.31.307.303.303 0 01-.304-.308zm3.11 1.596c-.2.081-.399.151-.59.16a1.245 1.245 0 01-.798-.254c-.274-.23-.47-.358-.552-.758a1.73 1.73 0 01.016-.588c.07-.327-.008-.537-.239-.727-.187-.156-.426-.199-.688-.199a.559.559 0 01-.254-.078c-.11-.054-.2-.19-.114-.358.028-.054.16-.186.192-.21.356-.202.767-.136 1.146.016.352.144.618.408 1.001.782.391.451.462.576.685.914.176.265.336.537.445.848.067.195-.019.354-.25.452z" fill="#4D6BFE"></path></svg>
\ No newline at end of file
diff --git a/packages/app/public/logos/kimi-color.svg b/packages/app/public/logos/kimi-color.svg
new file mode 100644
index 000000000..214eb84f3
--- /dev/null
+++ b/packages/app/public/logos/kimi-color.svg
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Kimi</title><rect width="24" height="24" rx="5.4" fill="#000"></rect><g transform="translate(1.8,2.4) scale(0.78)"><path d="M21.846 0a1.923 1.923 0 110 3.846H20.15a.226.226 0 01-.227-.226V1.923C19.923.861 20.784 0 21.846 0z" fill="#1783FF"></path><path d="M11.065 11.199l7.257-7.2c.137-.136.06-.41-.116-.41H14.3a.164.164 0 00-.117.051l-7.82 7.756c-.122.12-.302.013-.302-.179V3.82c0-.127-.083-.23-.185-.23H3.186c-.103 0-.186.103-.186.23V19.77c0 .128.083.23.186.23h2.69c.103 0 .186-.102.186-.23v-3.25c0-.069.025-.135.069-.178l2.424-2.406a.158.158 0 01.205-.023l6.484 4.772a7.677 7.677 0 003.453 1.283c.108.012.2-.095.2-.23v-3.06c0-.117-.07-.212-.164-.227a5.028 5.028 0 01-2.027-.807l-5.613-4.064c-.117-.078-.132-.279-.028-.381z" fill="#fff"></path></g></svg>
diff --git a/packages/app/public/logos/kvcache-ai.svg b/packages/app/public/logos/kvcache-ai.svg
new file mode 100644
index 000000000..e8e3bb33b
--- /dev/null
+++ b/packages/app/public/logos/kvcache-ai.svg
@@ -0,0 +1,5 @@
+<!-- Vectorized from the official KVCache.AI mark at https://kvcache.ai/media/icon_1.png. -->
+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 240 240" role="img" aria-label="KVCache.AI">
+  <path d="M0 0h68v44L0 108V0Zm0 132 68 64v44H0V132Z" />
+  <path fill="#9d0eb3" d="M157 0h83L115 120l125 120h-83L31 120 157 0Z" />
+</svg>
diff --git a/packages/app/public/logos/meta-color.svg b/packages/app/public/logos/meta-color.svg
new file mode 100644
index 000000000..a8a8d9485
--- /dev/null
+++ b/packages/app/public/logos/meta-color.svg
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Meta</title><path d="M6.897 4h-.024l-.031 2.615h.022c1.715 0 3.046 1.357 5.94 6.246l.175.297.012.02 1.62-2.438-.012-.019a48.763 48.763 0 00-1.098-1.716 28.01 28.01 0 00-1.175-1.629C10.413 4.932 8.812 4 6.896 4z" fill="url(#lobe-icons-meta-0-_R_0_)"></path><path d="M6.873 4C4.95 4.01 3.247 5.258 2.02 7.17a4.352 4.352 0 00-.01.017l2.254 1.231.011-.017c.718-1.083 1.61-1.774 2.568-1.785h.021L6.896 4h-.023z" fill="url(#lobe-icons-meta-1-_R_0_)"></path><path d="M2.019 7.17l-.011.017C1.2 8.447.598 9.995.274 11.664l-.005.022 2.534.6.004-.022c.27-1.467.786-2.828 1.456-3.845l.011-.017L2.02 7.17z" fill="url(#lobe-icons-meta-2-_R_0_)"></path><path d="M2.807 12.264l-2.533-.6-.005.022c-.177.918-.267 1.851-.269 2.786v.023l2.598.233v-.023a12.591 12.591 0 01.21-2.44z" fill="url(#lobe-icons-meta-3-_R_0_)"></path><path d="M2.677 15.537a5.462 5.462 0 01-.079-.813v-.022L0 14.468v.024a8.89 8.89 0 00.146 1.652l2.535-.585a4.106 4.106 0 01-.004-.022z" fill="url(#lobe-icons-meta-4-_R_0_)"></path><path d="M3.27 16.89c-.284-.31-.484-.756-.589-1.328l-.004-.021-2.535.585.004.021c.192 1.01.568 1.85 1.106 2.487l.014.017 2.018-1.745a2.106 2.106 0 01-.015-.016z" fill="url(#lobe-icons-meta-5-_R_0_)"></path><path d="M10.78 9.654c-1.528 2.35-2.454 3.825-2.454 3.825-2.035 3.2-2.739 3.917-3.871 3.917a1.545 1.545 0 01-1.186-.508l-2.017 1.744.014.017C2.01 19.518 3.058 20 4.356 20c1.963 0 3.374-.928 5.884-5.33l1.766-3.13a41.283 41.283 0 00-1.227-1.886z" fill="#0082FB"></path><path d="M13.502 5.946l-.016.016c-.4.43-.786.908-1.16 1.416.378.483.768 1.024 1.175 1.63.48-.743.928-1.345 1.367-1.807l.016-.016-1.382-1.24z" fill="url(#lobe-icons-meta-6-_R_0_)"></path><path d="M20.918 5.713C19.853 4.633 18.583 4 17.225 4c-1.432 0-2.637.787-3.723 1.944l-.016.016 1.382 1.24.016-.017c.715-.747 1.408-1.12 2.176-1.12.826 0 1.6.39 2.27 1.075l.015.016 1.589-1.425-.016-.016z" fill="#0082FB"></path><path d="M23.998 14.125c-.06-3.467-1.27-6.566-3.064-8.396l-.016-.016-1.588 1.424.015.016c1.35 1.392 2.277 3.98 2.361 6.971v.023h2.292v-.022z" fill="url(#lobe-icons-meta-7-_R_0_)"></path><path d="M23.998 14.15v-.023h-2.292v.022c.004.14.006.282.006.424 0 .815-.121 1.474-.368 1.95l-.011.022 1.708 1.782.013-.02c.62-.96.946-2.293.946-3.91 0-.083 0-.165-.002-.247z" fill="url(#lobe-icons-meta-8-_R_0_)"></path><path d="M21.344 16.52l-.011.02c-.214.402-.519.67-.917.787l.778 2.462a3.493 3.493 0 00.438-.182 3.558 3.558 0 001.366-1.218l.044-.065.012-.02-1.71-1.784z" fill="url(#lobe-icons-meta-9-_R_0_)"></path><path d="M19.92 17.393c-.262 0-.492-.039-.718-.14l-.798 2.522c.449.153.927.222 1.46.222.492 0 .943-.073 1.352-.215l-.78-2.462c-.167.05-.341.075-.517.073z" fill="url(#lobe-icons-meta-10-_R_0_)"></path><path d="M18.323 16.534l-.014-.017-1.836 1.914.016.017c.637.682 1.246 1.105 1.937 1.337l.797-2.52c-.291-.125-.573-.353-.9-.731z" fill="url(#lobe-icons-meta-11-_R_0_)"></path><path d="M18.309 16.515c-.55-.642-1.232-1.712-2.303-3.44l-1.396-2.336-.011-.02-1.62 2.438.012.02.989 1.668c.959 1.61 1.74 2.774 2.493 3.585l.016.016 1.834-1.914a2.353 2.353 0 01-.014-.017z" fill="url(#lobe-icons-meta-12-_R_0_)"></path><defs><linearGradient id="lobe-icons-meta-0-_R_0_" x1="75.897%" x2="26.312%" y1="89.199%" y2="12.194%"><stop offset=".06%" stop-color="#0867DF"></stop><stop offset="45.39%" stop-color="#0668E1"></stop><stop offset="85.91%" stop-color="#0064E0"></stop></linearGradient><linearGradient id="lobe-icons-meta-1-_R_0_" x1="21.67%" x2="97.068%" y1="75.874%" y2="23.985%"><stop offset="13.23%" stop-color="#0064DF"></stop><stop offset="99.88%" stop-color="#0064E0"></stop></linearGradient><linearGradient id="lobe-icons-meta-2-_R_0_" x1="38.263%" x2="60.895%" y1="89.127%" y2="16.131%"><stop offset="1.47%" stop-color="#0072EC"></stop><stop offset="68.81%" stop-color="#0064DF"></stop></linearGradient><linearGradient id="lobe-icons-meta-3-_R_0_" x1="47.032%" x2="52.15%" y1="90.19%" y2="15.745%"><stop offset="7.31%" stop-color="#007CF6"></stop><stop offset="99.43%" stop-color="#0072EC"></stop></linearGradient><linearGradient id="lobe-icons-meta-4-_R_0_" x1="52.155%" x2="47.591%" y1="58.301%" y2="37.004%"><stop offset="7.31%" stop-color="#007FF9"></stop><stop offset="100%" stop-color="#007CF6"></stop></linearGradient><linearGradient id="lobe-icons-meta-5-_R_0_" x1="37.689%" x2="61.961%" y1="12.502%" y2="63.624%"><stop offset="7.31%" stop-color="#007FF9"></stop><stop offset="100%" stop-color="#0082FB"></stop></linearGradient><linearGradient id="lobe-icons-meta-6-_R_0_" x1="34.808%" x2="62.313%" y1="68.859%" y2="23.174%"><stop offset="27.99%" stop-color="#007FF8"></stop><stop offset="91.41%" stop-color="#0082FB"></stop></linearGradient><linearGradient id="lobe-icons-meta-7-_R_0_" x1="43.762%" x2="57.602%" y1="6.235%" y2="98.514%"><stop offset="0%" stop-color="#0082FB"></stop><stop offset="99.95%" stop-color="#0081FA"></stop></linearGradient><linearGradient id="lobe-icons-meta-8-_R_0_" x1="60.055%" x2="39.88%" y1="4.661%" y2="69.077%"><stop offset="6.19%" stop-color="#0081FA"></stop><stop offset="100%" stop-color="#0080F9"></stop></linearGradient><linearGradient id="lobe-icons-meta-9-_R_0_" x1="30.282%" x2="61.081%" y1="59.32%" y2="33.244%"><stop offset="0%" stop-color="#027AF3"></stop><stop offset="100%" stop-color="#0080F9"></stop></linearGradient><linearGradient id="lobe-icons-meta-10-_R_0_" x1="20.433%" x2="82.112%" y1="50.001%" y2="50.001%"><stop offset="0%" stop-color="#0377EF"></stop><stop offset="99.94%" stop-color="#0279F1"></stop></linearGradient><linearGradient id="lobe-icons-meta-11-_R_0_" x1="40.303%" x2="72.394%" y1="35.298%" y2="57.811%"><stop offset=".19%" stop-color="#0471E9"></stop><stop offset="100%" stop-color="#0377EF"></stop></linearGradient><linearGradient id="lobe-icons-meta-12-_R_0_" x1="32.254%" x2="68.003%" y1="19.719%" y2="84.908%"><stop offset="27.65%" stop-color="#0867DF"></stop><stop offset="100%" stop-color="#0471E9"></stop></linearGradient></defs></svg>
\ No newline at end of file
diff --git a/packages/app/public/logos/minimax-color.svg b/packages/app/public/logos/minimax-color.svg
new file mode 100644
index 000000000..beb7adb86
--- /dev/null
+++ b/packages/app/public/logos/minimax-color.svg
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Minimax</title><defs><linearGradient id="lobe-icons-minimax-_R_0_" x1="0%" x2="100.182%" y1="50.057%" y2="50.057%"><stop offset="0%" stop-color="#E2167E"></stop><stop offset="100%" stop-color="#FE603C"></stop></linearGradient></defs><path d="M16.278 2c1.156 0 2.093.927 2.093 2.07v12.501a.74.74 0 00.744.709.74.74 0 00.743-.709V9.099a2.06 2.06 0 012.071-2.049A2.06 2.06 0 0124 9.1v6.561a.649.649 0 01-.652.645.649.649 0 01-.653-.645V9.1a.762.762 0 00-.766-.758.762.762 0 00-.766.758v7.472a2.037 2.037 0 01-2.048 2.026 2.037 2.037 0 01-2.048-2.026v-12.5a.785.785 0 00-.788-.753.785.785 0 00-.789.752l-.001 15.904A2.037 2.037 0 0113.441 22a2.037 2.037 0 01-2.048-2.026V18.04c0-.356.292-.645.652-.645.36 0 .652.289.652.645v1.934c0 .263.142.506.372.638.23.131.514.131.744 0a.734.734 0 00.372-.638V4.07c0-1.143.937-2.07 2.093-2.07zm-5.674 0c1.156 0 2.093.927 2.093 2.07v11.523a.648.648 0 01-.652.645.648.648 0 01-.652-.645V4.07a.785.785 0 00-.789-.78.785.785 0 00-.789.78v14.013a2.06 2.06 0 01-2.07 2.048 2.06 2.06 0 01-2.071-2.048V9.1a.762.762 0 00-.766-.758.762.762 0 00-.766.758v3.8a2.06 2.06 0 01-2.071 2.049A2.06 2.06 0 010 12.9v-1.378c0-.357.292-.646.652-.646.36 0 .653.29.653.646V12.9c0 .418.343.757.766.757s.766-.339.766-.757V9.099a2.06 2.06 0 012.07-2.048 2.06 2.06 0 012.071 2.048v8.984c0 .419.343.758.767.758.423 0 .766-.339.766-.758V4.07c0-1.143.937-2.07 2.093-2.07z" fill="url(#lobe-icons-minimax-_R_0_)" fill-rule="nonzero"></path></svg>
\ No newline at end of file
diff --git a/packages/app/public/logos/nvidia-color.svg b/packages/app/public/logos/nvidia-color.svg
new file mode 100644
index 000000000..f711e8830
--- /dev/null
+++ b/packages/app/public/logos/nvidia-color.svg
@@ -0,0 +1 @@
+<svg fill="#76B900" role="img" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><title>NVIDIA</title><path d="M8.948 8.798v-1.43a6.7 6.7 0 0 1 .424-.018c3.922-.124 6.493 3.374 6.493 3.374s-2.774 3.851-5.75 3.851c-.398 0-.787-.062-1.158-.185v-4.346c1.528.185 1.837.857 2.747 2.385l2.04-1.714s-1.492-1.952-4-1.952a6.016 6.016 0 0 0-.796.035m0-4.735v2.138l.424-.027c5.45-.185 9.01 4.47 9.01 4.47s-4.08 4.964-8.33 4.964c-.37 0-.733-.035-1.095-.097v1.325c.3.035.61.062.91.062 3.957 0 6.82-2.023 9.593-4.408.459.371 2.34 1.263 2.73 1.652-2.633 2.208-8.772 3.984-12.253 3.984-.335 0-.653-.018-.971-.053v1.864H24V4.063zm0 10.326v1.131c-3.657-.654-4.673-4.46-4.673-4.46s1.758-1.944 4.673-2.262v1.237H8.94c-1.528-.186-2.73 1.245-2.73 1.245s.68 2.412 2.739 3.11M2.456 10.9s2.164-3.197 6.5-3.533V6.201C4.153 6.59 0 10.653 0 10.653s2.35 6.802 8.948 7.42v-1.237c-4.84-.6-6.492-5.936-6.492-5.936z"/></svg>
diff --git a/packages/app/public/logos/qwen-color.svg b/packages/app/public/logos/qwen-color.svg
new file mode 100644
index 000000000..f2d0ada16
--- /dev/null
+++ b/packages/app/public/logos/qwen-color.svg
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Qwen</title><path d="M12.604 1.34c.393.69.784 1.382 1.174 2.075a.18.18 0 00.157.091h5.552c.174 0 .322.11.446.327l1.454 2.57c.19.337.24.478.024.837-.26.43-.513.864-.76 1.3l-.367.658c-.106.196-.223.28-.04.512l2.652 4.637c.172.301.111.494-.043.77-.437.785-.882 1.564-1.335 2.34-.159.272-.352.375-.68.37-.777-.016-1.552-.01-2.327.016a.099.099 0 00-.081.05 575.097 575.097 0 01-2.705 4.74c-.169.293-.38.363-.725.364-.997.003-2.002.004-3.017.002a.537.537 0 01-.465-.271l-1.335-2.323a.09.09 0 00-.083-.049H4.982c-.285.03-.553-.001-.805-.092l-1.603-2.77a.543.543 0 01-.002-.54l1.207-2.12a.198.198 0 000-.197 550.951 550.951 0 01-1.875-3.272l-.79-1.395c-.16-.31-.173-.496.095-.965.465-.813.927-1.625 1.387-2.436.132-.234.304-.334.584-.335a338.3 338.3 0 012.589-.001.124.124 0 00.107-.063l2.806-4.895a.488.488 0 01.422-.246c.524-.001 1.053 0 1.583-.006L11.704 1c.341-.003.724.032.9.34zm-3.432.403a.06.06 0 00-.052.03L6.254 6.788a.157.157 0 01-.135.078H3.253c-.056 0-.07.025-.041.074l5.81 10.156c.025.042.013.062-.034.063l-2.795.015a.218.218 0 00-.2.116l-1.32 2.31c-.044.078-.021.118.068.118l5.716.008c.046 0 .08.02.104.061l1.403 2.454c.046.081.092.082.139 0l5.006-8.76.783-1.382a.055.055 0 01.096 0l1.424 2.53a.122.122 0 00.107.062l2.763-.02a.04.04 0 00.035-.02.041.041 0 000-.04l-2.9-5.086a.108.108 0 010-.113l.293-.507 1.12-1.977c.024-.041.012-.062-.035-.062H9.2c-.059 0-.073-.026-.043-.077l1.434-2.505a.107.107 0 000-.114L9.225 1.774a.06.06 0 00-.053-.031zm6.29 8.02c.046 0 .058.02.034.06l-.832 1.465-2.613 4.585a.056.056 0 01-.05.029.058.058 0 01-.05-.029L8.498 9.841c-.02-.034-.01-.052.028-.054l.216-.012 6.722-.012z" fill="url(#lobe-icons-qwen-_R_0_)" fill-rule="nonzero"></path><defs><linearGradient id="lobe-icons-qwen-_R_0_" x1="0%" x2="100%" y1="0%" y2="0%"><stop offset="0%" stop-color="#6336E7" stop-opacity=".84"></stop><stop offset="100%" stop-color="#6F69F7" stop-opacity=".84"></stop></linearGradient></defs></svg>
\ No newline at end of file
diff --git a/packages/app/public/logos/zai-color.svg b/packages/app/public/logos/zai-color.svg
new file mode 100644
index 000000000..d46c329da
--- /dev/null
+++ b/packages/app/public/logos/zai-color.svg
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Z.ai</title><rect width="24" height="24" rx="5.4" fill="#0A0A0A"></rect><g transform="translate(4.2 4.2) scale(0.65)"><path d="M12.105 2L9.927 4.953H.653L2.83 2h9.276zM23.254 19.048L21.078 22h-9.242l2.174-2.952h9.244zM24 2L9.264 22H0L14.736 2H24z" fill="#fff" fill-rule="evenodd"></path></g></svg>
diff --git a/packages/app/public/promo-poster.webp b/packages/app/public/promo-poster.webp
new file mode 100644
index 000000000..108e035c7
Binary files /dev/null and b/packages/app/public/promo-poster.webp differ
diff --git a/packages/app/scripts/capture-cypress-fixtures.ts b/packages/app/scripts/capture-cypress-fixtures.ts
index e30c042d3..0b1e61595 100644
--- a/packages/app/scripts/capture-cypress-fixtures.ts
+++ b/packages/app/scripts/capture-cypress-fixtures.ts
@@ -11,7 +11,7 @@
  *   bun run --cwd packages/app capture:fixtures -- --collectivex-only         (synthetic multi-run data)
  */
 
-import { mkdir, writeFile } from 'node:fs/promises';
+import { mkdir, readdir, readFile, writeFile } from 'node:fs/promises';
 import { resolve } from 'node:path';
 
 import { buildRunSummary } from '@semianalysisai/inferencex-db/collectivex/reader';
@@ -20,6 +20,16 @@ import {
   makeCollectiveXDataset,
   makeRawShard,
 } from '@semianalysisai/inferencex-db/collectivex/test-fixture';
+import type { BenchmarkRow } from '@semianalysisai/inferencex-db/queries/benchmarks';
+import {
+  FIXTURE_MANIFEST_FILENAME,
+  FIXTURE_MANIFEST_SCHEMA_VERSION,
+  type FixtureManifest,
+  type FixtureManifestEntry,
+  assertFixtureContent,
+  fixtureSha256,
+  fixtureTopLevel,
+} from '../src/lib/test-fixture-manifest';
 
 const cliArgs = process.argv.filter((argument) => argument !== '--').slice(2);
 const collectiveXOnly = cliArgs.includes('--collectivex-only');
@@ -139,6 +149,44 @@ async function writeFixture(name: string, data: unknown): Promise<number> {
   await writeFile(resolve(fixturesDir, `${name}.json`), body);
   return body.length;
 }
+async function writeFixtureManifest(updatedNames: Set<string>, source: string): Promise<void> {
+  const manifestPath = resolve(fixturesDir, FIXTURE_MANIFEST_FILENAME);
+  let previous: FixtureManifest | null = null;
+  try {
+    previous = JSON.parse(await readFile(manifestPath, 'utf8')) as FixtureManifest;
+  } catch (error) {
+    if ((error as NodeJS.ErrnoException).code !== 'ENOENT') throw error;
+  }
+
+  const capturedAt = new Date().toISOString();
+  const directoryEntries = await readdir(fixturesDir);
+  const fixtureFiles = directoryEntries
+    .filter((name) => name.endsWith('.json') && name !== FIXTURE_MANIFEST_FILENAME)
+    .toSorted();
+  const fixtures: Record<string, FixtureManifestEntry> = {};
+  for (const filename of fixtureFiles) {
+    const name = filename.slice(0, -'.json'.length);
+    const body = await readFile(resolve(fixturesDir, filename), 'utf8');
+    const value = JSON.parse(body) as unknown;
+    assertFixtureContent(name, value);
+    const oldEntry = previous?.fixtures[name];
+    const wasUpdated = updatedNames.has(name);
+    fixtures[name] = {
+      bytes: Buffer.byteLength(body),
+      capturedAt: wasUpdated ? capturedAt : (oldEntry?.capturedAt ?? null),
+      sha256: fixtureSha256(body),
+      source: wasUpdated ? source : (oldEntry?.source ?? null),
+      topLevel: fixtureTopLevel(value),
+    };
+  }
+
+  const manifest: FixtureManifest = {
+    schemaVersion: FIXTURE_MANIFEST_SCHEMA_VERSION,
+    generatedAt: capturedAt,
+    fixtures,
+  };
+  await writeFile(manifestPath, `${JSON.stringify(manifest, null, 2)}\n`);
+}
 
 async function writeCollectiveXFixtures(): Promise<[string, number][]> {
   const latest = makeCollectiveXDataset();
@@ -182,7 +230,9 @@ async function main() {
   await mkdir(fixturesDir, { recursive: true });
   if (collectiveXOnly) {
     console.log('Generating synthetic CollectiveX fixtures');
-    printSizes(await writeCollectiveXFixtures());
+    const sizes = await writeCollectiveXFixtures();
+    await writeFixtureManifest(new Set(sizes.map(([name]) => name)), 'synthetic://collectivex');
+    printSizes(sizes);
     return;
   }
 
@@ -200,14 +250,6 @@ async function main() {
   // Latest-snapshot: already deduped to one row per config, no date filter.
   // ~20 conc levels per (hw, fw, prec, isl, osl) — sample down to keep the
   // scatter visually populated without writing every concurrency point.
-  interface BenchmarkRow {
-    conc: number;
-    hardware: string;
-    framework: string;
-    precision: string;
-    isl: number;
-    osl: number;
-  }
   const benchmarks = await fetchJson<BenchmarkRow[]>(
     `/api/v1/benchmarks?model=${encodeURIComponent(BENCHMARK_MODEL)}`,
   );
@@ -309,6 +351,12 @@ async function main() {
     ...collectiveXSizes,
   ];
 
+  const collectiveXNames = new Set(collectiveXSizes.map(([name]) => name));
+  await writeFixtureManifest(
+    new Set(sizes.map(([name]) => name).filter((name) => !collectiveXNames.has(name))),
+    baseUrl,
+  );
+  await writeFixtureManifest(collectiveXNames, 'synthetic://collectivex');
   printSizes(sizes);
 }
 
diff --git a/packages/app/scripts/check-typography.ts b/packages/app/scripts/check-typography.ts
new file mode 100644
index 000000000..3722881f7
--- /dev/null
+++ b/packages/app/scripts/check-typography.ts
@@ -0,0 +1,96 @@
+#!/usr/bin/env bun
+/**
+ * Typography ratchet gate — CLI.
+ *
+ *   bun run check:typography            # verify (CI + lefthook)
+ *   bun run check:typography --update   # shrink the allowlist after fixes
+ *
+ * Scans packages/app/src for arbitrary font sizes (text-[11px]), arbitrary
+ * letter-spacing (tracking-[0.16em]), and quoted font-size literals in the
+ * shared D3 chart library, then compares per-file counts against
+ * scripts/typography-allowlist.json. New offenders fail immediately; the
+ * allowlisted ones are the migration burndown and may only ever decrease.
+ * Rules and exemptions live in src/lib/typography-gate.ts; docs/typography.md
+ * explains the token system these rules point at.
+ */
+import { readdirSync, readFileSync, writeFileSync } from 'node:fs';
+import path from 'node:path';
+
+import {
+  compareToAllowlist,
+  countViolations,
+  scanSource,
+  type ViolationCounts,
+} from '../src/lib/typography-gate';
+
+const APP_ROOT = path.resolve(import.meta.dirname, '..');
+const ALLOWLIST_PATH = path.join(import.meta.dirname, 'typography-allowlist.json');
+const UPDATE = process.argv.includes('--update');
+
+function collectSourceFiles(dir: string, out: string[] = []): string[] {
+  for (const entry of readdirSync(dir, { withFileTypes: true })) {
+    const full = path.join(dir, entry.name);
+    if (entry.isDirectory()) {
+      collectSourceFiles(full, out);
+    } else if (/\.tsx?$/u.test(entry.name)) {
+      out.push(full);
+    }
+  }
+  return out;
+}
+
+const files = collectSourceFiles(path.join(APP_ROOT, 'src'))
+  .map((absPath) => ({
+    relPath: path.relative(APP_ROOT, absPath).replaceAll(path.sep, '/'),
+    source: readFileSync(absPath, 'utf8'),
+  }))
+  .sort((a, b) => a.relPath.localeCompare(b.relPath));
+
+const counts = countViolations(files);
+
+if (UPDATE) {
+  const sorted: ViolationCounts = Object.fromEntries(
+    Object.keys(counts)
+      .sort()
+      .map((file) => [file, counts[file]]),
+  );
+  writeFileSync(ALLOWLIST_PATH, `${JSON.stringify(sorted, null, 2)}\n`);
+  const total = Object.values(sorted)
+    .flatMap((rules) => Object.values(rules))
+    .reduce((sum, n) => sum + n, 0);
+  console.log(
+    `check-typography: allowlist rewritten with ${Object.keys(sorted).length} file(s), ${total} violation(s) remaining.`,
+  );
+  process.exit(0);
+}
+
+const allowlist: ViolationCounts = JSON.parse(readFileSync(ALLOWLIST_PATH, 'utf8'));
+const { errors, stale } = compareToAllowlist(counts, allowlist);
+
+if (errors.length > 0) {
+  console.error('check-typography: new typography violations (see docs/typography.md):\n');
+  for (const error of errors) {
+    console.error(`  ${error}`);
+    const [file] = error.split(':', 1);
+    const entry = files.find((candidate) => candidate.relPath === file);
+    if (!entry) continue;
+    for (const violation of scanSource(entry.relPath, entry.source)) {
+      console.error(
+        `    ${file}:${violation.line}  ${violation.match}  →  ${violation.suggestion}`,
+      );
+    }
+  }
+}
+
+if (stale.length > 0) {
+  console.error(
+    '\ncheck-typography: violations were fixed — ratchet down by running:\n  bun run check:typography --update\n',
+  );
+  for (const line of stale) console.error(`  ${line}`);
+}
+
+if (errors.length > 0 || stale.length > 0) {
+  process.exit(1);
+}
+
+console.log('check-typography: OK');
diff --git a/packages/app/scripts/typography-allowlist.json b/packages/app/scripts/typography-allowlist.json
new file mode 100644
index 000000000..bd3b12185
--- /dev/null
+++ b/packages/app/scripts/typography-allowlist.json
@@ -0,0 +1,26 @@
+{
+  "src/app/land-acknowledgement/page.tsx": {
+    "arbitrary-tracking": 2
+  },
+  "src/app/zh/land-acknowledgement/page.tsx": {
+    "arbitrary-tracking": 2
+  },
+  "src/components/compare/agentx-compare-hero.tsx": {
+    "arbitrary-text-size": 1
+  },
+  "src/components/evaluation/ui/EvaluationTable.tsx": {
+    "arbitrary-text-size": 1
+  },
+  "src/components/glossary/glossary-browser.tsx": {
+    "arbitrary-tracking": 1
+  },
+  "src/components/inference/agentic-point/request-timeline.tsx": {
+    "arbitrary-text-size": 2
+  },
+  "src/components/inference/agentic-point/timeline-tooltips.tsx": {
+    "arbitrary-text-size": 1
+  },
+  "src/lib/d3-chart/layers/perf-ruler.ts": {
+    "chart-font-literal": 1
+  }
+}
diff --git a/packages/app/src/app/(dashboard)/calculator/[model]/page.tsx b/packages/app/src/app/(dashboard)/calculator/[model]/page.tsx
new file mode 100644
index 000000000..7675cafc9
--- /dev/null
+++ b/packages/app/src/app/(dashboard)/calculator/[model]/page.tsx
@@ -0,0 +1,44 @@
+import type { Metadata } from 'next';
+import { notFound, permanentRedirect } from 'next/navigation';
+
+import ThroughputCalculatorDisplay from '@/components/calculator/ThroughputCalculatorDisplay';
+import { resolveCalculatorUrlSeed } from '@/components/calculator/url-seed';
+import { modelRoutePath, pathWithSearchParams, resolveModelRouteSlug } from '@/lib/model-routes';
+import { modelTabMetadata } from '@/lib/tab-meta';
+
+/**
+ * `/calculator/<model>` — the Throughput & TCO calculator seeded to a specific
+ * model, giving every model an indexable URL. Switching models in the UI
+ * rewrites the pathname in place via `history.replaceState` (no reload, no
+ * remount). No `generateStaticParams`: like the bare /calculator route this
+ * page reads searchParams, so it renders dynamically anyway.
+ */
+
+interface Props {
+  params: Promise<{ model: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { model } = await params;
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) return {};
+  return modelTabMetadata('calculator', resolved.route);
+}
+
+export default async function CalculatorModelPage({ params, searchParams }: Props) {
+  const [{ model }, sp] = await Promise.all([params, searchParams]);
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) notFound();
+  // Aliases and non-canonical casing 308 to the canonical slug, keeping any
+  // share-link params, mirroring /compare/[slug].
+  if (resolved.isAlias) {
+    permanentRedirect(pathWithSearchParams(modelRoutePath('calculator', resolved.route.slug), sp));
+  }
+  const seed = resolveCalculatorUrlSeed(sp);
+  // A legacy explicit ?g_model= wins over the path — same precedence as
+  // GlobalFilterProvider, which then canonicalizes the pathname client-side.
+  return (
+    <ThroughputCalculatorDisplay urlSeed={{ ...seed, model: seed.model ?? resolved.route.model }} />
+  );
+}
diff --git a/packages/app/src/app/(dashboard)/current-inferencex-image/page.tsx b/packages/app/src/app/(dashboard)/current-inferencex-image/page.tsx
index 00c9f32df..d573d5fba 100644
--- a/packages/app/src/app/(dashboard)/current-inferencex-image/page.tsx
+++ b/packages/app/src/app/(dashboard)/current-inferencex-image/page.tsx
@@ -1,19 +1,7 @@
-import type { Metadata } from 'next';
-
-import { SITE_URL } from '@semianalysisai/inferencex-constants';
 import { CurrentImageContent } from '@/components/latest-image/latest-image-content';
+import { tabMetadata } from '@/lib/tab-meta';
 
-export const metadata: Metadata = {
-  title: 'Current InferenceX Image',
-  description: 'Current InferenceX Docker image tags for each model, chip SKU, and configuration.',
-  alternates: { canonical: `${SITE_URL}/current-inferencex-image` },
-  openGraph: {
-    title: 'Current InferenceX Image | InferenceX by SemiAnalysis',
-    description:
-      'Current InferenceX Docker image tags for each model, chip SKU, and configuration.',
-    url: `${SITE_URL}/current-inferencex-image`,
-  },
-};
+export const metadata = tabMetadata('current-inferencex-image');
 
 export default function CurrentInferenceXImagePage() {
   return <CurrentImageContent />;
diff --git a/packages/app/src/app/(dashboard)/fleet/page.tsx b/packages/app/src/app/(dashboard)/fleet/page.tsx
new file mode 100644
index 000000000..29ff67ef9
--- /dev/null
+++ b/packages/app/src/app/(dashboard)/fleet/page.tsx
@@ -0,0 +1,17 @@
+import type { Metadata } from 'next';
+
+import FleetLifecycleDisplay from '@/components/calculator/FleetLifecycleDisplay';
+import { resolveCalculatorUrlSeed } from '@/components/calculator/url-seed';
+import { tabMetadata } from '@/lib/tab-meta';
+
+export const metadata: Metadata = tabMetadata('fleet');
+
+interface Props {
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export default async function FleetPage({ searchParams }: Props) {
+  const sp = await searchParams;
+  const seed = resolveCalculatorUrlSeed(sp);
+  return <FleetLifecycleDisplay urlSeed={seed} />;
+}
diff --git a/packages/app/src/app/(dashboard)/historical/[model]/page.tsx b/packages/app/src/app/(dashboard)/historical/[model]/page.tsx
new file mode 100644
index 000000000..0991b851b
--- /dev/null
+++ b/packages/app/src/app/(dashboard)/historical/[model]/page.tsx
@@ -0,0 +1,55 @@
+import type { Metadata } from 'next';
+import { notFound, permanentRedirect } from 'next/navigation';
+
+import { InferenceProvider } from '@/components/inference/InferenceContext';
+import HistoricalTrendsDisplay from '@/components/trends/HistoricalTrendsDisplay';
+import {
+  MODEL_ROUTES,
+  modelRoutePath,
+  pathWithSearchParams,
+  resolveModelRouteSlug,
+} from '@/lib/model-routes';
+import { modelTabMetadata } from '@/lib/tab-meta';
+
+/**
+ * `/historical/<model>` — the Historical Trends tab seeded to a specific
+ * model, giving every model an indexable URL. `GlobalFilterProvider` (mounted
+ * by the dashboard shell) reads the model straight from the pathname, so the
+ * page body is identical to the bare tab; switching models in the UI rewrites
+ * the pathname in place via `history.replaceState` (no reload, no remount).
+ */
+
+interface Props {
+  params: Promise<{ model: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export function generateStaticParams() {
+  return MODEL_ROUTES.map((route) => ({ model: route.slug }));
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { model } = await params;
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) return {};
+  return modelTabMetadata('historical', resolved.route);
+}
+
+export default async function HistoricalModelPage({ params, searchParams }: Props) {
+  const { model } = await params;
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) notFound();
+  // Aliases and non-canonical casing 308 to the canonical slug, keeping any
+  // share-link params, mirroring /compare/[slug]. `searchParams` is awaited
+  // only on this branch so the canonical slugs stay statically generated.
+  if (resolved.isAlias) {
+    permanentRedirect(
+      pathWithSearchParams(modelRoutePath('historical', resolved.route.slug), await searchParams),
+    );
+  }
+  return (
+    <InferenceProvider activeTab="historical">
+      <HistoricalTrendsDisplay />
+    </InferenceProvider>
+  );
+}
diff --git a/packages/app/src/app/(dashboard)/inference/[model]/page.tsx b/packages/app/src/app/(dashboard)/inference/[model]/page.tsx
new file mode 100644
index 000000000..02dce4651
--- /dev/null
+++ b/packages/app/src/app/(dashboard)/inference/[model]/page.tsx
@@ -0,0 +1,86 @@
+import type { Metadata } from 'next';
+import { notFound, permanentRedirect } from 'next/navigation';
+
+import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { InferenceProvider } from '@/components/inference/InferenceContext';
+import InferenceChartDisplay from '@/components/inference/ui/ChartDisplay';
+import { enAlternates } from '@/lib/i18n';
+import { inferenceModelMeta } from '@/lib/inference-model-meta';
+import {
+  getInferenceModelBySlug,
+  INFERENCE_MODEL_SLUGS,
+  inferenceModelPath,
+} from '@/lib/inference-model-slug';
+
+/**
+ * `/inference/<model>` — an indexable path form of `/inference?g_model=<model>`.
+ *
+ * The query form stays supported for share links, but this page gives every
+ * model its own crawlable URL with model-specific metadata, a self-canonical,
+ * and a `/zh` hreflang sibling. The chart body is identical to `/inference`;
+ * the model is pinned from the pathname by the dashboard shell
+ * (`GlobalFilterProvider` seeding via `inferenceModelForPathname`), and an
+ * explicit `?g_model=` param still wins over the path pin.
+ */
+
+interface Props {
+  params: Promise<{ model: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export function generateStaticParams(): { model: string }[] {
+  return INFERENCE_MODEL_SLUGS.map((entry) => ({ model: entry.slug }));
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { model } = await params;
+  const entry = getInferenceModelBySlug(model);
+  if (!entry) return {};
+  const { title, description } = inferenceModelMeta(entry);
+  const enPath = inferenceModelPath(entry.slug);
+  const url = `${SITE_URL}${enPath}`;
+  return {
+    title: { absolute: `${title} | ${SITE_NAME} by ${AUTHOR_NAME}` },
+    description,
+    alternates: enAlternates(enPath),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+      type: 'website',
+    },
+    twitter: {
+      card: 'summary_large_image',
+      title: `${title} | ${SITE_NAME}`,
+      description,
+    },
+  };
+}
+
+export default async function InferenceModelPage({ params, searchParams }: Props) {
+  const { model } = await params;
+  const entry = getInferenceModelBySlug(model);
+  if (!entry) notFound();
+  // Aliases (family names, superseded versions, raw `g_model` display names,
+  // uppercase variants) collapse onto the one canonical URL per model.
+  // Preserves the query string so share-link params like `?i_seq=` and
+  // `?i_prec=` survive the redirect — same treatment as the compare pages.
+  if (model !== entry.slug) {
+    const sp = await searchParams;
+    const qs = Object.entries(sp)
+      .flatMap(([k, v]) => {
+        if (Array.isArray(v)) return v.map((vv) => [k, vv] as const);
+        if (v === undefined) return [];
+        return [[k, v] as const];
+      })
+      .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
+      .join('&');
+    permanentRedirect(`${inferenceModelPath(entry.slug)}${qs ? `?${qs}` : ''}`);
+  }
+  return (
+    <InferenceProvider activeTab="inference">
+      <InferenceChartDisplay />
+    </InferenceProvider>
+  );
+}
diff --git a/packages/app/src/app/(dashboard)/inference/agentic/page.tsx b/packages/app/src/app/(dashboard)/inference/agentic/page.tsx
new file mode 100644
index 000000000..3c6a85d99
--- /dev/null
+++ b/packages/app/src/app/(dashboard)/inference/agentic/page.tsx
@@ -0,0 +1,49 @@
+import type { Metadata } from 'next';
+
+import { AgenticCatalogHero } from '@/components/inference/agentic-catalog/agentic-catalog-hero';
+import { AgenticCatalogList } from '@/components/inference/agentic-catalog/agentic-catalog-list';
+import { JsonLd } from '@/components/json-ld';
+import { getAgenticCatalogGroups } from '@/lib/agentic-catalog';
+import { enAlternates } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const TITLE = 'AgentX Telemetry';
+const DESCRIPTION =
+  'Browse every AgentX benchmark run with stored per-request telemetry, grouped by model and serving stack. Each configuration opens its point-detail charts, request timeline, and cache behavior.';
+
+export const metadata: Metadata = {
+  title: TITLE,
+  description: DESCRIPTION,
+  alternates: enAlternates('/inference/agentic'),
+  openGraph: {
+    title: `${TITLE} | InferenceX`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/inference/agentic`,
+  },
+  twitter: { title: `${TITLE} | InferenceX`, description: DESCRIPTION },
+};
+
+// The catalog reads the same live benchmark tables the dashboard does, so a
+// build-time snapshot would go stale the moment a new run lands.
+export const dynamic = 'force-dynamic';
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'CollectionPage',
+  name: `${TITLE} | InferenceX`,
+  description: DESCRIPTION,
+  url: `${SITE_URL}/inference/agentic`,
+  inLanguage: 'en',
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default async function AgenticTelemetryCatalogPage() {
+  const groups = await getAgenticCatalogGroups();
+  return (
+    <>
+      <JsonLd data={jsonLd} />
+      <AgenticCatalogHero locale="en" />
+      <AgenticCatalogList groups={groups} locale="en" />
+    </>
+  );
+}
diff --git a/packages/app/src/app/(dashboard)/inference/logs/[id]/page.tsx b/packages/app/src/app/(dashboard)/inference/logs/[id]/page.tsx
new file mode 100644
index 000000000..681af75ca
--- /dev/null
+++ b/packages/app/src/app/(dashboard)/inference/logs/[id]/page.tsx
@@ -0,0 +1,17 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { BenchmarkLogDetail } from '@/components/inference/log-viewer/benchmark-log-detail';
+import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
+
+export const metadata: Metadata = {
+  title: 'Benchmark logs | InferenceX',
+  robots: { index: false },
+};
+
+export default async function BenchmarkLogPage({ params }: { params: Promise<{ id: string }> }) {
+  const { id } = await params;
+  const numericId = Number(id);
+  if (!isPersistedBenchmarkId(numericId)) notFound();
+  return <BenchmarkLogDetail id={numericId} />;
+}
diff --git a/packages/app/src/app/about/page.tsx b/packages/app/src/app/about/page.tsx
index 5e9348afe..ae75c761e 100644
--- a/packages/app/src/app/about/page.tsx
+++ b/packages/app/src/app/about/page.tsx
@@ -4,38 +4,29 @@ import Link from 'next/link';
 import { Card } from '@/components/ui/card';
 import { FAQ_ITEMS } from '@/components/about/faq-data';
 import { AgentXFaq } from '@/components/about/agentx-faq';
+import { buildFaqJsonLd, FaqList } from '@/components/about/faq';
+import { PromoVideo } from '@/components/about/promo-video';
 import { JsonLd } from '@/components/json-ld';
 import { enAlternates } from '@/lib/i18n';
 import { GITHUB_OWNER, GITHUB_REPO, SITE_URL } from '@semianalysisai/inferencex-constants';
 
-const faqJsonLd = {
-  '@context': 'https://schema.org',
-  '@type': 'FAQPage',
-  mainEntity: FAQ_ITEMS.map((item) => ({
-    '@type': 'Question',
-    name: item.question,
-    acceptedAnswer: {
-      '@type': 'Answer',
-      text: [item.answer, item.link?.text, ...(item.list ?? [])].filter(Boolean).join(' '),
-    },
-  })),
-};
+const faqJsonLd = buildFaqJsonLd(FAQ_ITEMS);
+
+const ABOUT_DESCRIPTION =
+  'InferenceX benchmarks agentic and fixed-sequence AI inference across accelerators and serving stacks. AgentX is its long-context, multi-turn coding scenario.';
 
 export const metadata: Metadata = {
   title: 'About',
-  description:
-    'InferenceX is an independent, vendor neutral, reproducible benchmark which continuously benchmarks inference software across a wide range of AI accelerators.',
+  description: ABOUT_DESCRIPTION,
   alternates: enAlternates('/about'),
   openGraph: {
     title: 'About | InferenceX',
-    description:
-      'InferenceX is an independent, vendor neutral, reproducible benchmark which continuously benchmarks inference software across a wide range of AI accelerators.',
+    description: ABOUT_DESCRIPTION,
     url: `${SITE_URL}/about`,
   },
   twitter: {
     title: 'About | InferenceX',
-    description:
-      'InferenceX is an independent, vendor neutral, reproducible benchmark which continuously benchmarks inference software across a wide range of AI accelerators.',
+    description: ABOUT_DESCRIPTION,
   },
 };
 
@@ -44,11 +35,15 @@ export default function AboutPage() {
     <main className="relative">
       <JsonLd data={faqJsonLd} />
       <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
+        <section>
+          <PromoVideo />
+        </section>
+
         <section>
           <Card>
             <h2 className="text-lg font-semibold mb-2">
-              Open Source Continuous Inference Benchmark trusted by Operators of Trillion Dollar
-              GigaWatt Scale Token Factories
+              Open-Source Continuous Agentic Inference Benchmark Trusted by Operators of
+              Trillion-Dollar, GigaWatt-Scale Token Factories
             </h2>
             <p className="text-muted-foreground mb-2">
               As the world progresses exponentially towards AGI, software development and model
@@ -58,9 +53,9 @@ export default function AboutPage() {
             </p>
             <p className="text-muted-foreground mb-2">
               <strong>InferenceX&trade;</strong> (formerly InferenceMAX) is our independent, vendor
-              neutral, reproducible benchmark which addresses these issues by continuously
-              benchmarking inference software across a wide range of AI accelerators that are
-              actually available to the ML community.
+              neutral, reproducible benchmark. It measures both fixed-sequence serving and AgentX,
+              our long-context, multi-turn agentic coding workload, across AI accelerators and
+              serving stacks available to the ML community.
             </p>
             <p className="text-muted-foreground">
               Our open data & insights are widely adopted by the ML community, capacity planning
@@ -176,45 +171,7 @@ export default function AboutPage() {
           </Card>
         </section>
 
-        <section>
-          <Card>
-            <h2 className="text-lg font-semibold mb-4">Frequently Asked Questions</h2>
-            <dl className="divide-y divide-border">
-              {FAQ_ITEMS.map((item) => (
-                <div key={item.question} className="py-4 first:pt-0 last:pb-0">
-                  <dt className="font-medium mb-1">{item.question}</dt>
-                  <dd className="text-muted-foreground text-sm">
-                    {item.answer && (
-                      <p>
-                        {item.answer}
-                        {item.link && (
-                          <>
-                            {' '}
-                            <a
-                              href={item.link.href}
-                              target="_blank"
-                              rel="noopener noreferrer"
-                              className="text-brand hover:underline font-medium"
-                            >
-                              {item.link.text}
-                            </a>
-                          </>
-                        )}
-                      </p>
-                    )}
-                    {item.list && (
-                      <ul className="mt-1.5 ml-8 list-disc space-y-0.5">
-                        {item.list.map((li) => (
-                          <li key={li}>{li}</li>
-                        ))}
-                      </ul>
-                    )}
-                  </dd>
-                </div>
-              ))}
-            </dl>
-          </Card>
-        </section>
+        <FaqList title="Frequently Asked Questions" items={FAQ_ITEMS} />
       </div>
     </main>
   );
diff --git a/packages/app/src/app/datasets/[slug]/conversations/[convId]/page.tsx b/packages/app/src/app/agentx/[slug]/conversations/[convId]/page.tsx
similarity index 94%
rename from packages/app/src/app/datasets/[slug]/conversations/[convId]/page.tsx
rename to packages/app/src/app/agentx/[slug]/conversations/[convId]/page.tsx
index 8957c8b4d..8a9d79226 100644
--- a/packages/app/src/app/datasets/[slug]/conversations/[convId]/page.tsx
+++ b/packages/app/src/app/agentx/[slug]/conversations/[convId]/page.tsx
@@ -19,7 +19,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     title,
     description,
     alternates: {
-      canonical: `${SITE_URL}/datasets/${slug}/conversations/${encodeURIComponent(convId)}`,
+      canonical: `${SITE_URL}/agentx/${slug}/conversations/${encodeURIComponent(convId)}`,
     },
     robots: { index: false }, // per-conversation pages are too numerous to index
   };
diff --git a/packages/app/src/app/datasets/[slug]/page.tsx b/packages/app/src/app/agentx/[slug]/page.tsx
similarity index 83%
rename from packages/app/src/app/datasets/[slug]/page.tsx
rename to packages/app/src/app/agentx/[slug]/page.tsx
index 567f5588d..613ccde1a 100644
--- a/packages/app/src/app/datasets/[slug]/page.tsx
+++ b/packages/app/src/app/agentx/[slug]/page.tsx
@@ -10,16 +10,16 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
-  const title = `${slug} | Agentic Datasets`;
+  const title = `${slug} | AgentX Datasets`;
   const description = `Distributions, token statistics, and per-conversation flamegraphs for the ${slug} agentic trace dataset.`;
   return {
     title,
     description,
     alternates: {
-      canonical: `${SITE_URL}/datasets/${slug}`,
-      languages: languageAlternates(`/datasets/${slug}`),
+      canonical: `${SITE_URL}/agentx/${slug}`,
+      languages: languageAlternates(`/agentx/${slug}`),
     },
-    openGraph: { title: `${title} | InferenceX`, description, url: `${SITE_URL}/datasets/${slug}` },
+    openGraph: { title: `${title} | InferenceX`, description, url: `${SITE_URL}/agentx/${slug}` },
     twitter: { title: `${title} | InferenceX`, description },
   };
 }
diff --git a/packages/app/src/app/agentx/methodology/page.tsx b/packages/app/src/app/agentx/methodology/page.tsx
new file mode 100644
index 000000000..4aee25e6d
--- /dev/null
+++ b/packages/app/src/app/agentx/methodology/page.tsx
@@ -0,0 +1,42 @@
+import type { Metadata } from 'next';
+
+import { AgentXMethodologyArticle } from '@/components/datasets/agentx-methodology-article';
+import { JsonLd } from '@/components/json-ld';
+import { enAlternates } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const DESCRIPTION =
+  'How AgentX transforms opt-in coding-agent traces into replay graphs with controlled KV-cache warmup and benchmark settings.';
+
+export const metadata: Metadata = {
+  title: 'AgentX Methodology',
+  description: DESCRIPTION,
+  alternates: enAlternates('/agentx/methodology'),
+  openGraph: {
+    title: 'AgentX Methodology | InferenceX',
+    description: DESCRIPTION,
+    url: `${SITE_URL}/agentx/methodology`,
+  },
+  twitter: { title: 'AgentX Methodology | InferenceX', description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'TechArticle',
+  headline: 'AgentX Methodology',
+  description: DESCRIPTION,
+  url: `${SITE_URL}/agentx/methodology`,
+  inLanguage: 'en',
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default function AgentXMethodologyPage() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXMethodologyArticle locale="en" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/agentx/optimizations/[framework]/page.tsx b/packages/app/src/app/agentx/optimizations/[framework]/page.tsx
new file mode 100644
index 000000000..b6dff301e
--- /dev/null
+++ b/packages/app/src/app/agentx/optimizations/[framework]/page.tsx
@@ -0,0 +1,64 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { AgentXOptimizationsArticle } from '@/components/datasets/agentx-optimizations-article';
+import { JsonLd } from '@/components/json-ld';
+import {
+  AGENTX_OPTIMIZATION_FRAMEWORKS,
+  getOptimizationFramework,
+} from '@/lib/agentx-optimizations';
+import { enAlternates } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+interface Props {
+  params: Promise<{ framework: string }>;
+}
+
+export function generateStaticParams() {
+  return AGENTX_OPTIMIZATION_FRAMEWORKS.map((entry) => ({ framework: entry.slug }));
+}
+
+export const dynamicParams = false;
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { framework: slug } = await params;
+  const framework = getOptimizationFramework(slug);
+  if (!framework) return {};
+  const title = `${framework.name} Agentic Optimizations`;
+  return {
+    title,
+    description: framework.summary,
+    alternates: enAlternates(`/agentx/optimizations/${slug}`),
+    openGraph: {
+      title: `${title} | InferenceX`,
+      description: framework.summary,
+      url: `${SITE_URL}/agentx/optimizations/${slug}`,
+    },
+    twitter: { title: `${title} | InferenceX`, description: framework.summary },
+  };
+}
+
+export default async function AgentXOptimizationsFrameworkPage({ params }: Props) {
+  const { framework: slug } = await params;
+  const framework = getOptimizationFramework(slug);
+  if (!framework) notFound();
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'TechArticle',
+    headline: `${framework.name} Agentic Optimizations`,
+    description: framework.summary,
+    url: `${SITE_URL}/agentx/optimizations/${slug}`,
+    inLanguage: 'en',
+    isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXOptimizationsArticle slug={slug} locale="en" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/agentx/optimizations/page.tsx b/packages/app/src/app/agentx/optimizations/page.tsx
new file mode 100644
index 000000000..559df325a
--- /dev/null
+++ b/packages/app/src/app/agentx/optimizations/page.tsx
@@ -0,0 +1,45 @@
+import type { Metadata } from 'next';
+
+import { AgentXOptimizationsIndex } from '@/components/datasets/agentx-optimizations-article';
+import { JsonLd } from '@/components/json-ld';
+import { OPTIMIZATIONS_OVERVIEW } from '@/lib/agentx-optimizations';
+import { enAlternates } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const TITLE = 'AgentX Industry Impact: Optimizations for Agentic Workloads';
+const DESCRIPTION =
+  'The 50+ upstream pull requests AgentX drove across vLLM, SGLang, TensorRT-LLM, ATOM, AITER, Dynamo, LMCache, and Mooncake.';
+
+export const metadata: Metadata = {
+  title: 'AgentX Industry Impact',
+  description: DESCRIPTION,
+  alternates: enAlternates('/agentx/optimizations'),
+  openGraph: {
+    title: `${TITLE} | InferenceX`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/agentx/optimizations`,
+  },
+  twitter: { title: `${TITLE} | InferenceX`, description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'TechArticle',
+  headline: TITLE,
+  description: DESCRIPTION,
+  url: `${SITE_URL}/agentx/optimizations`,
+  inLanguage: 'en',
+  about: OPTIMIZATIONS_OVERVIEW.title,
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default function AgentXOptimizationsPage() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXOptimizationsIndex locale="en" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/agentx/page.tsx b/packages/app/src/app/agentx/page.tsx
new file mode 100644
index 000000000..41308e6ec
--- /dev/null
+++ b/packages/app/src/app/agentx/page.tsx
@@ -0,0 +1,48 @@
+import type { Metadata } from 'next';
+
+import { AgentXMethodology } from '@/components/datasets/agentx-methodology';
+import { DatasetList } from '@/components/datasets/dataset-list';
+import { JsonLd } from '@/components/json-ld';
+import { enAlternates } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const DESCRIPTION =
+  'AgentX replays workload shapes derived from opt-in Claude Code sessions. Explore its methodology, distributions, and per-conversation flamegraphs.';
+
+export const metadata: Metadata = {
+  title: 'AgentX Methodology and Datasets',
+  description: DESCRIPTION,
+  alternates: enAlternates('/agentx'),
+  openGraph: {
+    title: 'AgentX Methodology and Datasets | InferenceX',
+    description: DESCRIPTION,
+    url: `${SITE_URL}/agentx`,
+  },
+  twitter: { title: 'AgentX Methodology and Datasets | InferenceX', description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'CollectionPage',
+  name: 'InferenceX AgentX Datasets',
+  description: DESCRIPTION,
+  url: `${SITE_URL}/agentx`,
+};
+
+export default function AgentXPage() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto flex flex-col gap-6 px-4 pb-8 lg:px-8">
+        <section>
+          <AgentXMethodology locale="en" />
+        </section>
+
+        <section className="flex flex-col gap-3">
+          <h2 className="text-lg font-semibold text-foreground">Datasets</h2>
+          <DatasetList />
+        </section>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/agentx/telemetry/page.tsx b/packages/app/src/app/agentx/telemetry/page.tsx
new file mode 100644
index 000000000..374f1fbcb
--- /dev/null
+++ b/packages/app/src/app/agentx/telemetry/page.tsx
@@ -0,0 +1,45 @@
+import type { Metadata } from 'next';
+
+import { AgentXTelemetryArticle } from '@/components/datasets/agentx-telemetry-article';
+import { JsonLd } from '@/components/json-ld';
+import { AGENTX_TELEMETRY_GUIDE } from '@/lib/agentx-telemetry';
+import { enAlternates } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const TITLE = 'Exploring Agentic Workloads: Detailed Telemetry';
+const DESCRIPTION =
+  'How to read the telemetry behind a single AgentX point: the per-point charts, the request timeline, KV offload markers, and the per-conversation flamegraph.';
+
+export const metadata: Metadata = {
+  title: 'AgentX Telemetry Tutorial',
+  description: DESCRIPTION,
+  alternates: enAlternates('/agentx/telemetry'),
+  openGraph: {
+    title: `${TITLE} | InferenceX`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/agentx/telemetry`,
+  },
+  twitter: { title: `${TITLE} | InferenceX`, description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'TechArticle',
+  headline: TITLE,
+  description: DESCRIPTION,
+  url: `${SITE_URL}/agentx/telemetry`,
+  inLanguage: 'en',
+  about: AGENTX_TELEMETRY_GUIDE.title,
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default function AgentXTelemetryPage() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXTelemetryArticle locale="en" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/api/unofficial-run/route.test.ts b/packages/app/src/app/api/unofficial-run/route.test.ts
index b9c45b801..d4a8e5f4d 100644
--- a/packages/app/src/app/api/unofficial-run/route.test.ts
+++ b/packages/app/src/app/api/unofficial-run/route.test.ts
@@ -144,6 +144,70 @@ describe('normalizeArtifactRows', () => {
     expect(m.mean_e2el).toBe(1.5);
   });
 
+  it.each([
+    {
+      name: 'boolean verdict and junk-suffixed schema',
+      input: { power_valid: true, power_metric_schema_version: '2garbage' },
+      expectedVerdict: 0,
+      expectedSchema: undefined,
+    },
+    {
+      name: 'garbage verdict and valid numeric schema',
+      input: { power_valid: 'garbage', power_metric_schema_version: 2 },
+      expectedVerdict: 0,
+      expectedSchema: 2,
+    },
+    {
+      name: 'canonical numeric strings',
+      input: { power_valid: '1', power_metric_schema_version: '2' },
+      expectedVerdict: 1,
+      expectedSchema: 2,
+    },
+    {
+      name: 'explicit invalid verdict',
+      input: { power_valid: 0, power_metric_schema_version: 2 },
+      expectedVerdict: 0,
+      expectedSchema: 2,
+    },
+    {
+      name: 'legacy row without power contract fields',
+      input: {},
+      expectedVerdict: undefined,
+      expectedSchema: undefined,
+    },
+  ])(
+    'normalizes overlay power contract discriminators: $name',
+    ({ input, expectedVerdict, expectedSchema }) => {
+      const [row] = normalizeArtifactRows([rawRow(input)], '2026-03-01');
+
+      if (expectedVerdict === undefined) {
+        expect(row.metrics).not.toHaveProperty('power_valid');
+      } else {
+        expect(row.metrics.power_valid).toBe(expectedVerdict);
+      }
+      if (expectedSchema === undefined) {
+        expect(row.metrics).not.toHaveProperty('power_metric_schema_version');
+      } else {
+        expect(row.metrics.power_metric_schema_version).toBe(expectedSchema);
+      }
+    },
+  );
+
+  it('preserves recipe identity for unofficial overlays', () => {
+    const rows = normalizeArtifactRows(
+      [
+        rawRow({
+          recipe_fingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+        }),
+      ],
+      '2026-03-01',
+    );
+
+    expect(rows[0].recipe_fingerprint).toBe(
+      'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+    );
+  });
+
   it('surfaces pipeline-parallelism fields in metrics (auto-capture)', () => {
     // pp has no configs-table column: the frontend reads it from the metrics
     // JSONB (rowToAggDataEntry), so the overlay route must keep passing it
diff --git a/packages/app/src/app/api/unofficial-run/route.ts b/packages/app/src/app/api/unofficial-run/route.ts
index 168d05d51..b827b6d75 100644
--- a/packages/app/src/app/api/unofficial-run/route.ts
+++ b/packages/app/src/app/api/unofficial-run/route.ts
@@ -68,6 +68,7 @@ export function normalizeArtifactRows(
       osl: params.osl,
       conc: params.conc,
       image: params.image,
+      recipe_fingerprint: params.recipeFingerprint,
       metrics: params.metrics,
       // Surface the same per-worker payload the DB path emits so unofficial
       // overlays carry the multinode measured-power breakdown too.
diff --git a/packages/app/src/app/api/v1/agentic-aggregates/route.ts b/packages/app/src/app/api/v1/agentic-aggregates/route.ts
index 9808cec6e..f0cd38408 100644
--- a/packages/app/src/app/api/v1/agentic-aggregates/route.ts
+++ b/packages/app/src/app/api/v1/agentic-aggregates/route.ts
@@ -32,7 +32,7 @@ const getCachedAgenticAggregates = cachedQuery(
 /**
  * GET /api/v1/agentic-aggregates?ids=1,2,3
  *
- * Returns per-id mean/p50/p75/p90/p99 for ISL, OSL, KV cache utilization,
+ * Returns per-id mean/p50/p75/p90/p95/p99 for ISL, OSL, KV cache utilization,
  * and prefix cache hit rate — computed live from the stored aiperf
  * profile_export.jsonl + server_metrics_json blobs. Ids without a
  * trace_replay blob (or with no usable samples) get nulls.
diff --git a/packages/app/src/app/api/v1/agentic-cache-keys.test.ts b/packages/app/src/app/api/v1/agentic-cache-keys.test.ts
index 7177f9960..57be6bbcc 100644
--- a/packages/app/src/app/api/v1/agentic-cache-keys.test.ts
+++ b/packages/app/src/app/api/v1/agentic-cache-keys.test.ts
@@ -26,12 +26,17 @@ vi.mock('@/lib/api-cache', () => ({
 
 import { STATS_VERSION } from '@semianalysisai/inferencex-db/queries/agentic-aggregates';
 import { REQUEST_TIMELINE_VERSION } from '@semianalysisai/inferencex-db/etl/compute-request-timeline';
+import { CHART_SERIES_VERSION } from '@semianalysisai/inferencex-db/etl/compute-chart-series';
+import { REQUEST_CHART_DATA_VERSION } from '@semianalysisai/inferencex-db/queries/request-chart-data';
 import { TRACE_SERVER_METRICS_VERSION } from '@semianalysisai/inferencex-db/queries/trace-server-metrics';
 
 import { CACHE_KEY_PREFIX as agenticAggregatesKey } from './agentic-aggregates/route';
 import { CACHE_KEY_PREFIX as requestTimelineKey } from './request-timeline/route';
+import { CACHE_KEY_PREFIX as requestChartDataKey } from './request-chart-data/route';
+import { CACHE_KEY_PREFIX as traceServerMetricSourceKey } from './trace-server-metric-source/route';
 import { CACHE_KEY_PREFIX as traceServerMetricsKey } from './trace-server-metrics/route';
 import { CACHE_KEY_PREFIX as traceHistogramsKey } from './trace-histograms/route';
+import { CACHE_KEY_PREFIX as residentSequenceLengthsKey } from './resident-sequence-lengths/route';
 
 describe('agentic blob-cache keys are version-derived', () => {
   it('agentic-aggregates key embeds STATS_VERSION', () => {
@@ -42,20 +47,35 @@ describe('agentic blob-cache keys are version-derived', () => {
     expect(requestTimelineKey).toBe(`request-timeline-v${REQUEST_TIMELINE_VERSION}`);
   });
 
+  it('request-chart-data key embeds REQUEST_CHART_DATA_VERSION', () => {
+    expect(requestChartDataKey).toBe(`request-chart-data-v${REQUEST_CHART_DATA_VERSION}`);
+  });
+
   it('trace-server-metrics key embeds its composite response version', () => {
     expect(traceServerMetricsKey).toBe(`trace-server-metrics-v${TRACE_SERVER_METRICS_VERSION}`);
   });
 
+  it('trace-server-metric-source key embeds CHART_SERIES_VERSION', () => {
+    expect(traceServerMetricSourceKey).toBe(`trace-server-metric-source-v${CHART_SERIES_VERSION}`);
+  });
+
   it('trace-histograms key embeds REQUEST_TIMELINE_VERSION (its payload is read from request_timeline)', () => {
     expect(traceHistogramsKey).toBe(`trace-histograms-v${REQUEST_TIMELINE_VERSION}`);
   });
 
+  it('resident-sequence-lengths key embeds STATS_VERSION', () => {
+    expect(residentSequenceLengthsKey).toBe(`resident-sequence-lengths-v${STATS_VERSION}`);
+  });
+
   it('every key actually contains a version segment (no unversioned literals)', () => {
     for (const key of [
       agenticAggregatesKey,
       requestTimelineKey,
+      requestChartDataKey,
       traceServerMetricsKey,
+      traceServerMetricSourceKey,
       traceHistogramsKey,
+      residentSequenceLengthsKey,
     ]) {
       expect(key).toMatch(/-v\d+$/u);
     }
diff --git a/packages/app/src/app/api/v1/availability/route.ts b/packages/app/src/app/api/v1/availability/route.ts
index c3c4bab0a..b8db29c86 100644
--- a/packages/app/src/app/api/v1/availability/route.ts
+++ b/packages/app/src/app/api/v1/availability/route.ts
@@ -1,23 +1,15 @@
-import { NextResponse } from 'next/server';
-
-import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
+import { getDb } from '@semianalysisai/inferencex-db/connection';
 
 import { getAvailabilityData } from '@semianalysisai/inferencex-db/queries/workflow-info';
 
-import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { cachedReadRoute } from '@/lib/cached-read-route';
 import { loadFixture } from '@/lib/test-fixtures';
 
 export const dynamic = 'force-dynamic';
 
-const getCachedAvailability = cachedQuery(() => getAvailabilityData(getDb()), 'availability');
-
-export async function GET() {
-  if (FIXTURES_MODE) return cachedJson(loadFixture('availability'));
-  try {
-    const rows = await getCachedAvailability();
-    return cachedJson(rows);
-  } catch (error) {
-    console.error('Error fetching availability:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
-  }
-}
+export const GET = cachedReadRoute({
+  cacheKey: 'availability',
+  fetch: () => getAvailabilityData(getDb()),
+  fixture: () => loadFixture('availability'),
+  logLabel: 'availability',
+});
diff --git a/packages/app/src/app/api/v1/benchmark-siblings/route.ts b/packages/app/src/app/api/v1/benchmark-siblings/route.ts
index 1e7482278..a5a5ce5ec 100644
--- a/packages/app/src/app/api/v1/benchmark-siblings/route.ts
+++ b/packages/app/src/app/api/v1/benchmark-siblings/route.ts
@@ -13,9 +13,9 @@ export const dynamic = 'force-dynamic';
 
 const getCachedSiblings = cachedQuery(
   (id: number): Promise<BenchmarkSiblings | null> => getBenchmarkSiblings(getDb(), id),
-  // v2 adds PP and makes the aggregate topology explicit. Roll the blob
-  // namespace so old TP0 sibling payloads cannot survive the schema change.
-  'benchmark-siblings-v2',
+  // v3 adds DCP/PCP. Roll the blob namespace so cached sibling payloads gain
+  // the complete parallelism mapping used by chart point labels.
+  'benchmark-siblings-v3',
 );
 
 /**
diff --git a/packages/app/src/app/api/v1/benchmarks/history/route.ts b/packages/app/src/app/api/v1/benchmarks/history/route.ts
index 113a6ed65..28ba786ba 100644
--- a/packages/app/src/app/api/v1/benchmarks/history/route.ts
+++ b/packages/app/src/app/api/v1/benchmarks/history/route.ts
@@ -1,11 +1,13 @@
-import { type NextRequest, NextResponse } from 'next/server';
+import type { NextRequest } from 'next/server';
 
-import { DISPLAY_MODEL_TO_DB } from '@semianalysisai/inferencex-constants';
+import { DISPLAY_MODEL_TO_DB, islOslToSequence } from '@semianalysisai/inferencex-constants';
 import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
 
 import { getAllBenchmarksForHistory } from '@semianalysisai/inferencex-db/queries/benchmarks';
 
 import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { toCalculatorBenchmarkRows } from '@/lib/benchmark-api-view';
+import { PUBLIC_API_ERRORS, publicApiError } from '@/lib/public-api-errors';
 import { loadFixture } from '@/lib/test-fixtures';
 import { agenticWorkflowMetadataOnly } from '@/lib/agentic-workflow-metadata';
 
@@ -24,6 +26,34 @@ const getCachedAgenticBenchmarkHistory = cachedQuery(
   { blobOnly: true },
 );
 
+/**
+ * Trim a history response to the calculator's metric allowlist.
+ *
+ * Opt-in only. `CALCULATOR_METRIC_KEYS` excludes the measured-power metrics
+ * (`measured_avg_power_w`, `joules_per_output_token`, …) that Historical Trends
+ * plots, so applying this to every history response would silently blank those
+ * charts. The sequence is derived from the route's own isl/osl rather than
+ * taken as a parameter, so the two can never disagree.
+ *
+ * Agentic requests carry no isl/osl and so have no sequence to key on. Those
+ * responses pass through untrimmed rather than guessing a sequence — the trim is
+ * a payload optimisation, and the agentic payload is a fraction of a fixed one.
+ * The calculator's lifecycle section consumes them either way.
+ */
+function applyCalculatorView<
+  T extends {
+    benchmark_type: string;
+    isl: number | null;
+    osl: number | null;
+    metrics: Record<string, unknown>;
+  },
+>(rows: T[], isl: number | null, osl: number | null): T[] {
+  if (isl === null || osl === null) return rows;
+  const sequence = islOslToSequence(isl, osl);
+  if (!sequence) return rows;
+  return toCalculatorBenchmarkRows(rows, sequence);
+}
+
 export async function GET(request: NextRequest) {
   const model = request.nextUrl.searchParams.get('model') ?? '';
   const rawIsl = request.nextUrl.searchParams.get('isl');
@@ -32,26 +62,34 @@ export async function GET(request: NextRequest) {
   const isl = rawIsl === null ? null : Number(rawIsl);
   const osl = rawOsl === null ? null : Number(rawOsl);
   const isAgentic = benchmarkType === 'agentic_traces';
+  const calculatorView = request.nextUrl.searchParams.get('view') === 'calculator';
 
   if (!model) {
-    return NextResponse.json({ error: 'model, isl, and osl are required' }, { status: 400 });
+    return publicApiError(PUBLIC_API_ERRORS.benchmarkHistoryParameters, 400);
   }
   if (!isAgentic && (!isl || !osl)) {
-    return NextResponse.json({ error: 'model, isl, and osl are required' }, { status: 400 });
+    return publicApiError(PUBLIC_API_ERRORS.benchmarkHistoryParameters, 400);
+  }
+  if (FIXTURES_MODE) {
+    const fixture = loadFixture('benchmarks-history');
+    return cachedJson(
+      calculatorView && Array.isArray(fixture) ? applyCalculatorView(fixture, isl, osl) : fixture,
+    );
   }
-  if (FIXTURES_MODE) return cachedJson(loadFixture('benchmarks-history'));
 
   try {
     const modelKeys = DISPLAY_MODEL_TO_DB[model];
     if (!modelKeys || modelKeys.length === 0) {
-      return NextResponse.json({ error: 'Unknown model' }, { status: 400 });
+      return publicApiError(PUBLIC_API_ERRORS.unknownModel, 400);
     }
     const rows = isAgentic
       ? await getCachedAgenticBenchmarkHistory(modelKeys)
       : await getCachedBenchmarkHistory(modelKeys, isl!, osl!);
-    return cachedJson(agenticWorkflowMetadataOnly(rows));
+    // Both trims run after the cache, so every view shares one cached query.
+    const trimmed = agenticWorkflowMetadataOnly(rows);
+    return cachedJson(calculatorView ? applyCalculatorView(trimmed, isl, osl) : trimmed);
   } catch (error) {
     console.error('Error fetching benchmark history:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+    return publicApiError(PUBLIC_API_ERRORS.internal, 500);
   }
 }
diff --git a/packages/app/src/app/api/v1/benchmarks/route.ts b/packages/app/src/app/api/v1/benchmarks/route.ts
index 84ef529f8..a75260d7b 100644
--- a/packages/app/src/app/api/v1/benchmarks/route.ts
+++ b/packages/app/src/app/api/v1/benchmarks/route.ts
@@ -11,6 +11,7 @@ import {
 
 import { cachedJson, cachedQuery } from '@/lib/api-cache';
 import { toCalculatorBenchmarkRows } from '@/lib/benchmark-api-view';
+import { PUBLIC_API_ERRORS, publicApiError } from '@/lib/public-api-errors';
 import { agenticWorkflowMetadataOnly } from '@/lib/agentic-workflow-metadata';
 import { loadFixture } from '@/lib/test-fixtures';
 
@@ -23,8 +24,9 @@ const getCachedBenchmarks = cachedQuery(
   { blobOnly: true },
 );
 
-// Exactly one run's results (GPU comparison of individual same-day runs). Cached
-// under a distinct key prefix so it never collides with the latest/as-of query.
+// One logical run snapshot (GPU comparison of individual same-day runs). For an
+// append-only run this includes its same-image predecessor chain. Cached under a
+// distinct key prefix so it never collides with the latest/as-of query.
 const getCachedBenchmarksForRun = cachedQuery(
   (dbModelKeys: string[], runId: string) => getBenchmarksForRun(getDb(), dbModelKeys, runId),
   'benchmarks-run-agentic-run-metadata',
@@ -52,7 +54,7 @@ export async function GET(request: NextRequest) {
   const sequence = params.get('sequence') ?? '';
   const dbModelKeys = DISPLAY_MODEL_TO_DB[model];
   if (!dbModelKeys || dbModelKeys.length === 0) {
-    return NextResponse.json({ error: 'Unknown model' }, { status: 400 });
+    return publicApiError(PUBLIC_API_ERRORS.unknownModel, 400);
   }
   if (view === 'calculator' && !['1k/1k', '1k/8k', '8k/1k', 'agentic-traces'].includes(sequence)) {
     return NextResponse.json({ error: 'Unknown calculator sequence' }, { status: 400 });
@@ -74,6 +76,6 @@ export async function GET(request: NextRequest) {
     return cachedJson(agenticWorkflowMetadataOnly(rows));
   } catch (error) {
     console.error('Error fetching benchmarks:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+    return publicApiError(PUBLIC_API_ERRORS.internal, 500);
   }
 }
diff --git a/packages/app/src/app/api/v1/collectivex/latest/route.test.ts b/packages/app/src/app/api/v1/collectivex/latest/route.test.ts
index da3ee6929..b806694b6 100644
--- a/packages/app/src/app/api/v1/collectivex/latest/route.test.ts
+++ b/packages/app/src/app/api/v1/collectivex/latest/route.test.ts
@@ -2,15 +2,15 @@ import { beforeEach, describe, expect, it, vi } from 'vitest';
 
 import { makeCollectiveXDataset } from '@semianalysisai/inferencex-db/collectivex/test-fixture';
 
-const { mockGetLatest, mockFromRow, mockGetDb, mockEnsureLatest } = vi.hoisted(() => ({
+const { mockGetLatest, mockFromRow, mockGetWriteDb, mockEnsureLatest } = vi.hoisted(() => ({
   mockGetLatest: vi.fn(),
   mockFromRow: vi.fn(),
-  mockGetDb: vi.fn(() => 'mock-sql'),
+  mockGetWriteDb: vi.fn(() => 'mock-write-sql'),
   mockEnsureLatest: vi.fn(),
 }));
 
 vi.mock('@semianalysisai/inferencex-db/connection', () => ({
-  getCollectiveXDb: mockGetDb,
+  getCollectiveXWriteDb: mockGetWriteDb,
   FIXTURES_MODE: false,
 }));
 
@@ -73,7 +73,7 @@ describe('GET /api/v1/collectivex/latest', () => {
     expect(res.status).toBe(200);
     expect(await res.json()).toEqual(dataset);
     expect(mockEnsureLatest).toHaveBeenCalledWith(1);
-    expect(mockGetLatest).toHaveBeenCalledWith('mock-sql', 1);
+    expect(mockGetLatest).toHaveBeenCalledWith('mock-write-sql', 1);
   });
 
   it('serves the stored run when GitHub discovery fails', async () => {
diff --git a/packages/app/src/app/api/v1/collectivex/latest/route.ts b/packages/app/src/app/api/v1/collectivex/latest/route.ts
index 264b5a842..302063309 100644
--- a/packages/app/src/app/api/v1/collectivex/latest/route.ts
+++ b/packages/app/src/app/api/v1/collectivex/latest/route.ts
@@ -1,7 +1,7 @@
 import { type NextRequest, NextResponse } from 'next/server';
 
 import { parseCollectiveXVersion } from '@semianalysisai/inferencex-db/collectivex/types';
-import { FIXTURES_MODE, getCollectiveXDb } from '@semianalysisai/inferencex-db/connection';
+import { FIXTURES_MODE, getCollectiveXWriteDb } from '@semianalysisai/inferencex-db/connection';
 import {
   collectiveXDatasetFromRow,
   getLatestCollectiveXRun,
@@ -34,7 +34,8 @@ export async function GET(request: NextRequest) {
   }
 
   try {
-    const row = await getLatestCollectiveXRun(getCollectiveXDb(), version);
+    // Discovery may have persisted a run, so read from the same primary.
+    const row = await getLatestCollectiveXRun(getCollectiveXWriteDb(), version);
     if (row !== null) {
       if (ensureError)
         console.error('CollectiveX discovery failed; serving stored run:', ensureError);
diff --git a/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.fixtures.test.ts b/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.fixtures.test.ts
index 4c2b1b5ce..cd03846ed 100644
--- a/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.fixtures.test.ts
+++ b/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.fixtures.test.ts
@@ -12,7 +12,6 @@ const { mockLoadFixture } = vi.hoisted(() => ({
 
 vi.mock('@semianalysisai/inferencex-db/connection', () => ({
   FIXTURES_MODE: true,
-  getCollectiveXDb: vi.fn(),
   getCollectiveXWriteDb: vi.fn(),
 }));
 
diff --git a/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.test.ts b/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.test.ts
index c2f2c51ad..4ceed09db 100644
--- a/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.test.ts
+++ b/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.test.ts
@@ -2,19 +2,17 @@ import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
 import { makeCollectiveXDataset } from '@semianalysisai/inferencex-db/collectivex/test-fixture';
 
-const { mockGetRun, mockDelete, mockFromRow, mockGetDb, mockGetWriteDb, mockPurge, mockEnsureRun } =
+const { mockGetRun, mockDelete, mockFromRow, mockGetWriteDb, mockPurge, mockEnsureRun } =
   vi.hoisted(() => ({
     mockGetRun: vi.fn(),
     mockDelete: vi.fn(),
     mockFromRow: vi.fn(),
-    mockGetDb: vi.fn(() => 'mock-sql'),
     mockGetWriteDb: vi.fn(() => 'mock-write-sql'),
     mockPurge: vi.fn(() => Promise.resolve(0)),
     mockEnsureRun: vi.fn(),
   }));
 
 vi.mock('@semianalysisai/inferencex-db/connection', () => ({
-  getCollectiveXDb: mockGetDb,
   getCollectiveXWriteDb: mockGetWriteDb,
   FIXTURES_MODE: false,
 }));
@@ -99,7 +97,7 @@ describe('GET /api/v1/collectivex/runs/[runId]', () => {
     expect(res.status).toBe(200);
     expect(await res.json()).toEqual(dataset);
     expect(mockEnsureRun).toHaveBeenCalledWith(1, runId);
-    expect(mockGetRun).toHaveBeenCalledWith('mock-sql', 1, runId);
+    expect(mockGetRun).toHaveBeenCalledWith('mock-write-sql', 1, runId);
   });
 
   it.each([
diff --git a/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.ts b/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.ts
index e8150c1fb..43883cbd5 100644
--- a/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.ts
+++ b/packages/app/src/app/api/v1/collectivex/runs/[runId]/route.ts
@@ -1,5 +1,3 @@
-import { timingSafeEqual } from 'crypto';
-
 import { type NextRequest, NextResponse } from 'next/server';
 
 import {
@@ -7,11 +5,7 @@ import {
   type CollectiveXRunSummary,
   parseCollectiveXVersion,
 } from '@semianalysisai/inferencex-db/collectivex/types';
-import {
-  FIXTURES_MODE,
-  getCollectiveXDb,
-  getCollectiveXWriteDb,
-} from '@semianalysisai/inferencex-db/connection';
+import { FIXTURES_MODE, getCollectiveXWriteDb } from '@semianalysisai/inferencex-db/connection';
 import {
   collectiveXDatasetFromRow,
   deleteCollectiveXRun,
@@ -24,6 +18,7 @@ import {
   cachedJson,
   purgeCollectiveX,
 } from '@/lib/api-cache';
+import { bearerMatches } from '@/lib/bearer-auth';
 import { collectiveXSweepErrorStatus, ensureCollectiveXRun } from '@/lib/collectivex-lazy-ingest';
 import { loadFixture } from '@/lib/test-fixtures';
 
@@ -58,7 +53,8 @@ export async function GET(request: NextRequest, context: { params: Promise<{ run
   }
 
   try {
-    const row = await getCollectiveXRun(getCollectiveXDb(), version, runId);
+    // Ensure may have inserted or refreshed this run, so read from the same primary.
+    const row = await getCollectiveXRun(getCollectiveXWriteDb(), version, runId);
     if (row === null) {
       if (ensureError) {
         const status = collectiveXSweepErrorStatus(ensureError);
@@ -87,15 +83,6 @@ export async function GET(request: NextRequest, context: { params: Promise<{ run
   }
 }
 
-/** Constant-time Bearer check that tolerates multibyte header bytes. */
-function bearerMatches(header: string, secret: string): boolean {
-  const provided = Buffer.from(header);
-  const expected = Buffer.from(`Bearer ${secret}`);
-  // Compare BYTE lengths — a multibyte char can make the JS string lengths
-  // equal while the buffers differ, and timingSafeEqual throws on that.
-  return provided.length === expected.length && timingSafeEqual(provided, expected);
-}
-
 /**
  * Admin deletion of an ingested run. Authenticated with a dedicated Bearer
  * secret — the token is remembered in browser localStorage, so it must not
diff --git a/packages/app/src/app/api/v1/collectivex/runs/route.test.ts b/packages/app/src/app/api/v1/collectivex/runs/route.test.ts
index 658d7b178..b4230fcaf 100644
--- a/packages/app/src/app/api/v1/collectivex/runs/route.test.ts
+++ b/packages/app/src/app/api/v1/collectivex/runs/route.test.ts
@@ -3,15 +3,15 @@ import { beforeEach, describe, expect, it, vi } from 'vitest';
 import { buildRunSummary } from '@semianalysisai/inferencex-db/collectivex/reader';
 import { makeCollectiveXDataset } from '@semianalysisai/inferencex-db/collectivex/test-fixture';
 
-const { mockList, mockGetDb, mockEnsureList, mockCachedJson } = vi.hoisted(() => ({
+const { mockList, mockGetWriteDb, mockEnsureList, mockCachedJson } = vi.hoisted(() => ({
   mockList: vi.fn(),
-  mockGetDb: vi.fn(() => 'mock-sql'),
+  mockGetWriteDb: vi.fn(() => 'mock-write-sql'),
   mockEnsureList: vi.fn(),
   mockCachedJson: vi.fn((data: unknown) => Response.json(data)),
 }));
 
 vi.mock('@semianalysisai/inferencex-db/connection', () => ({
-  getCollectiveXDb: mockGetDb,
+  getCollectiveXWriteDb: mockGetWriteDb,
   FIXTURES_MODE: false,
 }));
 
@@ -76,7 +76,7 @@ describe('GET /api/v1/collectivex/runs', () => {
       discovery_complete: true,
     });
     expect(mockEnsureList).toHaveBeenCalledWith(1);
-    expect(mockList).toHaveBeenCalledWith('mock-sql', 1);
+    expect(mockList).toHaveBeenCalledWith('mock-write-sql', 1);
   });
 
   it('marks the list incomplete when another bounded discovery pass is needed', async () => {
diff --git a/packages/app/src/app/api/v1/collectivex/runs/route.ts b/packages/app/src/app/api/v1/collectivex/runs/route.ts
index fce3f2a2d..62b2ca22d 100644
--- a/packages/app/src/app/api/v1/collectivex/runs/route.ts
+++ b/packages/app/src/app/api/v1/collectivex/runs/route.ts
@@ -1,7 +1,7 @@
 import { type NextRequest, NextResponse } from 'next/server';
 
 import { parseCollectiveXVersion } from '@semianalysisai/inferencex-db/collectivex/types';
-import { FIXTURES_MODE, getCollectiveXDb } from '@semianalysisai/inferencex-db/connection';
+import { FIXTURES_MODE, getCollectiveXWriteDb } from '@semianalysisai/inferencex-db/connection';
 import { listCollectiveXRuns } from '@semianalysisai/inferencex-db/queries/collectivex';
 
 import { COLLECTIVEX_CACHE_CONTROL, cachedJson, collectiveXCacheTag } from '@/lib/api-cache';
@@ -35,7 +35,8 @@ export async function GET(request: NextRequest) {
   }
 
   try {
-    const runs = await listCollectiveXRuns(getCollectiveXDb(), version);
+    // Backfill may have persisted runs, so read from the same primary.
+    const runs = await listCollectiveXRuns(getCollectiveXWriteDb(), version);
     if (runs.length === 0 && ensureError) {
       console.error('CollectiveX run backfill failed with no stored fallback:', ensureError);
       const status = collectiveXSweepErrorStatus(ensureError) ?? 502;
diff --git a/packages/app/src/app/api/v1/datasets/route.ts b/packages/app/src/app/api/v1/datasets/route.ts
index 242a2b086..ec0a3160e 100644
--- a/packages/app/src/app/api/v1/datasets/route.ts
+++ b/packages/app/src/app/api/v1/datasets/route.ts
@@ -1,25 +1,14 @@
-import { NextResponse } from 'next/server';
-
 import { getDb } from '@semianalysisai/inferencex-db/connection';
 
 import { listDatasets, type DatasetRecord } from '@semianalysisai/inferencex-db/queries/datasets';
 
-import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { cachedReadRoute } from '@/lib/cached-read-route';
 
 export const dynamic = 'force-dynamic';
 
-const getCachedDatasets = cachedQuery(
-  (): Promise<DatasetRecord[]> => listDatasets(getDb()),
-  'datasets',
-);
-
 /** GET /api/v1/datasets — all ingested cc-traces-weka datasets (registry cards). */
-export async function GET() {
-  try {
-    const data = await getCachedDatasets();
-    return cachedJson(data);
-  } catch (error) {
-    console.error('Error fetching datasets:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
-  }
-}
+export const GET = cachedReadRoute<DatasetRecord[]>({
+  cacheKey: 'datasets',
+  fetch: () => listDatasets(getDb()),
+  logLabel: 'datasets',
+});
diff --git a/packages/app/src/app/api/v1/eval-samples-live/route.ts b/packages/app/src/app/api/v1/eval-samples-live/route.ts
index a98775dbf..e985f553e 100644
--- a/packages/app/src/app/api/v1/eval-samples-live/route.ts
+++ b/packages/app/src/app/api/v1/eval-samples-live/route.ts
@@ -6,6 +6,7 @@
 import { type NextRequest, NextResponse } from 'next/server';
 
 import { extractDemonstrations } from '@/lib/eval-sample-utils';
+import { parseEvalSampleWindow } from '@/lib/eval-sample-params';
 import {
   type EvalArtifactConfig,
   fetchAndParseSamples,
@@ -15,10 +16,6 @@ import { fetchGithubRunArtifacts, getGithubToken } from '@/lib/github-artifacts'
 
 export const dynamic = 'force-dynamic';
 
-const ALLOWED_FILTERS = new Set(['all', 'passed', 'failed']);
-const DEFAULT_LIMIT = 200;
-const MAX_LIMIT = 500;
-
 /**
  * GET /api/v1/eval-samples-live
  *   ?run_id=<numeric>
@@ -42,10 +39,6 @@ export async function GET(request: NextRequest) {
   const params = request.nextUrl.searchParams;
   const runId = params.get('run_id') ?? '';
   const task = params.get('task') ?? '';
-  const filterParam = params.get('filter') ?? 'all';
-  const offset = Math.max(0, Math.trunc(Number(params.get('offset') ?? '0')));
-  const requestedLimit = Math.trunc(Number(params.get('limit') ?? String(DEFAULT_LIMIT)));
-  const limit = Math.min(MAX_LIMIT, Math.max(1, requestedLimit || DEFAULT_LIMIT));
 
   if (!/^\d+$/u.test(runId)) {
     return NextResponse.json({ error: 'run_id must be a positive integer' }, { status: 400 });
@@ -53,13 +46,11 @@ export async function GET(request: NextRequest) {
   if (!task) {
     return NextResponse.json({ error: 'task is required' }, { status: 400 });
   }
-  if (!ALLOWED_FILTERS.has(filterParam)) {
-    return NextResponse.json(
-      { error: `filter must be one of: ${[...ALLOWED_FILTERS].join(', ')}` },
-      { status: 400 },
-    );
+  const window = parseEvalSampleWindow(params);
+  if ('error' in window) {
+    return NextResponse.json(window, { status: 400 });
   }
-  const filter = filterParam as 'all' | 'passed' | 'failed';
+  const { filter, offset, limit } = window;
 
   const config: EvalArtifactConfig = {
     model: params.get('model') ?? '',
diff --git a/packages/app/src/app/api/v1/eval-samples/route.ts b/packages/app/src/app/api/v1/eval-samples/route.ts
index 897eb5d2a..93d5f8670 100644
--- a/packages/app/src/app/api/v1/eval-samples/route.ts
+++ b/packages/app/src/app/api/v1/eval-samples/route.ts
@@ -8,13 +8,10 @@ import {
 
 import { cachedJson, cachedQuery } from '@/lib/api-cache';
 import { extractDemonstrations } from '@/lib/eval-sample-utils';
+import { parseEvalSampleWindow } from '@/lib/eval-sample-params';
 
 export const dynamic = 'force-dynamic';
 
-const ALLOWED_FILTERS = new Set(['all', 'passed', 'failed']);
-const DEFAULT_LIMIT = 200;
-const MAX_LIMIT = 500;
-
 const getCachedEvalSamples = cachedQuery(
   (evalResultId: number, filter: 'all' | 'passed' | 'failed', offset: number, limit: number) =>
     getEvalSamples(getDb(), evalResultId, filter, offset, limit),
@@ -38,12 +35,8 @@ const getCachedEvalSampleOffset = cachedQuery(
 export async function GET(request: NextRequest) {
   const params = request.nextUrl.searchParams;
   const evalResultId = Number(params.get('eval_result_id'));
-  const filterParam = params.get('filter') ?? 'all';
-  const offset = Math.max(0, Math.trunc(Number(params.get('offset') ?? '0')));
   const docIdParam = params.get('doc_id');
   const docId = docIdParam === null ? null : Math.trunc(Number(docIdParam));
-  const requestedLimit = Math.trunc(Number(params.get('limit') ?? String(DEFAULT_LIMIT)));
-  const limit = Math.min(MAX_LIMIT, Math.max(1, requestedLimit || DEFAULT_LIMIT));
 
   if (!evalResultId || !Number.isFinite(evalResultId) || evalResultId <= 0) {
     return NextResponse.json(
@@ -54,13 +47,11 @@ export async function GET(request: NextRequest) {
   if (docId !== null && (!Number.isFinite(docId) || docId < 0)) {
     return NextResponse.json({ error: 'doc_id must be a non-negative integer' }, { status: 400 });
   }
-  if (!ALLOWED_FILTERS.has(filterParam)) {
-    return NextResponse.json(
-      { error: `filter must be one of: ${[...ALLOWED_FILTERS].join(', ')}` },
-      { status: 400 },
-    );
+  const window = parseEvalSampleWindow(params);
+  if ('error' in window) {
+    return NextResponse.json(window, { status: 400 });
   }
-  const filter = filterParam as 'all' | 'passed' | 'failed';
+  const { filter, offset, limit } = window;
 
   try {
     const sampleOffset =
diff --git a/packages/app/src/app/api/v1/evaluations/route.ts b/packages/app/src/app/api/v1/evaluations/route.ts
index 9fafff4b0..23477a4a4 100644
--- a/packages/app/src/app/api/v1/evaluations/route.ts
+++ b/packages/app/src/app/api/v1/evaluations/route.ts
@@ -1,23 +1,15 @@
-import { NextResponse } from 'next/server';
-
-import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
+import { getDb } from '@semianalysisai/inferencex-db/connection';
 
 import { getAllEvalResults } from '@semianalysisai/inferencex-db/queries/evaluations';
 
-import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { cachedReadRoute } from '@/lib/cached-read-route';
 import { loadFixture } from '@/lib/test-fixtures';
 
 export const dynamic = 'force-dynamic';
 
-const getCachedEvaluations = cachedQuery(() => getAllEvalResults(getDb()), 'evaluations');
-
-export async function GET() {
-  if (FIXTURES_MODE) return cachedJson(loadFixture('evaluations'));
-  try {
-    const rows = await getCachedEvaluations();
-    return cachedJson(rows);
-  } catch (error) {
-    console.error('Error fetching evaluations:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
-  }
-}
+export const GET = cachedReadRoute({
+  cacheKey: 'evaluations',
+  fetch: () => getAllEvalResults(getDb()),
+  fixture: () => loadFixture('evaluations'),
+  logLabel: 'evaluations',
+});
diff --git a/packages/app/src/app/api/v1/id-routes.ts b/packages/app/src/app/api/v1/id-routes.ts
index fea9221bd..920e6ad4b 100644
--- a/packages/app/src/app/api/v1/id-routes.ts
+++ b/packages/app/src/app/api/v1/id-routes.ts
@@ -1,6 +1,7 @@
 import { type NextRequest, NextResponse } from 'next/server';
 
 import { cachedJson } from '@/lib/api-cache';
+import { PUBLIC_API_ERRORS, publicApiError, tooManyIdsError } from '@/lib/public-api-errors';
 
 /**
  * Shared GET-handler factories for the agentic benchmark routes, which all
@@ -22,7 +23,7 @@ import { cachedJson } from '@/lib/api-cache';
 export function parseIdsParam(request: NextRequest, maxIds: number): number[] | NextResponse {
   const raw = request.nextUrl.searchParams.get('ids');
   if (!raw) {
-    return NextResponse.json({ error: 'ids query param is required' }, { status: 400 });
+    return publicApiError(PUBLIC_API_ERRORS.idsRequired, 400);
   }
 
   const ids = [
@@ -34,10 +35,10 @@ export function parseIdsParam(request: NextRequest, maxIds: number): number[] |
     ),
   ];
   if (ids.length === 0) {
-    return NextResponse.json({ error: 'no valid ids provided' }, { status: 400 });
+    return publicApiError(PUBLIC_API_ERRORS.noValidIds, 400);
   }
   if (ids.length > maxIds) {
-    return NextResponse.json({ error: `too many ids (max ${maxIds})` }, { status: 400 });
+    return publicApiError(tooManyIdsError(maxIds), 400);
   }
   return ids.toSorted((a, b) => a - b);
 }
@@ -57,7 +58,7 @@ export function idsQueryRoute<T>(options: {
       return cachedJson(await fetch(ids));
     } catch (error) {
       console.error(`Error fetching ${logLabel}:`, error);
-      return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+      return publicApiError(PUBLIC_API_ERRORS.internal, 500);
     }
   };
 }
@@ -71,15 +72,15 @@ export function idQueryRoute<T>(options: {
   return async (request: NextRequest) => {
     const id = Number(request.nextUrl.searchParams.get('id'));
     if (!id || !Number.isFinite(id)) {
-      return NextResponse.json({ error: 'id is required (benchmark_result_id)' }, { status: 400 });
+      return publicApiError(PUBLIC_API_ERRORS.idRequired, 400);
     }
     try {
       const data = await fetch(id);
-      if (!data) return NextResponse.json({ error: 'Not found' }, { status: 404 });
+      if (!data) return publicApiError(PUBLIC_API_ERRORS.notFound, 404);
       return cachedJson(data);
     } catch (error) {
       console.error(`Error fetching ${logLabel}:`, error);
-      return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+      return publicApiError(PUBLIC_API_ERRORS.internal, 500);
     }
   };
 }
diff --git a/packages/app/src/app/api/v1/invalidate/route.ts b/packages/app/src/app/api/v1/invalidate/route.ts
index e648e75b1..129b3926e 100644
--- a/packages/app/src/app/api/v1/invalidate/route.ts
+++ b/packages/app/src/app/api/v1/invalidate/route.ts
@@ -1,14 +1,12 @@
-import { timingSafeEqual } from 'crypto';
-
 import { NextResponse } from 'next/server';
 
+import { bearerMatches } from '@/lib/bearer-auth';
+
 import { COLLECTIVEX_CACHE_SCOPE, purgeAll, purgeCollectiveX } from '@/lib/api-cache';
 
 export async function POST(request: Request) {
   const secret = process.env.INVALIDATE_SECRET;
   const authHeader = request.headers.get('Authorization') ?? '';
-  const provided = Buffer.from(authHeader);
-  const expected = Buffer.from(`Bearer ${secret}`);
   // The shared staging deployment is already protected by Vercel. Its CI
   // caller must present the project-scoped automation bypass before Vercel
   // forwards this header to the route, so a second app secret is redundant.
@@ -19,12 +17,7 @@ export async function POST(request: Request) {
     process.env.VERCEL_GIT_COMMIT_REF === 'staging' &&
     Boolean(request.headers.get('x-vercel-protection-bypass'));
 
-  // Compare BYTE lengths — a multibyte char can make the JS string lengths
-  // equal while the buffers differ, and timingSafeEqual throws on that.
-  if (
-    !isProtectedStagingRequest &&
-    (!secret || provided.length !== expected.length || !timingSafeEqual(provided, expected))
-  ) {
+  if (!isProtectedStagingRequest && (!secret || !bearerMatches(authHeader, secret))) {
     return new NextResponse('Unauthorized', { status: 401 });
   }
 
diff --git a/packages/app/src/app/api/v1/latest-images/route.ts b/packages/app/src/app/api/v1/latest-images/route.ts
index 001f01ea5..eaa15ff79 100644
--- a/packages/app/src/app/api/v1/latest-images/route.ts
+++ b/packages/app/src/app/api/v1/latest-images/route.ts
@@ -1,22 +1,13 @@
-import { NextResponse } from 'next/server';
-
 import { getDb } from '@semianalysisai/inferencex-db/connection';
 import { getLatestImages } from '@semianalysisai/inferencex-db/queries/latest-images';
 
-import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { cachedReadRoute } from '@/lib/cached-read-route';
 
 export const dynamic = 'force-dynamic';
 
-const getCachedLatestImages = cachedQuery(() => getLatestImages(getDb()), 'latest-images-v2', {
+export const GET = cachedReadRoute({
+  cacheKey: 'latest-images-v2',
+  fetch: () => getLatestImages(getDb()),
+  logLabel: 'latest images',
   blobOnly: true,
 });
-
-export async function GET() {
-  try {
-    const rows = await getCachedLatestImages();
-    return cachedJson(rows);
-  } catch (error) {
-    console.error('Error fetching latest images:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
-  }
-}
diff --git a/packages/app/src/app/api/v1/log-availability/route.test.ts b/packages/app/src/app/api/v1/log-availability/route.test.ts
new file mode 100644
index 000000000..ffc9e8ead
--- /dev/null
+++ b/packages/app/src/app/api/v1/log-availability/route.test.ts
@@ -0,0 +1,38 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+const { mockGetServerLogAvailability, mockGetDb } = vi.hoisted(() => ({
+  mockGetServerLogAvailability: vi.fn(),
+  mockGetDb: vi.fn(() => 'mock-sql'),
+}));
+
+vi.mock('@semianalysisai/inferencex-db/connection', () => ({ getDb: mockGetDb }));
+vi.mock('@semianalysisai/inferencex-db/queries/server-logs', () => ({
+  getServerLogAvailability: mockGetServerLogAvailability,
+}));
+vi.mock('@/lib/api-cache', () => ({
+  cachedQuery: (fn: (...args: any[]) => any) => fn,
+  cachedJson: (data: unknown) => Response.json(data),
+}));
+
+import { NextRequest } from 'next/server';
+import { GET } from './route';
+
+const req = (url: string) => new NextRequest(new URL(url, 'http://localhost'));
+
+beforeEach(() => vi.clearAllMocks());
+
+describe('GET /api/v1/log-availability', () => {
+  it('returns a presence map for valid, deduplicated ids', async () => {
+    mockGetServerLogAvailability.mockResolvedValueOnce({ 4: true });
+    const response = await GET(req('/api/v1/log-availability?ids=9,4,4'));
+    expect(response.status).toBe(200);
+    await expect(response.json()).resolves.toEqual({ 4: true });
+    expect(mockGetServerLogAvailability).toHaveBeenCalledWith('mock-sql', [4, 9]);
+  });
+
+  it('returns 400 when no valid ids are provided', async () => {
+    const response = await GET(req('/api/v1/log-availability?ids=bad,0'));
+    expect(response.status).toBe(400);
+    expect(mockGetServerLogAvailability).not.toHaveBeenCalled();
+  });
+});
diff --git a/packages/app/src/app/api/v1/log-availability/route.ts b/packages/app/src/app/api/v1/log-availability/route.ts
new file mode 100644
index 000000000..77e7c0a48
--- /dev/null
+++ b/packages/app/src/app/api/v1/log-availability/route.ts
@@ -0,0 +1,28 @@
+import { getDb } from '@semianalysisai/inferencex-db/connection';
+import {
+  getServerLogAvailability,
+  type ServerLogAvailabilityMap,
+} from '@semianalysisai/inferencex-db/queries/server-logs';
+
+import { cachedQuery } from '@/lib/api-cache';
+
+import { idsQueryRoute } from '../id-routes';
+
+export const dynamic = 'force-dynamic';
+
+const getCachedServerLogAvailability = cachedQuery(
+  (ids: number[]): Promise<ServerLogAvailabilityMap> => getServerLogAvailability(getDb(), ids),
+  'server-log-availability',
+);
+
+/**
+ * GET /api/v1/log-availability?ids=1,2,3
+ *
+ * Returns `{[id]: true}` for benchmark points with a linked server log. The
+ * chart uses this lightweight check before offering its "View logs" action.
+ */
+export const GET = idsQueryRoute({
+  maxIds: 500,
+  logLabel: 'server log availability',
+  fetch: getCachedServerLogAvailability,
+});
diff --git a/packages/app/src/app/api/v1/overview/route.test.ts b/packages/app/src/app/api/v1/overview/route.test.ts
index 2a0829d2b..e68107f73 100644
--- a/packages/app/src/app/api/v1/overview/route.test.ts
+++ b/packages/app/src/app/api/v1/overview/route.test.ts
@@ -37,12 +37,22 @@ describe('GET /api/v1/overview', () => {
     mockGetOverviewPageData.mockResolvedValueOnce(data);
 
     const response = await GET(
-      request('/api/v1/overview?tier=75&engine=all&compare=30d&ref=b300&models=all'),
+      request(
+        '/api/v1/overview?tier=75&engine=all&compare=30d&ref=b300&models=all&rows=changed&hwrows=priced',
+      ),
     );
 
     expect(response.status).toBe(200);
     expect(await response.json()).toEqual(data);
-    expect(mockGetOverviewPageData).toHaveBeenCalledWith(75, 'all', '30d', 'b300', 'all');
+    expect(mockGetOverviewPageData).toHaveBeenCalledWith(
+      75,
+      'all',
+      '30d',
+      'b300',
+      'all',
+      'changed',
+      'priced',
+    );
     expect(mockCachedJson).toHaveBeenCalledWith(data);
   });
 
@@ -50,7 +60,9 @@ describe('GET /api/v1/overview', () => {
     mockGetOverviewPageData.mockResolvedValueOnce({ models: [] });
 
     await GET(
-      request('/api/v1/overview?tier=999&engine=vendor&compare=weekly&ref=h100&models=inactive'),
+      request(
+        '/api/v1/overview?tier=999&engine=vendor&compare=weekly&ref=h100&models=inactive&rows=some&hwrows=blank',
+      ),
     );
 
     expect(mockGetOverviewPageData).toHaveBeenCalledWith(
@@ -59,6 +71,8 @@ describe('GET /api/v1/overview', () => {
       'hardware',
       'b200',
       'default',
+      'all',
+      'all',
     );
   });
 
diff --git a/packages/app/src/app/api/v1/overview/route.ts b/packages/app/src/app/api/v1/overview/route.ts
index ba6819160..3842b87c7 100644
--- a/packages/app/src/app/api/v1/overview/route.ts
+++ b/packages/app/src/app/api/v1/overview/route.ts
@@ -4,8 +4,10 @@ import { cachedJson } from '@/lib/api-cache';
 import {
   resolveOverviewComparisonMode,
   resolveOverviewEngineScope,
+  resolveOverviewHardwareRowScope,
   resolveOverviewModelScope,
   resolveOverviewReferenceHardware,
+  resolveOverviewRowScope,
   resolveOverviewTier,
 } from '@/lib/overview-data';
 import { getOverviewPageData } from '@/lib/overview-data.server';
@@ -22,6 +24,8 @@ export async function GET(request: NextRequest) {
       resolveOverviewComparisonMode(params.get('compare') ?? undefined),
       resolveOverviewReferenceHardware(params.get('ref') ?? undefined),
       resolveOverviewModelScope(params.get('models') ?? undefined),
+      resolveOverviewRowScope(params.get('rows') ?? undefined),
+      resolveOverviewHardwareRowScope(params.get('hwrows') ?? undefined),
     );
     return cachedJson(data);
   } catch (error) {
diff --git a/packages/app/src/app/api/v1/reliability/route.ts b/packages/app/src/app/api/v1/reliability/route.ts
index eb3641433..faa0ea6c1 100644
--- a/packages/app/src/app/api/v1/reliability/route.ts
+++ b/packages/app/src/app/api/v1/reliability/route.ts
@@ -1,23 +1,15 @@
-import { NextResponse } from 'next/server';
-
-import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
+import { getDb } from '@semianalysisai/inferencex-db/connection';
 
 import { getReliabilityStats } from '@semianalysisai/inferencex-db/queries/reliability';
 
-import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { cachedReadRoute } from '@/lib/cached-read-route';
 import { loadFixture } from '@/lib/test-fixtures';
 
 export const dynamic = 'force-dynamic';
 
-const getCachedReliability = cachedQuery(() => getReliabilityStats(getDb()), 'reliability');
-
-export async function GET() {
-  if (FIXTURES_MODE) return cachedJson(loadFixture('reliability'));
-  try {
-    const rows = await getCachedReliability();
-    return cachedJson(rows);
-  } catch (error) {
-    console.error('Error fetching reliability stats:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
-  }
-}
+export const GET = cachedReadRoute({
+  cacheKey: 'reliability',
+  fetch: () => getReliabilityStats(getDb()),
+  fixture: () => loadFixture('reliability'),
+  logLabel: 'reliability stats',
+});
diff --git a/packages/app/src/app/api/v1/request-chart-data/route.ts b/packages/app/src/app/api/v1/request-chart-data/route.ts
new file mode 100644
index 000000000..ddba93496
--- /dev/null
+++ b/packages/app/src/app/api/v1/request-chart-data/route.ts
@@ -0,0 +1,33 @@
+import { getDb } from '@semianalysisai/inferencex-db/connection';
+
+import {
+  getRequestChartData,
+  REQUEST_CHART_DATA_VERSION,
+  type RequestChartDataWire,
+} from '@semianalysisai/inferencex-db/queries/request-chart-data';
+
+import { cachedQuery } from '@/lib/api-cache';
+
+import { idQueryRoute } from '../id-routes';
+
+export const dynamic = 'force-dynamic';
+
+/** Version-derived blob-cache key namespace (exported for the key-derivation test). */
+export const CACHE_KEY_PREFIX = `request-chart-data-v${REQUEST_CHART_DATA_VERSION}`;
+
+const getCachedRequestChartData = cachedQuery(
+  (id: number): Promise<RequestChartDataWire | null> => getRequestChartData(getDb(), id),
+  CACHE_KEY_PREFIX,
+  { blobOnly: true },
+);
+
+/**
+ * GET /api/v1/request-chart-data?id=N
+ *
+ * Returns the compact request fields used by the default point charts. The
+ * full request timeline remains a separate, timeline-view-only download.
+ */
+export const GET = idQueryRoute({
+  logLabel: 'request chart data',
+  fetch: getCachedRequestChartData,
+});
diff --git a/packages/app/src/app/api/v1/resident-sequence-lengths/route.ts b/packages/app/src/app/api/v1/resident-sequence-lengths/route.ts
new file mode 100644
index 000000000..794a53357
--- /dev/null
+++ b/packages/app/src/app/api/v1/resident-sequence-lengths/route.ts
@@ -0,0 +1,32 @@
+import { getDb } from '@semianalysisai/inferencex-db/connection';
+import { STATS_VERSION } from '@semianalysisai/inferencex-db/queries/agentic-shared';
+import {
+  getResidentSequenceLengthSketches,
+  type ResidentSequenceLengthSketches,
+} from '@semianalysisai/inferencex-db/queries/resident-sequence-lengths';
+
+import { cachedQuery } from '@/lib/api-cache';
+
+import { idsQueryRoute } from '../id-routes';
+
+export const dynamic = 'force-dynamic';
+
+export const CACHE_KEY_PREFIX = `resident-sequence-lengths-v${STATS_VERSION}`;
+
+const getCachedResidentSequenceLengths = cachedQuery(
+  (ids: number[]): Promise<ResidentSequenceLengthSketches> =>
+    getResidentSequenceLengthSketches(getDb(), ids),
+  CACHE_KEY_PREFIX,
+);
+
+/**
+ * GET /api/v1/resident-sequence-lengths?ids=1,2,3
+ *
+ * Returns one mergeable ISL/OSL sketch for the requested resident chart
+ * points. The client chunks large charts and merges these bounded payloads.
+ */
+export const GET = idsQueryRoute({
+  maxIds: 200,
+  logLabel: 'resident sequence lengths',
+  fetch: getCachedResidentSequenceLengths,
+});
diff --git a/packages/app/src/app/api/v1/server-log-files/route.test.ts b/packages/app/src/app/api/v1/server-log-files/route.test.ts
new file mode 100644
index 000000000..36707c18f
--- /dev/null
+++ b/packages/app/src/app/api/v1/server-log-files/route.test.ts
@@ -0,0 +1,50 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+const { mockGetServerLogFileNames, mockGetDb } = vi.hoisted(() => ({
+  mockGetServerLogFileNames: vi.fn(),
+  mockGetDb: vi.fn(() => 'mock-sql'),
+}));
+
+vi.mock('@semianalysisai/inferencex-db/connection', () => ({
+  getDb: mockGetDb,
+  FIXTURES_MODE: false,
+}));
+
+vi.mock('@semianalysisai/inferencex-db/queries/server-logs', () => ({
+  getServerLogFileNames: mockGetServerLogFileNames,
+}));
+
+vi.mock('@/lib/api-cache', () => ({
+  cachedQuery: (fn: (...args: any[]) => any) => fn,
+  cachedJson: (data: unknown) => Response.json(data),
+}));
+
+import { NextRequest } from 'next/server';
+import { GET } from './route';
+
+const req = (url: string) => new NextRequest(new URL(url, 'http://localhost'));
+
+beforeEach(() => vi.clearAllMocks());
+
+describe('GET /api/v1/server-log-files', () => {
+  it('returns artifact-relative filenames in database order', async () => {
+    mockGetServerLogFileNames.mockResolvedValueOnce([
+      'results/server.log',
+      'results/benchmark.log',
+      'results/router.log',
+    ]);
+    const response = await GET(req('/api/v1/server-log-files?id=439516'));
+    expect(response.status).toBe(200);
+    await expect(response.json()).resolves.toEqual([
+      'results/server.log',
+      'results/benchmark.log',
+      'results/router.log',
+    ]);
+  });
+
+  it('returns 404 when no log bundle is linked', async () => {
+    mockGetServerLogFileNames.mockResolvedValueOnce(null);
+    const response = await GET(req('/api/v1/server-log-files?id=9'));
+    expect(response.status).toBe(404);
+  });
+});
diff --git a/packages/app/src/app/api/v1/server-log-files/route.ts b/packages/app/src/app/api/v1/server-log-files/route.ts
new file mode 100644
index 000000000..308ace8d9
--- /dev/null
+++ b/packages/app/src/app/api/v1/server-log-files/route.ts
@@ -0,0 +1,19 @@
+import { getDb } from '@semianalysisai/inferencex-db/connection';
+import { getServerLogFileNames } from '@semianalysisai/inferencex-db/queries/server-logs';
+
+import { cachedQuery } from '@/lib/api-cache';
+
+import { idQueryRoute } from '../id-routes';
+
+export const dynamic = 'force-dynamic';
+
+const getCachedServerLogFileNames = cachedQuery(
+  (id: number) => getServerLogFileNames(getDb(), id),
+  'server-log-files',
+);
+
+/** GET /api/v1/server-log-files?id=1 — artifact-relative .log/.out filenames. */
+export const GET = idQueryRoute({
+  logLabel: 'server log filenames',
+  fetch: getCachedServerLogFileNames,
+});
diff --git a/packages/app/src/app/api/v1/server-log-search/route.test.ts b/packages/app/src/app/api/v1/server-log-search/route.test.ts
new file mode 100644
index 000000000..2f78ffc79
--- /dev/null
+++ b/packages/app/src/app/api/v1/server-log-search/route.test.ts
@@ -0,0 +1,81 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+const { mockSearchServerLogs, mockGetDb } = vi.hoisted(() => ({
+  mockSearchServerLogs: vi.fn(),
+  mockGetDb: vi.fn(() => 'mock-sql'),
+}));
+
+vi.mock('@semianalysisai/inferencex-db/connection', () => ({
+  getDb: mockGetDb,
+  FIXTURES_MODE: false,
+}));
+
+vi.mock('@semianalysisai/inferencex-db/queries/server-logs', () => ({
+  searchServerLogs: mockSearchServerLogs,
+}));
+
+vi.mock('@/lib/api-cache', () => ({
+  cachedJson: (data: unknown) => Response.json(data),
+}));
+
+import { NextRequest } from 'next/server';
+import { GET } from './route';
+
+function req(url: string): NextRequest {
+  return new NextRequest(new URL(url, 'http://localhost'));
+}
+
+beforeEach(() => vi.clearAllMocks());
+
+describe('GET /api/v1/server-log-search', () => {
+  it('searches complete stored logs with a bounded result count', async () => {
+    mockSearchServerLogs.mockResolvedValueOnce({
+      matches: [
+        {
+          fileName: 'results/router.log',
+          offset: 12,
+          before: 'router ',
+          match: 'ready',
+          after: '\n',
+        },
+      ],
+      truncated: false,
+    });
+
+    const response = await GET(req('/api/v1/server-log-search?id=42&q=ready&limit=20'));
+    expect(response.status).toBe(200);
+    await expect(response.json()).resolves.toEqual({
+      id: 42,
+      query: 'ready',
+      matches: [
+        {
+          fileName: 'results/router.log',
+          offset: 12,
+          before: 'router ',
+          match: 'ready',
+          after: '\n',
+        },
+      ],
+      truncated: false,
+    });
+    expect(mockSearchServerLogs).toHaveBeenCalledWith('mock-sql', 42, 'ready', 20);
+  });
+
+  it.each([
+    '/api/v1/server-log-search?q=ready',
+    '/api/v1/server-log-search?id=42&q=',
+    `/api/v1/server-log-search?id=42&q=${'x'.repeat(257)}`,
+    '/api/v1/server-log-search?id=42&q=ready&limit=0',
+    '/api/v1/server-log-search?id=42&q=ready&limit=101',
+  ])('rejects invalid search request %s', async (url) => {
+    const response = await GET(req(url));
+    expect(response.status).toBe(400);
+    expect(mockSearchServerLogs).not.toHaveBeenCalled();
+  });
+
+  it('returns 500 when the database search fails', async () => {
+    mockSearchServerLogs.mockRejectedValueOnce(new Error('Connection reset'));
+    const response = await GET(req('/api/v1/server-log-search?id=42&q=ready'));
+    expect(response.status).toBe(500);
+  });
+});
diff --git a/packages/app/src/app/api/v1/server-log-search/route.ts b/packages/app/src/app/api/v1/server-log-search/route.ts
new file mode 100644
index 000000000..6a03a3728
--- /dev/null
+++ b/packages/app/src/app/api/v1/server-log-search/route.ts
@@ -0,0 +1,46 @@
+import { type NextRequest, NextResponse } from 'next/server';
+
+import { getDb } from '@semianalysisai/inferencex-db/connection';
+import { searchServerLogs } from '@semianalysisai/inferencex-db/queries/server-logs';
+
+import { cachedJson } from '@/lib/api-cache';
+
+export const dynamic = 'force-dynamic';
+
+const DEFAULT_RESULT_LIMIT = 50;
+const MAX_RESULT_LIMIT = 100;
+const MAX_SEARCH_LENGTH = 256;
+
+function parseLimit(value: string | null): number | null {
+  if (value === null) return DEFAULT_RESULT_LIMIT;
+  if (!/^\d+$/u.test(value)) return null;
+  const parsed = Number(value);
+  return Number.isSafeInteger(parsed) && parsed >= 1 && parsed <= MAX_RESULT_LIMIT ? parsed : null;
+}
+
+export async function GET(request: NextRequest) {
+  const id = Number(request.nextUrl.searchParams.get('id'));
+  if (!id || !Number.isFinite(id)) {
+    return NextResponse.json({ error: 'id is required (benchmark_result_id)' }, { status: 400 });
+  }
+
+  const query = request.nextUrl.searchParams.get('q')?.trim() ?? '';
+  if (query.length === 0 || query.length > MAX_SEARCH_LENGTH || query.includes('\u0000')) {
+    return NextResponse.json(
+      { error: `q must contain 1-${MAX_SEARCH_LENGTH} characters` },
+      { status: 400 },
+    );
+  }
+
+  const limit = parseLimit(request.nextUrl.searchParams.get('limit'));
+  if (limit === null) {
+    return NextResponse.json({ error: `limit must be 1-${MAX_RESULT_LIMIT}` }, { status: 400 });
+  }
+
+  try {
+    return cachedJson({ id, query, ...(await searchServerLogs(getDb(), id, query, limit)) });
+  } catch (error) {
+    console.error('Error searching server logs:', error);
+    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+  }
+}
diff --git a/packages/app/src/app/api/v1/server-log/route.test.ts b/packages/app/src/app/api/v1/server-log/route.test.ts
index e8159c846..dbd4b3aa0 100644
--- a/packages/app/src/app/api/v1/server-log/route.test.ts
+++ b/packages/app/src/app/api/v1/server-log/route.test.ts
@@ -1,7 +1,8 @@
 import { describe, expect, it, vi, beforeEach } from 'vitest';
 
-const { mockGetServerLog, mockGetDb } = vi.hoisted(() => ({
+const { mockGetServerLog, mockGetServerLogChunk, mockGetDb } = vi.hoisted(() => ({
   mockGetServerLog: vi.fn(),
+  mockGetServerLogChunk: vi.fn(),
   mockGetDb: vi.fn(() => 'mock-sql'),
 }));
 
@@ -12,6 +13,7 @@ vi.mock('@semianalysisai/inferencex-db/connection', () => ({
 
 vi.mock('@semianalysisai/inferencex-db/queries/server-logs', () => ({
   getServerLog: mockGetServerLog,
+  getServerLogChunk: mockGetServerLogChunk,
 }));
 
 vi.mock('@/lib/api-cache', () => ({
@@ -72,6 +74,137 @@ describe('GET /api/v1/server-log', () => {
     expect(mockGetServerLog).toHaveBeenCalledWith('mock-sql', 42);
   });
 
+  it('returns a bounded chunk when offset or limit is provided', async () => {
+    mockGetServerLogChunk.mockResolvedValueOnce({
+      fileName: 'results/router.log',
+      serverLog: 'INFO ready\n',
+      offset: 65536,
+      nextOffset: 65547,
+    });
+
+    const res = await GET(
+      req('/api/v1/server-log?id=42&file=results%2Frouter.log&offset=65536&limit=1024'),
+    );
+    expect(res.status).toBe(200);
+    await expect(res.json()).resolves.toEqual({
+      id: 42,
+      fileName: 'results/router.log',
+      serverLog: 'INFO ready\n',
+      offset: 65536,
+      nextOffset: 65547,
+    });
+    expect(mockGetServerLogChunk).toHaveBeenCalledWith(
+      'mock-sql',
+      42,
+      65536,
+      1024,
+      'results/router.log',
+    );
+    expect(mockGetServerLog).not.toHaveBeenCalled();
+  });
+
+  it('uses bounded defaults when only one chunk parameter is provided', async () => {
+    mockGetServerLogChunk.mockResolvedValueOnce({
+      fileName: 'server.log',
+      serverLog: 'start',
+      offset: 0,
+      nextOffset: null,
+    });
+
+    const res = await GET(req('/api/v1/server-log?id=42&offset=0'));
+    expect(res.status).toBe(200);
+    expect(mockGetServerLogChunk).toHaveBeenCalledWith('mock-sql', 42, 0, 64 * 1024, undefined);
+  });
+
+  it('streams the complete selected file as a download attachment', async () => {
+    mockGetServerLogChunk
+      .mockResolvedValueOnce({
+        fileName: 'nested/router log.out',
+        serverLog: 'INFO',
+        offset: 0,
+        nextOffset: 4,
+      })
+      .mockResolvedValueOnce({
+        fileName: 'nested/router log.out',
+        serverLog: ' ready\n',
+        offset: 4,
+        nextOffset: null,
+      });
+
+    const res = await GET(
+      req('/api/v1/server-log?id=42&file=nested%2Frouter%20log.out&download=1'),
+    );
+
+    expect(res.status).toBe(200);
+    expect(res.headers.get('content-type')).toBe('text/plain; charset=utf-8');
+    expect(res.headers.get('content-disposition')).toContain(
+      'attachment; filename="router log.out"',
+    );
+    await expect(res.text()).resolves.toBe('INFO ready\n');
+    expect(mockGetServerLogChunk).toHaveBeenNthCalledWith(
+      1,
+      'mock-sql',
+      42,
+      0,
+      256 * 1024,
+      'nested/router log.out',
+    );
+    expect(mockGetServerLogChunk).toHaveBeenNthCalledWith(
+      2,
+      'mock-sql',
+      42,
+      4,
+      256 * 1024,
+      'nested/router log.out',
+    );
+    expect(mockGetServerLog).not.toHaveBeenCalled();
+  });
+
+  it.each(['download=0', 'download=yes', 'download=1&offset=0', 'download=1&limit=1'])(
+    'returns 400 for invalid download parameters %s',
+    async (query) => {
+      const res = await GET(req(`/api/v1/server-log?id=42&${query}`));
+      expect(res.status).toBe(400);
+      expect(mockGetServerLogChunk).not.toHaveBeenCalled();
+    },
+  );
+
+  it.each([
+    'offset=-1',
+    'offset=1.5',
+    'offset=2000000001',
+    'limit=0',
+    'limit=262145',
+    'limit=nope',
+  ])('returns 400 for invalid chunk parameter %s', async (query) => {
+    const res = await GET(req(`/api/v1/server-log?id=42&${query}`));
+    expect(res.status).toBe(400);
+    expect(mockGetServerLogChunk).not.toHaveBeenCalled();
+  });
+
+  it('returns 400 for an empty filename', async () => {
+    const res = await GET(req('/api/v1/server-log?id=42&file='));
+    expect(res.status).toBe(400);
+    expect(mockGetServerLog).not.toHaveBeenCalled();
+  });
+
+  it('returns a complete named log for compatibility clients', async () => {
+    mockGetServerLog.mockResolvedValueOnce('router output');
+    const res = await GET(req('/api/v1/server-log?id=42&file=router.log'));
+    await expect(res.json()).resolves.toEqual({
+      id: 42,
+      fileName: 'router.log',
+      serverLog: 'router output',
+    });
+    expect(mockGetServerLog).toHaveBeenCalledWith('mock-sql', 42, 'router.log');
+  });
+
+  it('returns 404 when a requested chunk has no linked log', async () => {
+    mockGetServerLogChunk.mockResolvedValueOnce(null);
+    const res = await GET(req('/api/v1/server-log?id=999&offset=0'));
+    expect(res.status).toBe(404);
+  });
+
   it('returns 500 when query throws', async () => {
     mockGetServerLog.mockRejectedValueOnce(new Error('Connection reset'));
 
diff --git a/packages/app/src/app/api/v1/server-log/route.ts b/packages/app/src/app/api/v1/server-log/route.ts
index 55de025b4..dafaaff4f 100644
--- a/packages/app/src/app/api/v1/server-log/route.ts
+++ b/packages/app/src/app/api/v1/server-log/route.ts
@@ -2,33 +2,177 @@ import { type NextRequest, NextResponse } from 'next/server';
 
 import { getDb } from '@semianalysisai/inferencex-db/connection';
 
-import { getServerLog } from '@semianalysisai/inferencex-db/queries/server-logs';
+import { getServerLog, getServerLogChunk } from '@semianalysisai/inferencex-db/queries/server-logs';
 
 import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { PUBLIC_API_ERRORS, publicApiError } from '@/lib/public-api-errors';
 
 export const dynamic = 'force-dynamic';
 
 const getCachedServerLog = cachedQuery((id: number) => getServerLog(getDb(), id), 'server-log', {
   blobOnly: true,
 });
+const getCachedNamedServerLog = cachedQuery(
+  (id: number, fileName: string) => getServerLog(getDb(), id, fileName),
+  'server-log-file',
+  { blobOnly: true },
+);
+const getCachedServerLogChunk = cachedQuery(
+  (id: number, offset: number, limit: number, fileName: string | null) =>
+    getServerLogChunk(getDb(), id, offset, limit, fileName ?? undefined),
+  'server-log-chunk',
+);
+
+export const DEFAULT_SERVER_LOG_CHUNK_SIZE = 64 * 1024;
+export const MAX_SERVER_LOG_CHUNK_SIZE = 256 * 1024;
+const MAX_SERVER_LOG_OFFSET = 2_000_000_000;
+const MAX_SERVER_LOG_FILE_NAME_LENGTH = 1024;
+
+function downloadFileName(fileName: string): string {
+  const leafName = fileName.replaceAll('\\', '/').split('/').at(-1) || 'server.log';
+  const asciiName = [...leafName]
+    .map((character) => {
+      const codePoint = character.codePointAt(0) ?? 0;
+      return codePoint >= 32 && codePoint <= 126 && character !== '"' && character !== '\\'
+        ? character
+        : '_';
+    })
+    .join('')
+    .slice(0, 180);
+  const encodedName = encodeURIComponent(leafName).replaceAll(
+    /[!'()*]/gu,
+    (character) => `%${character.codePointAt(0)?.toString(16).toUpperCase() ?? ''}`,
+  );
+  return `attachment; filename="${asciiName || 'server.log'}"; filename*=UTF-8''${encodedName}`;
+}
+
+function parseIntegerParam(
+  value: string | null,
+  fallback: number,
+  minimum: number,
+  maximum: number,
+): number | null {
+  if (value === null) return fallback;
+  if (!/^\d+$/u.test(value)) return null;
+  const parsed = Number(value);
+  return Number.isSafeInteger(parsed) && parsed >= minimum && parsed <= maximum ? parsed : null;
+}
 
 export async function GET(request: NextRequest) {
   const id = Number(request.nextUrl.searchParams.get('id'));
 
   if (!id || !Number.isFinite(id)) {
-    return NextResponse.json({ error: 'id is required (benchmark_result_id)' }, { status: 400 });
+    return publicApiError(PUBLIC_API_ERRORS.idRequired, 400);
   }
 
   try {
-    const serverLog = await getCachedServerLog(id);
+    const fileName = request.nextUrl.searchParams.get('file');
+    if (
+      fileName !== null &&
+      (fileName.length === 0 ||
+        fileName.length > MAX_SERVER_LOG_FILE_NAME_LENGTH ||
+        fileName.includes('\u0000'))
+    ) {
+      return NextResponse.json({ error: 'file is invalid' }, { status: 400 });
+    }
+    const wantsChunk =
+      request.nextUrl.searchParams.has('offset') || request.nextUrl.searchParams.has('limit');
+    const download = request.nextUrl.searchParams.get('download');
+    if ((download !== null && download !== '1') || (download === '1' && wantsChunk)) {
+      return NextResponse.json(
+        { error: 'download must be 1 and cannot be combined with offset or limit' },
+        { status: 400 },
+      );
+    }
+
+    if (download === '1') {
+      const firstChunk = await getCachedServerLogChunk(id, 0, MAX_SERVER_LOG_CHUNK_SIZE, fileName);
+      if (firstChunk === null) {
+        return publicApiError(PUBLIC_API_ERRORS.notFound, 404);
+      }
+      const encoder = new TextEncoder();
+      let nextOffset = firstChunk.nextOffset;
+      const stream = new ReadableStream<Uint8Array>({
+        start(controller) {
+          if (firstChunk.serverLog.length > 0) {
+            controller.enqueue(encoder.encode(firstChunk.serverLog));
+          }
+          if (nextOffset === null) controller.close();
+        },
+        async pull(controller) {
+          if (nextOffset === null) return;
+          try {
+            const chunk = await getCachedServerLogChunk(
+              id,
+              nextOffset,
+              MAX_SERVER_LOG_CHUNK_SIZE,
+              fileName,
+            );
+            if (chunk === null) {
+              controller.error(new Error('Log file disappeared during download'));
+              return;
+            }
+            if (chunk.serverLog.length > 0) {
+              controller.enqueue(encoder.encode(chunk.serverLog));
+            }
+            nextOffset = chunk.nextOffset;
+            if (nextOffset === null) controller.close();
+          } catch (error) {
+            controller.error(error);
+          }
+        },
+      });
+      return new Response(stream, {
+        headers: {
+          'Cache-Control': 'private, no-cache, no-store, max-age=0',
+          'Content-Disposition': downloadFileName(fileName ?? 'server.log'),
+          'Content-Type': 'text/plain; charset=utf-8',
+          'X-Content-Type-Options': 'nosniff',
+        },
+      });
+    }
+
+    if (wantsChunk) {
+      const offset = parseIntegerParam(
+        request.nextUrl.searchParams.get('offset'),
+        0,
+        0,
+        MAX_SERVER_LOG_OFFSET,
+      );
+      const limit = parseIntegerParam(
+        request.nextUrl.searchParams.get('limit'),
+        DEFAULT_SERVER_LOG_CHUNK_SIZE,
+        1,
+        MAX_SERVER_LOG_CHUNK_SIZE,
+      );
+      if (offset === null || limit === null) {
+        return NextResponse.json(
+          {
+            error: `offset must be 0-${MAX_SERVER_LOG_OFFSET}; limit must be 1-${MAX_SERVER_LOG_CHUNK_SIZE}`,
+          },
+          { status: 400 },
+        );
+      }
+
+      const chunk = await getCachedServerLogChunk(id, offset, limit, fileName);
+      if (chunk === null) {
+        return publicApiError(PUBLIC_API_ERRORS.notFound, 404);
+      }
+      return cachedJson({ id, ...chunk });
+    }
+
+    const serverLog =
+      fileName === null
+        ? await getCachedServerLog(id)
+        : await getCachedNamedServerLog(id, fileName);
 
     if (serverLog === null) {
-      return NextResponse.json({ error: 'Not found' }, { status: 404 });
+      return publicApiError(PUBLIC_API_ERRORS.notFound, 404);
     }
 
-    return cachedJson({ id, serverLog });
+    return cachedJson(fileName === null ? { id, serverLog } : { id, fileName, serverLog });
   } catch (error) {
     console.error('Error fetching server log:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+    return publicApiError(PUBLIC_API_ERRORS.internal, 500);
   }
 }
diff --git a/packages/app/src/app/api/v1/trace-server-metric-source/route.ts b/packages/app/src/app/api/v1/trace-server-metric-source/route.ts
new file mode 100644
index 000000000..ed4bf8b83
--- /dev/null
+++ b/packages/app/src/app/api/v1/trace-server-metric-source/route.ts
@@ -0,0 +1,43 @@
+import { type NextRequest, NextResponse } from 'next/server';
+
+import { getDb } from '@semianalysisai/inferencex-db/connection';
+import { CHART_SERIES_VERSION } from '@semianalysisai/inferencex-db/etl/compute-chart-series';
+import {
+  getTraceServerMetricSource,
+  type MetricSourceSeries,
+} from '@semianalysisai/inferencex-db/queries/trace-server-metrics';
+
+import { cachedJson, cachedQuery } from '@/lib/api-cache';
+
+export const dynamic = 'force-dynamic';
+export const maxDuration = 300;
+
+/** Version-derived blob-cache key namespace (exported for the key-derivation test). */
+export const CACHE_KEY_PREFIX = `trace-server-metric-source-v${CHART_SERIES_VERSION}`;
+
+const getCachedMetricSource = cachedQuery(
+  (id: number, source: string): Promise<MetricSourceSeries | null> =>
+    getTraceServerMetricSource(getDb(), id, source),
+  CACHE_KEY_PREFIX,
+  { blobOnly: true },
+);
+
+/** GET /api/v1/trace-server-metric-source?id=N&source=SOURCE_ID */
+export async function GET(request: NextRequest): Promise<Response> {
+  const id = Number(request.nextUrl.searchParams.get('id'));
+  const source = request.nextUrl.searchParams.get('source')?.trim();
+  if (!id || !Number.isFinite(id)) {
+    return NextResponse.json({ error: 'id is required (benchmark_result_id)' }, { status: 400 });
+  }
+  if (!source || source.length > 512) {
+    return NextResponse.json({ error: 'source is required' }, { status: 400 });
+  }
+  try {
+    const data = await getCachedMetricSource(id, source);
+    if (!data) return NextResponse.json({ error: 'Not found' }, { status: 404 });
+    return cachedJson(data);
+  } catch (error) {
+    console.error('Error fetching trace server metric source:', error);
+    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+  }
+}
diff --git a/packages/app/src/app/api/v1/trace-server-metrics/route.ts b/packages/app/src/app/api/v1/trace-server-metrics/route.ts
index 1334b580e..a2adf3d54 100644
--- a/packages/app/src/app/api/v1/trace-server-metrics/route.ts
+++ b/packages/app/src/app/api/v1/trace-server-metrics/route.ts
@@ -12,6 +12,15 @@ import { idQueryRoute } from '../id-routes';
 
 export const dynamic = 'force-dynamic';
 
+// The slow path re-parses `server_metrics_json_gz`, which reaches 448 MB
+// compressed (several GB of JSON) on the largest agentic points and takes
+// minutes to stream. That only happens for rows whose stored `chart_series`
+// predates the current version — i.e. after a version bump, until
+// `db:backfill-chart-series` drains them — but the platform default would
+// cut those requests off mid-parse and the blob cache only populates on
+// success, so every visitor would re-pay. Ask for the maximum instead.
+export const maxDuration = 300;
+
 // Key derived from TRACE_SERVER_METRICS_VERSION (governs chart_series plus
 // the separately queried point-metadata payload).
 // The blob cache is write-once with no post-backfill purge, so the
diff --git a/packages/app/src/app/api/v1/workflow-info/route.ts b/packages/app/src/app/api/v1/workflow-info/route.ts
index d8d646e97..075c472e0 100644
--- a/packages/app/src/app/api/v1/workflow-info/route.ts
+++ b/packages/app/src/app/api/v1/workflow-info/route.ts
@@ -1,4 +1,4 @@
-import { type NextRequest, NextResponse } from 'next/server';
+import type { NextRequest } from 'next/server';
 
 import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
 
@@ -10,6 +10,7 @@ import {
 } from '@semianalysisai/inferencex-db/queries/workflow-info';
 
 import { cachedJson, cachedQuery } from '@/lib/api-cache';
+import { PUBLIC_API_ERRORS, publicApiError } from '@/lib/public-api-errors';
 import { loadFixture } from '@/lib/test-fixtures';
 
 export const dynamic = 'force-dynamic';
@@ -25,16 +26,9 @@ async function loadWorkflowInfo(date: string, benchmarkType?: 'agentic_traces')
   return { runs, changelogs, configs, runConfigs };
 }
 
-// Preserve the established public cache and response for calls without a
-// scenario. Scenario-scoped calls use a separate cache namespace.
-const getCachedWorkflowInfo = cachedQuery(
-  (date: string) => loadWorkflowInfo(date),
-  'workflow-info',
-);
-const getCachedScenarioWorkflowInfo = cachedQuery(
-  (date: string) => loadWorkflowInfo(date, 'agentic_traces'),
-  'workflow-info-scenario',
-);
+// Scenario is an explicit cache dimension: fixed-sequence and Agentic Traces
+// requests for the same date must never share a cached runConfigs payload.
+const getCachedWorkflowInfo = cachedQuery(loadWorkflowInfo, 'workflow-info-v2');
 
 export async function GET(request: NextRequest) {
   const date = request.nextUrl.searchParams.get('date') ?? '';
@@ -43,20 +37,15 @@ export async function GET(request: NextRequest) {
       ? 'agentic_traces'
       : undefined;
   if (date && !/^\d{4}-\d{2}-\d{2}$/u.test(date)) {
-    return NextResponse.json(
-      { error: 'Invalid date format (YYYY-MM-DD required)' },
-      { status: 400 },
-    );
+    return publicApiError(PUBLIC_API_ERRORS.invalidDate, 400);
   }
   if (FIXTURES_MODE) return cachedJson(loadFixture('workflow-info'));
 
   try {
-    const data = benchmarkType
-      ? await getCachedScenarioWorkflowInfo(date)
-      : await getCachedWorkflowInfo(date);
+    const data = await getCachedWorkflowInfo(date, benchmarkType);
     return cachedJson(data);
   } catch (error) {
     console.error('Error fetching workflow info:', error);
-    return NextResponse.json({ error: 'Internal server error' }, { status: 500 });
+    return publicApiError(PUBLIC_API_ERRORS.internal, 500);
   }
 }
diff --git a/packages/app/src/app/blog/[slug]/og-image-render.tsx b/packages/app/src/app/blog/[slug]/og-image-render.tsx
index 4d1e0732e..f758a6063 100644
--- a/packages/app/src/app/blog/[slug]/og-image-render.tsx
+++ b/packages/app/src/app/blog/[slug]/og-image-render.tsx
@@ -43,7 +43,16 @@ async function loadTiles() {
   return TILE_GRID.map((t) => (t ? { src: cache[t.file], rotate: t.rotate } : null));
 }
 
-export async function renderOgImage(meta: BlogPostMeta) {
+export async function renderOgImage(
+  meta: Pick<BlogPostMeta, 'title' | 'subtitle' | 'date'>,
+  opts?: {
+    /**
+     * Pre-formatted footer label. When set, `meta.date` is ignored — used by
+     * non-blog reusers (e.g. /model pages) whose dates are free-form text.
+     */
+    dateLabel?: string;
+  },
+) {
   const [logoSrc, tiles] = await Promise.all([
     readFile(join(process.cwd(), 'public/brand/logo-color.png')).then(
       (buf) => `data:image/png;base64,${buf.toString('base64')}`,
@@ -174,12 +183,13 @@ export async function renderOgImage(meta: BlogPostMeta) {
 
         <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'center' }}>
           <span style={{ fontSize: 28, color: '#d4d4d8' }}>
-            {new Date(`${meta.date}T00:00:00Z`).toLocaleDateString('en-US', {
-              year: 'numeric',
-              month: 'long',
-              day: 'numeric',
-              timeZone: 'UTC',
-            })}
+            {opts?.dateLabel ??
+              new Date(`${meta.date}T00:00:00Z`).toLocaleDateString('en-US', {
+                year: 'numeric',
+                month: 'long',
+                day: 'numeric',
+                timeZone: 'UTC',
+              })}
           </span>
           <img src={logoSrc} height={80} alt="SemiAnalysis" />
         </div>
diff --git a/packages/app/src/app/blog/[slug]/page.tsx b/packages/app/src/app/blog/[slug]/page.tsx
index 92baa0d65..4fa7a7bc2 100644
--- a/packages/app/src/app/blog/[slug]/page.tsx
+++ b/packages/app/src/app/blog/[slug]/page.tsx
@@ -1,12 +1,5 @@
 import type { Metadata } from 'next';
 import { notFound } from 'next/navigation';
-import { compileMDX } from 'next-mdx-remote/rsc';
-import rehypeShikiFromHighlighter from '@shikijs/rehype/core';
-import rehypeKatex from 'rehype-katex';
-import remarkGfm from 'remark-gfm';
-import remarkMath from 'remark-math';
-import { createHighlighterCore } from 'shiki/core';
-import { createOnigurumaEngine } from 'shiki/engine/oniguruma';
 
 import 'katex/dist/katex.min.css';
 
@@ -14,7 +7,6 @@ import { BlogBackLink } from '@/components/blog/blog-back-link';
 import { BlogPostNav } from '@/components/blog/blog-post-nav';
 import { BlogToc } from '@/components/blog/blog-toc';
 import { HashScroll } from '@/components/blog/hash-scroll';
-import { createMdxComponents } from '@/components/blog/mdx-components';
 import { ReadingProgressBar } from '@/components/blog/reading-progress-bar';
 import { ShareTwitterButton, ShareLinkedInButton } from '@/components/share-buttons';
 import { Card } from '@/components/ui/card';
@@ -29,6 +21,7 @@ import {
   getPostBySlug,
   hasZhTranslation,
 } from '@/lib/blog';
+import { compileBlogMdx } from '@/lib/blog-mdx';
 import { languageAlternates } from '@/lib/i18n';
 import {
   AUTHOR_HANDLE,
@@ -85,33 +78,6 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   };
 }
 
-let highlighterPromise: ReturnType<typeof createHighlighterCore> | null = null;
-
-function getHighlighter() {
-  if (!highlighterPromise) {
-    highlighterPromise = createHighlighterCore({
-      themes: [import('shiki/themes/github-dark.mjs'), import('shiki/themes/github-light.mjs')],
-      langs: [
-        import('shiki/langs/typescript.mjs'),
-        import('shiki/langs/javascript.mjs'),
-        import('shiki/langs/python.mjs'),
-        import('shiki/langs/bash.mjs'),
-        import('shiki/langs/json.mjs'),
-        import('shiki/langs/yaml.mjs'),
-        import('shiki/langs/css.mjs'),
-        import('shiki/langs/html.mjs'),
-        import('shiki/langs/tsx.mjs'),
-        import('shiki/langs/jsx.mjs'),
-        import('shiki/langs/sql.mjs'),
-        import('shiki/langs/go.mjs'),
-        import('shiki/langs/rust.mjs'),
-      ],
-      engine: createOnigurumaEngine(import('shiki/wasm')),
-    });
-  }
-  return highlighterPromise;
-}
-
 export default async function BlogPostPage({ params }: Props) {
   const { slug } = await params;
   const result = getPostBySlug(slug);
@@ -120,32 +86,7 @@ export default async function BlogPostPage({ params }: Props) {
   const { meta, raw } = result;
   const adjacent = getAdjacentPosts(slug);
   const headings = extractHeadings(raw);
-  const highlighter = await getHighlighter();
-
-  const { content } = await compileMDX({
-    source: raw,
-    components: createMdxComponents(),
-    options: {
-      mdxOptions: {
-        // `singleDollarTextMath: false` keeps prices like $1.95/GPU/hr from being parsed
-        // as inline math; math must be delimited with `$$`.
-        remarkPlugins: [remarkGfm, [remarkMath, { singleDollarTextMath: false }]],
-        rehypePlugins: [
-          // `strict: false` keeps a single malformed expression from failing the whole build;
-          // KaTeX renders it inline in red instead.
-          [rehypeKatex, { strict: false }],
-          [
-            rehypeShikiFromHighlighter,
-            highlighter,
-            {
-              themes: { dark: 'github-dark', light: 'github-light' },
-              defaultColor: false,
-            },
-          ],
-        ],
-      },
-    },
-  });
+  const { content } = await compileBlogMdx(raw);
 
   const jsonLd = buildBlogPostingJsonLd(meta, raw);
   const breadcrumbJsonLd = buildBlogBreadcrumbJsonLd(slug, meta.title);
diff --git a/packages/app/src/app/blog/page.tsx b/packages/app/src/app/blog/page.tsx
index bad1c01c1..b2a98ea2b 100644
--- a/packages/app/src/app/blog/page.tsx
+++ b/packages/app/src/app/blog/page.tsx
@@ -11,11 +11,11 @@ import { SITE_URL, SITE_NAME, AUTHOR_NAME } from '@semianalysisai/inferencex-con
 
 export const metadata: Metadata = {
   title: 'Articles',
-  description: `Technical articles from ${SITE_NAME} by ${AUTHOR_NAME} — AI inference benchmarking, chip performance analysis, and ML infrastructure insights.`,
+  description: `Technical articles from ${SITE_NAME} by ${AUTHOR_NAME} on agentic inference benchmarks, AgentX results, chip performance, and ML infrastructure.`,
   alternates: enAlternates('/blog'),
   openGraph: {
     title: `Articles | ${SITE_NAME} by ${AUTHOR_NAME}`,
-    description: 'AI inference benchmarking insights and chip performance analysis.',
+    description: 'Articles on agentic inference benchmarks, AgentX results, and chip performance.',
     url: `${SITE_URL}/blog`,
   },
 };
@@ -49,7 +49,8 @@ export default async function BlogPage({
           <Card>
             <h2 className="text-2xl lg:text-4xl font-bold tracking-tight">Articles</h2>
             <p className="mt-3 text-base lg:text-lg text-muted-foreground">
-              Insights on AI inference benchmarking, chip performance, and ML infrastructure.
+              Articles on agentic inference, AgentX results, chip performance, and ML
+              infrastructure.
             </p>
             <p className="text-sm text-muted-foreground">
               New to the terminology?{' '}
diff --git a/packages/app/src/app/chips/[slug]/page.tsx b/packages/app/src/app/chips/[slug]/page.tsx
new file mode 100644
index 000000000..9165e7f12
--- /dev/null
+++ b/packages/app/src/app/chips/[slug]/page.tsx
@@ -0,0 +1,113 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { ChipDetailContent, ChipVsContent } from '@/components/chips/chip-page-sections';
+import {
+  getAllChipRouteSlugs,
+  getChipPage,
+  getChipVsPage,
+  type ChipVsPage,
+} from '@/lib/chip-pages';
+import { enAlternates } from '@/lib/i18n';
+import {
+  AUTHOR_HANDLE,
+  AUTHOR_NAME,
+  SITE_NAME,
+  SITE_URL,
+  SUPPORTERS_LINE,
+} from '@semianalysisai/inferencex-constants';
+
+interface Props {
+  params: Promise<{ slug: string }>;
+}
+
+export const dynamicParams = false;
+
+export function generateStaticParams() {
+  return getAllChipRouteSlugs().map((slug) => ({ slug }));
+}
+
+function vsDescription(page: ChipVsPage): string {
+  return `${page.a.title} vs ${page.b.title}: memory, bandwidth, FP8/FP4 compute, TDP and cloud pricing side by side, with continuously measured LLM inference benchmarks on identical workloads. ${SUPPORTERS_LINE}`;
+}
+
+function vsKeywords(page: ChipVsPage): string[] {
+  const a = page.a.label;
+  const b = page.b.label;
+  return [
+    `${a} vs ${b}`,
+    `${b} vs ${a}`,
+    `${a} vs ${b} benchmark`,
+    `${a} vs ${b} specs`,
+    `${a} vs ${b} inference performance`,
+    `${a} vs ${b} price`,
+    `${a} GPU comparison`,
+    'AI inference benchmarks',
+  ];
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug } = await params;
+
+  const chip = getChipPage(slug);
+  if (chip) {
+    const title = `${chip.title} Specs, Pricing & AI Inference Benchmarks`;
+    const description = `${chip.summary} ${SUPPORTERS_LINE}`;
+    const url = `${SITE_URL}/chips/${chip.slug}`;
+    return {
+      title,
+      description,
+      keywords: [...chip.keywords],
+      authors: [{ name: AUTHOR_NAME }],
+      alternates: enAlternates(`/chips/${chip.slug}`),
+      openGraph: {
+        title: `${title} | ${SITE_NAME}`,
+        description,
+        url,
+        type: 'article',
+      },
+      twitter: {
+        card: 'summary_large_image',
+        title: `${title} | ${SITE_NAME}`,
+        description,
+        creator: AUTHOR_HANDLE,
+      },
+    };
+  }
+
+  const vsPage = getChipVsPage(slug);
+  if (!vsPage) return {};
+  const title = `${vsPage.a.label} vs ${vsPage.b.label}: Specs, Power & AI Inference Comparison`;
+  const description = vsDescription(vsPage);
+  const url = `${SITE_URL}/chips/${vsPage.slug}`;
+  return {
+    title,
+    description,
+    keywords: vsKeywords(vsPage),
+    authors: [{ name: AUTHOR_NAME }],
+    alternates: enAlternates(`/chips/${vsPage.slug}`),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+      type: 'article',
+    },
+    twitter: {
+      card: 'summary_large_image',
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      creator: AUTHOR_HANDLE,
+    },
+  };
+}
+
+export default async function ChipRoutePage({ params }: Props) {
+  const { slug } = await params;
+
+  const chip = getChipPage(slug);
+  if (chip) return <ChipDetailContent entry={chip} locale="en" />;
+
+  const vsPage = getChipVsPage(slug);
+  if (!vsPage) notFound();
+  return <ChipVsContent page={vsPage} locale="en" />;
+}
diff --git a/packages/app/src/app/chips/page.tsx b/packages/app/src/app/chips/page.tsx
new file mode 100644
index 000000000..bfb646858
--- /dev/null
+++ b/packages/app/src/app/chips/page.tsx
@@ -0,0 +1,39 @@
+import type { Metadata } from 'next';
+
+import { ChipsIndexContent } from '@/components/chips/chip-page-sections';
+import { enAlternates } from '@/lib/i18n';
+import { SITE_NAME, SITE_URL, SUPPORTERS_LINE } from '@semianalysisai/inferencex-constants';
+
+const TITLE = 'AI Chips for LLM Inference: Specs, Pricing & Benchmarks';
+const DESCRIPTION = `Specs, cloud pricing and continuously measured LLM inference benchmarks for NVIDIA Hopper and Blackwell (H100, H200, B200, B300, GB200 NVL72, GB300 NVL72) and AMD Instinct (MI300X, MI325X, MI355X) chips. ${SUPPORTERS_LINE}`;
+
+export const metadata: Metadata = {
+  title: TITLE,
+  description: DESCRIPTION,
+  keywords: [
+    'AI chips',
+    'AI GPU comparison',
+    'GPU specs',
+    'GPU price per hour',
+    'LLM inference hardware',
+    'NVIDIA vs AMD GPU',
+    'data center GPU pricing',
+    'AI accelerator comparison',
+  ],
+  alternates: enAlternates('/chips'),
+  openGraph: {
+    title: `${TITLE} | ${SITE_NAME}`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/chips`,
+    type: 'website',
+  },
+  twitter: {
+    card: 'summary_large_image',
+    title: `${TITLE} | ${SITE_NAME}`,
+    description: DESCRIPTION,
+  },
+};
+
+export default function ChipsIndexPage() {
+  return <ChipsIndexContent locale="en" />;
+}
diff --git a/packages/app/src/app/compare-per-dollar/[slug]/[scenario]/page.tsx b/packages/app/src/app/compare-per-dollar/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..ad17ae939
--- /dev/null
+++ b/packages/app/src/app/compare-per-dollar/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildPerDollarMetadata, renderPerDollarPage } from '../page';
+
+/**
+ * `/compare-per-dollar/<slug>/<scenario>` — the same comparison with its workload pinned by
+ * the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments rather than silently falling back to the pair's
+ *  default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildPerDollarMetadata(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function PerDollarScenarioPage({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderPerDollarPage(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/compare-per-dollar/[slug]/opengraph-image.tsx b/packages/app/src/app/compare-per-dollar/[slug]/opengraph-image.tsx
index 21bbd5642..b4f666157 100644
--- a/packages/app/src/app/compare-per-dollar/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/compare-per-dollar/[slug]/opengraph-image.tsx
@@ -1,201 +1,32 @@
-/**
- * Per-dollar compare OG image — identical circuit tile sidebar layout to the
- * /compare OG and the blog OG, but the eyebrow and bottom banner are swapped
- * to surface the "Performance per Dollar" framing.
- */
-import { readFile } from 'node:fs/promises';
-import { join } from 'node:path';
-
 import { notFound } from 'next/navigation';
-import { ImageResponse } from 'next/og';
 
 import { getAllComparableCompareSlugs } from '@/lib/compare-availability';
+import { renderCompareOg } from '@/lib/compare-og';
 import { canonicalCompareSlug, compareDisplayLabel, parseCompareSlug } from '@/lib/compare-slug';
 
 export const alt = 'Chip performance-per-dollar inference benchmark comparison';
 export const size = { width: 1200, height: 630 };
 export const contentType = 'image/png';
 
-const BLUE = '#0B86D1';
-const BG = '#131416';
-const PANEL_BG = '#0F1214';
-
-// Same tile grid as the /compare and blog OG so all three OG types read as a
-// family.
-const TILE_GRID: ({ file: string; rotate?: number } | null)[] = [
-  { file: 'teal-chevron.png', rotate: 180 },
-  { file: 'gold-diagonal.png' },
-  { file: 'teal-circuit.png' },
-  null,
-  { file: 'gold-wavy.png' },
-  { file: 'teal-chip.png' },
-  { file: 'teal-chevron.png', rotate: 90 },
-  { file: 'teal-organic.png' },
-  null,
-  { file: 'gold-circuit.png' },
-  { file: 'teal-circuit.png', rotate: 180 },
-  { file: 'teal-organic.png', rotate: 180 },
-];
-
 export async function generateStaticParams() {
   const slugs = await getAllComparableCompareSlugs();
   return slugs.map(({ modelSlug, a, b }) => ({ slug: canonicalCompareSlug(modelSlug, a, b) }));
 }
 
-let logoSrcPromise: Promise<string | null> | undefined;
-function getLogoSrc(): Promise<string | null> {
-  if (!logoSrcPromise) {
-    logoSrcPromise = readFile(join(process.cwd(), 'public/brand/logo-color.png'))
-      .then((buf) => `data:image/png;base64,${buf.toString('base64')}`)
-      .catch(() => null);
-  }
-  return logoSrcPromise;
-}
-
-let tilesPromise: Promise<({ src: string; rotate?: number } | null)[]> | undefined;
-function getTiles(): Promise<({ src: string; rotate?: number } | null)[]> {
-  if (!tilesPromise) {
-    const uniqueFiles = [...new Set(TILE_GRID.filter(Boolean).map((t) => t!.file))];
-    tilesPromise = Promise.all(
-      uniqueFiles.map(async (f) => {
-        const src = await readFile(join(process.cwd(), 'public/brand/og-tiles', f))
-          .then((buf) => `data:image/png;base64,${buf.toString('base64')}`)
-          .catch(() => null);
-        return [f, src] as const;
-      }),
-    ).then((loaded) => {
-      const cache = Object.fromEntries(loaded);
-      return TILE_GRID.map((t) => {
-        if (!t) return null;
-        const src = cache[t.file];
-        return src ? { src, rotate: t.rotate } : null;
-      });
-    });
-  }
-  return tilesPromise;
-}
-
 export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
   const { slug } = await params;
   const parsed = parseCompareSlug(slug);
   if (!parsed) notFound();
-  const [logoSrc, tiles] = await Promise.all([getLogoSrc(), getTiles()]);
 
   const title = compareDisplayLabel(parsed.a, parsed.b);
-  const eyebrow = `${parsed.model.label} · Performance per Dollar`;
   const titleSize = title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
 
-  return new ImageResponse(
-    <div
-      style={{
-        display: 'flex',
-        width: '100%',
-        height: '100%',
-        backgroundColor: BG,
-        color: '#EAEBEC',
-        overflow: 'hidden',
-      }}
-    >
-      <div
-        style={{
-          display: 'flex',
-          width: 195,
-          height: '100%',
-          backgroundColor: PANEL_BG,
-          position: 'relative',
-        }}
-      >
-        {tiles.map((tile, i) => {
-          if (!tile) return null;
-          const row = Math.floor(i / 2);
-          const col = i % 2;
-          return (
-            <img
-              key={i}
-              src={tile.src}
-              style={{
-                position: 'absolute',
-                left: 12 + col * 90,
-                top: 12 + row * 104,
-                width: 78,
-                height: 86,
-                borderRadius: 4,
-                objectFit: 'cover',
-                ...(tile.rotate ? { transform: `rotate(${tile.rotate}deg)` } : {}),
-              }}
-            />
-          );
-        })}
-        <div
-          style={{
-            position: 'absolute',
-            right: 0,
-            top: 0,
-            width: 3,
-            height: '100%',
-            backgroundColor: BLUE,
-            display: 'flex',
-          }}
-        />
-      </div>
-
-      <div
-        style={{
-          display: 'flex',
-          flexDirection: 'column',
-          justifyContent: 'space-between',
-          flex: 1,
-          padding: '48px 55px 20px 55px',
-        }}
-      >
-        <div
-          style={{
-            display: 'flex',
-            flexDirection: 'column',
-            gap: 24,
-            flex: 1,
-            justifyContent: 'center',
-          }}
-        >
-          <div
-            style={{
-              fontSize: 24,
-              color: '#9BA0A6',
-              letterSpacing: '0.18em',
-              textTransform: 'uppercase',
-              display: 'flex',
-            }}
-          >
-            {eyebrow}
-          </div>
-
-          <div
-            style={{
-              fontSize: titleSize,
-              fontWeight: 800,
-              color: '#FFFFFF',
-              lineHeight: 1.1,
-              display: 'flex',
-            }}
-          >
-            {title}
-          </div>
-        </div>
-
-        <div
-          style={{
-            display: 'flex',
-            justifyContent: 'space-between',
-            alignItems: 'center',
-          }}
-        >
-          <span style={{ fontSize: 26, color: '#9BA0A6', display: 'flex' }}>
-            AI inference benchmark · cost per million tokens
-          </span>
-          {logoSrc && <img src={logoSrc} height={72} />}
-        </div>
-      </div>
-    </div>,
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · Performance per Dollar`,
+    title,
+    titleSize,
+    footer: 'AI inference benchmark · cost per million tokens',
+    language: 'en',
     size,
-  );
+  });
 }
diff --git a/packages/app/src/app/compare-per-dollar/[slug]/page-client.tsx b/packages/app/src/app/compare-per-dollar/[slug]/page-client.tsx
index 3fa64bfba..44633ff7b 100644
--- a/packages/app/src/app/compare-per-dollar/[slug]/page-client.tsx
+++ b/packages/app/src/app/compare-per-dollar/[slug]/page-client.tsx
@@ -3,35 +3,45 @@
 import Link from 'next/link';
 import { useEffect, useMemo } from 'react';
 
-import type { GPUDataPoint, InterpolatedResult } from '@/components/calculator/types';
+import { TCO_SOURCE_TITLE, TCO_SOURCE_URL } from '@semianalysisai/inferencex-constants';
+
+import type { GPUDataPoint } from '@/components/calculator/types';
 import { useThroughputData } from '@/components/calculator/useThroughputData';
-import { CompareInterpolatedTable } from '@/components/compare/compare-interpolated-table';
-import { useGlobalFilters, GlobalFilterProvider } from '@/components/GlobalFilterContext';
+import {
+  CompareTableRenderer,
+  type CompareTableData,
+} from '@/components/compare/compare-table-renderer';
+import {
+  GlobalFilterProvider,
+  useGlobalFilterAvailability,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+} from '@/components/GlobalFilterContext';
 import { InferenceProvider } from '@/components/inference/InferenceContext';
 import InferenceChartDisplay from '@/components/inference/ui/ChartDisplay';
 import { Card } from '@/components/ui/card';
 import { track } from '@/lib/analytics';
+import type { BenchmarkRow } from '@/lib/api';
+import { toCalculatorBenchmarkRows } from '@/lib/benchmark-api-view';
 import { toModel, toPrecisions, toSequence } from '@/lib/compare-enum-coerce';
-
-interface SsrTableData {
-  defaultTargets: number[];
-  ssrRows: { target: number; a: InterpolatedResult | null; b: InterpolatedResult | null }[];
-  interactivityRange: { min: number; max: number };
-}
+import type { AgenticScenarioIntro } from '@/lib/compare-ssr';
 
 /** Only show Cost + Concurrency in the interpolated table — the rest of the
  *  metric rows (Throughput, tok/s/MW) live on the sibling /compare page. */
 const PER_DOLLAR_TABLE_METRICS = ['Cost ($/M tok)', 'Concurrency'];
 
-/** Rename "Cost ($/M tok)" to the full-English "Dollar per Million Tokens"
- *  in the per-dollar table so the cell reads in line with the page's
- *  "Performance per Dollar" framing and surfaces the SEO term verbatim. */
+/** Localized display labels keyed by the table's stable English metric key. */
 const PER_DOLLAR_LABEL_OVERRIDES = {
-  'Cost ($/M tok)': 'Dollar per Million Tokens',
-};
+  en: {
+    'Cost ($/M tok)': 'Dollar per Million Tokens',
+  },
+  zh: {
+    'Cost ($/M tok)': '每百万 token 美元成本',
+  },
+} as const;
 
 /** y_costh = Cost per Million Total Tokens (Owning - Hyperscaler). Defined in
- *  packages/app/src/components/inference/inference-chart-config.json. */
+ *  packages/app/src/components/inference/metric-registry.ts. */
 const PER_DOLLAR_DEFAULT_Y_AXIS = 'y_costh';
 
 const STRINGS = {
@@ -54,10 +64,10 @@ const STRINGS = {
     fullComparisonLinkText: '查看完整延迟与吞吐量对比 →',
     caveatSeqFallback: '序列',
     caveatPrecFallback: '精度',
-    pricingLabel: 'Chip 定价（所属云服务商）：',
+    pricingLabel: '芯片定价（所属云服务商）：',
     pricingSource: '来源：',
     emptyState:
-      '当前默认模型在此 Chip 组合上没有可用的插值每 token 成本数据。请使用下方图表控件选择一个两款 Chip 均有基准测试数据的模型和精度。',
+      '当前默认模型在此芯片组合上没有可用的插值每 token 成本数据。请使用下方图表控件选择一个两款芯片均有基准测试数据的模型和精度。',
   },
 } as const;
 
@@ -72,13 +82,17 @@ interface ComparePerDollarPageClientProps {
   defaultModel: string;
   defaultSequence: string | null;
   defaultPrecision: string | null;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
+  initialBenchmarkRows?: BenchmarkRow[];
   /** One SSR-rendered prose paragraph per interpolated-table row (default
    *  interactivity target). Each paragraph picks a template variant
    *  deterministically from the slug so prose stays stable across renders
    *  but varies across pages in the catalog. Empty array when there's no
    *  comparable data. */
   narrative: string[];
+  /** Set only when the page is rendering the agentic workload. Explains
+   *  what AgentX measures before the head-to-head numbers. */
+  agenticIntro?: AgenticScenarioIntro | null;
   aLabel: string;
   bLabel: string;
   aVendor: string;
@@ -105,7 +119,9 @@ export default function ComparePerDollarPageClient({
   defaultSequence,
   defaultPrecision,
   ssrTableData,
+  initialBenchmarkRows,
   narrative,
+  agenticIntro = null,
   aLabel,
   bLabel,
   aVendor,
@@ -125,6 +141,14 @@ export default function ComparePerDollarPageClient({
   const initialModel = toModel(defaultModel);
   const initialSequence = toSequence(defaultSequence);
   const initialPrecisions = toPrecisions(defaultPrecision);
+  const benchmarkQueryScope = `compare-pair:${a}:${b}`;
+  const initialCalculatorRows = useMemo(
+    () =>
+      initialBenchmarkRows && defaultSequence
+        ? toCalculatorBenchmarkRows(initialBenchmarkRows, defaultSequence)
+        : undefined,
+    [defaultSequence, initialBenchmarkRows],
+  );
   const t = STRINGS[locale];
   const isZh = locale === 'zh';
 
@@ -139,6 +163,9 @@ export default function ComparePerDollarPageClient({
         initialActiveHwTypes={[a, b]}
         compareGpuPair={compareGpuPair}
         initialYAxisMetric={PER_DOLLAR_DEFAULT_Y_AXIS}
+        benchmarkQueryScope={benchmarkQueryScope}
+        initialBenchmarkModel={initialModel}
+        initialBenchmarkRows={initialBenchmarkRows}
       >
         <div className="flex flex-col gap-4">
           <Card className="flex w-full min-w-0 flex-col gap-3">
@@ -175,6 +202,21 @@ export default function ComparePerDollarPageClient({
                   .
                 </p>
               )}
+              {agenticIntro && (
+                <p
+                  className="mt-3 max-w-3xl text-sm text-foreground/80"
+                  data-testid="compare-agentic-intro"
+                >
+                  {agenticIntro.paragraph}{' '}
+                  <Link
+                    href={agenticIntro.href}
+                    data-testid="compare-agentic-intro-link"
+                    className="font-medium text-brand underline underline-offset-4 hover:no-underline"
+                  >
+                    {agenticIntro.linkLabel} →
+                  </Link>
+                </p>
+              )}
               {narrative.length > 0 && (
                 <div
                   className="mt-3 flex flex-col gap-2"
@@ -208,13 +250,13 @@ export default function ComparePerDollarPageClient({
                   {bCostPerGpuHr > 0 ? `$${bCostPerGpuHr.toFixed(2)}/chip/hr` : '—'}.{' '}
                   {t.pricingSource}{' '}
                   <a
-                    href="https://semianalysis.com/ai-cloud-tco-model/"
+                    href={TCO_SOURCE_URL}
                     target="_blank"
                     rel="noopener noreferrer"
                     className="underline hover:text-primary"
                     onClick={() => track('compare_per_dollar_tco_source_clicked', { slug })}
                   >
-                    SemiAnalysis Market July 2026 Pricing Surveys &amp; AI Cloud TCO Model
+                    {TCO_SOURCE_TITLE}
                   </a>
                   .
                 </p>
@@ -258,7 +300,11 @@ export default function ComparePerDollarPageClient({
               aLabel={aLabel}
               bLabel={bLabel}
               ssrTableData={ssrTableData}
+              initialModel={initialModel}
+              initialSequence={initialSequence}
+              initialCalculatorRows={initialCalculatorRows}
               emptyStateText={t.emptyState}
+              isZh={isZh}
             />
           </Card>
           <InferenceChartDisplay />
@@ -274,23 +320,48 @@ function CompareTableSection({
   aLabel,
   bLabel,
   ssrTableData,
+  initialModel,
+  initialSequence,
+  initialCalculatorRows,
   emptyStateText,
+  isZh,
 }: {
   a: string;
   b: string;
   aLabel: string;
   bLabel: string;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
+  initialModel: ReturnType<typeof toModel>;
+  initialSequence: ReturnType<typeof toSequence>;
+  initialCalculatorRows?: BenchmarkRow[];
   emptyStateText: string;
+  isZh: boolean;
 }) {
-  const { effectiveSequence, effectivePrecisions, selectedRunDate, selectedModel } =
-    useGlobalFilters();
+  const { effectiveSequence, effectivePrecisions, selectedModel, sequenceResolved } =
+    useGlobalFilterSelection();
+  const { availableDates } = useGlobalFilterAvailability();
+  const { selectedRunDate } = useGlobalFilterRun();
+  const latestAvailableDate = availableDates.at(-1) ?? '';
+  const calculatorRunDate =
+    selectedRunDate && selectedRunDate === latestAvailableDate ? '' : selectedRunDate;
+  const initialRows =
+    initialCalculatorRows &&
+    sequenceResolved &&
+    selectedModel === initialModel &&
+    effectiveSequence === initialSequence &&
+    calculatorRunDate === ''
+      ? initialCalculatorRows
+      : undefined;
 
   const { gpuDataByGroupKey, ranges, hasData } = useThroughputData(
     selectedModel,
     effectiveSequence,
     effectivePrecisions,
-    selectedRunDate,
+    calculatorRunDate,
+    undefined,
+    undefined,
+    initialRows,
+    sequenceResolved,
   );
 
   const { pointsA, pointsB } = useMemo(() => {
@@ -306,25 +377,17 @@ function CompareTableSection({
 
   const clientRange = hasData ? ranges.interactivity : ssrTableData.interactivityRange;
 
-  if (ssrTableData.defaultTargets.length === 0) {
-    return (
-      <div className="border border-border/50 rounded-md px-4 py-3 text-sm text-muted-foreground bg-muted/30">
-        {emptyStateText}
-      </div>
-    );
-  }
-
   return (
-    <CompareInterpolatedTable
+    <CompareTableRenderer
       aLabel={aLabel}
       bLabel={bLabel}
-      ssrRows={ssrTableData.ssrRows}
-      defaultTargets={ssrTableData.defaultTargets}
+      ssrTableData={ssrTableData}
       interactivityRange={clientRange}
       gpuDataPointsA={pointsA}
       gpuDataPointsB={pointsB}
+      emptyStateText={emptyStateText}
+      metricLabelOverrides={PER_DOLLAR_LABEL_OVERRIDES[isZh ? 'zh' : 'en']}
       visibleMetricLabels={PER_DOLLAR_TABLE_METRICS}
-      metricLabelOverrides={PER_DOLLAR_LABEL_OVERRIDES}
     />
   );
 }
diff --git a/packages/app/src/app/compare-per-dollar/[slug]/page.tsx b/packages/app/src/app/compare-per-dollar/[slug]/page.tsx
index 0b3584ff2..e85cf15de 100644
--- a/packages/app/src/app/compare-per-dollar/[slug]/page.tsx
+++ b/packages/app/src/app/compare-per-dollar/[slug]/page.tsx
@@ -9,8 +9,12 @@ import {
 } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
 import { languageAlternates } from '@/lib/i18n';
-import { pickPairDefaults } from '@/lib/compare-pair-defaults';
 import {
   canonicalCompareSlug,
   compareDisplayLabel,
@@ -19,23 +23,40 @@ import {
 } from '@/lib/compare-slug';
 import { getGpuSpecs } from '@/lib/constants';
 import {
+  AGENTIC_SCENARIO_INTRO,
   buildBreadcrumbJsonLd,
   buildJsonLd,
   compareTableNarrative,
-  computeCompareTableData,
-  dateRangeForPair,
-  getCachedBenchmarks,
   KNOWN_MODELS,
   KNOWN_PRECISIONS,
   KNOWN_SEQUENCES,
   pickString,
-  summarize,
 } from '@/lib/compare-ssr';
+import {
+  getComparePageDerivedData,
+  initialCompareBenchmarkRows,
+} from '@/lib/compare-page-data.server';
 
 import ComparePerDollarPageClient from './page-client';
 
 export const dynamic = 'force-dynamic';
 
+/**
+ * `/compare-per-dollar/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than duplicated
+ * in a parallel route file so the body — redirects, JSON-LD, metadata, client
+ * props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
+function scenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/compare-per-dollar/${canonical}/${scenarioSegment}`
+    : `/compare-per-dollar/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
@@ -43,11 +64,26 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildPerDollarMetadata(slug, {});
+}
+
+export function buildPerDollarMetadata(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Metadata {
   const parsed = parseCompareSlug(slug);
   if (!parsed) return {};
   const fullLabel = compareModelDisplayLabel(parsed.model, parsed.a, parsed.b);
   const gpuLabel = compareDisplayLabel(parsed.a, parsed.b);
-  const url = `${SITE_URL}/compare-per-dollar/${canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b)}`;
+  const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const canonicalPath = scenarioPath(canonical);
+  const url = `${SITE_URL}${routePath}`;
+  const canonicalUrl = `${SITE_URL}${canonicalPath}`;
   // Description leads with the searched GPU pair + model, then the trust
   // signals (verified/reproducible, what InferenceX is, named supporters)
   // before weaving the per-dollar SEO terms ("performance per dollar", "cost
@@ -57,10 +93,8 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     title: `${fullLabel} — Performance per Dollar`,
     description,
     alternates: {
-      canonical: url,
-      languages: languageAlternates(
-        `/compare-per-dollar/${canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b)}`,
-      ),
+      canonical: canonicalUrl,
+      languages: languageAlternates(canonicalPath),
     },
     openGraph: {
       title: `${fullLabel} — Performance per Dollar | ${SITE_NAME}`,
@@ -78,11 +112,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function ComparePerDollarPage({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderPerDollarPage(slug, await searchParams, {});
+}
+
+export async function renderPerDollarPage(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parseCompareSlug(slug);
   if (!parsed) notFound();
 
-  const sp = await searchParams;
-
   // Same one-hop 308 normalization as /compare/[slug] — bare-slug fallback,
   // alias model resolution, GPU alphabetical order — but redirect target lives
   // under /compare-per-dollar/. Query string is preserved across the hop.
@@ -98,37 +138,46 @@ export default async function ComparePerDollarPage({ params, searchParams }: Pro
       })
       .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
       .join('&');
-    permanentRedirect(`/compare-per-dollar/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // the reader to the pair's default workload instead.
+    permanentRedirect(`${scenarioPath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
-  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
-  const summaryA = summarize(rows, parsed.a);
-  const summaryB = summarize(rows, parsed.b);
-  const { sequence: pickedSequence, precision: pickedPrecision } = pickPairDefaults(
-    rows,
-    parsed.a,
-    parsed.b,
-  );
+  const fallbackSequence = null;
 
   const urlSeq = pickString(sp.i_seq);
   const urlPrec = pickString(sp.i_prec);
   const urlModel = pickString(sp.g_model);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence;
-  const effectivePrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : pickedPrecision;
   const effectiveModel =
     urlModel && KNOWN_MODELS.has(urlModel) ? urlModel : parsed.model.displayName;
-
-  const { defaultTargets, ssrRows, interactivityRange } = computeCompareTableData(
-    rows,
+  // Path beats query beats the pair's default: a scenario segment is an
+  // explicit address for one workload, so it outranks a stale `?i_seq=`.
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const requestedSequence = pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : null);
+  const requestedPrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : null;
+  const {
+    sequence: effectiveSequence,
+    precision: effectivePrecision,
+    summaryA,
+    summaryB,
+    defaultTargets,
+    ssrRows,
+    interactivityRange,
+    oldest,
+    newest,
+    initialPairBenchmarkRows,
+  } = await getComparePageDerivedData(
+    parsed.model.dbKeys,
     parsed.a,
     parsed.b,
-    effectiveSequence,
-    effectivePrecision,
+    requestedSequence,
+    requestedPrecision,
+    fallbackSequence,
   );
 
-  const url = `${SITE_URL}/compare-per-dollar/${canonical}`;
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const url = `${SITE_URL}${routePath}`;
   const imageUrl = `${url}/performance-per-dollar.png`;
-  const { oldest, newest } = dateRangeForPair(rows, parsed.a, parsed.b);
   const jsonLd = buildJsonLd(
     'per-dollar',
     parsed.model,
@@ -182,7 +231,13 @@ export default async function ComparePerDollarPage({ params, searchParams }: Pro
         defaultSequence={effectiveSequence}
         defaultPrecision={effectivePrecision}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
+        initialBenchmarkRows={initialCompareBenchmarkRows(
+          parsed.model.displayName,
+          effectiveModel,
+          initialPairBenchmarkRows,
+        )}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO : null}
         aLabel={aLabel}
         bLabel={bLabel}
         aVendor={aMeta?.vendor ?? ''}
diff --git a/packages/app/src/app/compare-per-dollar/[slug]/performance-per-dollar.png/route.tsx b/packages/app/src/app/compare-per-dollar/[slug]/performance-per-dollar.png/route.tsx
index 9d3fad942..bc00c051b 100644
--- a/packages/app/src/app/compare-per-dollar/[slug]/performance-per-dollar.png/route.tsx
+++ b/packages/app/src/app/compare-per-dollar/[slug]/performance-per-dollar.png/route.tsx
@@ -1,8 +1,3 @@
-import { readFile } from 'node:fs/promises';
-import { join } from 'node:path';
-
-import { ImageResponse } from 'next/og';
-
 import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
 
 import { trackServer } from '@/lib/analytics-server';
@@ -12,99 +7,13 @@ import {
   computeCompareImageRows,
   computeCompareTableData,
   getCachedBenchmarks,
-  type SsrInterpolatedRow,
 } from '@/lib/compare-ssr';
+import { getLogoSrc } from '@/lib/og-assets';
+import { renderComparePngChart } from '@/lib/png-chart';
 
 export const dynamic = 'force-dynamic';
 export const runtime = 'nodejs';
 
-// Render natively at high-DPI. CSS `transform: scale()` causes Satori to rasterize
-// SVG/text at the source size and bitmap-upsample, which produces a blurry chart.
-// Multiplying every pixel value by R keeps glyphs and strokes as vectors at full res.
-const R = 2;
-const SIZE = { width: 1200 * R, height: 675 * R };
-const CHART_FRAME = { left: 0, top: 18 * R, width: 746 * R, height: 382 * R };
-const CHART = { left: 96 * R, top: 42 * R, width: 630 * R, height: 260 * R };
-const COLORS = {
-  background: '#0d1117',
-  panel: '#121a23',
-  border: '#23303d',
-  muted: '#9aa7b5',
-  faint: '#5f6e7d',
-  text: '#f3f7fb',
-  a: '#38d9a9',
-  b: '#f7b041',
-  grid: '#263544',
-  blue: '#0b86d1',
-};
-
-interface Point {
-  x: number;
-  y: number;
-}
-
-interface TargetedPoint extends Point {
-  target: number;
-}
-
-function money(value: number): string {
-  if (value >= 10) return `$${value.toFixed(1)}`;
-  if (value >= 1) return `$${value.toFixed(2)}`;
-  return `$${value.toFixed(3)}`;
-}
-
-/** Decimals chosen from the tick step so every label in the axis prints with
- * the same precision (no $0.000/$9.01/$18.0 mix). */
-function decimalsForStep(step: number): number {
-  if (step >= 1) return 0;
-  return Math.max(0, Math.ceil(-Math.log10(step)));
-}
-
-function moneyForStep(value: number, step: number): string {
-  return `$${value.toFixed(decimalsForStep(step))}`;
-}
-
-/** "Nice" step in the 1/2/5 × 10ⁿ family, the same convention d3 uses. */
-function niceStep(span: number, targetCount: number): number {
-  const rawStep = span / Math.max(1, targetCount - 1);
-  const mag = 10 ** Math.floor(Math.log10(rawStep));
-  const normalized = rawStep / mag;
-  if (normalized < 1.5) return mag;
-  if (normalized < 3) return 2 * mag;
-  if (normalized < 7) return 5 * mag;
-  return 10 * mag;
-}
-
-function niceAxis(
-  min: number,
-  max: number,
-  targetCount = 5,
-): { min: number; max: number; step: number; ticks: number[] } {
-  if (max <= min) return { min, max: min + 1, step: 1, ticks: [min] };
-  const step = niceStep(max - min, targetCount);
-  const niceMin = Math.floor(min / step) * step;
-  const niceMax = Math.ceil(max / step) * step;
-  const ticks: number[] = [];
-  for (let t = niceMin; t <= niceMax + step * 1e-6; t += step) {
-    ticks.push(Number(t.toFixed(10)));
-  }
-  return { min: niceMin, max: niceMax, step, ticks };
-}
-
-function pointsPath(points: Point[]): string {
-  return points.map((point, index) => `${index === 0 ? 'M' : 'L'} ${point.x} ${point.y}`).join(' ');
-}
-
-let logoSrcPromise: Promise<string | null> | undefined;
-function getLogoSrc(): Promise<string | null> {
-  if (!logoSrcPromise) {
-    logoSrcPromise = readFile(join(process.cwd(), 'public/brand/logo-color.png'))
-      .then((buf) => `data:image/png;base64,${buf.toString('base64')}`)
-      .catch(() => null);
-  }
-  return logoSrcPromise;
-}
-
 export async function GET(
   _request: Request,
   { params }: { params: Promise<{ slug: string }> },
@@ -138,431 +47,30 @@ export async function GET(
     sequence,
     precision,
     interactivityRange,
-    plottedRows.map((r) => r.target),
+    plottedRows.map((row) => row.target),
   ).filter((row) => row.a || row.b);
   const curveRows = imageRows.length > 0 ? imageRows : plottedRows;
 
   const aLabel = HW_REGISTRY[parsed.a]?.label ?? parsed.a.toUpperCase();
   const bLabel = HW_REGISTRY[parsed.b]?.label ?? parsed.b.toUpperCase();
-  const costs = curveRows
-    .flatMap((row) => [row.a?.cost, row.b?.cost])
-    .filter((cost): cost is number => typeof cost === 'number' && Number.isFinite(cost));
-  const costMin = costs.length > 0 ? Math.min(...costs) : 0;
-  const costMax = costs.length > 0 ? Math.max(...costs) : 1;
-  const yAxis = niceAxis(Math.min(0, costMin), costMax);
-  const yMin = yAxis.min;
-  const yMax = yAxis.max;
-  const yStep = yAxis.step;
-  const xMin = curveRows.at(0)?.target ?? 0;
-  const xMax = curveRows.at(-1)?.target ?? 100;
-  const matchedMin = plottedRows.at(0)?.target ?? xMin;
-  const matchedMax = plottedRows.at(-1)?.target ?? xMax;
-  const hasLeftExtension = matchedMin - xMin >= 0.5;
-  const hasRightExtension = xMax - matchedMax >= 0.5;
-  const scaleX = (value: number) =>
-    CHART.left + (xMax === xMin ? CHART.width / 2 : ((value - xMin) / (xMax - xMin)) * CHART.width);
-  const scaleY = (value: number) =>
-    CHART.top +
-    CHART.height -
-    (yMax === yMin ? CHART.height / 2 : ((value - yMin) / (yMax - yMin)) * CHART.height);
-
-  function buildSeriesPoints(getCost: (row: SsrInterpolatedRow) => number | null): TargetedPoint[] {
-    return curveRows
-      .map((row) => ({ target: row.target, cost: getCost(row) }))
-      .filter((p): p is { target: number; cost: number } => p.cost !== null)
-      .map((p) => ({ x: scaleX(p.target), y: scaleY(p.cost), target: p.target }));
-  }
-
-  function splitByMatchRange(points: TargetedPoint[]) {
-    return {
-      matched: points.filter((p) => p.target >= matchedMin && p.target <= matchedMax),
-      leftExt: points.filter((p) => p.target <= matchedMin),
-      rightExt: points.filter((p) => p.target >= matchedMax),
-    };
-  }
-
-  const aSeries = splitByMatchRange(buildSeriesPoints((r) => r.a?.cost ?? null));
-  const bSeries = splitByMatchRange(buildSeriesPoints((r) => r.b?.cost ?? null));
-  const aHighlightPoints = plottedRows
-    .filter((row) => row.a)
-    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.a!.cost) }));
-  const bHighlightPoints = plottedRows
-    .filter((row) => row.b)
-    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.b!.cost) }));
   const workload = [sequence, precision?.toUpperCase()].filter(Boolean).join(' / ');
-  const showRangeEndpoints = hasLeftExtension || hasRightExtension;
-  const svgWidth = 760 * R;
-  const svgHeight = 406 * R;
-
-  function renderSeriesPath(points: Point[], stroke: string, dashed: boolean) {
-    if (points.length < 2) return null;
-    return (
-      <path
-        d={pointsPath(points)}
-        fill="none"
-        stroke={stroke}
-        strokeWidth={9 * R}
-        strokeOpacity={dashed ? 0.55 : 1}
-        strokeDasharray={dashed ? `${14 * R} ${10 * R}` : undefined}
-        strokeLinejoin="round"
-        strokeLinecap="round"
-      />
-    );
-  }
 
   try {
-    return new ImageResponse(
-      <div
-        style={{
-          display: 'flex',
-          flexDirection: 'column',
-          width: SIZE.width,
-          height: SIZE.height,
-          padding: `${38 * R}px ${46 * R}px ${26 * R}px`,
-          background: COLORS.background,
-          color: COLORS.text,
-          fontFamily: 'Arial, sans-serif',
-        }}
-      >
-        <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'flex-start' }}>
-          <div style={{ display: 'flex', flexDirection: 'column', gap: 7 * R }}>
-            <div
-              style={{
-                display: 'flex',
-                fontSize: 19 * R,
-                fontWeight: 700,
-                letterSpacing: '0.13em',
-                textTransform: 'uppercase',
-                color: COLORS.blue,
-              }}
-            >
-              InferenceX Performance per Dollar
-            </div>
-            <div style={{ display: 'flex', fontSize: 41 * R, fontWeight: 800 }}>
-              {parsed.model.label}
-            </div>
-            <div style={{ display: 'flex', fontSize: 25 * R, color: COLORS.muted }}>
-              {aLabel} vs {bLabel} | Cost per Million Tokens
-            </div>
-          </div>
-          <div
-            style={{
-              display: 'flex',
-              flexDirection: 'column',
-              alignItems: 'flex-end',
-              border: `${R}px solid ${COLORS.border}`,
-              borderRadius: 12 * R,
-              padding: `${13 * R}px ${17 * R}px`,
-              background: COLORS.panel,
-              gap: 5 * R,
-            }}
-          >
-            <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
-              DEFAULT WORKLOAD
-            </div>
-            <div style={{ display: 'flex', fontSize: 21 * R, fontWeight: 700 }}>
-              {workload || 'Default comparison'}
-            </div>
-            <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
-              Lower cost is better
-            </div>
-          </div>
-        </div>
-
-        <div style={{ display: 'flex', flex: 1, gap: 34 * R, marginTop: 22 * R }}>
-          <div
-            style={{ display: 'flex', position: 'relative', width: svgWidth, height: svgHeight }}
-          >
-            <svg
-              width={svgWidth}
-              height={svgHeight}
-              viewBox={`0 0 ${svgWidth} ${svgHeight}`}
-              style={{ position: 'absolute', left: 0, top: 0 }}
-            >
-              <rect
-                x={CHART_FRAME.left}
-                y={CHART_FRAME.top}
-                width={CHART_FRAME.width}
-                height={CHART_FRAME.height}
-                rx={13 * R}
-                fill={COLORS.panel}
-                stroke={COLORS.border}
-              />
-              {yAxis.ticks.map((tick) => {
-                const y = scaleY(tick);
-                return (
-                  <line
-                    key={tick}
-                    x1={CHART.left}
-                    x2={CHART.left + CHART.width}
-                    y1={y}
-                    y2={y}
-                    stroke={COLORS.grid}
-                    strokeWidth={2 * R}
-                  />
-                );
-              })}
-              {plottedRows.map((row) => {
-                const x = scaleX(row.target);
-                return (
-                  <line
-                    key={`mark-${row.target}`}
-                    x1={x}
-                    x2={x}
-                    y1={CHART.top + CHART.height}
-                    y2={CHART.top + CHART.height + 6 * R}
-                    stroke={COLORS.muted}
-                    strokeWidth={2 * R}
-                  />
-                );
-              })}
-              {renderSeriesPath(aSeries.leftExt, COLORS.a, true)}
-              {renderSeriesPath(aSeries.rightExt, COLORS.a, true)}
-              {renderSeriesPath(aSeries.matched, COLORS.a, false)}
-              {renderSeriesPath(bSeries.leftExt, COLORS.b, true)}
-              {renderSeriesPath(bSeries.rightExt, COLORS.b, true)}
-              {renderSeriesPath(bSeries.matched, COLORS.b, false)}
-              {aHighlightPoints.map((point, index) => (
-                <circle
-                  key={`a-${index}`}
-                  cx={point.x}
-                  cy={point.y}
-                  r={10 * R}
-                  fill={COLORS.a}
-                  stroke={COLORS.background}
-                  strokeWidth={4 * R}
-                />
-              ))}
-              {bHighlightPoints.map((point, index) => (
-                <circle
-                  key={`b-${index}`}
-                  cx={point.x}
-                  cy={point.y}
-                  r={10 * R}
-                  fill={COLORS.b}
-                  stroke={COLORS.background}
-                  strokeWidth={4 * R}
-                />
-              ))}
-            </svg>
-            {logoSrc && (
-              <img
-                src={logoSrc}
-                alt=""
-                height={144 * R}
-                style={{
-                  position: 'absolute',
-                  left: CHART.left + CHART.width / 2 - 168 * R,
-                  top: CHART.top + CHART.height / 2 - 72 * R,
-                  opacity: 0.12,
-                }}
-              />
-            )}
-            {yAxis.ticks.map((tick) => (
-              <div
-                key={`y-label-${tick}`}
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: CHART_FRAME.left + 14 * R,
-                  top: scaleY(tick) - 9 * R,
-                  width: CHART.left - CHART_FRAME.left - 28 * R,
-                  justifyContent: 'flex-end',
-                  color: COLORS.muted,
-                  fontSize: 15 * R,
-                }}
-              >
-                {moneyForStep(tick, yStep)}
-              </div>
-            ))}
-            {plottedRows.map((row) => (
-              <div
-                key={`x-label-${row.target}`}
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(row.target) - 32 * R,
-                  top: CHART.top + CHART.height + 15 * R,
-                  width: 64 * R,
-                  justifyContent: 'center',
-                  color: COLORS.muted,
-                  fontSize: 16 * R,
-                  fontWeight: 600,
-                }}
-              >
-                {row.target}
-              </div>
-            ))}
-            {showRangeEndpoints && hasLeftExtension && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(xMin) - 4 * R,
-                  top: CHART.top + CHART.height + 16 * R,
-                  width: 56 * R,
-                  justifyContent: 'flex-start',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                {Math.round(xMin)}
-              </div>
-            )}
-            {showRangeEndpoints && hasRightExtension && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(xMax) - 52 * R,
-                  top: CHART.top + CHART.height + 16 * R,
-                  width: 56 * R,
-                  justifyContent: 'flex-end',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                {Math.round(xMax)}
-              </div>
-            )}
-            <div
-              style={{
-                display: 'flex',
-                position: 'absolute',
-                left: CHART.left,
-                top: CHART.top + CHART.height + 38 * R,
-                width: CHART.width,
-                justifyContent: 'center',
-                color: COLORS.muted,
-                fontSize: 15 * R,
-                fontWeight: 600,
-              }}
-            >
-              Interactivity (tok/s/user)
-            </div>
-            {showRangeEndpoints && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: CHART.left,
-                  top: CHART.top + CHART.height + 62 * R,
-                  width: CHART.width,
-                  justifyContent: 'center',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                Dashed segments extend to each SKU's operating envelope, where cost rises steeply
-              </div>
-            )}
-          </div>
-
-          <div
-            style={{
-              display: 'flex',
-              flex: 1,
-              flexDirection: 'column',
-              gap: 17 * R,
-              paddingTop: 18 * R,
-            }}
-          >
-            <div style={{ display: 'flex', fontSize: 18 * R, fontWeight: 700 }}>
-              Matched Interactivity
-            </div>
-            <div style={{ display: 'flex', gap: 20 * R, fontSize: 15 * R, color: COLORS.muted }}>
-              <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
-                <span
-                  style={{
-                    display: 'flex',
-                    width: 19 * R,
-                    height: 6 * R,
-                    borderRadius: 3 * R,
-                    background: COLORS.a,
-                  }}
-                />
-                {aLabel}
-              </span>
-              <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
-                <span
-                  style={{
-                    display: 'flex',
-                    width: 19 * R,
-                    height: 6 * R,
-                    borderRadius: 3 * R,
-                    background: COLORS.b,
-                  }}
-                />
-                {bLabel}
-              </span>
-            </div>
-            {plottedRows.length > 0 ? (
-              plottedRows.map((row) => (
-                <div
-                  key={`row-${row.target}`}
-                  style={{
-                    display: 'flex',
-                    flexDirection: 'column',
-                    gap: 6 * R,
-                    border: `${R}px solid ${COLORS.border}`,
-                    borderRadius: 10 * R,
-                    padding: `${11 * R}px ${13 * R}px`,
-                    background: COLORS.panel,
-                  }}
-                >
-                  <div style={{ display: 'flex', color: COLORS.muted, fontSize: 13 * R }}>
-                    {row.target} tok/s/user
-                  </div>
-                  <div style={{ display: 'flex', gap: 15 * R, fontSize: 19 * R, fontWeight: 700 }}>
-                    <span style={{ display: 'flex', color: COLORS.a }}>
-                      {row.a ? money(row.a.cost) : 'N/A'}
-                    </span>
-                    <span style={{ display: 'flex', color: COLORS.b }}>
-                      {row.b ? money(row.b.cost) : 'N/A'}
-                    </span>
-                  </div>
-                </div>
-              ))
-            ) : (
-              <div style={{ display: 'flex', fontSize: 18 * R, color: COLORS.muted }}>
-                No matched cost data available.
-              </div>
-            )}
-          </div>
-        </div>
-
-        <div
-          style={{
-            display: 'flex',
-            justifyContent: 'space-between',
-            paddingTop: 9 * R,
-            fontSize: 15 * R,
-            color: COLORS.muted,
-          }}
-        >
-          <span style={{ display: 'flex' }}>
-            Owning-hyperscaler TCO | interpolated from benchmark results
-          </span>
-          <span style={{ display: 'flex', color: COLORS.text, fontWeight: 700 }}>
-            inferencex.semianalysis.com
-          </span>
-        </div>
-      </div>,
-      {
-        ...SIZE,
-        headers: {
-          'Cache-Control': 'public, s-maxage=3600, stale-while-revalidate=86400',
-        },
-      },
-    );
+    return renderComparePngChart({
+      curveRows,
+      plottedRows,
+      logoSrc,
+      aLabel,
+      bLabel,
+      eyebrow: 'InferenceX Performance per Dollar',
+      title: parsed.model.label,
+      subtitle: `${aLabel} vs ${bLabel} | Cost per Million Tokens`,
+      workload,
+      rangeNote:
+        "Dashed segments extend to each SKU's operating envelope, where cost rises steeply",
+      footer: 'Owning-hyperscaler TCO | interpolated from benchmark results',
+    });
   } catch (error) {
-    // Satori can throw on a font fetch failure, malformed JSX layout, or a
-    // dataset that produces NaN/Infinity geometry. Capture which slug broke so
-    // we can find broken categories before a crawler hits them — without this,
-    // failures only surface as opaque Vercel 500s.
     const message = error instanceof Error ? error.message : String(error);
     trackServer('compare_per_dollar_png_render_failed', {
       slug,
@@ -574,9 +82,6 @@ export async function GET(
       error_name: error instanceof Error ? error.name : 'Unknown',
       error_message: message.slice(0, 500),
     });
-    // 502 (not 500): the route itself is reachable, the downstream renderer
-    // failed. Short cache so a retry within the hour pulls a fixed render
-    // instead of pinning the failure.
     return new Response('PNG render failed', {
       status: 502,
       headers: { 'Cache-Control': 'public, s-maxage=60' },
diff --git a/packages/app/src/app/compare-per-dollar/layout.tsx b/packages/app/src/app/compare-per-dollar/layout.tsx
index c4e14ab1a..4ff2b2492 100644
--- a/packages/app/src/app/compare-per-dollar/layout.tsx
+++ b/packages/app/src/app/compare-per-dollar/layout.tsx
@@ -1,20 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
-/**
- * Wraps `/compare-per-dollar/*` pages in UnofficialRunProvider but skips
- * DashboardShell so we get a focused single-purpose page (no TabNav). Mirrors
- * the structure of `/compare/layout.tsx` — the GlobalFilterProvider is mounted
- * inside the page's client component so it can seed initial precision/sequence
- * based on which combos actually have data for the GPU pair from the slug.
- */
 export default function ComparePerDollarLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/compare-per-dollar/page.tsx b/packages/app/src/app/compare-per-dollar/page.tsx
index 1a2e24ed2..de359e045 100644
--- a/packages/app/src/app/compare-per-dollar/page.tsx
+++ b/packages/app/src/app/compare-per-dollar/page.tsx
@@ -13,6 +13,7 @@ import { enAlternates } from '@/lib/i18n';
 import { ComparePairCardLink } from '@/components/compare/compare-pair-card-link';
 import { JsonLd } from '@/components/json-ld';
 import { Card } from '@/components/ui/card';
+import { comparisonPairHref, comparisonScenarioForModel } from '@/lib/compare-agentx';
 import { getComparablePairsByModelSlug } from '@/lib/compare-availability';
 import { type ComparePair, COMPARE_MODEL_SLUGS, type CompareModelSlug } from '@/lib/compare-slug';
 import { bucketComparePairsByVendor, formatModelList } from '@/lib/compare-ssr';
@@ -108,6 +109,11 @@ export default async function ComparePerDollarIndexPage() {
             each page renders the cost-per-token chart and an interpolated dollars-per-million
             comparison table so you can pick the cheaper SKU at any target interactivity level.
           </p>
+          <p className="mt-3 text-base lg:text-lg text-muted-foreground max-w-3xl">
+            Models with AgentX data open long-context, multi-turn trace replay results. Models not
+            yet covered by AgentX open the controlled 8K→1K workload. Each card identifies its
+            scenario.
+          </p>
           <div className="mt-6 flex flex-wrap gap-3">
             <Link
               data-testid="compare-per-dollar-index-compare-link"
@@ -146,6 +152,9 @@ export default async function ComparePerDollarIndexPage() {
       {modelsWithPairs.map((model) => {
         const pairs = comparablePairsByModel.get(model.slug) ?? [];
         const groups = groupPairsByVendorForModel(model, pairs);
+        // Same scenario split /compare uses: models with AgentX data open the
+        // agentic trace replay, the rest open the fixed 8K→1K workload.
+        const scenario = comparisonScenarioForModel(model);
         return (
           <section key={model.slug} id={model.slug}>
             <Card className="flex flex-col gap-4">
@@ -170,10 +179,11 @@ export default async function ComparePerDollarIndexPage() {
                       return (
                         <ComparePairCardLink
                           key={slug}
-                          href={`/compare-per-dollar/${slug}`}
+                          href={comparisonPairHref('en', slug, model, 'compare-per-dollar')}
                           slug={slug}
                           label={label}
                           archLine={archLine}
+                          scenarioLabel={scenario.label}
                         />
                       );
                     })}
diff --git a/packages/app/src/app/compare-precision/[slug]/[scenario]/page.tsx b/packages/app/src/app/compare-precision/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..f709b74a6
--- /dev/null
+++ b/packages/app/src/app/compare-precision/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildPrecisionMetadata, renderPrecisionPage } from '../page';
+
+/**
+ * `/compare-precision/<slug>/<scenario>` — the same comparison with its workload pinned by
+ * the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments rather than silently falling back to the pair's
+ *  default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildPrecisionMetadata(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function PrecisionScenarioPage({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderPrecisionPage(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/compare-precision/[slug]/opengraph-image.tsx b/packages/app/src/app/compare-precision/[slug]/opengraph-image.tsx
index b6a65d259..3b54414ce 100644
--- a/packages/app/src/app/compare-precision/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/compare-precision/[slug]/opengraph-image.tsx
@@ -1,20 +1,14 @@
-/**
- * Precision compare OG image — identical circuit tile sidebar layout to the
- * /compare OG and the blog OG, but the eyebrow and bottom banner surface
- * the "Precision Comparison" framing.
- */
 import { notFound } from 'next/navigation';
-import { ImageResponse } from 'next/og';
 
 import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
 
+import { renderCompareOg } from '@/lib/compare-og';
 import { getAllComparablePrecisionSlugs } from '@/lib/compare-variant-availability';
 import {
   canonicalPrecisionCompareSlug,
   parsePrecisionCompareSlug,
   precisionDisplayLabel,
 } from '@/lib/compare-variant-slug';
-import { BG, BLUE, getLogoSrc, getTiles, PANEL_BG } from '@/lib/og-assets';
 
 export const alt = 'Chip precision inference benchmark comparison';
 export const size = { width: 1200, height: 630 };
@@ -30,9 +24,6 @@ export async function generateStaticParams() {
 export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
   const { slug } = await params;
   const parsed = parsePrecisionCompareSlug(slug);
-  // Reject non-canonical slugs (reversed order, aliases, mixed case) so OG
-  // URLs can't render labels that disagree with the canonical page — same
-  // strictness as the sibling PNG route.
   if (
     !parsed ||
     canonicalPrecisionCompareSlug(parsed.model.slug, parsed.gpu, parsed.precA, parsed.precB) !==
@@ -40,127 +31,19 @@ export default async function OgImage({ params }: { params: Promise<{ slug: stri
   ) {
     notFound();
   }
-  const [logoSrc, tiles] = await Promise.all([getLogoSrc(), getTiles()]);
 
-  const gpuMeta = HW_REGISTRY[parsed.gpu];
-  const gpuLabel = gpuMeta?.label ?? parsed.gpu.toUpperCase();
+  const gpuLabel = HW_REGISTRY[parsed.gpu]?.label ?? parsed.gpu.toUpperCase();
   const aLabel = precisionDisplayLabel(parsed.precA);
   const bLabel = precisionDisplayLabel(parsed.precB);
   const title = `${gpuLabel}: ${aLabel} vs ${bLabel}`;
-  const eyebrow = `${parsed.model.label} · Precision Comparison`;
   const titleSize = title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
 
-  return new ImageResponse(
-    <div
-      style={{
-        display: 'flex',
-        width: '100%',
-        height: '100%',
-        backgroundColor: BG,
-        color: '#EAEBEC',
-        overflow: 'hidden',
-      }}
-    >
-      <div
-        style={{
-          display: 'flex',
-          width: 195,
-          height: '100%',
-          backgroundColor: PANEL_BG,
-          position: 'relative',
-        }}
-      >
-        {tiles.map((tile, i) => {
-          if (!tile) return null;
-          const row = Math.floor(i / 2);
-          const col = i % 2;
-          return (
-            <img
-              key={i}
-              src={tile.src}
-              style={{
-                position: 'absolute',
-                left: 12 + col * 90,
-                top: 12 + row * 104,
-                width: 78,
-                height: 86,
-                borderRadius: 4,
-                objectFit: 'cover',
-                ...(tile.rotate ? { transform: `rotate(${tile.rotate}deg)` } : {}),
-              }}
-            />
-          );
-        })}
-        <div
-          style={{
-            position: 'absolute',
-            right: 0,
-            top: 0,
-            width: 3,
-            height: '100%',
-            backgroundColor: BLUE,
-            display: 'flex',
-          }}
-        />
-      </div>
-
-      <div
-        style={{
-          display: 'flex',
-          flexDirection: 'column',
-          justifyContent: 'space-between',
-          flex: 1,
-          padding: '48px 55px 20px 55px',
-        }}
-      >
-        <div
-          style={{
-            display: 'flex',
-            flexDirection: 'column',
-            gap: 24,
-            flex: 1,
-            justifyContent: 'center',
-          }}
-        >
-          <div
-            style={{
-              fontSize: 24,
-              color: '#9BA0A6',
-              letterSpacing: '0.18em',
-              textTransform: 'uppercase',
-              display: 'flex',
-            }}
-          >
-            {eyebrow}
-          </div>
-
-          <div
-            style={{
-              fontSize: titleSize,
-              fontWeight: 800,
-              color: '#FFFFFF',
-              lineHeight: 1.1,
-              display: 'flex',
-            }}
-          >
-            {title}
-          </div>
-        </div>
-
-        <div
-          style={{
-            display: 'flex',
-            justifyContent: 'space-between',
-            alignItems: 'center',
-          }}
-        >
-          <span style={{ fontSize: 26, color: '#9BA0A6', display: 'flex' }}>
-            AI inference benchmark · precision comparison
-          </span>
-          {logoSrc && <img src={logoSrc} height={72} />}
-        </div>
-      </div>
-    </div>,
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · Precision Comparison`,
+    title,
+    titleSize,
+    footer: 'AI inference benchmark · precision comparison',
+    language: 'en',
     size,
-  );
+  });
 }
diff --git a/packages/app/src/app/compare-precision/[slug]/page-client.tsx b/packages/app/src/app/compare-precision/[slug]/page-client.tsx
index f7cd67409..11e2704f4 100644
--- a/packages/app/src/app/compare-precision/[slug]/page-client.tsx
+++ b/packages/app/src/app/compare-precision/[slug]/page-client.tsx
@@ -3,21 +3,23 @@
 import Link from 'next/link';
 import { useEffect, useMemo } from 'react';
 
-import type { GPUDataPoint, InterpolatedResult } from '@/components/calculator/types';
+import type { GPUDataPoint } from '@/components/calculator/types';
 import { useThroughputData } from '@/components/calculator/useThroughputData';
-import { CompareInterpolatedTable } from '@/components/compare/compare-interpolated-table';
-import { useGlobalFilters, GlobalFilterProvider } from '@/components/GlobalFilterContext';
+import {
+  CompareTableRenderer,
+  type CompareTableData,
+} from '@/components/compare/compare-table-renderer';
+import {
+  GlobalFilterProvider,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+} from '@/components/GlobalFilterContext';
 import { InferenceProvider } from '@/components/inference/InferenceContext';
 import InferenceChartDisplay from '@/components/inference/ui/ChartDisplay';
 import { Card } from '@/components/ui/card';
 import { track } from '@/lib/analytics';
 import { toModel, toSequence } from '@/lib/compare-enum-coerce';
-
-interface SsrTableData {
-  defaultTargets: number[];
-  ssrRows: { target: number; a: InterpolatedResult | null; b: InterpolatedResult | null }[];
-  interactivityRange: { min: number; max: number };
-}
+import type { AgenticScenarioIntro } from '@/lib/compare-ssr';
 
 const STRINGS = {
   en: {
@@ -50,8 +52,11 @@ interface ComparePrecisionPageClientProps {
   defaultSequence: string | null;
   precA: string;
   precB: string;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
   narrative: string[];
+  /** Set only when the page is rendering the agentic workload. Explains
+   *  what AgentX measures before the head-to-head numbers. */
+  agenticIntro?: AgenticScenarioIntro | null;
   gpuLabel: string;
   gpuVendor: string;
   gpuArch: string;
@@ -71,6 +76,7 @@ export default function ComparePrecisionPageClient({
   precB,
   ssrTableData,
   narrative,
+  agenticIntro = null,
   gpuLabel,
   gpuVendor,
   gpuArch,
@@ -135,6 +141,21 @@ export default function ComparePrecisionPageClient({
                   .
                 </p>
               )}
+              {agenticIntro && (
+                <p
+                  className="mt-3 max-w-3xl text-sm text-foreground/80"
+                  data-testid="compare-agentic-intro"
+                >
+                  {agenticIntro.paragraph}{' '}
+                  <Link
+                    href={agenticIntro.href}
+                    data-testid="compare-agentic-intro-link"
+                    className="font-medium text-brand underline underline-offset-4 hover:no-underline"
+                  >
+                    {agenticIntro.linkLabel} →
+                  </Link>
+                </p>
+              )}
               {narrative.length > 0 && (
                 <div
                   className="mt-3 flex flex-col gap-2 max-w-3xl"
@@ -228,11 +249,11 @@ function CompareTableSection({
   precB: string;
   aLabel: string;
   bLabel: string;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
   emptyStateText: string;
 }) {
-  const { effectiveSequence, effectivePrecisions, selectedRunDate, selectedModel } =
-    useGlobalFilters();
+  const { effectiveSequence, effectivePrecisions, selectedModel } = useGlobalFilterSelection();
+  const { selectedRunDate } = useGlobalFilterRun();
 
   const { gpuDataByGroupKey, ranges, hasData } = useThroughputData(
     selectedModel,
@@ -260,23 +281,15 @@ function CompareTableSection({
 
   const clientRange = hasData ? ranges.interactivity : ssrTableData.interactivityRange;
 
-  if (ssrTableData.defaultTargets.length === 0) {
-    return (
-      <div className="border border-border/50 rounded-md px-4 py-3 text-sm text-muted-foreground bg-muted/30">
-        {emptyStateText}
-      </div>
-    );
-  }
-
   return (
-    <CompareInterpolatedTable
+    <CompareTableRenderer
       aLabel={aLabel}
       bLabel={bLabel}
-      ssrRows={ssrTableData.ssrRows}
-      defaultTargets={ssrTableData.defaultTargets}
+      ssrTableData={ssrTableData}
       interactivityRange={clientRange}
       gpuDataPointsA={pointsA}
       gpuDataPointsB={pointsB}
+      emptyStateText={emptyStateText}
     />
   );
 }
diff --git a/packages/app/src/app/compare-precision/[slug]/page.tsx b/packages/app/src/app/compare-precision/[slug]/page.tsx
index 23f4922b9..b0f8b0e78 100644
--- a/packages/app/src/app/compare-precision/[slug]/page.tsx
+++ b/packages/app/src/app/compare-precision/[slug]/page.tsx
@@ -9,8 +9,18 @@ import {
 } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
 import { languageAlternates } from '@/lib/i18n';
-import { getCachedBenchmarks, KNOWN_SEQUENCES, pickString } from '@/lib/compare-ssr';
+import {
+  AGENTIC_SCENARIO_INTRO,
+  getCachedBenchmarks,
+  KNOWN_SEQUENCES,
+  pickString,
+} from '@/lib/compare-ssr';
 import {
   canonicalPrecisionCompareSlug,
   parsePrecisionCompareSlug,
@@ -30,6 +40,22 @@ import ComparePrecisionPageClient from './page-client';
 
 export const dynamic = 'force-dynamic';
 
+/**
+ * `/compare-precision/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than duplicated
+ * in a parallel route file so the body — redirects, JSON-LD, metadata, client
+ * props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
+function scenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/compare-precision/${canonical}/${scenarioSegment}`
+    : `/compare-precision/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
@@ -37,6 +63,13 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildPrecisionMetadata(slug, {});
+}
+
+export function buildPrecisionMetadata(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Metadata {
   const parsed = parsePrecisionCompareSlug(slug);
   if (!parsed) return {};
   const gpuMeta = HW_REGISTRY[parsed.gpu];
@@ -49,14 +82,21 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     parsed.precA,
     parsed.precB,
   );
-  const url = `${SITE_URL}/compare-precision/${canonical}`;
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const canonicalPath = scenarioPath(canonical);
+  const url = `${SITE_URL}${routePath}`;
+  const canonicalUrl = `${SITE_URL}${canonicalPath}`;
   const description = `${gpuLabel} precision comparison of ${aLabel} versus ${bLabel} on ${parsed.model.label}: verified, reproducible results from InferenceX, the independent open-source benchmark by SemiAnalysis. ${SUPPORTERS_LINE} See which quantization level delivers better throughput and cost at every interactivity level.`;
   return {
     title: `${parsed.model.label} — ${gpuLabel} ${aLabel} vs ${bLabel} — Precision Comparison`,
     description,
     alternates: {
-      canonical: url,
-      languages: languageAlternates(`/compare-precision/${canonical}`),
+      canonical: canonicalUrl,
+      languages: languageAlternates(canonicalPath),
     },
     openGraph: {
       title: `${parsed.model.label} — ${gpuLabel} ${aLabel} vs ${bLabel} — Precision Comparison | ${SITE_NAME}`,
@@ -74,11 +114,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function ComparePrecisionPage({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderPrecisionPage(slug, await searchParams, {});
+}
+
+export async function renderPrecisionPage(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parsePrecisionCompareSlug(slug);
   if (!parsed) notFound();
 
-  const sp = await searchParams;
-
   // 308 redirect to canonical — normalizes alias models and precision order.
   const canonical = canonicalPrecisionCompareSlug(
     parsed.model.slug,
@@ -95,7 +141,9 @@ export default async function ComparePrecisionPage({ params, searchParams }: Pro
       })
       .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
       .join('&');
-    permanentRedirect(`/compare-precision/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // the reader to the pair's default workload instead.
+    permanentRedirect(`${scenarioPath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
   const rows = await getCachedBenchmarks(parsed.model.dbKeys);
@@ -111,7 +159,11 @@ export default async function ComparePrecisionPage({ params, searchParams }: Pro
   );
 
   const urlSeq = pickString(sp.i_seq);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence;
+  // Path beats query beats the pair's default: a scenario segment is an
+  // explicit address for one workload, so it outranks a stale `?i_seq=`.
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const effectiveSequence =
+    pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence);
 
   const { defaultTargets, ssrRows, interactivityRange } = computeVariantCompareTableData(
     rows,
@@ -129,7 +181,7 @@ export default async function ComparePrecisionPage({ params, searchParams }: Pro
   const aLabel = precisionDisplayLabel(parsed.precA);
   const bLabel = precisionDisplayLabel(parsed.precB);
 
-  const url = `${SITE_URL}/compare-precision/${canonical}`;
+  const url = `${SITE_URL}${scenarioPath(canonical, scenarioSegment)}`;
   const imageUrl = `${url}/precision-comparison.png`;
   const { oldest, newest } = dateRangeForVariantPair(rows, parsed.gpu, sideA, sideB);
   const jsonLd = buildVariantJsonLd(
@@ -172,6 +224,7 @@ export default async function ComparePrecisionPage({ params, searchParams }: Pro
         precB={parsed.precB}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO : null}
         gpuLabel={gpuLabel}
         gpuVendor={gpuMeta?.vendor ?? ''}
         gpuArch={gpuMeta?.arch ?? ''}
diff --git a/packages/app/src/app/compare-precision/[slug]/precision-comparison.png/route.tsx b/packages/app/src/app/compare-precision/[slug]/precision-comparison.png/route.tsx
index 95552ac1c..f081f33d4 100644
--- a/packages/app/src/app/compare-precision/[slug]/precision-comparison.png/route.tsx
+++ b/packages/app/src/app/compare-precision/[slug]/precision-comparison.png/route.tsx
@@ -1,5 +1,3 @@
-import { ImageResponse } from 'next/og';
-
 import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
 
 import { trackServer } from '@/lib/analytics-server';
@@ -15,19 +13,7 @@ import {
   precisionDisplayLabel,
 } from '@/lib/compare-variant-slug';
 import { getLogoSrc } from '@/lib/og-assets';
-import {
-  buildSeriesPoints,
-  CHART,
-  CHART_FRAME,
-  COLORS,
-  money,
-  moneyForStep,
-  niceAxis,
-  R,
-  renderSeriesPath,
-  SIZE,
-  splitByMatchRange,
-} from '@/lib/png-chart';
+import { renderComparePngChart } from '@/lib/png-chart';
 
 export const dynamic = 'force-dynamic';
 export const runtime = 'nodejs';
@@ -38,8 +24,8 @@ export async function GET(
 ): Promise<Response> {
   const { slug } = await params;
   const parsed = parsePrecisionCompareSlug(slug);
-  // Compare case-insensitively — the HTML pages lowercase before their 308,
-  // so a mixed-case PNG URL should serve rather than 404.
+  // Compare case-insensitively. The HTML pages lowercase before their 308, so
+  // a mixed-case PNG URL should serve rather than 404.
   if (
     !parsed ||
     canonicalPrecisionCompareSlug(parsed.model.slug, parsed.gpu, parsed.precA, parsed.precB) !==
@@ -50,12 +36,10 @@ export async function GET(
 
   const sideA = { precision: parsed.precA };
   const sideB = { precision: parsed.precB };
-
   const [rows, logoSrc] = await Promise.all([
     getCachedBenchmarks(parsed.model.dbKeys),
     getLogoSrc(),
   ]);
-
   const { sequence } = pickVariantPairDefaults('precision', rows, parsed.gpu, sideA, sideB);
   const { ssrRows, interactivityRange } = computeVariantCompareTableData(
     rows,
@@ -72,7 +56,7 @@ export async function GET(
     sideA,
     sideB,
     interactivityRange,
-    plottedRows.map((r) => r.target),
+    plottedRows.map((row) => row.target),
   ).filter((row) => row.a || row.b);
   const curveRows = imageRows.length > 0 ? imageRows : plottedRows;
 
@@ -80,397 +64,21 @@ export async function GET(
   const gpuLabel = gpuMeta?.label ?? parsed.gpu.toUpperCase();
   const aLabel = precisionDisplayLabel(parsed.precA);
   const bLabel = precisionDisplayLabel(parsed.precB);
-  const costs = curveRows
-    .flatMap((row) => [row.a?.cost, row.b?.cost])
-    .filter((cost): cost is number => typeof cost === 'number' && Number.isFinite(cost));
-  const costMin = costs.length > 0 ? Math.min(...costs) : 0;
-  const costMax = costs.length > 0 ? Math.max(...costs) : 1;
-  const yAxis = niceAxis(Math.min(0, costMin), costMax);
-  const yMin = yAxis.min;
-  const yMax = yAxis.max;
-  const yStep = yAxis.step;
-  const xMin = curveRows.at(0)?.target ?? 0;
-  const xMax = curveRows.at(-1)?.target ?? 100;
-  const matchedMin = plottedRows.at(0)?.target ?? xMin;
-  const matchedMax = plottedRows.at(-1)?.target ?? xMax;
-  const hasLeftExtension = matchedMin - xMin >= 0.5;
-  const hasRightExtension = xMax - matchedMax >= 0.5;
-  const scaleX = (value: number) =>
-    CHART.left + (xMax === xMin ? CHART.width / 2 : ((value - xMin) / (xMax - xMin)) * CHART.width);
-  const scaleY = (value: number) =>
-    CHART.top +
-    CHART.height -
-    (yMax === yMin ? CHART.height / 2 : ((value - yMin) / (yMax - yMin)) * CHART.height);
-
-  const aSeries = splitByMatchRange(
-    buildSeriesPoints(curveRows, (r) => r.a?.cost ?? null, scaleX, scaleY),
-    matchedMin,
-    matchedMax,
-  );
-  const bSeries = splitByMatchRange(
-    buildSeriesPoints(curveRows, (r) => r.b?.cost ?? null, scaleX, scaleY),
-    matchedMin,
-    matchedMax,
-  );
-  const aHighlightPoints = plottedRows
-    .filter((row) => row.a)
-    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.a!.cost) }));
-  const bHighlightPoints = plottedRows
-    .filter((row) => row.b)
-    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.b!.cost) }));
-  const workload = sequence ?? '';
-  const showRangeEndpoints = hasLeftExtension || hasRightExtension;
-  const svgWidth = 760 * R;
-  const svgHeight = 406 * R;
 
   try {
-    return new ImageResponse(
-      <div
-        style={{
-          display: 'flex',
-          flexDirection: 'column',
-          width: SIZE.width,
-          height: SIZE.height,
-          padding: `${38 * R}px ${46 * R}px ${26 * R}px`,
-          background: COLORS.background,
-          color: COLORS.text,
-          fontFamily: 'Arial, sans-serif',
-        }}
-      >
-        <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'flex-start' }}>
-          <div style={{ display: 'flex', flexDirection: 'column', gap: 7 * R }}>
-            <div
-              style={{
-                display: 'flex',
-                fontSize: 19 * R,
-                fontWeight: 700,
-                letterSpacing: '0.13em',
-                textTransform: 'uppercase',
-                color: COLORS.blue,
-              }}
-            >
-              InferenceX Precision Comparison
-            </div>
-            <div style={{ display: 'flex', fontSize: 41 * R, fontWeight: 800 }}>
-              {parsed.model.label} · {gpuLabel}
-            </div>
-            <div style={{ display: 'flex', fontSize: 25 * R, color: COLORS.muted }}>
-              {aLabel} vs {bLabel} | Cost per Million Tokens
-            </div>
-          </div>
-          <div
-            style={{
-              display: 'flex',
-              flexDirection: 'column',
-              alignItems: 'flex-end',
-              border: `${R}px solid ${COLORS.border}`,
-              borderRadius: 12 * R,
-              padding: `${13 * R}px ${17 * R}px`,
-              background: COLORS.panel,
-              gap: 5 * R,
-            }}
-          >
-            <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
-              DEFAULT WORKLOAD
-            </div>
-            <div style={{ display: 'flex', fontSize: 21 * R, fontWeight: 700 }}>
-              {workload || 'Default comparison'}
-            </div>
-            <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
-              Lower cost is better
-            </div>
-          </div>
-        </div>
-
-        <div style={{ display: 'flex', flex: 1, gap: 34 * R, marginTop: 22 * R }}>
-          <div
-            style={{ display: 'flex', position: 'relative', width: svgWidth, height: svgHeight }}
-          >
-            <svg
-              width={svgWidth}
-              height={svgHeight}
-              viewBox={`0 0 ${svgWidth} ${svgHeight}`}
-              style={{ position: 'absolute', left: 0, top: 0 }}
-            >
-              <rect
-                x={CHART_FRAME.left}
-                y={CHART_FRAME.top}
-                width={CHART_FRAME.width}
-                height={CHART_FRAME.height}
-                rx={13 * R}
-                fill={COLORS.panel}
-                stroke={COLORS.border}
-              />
-              {yAxis.ticks.map((tick) => {
-                const y = scaleY(tick);
-                return (
-                  <line
-                    key={tick}
-                    x1={CHART.left}
-                    x2={CHART.left + CHART.width}
-                    y1={y}
-                    y2={y}
-                    stroke={COLORS.grid}
-                    strokeWidth={2 * R}
-                  />
-                );
-              })}
-              {plottedRows.map((row) => {
-                const x = scaleX(row.target);
-                return (
-                  <line
-                    key={`mark-${row.target}`}
-                    x1={x}
-                    x2={x}
-                    y1={CHART.top + CHART.height}
-                    y2={CHART.top + CHART.height + 6 * R}
-                    stroke={COLORS.muted}
-                    strokeWidth={2 * R}
-                  />
-                );
-              })}
-              {renderSeriesPath(aSeries.leftExt, COLORS.a, true)}
-              {renderSeriesPath(aSeries.rightExt, COLORS.a, true)}
-              {renderSeriesPath(aSeries.matched, COLORS.a, false)}
-              {renderSeriesPath(bSeries.leftExt, COLORS.b, true)}
-              {renderSeriesPath(bSeries.rightExt, COLORS.b, true)}
-              {renderSeriesPath(bSeries.matched, COLORS.b, false)}
-              {aHighlightPoints.map((point, index) => (
-                <circle
-                  key={`a-${index}`}
-                  cx={point.x}
-                  cy={point.y}
-                  r={10 * R}
-                  fill={COLORS.a}
-                  stroke={COLORS.background}
-                  strokeWidth={4 * R}
-                />
-              ))}
-              {bHighlightPoints.map((point, index) => (
-                <circle
-                  key={`b-${index}`}
-                  cx={point.x}
-                  cy={point.y}
-                  r={10 * R}
-                  fill={COLORS.b}
-                  stroke={COLORS.background}
-                  strokeWidth={4 * R}
-                />
-              ))}
-            </svg>
-            {logoSrc && (
-              <img
-                src={logoSrc}
-                alt=""
-                height={144 * R}
-                style={{
-                  position: 'absolute',
-                  left: CHART.left + CHART.width / 2 - 168 * R,
-                  top: CHART.top + CHART.height / 2 - 72 * R,
-                  opacity: 0.12,
-                }}
-              />
-            )}
-            {yAxis.ticks.map((tick) => (
-              <div
-                key={`y-label-${tick}`}
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: CHART_FRAME.left + 14 * R,
-                  top: scaleY(tick) - 9 * R,
-                  width: CHART.left - CHART_FRAME.left - 28 * R,
-                  justifyContent: 'flex-end',
-                  color: COLORS.muted,
-                  fontSize: 15 * R,
-                }}
-              >
-                {moneyForStep(tick, yStep)}
-              </div>
-            ))}
-            {plottedRows.map((row) => (
-              <div
-                key={`x-label-${row.target}`}
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(row.target) - 32 * R,
-                  top: CHART.top + CHART.height + 15 * R,
-                  width: 64 * R,
-                  justifyContent: 'center',
-                  color: COLORS.muted,
-                  fontSize: 16 * R,
-                  fontWeight: 600,
-                }}
-              >
-                {row.target}
-              </div>
-            ))}
-            {showRangeEndpoints && hasLeftExtension && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(xMin) - 4 * R,
-                  top: CHART.top + CHART.height + 16 * R,
-                  width: 56 * R,
-                  justifyContent: 'flex-start',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                {Math.round(xMin)}
-              </div>
-            )}
-            {showRangeEndpoints && hasRightExtension && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(xMax) - 52 * R,
-                  top: CHART.top + CHART.height + 16 * R,
-                  width: 56 * R,
-                  justifyContent: 'flex-end',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                {Math.round(xMax)}
-              </div>
-            )}
-            <div
-              style={{
-                display: 'flex',
-                position: 'absolute',
-                left: CHART.left,
-                top: CHART.top + CHART.height + 38 * R,
-                width: CHART.width,
-                justifyContent: 'center',
-                color: COLORS.muted,
-                fontSize: 15 * R,
-                fontWeight: 600,
-              }}
-            >
-              Interactivity (tok/s/user)
-            </div>
-            {showRangeEndpoints && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: CHART.left,
-                  top: CHART.top + CHART.height + 62 * R,
-                  width: CHART.width,
-                  justifyContent: 'center',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                Dashed segments extend to each precision's operating envelope
-              </div>
-            )}
-          </div>
-
-          <div
-            style={{
-              display: 'flex',
-              flex: 1,
-              flexDirection: 'column',
-              gap: 17 * R,
-              paddingTop: 18 * R,
-            }}
-          >
-            <div style={{ display: 'flex', fontSize: 18 * R, fontWeight: 700 }}>
-              Matched Interactivity
-            </div>
-            <div style={{ display: 'flex', gap: 20 * R, fontSize: 15 * R, color: COLORS.muted }}>
-              <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
-                <span
-                  style={{
-                    display: 'flex',
-                    width: 19 * R,
-                    height: 6 * R,
-                    borderRadius: 3 * R,
-                    background: COLORS.a,
-                  }}
-                />
-                {aLabel}
-              </span>
-              <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
-                <span
-                  style={{
-                    display: 'flex',
-                    width: 19 * R,
-                    height: 6 * R,
-                    borderRadius: 3 * R,
-                    background: COLORS.b,
-                  }}
-                />
-                {bLabel}
-              </span>
-            </div>
-            {plottedRows.length > 0 ? (
-              plottedRows.map((row) => (
-                <div
-                  key={`row-${row.target}`}
-                  style={{
-                    display: 'flex',
-                    flexDirection: 'column',
-                    gap: 6 * R,
-                    border: `${R}px solid ${COLORS.border}`,
-                    borderRadius: 10 * R,
-                    padding: `${11 * R}px ${13 * R}px`,
-                    background: COLORS.panel,
-                  }}
-                >
-                  <div style={{ display: 'flex', color: COLORS.muted, fontSize: 13 * R }}>
-                    {row.target} tok/s/user
-                  </div>
-                  <div style={{ display: 'flex', gap: 15 * R, fontSize: 19 * R, fontWeight: 700 }}>
-                    <span style={{ display: 'flex', color: COLORS.a }}>
-                      {row.a ? money(row.a.cost) : 'N/A'}
-                    </span>
-                    <span style={{ display: 'flex', color: COLORS.b }}>
-                      {row.b ? money(row.b.cost) : 'N/A'}
-                    </span>
-                  </div>
-                </div>
-              ))
-            ) : (
-              <div style={{ display: 'flex', fontSize: 18 * R, color: COLORS.muted }}>
-                No matched cost data available.
-              </div>
-            )}
-          </div>
-        </div>
-
-        <div
-          style={{
-            display: 'flex',
-            justifyContent: 'space-between',
-            paddingTop: 9 * R,
-            fontSize: 15 * R,
-            color: COLORS.muted,
-          }}
-        >
-          <span style={{ display: 'flex' }}>
-            Precision comparison | interpolated from benchmark results
-          </span>
-          <span style={{ display: 'flex', color: COLORS.text, fontWeight: 700 }}>
-            inferencex.semianalysis.com
-          </span>
-        </div>
-      </div>,
-      {
-        ...SIZE,
-        headers: {
-          'Cache-Control': 'public, s-maxage=3600, stale-while-revalidate=86400',
-        },
-      },
-    );
+    return renderComparePngChart({
+      curveRows,
+      plottedRows,
+      logoSrc,
+      aLabel,
+      bLabel,
+      eyebrow: 'InferenceX Precision Comparison',
+      title: `${parsed.model.label} · ${gpuLabel}`,
+      subtitle: `${aLabel} vs ${bLabel} | Cost per Million Tokens`,
+      workload: sequence ?? '',
+      rangeNote: "Dashed segments extend to each precision's operating envelope",
+      footer: 'Precision comparison | interpolated from benchmark results',
+    });
   } catch (error) {
     const message = error instanceof Error ? error.message : String(error);
     trackServer('compare_precision_png_render_failed', {
diff --git a/packages/app/src/app/compare-precision/layout.tsx b/packages/app/src/app/compare-precision/layout.tsx
index 6ed48e39a..0706bbeb3 100644
--- a/packages/app/src/app/compare-precision/layout.tsx
+++ b/packages/app/src/app/compare-precision/layout.tsx
@@ -1,20 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
-/**
- * Wraps `/compare-precision/*` pages in UnofficialRunProvider but skips
- * DashboardShell so we get a focused single-purpose page (no TabNav). Mirrors
- * the structure of `/compare-per-dollar/layout.tsx` — the GlobalFilterProvider
- * is mounted inside the page's client component so it can seed initial
- * precision/sequence based on which combos actually have data from the slug.
- */
 export default function ComparePrecisionLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/compare-spec-decode/[slug]/[scenario]/page.tsx b/packages/app/src/app/compare-spec-decode/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..ed071d34c
--- /dev/null
+++ b/packages/app/src/app/compare-spec-decode/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildSpecDecodeMetadata, renderSpecDecodePage } from '../page';
+
+/**
+ * `/compare-spec-decode/<slug>/<scenario>` — the same comparison with its workload pinned by
+ * the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments rather than silently falling back to the pair's
+ *  default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildSpecDecodeMetadata(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function SpecDecodeScenarioPage({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderSpecDecodePage(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/compare-spec-decode/[slug]/opengraph-image.tsx b/packages/app/src/app/compare-spec-decode/[slug]/opengraph-image.tsx
index aff210426..be4340a26 100644
--- a/packages/app/src/app/compare-spec-decode/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/compare-spec-decode/[slug]/opengraph-image.tsx
@@ -1,13 +1,8 @@
-/**
- * Spec-decode compare OG image — identical circuit tile sidebar layout to the
- * /compare OG and the blog OG, but the eyebrow and bottom banner are swapped
- * to surface the "Speculative Decoding" framing.
- */
 import { notFound } from 'next/navigation';
-import { ImageResponse } from 'next/og';
 
 import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
 
+import { renderCompareOg } from '@/lib/compare-og';
 import { getAllComparableSpecDecodeSlugs } from '@/lib/compare-variant-availability';
 import {
   canonicalSpecDecodeCompareSlug,
@@ -15,7 +10,6 @@ import {
   precisionDisplayLabel,
   specMethodDisplayLabel,
 } from '@/lib/compare-variant-slug';
-import { BG, BLUE, getLogoSrc, getTiles, PANEL_BG } from '@/lib/og-assets';
 
 export const alt = 'Chip speculative decoding inference benchmark comparison';
 export const size = { width: 1200, height: 630 };
@@ -31,9 +25,6 @@ export async function generateStaticParams() {
 export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
   const { slug } = await params;
   const parsed = parseSpecDecodeCompareSlug(slug);
-  // Reject non-canonical slugs (reversed order, aliases, mixed case) so OG
-  // URLs can't render labels that disagree with the canonical page — same
-  // strictness as the sibling PNG route.
   if (
     !parsed ||
     canonicalSpecDecodeCompareSlug(
@@ -45,127 +36,19 @@ export default async function OgImage({ params }: { params: Promise<{ slug: stri
   ) {
     notFound();
   }
-  const [logoSrc, tiles] = await Promise.all([getLogoSrc(), getTiles()]);
 
-  const gpuMeta = HW_REGISTRY[parsed.gpu];
-  const gpuLabel = gpuMeta?.label ?? parsed.gpu.toUpperCase();
-  const precLabel = precisionDisplayLabel(parsed.precision);
-  const aLabel = specMethodDisplayLabel(parsed.model.displayName, parsed.method);
-  const title = `${gpuLabel} ${precLabel}: ${aLabel} vs Off`;
-  const eyebrow = `${parsed.model.label} · Speculative Decoding`;
+  const gpuLabel = HW_REGISTRY[parsed.gpu]?.label ?? parsed.gpu.toUpperCase();
+  const precisionLabel = precisionDisplayLabel(parsed.precision);
+  const methodLabel = specMethodDisplayLabel(parsed.model.displayName, parsed.method);
+  const title = `${gpuLabel} ${precisionLabel}: ${methodLabel} vs Off`;
   const titleSize = title.length > 34 ? 72 : title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
 
-  return new ImageResponse(
-    <div
-      style={{
-        display: 'flex',
-        width: '100%',
-        height: '100%',
-        backgroundColor: BG,
-        color: '#EAEBEC',
-        overflow: 'hidden',
-      }}
-    >
-      <div
-        style={{
-          display: 'flex',
-          width: 195,
-          height: '100%',
-          backgroundColor: PANEL_BG,
-          position: 'relative',
-        }}
-      >
-        {tiles.map((tile, i) => {
-          if (!tile) return null;
-          const row = Math.floor(i / 2);
-          const col = i % 2;
-          return (
-            <img
-              key={i}
-              src={tile.src}
-              style={{
-                position: 'absolute',
-                left: 12 + col * 90,
-                top: 12 + row * 104,
-                width: 78,
-                height: 86,
-                borderRadius: 4,
-                objectFit: 'cover',
-                ...(tile.rotate ? { transform: `rotate(${tile.rotate}deg)` } : {}),
-              }}
-            />
-          );
-        })}
-        <div
-          style={{
-            position: 'absolute',
-            right: 0,
-            top: 0,
-            width: 3,
-            height: '100%',
-            backgroundColor: BLUE,
-            display: 'flex',
-          }}
-        />
-      </div>
-
-      <div
-        style={{
-          display: 'flex',
-          flexDirection: 'column',
-          justifyContent: 'space-between',
-          flex: 1,
-          padding: '48px 55px 20px 55px',
-        }}
-      >
-        <div
-          style={{
-            display: 'flex',
-            flexDirection: 'column',
-            gap: 24,
-            flex: 1,
-            justifyContent: 'center',
-          }}
-        >
-          <div
-            style={{
-              fontSize: 24,
-              color: '#9BA0A6',
-              letterSpacing: '0.18em',
-              textTransform: 'uppercase',
-              display: 'flex',
-            }}
-          >
-            {eyebrow}
-          </div>
-
-          <div
-            style={{
-              fontSize: titleSize,
-              fontWeight: 800,
-              color: '#FFFFFF',
-              lineHeight: 1.1,
-              display: 'flex',
-            }}
-          >
-            {title}
-          </div>
-        </div>
-
-        <div
-          style={{
-            display: 'flex',
-            justifyContent: 'space-between',
-            alignItems: 'center',
-          }}
-        >
-          <span style={{ fontSize: 26, color: '#9BA0A6', display: 'flex' }}>
-            AI inference benchmark · speculative decoding comparison
-          </span>
-          {logoSrc && <img src={logoSrc} height={72} />}
-        </div>
-      </div>
-    </div>,
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · Speculative Decoding`,
+    title,
+    titleSize,
+    footer: 'AI inference benchmark · speculative decoding comparison',
+    language: 'en',
     size,
-  );
+  });
 }
diff --git a/packages/app/src/app/compare-spec-decode/[slug]/page-client.tsx b/packages/app/src/app/compare-spec-decode/[slug]/page-client.tsx
index 4a9f11a6b..fff64ec8c 100644
--- a/packages/app/src/app/compare-spec-decode/[slug]/page-client.tsx
+++ b/packages/app/src/app/compare-spec-decode/[slug]/page-client.tsx
@@ -3,22 +3,24 @@
 import Link from 'next/link';
 import { useEffect, useMemo } from 'react';
 
-import type { GPUDataPoint, InterpolatedResult } from '@/components/calculator/types';
+import type { GPUDataPoint } from '@/components/calculator/types';
 import { useThroughputData } from '@/components/calculator/useThroughputData';
-import { CompareInterpolatedTable } from '@/components/compare/compare-interpolated-table';
-import { useGlobalFilters, GlobalFilterProvider } from '@/components/GlobalFilterContext';
+import {
+  CompareTableRenderer,
+  type CompareTableData,
+} from '@/components/compare/compare-table-renderer';
+import {
+  GlobalFilterProvider,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+} from '@/components/GlobalFilterContext';
 import { InferenceProvider } from '@/components/inference/InferenceContext';
 import InferenceChartDisplay from '@/components/inference/ui/ChartDisplay';
 import { Card } from '@/components/ui/card';
 import { track } from '@/lib/analytics';
 import { toModel, toPrecisions, toSequence } from '@/lib/compare-enum-coerce';
 import { SPEC_METHODS_ACTIVE } from '@/lib/compare-variant-slug';
-
-interface SsrTableData {
-  defaultTargets: number[];
-  ssrRows: { target: number; a: InterpolatedResult | null; b: InterpolatedResult | null }[];
-  interactivityRange: { min: number; max: number };
-}
+import type { AgenticScenarioIntro } from '@/lib/compare-ssr';
 
 const STRINGS = {
   en: {
@@ -62,8 +64,11 @@ interface CompareSpecDecodePageClientProps {
   modelDisplayName: string;
   defaultSequence: string | null;
   defaultPrecision: string | null;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
   narrative: string[];
+  /** Set only when the page is rendering the agentic workload. Explains
+   *  what AgentX measures before the head-to-head numbers. */
+  agenticIntro?: AgenticScenarioIntro | null;
   gpuLabel: string;
   precisionLabel: string;
   gpuArch: string;
@@ -84,6 +89,7 @@ export default function CompareSpecDecodePageClient({
   defaultPrecision,
   ssrTableData,
   narrative,
+  agenticIntro = null,
   gpuLabel,
   precisionLabel,
   gpuArch,
@@ -163,6 +169,21 @@ export default function CompareSpecDecodePageClient({
                   </p>
                 </div>
               </div>
+              {agenticIntro && (
+                <p
+                  className="mt-3 max-w-3xl text-sm text-foreground/80"
+                  data-testid="compare-agentic-intro"
+                >
+                  {agenticIntro.paragraph}{' '}
+                  <Link
+                    href={agenticIntro.href}
+                    data-testid="compare-agentic-intro-link"
+                    className="font-medium text-brand underline underline-offset-4 hover:no-underline"
+                  >
+                    {agenticIntro.linkLabel} →
+                  </Link>
+                </p>
+              )}
               {narrative.length > 0 && (
                 <div
                   className="mt-3 flex flex-col gap-2"
@@ -256,11 +277,11 @@ function CompareTableSection({
   precision: string | null;
   aLabel: string;
   bLabel: string;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
   emptyStateText: string;
 }) {
-  const { effectiveSequence, effectivePrecisions, selectedRunDate, selectedModel } =
-    useGlobalFilters();
+  const { effectiveSequence, effectivePrecisions, selectedModel } = useGlobalFilterSelection();
+  const { selectedRunDate } = useGlobalFilterRun();
 
   const { gpuDataByGroupKey, ranges, hasData } = useThroughputData(
     selectedModel,
@@ -297,23 +318,15 @@ function CompareTableSection({
 
   const clientRange = hasData ? ranges.interactivity : ssrTableData.interactivityRange;
 
-  if (ssrTableData.defaultTargets.length === 0) {
-    return (
-      <div className="border border-border/50 rounded-md px-4 py-3 text-sm text-muted-foreground bg-muted/30">
-        {emptyStateText}
-      </div>
-    );
-  }
-
   return (
-    <CompareInterpolatedTable
+    <CompareTableRenderer
       aLabel={aLabel}
       bLabel={bLabel}
-      ssrRows={ssrTableData.ssrRows}
-      defaultTargets={ssrTableData.defaultTargets}
+      ssrTableData={ssrTableData}
       interactivityRange={clientRange}
       gpuDataPointsA={pointsA}
       gpuDataPointsB={pointsB}
+      emptyStateText={emptyStateText}
     />
   );
 }
diff --git a/packages/app/src/app/compare-spec-decode/[slug]/page.tsx b/packages/app/src/app/compare-spec-decode/[slug]/page.tsx
index 352c14fb8..06b86055a 100644
--- a/packages/app/src/app/compare-spec-decode/[slug]/page.tsx
+++ b/packages/app/src/app/compare-spec-decode/[slug]/page.tsx
@@ -9,8 +9,18 @@ import {
 } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
 import { languageAlternates } from '@/lib/i18n';
-import { getCachedBenchmarks, KNOWN_SEQUENCES, pickString } from '@/lib/compare-ssr';
+import {
+  AGENTIC_SCENARIO_INTRO,
+  getCachedBenchmarks,
+  KNOWN_SEQUENCES,
+  pickString,
+} from '@/lib/compare-ssr';
 import {
   canonicalSpecDecodeCompareSlug,
   parseSpecDecodeCompareSlug,
@@ -32,6 +42,22 @@ import CompareSpecDecodePageClient from './page-client';
 
 export const dynamic = 'force-dynamic';
 
+/**
+ * `/compare-spec-decode/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than duplicated
+ * in a parallel route file so the body — redirects, JSON-LD, metadata, client
+ * props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
+function scenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/compare-spec-decode/${canonical}/${scenarioSegment}`
+    : `/compare-spec-decode/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
@@ -39,6 +65,13 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildSpecDecodeMetadata(slug, {});
+}
+
+export function buildSpecDecodeMetadata(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Metadata {
   const parsed = parseSpecDecodeCompareSlug(slug);
   if (!parsed) return {};
   const gpuMeta = HW_REGISTRY[parsed.gpu];
@@ -51,15 +84,22 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     parsed.precision,
     parsed.method,
   );
-  const url = `${SITE_URL}/compare-spec-decode/${canonical}`;
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const canonicalPath = scenarioPath(canonical);
+  const url = `${SITE_URL}${routePath}`;
+  const canonicalUrl = `${SITE_URL}${canonicalPath}`;
   const description = `${parsed.model.label} on ${gpuLabel} ${precLabel}: ${aLabel} vs Off speculative decoding comparison. Verified, reproducible results from InferenceX, the independent open-source benchmark by SemiAnalysis. ${SUPPORTERS_LINE} See whether speculative decoding improves throughput and cost at every interactivity level.`;
   const title = `${parsed.model.label} — ${gpuLabel} ${precLabel}: ${aLabel} vs Off — Speculative Decoding`;
   return {
     title,
     description,
     alternates: {
-      canonical: url,
-      languages: languageAlternates(`/compare-spec-decode/${canonical}`),
+      canonical: canonicalUrl,
+      languages: languageAlternates(canonicalPath),
     },
     openGraph: {
       title: `${title} | ${SITE_NAME}`,
@@ -77,11 +117,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function CompareSpecDecodePage({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderSpecDecodePage(slug, await searchParams, {});
+}
+
+export async function renderSpecDecodePage(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parseSpecDecodeCompareSlug(slug);
   if (!parsed) notFound();
 
-  const sp = await searchParams;
-
   // 308 canonical redirect: alias model, reversed 'none-vs-mtp' form.
   const canonical = canonicalSpecDecodeCompareSlug(
     parsed.model.slug,
@@ -98,7 +144,9 @@ export default async function CompareSpecDecodePage({ params, searchParams }: Pr
       })
       .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
       .join('&');
-    permanentRedirect(`/compare-spec-decode/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // the reader to the pair's default workload instead.
+    permanentRedirect(`${scenarioPath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
   const rows = await getCachedBenchmarks(parsed.model.dbKeys);
@@ -115,7 +163,9 @@ export default async function CompareSpecDecodePage({ params, searchParams }: Pr
   const defaults = pickVariantPairDefaults('spec-decode', rows, parsed.gpu, sideA, sideB);
 
   const urlSeq = pickString(sp.i_seq);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : defaults.sequence;
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const effectiveSequence =
+    pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : defaults.sequence);
   const effectivePrecision = parsed.precision;
 
   const sideAFull: VariantCompareSide = {
@@ -139,7 +189,7 @@ export default async function CompareSpecDecodePage({ params, searchParams }: Pr
   const summaryB = summarizeVariantSide(rows, parsed.gpu, sideBFull);
   const { oldest, newest } = dateRangeForVariantPair(rows, parsed.gpu, sideAFull, sideBFull);
 
-  const url = `${SITE_URL}/compare-spec-decode/${canonical}`;
+  const url = `${SITE_URL}${scenarioPath(canonical, scenarioSegment)}`;
   const imageUrl = `${url}/spec-decode-comparison.png`;
 
   const jsonLd = buildVariantJsonLd(
@@ -185,6 +235,7 @@ export default async function CompareSpecDecodePage({ params, searchParams }: Pr
         defaultPrecision={effectivePrecision}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO : null}
         gpuLabel={gpuLabel}
         precisionLabel={precLabel}
         gpuArch={gpuMeta?.arch ?? ''}
diff --git a/packages/app/src/app/compare-spec-decode/[slug]/spec-decode-comparison.png/route.tsx b/packages/app/src/app/compare-spec-decode/[slug]/spec-decode-comparison.png/route.tsx
index 529e3d4d8..2e779fc38 100644
--- a/packages/app/src/app/compare-spec-decode/[slug]/spec-decode-comparison.png/route.tsx
+++ b/packages/app/src/app/compare-spec-decode/[slug]/spec-decode-comparison.png/route.tsx
@@ -1,5 +1,3 @@
-import { ImageResponse } from 'next/og';
-
 import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
 
 import { trackServer } from '@/lib/analytics-server';
@@ -17,19 +15,7 @@ import {
   specMethodDisplayLabel,
 } from '@/lib/compare-variant-slug';
 import { getLogoSrc } from '@/lib/og-assets';
-import {
-  buildSeriesPoints,
-  CHART,
-  CHART_FRAME,
-  COLORS,
-  money,
-  moneyForStep,
-  niceAxis,
-  R,
-  renderSeriesPath,
-  SIZE,
-  splitByMatchRange,
-} from '@/lib/png-chart';
+import { renderComparePngChart } from '@/lib/png-chart';
 
 export const dynamic = 'force-dynamic';
 export const runtime = 'nodejs';
@@ -56,14 +42,11 @@ export async function GET(
     getCachedBenchmarks(parsed.model.dbKeys),
     getLogoSrc(),
   ]);
-
-  // Precision is fixed by the slug — both sides share it.
+  // Precision is fixed by the slug, and both sides share it.
   const sideA: VariantCompareSide = { specMethod: parsed.method, precision: parsed.precision };
   const sideB: VariantCompareSide = { specMethod: 'none', precision: parsed.precision };
-  const defaults = pickVariantPairDefaults('spec-decode', rows, parsed.gpu, sideA, sideB);
-  const { sequence } = defaults;
+  const { sequence } = pickVariantPairDefaults('spec-decode', rows, parsed.gpu, sideA, sideB);
   const precision = parsed.precision;
-
   const { ssrRows, interactivityRange } = computeVariantCompareTableData(
     rows,
     parsed.gpu,
@@ -79,7 +62,7 @@ export async function GET(
     sideA,
     sideB,
     interactivityRange,
-    plottedRows.map((r) => r.target),
+    plottedRows.map((row) => row.target),
   ).filter((row) => row.a || row.b);
   const curveRows = imageRows.length > 0 ? imageRows : plottedRows;
 
@@ -87,398 +70,23 @@ export async function GET(
   const precLabel = precisionDisplayLabel(parsed.precision);
   const aLabel = specMethodDisplayLabel(parsed.model.displayName, parsed.method);
   const bLabel = 'Off';
-
-  const costs = curveRows
-    .flatMap((row) => [row.a?.cost, row.b?.cost])
-    .filter((cost): cost is number => typeof cost === 'number' && Number.isFinite(cost));
-  const costMin = costs.length > 0 ? Math.min(...costs) : 0;
-  const costMax = costs.length > 0 ? Math.max(...costs) : 1;
-  const yAxis = niceAxis(Math.min(0, costMin), costMax);
-  const yMin = yAxis.min;
-  const yMax = yAxis.max;
-  const yStep = yAxis.step;
-  const xMin = curveRows.at(0)?.target ?? 0;
-  const xMax = curveRows.at(-1)?.target ?? 100;
-  const matchedMin = plottedRows.at(0)?.target ?? xMin;
-  const matchedMax = plottedRows.at(-1)?.target ?? xMax;
-  const hasLeftExtension = matchedMin - xMin >= 0.5;
-  const hasRightExtension = xMax - matchedMax >= 0.5;
-  const scaleX = (value: number) =>
-    CHART.left + (xMax === xMin ? CHART.width / 2 : ((value - xMin) / (xMax - xMin)) * CHART.width);
-  const scaleY = (value: number) =>
-    CHART.top +
-    CHART.height -
-    (yMax === yMin ? CHART.height / 2 : ((value - yMin) / (yMax - yMin)) * CHART.height);
-
-  const aSeries = splitByMatchRange(
-    buildSeriesPoints(curveRows, (r) => r.a?.cost ?? null, scaleX, scaleY),
-    matchedMin,
-    matchedMax,
-  );
-  const bSeries = splitByMatchRange(
-    buildSeriesPoints(curveRows, (r) => r.b?.cost ?? null, scaleX, scaleY),
-    matchedMin,
-    matchedMax,
-  );
-  const aHighlightPoints = plottedRows
-    .filter((row) => row.a)
-    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.a!.cost) }));
-  const bHighlightPoints = plottedRows
-    .filter((row) => row.b)
-    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.b!.cost) }));
   const workload = [sequence, precision?.toUpperCase()].filter(Boolean).join(' / ');
-  const showRangeEndpoints = hasLeftExtension || hasRightExtension;
-  const svgWidth = 760 * R;
-  const svgHeight = 406 * R;
 
   try {
-    return new ImageResponse(
-      <div
-        style={{
-          display: 'flex',
-          flexDirection: 'column',
-          width: SIZE.width,
-          height: SIZE.height,
-          padding: `${38 * R}px ${46 * R}px ${26 * R}px`,
-          background: COLORS.background,
-          color: COLORS.text,
-          fontFamily: 'Arial, sans-serif',
-        }}
-      >
-        <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'flex-start' }}>
-          <div style={{ display: 'flex', flexDirection: 'column', gap: 7 * R }}>
-            <div
-              style={{
-                display: 'flex',
-                fontSize: 19 * R,
-                fontWeight: 700,
-                letterSpacing: '0.13em',
-                textTransform: 'uppercase',
-                color: COLORS.blue,
-              }}
-            >
-              InferenceX Speculative Decoding Comparison
-            </div>
-            <div style={{ display: 'flex', fontSize: 41 * R, fontWeight: 800 }}>
-              {parsed.model.label}
-            </div>
-            <div style={{ display: 'flex', fontSize: 25 * R, color: COLORS.muted }}>
-              {gpuLabel} {precLabel}: {aLabel} vs {bLabel} | Cost per Million Tokens
-            </div>
-          </div>
-          <div
-            style={{
-              display: 'flex',
-              flexDirection: 'column',
-              alignItems: 'flex-end',
-              border: `${R}px solid ${COLORS.border}`,
-              borderRadius: 12 * R,
-              padding: `${13 * R}px ${17 * R}px`,
-              background: COLORS.panel,
-              gap: 5 * R,
-            }}
-          >
-            <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
-              DEFAULT WORKLOAD
-            </div>
-            <div style={{ display: 'flex', fontSize: 21 * R, fontWeight: 700 }}>
-              {workload || 'Default comparison'}
-            </div>
-            <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
-              Lower cost is better
-            </div>
-          </div>
-        </div>
-
-        <div style={{ display: 'flex', flex: 1, gap: 34 * R, marginTop: 22 * R }}>
-          <div
-            style={{ display: 'flex', position: 'relative', width: svgWidth, height: svgHeight }}
-          >
-            <svg
-              width={svgWidth}
-              height={svgHeight}
-              viewBox={`0 0 ${svgWidth} ${svgHeight}`}
-              style={{ position: 'absolute', left: 0, top: 0 }}
-            >
-              <rect
-                x={CHART_FRAME.left}
-                y={CHART_FRAME.top}
-                width={CHART_FRAME.width}
-                height={CHART_FRAME.height}
-                rx={13 * R}
-                fill={COLORS.panel}
-                stroke={COLORS.border}
-              />
-              {yAxis.ticks.map((tick) => {
-                const y = scaleY(tick);
-                return (
-                  <line
-                    key={tick}
-                    x1={CHART.left}
-                    x2={CHART.left + CHART.width}
-                    y1={y}
-                    y2={y}
-                    stroke={COLORS.grid}
-                    strokeWidth={2 * R}
-                  />
-                );
-              })}
-              {plottedRows.map((row) => {
-                const x = scaleX(row.target);
-                return (
-                  <line
-                    key={`mark-${row.target}`}
-                    x1={x}
-                    x2={x}
-                    y1={CHART.top + CHART.height}
-                    y2={CHART.top + CHART.height + 6 * R}
-                    stroke={COLORS.muted}
-                    strokeWidth={2 * R}
-                  />
-                );
-              })}
-              {renderSeriesPath(aSeries.leftExt, COLORS.a, true)}
-              {renderSeriesPath(aSeries.rightExt, COLORS.a, true)}
-              {renderSeriesPath(aSeries.matched, COLORS.a, false)}
-              {renderSeriesPath(bSeries.leftExt, COLORS.b, true)}
-              {renderSeriesPath(bSeries.rightExt, COLORS.b, true)}
-              {renderSeriesPath(bSeries.matched, COLORS.b, false)}
-              {aHighlightPoints.map((point, index) => (
-                <circle
-                  key={`a-${index}`}
-                  cx={point.x}
-                  cy={point.y}
-                  r={10 * R}
-                  fill={COLORS.a}
-                  stroke={COLORS.background}
-                  strokeWidth={4 * R}
-                />
-              ))}
-              {bHighlightPoints.map((point, index) => (
-                <circle
-                  key={`b-${index}`}
-                  cx={point.x}
-                  cy={point.y}
-                  r={10 * R}
-                  fill={COLORS.b}
-                  stroke={COLORS.background}
-                  strokeWidth={4 * R}
-                />
-              ))}
-            </svg>
-            {logoSrc && (
-              <img
-                src={logoSrc}
-                alt=""
-                height={144 * R}
-                style={{
-                  position: 'absolute',
-                  left: CHART.left + CHART.width / 2 - 168 * R,
-                  top: CHART.top + CHART.height / 2 - 72 * R,
-                  opacity: 0.12,
-                }}
-              />
-            )}
-            {yAxis.ticks.map((tick) => (
-              <div
-                key={`y-label-${tick}`}
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: CHART_FRAME.left + 14 * R,
-                  top: scaleY(tick) - 9 * R,
-                  width: CHART.left - CHART_FRAME.left - 28 * R,
-                  justifyContent: 'flex-end',
-                  color: COLORS.muted,
-                  fontSize: 15 * R,
-                }}
-              >
-                {moneyForStep(tick, yStep)}
-              </div>
-            ))}
-            {plottedRows.map((row) => (
-              <div
-                key={`x-label-${row.target}`}
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(row.target) - 32 * R,
-                  top: CHART.top + CHART.height + 15 * R,
-                  width: 64 * R,
-                  justifyContent: 'center',
-                  color: COLORS.muted,
-                  fontSize: 16 * R,
-                  fontWeight: 600,
-                }}
-              >
-                {row.target}
-              </div>
-            ))}
-            {showRangeEndpoints && hasLeftExtension && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(xMin) - 4 * R,
-                  top: CHART.top + CHART.height + 16 * R,
-                  width: 56 * R,
-                  justifyContent: 'flex-start',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                {Math.round(xMin)}
-              </div>
-            )}
-            {showRangeEndpoints && hasRightExtension && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: scaleX(xMax) - 52 * R,
-                  top: CHART.top + CHART.height + 16 * R,
-                  width: 56 * R,
-                  justifyContent: 'flex-end',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                {Math.round(xMax)}
-              </div>
-            )}
-            <div
-              style={{
-                display: 'flex',
-                position: 'absolute',
-                left: CHART.left,
-                top: CHART.top + CHART.height + 38 * R,
-                width: CHART.width,
-                justifyContent: 'center',
-                color: COLORS.muted,
-                fontSize: 15 * R,
-                fontWeight: 600,
-              }}
-            >
-              Interactivity (tok/s/user)
-            </div>
-            {showRangeEndpoints && (
-              <div
-                style={{
-                  display: 'flex',
-                  position: 'absolute',
-                  left: CHART.left,
-                  top: CHART.top + CHART.height + 62 * R,
-                  width: CHART.width,
-                  justifyContent: 'center',
-                  color: COLORS.faint,
-                  fontSize: 13 * R,
-                  fontStyle: 'italic',
-                }}
-              >
-                Dashed segments extend to each config's operating envelope, where cost rises steeply
-              </div>
-            )}
-          </div>
-
-          <div
-            style={{
-              display: 'flex',
-              flex: 1,
-              flexDirection: 'column',
-              gap: 17 * R,
-              paddingTop: 18 * R,
-            }}
-          >
-            <div style={{ display: 'flex', fontSize: 18 * R, fontWeight: 700 }}>
-              Matched Interactivity
-            </div>
-            <div style={{ display: 'flex', gap: 20 * R, fontSize: 15 * R, color: COLORS.muted }}>
-              <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
-                <span
-                  style={{
-                    display: 'flex',
-                    width: 19 * R,
-                    height: 6 * R,
-                    borderRadius: 3 * R,
-                    background: COLORS.a,
-                  }}
-                />
-                {aLabel}
-              </span>
-              <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
-                <span
-                  style={{
-                    display: 'flex',
-                    width: 19 * R,
-                    height: 6 * R,
-                    borderRadius: 3 * R,
-                    background: COLORS.b,
-                  }}
-                />
-                {bLabel}
-              </span>
-            </div>
-            {plottedRows.length > 0 ? (
-              plottedRows.map((row) => (
-                <div
-                  key={`row-${row.target}`}
-                  style={{
-                    display: 'flex',
-                    flexDirection: 'column',
-                    gap: 6 * R,
-                    border: `${R}px solid ${COLORS.border}`,
-                    borderRadius: 10 * R,
-                    padding: `${11 * R}px ${13 * R}px`,
-                    background: COLORS.panel,
-                  }}
-                >
-                  <div style={{ display: 'flex', color: COLORS.muted, fontSize: 13 * R }}>
-                    {row.target} tok/s/user
-                  </div>
-                  <div style={{ display: 'flex', gap: 15 * R, fontSize: 19 * R, fontWeight: 700 }}>
-                    <span style={{ display: 'flex', color: COLORS.a }}>
-                      {row.a ? money(row.a.cost) : 'N/A'}
-                    </span>
-                    <span style={{ display: 'flex', color: COLORS.b }}>
-                      {row.b ? money(row.b.cost) : 'N/A'}
-                    </span>
-                  </div>
-                </div>
-              ))
-            ) : (
-              <div style={{ display: 'flex', fontSize: 18 * R, color: COLORS.muted }}>
-                No matched cost data available.
-              </div>
-            )}
-          </div>
-        </div>
-
-        <div
-          style={{
-            display: 'flex',
-            justifyContent: 'space-between',
-            paddingTop: 9 * R,
-            fontSize: 15 * R,
-            color: COLORS.muted,
-          }}
-        >
-          <span style={{ display: 'flex' }}>
-            Speculative decoding comparison | interpolated from benchmark results
-          </span>
-          <span style={{ display: 'flex', color: COLORS.text, fontWeight: 700 }}>
-            inferencex.semianalysis.com
-          </span>
-        </div>
-      </div>,
-      {
-        ...SIZE,
-        headers: {
-          'Cache-Control': 'public, s-maxage=3600, stale-while-revalidate=86400',
-        },
-      },
-    );
+    return renderComparePngChart({
+      curveRows,
+      plottedRows,
+      logoSrc,
+      aLabel,
+      bLabel,
+      eyebrow: 'InferenceX Speculative Decoding Comparison',
+      title: parsed.model.label,
+      subtitle: `${gpuLabel} ${precLabel}: ${aLabel} vs ${bLabel} | Cost per Million Tokens`,
+      workload,
+      rangeNote:
+        "Dashed segments extend to each config's operating envelope, where cost rises steeply",
+      footer: 'Speculative decoding comparison | interpolated from benchmark results',
+    });
   } catch (error) {
     const message = error instanceof Error ? error.message : String(error);
     trackServer('compare_spec_decode_png_render_failed', {
diff --git a/packages/app/src/app/compare-spec-decode/layout.tsx b/packages/app/src/app/compare-spec-decode/layout.tsx
index ad681bd9a..f0a2fdb0d 100644
--- a/packages/app/src/app/compare-spec-decode/layout.tsx
+++ b/packages/app/src/app/compare-spec-decode/layout.tsx
@@ -1,20 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
-/**
- * Wraps `/compare-spec-decode/*` pages in UnofficialRunProvider but skips
- * DashboardShell so we get a focused single-purpose page (no TabNav). Mirrors
- * the structure of `/compare-per-dollar/layout.tsx` — the GlobalFilterProvider
- * is mounted inside the page's client component so it can seed initial
- * precision/sequence based on which combos actually have data from the slug.
- */
 export default function CompareSpecDecodeLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/compare/[slug]/[scenario]/page.tsx b/packages/app/src/app/compare/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..3d97fd9e8
--- /dev/null
+++ b/packages/app/src/app/compare/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildCompareMetadata, renderComparePage } from '../page';
+
+/**
+ * `/compare/<slug>/<scenario>` — the same comparison with its workload pinned
+ * by the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments here rather than silently falling back to the
+ *  pair's default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildCompareMetadata(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function CompareScenarioPage({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderComparePage(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/compare/[slug]/opengraph-image.tsx b/packages/app/src/app/compare/[slug]/opengraph-image.tsx
index 709b7e975..0e7762483 100644
--- a/packages/app/src/app/compare/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/compare/[slug]/opengraph-image.tsx
@@ -1,206 +1,32 @@
-/**
- * Compare OG image — same circuit tile sidebar layout as the blog OG.
- */
-import { readFile } from 'node:fs/promises';
-import { join } from 'node:path';
-
 import { notFound } from 'next/navigation';
-import { ImageResponse } from 'next/og';
 
 import { getAllComparableCompareSlugs } from '@/lib/compare-availability';
+import { renderCompareOg } from '@/lib/compare-og';
 import { canonicalCompareSlug, compareDisplayLabel, parseCompareSlug } from '@/lib/compare-slug';
 
 export const alt = 'Chip inference benchmark comparison';
 export const size = { width: 1200, height: 630 };
 export const contentType = 'image/png';
 
-const BLUE = '#0B86D1';
-const BG = '#131416';
-const PANEL_BG = '#0F1214';
-
-// Same tile grid as the blog OG so the two OG types read as a family.
-const TILE_GRID: ({ file: string; rotate?: number } | null)[] = [
-  { file: 'teal-chevron.png', rotate: 180 },
-  { file: 'gold-diagonal.png' },
-  { file: 'teal-circuit.png' },
-  null,
-  { file: 'gold-wavy.png' },
-  { file: 'teal-chip.png' },
-  { file: 'teal-chevron.png', rotate: 90 },
-  { file: 'teal-organic.png' },
-  null,
-  { file: 'gold-circuit.png' },
-  { file: 'teal-circuit.png', rotate: 180 },
-  { file: 'teal-organic.png', rotate: 180 },
-];
-
 export async function generateStaticParams() {
-  // Mirror the SSR page's static params — only emit (model, pair) combos
-  // with benchmark data on both sides so we don't generate OG images for
-  // empty pages.
   const slugs = await getAllComparableCompareSlugs();
   return slugs.map(({ modelSlug, a, b }) => ({ slug: canonicalCompareSlug(modelSlug, a, b) }));
 }
 
-// Read once at module load; a missing asset must not 500 every OG route.
-let logoSrcPromise: Promise<string | null> | undefined;
-function getLogoSrc(): Promise<string | null> {
-  if (!logoSrcPromise) {
-    logoSrcPromise = readFile(join(process.cwd(), 'public/brand/logo-color.png'))
-      .then((buf) => `data:image/png;base64,${buf.toString('base64')}`)
-      .catch(() => null);
-  }
-  return logoSrcPromise;
-}
-
-let tilesPromise: Promise<({ src: string; rotate?: number } | null)[]> | undefined;
-function getTiles(): Promise<({ src: string; rotate?: number } | null)[]> {
-  if (!tilesPromise) {
-    const uniqueFiles = [...new Set(TILE_GRID.filter(Boolean).map((t) => t!.file))];
-    tilesPromise = Promise.all(
-      uniqueFiles.map(async (f) => {
-        const src = await readFile(join(process.cwd(), 'public/brand/og-tiles', f))
-          .then((buf) => `data:image/png;base64,${buf.toString('base64')}`)
-          .catch(() => null);
-        return [f, src] as const;
-      }),
-    ).then((loaded) => {
-      const cache = Object.fromEntries(loaded);
-      return TILE_GRID.map((t) => {
-        if (!t) return null;
-        const src = cache[t.file];
-        return src ? { src, rotate: t.rotate } : null;
-      });
-    });
-  }
-  return tilesPromise;
-}
-
 export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
   const { slug } = await params;
   const parsed = parseCompareSlug(slug);
   if (!parsed) notFound();
-  const [logoSrc, tiles] = await Promise.all([getLogoSrc(), getTiles()]);
 
   const title = compareDisplayLabel(parsed.a, parsed.b);
-  const eyebrow = `${parsed.model.label} · Head-to-head Chip benchmark`;
-  // Content area is ~895px wide (1200 - 195 panel - 55*2 padding). Scale the
-  // title size down for longer labels so it fits without truncating.
   const titleSize = title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
 
-  return new ImageResponse(
-    <div
-      style={{
-        display: 'flex',
-        width: '100%',
-        height: '100%',
-        backgroundColor: BG,
-        color: '#EAEBEC',
-        overflow: 'hidden',
-      }}
-    >
-      {/* Left tile panel — identical pattern to blog OG */}
-      <div
-        style={{
-          display: 'flex',
-          width: 195,
-          height: '100%',
-          backgroundColor: PANEL_BG,
-          position: 'relative',
-        }}
-      >
-        {tiles.map((tile, i) => {
-          if (!tile) return null;
-          const row = Math.floor(i / 2);
-          const col = i % 2;
-          return (
-            <img
-              key={i}
-              src={tile.src}
-              style={{
-                position: 'absolute',
-                left: 12 + col * 90,
-                top: 12 + row * 104,
-                width: 78,
-                height: 86,
-                borderRadius: 4,
-                objectFit: 'cover',
-                ...(tile.rotate ? { transform: `rotate(${tile.rotate}deg)` } : {}),
-              }}
-            />
-          );
-        })}
-        <div
-          style={{
-            position: 'absolute',
-            right: 0,
-            top: 0,
-            width: 3,
-            height: '100%',
-            backgroundColor: BLUE,
-            display: 'flex',
-          }}
-        />
-      </div>
-
-      {/* Content */}
-      <div
-        style={{
-          display: 'flex',
-          flexDirection: 'column',
-          justifyContent: 'space-between',
-          flex: 1,
-          padding: '48px 55px 20px 55px',
-        }}
-      >
-        <div
-          style={{
-            display: 'flex',
-            flexDirection: 'column',
-            gap: 24,
-            flex: 1,
-            justifyContent: 'center',
-          }}
-        >
-          <div
-            style={{
-              fontSize: 24,
-              color: '#9BA0A6',
-              letterSpacing: '0.18em',
-              textTransform: 'uppercase',
-              display: 'flex',
-            }}
-          >
-            {eyebrow}
-          </div>
-
-          <div
-            style={{
-              fontSize: titleSize,
-              fontWeight: 800,
-              color: '#FFFFFF',
-              lineHeight: 1.1,
-              display: 'flex',
-            }}
-          >
-            {title}
-          </div>
-        </div>
-
-        <div
-          style={{
-            display: 'flex',
-            justifyContent: 'space-between',
-            alignItems: 'center',
-          }}
-        >
-          <span style={{ fontSize: 26, color: '#9BA0A6', display: 'flex' }}>
-            AI inference benchmark · latency, throughput, cost
-          </span>
-          {logoSrc && <img src={logoSrc} height={72} />}
-        </div>
-      </div>
-    </div>,
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · Head-to-head Chip benchmark`,
+    title,
+    titleSize,
+    footer: 'AI inference benchmark · latency, throughput, cost',
+    language: 'en',
     size,
-  );
+  });
 }
diff --git a/packages/app/src/app/compare/[slug]/page-client.tsx b/packages/app/src/app/compare/[slug]/page-client.tsx
index 81e2010e6..6e963de1a 100644
--- a/packages/app/src/app/compare/[slug]/page-client.tsx
+++ b/packages/app/src/app/compare/[slug]/page-client.tsx
@@ -3,21 +3,26 @@
 import Link from 'next/link';
 import { useEffect, useMemo } from 'react';
 
-import type { GPUDataPoint, InterpolatedResult } from '@/components/calculator/types';
+import type { GPUDataPoint } from '@/components/calculator/types';
 import { useThroughputData } from '@/components/calculator/useThroughputData';
-import { CompareInterpolatedTable } from '@/components/compare/compare-interpolated-table';
-import { useGlobalFilters, GlobalFilterProvider } from '@/components/GlobalFilterContext';
+import {
+  CompareTableRenderer,
+  type CompareTableData,
+} from '@/components/compare/compare-table-renderer';
+import {
+  GlobalFilterProvider,
+  useGlobalFilterAvailability,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+} from '@/components/GlobalFilterContext';
 import { InferenceProvider } from '@/components/inference/InferenceContext';
 import InferenceChartDisplay from '@/components/inference/ui/ChartDisplay';
 import { Card } from '@/components/ui/card';
 import { track } from '@/lib/analytics';
+import type { BenchmarkRow } from '@/lib/api';
+import { toCalculatorBenchmarkRows } from '@/lib/benchmark-api-view';
 import { toModel, toPrecisions, toSequence } from '@/lib/compare-enum-coerce';
-
-interface SsrTableData {
-  defaultTargets: number[];
-  ssrRows: { target: number; a: InterpolatedResult | null; b: InterpolatedResult | null }[];
-  interactivityRange: { min: number; max: number };
-}
+import type { AgenticScenarioIntro } from '@/lib/compare-ssr';
 
 const STRINGS = {
   en: {
@@ -30,13 +35,13 @@ const STRINGS = {
       'No interpolated comparison data available for the default model. Use the chart controls below to select a model with benchmark data for both chips.',
   },
   zh: {
-    eyebrowSuffix: 'Chip 对比',
+    eyebrowSuffix: '芯片对比',
     mainChartLinkText: '主推理图表',
     perDollarLinkText: '查看每美元性能对比 →',
     caveatSeqFallback: '序列',
     caveatPrecFallback: '精度',
     emptyState:
-      '当前默认模型没有可用的插值对比数据。请使用下方图表控件选择一个两款 Chip 均有基准测试数据的模型。',
+      '当前默认模型没有可用的插值对比数据。请使用下方图表控件选择一个两款芯片均有基准测试数据的模型。',
   },
 } as const;
 
@@ -54,13 +59,18 @@ interface ComparePageClientProps {
   defaultModel: string;
   defaultSequence: string | null;
   defaultPrecision: string | null;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
+  initialBenchmarkRows?: BenchmarkRow[];
   /** One SSR-rendered prose paragraph per interpolated-table row (default
    *  interactivity target). Each paragraph picks a template variant
    *  deterministically from the slug so prose stays stable across renders
    *  but varies across pages in the catalog. Empty array when there's no
    *  comparable data. */
   narrative: string[];
+  /** Set only when the page is rendering the agentic workload. Explains what
+   *  AgentX measures before the head-to-head numbers, since a reader who
+   *  arrived from a GPU query has no reason to expect a replayed trace. */
+  agenticIntro?: AgenticScenarioIntro | null;
   aLabel: string;
   bLabel: string;
   aVendor: string;
@@ -80,7 +90,9 @@ export default function ComparePageClient({
   defaultSequence,
   defaultPrecision,
   ssrTableData,
+  initialBenchmarkRows,
   narrative,
+  agenticIntro = null,
   aLabel,
   bLabel,
   aVendor,
@@ -96,6 +108,14 @@ export default function ComparePageClient({
   const compareGpuPair = useMemo(() => [a, b] as const, [a, b]);
   const initialModel = toModel(defaultModel);
   const initialSequence = toSequence(defaultSequence);
+  const benchmarkQueryScope = `compare-pair:${a}:${b}`;
+  const initialCalculatorRows = useMemo(
+    () =>
+      initialBenchmarkRows && defaultSequence
+        ? toCalculatorBenchmarkRows(initialBenchmarkRows, defaultSequence)
+        : undefined,
+    [defaultSequence, initialBenchmarkRows],
+  );
   const initialPrecisions = toPrecisions(defaultPrecision);
   const t = STRINGS[locale];
   const isZh = locale === 'zh';
@@ -110,6 +130,9 @@ export default function ComparePageClient({
         activeTab="compare"
         initialActiveHwTypes={[a, b]}
         compareGpuPair={compareGpuPair}
+        benchmarkQueryScope={benchmarkQueryScope}
+        initialBenchmarkModel={initialModel}
+        initialBenchmarkRows={initialBenchmarkRows}
       >
         <div className="flex flex-col gap-4">
           <Card className="flex flex-col gap-3">
@@ -142,6 +165,21 @@ export default function ComparePageClient({
                   .
                 </p>
               )}
+              {agenticIntro && (
+                <p
+                  className="mt-3 max-w-3xl text-sm text-foreground/80"
+                  data-testid="compare-agentic-intro"
+                >
+                  {agenticIntro.paragraph}{' '}
+                  <Link
+                    href={agenticIntro.href}
+                    data-testid="compare-agentic-intro-link"
+                    className="font-medium text-brand underline underline-offset-4 hover:no-underline"
+                  >
+                    {agenticIntro.linkLabel} →
+                  </Link>
+                </p>
+              )}
               {narrative.length > 0 && (
                 <div className="mt-3 flex flex-col gap-2 max-w-3xl" data-testid="compare-narrative">
                   {narrative.map((para, i) => (
@@ -177,6 +215,9 @@ export default function ComparePageClient({
               aLabel={aLabel}
               bLabel={bLabel}
               ssrTableData={ssrTableData}
+              initialModel={initialModel}
+              initialSequence={initialSequence}
+              initialCalculatorRows={initialCalculatorRows}
               emptyStateText={t.emptyState}
             />
           </Card>
@@ -193,23 +234,46 @@ function CompareTableSection({
   aLabel,
   bLabel,
   ssrTableData,
+  initialModel,
+  initialSequence,
+  initialCalculatorRows,
   emptyStateText,
 }: {
   a: string;
   b: string;
   aLabel: string;
   bLabel: string;
-  ssrTableData: SsrTableData;
+  ssrTableData: CompareTableData;
+  initialModel: ReturnType<typeof toModel>;
+  initialSequence: ReturnType<typeof toSequence>;
+  initialCalculatorRows?: BenchmarkRow[];
   emptyStateText: string;
 }) {
-  const { effectiveSequence, effectivePrecisions, selectedRunDate, selectedModel } =
-    useGlobalFilters();
+  const { effectiveSequence, effectivePrecisions, selectedModel, sequenceResolved } =
+    useGlobalFilterSelection();
+  const { availableDates } = useGlobalFilterAvailability();
+  const { selectedRunDate } = useGlobalFilterRun();
+  const latestAvailableDate = availableDates.at(-1) ?? '';
+  const calculatorRunDate =
+    selectedRunDate && selectedRunDate === latestAvailableDate ? '' : selectedRunDate;
+  const initialRows =
+    initialCalculatorRows &&
+    sequenceResolved &&
+    selectedModel === initialModel &&
+    effectiveSequence === initialSequence &&
+    calculatorRunDate === ''
+      ? initialCalculatorRows
+      : undefined;
 
   const { gpuDataByGroupKey, ranges, hasData } = useThroughputData(
     selectedModel,
     effectiveSequence,
     effectivePrecisions,
-    selectedRunDate,
+    calculatorRunDate,
+    undefined,
+    undefined,
+    initialRows,
+    sequenceResolved,
   );
 
   // Extract GPUDataPoint arrays for just the two GPUs in the pair.
@@ -229,23 +293,15 @@ function CompareTableSection({
 
   const clientRange = hasData ? ranges.interactivity : ssrTableData.interactivityRange;
 
-  if (ssrTableData.defaultTargets.length === 0) {
-    return (
-      <div className="border border-border/50 rounded-md px-4 py-3 text-sm text-muted-foreground bg-muted/30">
-        {emptyStateText}
-      </div>
-    );
-  }
-
   return (
-    <CompareInterpolatedTable
+    <CompareTableRenderer
       aLabel={aLabel}
       bLabel={bLabel}
-      ssrRows={ssrTableData.ssrRows}
-      defaultTargets={ssrTableData.defaultTargets}
+      ssrTableData={ssrTableData}
       interactivityRange={clientRange}
       gpuDataPointsA={pointsA}
       gpuDataPointsB={pointsB}
+      emptyStateText={emptyStateText}
     />
   );
 }
diff --git a/packages/app/src/app/compare/[slug]/page.tsx b/packages/app/src/app/compare/[slug]/page.tsx
index b11011d5d..aa70cc676 100644
--- a/packages/app/src/app/compare/[slug]/page.tsx
+++ b/packages/app/src/app/compare/[slug]/page.tsx
@@ -4,8 +4,13 @@ import { notFound, permanentRedirect } from 'next/navigation';
 import { HW_REGISTRY, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
+import { comparisonScenarioForModel } from '@/lib/compare-agentx';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
 import { languageAlternates } from '@/lib/i18n';
-import { pickPairDefaults } from '@/lib/compare-pair-defaults';
 import {
   canonicalCompareSlug,
   compareDisplayLabel,
@@ -15,38 +20,67 @@ import {
   parseCompareSlug,
 } from '@/lib/compare-slug';
 import {
+  AGENTIC_SCENARIO_INTRO,
   buildBreadcrumbJsonLd,
   buildJsonLd,
   compareMetaDescription,
   compareTableNarrative,
-  computeCompareTableData,
-  dateRangeForPair,
-  getCachedBenchmarks,
   KNOWN_MODELS,
   KNOWN_PRECISIONS,
   KNOWN_SEQUENCES,
   pickString,
-  summarize,
 } from '@/lib/compare-ssr';
+import {
+  getComparePageDerivedData,
+  initialCompareBenchmarkRows,
+} from '@/lib/compare-page-data.server';
 
 import ComparePageClient from './page-client';
 
 export const dynamic = 'force-dynamic';
 
+function comparePath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment ? `/compare/${canonical}/${scenarioSegment}` : `/compare/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
 }
 
+/**
+ * `/compare/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than handled in
+ * a parallel route file so the body — redirects, JSON-LD, metadata, the
+ * client props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildCompareMetadata(slug, {});
+}
+
+export async function buildCompareMetadata(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Promise<Metadata> {
   const parsed = parseCompareSlug(slug);
   if (!parsed) return {};
   const fullLabel = compareModelDisplayLabel(parsed.model, parsed.a, parsed.b);
   const gpuLabel = compareDisplayLabel(parsed.a, parsed.b);
   const modelSeoName = compareModelSeoName(parsed.model);
   const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
-  const url = `${SITE_URL}/compare/${canonical}`;
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = comparePath(canonical, scenarioSegment);
+  const canonicalPath = comparePath(canonical);
+  const url = `${SITE_URL}${routePath}`;
+  const canonicalUrl = `${SITE_URL}${canonicalPath}`;
 
   // Lead the SEO title with the GPU pair — that's the phrase people search
   // ("b200 vs b300") and must survive Google's ~60-char SERP truncation. The
@@ -54,20 +88,25 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   // template so the query isn't pushed off the end.
   const title = `${compareSeoTitle(gpuLabel, modelSeoName)} | ${SITE_NAME}`;
 
-  // Stat-led meta description built from the interpolated head-to-head numbers
-  // at the slug's default operating point (falls back to boilerplate for
-  // sparse-data pairs). Fetch is blob-cached and shared with the page render.
-  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
-  const { sequence, precision } = pickPairDefaults(rows, parsed.a, parsed.b);
-  const { ssrRows } = computeCompareTableData(rows, parsed.a, parsed.b, sequence, precision);
+  // Stat-led meta description from the default interpolated head-to-head
+  // numbers. The compact derived payload is shared with the page render and
+  // invalidated with the database cache tag.
+  const { ssrRows } = await getComparePageDerivedData(
+    parsed.model.dbKeys,
+    parsed.a,
+    parsed.b,
+    null,
+    null,
+    comparisonScenarioForModel(parsed.model).sequence,
+  );
   const description = compareMetaDescription(parsed.model, parsed.a, parsed.b, ssrRows);
 
   return {
     title: { absolute: title },
     description,
     alternates: {
-      canonical: url,
-      languages: languageAlternates(`/compare/${canonical}`),
+      canonical: canonicalUrl,
+      languages: languageAlternates(canonicalPath),
     },
     openGraph: {
       title: `${fullLabel} | ${SITE_NAME}`,
@@ -85,13 +124,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function ComparePage({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderComparePage(slug, await searchParams, {});
+}
+
+export async function renderComparePage(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parseCompareSlug(slug);
   if (!parsed) notFound();
 
-  // Await searchParams once so we can both preserve them on redirect and read
-  // them for URL-param overrides further down.
-  const sp = await searchParams;
-
   // One-hop redirect to the fully canonical URL. Handles all three normalization
   // cases in a single 308:
   //   - legacy bare slug:   `h100-vs-h200`              → `deepseek-r1-h100-vs-h200`
@@ -118,42 +161,46 @@ export default async function ComparePage({ params, searchParams }: Props) {
     // all permanent decisions — using a permanent redirect lets search engines
     // consolidate link equity onto the canonical URL instead of keeping the
     // alias URL in the index alongside the canonical one.
-    permanentRedirect(`/compare/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // `/compare/h100-vs-h200/agentic` to the pair's default workload.
+    permanentRedirect(`${comparePath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
-  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
-  const summaryA = summarize(rows, parsed.a);
-  const summaryB = summarize(rows, parsed.b);
-  const { sequence: pickedSequence, precision: pickedPrecision } = pickPairDefaults(
-    rows,
-    parsed.a,
-    parsed.b,
-  );
-
+  const fallbackSequence = comparisonScenarioForModel(parsed.model).sequence;
   // URL params win over slug-derived defaults; this baking-into-SSR avoids the
   // hydration flash where the client upgrades seeded defaults to URL values.
   // `sp` was already awaited above for the redirect-query-preservation path.
   const urlSeq = pickString(sp.i_seq);
   const urlPrec = pickString(sp.i_prec);
   const urlModel = pickString(sp.g_model);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence;
-  const effectivePrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : pickedPrecision;
-  // `?g_model=` is honored only if it matches a known model — but the slug's
-  // model is the canonical default. Disregard URL param if user wants to
-  // explicitly override (rare).
   const effectiveModel =
     urlModel && KNOWN_MODELS.has(urlModel) ? urlModel : parsed.model.displayName;
-
-  const { defaultTargets, ssrRows, interactivityRange } = computeCompareTableData(
-    rows,
+  // Path beats query beats the pair's default: a scenario segment is an
+  // explicit address for one workload, so it outranks a stale `?i_seq=`.
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const requestedSequence = pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : null);
+  const requestedPrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : null;
+  const {
+    sequence: effectiveSequence,
+    precision: effectivePrecision,
+    summaryA,
+    summaryB,
+    defaultTargets,
+    ssrRows,
+    interactivityRange,
+    oldest,
+    newest,
+    initialPairBenchmarkRows,
+  } = await getComparePageDerivedData(
+    parsed.model.dbKeys,
     parsed.a,
     parsed.b,
-    effectiveSequence,
-    effectivePrecision,
+    requestedSequence,
+    requestedPrecision,
+    fallbackSequence,
   );
 
-  const url = `${SITE_URL}/compare/${canonical}`;
-  const { oldest, newest } = dateRangeForPair(rows, parsed.a, parsed.b);
+  const url = `${SITE_URL}${comparePath(canonical, scenarioSegment)}`;
   const jsonLd = buildJsonLd(
     'full',
     parsed.model,
@@ -200,8 +247,14 @@ export default async function ComparePage({ params, searchParams }: Props) {
         defaultModel={effectiveModel}
         defaultSequence={effectiveSequence}
         defaultPrecision={effectivePrecision}
+        initialBenchmarkRows={initialCompareBenchmarkRows(
+          parsed.model.displayName,
+          effectiveModel,
+          initialPairBenchmarkRows,
+        )}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO : null}
         aLabel={aLabel}
         bLabel={bLabel}
         aVendor={aMeta?.vendor ?? ''}
diff --git a/packages/app/src/app/compare/layout.tsx b/packages/app/src/app/compare/layout.tsx
index e3f01ddc4..76217427d 100644
--- a/packages/app/src/app/compare/layout.tsx
+++ b/packages/app/src/app/compare/layout.tsx
@@ -1,20 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
-/**
- * Wraps `/compare/*` pages in UnofficialRunProvider but skips DashboardShell so
- * we get a focused single-purpose page (no TabNav). The GlobalFilterProvider is
- * mounted inside the page's client component instead — that lets the page seed
- * its initial precision/sequence based on which combos actually have data for
- * the GPU pair from the slug.
- */
 export default function CompareLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/compare/page.tsx b/packages/app/src/app/compare/page.tsx
index 2f1a0cb0c..99e2f560c 100644
--- a/packages/app/src/app/compare/page.tsx
+++ b/packages/app/src/app/compare/page.tsx
@@ -1,39 +1,39 @@
 import type { Metadata } from 'next';
 import Link from 'next/link';
 
-import {
-  HW_REGISTRY,
-  SITE_NAME,
-  SITE_URL,
-  SUPPORTERS_LINE,
-} from '@semianalysisai/inferencex-constants';
+import { HW_REGISTRY, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { enAlternates } from '@/lib/i18n';
 
+import { AgentXCompareHero } from '@/components/compare/agentx-compare-hero';
 import { ComparePairCardLink } from '@/components/compare/compare-pair-card-link';
 import { JsonLd } from '@/components/json-ld';
 import { Card } from '@/components/ui/card';
+import { ModelLogo } from '@/components/ui/model-logo';
+import { comparisonPairHref, comparisonScenarioForModel } from '@/lib/compare-agentx';
 import { getComparablePairsByModelSlug } from '@/lib/compare-availability';
 import { type ComparePair, COMPARE_MODEL_SLUGS, type CompareModelSlug } from '@/lib/compare-slug';
 import { bucketComparePairsByVendor, formatModelList } from '@/lib/compare-ssr';
+import { type Model } from '@/lib/data-mappings';
 
 export const dynamic = 'force-dynamic';
 
-const DESCRIPTION = `InferenceX is the independent, open-source chip inference benchmark from SemiAnalysis, with verified, reproducible nightly results. ${SUPPORTERS_LINE} Compare latency, throughput & cost head-to-head across DeepSeek V4 Pro, DeepSeek R1, Kimi K2, MiniMax M3, GLM 5, Qwen 3.5 & more.`;
+const DESCRIPTION =
+  'Compare AgentX agentic inference results for Kimi K3, DeepSeek V4 Pro, MiniMax M3, Qwen 3.5, and GLM 5.3, plus fixed-sequence chip comparisons.';
 
 export const metadata: Metadata = {
-  title: 'Chip Comparisons',
+  title: 'AgentX Inference Comparisons',
   description: DESCRIPTION,
   alternates: enAlternates('/compare'),
   openGraph: {
-    title: `Chip Comparisons | ${SITE_NAME}`,
+    title: `AgentX Inference Comparisons | ${SITE_NAME}`,
     description: DESCRIPTION,
     url: `${SITE_URL}/compare`,
     type: 'website',
   },
   twitter: {
     card: 'summary_large_image',
-    title: `Chip Comparisons | ${SITE_NAME}`,
+    title: `AgentX Inference Comparisons | ${SITE_NAME}`,
     description: DESCRIPTION,
   },
 };
@@ -77,7 +77,7 @@ function groupPairsByVendorForModel(
 const jsonLd = {
   '@context': 'https://schema.org',
   '@type': 'CollectionPage',
-  name: `Chip Comparisons | ${SITE_NAME}`,
+  name: `AgentX Inference Comparisons | ${SITE_NAME}`,
   description: DESCRIPTION,
   url: `${SITE_URL}/compare`,
 };
@@ -97,19 +97,27 @@ export default async function CompareIndexPage() {
   return (
     <>
       <JsonLd data={jsonLd} />
-      <section>
+      <AgentXCompareHero locale="en" />
+
+      <section id="model-comparisons" data-testid="compare-model-catalog">
         <Card>
-          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">Chip Comparisons</h1>
+          <p className="font-mono text-xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
+            Comparison catalog
+          </p>
+          <h2 className="mt-2 text-2xl font-bold tracking-tight lg:text-3xl">
+            AgentX and 8K→1K results
+          </h2>
           <p className="mt-3 text-base lg:text-lg text-muted-foreground max-w-3xl">
             {totalUrls.toLocaleString()} head-to-head inference benchmark comparisons across{' '}
-            {formatModelList(modelsWithPairs)}. Each page includes interactive charts for latency,
-            throughput, and cost metrics, plus an interpolated comparison table.
+            {formatModelList(modelsWithPairs)}. Models with AgentX data open long-context,
+            multi-turn trace replay results. Models not yet covered by AgentX open the controlled
+            8K→1K workload. Each card identifies its scenario.
           </p>
           <div className="mt-6 flex flex-wrap gap-3">
             <Link
               data-testid="compare-index-per-dollar-link"
               href="/compare-per-dollar"
-              className="inline-flex items-center gap-2 rounded-md bg-brand px-5 py-3 text-base lg:text-lg font-semibold text-primary-foreground shadow-sm transition-colors hover:bg-brand/90"
+              className="inline-flex items-center gap-2 rounded-md border border-border px-5 py-3 text-base lg:text-lg font-semibold text-foreground shadow-sm transition-colors hover:bg-muted"
             >
               Compare Chip performance per dollar
               <span aria-hidden="true" className="text-lg lg:text-xl">
@@ -143,11 +151,20 @@ export default async function CompareIndexPage() {
       {modelsWithPairs.map((model) => {
         const pairs = comparablePairsByModel.get(model.slug) ?? [];
         const groups = groupPairsByVendorForModel(model, pairs);
+        const scenario = comparisonScenarioForModel(model);
         return (
           <section key={model.slug} id={model.slug}>
             <Card className="flex flex-col gap-4">
               <div>
-                <h2 className="text-xl lg:text-2xl font-bold tracking-tight">{model.label}</h2>
+                {/* `displayName` is the Model enum value by contract (see
+                    CompareModelSlug in compare-slug.ts), so the shared
+                    ModelLogo resolves each section's brand mark (DeepSeek,
+                    MiniMax, Kimi, ...) from MODEL_CONFIG and renders nothing
+                    for models without one. */}
+                <h2 className="flex items-center gap-2.5 text-xl lg:text-2xl font-bold tracking-tight">
+                  <ModelLogo model={model.displayName as Model} className="size-6 lg:size-7" />
+                  {model.label}
+                </h2>
                 <p className="text-sm text-muted-foreground mt-1">
                   {pairs.length} chip pair{pairs.length === 1 ? '' : 's'} with benchmark data on{' '}
                   {model.label}.
@@ -167,10 +184,19 @@ export default async function CompareIndexPage() {
                       return (
                         <ComparePairCardLink
                           key={slug}
-                          href={`/compare/${slug}`}
+                          href={comparisonPairHref('en', slug, model)}
                           slug={slug}
                           label={label}
                           archLine={archLine}
+                          scenarioLabel={scenario.label}
+                          hardwareA={{
+                            label: aMeta?.label ?? a.toUpperCase(),
+                            vendor: aMeta?.vendor,
+                          }}
+                          hardwareB={{
+                            label: bMeta?.label ?? b.toUpperCase(),
+                            vendor: bMeta?.vendor,
+                          }}
                         />
                       );
                     })}
diff --git a/packages/app/src/app/datasets/page.tsx b/packages/app/src/app/datasets/page.tsx
deleted file mode 100644
index 6bd33e439..000000000
--- a/packages/app/src/app/datasets/page.tsx
+++ /dev/null
@@ -1,100 +0,0 @@
-import type { Metadata } from 'next';
-
-import { Card } from '@/components/ui/card';
-import { JsonLd } from '@/components/json-ld';
-import { DatasetList } from '@/components/datasets/dataset-list';
-import { enAlternates } from '@/lib/i18n';
-import { SITE_URL } from '@semianalysisai/inferencex-constants';
-
-const DESCRIPTION =
-  'The real Claude Code agentic conversation traces that the InferenceX agentic benchmark replays — methodology, distributions, and per-conversation flamegraphs.';
-
-export const metadata: Metadata = {
-  title: 'Agentic Datasets',
-  description: DESCRIPTION,
-  alternates: enAlternates('/datasets'),
-  openGraph: {
-    title: 'Agentic Datasets | InferenceX',
-    description: DESCRIPTION,
-    url: `${SITE_URL}/datasets`,
-  },
-  twitter: { title: 'Agentic Datasets | InferenceX', description: DESCRIPTION },
-};
-
-const jsonLd = {
-  '@context': 'https://schema.org',
-  '@type': 'CollectionPage',
-  name: 'InferenceX Agentic Datasets',
-  description: DESCRIPTION,
-  url: `${SITE_URL}/datasets`,
-};
-
-export default function DatasetsPage() {
-  return (
-    <main className="relative">
-      <JsonLd data={jsonLd} />
-      <div className="container mx-auto flex flex-col gap-6 px-4 pb-8 lg:px-8">
-        <section>
-          <Card>
-            <h1 className="mb-2 text-xl font-semibold text-foreground">
-              Agentic Benchmark Datasets
-            </h1>
-            <p className="mb-3 text-sm text-muted-foreground">
-              InferenceX&apos;s agentic benchmark doesn&apos;t replay synthetic prompts — it replays
-              real Claude Code coding sessions captured as <strong>conversation traces</strong>.
-              Each trace is a full multi-turn session: the main agent&apos;s turns plus any
-              subagents it spawned, with per-turn input/output token counts and the 64-token
-              KV-cache block hashes needed to reconstruct prefix-cache reuse. The traces are
-              published openly on HuggingFace under <code>semianalysisai/cc-traces-weka-*</code>{' '}
-              (apache-2.0).
-            </p>
-
-            <h2 className="mb-1.5 mt-4 text-sm font-semibold text-foreground">
-              How traces are captured
-            </h2>
-            <p className="mb-3 text-sm text-muted-foreground">
-              Production Claude Code sessions are recorded through a logging proxy that captures
-              every API request: its input and output token counts, the model used, timing (TTFT,
-              inter-token latency), and a list of <code>hash_ids</code> — one per 64-token KV block
-              of the request&apos;s input. Subagent invocations are grouped under their parent turn.
-              No prompt or completion text is stored; only token counts and block hashes, so the
-              corpus is shareable while remaining a faithful workload for replay.
-            </p>
-
-            <h2 className="mb-1.5 mt-4 text-sm font-semibold text-foreground">
-              Cached prefix vs uncached suffix
-            </h2>
-            <p className="mb-3 text-sm text-muted-foreground">
-              Agentic workloads are dominated by prefix reuse: each turn resends the growing
-              conversation, so most of its input is already in the KV cache from prior turns. We
-              reconstruct this exactly. Walking a conversation in order under an idealized infinite
-              cache, a turn&apos;s <strong>cached prefix</strong> is its longest run of leading{' '}
-              <code>hash_ids</code> already seen; the rest is the <strong>uncached suffix</strong>{' '}
-              that must be (re)computed. Blocks are 64 tokens; the split is clamped so cached +
-              uncached equals the turn&apos;s effective input even on a partial final block.
-              Subagents run against a snapshot of the parent cache at spawn (their context is
-              separate and is not folded back into the parent).
-            </p>
-
-            <h2 className="mb-1.5 mt-4 text-sm font-semibold text-foreground">Dataset variants</h2>
-            <ul className="mb-1 list-disc space-y-1 pl-5 text-sm text-muted-foreground">
-              <li>
-                <strong>full</strong> — every captured request, unmodified.
-              </li>
-              <li>
-                <strong>256k</strong> — requests whose input + output exceeds 256,000 tokens are
-                dropped so every turn fits a 256k context window (used when benchmarking engines
-                configured for a 256k max context).
-              </li>
-            </ul>
-          </Card>
-        </section>
-
-        <section className="flex flex-col gap-3">
-          <h2 className="text-lg font-semibold text-foreground">Datasets</h2>
-          <DatasetList />
-        </section>
-      </div>
-    </main>
-  );
-}
diff --git a/packages/app/src/app/globals.css b/packages/app/src/app/globals.css
index 4615ae1d6..ed5c23d66 100644
--- a/packages/app/src/app/globals.css
+++ b/packages/app/src/app/globals.css
@@ -40,20 +40,20 @@
   content: none;
 }
 
-/* Pride theme for the InferenceX wordmark — celebrating July Disability Pride Month.
-   Paints the brand text with the Disability Pride flag's 5 stripe colors via a
-   clipped diagonal gradient (Ann Magill's 2021 revised, low-sensitivity palette):
-   red = physical, gold = neurodivergent, gray = invisible/undiagnosed,
-   blue = psychiatric, green = sensory. The flag's white stripe is darkened to a
-   readable slate here so the clipped wordmark stays legible on the light header. */
+/* Pride theme for the InferenceX wordmark — the transgender flag's stripe colors
+   painted onto the brand text via a clipped gradient: light blue, pink, white,
+   pink, light blue. The white stripe is darkened to a readable slate here, the
+   same treatment the earlier flag themes used, so the clipped wordmark stays
+   legible on the light header. Fixed flag colors, independent of the theme
+   tokens, so it renders the same in light, dark, and minecraft modes. */
 .pride-wordmark {
   background-image: linear-gradient(
-    135deg,
-    #cf7280 0%,
-    #d9ad2e 28%,
+    90deg,
+    #5bcefa 0%,
+    #f5a9b8 25%,
     #8a8f99 50%,
-    #4aa6cf 72%,
-    #3aa660 100%
+    #f5a9b8 75%,
+    #5bcefa 100%
   );
   background-clip: text;
   -webkit-background-clip: text;
@@ -61,6 +61,21 @@
   -webkit-text-fill-color: transparent;
 }
 
+/* Benchmark log viewer gutter. Line numbers come from a CSS counter rather than
+   DOM text so they stay out of the element's text content and out of anything
+   the reader copies from the pane. */
+.log-lines {
+  counter-reset: log-line;
+}
+
+.log-lines > .log-gutter {
+  counter-increment: log-line;
+}
+
+.log-lines > .log-gutter::before {
+  content: counter(log-line);
+}
+
 /* Shiki dual-theme: light by default, dark when .dark class is active */
 .shiki,
 .shiki span {
@@ -105,7 +120,7 @@
 
   /* Fonts */
   --font-sans: var(--font-dm-sans);
-  --font-mono: monospace;
+  --font-mono: ui-monospace, 'SFMono-Regular', 'Menlo', 'Consolas', 'Liberation Mono', monospace;
 
   /* Breakpoints */
   --breakpoint-2xl: 80rem;
@@ -122,6 +137,24 @@
   --color-sidebar: var(--sidebar);
 }
 
+/* Typography tokens — the governed alternative to arbitrary values like
+   text-[11px] or tracking-[0.16em]. New arbitrary font-size/tracking values
+   are blocked by scripts/check-typography.ts; use these instead.
+   See docs/typography.md. */
+@theme {
+  /* Micro label sizes (smaller than text-xs). Deliberately no paired
+     --text-*--line-height: the arbitrary values these replace set only
+     font-size and inherited line-height, and dense tables depend on that. */
+  --text-2xs: 0.6875rem; /* 11px */
+  --text-3xs: 0.625rem; /* 10px */
+
+  /* Named letter-spacing for the recurring uppercase-eyebrow and
+     tightened-heading recipes. */
+  --tracking-eyebrow: 0.16em;
+  --tracking-eyebrow-wide: 0.2em;
+  --tracking-heading: -0.04em;
+}
+
 :root {
   /* Setup variables */
   --radius: 0.625rem;
@@ -410,11 +443,21 @@
 
 .splash-text {
   display: inline-block;
-  color: #ffff00;
+  /* Light mode paints this over a white card, where flat #ffff00 all but
+     disappears — keep the Minecraft yellow but ring it with a dark outline. */
+  color: #ffcc00;
   font-weight: bold;
   font-size: 0.85rem;
   white-space: nowrap;
-  text-shadow: 2px 2px 0px #3f3f00;
+  text-shadow:
+    1px 1px 0 #4a3800,
+    1px -1px 0 #4a3800,
+    -1px 1px 0 #4a3800,
+    -1px -1px 0 #4a3800,
+    2px 2px 0 #4a3800;
+  /* The keyframes re-declare this; keeping it on the element means the splash
+     still sits where it belongs when the animation is off (reduced motion). */
+  transform: translate(2rem, 2.25rem) rotate(10deg);
   animation: splash-bounce 1.5s ease-in-out infinite;
   transform-origin: center;
   /* Always use Minecraft font & style regardless of active theme */
@@ -423,6 +466,19 @@
   -moz-osx-font-smoothing: unset;
 }
 
+/* Dark backgrounds need no outline — this is the original title-screen look. */
+.dark .splash-text,
+.minecraft .splash-text {
+  color: #ffff00;
+  text-shadow: 2px 2px 0px #3f3f00;
+}
+
+@media (prefers-reduced-motion: reduce) {
+  .splash-text {
+    animation: none;
+  }
+}
+
 @media (min-width: 640px) {
   .splash-text {
     font-size: 1.1rem;
diff --git a/packages/app/src/app/glossary/[slug]/page.tsx b/packages/app/src/app/glossary/[slug]/page.tsx
index 2e7b24f8b..91722f144 100644
--- a/packages/app/src/app/glossary/[slug]/page.tsx
+++ b/packages/app/src/app/glossary/[slug]/page.tsx
@@ -142,16 +142,16 @@ export default async function GlossaryTermPage({ params }: Props) {
                 AI inference glossary
               </Link>
               <div className="mt-8 flex flex-wrap items-center gap-2">
-                <span className="rounded-full border border-brand/25 bg-brand/8 px-3 py-1 text-xs font-semibold tracking-[0.14em] text-brand uppercase">
+                <span className="rounded-full border border-brand/25 bg-brand/8 px-3 py-1 text-xs font-semibold tracking-eyebrow text-brand uppercase">
                   {entry.category}
                 </span>
                 {entry.abbreviation && (
-                  <span className="font-mono text-xs tracking-[0.16em] text-muted-foreground uppercase">
+                  <span className="font-mono text-xs tracking-eyebrow text-muted-foreground uppercase">
                     {entry.abbreviation}
                   </span>
                 )}
               </div>
-              <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-[-0.035em] text-balance md:text-6xl">
+              <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-6xl">
                 {entry.term}
               </h1>
               {entry.aliases && entry.aliases.length > 0 && (
@@ -169,7 +169,7 @@ export default async function GlossaryTermPage({ params }: Props) {
                 >
                   <p
                     id="plain-english-heading"
-                    className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase"
+                    className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase"
                   >
                     In plain English
                   </p>
@@ -190,7 +190,7 @@ export default async function GlossaryTermPage({ params }: Props) {
 
                 {entry.measurement && (
                   <div className="mt-8 border-l-2 border-brand bg-brand/6 px-5 py-4">
-                    <p className="text-xs font-semibold tracking-[0.14em] text-muted-foreground uppercase">
+                    <p className="text-xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
                       {entry.measurement.label}
                     </p>
                     <p className="mt-1 font-mono text-sm leading-relaxed text-foreground">
@@ -222,7 +222,7 @@ export default async function GlossaryTermPage({ params }: Props) {
               </div>
 
               <aside className="border-t border-border/50 bg-muted/10 px-5 py-8 lg:border-t-0 lg:border-l lg:px-6 lg:py-12">
-                <p className="font-mono text-xs font-semibold tracking-[0.16em] text-muted-foreground uppercase">
+                <p className="font-mono text-xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
                   Related terms
                 </p>
                 <nav aria-label="Related glossary terms" className="mt-4 flex flex-col">
@@ -253,7 +253,7 @@ export default async function GlossaryTermPage({ params }: Props) {
               <Card>
                 <div className="flex flex-col gap-2 border-b border-border/50 pb-5 md:flex-row md:items-end md:justify-between">
                   <div>
-                    <p className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase">
+                    <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
                       Source material
                     </p>
                     <h2 id="further-reading" className="mt-2 text-2xl font-semibold tracking-tight">
@@ -298,7 +298,7 @@ export default async function GlossaryTermPage({ params }: Props) {
                 href={`/glossary/${adjacent.previous.slug}`}
                 className="rounded-xl border border-border/40 bg-background/20 p-5 backdrop-blur-[2px] transition-colors hover:border-brand/40 hover:bg-brand/5"
               >
-                <span className="text-xs tracking-[0.14em] text-muted-foreground uppercase">
+                <span className="text-xs tracking-eyebrow text-muted-foreground uppercase">
                   ← Previous
                 </span>
                 <span className="mt-2 block font-semibold">{adjacent.previous.term}</span>
@@ -311,7 +311,7 @@ export default async function GlossaryTermPage({ params }: Props) {
                 href={`/glossary/${adjacent.next.slug}`}
                 className="rounded-xl border border-border/40 bg-background/20 p-5 text-right backdrop-blur-[2px] transition-colors hover:border-brand/40 hover:bg-brand/5"
               >
-                <span className="text-xs tracking-[0.14em] text-muted-foreground uppercase">
+                <span className="text-xs tracking-eyebrow text-muted-foreground uppercase">
                   Next →
                 </span>
                 <span className="mt-2 block font-semibold">{adjacent.next.term}</span>
diff --git a/packages/app/src/app/glossary/page.tsx b/packages/app/src/app/glossary/page.tsx
index c2b231bce..35d88fb60 100644
--- a/packages/app/src/app/glossary/page.tsx
+++ b/packages/app/src/app/glossary/page.tsx
@@ -11,13 +11,15 @@ import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-con
 
 const title = 'AI Inference Glossary';
 const description =
-  'Clear, technically grounded definitions for LLM inference benchmarks, serving metrics, distributed parallelism, numerical precision, chip hardware, and inference software.';
+  'Clear, technically grounded definitions for agentic inference, LLM benchmarks, serving metrics, distributed parallelism, numerical precision, chip hardware, and inference software.';
 
 export const metadata: Metadata = {
   title,
   description,
   keywords: [
     'AI inference glossary',
+    'agentic inference glossary',
+    'AgentX benchmark terms',
     'LLM inference terms',
     'GPU benchmark terminology',
     'inference serving glossary',
@@ -93,28 +95,28 @@ export default function GlossaryPage() {
             />
             <div className="grid gap-10 lg:grid-cols-[minmax(0,1fr)_18rem] lg:items-end">
               <div>
-                <p className="font-mono text-xs font-semibold tracking-[0.2em] text-brand uppercase">
+                <p className="font-mono text-xs font-semibold tracking-eyebrow-wide text-brand uppercase">
                   Field guide / AI infrastructure
                 </p>
-                <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-[-0.045em] text-balance md:text-6xl lg:text-7xl">
+                <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-6xl lg:text-7xl">
                   The language behind the inference curve.
                 </h1>
                 <p className="mt-6 max-w-3xl text-base leading-7 text-muted-foreground md:text-lg md:leading-8">
-                  Definitions for the metrics, serving techniques, numerical formats, and
-                  distributed systems concepts used across InferenceX. Based on measured behavior,
-                  not vendor peak specifications.
+                  Definitions for agentic inference, benchmark metrics, serving techniques,
+                  numerical formats, and distributed systems concepts used across InferenceX.
+                  Entries describe measured behavior and published benchmark recipes.
                 </p>
               </div>
 
               <dl className="grid grid-cols-3 gap-px overflow-hidden rounded-lg border border-border/50 bg-border/50 lg:grid-cols-1">
                 <div className="bg-background/70 p-4">
-                  <dt className="font-mono text-[0.65rem] tracking-[0.16em] text-muted-foreground uppercase">
+                  <dt className="font-mono text-3xs tracking-eyebrow text-muted-foreground uppercase">
                     Terms
                   </dt>
                   <dd className="mt-1 text-2xl font-semibold tabular-nums">{entries.length}</dd>
                 </div>
                 <div className="bg-background/70 p-4">
-                  <dt className="font-mono text-[0.65rem] tracking-[0.16em] text-muted-foreground uppercase">
+                  <dt className="font-mono text-3xs tracking-eyebrow text-muted-foreground uppercase">
                     Categories
                   </dt>
                   <dd className="mt-1 text-2xl font-semibold tabular-nums">
@@ -122,7 +124,7 @@ export default function GlossaryPage() {
                   </dd>
                 </div>
                 <div className="bg-background/70 p-4">
-                  <dt className="font-mono text-[0.65rem] tracking-[0.16em] text-muted-foreground uppercase">
+                  <dt className="font-mono text-3xs tracking-eyebrow text-muted-foreground uppercase">
                     Articles reviewed
                   </dt>
                   <dd className="mt-1 text-2xl font-semibold tabular-nums">{articleCount}</dd>
@@ -136,7 +138,7 @@ export default function GlossaryPage() {
 
         <section className="mt-4 grid gap-4 md:grid-cols-2">
           <Card>
-            <p className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase">
+            <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
               Reading the benchmark
             </p>
             <h2 className="mt-3 text-2xl font-semibold tracking-tight">
@@ -161,7 +163,7 @@ export default function GlossaryPage() {
           </Card>
 
           <Card>
-            <p className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase">
+            <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
               Grounded in measurements
             </p>
             <h2 className="mt-3 text-2xl font-semibold tracking-tight">
@@ -173,6 +175,9 @@ export default function GlossaryPage() {
               speedups on unchanged chips.
             </p>
             <div className="mt-5 flex flex-wrap gap-3 text-sm font-medium">
+              <Link href="/agentx/methodology" className="text-brand hover:underline">
+                AgentX methodology →
+              </Link>
               <Link href="/blog" className="text-brand hover:underline">
                 Browse technical articles →
               </Link>
diff --git a/packages/app/src/app/land-acknowledgement/page.tsx b/packages/app/src/app/land-acknowledgement/page.tsx
index bb85154c4..0fdbd2028 100644
--- a/packages/app/src/app/land-acknowledgement/page.tsx
+++ b/packages/app/src/app/land-acknowledgement/page.tsx
@@ -53,7 +53,7 @@ export default function LandAcknowledgementPage() {
             <p className="mb-3 text-xs font-semibold uppercase tracking-[0.32em] text-brand">
               Land Acknowledgement
             </p>
-            <h1 className="text-4xl font-semibold tracking-[-0.04em] text-foreground md:text-5xl">
+            <h1 className="text-4xl font-semibold tracking-heading text-foreground md:text-5xl">
               We recognize the Indigenous homelands connected to our US infrastructure.
             </h1>
             <p className="mt-4 text-sm leading-6 text-muted-foreground md:text-base">
@@ -79,7 +79,7 @@ export default function LandAcknowledgementPage() {
                 <p className="text-xs font-semibold uppercase tracking-[0.28em] text-muted-foreground">
                   {entry.region}
                 </p>
-                <h2 className="mt-3 text-xl font-semibold tracking-[-0.04em] text-foreground">
+                <h2 className="mt-3 text-xl font-semibold tracking-heading text-foreground">
                   {entry.peoples}
                 </h2>
                 <p className="mt-4 text-sm leading-6 text-muted-foreground">
diff --git a/packages/app/src/app/layout.tsx b/packages/app/src/app/layout.tsx
index 0f8a9e5cd..841db6394 100644
--- a/packages/app/src/app/layout.tsx
+++ b/packages/app/src/app/layout.tsx
@@ -54,6 +54,11 @@ export const metadata: Metadata = {
   },
   description: DESCRIPTION,
   keywords: [
+    'AgentX scenario',
+    'agentic inference benchmark',
+    'agentic coding benchmark',
+    'long context inference benchmark',
+    'multi-turn inference benchmark',
     'AI inference benchmark',
     'GPU benchmark',
     'LLM benchmark',
@@ -160,7 +165,7 @@ const jsonLd = {
       '@id': `${SITE_URL}/#application`,
       name: SITE_NAME,
       description:
-        'Open-source AI inference benchmark dashboard. Compare chip performance for LLM inference across NVIDIA GB200, H100, AMD MI355X, and more.',
+        'InferenceX is an open-source agentic inference benchmark dashboard. It compares the AgentX long-context, multi-turn coding scenario with fixed-sequence serving across NVIDIA, AMD, and other accelerators.',
       url: SITE_URL,
       applicationCategory: 'DeveloperApplication',
       operatingSystem: 'Web',
diff --git a/packages/app/src/app/llms-full.txt/route.ts b/packages/app/src/app/llms-full.txt/route.ts
index 9648617c2..ccf5b9f29 100644
--- a/packages/app/src/app/llms-full.txt/route.ts
+++ b/packages/app/src/app/llms-full.txt/route.ts
@@ -28,7 +28,7 @@ export async function GET() {
     `# ${SITE_NAME} Articles — Full Content`,
     `> By ${AUTHOR_NAME}`,
     '',
-    `This file contains the full text of all articles from ${SITE_NAME} (${SITE_URL}/blog).`,
+    `This file contains the full text of all articles from ${SITE_NAME} (${SITE_URL}/blog), covering agentic inference benchmarks including AgentX results, fixed-sequence AI inference, chip performance, and ML infrastructure.`,
     `It is intended for consumption by large language models and AI assistants.`,
     '',
     '---',
diff --git a/packages/app/src/app/llms.txt/route.ts b/packages/app/src/app/llms.txt/route.ts
index a53149eed..3f3dd0966 100644
--- a/packages/app/src/app/llms.txt/route.ts
+++ b/packages/app/src/app/llms.txt/route.ts
@@ -1,4 +1,8 @@
 import { getAllPosts } from '@/lib/blog';
+import { getAllChipPages, getAllChipVsPages } from '@/lib/chip-pages';
+import { getAllRankingPageEntries, rankingPageHeading } from '@/lib/rankings';
+import { inferenceModelMeta } from '@/lib/inference-model-meta';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
 import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 // oxlint-disable-next-line require-await
@@ -8,11 +12,13 @@ export async function GET() {
   const lines = [
     `# ${SITE_NAME} by ${AUTHOR_NAME}`,
     '',
-    `> ${SITE_NAME} is the open-source AI inference benchmark dashboard. We compare chip performance for LLM inference across NVIDIA GB200, H100, AMD MI355X, and more.`,
+    `> ${SITE_NAME} is an open-source agentic inference benchmark dashboard. It compares the AgentX long-context, multi-turn coding scenario with fixed-sequence serving across NVIDIA, AMD, and other accelerators using public runs.`,
     '',
     `## Links`,
     '',
     `- [Dashboard](${SITE_URL})`,
+    `- [AgentX](${SITE_URL}/agentx)`,
+    `- [AgentX Methodology](${SITE_URL}/agentx/methodology)`,
     `- [Articles](${SITE_URL}/blog)`,
     `- [API Reference](${SITE_URL}/api)`,
     `- [OpenAPI 3.1 Specification](${SITE_URL}/api/openapi.json)`,
@@ -20,6 +26,35 @@ export async function GET() {
     `- [Full content for LLMs](${SITE_URL}/llms-full.txt)`,
     `- [GitHub](https://github.com/SemiAnalysisAI/InferenceX)`,
     '',
+    `## Model Benchmark Pages`,
+    '',
+    ...INFERENCE_MODEL_SLUGS.map(
+      (entry) =>
+        `- [${inferenceModelMeta(entry).title}](${SITE_URL}/inference/${entry.slug}): ${entry.label}`,
+    ),
+    '',
+    `## Chip Pages`,
+    '',
+    ...getAllChipPages().map(
+      (chip) => `- [${chip.title}](${SITE_URL}/chips/${chip.slug}): specs, pricing and benchmarks`,
+    ),
+    ...getAllChipVsPages().map(
+      (page) =>
+        `- [${page.a.label} vs ${page.b.label}](${SITE_URL}/chips/${page.slug}): head-to-head comparison`,
+    ),
+    '',
+    `## GPU Rankings`,
+    '',
+    `- [GPU Rankings for LLM Inference](${SITE_URL}/rankings): live fastest and cheapest GPU leaderboards per model`,
+    ...getAllRankingPageEntries().map(
+      (entry) =>
+        `- [${rankingPageHeading(entry)}](${SITE_URL}/rankings/${entry.slug}): live benchmark ranking`,
+    ),
+    '',
+    `## Model on GPU Results`,
+    '',
+    `- [Run Any Model on Any GPU](${SITE_URL}/run): measured throughput and cost for every benchmarked model and GPU pairing`,
+    '',
     `## Articles`,
     '',
     ...posts.map((post) => `- [${post.title}](${SITE_URL}/blog/${post.slug}): ${post.subtitle}`),
diff --git a/packages/app/src/app/manifest.ts b/packages/app/src/app/manifest.ts
index 26b9d5e46..fd80ad822 100644
--- a/packages/app/src/app/manifest.ts
+++ b/packages/app/src/app/manifest.ts
@@ -7,7 +7,7 @@ export default function manifest(): MetadataRoute.Manifest {
     name: `${SITE_NAME} by ${AUTHOR_NAME}`,
     short_name: SITE_NAME,
     description:
-      'Open-source AI inference benchmark. Compare chip performance across NVIDIA, AMD, and more.',
+      'Open-source agentic inference benchmark. Compare the AgentX scenario and fixed-sequence serving performance across NVIDIA, AMD, and more.',
     start_url: '/',
     display: 'standalone',
     background_color: '#09090b',
diff --git a/packages/app/src/app/model/[slug]/opengraph-image.tsx b/packages/app/src/app/model/[slug]/opengraph-image.tsx
new file mode 100644
index 000000000..cc2840d0d
--- /dev/null
+++ b/packages/app/src/app/model/[slug]/opengraph-image.tsx
@@ -0,0 +1,48 @@
+import { ImageResponse } from 'next/og';
+
+import { renderOgImage, size } from '@/app/blog/[slug]/og-image-render';
+import { getModelPage, getModelPageSlugs } from '@/lib/model-pages';
+
+export const alt = 'InferenceX Model Architecture Deep-Dive';
+export { size };
+export const contentType = 'image/png';
+
+export function generateStaticParams() {
+  return getModelPageSlugs().map((slug) => ({ slug }));
+}
+
+export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
+  const { slug } = await params;
+  const page = getModelPage(slug);
+
+  if (!page) {
+    return new ImageResponse(
+      <div
+        style={{
+          display: 'flex',
+          alignItems: 'center',
+          justifyContent: 'center',
+          width: '100%',
+          height: '100%',
+          backgroundColor: '#18181b',
+          color: '#fafafa',
+          fontSize: 48,
+        }}
+      >
+        InferenceX
+      </div>,
+      size,
+    );
+  }
+
+  return renderOgImage(
+    {
+      title: `${page.meta.title} — Architecture & Evals`,
+      subtitle: page.meta.description,
+      date: '',
+    },
+    // releaseDate is free-form text (may hold multiple versions), so it is
+    // passed through verbatim instead of being date-formatted.
+    { dateLabel: `${page.meta.developer} · ${page.meta.releaseDate}` },
+  );
+}
diff --git a/packages/app/src/app/model/[slug]/page.tsx b/packages/app/src/app/model/[slug]/page.tsx
new file mode 100644
index 000000000..bd673833c
--- /dev/null
+++ b/packages/app/src/app/model/[slug]/page.tsx
@@ -0,0 +1,155 @@
+import type { Metadata } from 'next';
+import Link from 'next/link';
+import { notFound, redirect } from 'next/navigation';
+
+import 'katex/dist/katex.min.css';
+
+import { JsonLd } from '@/components/json-ld';
+import EmbeddedModelDashboard from '@/components/model/EmbeddedModelDashboard';
+import ModelArchitectureInline from '@/components/model/ModelArchitectureInline';
+import ModelDeveloperLogo from '@/components/model/ModelDeveloperLogo';
+import { Card } from '@/components/ui/card';
+import { Heading } from '@/components/ui/heading';
+import { compileBlogMdx } from '@/lib/blog-mdx';
+import { comparisonScenarioForModel } from '@/lib/compare-agentx';
+import { COMPARE_MODEL_ALIASES } from '@/lib/compare-slug';
+import { getModelPage, getModelPageSlugs } from '@/lib/model-pages';
+import { AUTHOR_HANDLE, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+interface Props {
+  params: Promise<{ slug: string }>;
+}
+
+export function generateStaticParams() {
+  return getModelPageSlugs().map((slug) => ({ slug }));
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug } = await params;
+  const page = getModelPage(slug);
+  if (!page) return {};
+  const title = `${page.meta.title} — Architecture, Evals & Inference Performance`;
+  return {
+    title: { absolute: `${title} | ${SITE_NAME}` },
+    description: page.meta.description,
+    // English-only page: canonical without hreflang alternates (no /zh sibling).
+    alternates: { canonical: `${SITE_URL}/model/${slug}` },
+    // og:image / twitter:image come from the colocated opengraph-image.tsx.
+    openGraph: {
+      title,
+      description: page.meta.description,
+      url: `${SITE_URL}/model/${slug}`,
+      siteName: SITE_NAME,
+      type: 'article',
+      locale: 'en_US',
+    },
+    twitter: {
+      card: 'summary_large_image',
+      title,
+      description: page.meta.description,
+      site: AUTHOR_HANDLE,
+      creator: AUTHOR_HANDLE,
+    },
+  };
+}
+
+export default async function ModelPage({ params }: Props) {
+  const { slug } = await params;
+  // Alias slugs (e.g. /model/glm-5) resolve to their canonical page, matching
+  // the /compare route family's alias behavior.
+  const canonical = COMPARE_MODEL_ALIASES[slug];
+  if (canonical) redirect(`/model/${canonical}`);
+
+  const page = getModelPage(slug);
+  if (!page) notFound();
+
+  const { meta, entry, raw } = page;
+  const { content } = await compileBlogMdx(raw);
+  const scenario = comparisonScenarioForModel(entry);
+  const dashboardQuery = new URLSearchParams({
+    g_model: entry.displayName,
+    i_seq: scenario.sequence,
+    i_optimal: '1',
+  });
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'WebPage',
+    name: `${meta.title} — Architecture, Evals & Inference Performance`,
+    description: meta.description,
+    url: `${SITE_URL}/model/${slug}`,
+    isPartOf: { '@type': 'WebSite', name: SITE_NAME, url: SITE_URL },
+  };
+
+  const breadcrumbJsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'BreadcrumbList',
+    itemListElement: [
+      { '@type': 'ListItem', position: 1, name: SITE_NAME, item: SITE_URL },
+      { '@type': 'ListItem', position: 2, name: 'Model Architectures', item: `${SITE_URL}/model` },
+      { '@type': 'ListItem', position: 3, name: meta.title, item: `${SITE_URL}/model/${slug}` },
+    ],
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <JsonLd data={breadcrumbJsonLd} />
+      <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-4">
+        <Card>
+          <header>
+            <p className="text-sm text-muted-foreground mb-2">
+              <Link href="/inference" className="hover:text-foreground transition-colors">
+                Inference Dashboard
+              </Link>{' '}
+              /{' '}
+              <Link href="/model" className="hover:text-foreground transition-colors">
+                Model
+              </Link>
+            </p>
+            <div className="flex items-center gap-3">
+              <ModelDeveloperLogo developer={meta.developer} className="size-8 lg:size-10" />
+              <Heading as="h1" level="page">
+                {meta.title}
+              </Heading>
+            </div>
+            <p className="mt-3 text-base lg:text-lg text-muted-foreground">{meta.description}</p>
+            <div className="flex flex-wrap items-center gap-3 text-sm text-muted-foreground mt-3">
+              <span>{meta.developer}</span>
+              <span>&middot;</span>
+              <span>Released {meta.releaseDate}</span>
+            </div>
+          </header>
+          <div className="mt-6">
+            <ModelArchitectureInline displayName={entry.displayName} />
+          </div>
+          <div className="mt-6 pt-6 border-t border-border/40">
+            <article className="prose prose-neutral dark:prose-invert max-w-none blog-prose">
+              {content}
+            </article>
+          </div>
+        </Card>
+        <Card data-testid="model-page-dashboard">
+          <div className="flex flex-wrap items-baseline justify-between gap-2 mb-4">
+            <div>
+              <h2 className="text-lg font-semibold mb-1">
+                {meta.title} inference performance ({scenario.label})
+              </h2>
+              <p className="text-muted-foreground text-sm">
+                Live InferenceX benchmark data for {entry.label} on the {scenario.label} workload,
+                measured in total tokens per dollar across every chip config with data.
+              </p>
+            </div>
+            <Link
+              href={`/inference?${dashboardQuery}`}
+              className="text-sm text-primary hover:underline whitespace-nowrap"
+            >
+              Open in full dashboard →
+            </Link>
+          </div>
+          <EmbeddedModelDashboard displayName={entry.displayName} sequence={scenario.sequence} />
+        </Card>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/model/layout.tsx b/packages/app/src/app/model/layout.tsx
new file mode 100644
index 000000000..cedf05200
--- /dev/null
+++ b/packages/app/src/app/model/layout.tsx
@@ -0,0 +1,12 @@
+import type { ReactNode } from 'react';
+
+import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+
+/**
+ * Model deep-dive shell. `UnofficialRunProvider` is required by the embedded
+ * dashboard's filter contexts (same dependency the /compare routes satisfy in
+ * their layout).
+ */
+export default function ModelLayout({ children }: { children: ReactNode }) {
+  return <UnofficialRunProvider>{children}</UnofficialRunProvider>;
+}
diff --git a/packages/app/src/app/model/opengraph-image.tsx b/packages/app/src/app/model/opengraph-image.tsx
new file mode 100644
index 000000000..0476adb1f
--- /dev/null
+++ b/packages/app/src/app/model/opengraph-image.tsx
@@ -0,0 +1,18 @@
+import { renderOgImage, size } from '@/app/blog/[slug]/og-image-render';
+import { getModelPageSlugs } from '@/lib/model-pages';
+
+export const alt = 'InferenceX Model Architectures';
+export { size };
+export const contentType = 'image/png';
+
+export default function OgImage() {
+  return renderOgImage(
+    {
+      title: 'Model Architectures',
+      subtitle:
+        'MoE and attention design, official vendor eval scores, and live inference performance for every model benchmarked on InferenceX.',
+      date: '',
+    },
+    { dateLabel: `${getModelPageSlugs().length} models` },
+  );
+}
diff --git a/packages/app/src/app/model/page.tsx b/packages/app/src/app/model/page.tsx
new file mode 100644
index 000000000..c5e949cd4
--- /dev/null
+++ b/packages/app/src/app/model/page.tsx
@@ -0,0 +1,130 @@
+import type { Metadata } from 'next';
+import Link from 'next/link';
+
+import { JsonLd } from '@/components/json-ld';
+import ModelDeveloperLogo from '@/components/model/ModelDeveloperLogo';
+import { Badge } from '@/components/ui/badge';
+import { Card } from '@/components/ui/card';
+import { toModel } from '@/lib/compare-enum-coerce';
+import { formatParamCount, getModelArchitecture } from '@/lib/model-architectures';
+import { getModelPage, getModelPageSlugs } from '@/lib/model-pages';
+import { AUTHOR_HANDLE, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const PAGE_TITLE = 'Model Architectures';
+const PAGE_DESCRIPTION =
+  'Architecture deep-dives for every model benchmarked on InferenceX: MoE and attention design, official vendor eval scores, and live inference performance data.';
+
+export const metadata: Metadata = {
+  title: { absolute: `${PAGE_TITLE} — MoE, Attention & Evals per Model | ${SITE_NAME}` },
+  description: PAGE_DESCRIPTION,
+  // English-only page: canonical without hreflang alternates (no /zh sibling).
+  alternates: { canonical: `${SITE_URL}/model` },
+  // og:image / twitter:image come from the colocated opengraph-image.tsx.
+  openGraph: {
+    title: `${PAGE_TITLE} — MoE, Attention & Evals per Model`,
+    description: PAGE_DESCRIPTION,
+    url: `${SITE_URL}/model`,
+    siteName: SITE_NAME,
+    type: 'website',
+    locale: 'en_US',
+  },
+  twitter: {
+    card: 'summary_large_image',
+    title: `${PAGE_TITLE} — MoE, Attention & Evals per Model`,
+    description: PAGE_DESCRIPTION,
+    site: AUTHOR_HANDLE,
+    creator: AUTHOR_HANDLE,
+  },
+};
+
+export default function ModelIndexPage() {
+  const pages = getModelPageSlugs()
+    .map((slug) => ({ slug, page: getModelPage(slug) }))
+    .filter((p): p is { slug: string; page: NonNullable<ReturnType<typeof getModelPage>> } =>
+      Boolean(p.page),
+    );
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'CollectionPage',
+    name: PAGE_TITLE,
+    description: PAGE_DESCRIPTION,
+    url: `${SITE_URL}/model`,
+    isPartOf: { '@type': 'WebSite', name: SITE_NAME, url: SITE_URL },
+    hasPart: pages.map(({ slug, page }) => ({
+      '@type': 'WebPage',
+      name: page.meta.title,
+      description: page.meta.description,
+      url: `${SITE_URL}/model/${slug}`,
+    })),
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-4">
+        <Card>
+          <header>
+            <p className="text-sm text-muted-foreground mb-2">
+              <Link href="/inference" className="hover:text-foreground transition-colors">
+                Inference Dashboard
+              </Link>{' '}
+              / Model
+            </p>
+            <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">{PAGE_TITLE}</h1>
+            <p className="mt-3 text-base lg:text-lg text-muted-foreground">{PAGE_DESCRIPTION}</p>
+          </header>
+          <ul
+            data-testid="model-index-list"
+            className="mt-6 grid grid-cols-1 md:grid-cols-2 gap-3 list-none p-0"
+          >
+            {pages.map(({ slug, page }) => {
+              const model = toModel(page.entry.displayName);
+              const arch = model ? getModelArchitecture(model) : undefined;
+              return (
+                <li key={slug}>
+                  <Link
+                    href={`/model/${slug}`}
+                    data-testid={`model-index-link-${slug}`}
+                    className="group h-full rounded-lg border border-border/50 bg-muted/30 px-4 py-3 flex flex-col gap-2 hover:bg-muted/50 transition-colors"
+                  >
+                    <div className="flex items-center justify-between gap-3">
+                      <span className="flex items-center gap-2 text-base font-semibold group-hover:text-primary transition-colors">
+                        <ModelDeveloperLogo developer={page.meta.developer} className="size-5" />
+                        {page.meta.title}
+                      </span>
+                      <span className="text-sm shrink-0 text-muted-foreground group-hover:text-foreground transition-colors">
+                        →
+                      </span>
+                    </div>
+                    {arch && (
+                      <span className="flex flex-wrap items-center gap-1.5">
+                        <Badge variant="outline" className="text-xs py-0">
+                          {arch.architectureType === 'moe' ? 'MoE' : 'Dense'}
+                        </Badge>
+                        <Badge variant="outline" className="text-xs py-0">
+                          {arch.attentionType === 'AlternatingSinkGQA'
+                            ? 'Sink/Full GQA'
+                            : arch.attentionType}
+                        </Badge>
+                        <Badge variant="outline" className="text-xs py-0">
+                          {formatParamCount(arch.totalParams)}
+                        </Badge>
+                      </span>
+                    )}
+                    <span className="text-sm text-muted-foreground line-clamp-3">
+                      {page.meta.description}
+                    </span>
+                    <span className="text-xs text-muted-foreground mt-auto">
+                      {page.meta.developer} &middot; Released {page.meta.releaseDate}
+                    </span>
+                  </Link>
+                </li>
+              );
+            })}
+          </ul>
+        </Card>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/overview/page.tsx b/packages/app/src/app/overview/page.tsx
index a0c044466..7ff2e9f8f 100644
--- a/packages/app/src/app/overview/page.tsx
+++ b/packages/app/src/app/overview/page.tsx
@@ -1,52 +1,13 @@
-import type { Metadata } from 'next';
-
-import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
-
-import { OverviewPageContent } from '@/components/overview/overview-page';
-import { enAlternates } from '@/lib/i18n';
 import {
-  resolveOverviewComparisonMode,
-  resolveOverviewEngineScope,
-  resolveOverviewModelScope,
-  resolveOverviewReferenceHardware,
-  resolveOverviewTier,
-} from '@/lib/overview-data';
-import { getOverviewPageData } from '@/lib/overview-data.server';
+  buildOverviewMetadata,
+  type OverviewRoutePageProps,
+  renderOverviewPage,
+} from '@/lib/overview-route.server';
 
 export const dynamic = 'force-dynamic';
 
-const DESCRIPTION =
-  'Compare hyperscaler cost per million total tokens across MI355X, B200, B300, GB200 and GB300 using the scenario shown for each active model.';
-
-export const metadata: Metadata = {
-  title: 'Inference Cost Overview',
-  description: DESCRIPTION,
-  alternates: enAlternates('/overview'),
-  openGraph: {
-    title: `Inference Cost Overview | ${SITE_NAME}`,
-    description: DESCRIPTION,
-    url: `${SITE_URL}/overview`,
-    type: 'website',
-  },
-  twitter: {
-    card: 'summary_large_image',
-    title: `Inference Cost Overview | ${SITE_NAME}`,
-    description: DESCRIPTION,
-  },
-};
-
-interface Props {
-  searchParams: Promise<Record<string, string | string[] | undefined>>;
-}
+export const metadata = buildOverviewMetadata('en');
 
-export default async function OverviewPage({ searchParams }: Props) {
-  const sp = await searchParams;
-  const data = await getOverviewPageData(
-    resolveOverviewTier(sp.tier),
-    resolveOverviewEngineScope(sp.engine),
-    resolveOverviewComparisonMode(sp.compare),
-    resolveOverviewReferenceHardware(sp.ref),
-    resolveOverviewModelScope(sp.models),
-  );
-  return <OverviewPageContent data={data} locale="en" />;
+export default function OverviewPage({ searchParams }: OverviewRoutePageProps) {
+  return renderOverviewPage({ locale: 'en', searchParams });
 }
diff --git a/packages/app/src/app/rankings/[slug]/page.tsx b/packages/app/src/app/rankings/[slug]/page.tsx
new file mode 100644
index 000000000..57e9587f0
--- /dev/null
+++ b/packages/app/src/app/rankings/[slug]/page.tsx
@@ -0,0 +1,210 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import {
+  AUTHOR_HANDLE,
+  AUTHOR_NAME,
+  SITE_NAME,
+  SITE_URL,
+  SUPPORTERS_LINE,
+} from '@semianalysisai/inferencex-constants';
+
+import { fmtCostPerMtok, fmtThroughput } from '@/components/live-seo/format';
+import {
+  RankingsDetailContent,
+  type RankingsStrings,
+} from '@/components/live-seo/rankings-page-sections';
+import { JsonLd } from '@/components/json-ld';
+import { enAlternates } from '@/lib/i18n';
+import {
+  getRankingPageEntry,
+  rankingPageDescription,
+  rankingPageHeading,
+  rankingPageKeywords,
+  rankingPageTitle,
+  scenarioLabel,
+  type RankingPageEntry,
+} from '@/lib/rankings';
+import { getRankingPageData, type RankingPageData } from '@/lib/run-rankings-data.server';
+
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string }>;
+}
+
+function statLedDescription(entry: RankingPageEntry, data: RankingPageData): string {
+  const [first] = data.rows;
+  if (!first) return rankingPageDescription(entry);
+  if (entry.kind === 'fastest-gpu' && first.throughputPerGpu !== null) {
+    return `${first.hardwareLabel} currently serves ${entry.model.seoName} fastest at ${fmtThroughput(first.throughputPerGpu)} tokens/s per GPU. Live ranking of ${data.rows.length} platforms, re-benchmarked continuously.`;
+  }
+  if (entry.kind === 'cheapest-gpu' && first.costPerMtok !== null) {
+    return `${first.hardwareLabel} currently serves ${entry.model.seoName} cheapest at ${fmtCostPerMtok(first.costPerMtok)} per million tokens. Live ranking of ${data.rows.length} platforms, re-benchmarked continuously.`;
+  }
+  return rankingPageDescription(entry);
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug } = await params;
+  const entry = getRankingPageEntry(slug);
+  if (!entry) return {};
+  const data = await getRankingPageData(entry.slug);
+  const title = rankingPageTitle(entry);
+  const description = `${data ? statLedDescription(entry, data) : rankingPageDescription(entry)} ${SUPPORTERS_LINE}`;
+  const url = `${SITE_URL}/rankings/${entry.slug}`;
+  return {
+    title,
+    description,
+    keywords: rankingPageKeywords(entry),
+    authors: [{ name: AUTHOR_NAME }],
+    alternates: enAlternates(`/rankings/${entry.slug}`),
+    openGraph: { title: `${title} | ${SITE_NAME}`, description, url, type: 'article' },
+    twitter: {
+      card: 'summary_large_image',
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      creator: AUTHOR_HANDLE,
+    },
+  };
+}
+
+function buildFaq(
+  entry: RankingPageEntry,
+  data: RankingPageData,
+): { question: string; answer: string }[] {
+  const model = entry.model.seoName;
+  const [first] = data.rows;
+  const workload = scenarioLabel(data.scenario, 'en');
+  const fallbackLeader = `The ranking is refreshed continuously from InferenceX benchmark runs; check the table above for the current leader.`;
+  const kindFaq =
+    entry.kind === 'fastest-gpu'
+      ? [
+          {
+            question: `What is the fastest GPU for ${model} inference?`,
+            answer:
+              first && typeof first.throughputPerGpu === 'number'
+                ? `As of the latest benchmark runs, ${first.hardwareLabel} leads at ${fmtThroughput(first.throughputPerGpu)} tokens/s per GPU on ${workload}, measured at a matched interactivity target of ${data.tier} tokens/s per user.`
+                : fallbackLeader,
+          },
+          {
+            question: `How is "fastest" measured?`,
+            answer: `Every platform is read at the same interactivity tier (${data.tier} tokens/s per user) on ${workload}, then ranked by measured throughput per GPU. This is the same derivation the InferenceX overview leaderboard renders, so the ranking can never disagree with the dashboard.`,
+          },
+        ]
+      : [
+          {
+            question: `What is the cheapest GPU to run ${model}?`,
+            answer:
+              first && typeof first.costPerMtok === 'number'
+                ? `As of the latest benchmark runs, ${first.hardwareLabel} is cheapest at ${fmtCostPerMtok(first.costPerMtok)} per million total tokens on ${workload}, at hyperscaler $/GPU/hr pricing and a matched interactivity target of ${data.tier} tokens/s per user.`
+                : fallbackLeader,
+          },
+          {
+            question: `How is cost per million tokens calculated?`,
+            answer: `Measured throughput per GPU at the ${data.tier} tokens/s per user tier is converted to $ per million total (input plus output) tokens using hyperscaler $/GPU/hr rates from the SemiAnalysis AI Cloud TCO model. Slower interactivity targets or cheaper rental tiers change the absolute numbers but rarely the order.`,
+          },
+        ];
+  return [
+    ...kindFaq,
+    {
+      question: `How often is this ranking updated?`,
+      answer: data.newest
+        ? `Benchmarks re-run continuously on the InferenceX cluster fleet; the newest result feeding this ranking landed on ${data.newest}. The page always renders the latest data.`
+        : `Benchmarks re-run continuously on the InferenceX cluster fleet, and the page always renders the latest data.`,
+    },
+  ];
+}
+
+export default async function RankingPage({ params }: Props) {
+  const { slug } = await params;
+  const entry = getRankingPageEntry(slug);
+  if (!entry) notFound();
+  const data = await getRankingPageData(entry.slug);
+  if (!data) notFound();
+
+  const url = `${SITE_URL}/rankings/${entry.slug}`;
+  const heading = rankingPageHeading(entry);
+  const faq = buildFaq(entry, data);
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@graph': [
+      {
+        '@type': 'ItemList',
+        name: rankingPageTitle(entry),
+        itemListOrder: 'https://schema.org/ItemListOrderAscending',
+        numberOfItems: data.rows.length,
+        itemListElement: data.rows.map((row) => ({
+          '@type': 'ListItem',
+          position: row.rank,
+          name: row.hardwareLabel,
+          ...(row.chip && { url: `${SITE_URL}/chips/${row.chip.slug}` }),
+        })),
+      },
+      {
+        '@type': 'FAQPage',
+        mainEntity: faq.map((item) => ({
+          '@type': 'Question',
+          name: item.question,
+          acceptedAnswer: { '@type': 'Answer', text: item.answer },
+        })),
+      },
+      {
+        '@type': 'BreadcrumbList',
+        itemListElement: [
+          { '@type': 'ListItem', position: 1, name: 'Rankings', item: `${SITE_URL}/rankings` },
+          { '@type': 'ListItem', position: 2, name: heading, item: url },
+        ],
+      },
+    ],
+  };
+
+  const t: RankingsStrings = {
+    backLabel: 'GPU rankings',
+    heading,
+    scenarioNote: `Measured on ${scenarioLabel(data.scenario, 'en')} at a matched interactivity target of ${data.tier} tokens/s per user. Hardware without a measurement at this operating point is not ranked.`,
+    tableCaption: `${heading}: live benchmark ranking`,
+    colRank: 'Rank',
+    colGpu: 'GPU',
+    colThroughput: 'Tokens/s per GPU',
+    colCost: '$ / 1M tokens',
+    colPrecision: 'Precision',
+    colEngine: 'Engine',
+    emptyState: `No benchmark measurements are available for ${entry.model.seoName} at this operating point yet. The InferenceX fleet re-benchmarks continuously; this page fills in automatically as soon as runs land.`,
+    siblingLead:
+      entry.kind === 'fastest-gpu'
+        ? 'Optimizing for cost instead of speed?'
+        : 'Optimizing for speed instead of cost?',
+    siblingLabel:
+      entry.kind === 'fastest-gpu'
+        ? `See the cheapest GPU for ${entry.model.seoName}`
+        : `See the fastest GPU for ${entry.model.seoName}`,
+    methodologyHeading: 'Methodology',
+    methodologyBody: [
+      `Every number on this page is a measurement, not a spec-sheet estimate. The InferenceX fleet serves ${entry.model.seoName} on real hardware with community serving engines, sweeping concurrency to trace each platform's throughput-versus-interactivity frontier.`,
+      `Platforms are then read at the same operating point (${data.tier} tokens/s per user) so the comparison is iso-interactivity: a GPU cannot win by quoting throughput at an unusably slow per-user speed. Cost converts measured throughput to $ per million total tokens using $/GPU/hr rates from the SemiAnalysis AI Cloud TCO model.`,
+      `The derivation is shared with the InferenceX overview leaderboard, and results re-run continuously, so this ranking updates as new engine releases and configs land.`,
+    ],
+    faqHeading: 'Frequently asked questions',
+    faq,
+    exploreHeading: 'Explore the data',
+    exploreLinks: [
+      {
+        href: `/model/${entry.model.slug}`,
+        label: `${entry.model.seoName} deep dive: architecture, evals, and performance`,
+      },
+      { href: '/inference', label: 'Interactive inference dashboard' },
+      { href: '/overview', label: 'Cross-model inference leaderboard' },
+      { href: '/compare', label: 'Head-to-head GPU comparisons' },
+      { href: '/chips', label: 'GPU specs and pricing' },
+    ],
+  };
+
+  return (
+    <>
+      <JsonLd data={jsonLd} />
+      <RankingsDetailContent entry={entry} data={data} t={t} locale="en" />
+    </>
+  );
+}
diff --git a/packages/app/src/app/rankings/page.tsx b/packages/app/src/app/rankings/page.tsx
new file mode 100644
index 000000000..395f193b3
--- /dev/null
+++ b/packages/app/src/app/rankings/page.tsx
@@ -0,0 +1,103 @@
+import type { Metadata } from 'next';
+import Link from 'next/link';
+
+import { SITE_NAME, SITE_URL, SUPPORTERS_LINE } from '@semianalysisai/inferencex-constants';
+
+import { JsonLd } from '@/components/json-ld';
+import { Card } from '@/components/ui/card';
+import { enAlternates } from '@/lib/i18n';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
+import { getAllRankingPageEntries } from '@/lib/rankings';
+
+const title = 'GPU Rankings for LLM Inference';
+const description =
+  'Live leaderboards of the fastest and cheapest GPUs for serving open LLMs like DeepSeek, Kimi, GLM, MiniMax, and Qwen. Ranked by measured tokens/s per GPU and $ per million tokens, re-benchmarked continuously.';
+
+export const metadata: Metadata = {
+  title,
+  description: `${description} ${SUPPORTERS_LINE}`,
+  keywords: [
+    'fastest GPU for LLM inference',
+    'cheapest GPU for LLM inference',
+    'GPU leaderboard LLM',
+    'best GPU for AI inference',
+    'LLM inference cost per token',
+    'H100 vs B200 vs MI355X',
+    'tokens per second per GPU',
+    'GPU inference rankings',
+  ],
+  alternates: enAlternates('/rankings'),
+  openGraph: {
+    title: `${title} | ${SITE_NAME}`,
+    description,
+    url: `${SITE_URL}/rankings`,
+    type: 'website',
+  },
+  twitter: { card: 'summary_large_image', title, description },
+};
+
+export default function RankingsIndexPage() {
+  const entries = getAllRankingPageEntries();
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'CollectionPage',
+    name: title,
+    description,
+    url: `${SITE_URL}/rankings`,
+    hasPart: entries.map((entry) => ({
+      '@type': 'WebPage',
+      name:
+        entry.kind === 'fastest-gpu'
+          ? `Fastest GPU for ${entry.model.seoName}`
+          : `Cheapest GPU for ${entry.model.seoName}`,
+      url: `${SITE_URL}/rankings/${entry.slug}`,
+    })),
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8">
+        <div className="mx-auto max-w-5xl">
+          <header className="pt-8 md:pt-12">
+            <h1 className="max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-5xl">
+              {title}
+            </h1>
+            <p className="mt-4 max-w-3xl text-lg leading-relaxed text-muted-foreground">
+              Which GPU should serve your model? Every ranking below is derived from measured
+              benchmark runs at matched interactivity, not spec sheets, and re-renders as new
+              results land.
+            </p>
+          </header>
+
+          <section className="mt-10 mb-16 grid gap-4 sm:grid-cols-2">
+            {INFERENCE_MODEL_SLUGS.map((model) => (
+              <Card key={model.slug} className="p-5">
+                <h2 className="text-lg font-semibold tracking-tight">{model.seoName}</h2>
+                <ul className="mt-3 space-y-1">
+                  <li>
+                    <Link
+                      href={`/rankings/fastest-gpu-for-${model.slug}`}
+                      className="text-sm font-medium text-brand hover:underline"
+                    >
+                      Fastest GPU for {model.seoName}
+                    </Link>
+                  </li>
+                  <li>
+                    <Link
+                      href={`/rankings/cheapest-gpu-for-${model.slug}`}
+                      className="text-sm font-medium text-brand hover:underline"
+                    >
+                      Cheapest GPU for {model.seoName}
+                    </Link>
+                  </li>
+                </ul>
+              </Card>
+            ))}
+          </section>
+        </div>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/run/[slug]/page.tsx b/packages/app/src/app/run/[slug]/page.tsx
new file mode 100644
index 000000000..dc4023934
--- /dev/null
+++ b/packages/app/src/app/run/[slug]/page.tsx
@@ -0,0 +1,222 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import {
+  AUTHOR_HANDLE,
+  AUTHOR_NAME,
+  SITE_NAME,
+  SITE_URL,
+  SUPPORTERS_LINE,
+} from '@semianalysisai/inferencex-constants';
+
+import { fmtCostPerMtok, fmtThroughput } from '@/components/live-seo/format';
+import {
+  latencyQuote,
+  RunDetailContent,
+  type RunStrings,
+} from '@/components/live-seo/run-page-sections';
+import { JsonLd } from '@/components/json-ld';
+import { enAlternates } from '@/lib/i18n';
+import { scenarioLabel } from '@/lib/rankings';
+import {
+  getRunPageEntry,
+  runPageDescription,
+  runPageFaqQuestions,
+  runPageHeading,
+  runPageKeywords,
+  runPageTitle,
+  type RunPageEntry,
+} from '@/lib/run-pages';
+import { getRunPageData, type RunPageData } from '@/lib/run-rankings-data.server';
+
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string }>;
+}
+
+function primaryRead(data: RunPageData) {
+  return data.tierLadder.find((read) => read.tier === data.primaryTier);
+}
+
+function statLedDescription(entry: RunPageEntry, data: RunPageData): string {
+  const read = primaryRead(data);
+  if (!data.hasData || !read?.throughputPerGpu) return runPageDescription(entry);
+  const cost =
+    read.costPerMtok === null
+      ? ''
+      : `, ${fmtCostPerMtok(read.costPerMtok)} per million tokens at hyperscaler pricing`;
+  return `Measured: ${entry.model.seoName} sustains ${fmtThroughput(read.throughputPerGpu)} tokens/s per GPU on ${entry.chip.label} at ${data.primaryTier} tokens/s per user${cost}. Live data from ${data.configCount} benchmarked configs.`;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug } = await params;
+  const entry = getRunPageEntry(slug);
+  if (!entry) return {};
+  const data = await getRunPageData(entry.slug);
+  const title = runPageTitle(entry);
+  const description = `${data ? statLedDescription(entry, data) : runPageDescription(entry)} ${SUPPORTERS_LINE}`;
+  const url = `${SITE_URL}/run/${entry.slug}`;
+  return {
+    title,
+    description,
+    keywords: runPageKeywords(entry),
+    authors: [{ name: AUTHOR_NAME }],
+    // Valid slugs without benchmark data render an empty state; keep those
+    // out of the index (the sitemap already excludes them).
+    ...(data?.hasData ? {} : { robots: { index: false, follow: true } }),
+    alternates: enAlternates(`/run/${entry.slug}`),
+    openGraph: { title: `${title} | ${SITE_NAME}`, description, url, type: 'article' },
+    twitter: {
+      card: 'summary_large_image',
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      creator: AUTHOR_HANDLE,
+    },
+  };
+}
+
+function buildFaq(entry: RunPageEntry, data: RunPageData): { question: string; answer: string }[] {
+  const questions = runPageFaqQuestions(entry);
+  const model = entry.model.seoName;
+  const chip = entry.chip.label;
+  const read = primaryRead(data);
+  const workload = scenarioLabel(data.scenario, 'en');
+
+  const peak =
+    data.bestThroughputPerGpu === null ? 'n/a' : fmtThroughput(data.bestThroughputPerGpu);
+  const throughputAnswer =
+    typeof read?.throughputPerGpu === 'number'
+      ? `At an interactivity target of ${data.primaryTier} tokens/s per user on ${workload}, ${chip} sustains ${fmtThroughput(read.throughputPerGpu)} tokens/s per GPU serving ${model}${read.framework ? ` with ${read.framework}` : ''}${read.precision ? ` in ${read.precision.toUpperCase()}` : ''}. Peak measured throughput across all configs is ${peak} tokens/s per GPU.`
+      : `The InferenceX fleet has ${data.configCount} benchmarked configs for this pairing; see the interactivity ladder above for the operating points reached so far.`;
+
+  const costAnswer =
+    typeof read?.costPerMtok === 'number'
+      ? `${fmtCostPerMtok(read.costPerMtok)} per million total tokens at hyperscaler $/GPU/hr pricing, at ${data.primaryTier} tokens/s per user. Neocloud and retail rental tiers are tabulated above; slower interactivity targets lower the cost further.`
+      : `Cost per million tokens is derived from measured throughput and $/GPU/hr rates from the SemiAnalysis AI Cloud TCO model; it appears once this pairing reaches the primary interactivity tier.`;
+
+  const servingAnswer = `The runs behind this page used ${data.frameworks.join(', ')} in ${data.precisions.map((p) => p.toUpperCase()).join(', ')}${data.hasDisagg ? ', including disaggregated prefill' : ''}${data.hasMultinode ? ' and multi-node serving' : ''}. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.`;
+
+  const methodologyAnswer = `Every number is measured on real ${chip} hardware by the InferenceX fleet, sweeping concurrency on ${workload} to trace the throughput-versus-interactivity frontier${data.newest ? `; the newest run landed on ${data.newest}` : ''}. The same derivation powers the InferenceX overview leaderboard.`;
+
+  return [
+    { question: questions.throughput, answer: throughputAnswer },
+    { question: questions.cost, answer: costAnswer },
+    { question: questions.serving, answer: servingAnswer },
+    { question: questions.methodology, answer: methodologyAnswer },
+  ];
+}
+
+export default async function RunPage({ params }: Props) {
+  const { slug } = await params;
+  const entry = getRunPageEntry(slug);
+  if (!entry) notFound();
+  const data = await getRunPageData(entry.slug);
+  if (!data) notFound();
+
+  const url = `${SITE_URL}/run/${entry.slug}`;
+  const heading = runPageHeading(entry);
+  const faq = data.hasData ? buildFaq(entry, data) : [];
+  const read = primaryRead(data);
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@graph': [
+      {
+        '@type': 'WebPage',
+        name: runPageTitle(entry),
+        url,
+        ...(data.newest && { dateModified: data.newest }),
+        ...(data.oldest && { datePublished: data.oldest }),
+      },
+      ...(faq.length > 0
+        ? [
+            {
+              '@type': 'FAQPage',
+              mainEntity: faq.map((item) => ({
+                '@type': 'Question',
+                name: item.question,
+                acceptedAnswer: { '@type': 'Answer', text: item.answer },
+              })),
+            },
+          ]
+        : []),
+      {
+        '@type': 'BreadcrumbList',
+        itemListElement: [
+          { '@type': 'ListItem', position: 1, name: 'Run', item: `${SITE_URL}/run` },
+          { '@type': 'ListItem', position: 2, name: heading, item: url },
+        ],
+      },
+    ],
+  };
+
+  const quickAnswerCost =
+    typeof read?.costPerMtok === 'number'
+      ? `, which works out to ${fmtCostPerMtok(read.costPerMtok)} per million tokens at hyperscaler pricing`
+      : '';
+  const latency = latencyQuote(data);
+  const quickAnswerLatency = latency
+    ? ` Fastest measured TTFT: ${latency.ttft}; fastest TPOT: ${latency.tpot} (each the best across all configs, not one run).`
+    : '';
+  const quickAnswer =
+    typeof read?.throughputPerGpu === 'number'
+      ? `${entry.model.seoName} sustains ${fmtThroughput(read.throughputPerGpu)} tokens/s per GPU on ${entry.chip.label} at ${data.primaryTier} tokens/s per user${quickAnswerCost}${read.framework ? `, served by ${read.framework}` : ''}.${quickAnswerLatency}`
+      : `${entry.model.seoName} runs on ${entry.chip.label}: ${data.configCount} benchmarked configs so far. See the interactivity ladder below for measured operating points.`;
+
+  const t: RunStrings = {
+    backHref: '/run',
+    backLabel: 'All model and GPU pairings',
+    heading,
+    quickAnswerLabel: 'Quick answer',
+    quickAnswer,
+    statConfigs: 'Benchmarked configs',
+    statEngines: 'Serving engines',
+    statPrecisions: 'Precisions',
+    statFreshness: 'Run dates',
+    ladderHeading: 'Throughput at every interactivity target',
+    ladderIntro: `Serving is a trade-off: push more concurrent users through a GPU and each user's tokens arrive slower. The ladder below reads the measured frontier at each per-user speed target on ${scenarioLabel(data.scenario, 'en')}, using the best engine and precision at that point.`,
+    colTier: 'Per-user target',
+    colThroughput: 'Tokens/s per GPU',
+    colCost: '$ / 1M tokens',
+    colEngine: 'Engine',
+    colPrecision: 'Precision',
+    costHeading: 'What serving actually costs',
+    costIntro: `Converting the ${data.primaryTier} tokens/s per user operating point to $ per million total tokens across rental pricing tiers from the SemiAnalysis AI Cloud TCO model.`,
+    colPriceTier: 'Pricing tier',
+    colGpuHour: '$ / GPU / hr',
+    colCostPerMtok: '$ / 1M tokens',
+    priceTierLabels: {
+      hyperscaler: 'Hyperscaler',
+      neocloud: 'Neocloud',
+      retail: 'Retail',
+    },
+    emptyState: `The InferenceX fleet has not published benchmark runs for ${entry.model.seoName} on ${entry.chip.label} yet. Benchmarks re-run continuously; this page fills in automatically as soon as results land.`,
+    faqHeading: 'Frequently asked questions',
+    faq,
+    exploreHeading: 'Explore the data',
+    exploreLinks: [
+      {
+        href: `/rankings/fastest-gpu-for-${entry.model.slug}`,
+        label: `Fastest GPU for ${entry.model.seoName}: full ranking`,
+      },
+      {
+        href: `/rankings/cheapest-gpu-for-${entry.model.slug}`,
+        label: `Cheapest GPU for ${entry.model.seoName}: full ranking`,
+      },
+      {
+        href: `/model/${entry.model.slug}`,
+        label: `${entry.model.seoName} deep dive: architecture, evals, and performance`,
+      },
+      { href: `/chips/${entry.chip.slug}`, label: `${entry.chip.title} specs and pricing` },
+      { href: '/inference', label: 'Interactive inference dashboard' },
+    ],
+  };
+
+  return (
+    <>
+      <JsonLd data={jsonLd} />
+      <RunDetailContent entry={entry} data={data} t={t} />
+    </>
+  );
+}
diff --git a/packages/app/src/app/run/page.tsx b/packages/app/src/app/run/page.tsx
new file mode 100644
index 000000000..186908152
--- /dev/null
+++ b/packages/app/src/app/run/page.tsx
@@ -0,0 +1,106 @@
+import type { Metadata } from 'next';
+import Link from 'next/link';
+
+import { SITE_NAME, SITE_URL, SUPPORTERS_LINE } from '@semianalysisai/inferencex-constants';
+
+import { JsonLd } from '@/components/json-ld';
+import { Card } from '@/components/ui/card';
+import { enAlternates } from '@/lib/i18n';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
+import { runPageHeading, type RunPageEntry } from '@/lib/run-pages';
+import { getAvailableRunEntries } from '@/lib/run-rankings-data.server';
+
+export const dynamic = 'force-dynamic';
+
+const title = 'Run Any Model on Any GPU: Measured Results';
+const description =
+  'Measured throughput, latency, and cost for every open LLM and GPU pairing the InferenceX fleet benchmarks: DeepSeek, Kimi, GLM, MiniMax, and Qwen on H100, H200, B200, GB200 NVL72, MI300X, MI325X, MI355X, and more.';
+
+export function generateMetadata(): Metadata {
+  return {
+    title,
+    description: `${description} ${SUPPORTERS_LINE}`,
+    keywords: [
+      'run LLM on GPU benchmark',
+      'LLM GPU throughput',
+      'model on GPU performance',
+      'DeepSeek GPU benchmark',
+      'Kimi GPU benchmark',
+      'LLM inference cost per GPU',
+      'H100 LLM benchmark',
+      'MI355X LLM benchmark',
+    ],
+    alternates: enAlternates('/run'),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url: `${SITE_URL}/run`,
+      type: 'website',
+    },
+    twitter: { card: 'summary_large_image', title, description },
+  };
+}
+
+export default async function RunIndexPage() {
+  const entries = await getAvailableRunEntries();
+  const byModel = new Map<string, RunPageEntry[]>();
+  for (const entry of entries) {
+    const list = byModel.get(entry.model.slug) ?? [];
+    list.push(entry);
+    byModel.set(entry.model.slug, list);
+  }
+  const models = INFERENCE_MODEL_SLUGS.filter((model) => byModel.has(model.slug));
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'CollectionPage',
+    name: title,
+    description,
+    url: `${SITE_URL}/run`,
+    hasPart: entries.map((entry) => ({
+      '@type': 'WebPage',
+      name: runPageHeading(entry),
+      url: `${SITE_URL}/run/${entry.slug}`,
+    })),
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8">
+        <div className="mx-auto max-w-5xl">
+          <header className="pt-8 md:pt-12">
+            <h1 className="max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-5xl">
+              {title}
+            </h1>
+            <p className="mt-4 max-w-3xl text-lg leading-relaxed text-muted-foreground">
+              Pick a model and a GPU: each page below answers how fast it runs, what it costs per
+              million tokens, and which serving engine produced the number, from continuously
+              re-benchmarked runs on real hardware.
+            </p>
+          </header>
+
+          <section className="mt-10 mb-16 space-y-8">
+            {models.map((model) => (
+              <Card key={model.slug} className="p-5">
+                <h2 className="text-lg font-semibold tracking-tight">{model.seoName}</h2>
+                <ul className="mt-3 grid gap-1 sm:grid-cols-2">
+                  {(byModel.get(model.slug) ?? []).map((entry) => (
+                    <li key={entry.slug}>
+                      <Link
+                        href={`/run/${entry.slug}`}
+                        className="text-sm font-medium text-brand hover:underline"
+                      >
+                        {model.seoName} on {entry.chip.label}
+                      </Link>
+                    </li>
+                  ))}
+                </ul>
+              </Card>
+            ))}
+          </section>
+        </div>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/sitemap.test.ts b/packages/app/src/app/sitemap.test.ts
new file mode 100644
index 000000000..c9c48bb66
--- /dev/null
+++ b/packages/app/src/app/sitemap.test.ts
@@ -0,0 +1,126 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+import { DASHBOARD_ROUTES } from '@/lib/dashboard-routes';
+import { getAllChipRouteSlugs } from '@/lib/chip-pages';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
+import { zhPath } from '@/lib/i18n';
+const mocks = vi.hoisted(() => ({
+  fixturesMode: false,
+  getDb: vi.fn(() => ({})),
+  listDatasets: vi.fn(() => Promise.resolve([{ slug: 'agentx-fixture' }])),
+}));
+
+vi.mock('@semianalysisai/inferencex-db/connection', () => ({
+  get FIXTURES_MODE() {
+    return mocks.fixturesMode;
+  },
+  getDb: mocks.getDb,
+}));
+vi.mock('@semianalysisai/inferencex-db/queries/datasets', () => ({
+  listDatasets: mocks.listDatasets,
+}));
+vi.mock('@/lib/agentx-optimizations', () => ({ AGENTX_OPTIMIZATION_SLUGS: [] }));
+vi.mock('@/lib/blog', () => ({ getAllPosts: () => [] }));
+vi.mock('@/lib/compare-availability', () => ({
+  getAllComparableCompareSlugs: () => Promise.resolve([]),
+}));
+vi.mock('@/lib/compare-variant-availability', () => ({
+  getAllComparablePrecisionSlugs: () => Promise.resolve([]),
+  getAllComparableSpecDecodeSlugs: () => Promise.resolve([]),
+}));
+vi.mock('@/lib/glossary', () => ({ getAllGlossaryEntries: () => [] }));
+vi.mock('@/lib/run-rankings-data.server', () => ({
+  getAvailableRunEntries: () => Promise.resolve([]),
+}));
+
+import { getAllRankingPageEntries } from '@/lib/rankings';
+
+import sitemap from './sitemap';
+
+describe('sitemap locale parity', () => {
+  beforeEach(() => {
+    mocks.fixturesMode = false;
+    vi.clearAllMocks();
+  });
+
+  it('emits both locales for every indexable dashboard route and omits non-indexable routes', async () => {
+    const entries = await sitemap();
+    const urls = new Set(entries.map((entry) => entry.url));
+
+    for (const route of DASHBOARD_ROUTES) {
+      const englishUrl =
+        route.canonicalPath === '/' ? SITE_URL : `${SITE_URL}${route.canonicalPath}`;
+      const chineseUrl = `${SITE_URL}${zhPath(route.canonicalPath)}`;
+
+      expect(urls.has(englishUrl)).toBe(route.indexable);
+      expect(urls.has(chineseUrl)).toBe(route.indexable && route.localeMirrored);
+    }
+  });
+
+  it('emits both locales for every /inference/<model> page', async () => {
+    const entries = await sitemap();
+    const urls = new Set(entries.map((entry) => entry.url));
+    expect(INFERENCE_MODEL_SLUGS.length).toBeGreaterThan(0);
+    for (const entry of INFERENCE_MODEL_SLUGS) {
+      expect(urls.has(`${SITE_URL}/inference/${entry.slug}`)).toBe(true);
+      expect(urls.has(`${SITE_URL}${zhPath(`/inference/${entry.slug}`)}`)).toBe(true);
+    }
+  });
+
+  it('emits both locales for the chips index and every chip page', async () => {
+    const entries = await sitemap();
+    const urls = new Set(entries.map((entry) => entry.url));
+    expect(urls.has(`${SITE_URL}/chips`)).toBe(true);
+    expect(urls.has(`${SITE_URL}${zhPath('/chips')}`)).toBe(true);
+    const slugs = getAllChipRouteSlugs();
+    expect(slugs.length).toBe(21);
+    for (const slug of slugs) {
+      expect(urls.has(`${SITE_URL}/chips/${slug}`)).toBe(true);
+      expect(urls.has(`${SITE_URL}${zhPath(`/chips/${slug}`)}`)).toBe(true);
+    }
+  });
+
+  it('emits both locales for the rankings index and every ranking page', async () => {
+    const entries = await sitemap();
+    const urls = new Set(entries.map((entry) => entry.url));
+    expect(urls.has(`${SITE_URL}/rankings`)).toBe(true);
+    expect(urls.has(`${SITE_URL}${zhPath('/rankings')}`)).toBe(true);
+    const rankingEntries = getAllRankingPageEntries();
+    expect(rankingEntries.length).toBeGreaterThan(0);
+    for (const entry of rankingEntries) {
+      expect(urls.has(`${SITE_URL}/rankings/${entry.slug}`)).toBe(true);
+      expect(urls.has(`${SITE_URL}${zhPath(`/rankings/${entry.slug}`)}`)).toBe(true);
+    }
+  });
+
+  it('emits the /run index in both locales and only available run pages', async () => {
+    const entries = await sitemap();
+    const urls = new Set(entries.map((entry) => entry.url));
+    expect(urls.has(`${SITE_URL}/run`)).toBe(true);
+    expect(urls.has(`${SITE_URL}${zhPath('/run')}`)).toBe(true);
+    // Availability mock returns no pairs, so no /run/<pair> URLs may leak in.
+    for (const url of urls) {
+      expect(url.startsWith(`${SITE_URL}/run/`)).toBe(false);
+      expect(url.startsWith(`${SITE_URL}/zh/run/`)).toBe(false);
+    }
+  });
+
+  it('emits both locales for indexable AgentX dataset pages', async () => {
+    const entries = await sitemap();
+    const urls = new Set(entries.map((entry) => entry.url));
+    expect(urls.has(`${SITE_URL}/agentx/agentx-fixture`)).toBe(true);
+    expect(urls.has(`${SITE_URL}/zh/agentx/agentx-fixture`)).toBe(true);
+  });
+
+  it('does not require a database connection in fixture mode', async () => {
+    mocks.fixturesMode = true;
+
+    const entries = await sitemap();
+    const urls = new Set(entries.map((entry) => entry.url));
+
+    expect(mocks.getDb).not.toHaveBeenCalled();
+    expect(mocks.listDatasets).not.toHaveBeenCalled();
+    expect(urls.has(`${SITE_URL}/agentx/agentx-fixture`)).toBe(false);
+  });
+});
diff --git a/packages/app/src/app/sitemap.ts b/packages/app/src/app/sitemap.ts
index d85b07299..b05f2713e 100644
--- a/packages/app/src/app/sitemap.ts
+++ b/packages/app/src/app/sitemap.ts
@@ -1,6 +1,11 @@
 import type { MetadataRoute } from 'next';
+import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
+import { listDatasets } from '@semianalysisai/inferencex-db/queries/datasets';
 
+import { AGENTX_OPTIMIZATION_SLUGS } from '@/lib/agentx-optimizations';
+import { DASHBOARD_ROUTES } from '@/lib/dashboard-routes';
 import { getAllPosts } from '@/lib/blog';
+import { getModelPageSlugs } from '@/lib/model-pages';
 import { getAllComparableCompareSlugs } from '@/lib/compare-availability';
 import { canonicalCompareSlug } from '@/lib/compare-slug';
 import {
@@ -11,20 +16,20 @@ import {
   canonicalPrecisionCompareSlug,
   canonicalSpecDecodeCompareSlug,
 } from '@/lib/compare-variant-slug';
+import { getAllChipRouteSlugs } from '@/lib/chip-pages';
 import { getAllGlossaryEntries } from '@/lib/glossary';
+import { ACTIVE_INFERENCE_MODEL_SLUGS, INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
 import { languageAlternates, zhPath } from '@/lib/i18n';
+import {
+  DEFAULT_ROUTE_MODEL,
+  MODEL_ROUTE_TABS,
+  MODEL_ROUTES,
+  modelRoutePath,
+} from '@/lib/model-routes';
+import { getAllRankingPageEntries } from '@/lib/rankings';
+import { getAvailableRunEntries } from '@/lib/run-rankings-data.server';
 import { SITE_URL as BASE_URL } from '@semianalysisai/inferencex-constants';
 
-const TABS = [
-  'evaluation',
-  'historical',
-  'calculator',
-  'reliability',
-  'gpu-specs',
-  'gpu-metrics',
-  'collectivex',
-] as const;
-
 type SitemapEntry = MetadataRoute.Sitemap[number];
 
 /**
@@ -50,27 +55,40 @@ export default async function sitemap(): Promise<MetadataRoute.Sitemap> {
   const now = new Date().toISOString();
   // Only emit (model, pair) URLs that have benchmark data on both sides —
   // avoids polluting the sitemap with empty pages that hurt crawl budget.
-  const [compareSlugs, precisionSlugs, specDecodeSlugs] = await Promise.all([
+  const [compareSlugs, precisionSlugs, specDecodeSlugs, datasets, runEntries] = await Promise.all([
     getAllComparableCompareSlugs(),
     getAllComparablePrecisionSlugs(),
     getAllComparableSpecDecodeSlugs(),
+    FIXTURES_MODE ? Promise.resolve([]) : listDatasets(getDb()),
+    getAvailableRunEntries(),
   ]);
   const zhPosts = new Set(getAllPosts('zh').map((post) => post.slug));
 
   return [
-    ...localizedPair('/', { lastModified: now, changeFrequency: 'daily', priority: 1 }),
+    ...DASHBOARD_ROUTES.filter((route) => route.indexable).flatMap((route) =>
+      localizedPair(route.canonicalPath, {
+        lastModified: now,
+        changeFrequency: 'daily' as const,
+        priority: route.canonicalPath === '/' ? 1 : 0.9,
+      }),
+    ),
+    // Per-model tab routes (/calculator/<slug>, /historical/<slug>). The
+    // default model's pages canonicalize to the bare tab paths (already
+    // emitted above), so they stay out of the sitemap.
+    ...MODEL_ROUTE_TABS.flatMap((tab) =>
+      MODEL_ROUTES.filter((route) => route.model !== DEFAULT_ROUTE_MODEL).flatMap((route) =>
+        localizedPair(modelRoutePath(tab, route.slug), {
+          lastModified: now,
+          changeFrequency: 'daily' as const,
+          priority: 0.8,
+        }),
+      ),
+    ),
     ...localizedPair('/overview', {
       lastModified: now,
       changeFrequency: 'daily',
       priority: 0.9,
     }),
-    ...TABS.flatMap((tab) =>
-      localizedPair(`/${tab}`, {
-        lastModified: now,
-        changeFrequency: 'daily' as const,
-        priority: 0.9,
-      }),
-    ),
     ...localizedPair('/quotes', { lastModified: now, changeFrequency: 'monthly', priority: 0.6 }),
     ...localizedPair('/about', { lastModified: now, changeFrequency: 'monthly', priority: 0.6 }),
     ...localizedPair('/land-acknowledgement', {
@@ -78,6 +96,23 @@ export default async function sitemap(): Promise<MetadataRoute.Sitemap> {
       changeFrequency: 'yearly',
       priority: 0.4,
     }),
+    // Per-model inference pages — the indexable path form of
+    // `/inference?g_model=…`. Deprecated models keep their pages (historical
+    // data stays reachable) at a lower priority than actively benchmarked ones.
+    ...INFERENCE_MODEL_SLUGS.flatMap((entry) =>
+      localizedPair(`/inference/${entry.slug}`, {
+        lastModified: now,
+        changeFrequency: 'daily' as const,
+        priority: ACTIVE_INFERENCE_MODEL_SLUGS.includes(entry) ? 0.8 : 0.5,
+      }),
+    ),
+    // The catalog index is indexable; the per-point detail pages it links to
+    // stay noindex, so only this page enters the sitemap.
+    ...localizedPair('/inference/agentic', {
+      lastModified: now,
+      changeFrequency: 'daily',
+      priority: 0.7,
+    }),
     ...localizedPair('/compare', { lastModified: now, changeFrequency: 'daily', priority: 0.8 }),
     ...localizedPair('/compare-per-dollar', {
       lastModified: now,
@@ -94,7 +129,36 @@ export default async function sitemap(): Promise<MetadataRoute.Sitemap> {
       changeFrequency: 'daily',
       priority: 0.8,
     }),
-    ...localizedPair('/datasets', { lastModified: now, changeFrequency: 'weekly', priority: 0.6 }),
+    ...localizedPair('/agentx', { lastModified: now, changeFrequency: 'weekly', priority: 0.6 }),
+    ...datasets.flatMap((dataset) =>
+      localizedPair(`/agentx/${dataset.slug}`, {
+        lastModified: now,
+        changeFrequency: 'monthly' as const,
+        priority: 0.5,
+      }),
+    ),
+    ...localizedPair('/agentx/methodology', {
+      lastModified: now,
+      changeFrequency: 'monthly',
+      priority: 0.6,
+    }),
+    ...localizedPair('/agentx/telemetry', {
+      lastModified: now,
+      changeFrequency: 'monthly',
+      priority: 0.6,
+    }),
+    ...localizedPair('/agentx/optimizations', {
+      lastModified: now,
+      changeFrequency: 'monthly',
+      priority: 0.7,
+    }),
+    ...AGENTX_OPTIMIZATION_SLUGS.flatMap((slug) =>
+      localizedPair(`/agentx/optimizations/${slug}`, {
+        lastModified: now,
+        changeFrequency: 'monthly' as const,
+        priority: 0.6,
+      }),
+    ),
     ...localizedPair('/api', { lastModified: now, changeFrequency: 'monthly', priority: 0.7 }),
     ...localizedPair('/blog', { lastModified: now, changeFrequency: 'weekly', priority: 0.8 }),
     ...localizedPair('/glossary', {
@@ -109,6 +173,38 @@ export default async function sitemap(): Promise<MetadataRoute.Sitemap> {
         priority: 0.6,
       }),
     ),
+    ...localizedPair('/chips', {
+      lastModified: now,
+      changeFrequency: 'weekly',
+      priority: 0.8,
+    }),
+    ...getAllChipRouteSlugs().flatMap((slug) =>
+      localizedPair(`/chips/${slug}`, {
+        lastModified: now,
+        changeFrequency: 'weekly' as const,
+        priority: 0.7,
+      }),
+    ),
+    // Live GPU rankings per model (fastest / cheapest). The registry is
+    // static; page bodies render from live benchmark data on request.
+    ...localizedPair('/rankings', { lastModified: now, changeFrequency: 'daily', priority: 0.8 }),
+    ...getAllRankingPageEntries().flatMap((entry) =>
+      localizedPair(`/rankings/${entry.slug}`, {
+        lastModified: now,
+        changeFrequency: 'daily' as const,
+        priority: 0.7,
+      }),
+    ),
+    // Model-on-hardware pages, availability-filtered so only pairings with
+    // benchmark rows enter the sitemap (same crawl-budget rule as /compare).
+    ...localizedPair('/run', { lastModified: now, changeFrequency: 'daily', priority: 0.8 }),
+    ...runEntries.flatMap((entry) =>
+      localizedPair(`/run/${entry.slug}`, {
+        lastModified: now,
+        changeFrequency: 'daily' as const,
+        priority: 0.7,
+      }),
+    ),
     ...getAllPosts().flatMap((post) => {
       const entry = {
         lastModified: new Date(`${post.modifiedDate ?? post.date}T00:00:00Z`).toISOString(),
@@ -119,6 +215,20 @@ export default async function sitemap(): Promise<MetadataRoute.Sitemap> {
       if (!zhPosts.has(post.slug)) return [{ ...entry, url: `${BASE_URL}/blog/${post.slug}` }];
       return localizedPair(`/blog/${post.slug}`, entry);
     }),
+    // Model deep-dive pages (architecture + vendor evals + embedded dashboard).
+    // English-only: no /zh sibling, so no localizedPair.
+    {
+      url: `${BASE_URL}/model`,
+      lastModified: now,
+      changeFrequency: 'weekly' as const,
+      priority: 0.7,
+    },
+    ...getModelPageSlugs().map((slug) => ({
+      url: `${BASE_URL}/model/${slug}`,
+      lastModified: now,
+      changeFrequency: 'weekly' as const,
+      priority: 0.7,
+    })),
     ...compareSlugs.flatMap(({ modelSlug, a, b }) =>
       localizedPair(`/compare/${canonicalCompareSlug(modelSlug, a, b)}`, {
         lastModified: now,
diff --git a/packages/app/src/app/zh/(dashboard)/calculator/[model]/page.tsx b/packages/app/src/app/zh/(dashboard)/calculator/[model]/page.tsx
new file mode 100644
index 000000000..2fb1f688e
--- /dev/null
+++ b/packages/app/src/app/zh/(dashboard)/calculator/[model]/page.tsx
@@ -0,0 +1,54 @@
+import type { Metadata } from 'next';
+import { notFound, permanentRedirect } from 'next/navigation';
+
+import ThroughputCalculatorDisplay from '@/components/calculator/ThroughputCalculatorDisplay';
+import { resolveCalculatorUrlSeed } from '@/components/calculator/url-seed';
+import { ZhTabIntro } from '@/components/zh/zh-tab-intro';
+import { modelRoutePath, pathWithSearchParams, resolveModelRouteSlug } from '@/lib/model-routes';
+import { modelTabCanonicalPath } from '@/lib/tab-meta';
+import { MODEL_TAB_META_ZH, modelTabMetadataZh } from '@/lib/tab-meta-zh';
+
+/** Chinese sibling of `/calculator/<model>` (see that page for the routing
+ *  and client-side model-switch behavior). */
+
+interface Props {
+  params: Promise<{ model: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { model } = await params;
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) return {};
+  return modelTabMetadataZh(
+    'calculator',
+    resolved.route,
+    modelTabCanonicalPath('calculator', resolved.route),
+  );
+}
+
+export default async function ZhCalculatorModelPage({ params, searchParams }: Props) {
+  const [{ model }, sp] = await Promise.all([params, searchParams]);
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) notFound();
+  // Keep share-link params through the 308.
+  if (resolved.isAlias) {
+    permanentRedirect(
+      pathWithSearchParams(`/zh${modelRoutePath('calculator', resolved.route.slug)}`, sp),
+    );
+  }
+  const seed = resolveCalculatorUrlSeed(sp);
+  const meta = MODEL_TAB_META_ZH.calculator;
+  return (
+    <>
+      <ZhTabIntro
+        tab="calculator"
+        title={meta.title(resolved.route.seoName)}
+        intro={meta.intro(resolved.route.seoName)}
+      />
+      <ThroughputCalculatorDisplay
+        urlSeed={{ ...seed, model: seed.model ?? resolved.route.model }}
+      />
+    </>
+  );
+}
diff --git a/packages/app/src/app/zh/(dashboard)/fleet/page.tsx b/packages/app/src/app/zh/(dashboard)/fleet/page.tsx
new file mode 100644
index 000000000..474ba7cd9
--- /dev/null
+++ b/packages/app/src/app/zh/(dashboard)/fleet/page.tsx
@@ -0,0 +1,23 @@
+import type { Metadata } from 'next';
+
+import FleetLifecycleDisplay from '@/components/calculator/FleetLifecycleDisplay';
+import { resolveCalculatorUrlSeed } from '@/components/calculator/url-seed';
+import { ZhTabIntro } from '@/components/zh/zh-tab-intro';
+import { tabMetadataZh } from '@/lib/tab-meta-zh';
+
+export const metadata: Metadata = tabMetadataZh('fleet');
+
+interface Props {
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export default async function ZhFleetPage({ searchParams }: Props) {
+  const sp = await searchParams;
+  const seed = resolveCalculatorUrlSeed(sp);
+  return (
+    <>
+      <ZhTabIntro tab="fleet" />
+      <FleetLifecycleDisplay urlSeed={seed} />
+    </>
+  );
+}
diff --git a/packages/app/src/app/zh/(dashboard)/historical/[model]/page.tsx b/packages/app/src/app/zh/(dashboard)/historical/[model]/page.tsx
new file mode 100644
index 000000000..ce3e20d4f
--- /dev/null
+++ b/packages/app/src/app/zh/(dashboard)/historical/[model]/page.tsx
@@ -0,0 +1,66 @@
+import type { Metadata } from 'next';
+import { notFound, permanentRedirect } from 'next/navigation';
+
+import { InferenceProvider } from '@/components/inference/InferenceContext';
+import HistoricalTrendsDisplay from '@/components/trends/HistoricalTrendsDisplay';
+import { ZhTabIntro } from '@/components/zh/zh-tab-intro';
+import {
+  MODEL_ROUTES,
+  modelRoutePath,
+  pathWithSearchParams,
+  resolveModelRouteSlug,
+} from '@/lib/model-routes';
+import { modelTabCanonicalPath } from '@/lib/tab-meta';
+import { MODEL_TAB_META_ZH, modelTabMetadataZh } from '@/lib/tab-meta-zh';
+
+/** Chinese sibling of `/historical/<model>` (see that page for the routing
+ *  and client-side model-switch behavior). */
+
+interface Props {
+  params: Promise<{ model: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export function generateStaticParams() {
+  return MODEL_ROUTES.map((route) => ({ model: route.slug }));
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { model } = await params;
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) return {};
+  return modelTabMetadataZh(
+    'historical',
+    resolved.route,
+    modelTabCanonicalPath('historical', resolved.route),
+  );
+}
+
+export default async function ZhHistoricalModelPage({ params, searchParams }: Props) {
+  const { model } = await params;
+  const resolved = resolveModelRouteSlug(model);
+  if (!resolved) notFound();
+  // Keep share-link params through the 308; awaited only on this branch so
+  // the canonical slugs stay statically generated.
+  if (resolved.isAlias) {
+    permanentRedirect(
+      pathWithSearchParams(
+        `/zh${modelRoutePath('historical', resolved.route.slug)}`,
+        await searchParams,
+      ),
+    );
+  }
+  const meta = MODEL_TAB_META_ZH.historical;
+  return (
+    <>
+      <ZhTabIntro
+        tab="historical"
+        title={meta.title(resolved.route.seoName)}
+        intro={meta.intro(resolved.route.seoName)}
+      />
+      <InferenceProvider activeTab="historical">
+        <HistoricalTrendsDisplay />
+      </InferenceProvider>
+    </>
+  );
+}
diff --git a/packages/app/src/app/zh/(dashboard)/inference/[model]/page.tsx b/packages/app/src/app/zh/(dashboard)/inference/[model]/page.tsx
new file mode 100644
index 000000000..753c82ba0
--- /dev/null
+++ b/packages/app/src/app/zh/(dashboard)/inference/[model]/page.tsx
@@ -0,0 +1,83 @@
+import type { Metadata } from 'next';
+import { notFound, permanentRedirect } from 'next/navigation';
+
+import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { InferenceProvider } from '@/components/inference/InferenceContext';
+import InferenceChartDisplay from '@/components/inference/ui/ChartDisplay';
+import { ZhTabIntro } from '@/components/zh/zh-tab-intro';
+import { ZH_OG_LOCALE, zhAlternates, zhPath } from '@/lib/i18n';
+import { inferenceModelMetaZh } from '@/lib/inference-model-meta';
+import {
+  getInferenceModelBySlug,
+  INFERENCE_MODEL_SLUGS,
+  inferenceModelPath,
+} from '@/lib/inference-model-slug';
+
+/**
+ * `/zh/inference/<model>` — Chinese sibling of `/inference/<model>`. Same
+ * pinned-model chart body; metadata and intro copy are Simplified Chinese.
+ */
+
+interface Props {
+  params: Promise<{ model: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+export function generateStaticParams(): { model: string }[] {
+  return INFERENCE_MODEL_SLUGS.map((entry) => ({ model: entry.slug }));
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { model } = await params;
+  const entry = getInferenceModelBySlug(model);
+  if (!entry) return {};
+  const { title, description } = inferenceModelMetaZh(entry);
+  const enPath = inferenceModelPath(entry.slug);
+  const url = `${SITE_URL}${zhPath(enPath)}`;
+  return {
+    title: { absolute: `${title} | ${SITE_NAME} by ${AUTHOR_NAME}` },
+    description,
+    alternates: zhAlternates(enPath),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+      type: 'website',
+      locale: ZH_OG_LOCALE,
+    },
+    twitter: {
+      card: 'summary_large_image',
+      title: `${title} | ${SITE_NAME}`,
+      description,
+    },
+  };
+}
+
+export default async function ZhInferenceModelPage({ params, searchParams }: Props) {
+  const { model } = await params;
+  const entry = getInferenceModelBySlug(model);
+  if (!entry) notFound();
+  // Preserves the query string so share-link params like `?i_seq=` and
+  // `?i_prec=` survive the redirect — same treatment as the compare pages.
+  if (model !== entry.slug) {
+    const sp = await searchParams;
+    const qs = Object.entries(sp)
+      .flatMap(([k, v]) => {
+        if (Array.isArray(v)) return v.map((vv) => [k, vv] as const);
+        if (v === undefined) return [];
+        return [[k, v] as const];
+      })
+      .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
+      .join('&');
+    permanentRedirect(`${zhPath(inferenceModelPath(entry.slug))}${qs ? `?${qs}` : ''}`);
+  }
+  return (
+    <>
+      <ZhTabIntro tab="inference" />
+      <InferenceProvider activeTab="inference">
+        <InferenceChartDisplay />
+      </InferenceProvider>
+    </>
+  );
+}
diff --git a/packages/app/src/app/zh/(dashboard)/inference/agentic/page.tsx b/packages/app/src/app/zh/(dashboard)/inference/agentic/page.tsx
new file mode 100644
index 000000000..0acfae29e
--- /dev/null
+++ b/packages/app/src/app/zh/(dashboard)/inference/agentic/page.tsx
@@ -0,0 +1,48 @@
+import type { Metadata } from 'next';
+
+import { AgenticCatalogHero } from '@/components/inference/agentic-catalog/agentic-catalog-hero';
+import { AgenticCatalogList } from '@/components/inference/agentic-catalog/agentic-catalog-list';
+import { JsonLd } from '@/components/json-ld';
+import { getAgenticCatalogGroups } from '@/lib/agentic-catalog';
+import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const TITLE = 'AgentX 遥测数据';
+const DESCRIPTION =
+  '浏览所有存储了逐请求遥测数据的 AgentX 基准测试运行，按模型与推理服务栈分组。每种配置都可打开其数据点详情图表、请求时间线与 cache 行为。';
+
+export const metadata: Metadata = {
+  title: TITLE,
+  description: DESCRIPTION,
+  alternates: zhAlternates('/inference/agentic'),
+  openGraph: {
+    title: `${TITLE} | InferenceX`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/zh/inference/agentic`,
+    locale: ZH_OG_LOCALE,
+  },
+  twitter: { title: `${TITLE} | InferenceX`, description: DESCRIPTION },
+};
+
+export const dynamic = 'force-dynamic';
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'CollectionPage',
+  name: `${TITLE} | InferenceX`,
+  description: DESCRIPTION,
+  url: `${SITE_URL}/zh/inference/agentic`,
+  inLanguage: 'zh-CN',
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default async function AgenticTelemetryCatalogPageZh() {
+  const groups = await getAgenticCatalogGroups();
+  return (
+    <>
+      <JsonLd data={jsonLd} />
+      <AgenticCatalogHero locale="zh" />
+      <AgenticCatalogList groups={groups} locale="zh" />
+    </>
+  );
+}
diff --git a/packages/app/src/app/zh/(dashboard)/inference/logs/[id]/page.tsx b/packages/app/src/app/zh/(dashboard)/inference/logs/[id]/page.tsx
new file mode 100644
index 000000000..fbc7f14c3
--- /dev/null
+++ b/packages/app/src/app/zh/(dashboard)/inference/logs/[id]/page.tsx
@@ -0,0 +1,21 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { BenchmarkLogDetail } from '@/components/inference/log-viewer/benchmark-log-detail';
+import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
+
+export const metadata: Metadata = {
+  title: '基准测试日志 | InferenceX',
+  robots: { index: false },
+};
+
+export default async function ZhBenchmarkLogPage({
+  params,
+}: {
+  params: Promise<{ id: string }>;
+}) {
+  const { id } = await params;
+  const numericId = Number(id);
+  if (!isPersistedBenchmarkId(numericId)) notFound();
+  return <BenchmarkLogDetail id={numericId} />;
+}
diff --git a/packages/app/src/app/zh/about/page.tsx b/packages/app/src/app/zh/about/page.tsx
index 91209d257..e34141ffb 100644
--- a/packages/app/src/app/zh/about/page.tsx
+++ b/packages/app/src/app/zh/about/page.tsx
@@ -4,40 +4,30 @@ import Link from 'next/link';
 import { Card } from '@/components/ui/card';
 import { FAQ_ITEMS_ZH } from '@/components/about/faq-data-zh';
 import { AgentXFaq } from '@/components/about/agentx-faq';
+import { buildFaqJsonLd, FaqList } from '@/components/about/faq';
+import { PromoVideo } from '@/components/about/promo-video';
 import { JsonLd } from '@/components/json-ld';
 import { zhAlternates, ZH_OG_LOCALE, ZH_LANG_TAG } from '@/lib/i18n';
 import { GITHUB_OWNER, GITHUB_REPO, SITE_URL } from '@semianalysisai/inferencex-constants';
 
-const faqJsonLd = {
-  '@context': 'https://schema.org',
-  '@type': 'FAQPage',
-  inLanguage: ZH_LANG_TAG,
-  mainEntity: FAQ_ITEMS_ZH.map((item) => ({
-    '@type': 'Question',
-    name: item.question,
-    acceptedAnswer: {
-      '@type': 'Answer',
-      text: [item.answer, item.link?.text, ...(item.list ?? [])].filter(Boolean).join(' '),
-    },
-  })),
-};
+const faqJsonLd = buildFaqJsonLd(FAQ_ITEMS_ZH, ZH_LANG_TAG);
+
+const ABOUT_DESCRIPTION =
+  'InferenceX 对比各类加速器与服务栈的智能体推理和固定序列 AI 推理性能。AgentX 是其长上下文多轮编码场景。';
 
 export const metadata: Metadata = {
   title: '关于',
-  description:
-    'InferenceX 是一个独立、厂商中立、可复现的基准测试平台，持续测试各类 AI 加速器上的推理软件性能。',
+  description: ABOUT_DESCRIPTION,
   alternates: zhAlternates('/about'),
   openGraph: {
     title: '关于 | InferenceX',
-    description:
-      'InferenceX 是一个独立、厂商中立、可复现的基准测试平台，持续测试各类 AI 加速器上的推理软件性能。',
+    description: ABOUT_DESCRIPTION,
     url: `${SITE_URL}/zh/about`,
     locale: ZH_OG_LOCALE,
   },
   twitter: {
     title: '关于 | InferenceX',
-    description:
-      'InferenceX 是一个独立、厂商中立、可复现的基准测试平台，持续测试各类 AI 加速器上的推理软件性能。',
+    description: ABOUT_DESCRIPTION,
   },
 };
 
@@ -46,10 +36,14 @@ export default function AboutPageZh() {
     <main className="relative">
       <JsonLd data={faqJsonLd} />
       <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
+        <section>
+          <PromoVideo />
+        </section>
+
         <section>
           <Card>
             <h2 className="text-lg font-semibold mb-2">
-              开源持续推理基准测试——受万亿美元级吉瓦规模 Token 工厂运营者的信赖
+              开源持续智能体推理基准测试——受万亿美元级吉瓦规模 Token 工厂运营者的信赖
             </h2>
             <p className="text-muted-foreground mb-2">
               随着世界以指数级速度迈向
@@ -57,8 +51,8 @@ export default function AboutPageZh() {
             </p>
             <p className="text-muted-foreground mb-2">
               <strong>InferenceX&trade;</strong>（原名
-              InferenceMAX）是我们独立、厂商中立、可复现的基准测试平台，通过持续测试实际可用于 ML
-              社区的各类 AI 加速器上的推理软件来解决这些问题。
+              InferenceMAX）是我们独立、厂商中立、可复现的基准测试平台。它测试固定序列推理服务和
+              AgentX 长上下文多轮智能体编码工作负载，覆盖 ML 社区实际可用的各类 AI 加速器与服务栈。
             </p>
             <p className="text-muted-foreground">
               我们的开放数据与洞察已被 ML 社区广泛采用，包括万亿美元级 Token 工厂和 AI
@@ -66,16 +60,18 @@ export default function AboutPageZh() {
               NeoCloud。了解更多详情请阅读我们的文章：{' '}
               <Link
                 href="/blog/inferencemax-open-source-inference-benchmarking"
+                hrefLang="en"
                 className="text-brand hover:underline font-medium"
               >
-                InferenceX v1
+                InferenceX v1（英文文章）
               </Link>
               、{' '}
               <Link
                 href="/blog/inferencex-v2-nvidia-blackwell-vs-amd-vs-hopper"
+                hrefLang="en"
                 className="text-brand hover:underline font-medium"
               >
-                InferenceX v2
+                InferenceX v2（英文文章）
               </Link>
               。
             </p>
@@ -118,7 +114,7 @@ export default function AboutPageZh() {
                   3
                 </span>
                 <div>
-                  <strong className="text-foreground">上传产物。</strong> 请求延迟、token 计数、Chip
+                  <strong className="text-foreground">上传产物。</strong> 请求延迟、token 计数、芯片
                   功耗遥测数据和评估样本均附加到运行页面。GitHub Actions 保留这些产物 90
                   天，同时每周发布完整基准测试数据库的快照作为公开的 GitHub
                   Release，以实现更长期的可审计性。
@@ -173,45 +169,7 @@ export default function AboutPageZh() {
           </Card>
         </section>
 
-        <section>
-          <Card>
-            <h2 className="text-lg font-semibold mb-4">常见问题</h2>
-            <dl className="divide-y divide-border">
-              {FAQ_ITEMS_ZH.map((item) => (
-                <div key={item.question} className="py-4 first:pt-0 last:pb-0">
-                  <dt className="font-medium mb-1">{item.question}</dt>
-                  <dd className="text-muted-foreground text-sm">
-                    {item.answer && (
-                      <p>
-                        {item.answer}
-                        {item.link && (
-                          <>
-                            {' '}
-                            <a
-                              href={item.link.href}
-                              target="_blank"
-                              rel="noopener noreferrer"
-                              className="text-brand hover:underline font-medium"
-                            >
-                              {item.link.text}
-                            </a>
-                          </>
-                        )}
-                      </p>
-                    )}
-                    {item.list && (
-                      <ul className="mt-1.5 ml-8 list-disc space-y-0.5">
-                        {item.list.map((li) => (
-                          <li key={li}>{li}</li>
-                        ))}
-                      </ul>
-                    )}
-                  </dd>
-                </div>
-              ))}
-            </dl>
-          </Card>
-        </section>
+        <FaqList title="常见问题" items={FAQ_ITEMS_ZH} />
       </div>
     </main>
   );
diff --git a/packages/app/src/app/zh/datasets/[slug]/conversations/[convId]/page.tsx b/packages/app/src/app/zh/agentx/[slug]/conversations/[convId]/page.tsx
similarity index 91%
rename from packages/app/src/app/zh/datasets/[slug]/conversations/[convId]/page.tsx
rename to packages/app/src/app/zh/agentx/[slug]/conversations/[convId]/page.tsx
index cbb99dec8..8704e889c 100644
--- a/packages/app/src/app/zh/datasets/[slug]/conversations/[convId]/page.tsx
+++ b/packages/app/src/app/zh/agentx/[slug]/conversations/[convId]/page.tsx
@@ -17,7 +17,7 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     title,
     description,
     alternates: {
-      canonical: `${SITE_URL}/zh/datasets/${slug}/conversations/${encodeURIComponent(convId)}`,
+      canonical: `${SITE_URL}/zh/agentx/${slug}/conversations/${encodeURIComponent(convId)}`,
     },
     robots: { index: false },
   };
diff --git a/packages/app/src/app/zh/datasets/[slug]/page.tsx b/packages/app/src/app/zh/agentx/[slug]/page.tsx
similarity index 87%
rename from packages/app/src/app/zh/datasets/[slug]/page.tsx
rename to packages/app/src/app/zh/agentx/[slug]/page.tsx
index a2aedc092..5291637f9 100644
--- a/packages/app/src/app/zh/datasets/[slug]/page.tsx
+++ b/packages/app/src/app/zh/agentx/[slug]/page.tsx
@@ -10,16 +10,16 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
-  const title = `${slug} | Agentic 数据集`;
+  const title = `${slug} | AgentX 数据集`;
   const description = `${slug} agentic trace 数据集的分布、token 统计及逐对话火焰图。`;
   return {
     title,
     description,
-    alternates: zhAlternates(`/datasets/${slug}`),
+    alternates: zhAlternates(`/agentx/${slug}`),
     openGraph: {
       title: `${title} | InferenceX`,
       description,
-      url: `${SITE_URL}/zh/datasets/${slug}`,
+      url: `${SITE_URL}/zh/agentx/${slug}`,
       locale: ZH_OG_LOCALE,
     },
     twitter: { title: `${title} | InferenceX`, description },
diff --git a/packages/app/src/app/zh/agentx/methodology/page.tsx b/packages/app/src/app/zh/agentx/methodology/page.tsx
new file mode 100644
index 000000000..1cb63707f
--- /dev/null
+++ b/packages/app/src/app/zh/agentx/methodology/page.tsx
@@ -0,0 +1,43 @@
+import type { Metadata } from 'next';
+
+import { AgentXMethodologyArticle } from '@/components/datasets/agentx-methodology-article';
+import { JsonLd } from '@/components/json-ld';
+import { zhAlternates, ZH_LANG_TAG, ZH_OG_LOCALE } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const DESCRIPTION =
+  'AgentX 如何将自愿采集的编码智能体 trace 转换为回放图，并统一 KV cache warmup 与基准测试配置。';
+
+export const metadata: Metadata = {
+  title: 'AgentX 方法论',
+  description: DESCRIPTION,
+  alternates: zhAlternates('/agentx/methodology'),
+  openGraph: {
+    title: 'AgentX 方法论 | InferenceX',
+    description: DESCRIPTION,
+    url: `${SITE_URL}/zh/agentx/methodology`,
+    locale: ZH_OG_LOCALE,
+  },
+  twitter: { title: 'AgentX 方法论 | InferenceX', description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'TechArticle',
+  headline: 'AgentX 方法论',
+  description: DESCRIPTION,
+  url: `${SITE_URL}/zh/agentx/methodology`,
+  inLanguage: ZH_LANG_TAG,
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default function AgentXMethodologyPageZh() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXMethodologyArticle locale="zh" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/zh/agentx/optimizations/[framework]/page.tsx b/packages/app/src/app/zh/agentx/optimizations/[framework]/page.tsx
new file mode 100644
index 000000000..7c8d64da3
--- /dev/null
+++ b/packages/app/src/app/zh/agentx/optimizations/[framework]/page.tsx
@@ -0,0 +1,68 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { AgentXOptimizationsArticle } from '@/components/datasets/agentx-optimizations-article';
+import { JsonLd } from '@/components/json-ld';
+import {
+  AGENTX_OPTIMIZATION_FRAMEWORKS,
+  getOptimizationFramework,
+} from '@/lib/agentx-optimizations';
+import { getLocalizedFramework } from '@/lib/agentx-optimizations-zh';
+import { zhAlternates, ZH_LANG_TAG, ZH_OG_LOCALE } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+interface Props {
+  params: Promise<{ framework: string }>;
+}
+
+export function generateStaticParams() {
+  return AGENTX_OPTIMIZATION_FRAMEWORKS.map((entry) => ({ framework: entry.slug }));
+}
+
+export const dynamicParams = false;
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { framework: slug } = await params;
+  const framework = getOptimizationFramework(slug);
+  if (!framework) return {};
+  const localized = getLocalizedFramework(framework, 'zh');
+  const title = `${framework.name} 智能体负载优化`;
+  return {
+    title,
+    description: localized.summary,
+    alternates: zhAlternates(`/agentx/optimizations/${slug}`),
+    openGraph: {
+      title: `${title} | InferenceX`,
+      description: localized.summary,
+      url: `${SITE_URL}/zh/agentx/optimizations/${slug}`,
+      locale: ZH_OG_LOCALE,
+    },
+    twitter: { title: `${title} | InferenceX`, description: localized.summary },
+  };
+}
+
+export default async function AgentXOptimizationsFrameworkPageZh({ params }: Props) {
+  const { framework: slug } = await params;
+  const framework = getOptimizationFramework(slug);
+  if (!framework) notFound();
+
+  const localized = getLocalizedFramework(framework, 'zh');
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'TechArticle',
+    headline: `${framework.name} 智能体负载优化`,
+    description: localized.summary,
+    url: `${SITE_URL}/zh/agentx/optimizations/${slug}`,
+    inLanguage: ZH_LANG_TAG,
+    isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXOptimizationsArticle slug={slug} locale="zh" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/zh/agentx/optimizations/page.tsx b/packages/app/src/app/zh/agentx/optimizations/page.tsx
new file mode 100644
index 000000000..e57a5b159
--- /dev/null
+++ b/packages/app/src/app/zh/agentx/optimizations/page.tsx
@@ -0,0 +1,44 @@
+import type { Metadata } from 'next';
+
+import { AgentXOptimizationsIndex } from '@/components/datasets/agentx-optimizations-article';
+import { JsonLd } from '@/components/json-ld';
+import { zhAlternates, ZH_LANG_TAG, ZH_OG_LOCALE } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const TITLE = 'AgentX 行业影响：面向智能体负载的优化';
+const DESCRIPTION =
+  'AgentX 推动 vLLM、SGLang、TensorRT-LLM、ATOM、AITER、Dynamo、LMCache 与 Mooncake 提交的 50 多个上游 PR。';
+
+export const metadata: Metadata = {
+  title: 'AgentX 行业影响',
+  description: DESCRIPTION,
+  alternates: zhAlternates('/agentx/optimizations'),
+  openGraph: {
+    title: `${TITLE} | InferenceX`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/zh/agentx/optimizations`,
+    locale: ZH_OG_LOCALE,
+  },
+  twitter: { title: `${TITLE} | InferenceX`, description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'TechArticle',
+  headline: TITLE,
+  description: DESCRIPTION,
+  url: `${SITE_URL}/zh/agentx/optimizations`,
+  inLanguage: ZH_LANG_TAG,
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default function AgentXOptimizationsPageZh() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXOptimizationsIndex locale="zh" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/zh/agentx/page.tsx b/packages/app/src/app/zh/agentx/page.tsx
new file mode 100644
index 000000000..40d446d74
--- /dev/null
+++ b/packages/app/src/app/zh/agentx/page.tsx
@@ -0,0 +1,50 @@
+import type { Metadata } from 'next';
+
+import { AgentXMethodology } from '@/components/datasets/agentx-methodology';
+import { DatasetList } from '@/components/datasets/dataset-list';
+import { JsonLd } from '@/components/json-ld';
+import { zhAlternates, ZH_LANG_TAG, ZH_OG_LOCALE } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const DESCRIPTION =
+  'AgentX 根据参与者自愿提供的 Claude Code 会话构建回放工作负载。可查看测试方法、数据分布和单会话火焰图。';
+
+export const metadata: Metadata = {
+  title: 'AgentX 测试方法与数据集',
+  description: DESCRIPTION,
+  alternates: zhAlternates('/agentx'),
+  openGraph: {
+    title: 'AgentX 测试方法与数据集 | InferenceX',
+    description: DESCRIPTION,
+    url: `${SITE_URL}/zh/agentx`,
+    locale: ZH_OG_LOCALE,
+  },
+  twitter: { title: 'AgentX 测试方法与数据集 | InferenceX', description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'CollectionPage',
+  name: 'InferenceX AgentX 数据集',
+  description: DESCRIPTION,
+  url: `${SITE_URL}/zh/agentx`,
+  inLanguage: ZH_LANG_TAG,
+};
+
+export default function AgentXPageZh() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto flex flex-col gap-6 px-4 pb-8 lg:px-8">
+        <section>
+          <AgentXMethodology locale="zh" />
+        </section>
+
+        <section className="flex flex-col gap-3">
+          <h2 className="text-lg font-semibold text-foreground">数据集</h2>
+          <DatasetList />
+        </section>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/zh/agentx/telemetry/page.tsx b/packages/app/src/app/zh/agentx/telemetry/page.tsx
new file mode 100644
index 000000000..0af65a476
--- /dev/null
+++ b/packages/app/src/app/zh/agentx/telemetry/page.tsx
@@ -0,0 +1,46 @@
+import type { Metadata } from 'next';
+
+import { AgentXTelemetryArticle } from '@/components/datasets/agentx-telemetry-article';
+import { JsonLd } from '@/components/json-ld';
+import { getTelemetryGuide } from '@/lib/agentx-telemetry-zh';
+import { zhAlternates, ZH_LANG_TAG, ZH_OG_LOCALE } from '@/lib/i18n';
+import { SITE_URL } from '@semianalysisai/inferencex-constants';
+
+const TITLE = '通过详细遥测数据解析智能体负载';
+const DESCRIPTION =
+  '如何解读 AgentX 单个数据点背后的遥测数据：单点图表、请求时间线、KV offload 标记，以及单会话火焰图。';
+
+export const metadata: Metadata = {
+  title: 'AgentX 遥测数据教程',
+  description: DESCRIPTION,
+  alternates: zhAlternates('/agentx/telemetry'),
+  openGraph: {
+    title: `${TITLE} | InferenceX`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/zh/agentx/telemetry`,
+    locale: ZH_OG_LOCALE,
+  },
+  twitter: { title: `${TITLE} | InferenceX`, description: DESCRIPTION },
+};
+
+const jsonLd = {
+  '@context': 'https://schema.org',
+  '@type': 'TechArticle',
+  headline: TITLE,
+  description: DESCRIPTION,
+  url: `${SITE_URL}/zh/agentx/telemetry`,
+  inLanguage: ZH_LANG_TAG,
+  about: getTelemetryGuide('zh').title,
+  isPartOf: { '@type': 'WebSite', name: 'InferenceX', url: SITE_URL },
+};
+
+export default function AgentXTelemetryPageZh() {
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto max-w-6xl px-4 pb-12 lg:px-8">
+        <AgentXTelemetryArticle locale="zh" />
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/zh/blog/[slug]/page.tsx b/packages/app/src/app/zh/blog/[slug]/page.tsx
index 071d48b04..5b6deabe6 100644
--- a/packages/app/src/app/zh/blog/[slug]/page.tsx
+++ b/packages/app/src/app/zh/blog/[slug]/page.tsx
@@ -1,13 +1,6 @@
 import type { Metadata } from 'next';
 import Link from 'next/link';
 import { notFound } from 'next/navigation';
-import { compileMDX } from 'next-mdx-remote/rsc';
-import rehypeShikiFromHighlighter from '@shikijs/rehype/core';
-import rehypeKatex from 'rehype-katex';
-import remarkGfm from 'remark-gfm';
-import remarkMath from 'remark-math';
-import { createHighlighterCore } from 'shiki/core';
-import { createOnigurumaEngine } from 'shiki/engine/oniguruma';
 
 import 'katex/dist/katex.min.css';
 
@@ -15,7 +8,6 @@ import { BlogBackLink } from '@/components/blog/blog-back-link';
 import { BlogPostNav } from '@/components/blog/blog-post-nav';
 import { BlogToc } from '@/components/blog/blog-toc';
 import { HashScroll } from '@/components/blog/hash-scroll';
-import { createMdxComponents } from '@/components/blog/mdx-components';
 import { ReadingProgressBar } from '@/components/blog/reading-progress-bar';
 import { ShareTwitterButton, ShareLinkedInButton } from '@/components/share-buttons';
 import { Card } from '@/components/ui/card';
@@ -29,6 +21,7 @@ import {
   extractHeadings,
   getPostBySlug,
 } from '@/lib/blog';
+import { compileBlogMdx } from '@/lib/blog-mdx';
 import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 import {
   AUTHOR_HANDLE,
@@ -81,33 +74,6 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
   };
 }
 
-let highlighterPromise: ReturnType<typeof createHighlighterCore> | null = null;
-
-function getHighlighter() {
-  if (!highlighterPromise) {
-    highlighterPromise = createHighlighterCore({
-      themes: [import('shiki/themes/github-dark.mjs'), import('shiki/themes/github-light.mjs')],
-      langs: [
-        import('shiki/langs/typescript.mjs'),
-        import('shiki/langs/javascript.mjs'),
-        import('shiki/langs/python.mjs'),
-        import('shiki/langs/bash.mjs'),
-        import('shiki/langs/json.mjs'),
-        import('shiki/langs/yaml.mjs'),
-        import('shiki/langs/css.mjs'),
-        import('shiki/langs/html.mjs'),
-        import('shiki/langs/tsx.mjs'),
-        import('shiki/langs/jsx.mjs'),
-        import('shiki/langs/sql.mjs'),
-        import('shiki/langs/go.mjs'),
-        import('shiki/langs/rust.mjs'),
-      ],
-      engine: createOnigurumaEngine(import('shiki/wasm')),
-    });
-  }
-  return highlighterPromise;
-}
-
 export default async function ZhBlogPostPage({ params }: Props) {
   const { slug } = await params;
   const result = getPostBySlug(slug, 'zh');
@@ -116,32 +82,7 @@ export default async function ZhBlogPostPage({ params }: Props) {
   const { meta, raw } = result;
   const adjacent = getAdjacentPosts(slug, 'zh');
   const headings = extractHeadings(raw);
-  const highlighter = await getHighlighter();
-
-  const { content } = await compileMDX({
-    source: raw,
-    components: createMdxComponents('zh'),
-    options: {
-      mdxOptions: {
-        // `singleDollarTextMath: false` keeps prices like $1.95/GPU/hr from being parsed
-        // as inline math; math must be delimited with `$$`.
-        remarkPlugins: [remarkGfm, [remarkMath, { singleDollarTextMath: false }]],
-        rehypePlugins: [
-          // `strict: false` keeps a single malformed expression from failing the whole build;
-          // KaTeX renders it inline in red instead.
-          [rehypeKatex, { strict: false }],
-          [
-            rehypeShikiFromHighlighter,
-            highlighter,
-            {
-              themes: { dark: 'github-dark', light: 'github-light' },
-              defaultColor: false,
-            },
-          ],
-        ],
-      },
-    },
-  });
+  const { content } = await compileBlogMdx(raw, 'zh');
 
   const jsonLd = buildBlogPostingJsonLd(meta, raw, 'zh');
   const breadcrumbJsonLd = buildBlogBreadcrumbJsonLdZh(slug, meta.title);
diff --git a/packages/app/src/app/zh/blog/page.tsx b/packages/app/src/app/zh/blog/page.tsx
index ef6b77906..4e040d32b 100644
--- a/packages/app/src/app/zh/blog/page.tsx
+++ b/packages/app/src/app/zh/blog/page.tsx
@@ -11,11 +11,11 @@ import { SITE_URL, SITE_NAME, AUTHOR_NAME } from '@semianalysisai/inferencex-con
 
 export const metadata: Metadata = {
   title: '文章',
-  description: `${SITE_NAME} by ${AUTHOR_NAME} 的技术文章——AI 推理基准测试、Chip 性能分析与 ML 基础设施洞见。`,
+  description: `${SITE_NAME} by ${AUTHOR_NAME} 发布关于智能体推理基准测试、AgentX 结果、芯片性能与 ML 基础设施的技术文章。`,
   alternates: zhAlternates('/blog'),
   openGraph: {
     title: `文章 | ${SITE_NAME} by ${AUTHOR_NAME}`,
-    description: 'AI 推理基准测试洞见与 Chip 性能分析。',
+    description: '关于智能体推理基准测试、AgentX 结果与芯片性能的文章。',
     url: `${SITE_URL}/zh/blog`,
     locale: ZH_OG_LOCALE,
   },
@@ -51,7 +51,7 @@ export default async function ZhBlogPage({
           <Card>
             <h2 className="text-2xl lg:text-4xl font-bold tracking-tight">文章</h2>
             <p className="mt-3 text-base lg:text-lg text-muted-foreground">
-              关于 AI 推理基准测试、Chip 性能与 ML 基础设施的深度洞见。
+              关于智能体推理、AgentX 结果、芯片性能与 ML 基础设施的文章。
             </p>
             <p className="text-sm text-muted-foreground">
               不熟悉相关概念？{' '}
diff --git a/packages/app/src/app/zh/chips/[slug]/page.tsx b/packages/app/src/app/zh/chips/[slug]/page.tsx
new file mode 100644
index 000000000..6c0d40019
--- /dev/null
+++ b/packages/app/src/app/zh/chips/[slug]/page.tsx
@@ -0,0 +1,118 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { ChipDetailContent, ChipVsContent } from '@/components/chips/chip-page-sections';
+import {
+  getAllChipRouteSlugs,
+  getChipPage,
+  getChipVsPage,
+  type ChipVsPage,
+} from '@/lib/chip-pages';
+import { getZhChipTranslation } from '@/lib/chip-pages-zh';
+import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import {
+  AUTHOR_HANDLE,
+  AUTHOR_NAME,
+  SITE_NAME,
+  SITE_URL,
+  SUPPORTERS_LINE_ZH,
+} from '@semianalysisai/inferencex-constants';
+
+interface Props {
+  params: Promise<{ slug: string }>;
+}
+
+export const dynamicParams = false;
+
+export function generateStaticParams() {
+  return getAllChipRouteSlugs().map((slug) => ({ slug }));
+}
+
+function vsDescriptionZh(page: ChipVsPage): string {
+  return `${page.a.title} 对比 ${page.b.title}：显存、内存带宽、FP8/FP4 算力、TDP 与云端价格并列呈现，并附相同工作负载上持续测量的 LLM 推理基准测试。${SUPPORTERS_LINE_ZH}`;
+}
+
+function vsKeywordsZh(page: ChipVsPage): string[] {
+  const a = page.a.label;
+  const b = page.b.label;
+  return [
+    `${a} vs ${b}`,
+    `${b} vs ${a}`,
+    `${a} 对比 ${b}`,
+    `${a} vs ${b} 基准测试`,
+    `${a} vs ${b} 规格`,
+    `${a} vs ${b} 价格`,
+    'AI 芯片对比',
+    'AI 推理基准测试',
+  ];
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug } = await params;
+
+  const entry = getChipPage(slug);
+  if (entry) {
+    const translation = getZhChipTranslation(entry.slug);
+    if (!translation) return {};
+    const title = `${entry.title} 规格、价格与 AI 推理基准测试`;
+    const description = `${translation.summary}${SUPPORTERS_LINE_ZH}`;
+    const url = `${SITE_URL}/zh/chips/${entry.slug}`;
+    return {
+      title,
+      description,
+      keywords: [...translation.keywords],
+      authors: [{ name: AUTHOR_NAME }],
+      alternates: zhAlternates(`/chips/${entry.slug}`),
+      openGraph: {
+        title: `${title} | ${SITE_NAME}`,
+        description,
+        url,
+        locale: ZH_OG_LOCALE,
+        type: 'article',
+      },
+      twitter: {
+        card: 'summary_large_image',
+        title: `${title} | ${SITE_NAME}`,
+        description,
+        creator: AUTHOR_HANDLE,
+      },
+    };
+  }
+
+  const vsPage = getChipVsPage(slug);
+  if (!vsPage) return {};
+  const title = `${vsPage.a.label} vs ${vsPage.b.label}：规格、功耗与 AI 推理对比`;
+  const description = vsDescriptionZh(vsPage);
+  const url = `${SITE_URL}/zh/chips/${vsPage.slug}`;
+  return {
+    title,
+    description,
+    keywords: vsKeywordsZh(vsPage),
+    authors: [{ name: AUTHOR_NAME }],
+    alternates: zhAlternates(`/chips/${vsPage.slug}`),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+      locale: ZH_OG_LOCALE,
+      type: 'article',
+    },
+    twitter: {
+      card: 'summary_large_image',
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      creator: AUTHOR_HANDLE,
+    },
+  };
+}
+
+export default async function ZhChipRoutePage({ params }: Props) {
+  const { slug } = await params;
+
+  const entry = getChipPage(slug);
+  if (entry) return <ChipDetailContent entry={entry} locale="zh" />;
+
+  const vsPage = getChipVsPage(slug);
+  if (!vsPage) notFound();
+  return <ChipVsContent page={vsPage} locale="zh" />;
+}
diff --git a/packages/app/src/app/zh/chips/page.tsx b/packages/app/src/app/zh/chips/page.tsx
new file mode 100644
index 000000000..ca34231d6
--- /dev/null
+++ b/packages/app/src/app/zh/chips/page.tsx
@@ -0,0 +1,40 @@
+import type { Metadata } from 'next';
+
+import { ChipsIndexContent } from '@/components/chips/chip-page-sections';
+import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { SITE_NAME, SITE_URL, SUPPORTERS_LINE_ZH } from '@semianalysisai/inferencex-constants';
+
+const TITLE = '面向 LLM 推理的 AI 芯片：规格、价格与基准测试';
+const DESCRIPTION = `NVIDIA Hopper 与 Blackwell（H100、H200、B200、B300、GB200 NVL72、GB300 NVL72）以及 AMD Instinct（MI300X、MI325X、MI355X）芯片的规格、云端价格与持续测量的 LLM 推理基准测试。${SUPPORTERS_LINE_ZH}`;
+
+export const metadata: Metadata = {
+  title: TITLE,
+  description: DESCRIPTION,
+  keywords: [
+    'AI 芯片',
+    'AI 芯片对比',
+    'GPU 规格',
+    'GPU 每小时价格',
+    'LLM 推理硬件',
+    'NVIDIA 对比 AMD',
+    '数据中心芯片价格',
+    'AI 加速器对比',
+  ],
+  alternates: zhAlternates('/chips'),
+  openGraph: {
+    title: `${TITLE} | ${SITE_NAME}`,
+    description: DESCRIPTION,
+    url: `${SITE_URL}/zh/chips`,
+    locale: ZH_OG_LOCALE,
+    type: 'website',
+  },
+  twitter: {
+    card: 'summary_large_image',
+    title: `${TITLE} | ${SITE_NAME}`,
+    description: DESCRIPTION,
+  },
+};
+
+export default function ZhChipsIndexPage() {
+  return <ChipsIndexContent locale="zh" />;
+}
diff --git a/packages/app/src/app/zh/compare-per-dollar/[slug]/[scenario]/page.tsx b/packages/app/src/app/zh/compare-per-dollar/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..0e52e1f09
--- /dev/null
+++ b/packages/app/src/app/zh/compare-per-dollar/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildPerDollarMetadataZh, renderPerDollarPageZh } from '../page';
+
+/**
+ * `/zh/compare-per-dollar/<slug>/<scenario>` — the same comparison with its workload pinned by
+ * the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments rather than silently falling back to the pair's
+ *  default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildPerDollarMetadataZh(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function PerDollarScenarioPageZh({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderPerDollarPageZh(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/zh/compare-per-dollar/[slug]/opengraph-image.tsx b/packages/app/src/app/zh/compare-per-dollar/[slug]/opengraph-image.tsx
index 0f1b62f36..89cf95a67 100644
--- a/packages/app/src/app/zh/compare-per-dollar/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/zh/compare-per-dollar/[slug]/opengraph-image.tsx
@@ -1,7 +1,36 @@
-export {
-  default,
-  alt,
-  size,
-  contentType,
-  generateStaticParams,
-} from '../../../compare-per-dollar/[slug]/opengraph-image';
+import { notFound } from 'next/navigation';
+
+import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
+
+import { getAllComparableCompareSlugs } from '@/lib/compare-availability';
+import { renderCompareOg } from '@/lib/compare-og';
+import { canonicalCompareSlug, parseCompareSlug } from '@/lib/compare-slug';
+
+export const alt = '芯片每美元推理性能基准测试对比';
+export const size = { width: 1200, height: 630 };
+export const contentType = 'image/png';
+
+export async function generateStaticParams() {
+  const slugs = await getAllComparableCompareSlugs();
+  return slugs.map(({ modelSlug, a, b }) => ({ slug: canonicalCompareSlug(modelSlug, a, b) }));
+}
+
+export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
+  const { slug } = await params;
+  const parsed = parseCompareSlug(slug);
+  if (!parsed) notFound();
+
+  const aLabel = HW_REGISTRY[parsed.a]?.label ?? parsed.a.toUpperCase();
+  const bLabel = HW_REGISTRY[parsed.b]?.label ?? parsed.b.toUpperCase();
+  const title = `${aLabel} 对比 ${bLabel}`;
+  const titleSize = title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
+
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · 每美元性能对比`,
+    title,
+    titleSize,
+    footer: 'AI 推理基准测试 · 每百万 token 成本',
+    language: 'zh',
+    size,
+  });
+}
diff --git a/packages/app/src/app/zh/compare-per-dollar/[slug]/page.tsx b/packages/app/src/app/zh/compare-per-dollar/[slug]/page.tsx
index aa42dadc6..e6a2067cb 100644
--- a/packages/app/src/app/zh/compare-per-dollar/[slug]/page.tsx
+++ b/packages/app/src/app/zh/compare-per-dollar/[slug]/page.tsx
@@ -9,7 +9,11 @@ import {
 } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
-import { pickPairDefaults } from '@/lib/compare-pair-defaults';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
 import {
   canonicalCompareSlug,
   compareDisplayLabel,
@@ -17,17 +21,13 @@ import {
   parseCompareSlug,
 } from '@/lib/compare-slug';
 import { getGpuSpecs } from '@/lib/constants';
+import { KNOWN_MODELS, KNOWN_PRECISIONS, KNOWN_SEQUENCES, pickString } from '@/lib/compare-ssr';
 import {
-  computeCompareTableData,
-  dateRangeForPair,
-  getCachedBenchmarks,
-  KNOWN_MODELS,
-  KNOWN_PRECISIONS,
-  KNOWN_SEQUENCES,
-  pickString,
-  summarize,
-} from '@/lib/compare-ssr';
+  getComparePageDerivedData,
+  initialCompareBenchmarkRows,
+} from '@/lib/compare-page-data.server';
 import {
+  AGENTIC_SCENARIO_INTRO_ZH,
   buildBreadcrumbJsonLdZh,
   buildJsonLdZh,
   compareTableNarrativeZh,
@@ -38,6 +38,30 @@ import ComparePerDollarPageClient from '../../../compare-per-dollar/[slug]/page-
 
 export const dynamic = 'force-dynamic';
 
+/**
+ * `/zh/compare-per-dollar/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than duplicated
+ * in a parallel route file so the body — redirects, JSON-LD, metadata, client
+ * props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
+/** English twin of `scenarioPath` — hreflang pairs are keyed off the
+ *  English route, so the segment has to survive the locale swap. */
+function enScenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/compare-per-dollar/${canonical}/${scenarioSegment}`
+    : `/compare-per-dollar/${canonical}`;
+}
+
+function scenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/zh/compare-per-dollar/${canonical}/${scenarioSegment}`
+    : `/zh/compare-per-dollar/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
@@ -45,17 +69,29 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildPerDollarMetadataZh(slug, {});
+}
+
+export function buildPerDollarMetadataZh(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Metadata {
   const parsed = parseCompareSlug(slug);
   if (!parsed) return {};
   const fullLabel = compareModelDisplayLabel(parsed.model, parsed.a, parsed.b);
   const gpuLabel = compareDisplayLabel(parsed.a, parsed.b);
   const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
-  const url = `${SITE_URL}/zh/compare-per-dollar/${canonical}`;
-  const description = `${gpuLabel} 在 ${parsed.model.label} 上的每美元性能：来自 InferenceX（SemiAnalysis 推出的独立开源基准测试平台）的经验证、可复现的每百万 token 成本结果，基于云服务商 TCO 归一化。${SUPPORTERS_LINE_ZH}查看哪款 Chip 在各交互性水平下更经济。`;
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const url = `${SITE_URL}${routePath}`;
+  const description = `${gpuLabel} 在 ${parsed.model.label} 上的每美元性能：来自 InferenceX（SemiAnalysis 推出的独立开源基准测试平台）的经验证、可复现的每百万 token 成本结果，基于云服务商 TCO 归一化。${SUPPORTERS_LINE_ZH}查看哪款芯片在各交互性水平下更经济。`;
   return {
     title: `${fullLabel} — 每美元性能`,
     description,
-    alternates: zhAlternates(`/compare-per-dollar/${canonical}`),
+    alternates: zhAlternates(enScenarioPath(canonical)),
     openGraph: {
       title: `${fullLabel} — 每美元性能 | ${SITE_NAME}`,
       description,
@@ -73,11 +109,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function ComparePerDollarPageZh({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderPerDollarPageZh(slug, await searchParams, {});
+}
+
+export async function renderPerDollarPageZh(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parseCompareSlug(slug);
   if (!parsed) notFound();
 
-  const sp = await searchParams;
-
   const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
   if (canonical !== slug.toLowerCase()) {
     const qs = Object.entries(sp)
@@ -88,38 +130,46 @@ export default async function ComparePerDollarPageZh({ params, searchParams }: P
       })
       .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
       .join('&');
-    permanentRedirect(`/zh/compare-per-dollar/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // the reader to the pair's default workload instead.
+    permanentRedirect(`${scenarioPath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
-  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
-  const summaryA = summarize(rows, parsed.a);
-  const summaryB = summarize(rows, parsed.b);
-  const { sequence: pickedSequence, precision: pickedPrecision } = pickPairDefaults(
-    rows,
-    parsed.a,
-    parsed.b,
-  );
+  const fallbackSequence = null;
 
   const urlSeq = pickString(sp.i_seq);
   const urlPrec = pickString(sp.i_prec);
   const urlModel = pickString(sp.g_model);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence;
-  const effectivePrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : pickedPrecision;
   const effectiveModel =
     urlModel && KNOWN_MODELS.has(urlModel) ? urlModel : parsed.model.displayName;
-
-  const { defaultTargets, ssrRows, interactivityRange } = computeCompareTableData(
-    rows,
+  // Path beats query beats the pair's default: a scenario segment is an
+  // explicit address for one workload, so it outranks a stale `?i_seq=`.
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const requestedSequence = pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : null);
+  const requestedPrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : null;
+  const {
+    sequence: effectiveSequence,
+    precision: effectivePrecision,
+    summaryA,
+    summaryB,
+    defaultTargets,
+    ssrRows,
+    interactivityRange,
+    oldest,
+    newest,
+    initialPairBenchmarkRows,
+  } = await getComparePageDerivedData(
+    parsed.model.dbKeys,
     parsed.a,
     parsed.b,
-    effectiveSequence,
-    effectivePrecision,
+    requestedSequence,
+    requestedPrecision,
+    fallbackSequence,
   );
 
-  const url = `${SITE_URL}/zh/compare-per-dollar/${canonical}`;
+  const url = `${SITE_URL}${scenarioPath(canonical, scenarioSegment)}`;
   // The PNG route exists only under the EN tree; zh JSON-LD references it there.
   const imageUrl = `${SITE_URL}/compare-per-dollar/${canonical}/performance-per-dollar.png`;
-  const { oldest, newest } = dateRangeForPair(rows, parsed.a, parsed.b);
   const jsonLd = buildJsonLdZh(
     'per-dollar',
     parsed.model,
@@ -169,7 +219,13 @@ export default async function ComparePerDollarPageZh({ params, searchParams }: P
         defaultSequence={effectiveSequence}
         defaultPrecision={effectivePrecision}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
+        initialBenchmarkRows={initialCompareBenchmarkRows(
+          parsed.model.displayName,
+          effectiveModel,
+          initialPairBenchmarkRows,
+        )}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO_ZH : null}
         aLabel={aLabel}
         bLabel={bLabel}
         aVendor={aMeta?.vendor ?? ''}
diff --git a/packages/app/src/app/zh/compare-per-dollar/layout.tsx b/packages/app/src/app/zh/compare-per-dollar/layout.tsx
index 0812858a0..4ff2b2492 100644
--- a/packages/app/src/app/zh/compare-per-dollar/layout.tsx
+++ b/packages/app/src/app/zh/compare-per-dollar/layout.tsx
@@ -1,13 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
 export default function ComparePerDollarLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/zh/compare-per-dollar/page.tsx b/packages/app/src/app/zh/compare-per-dollar/page.tsx
index 7dbcfa8cf..f505a9a71 100644
--- a/packages/app/src/app/zh/compare-per-dollar/page.tsx
+++ b/packages/app/src/app/zh/compare-per-dollar/page.tsx
@@ -11,6 +11,7 @@ import {
 import { ComparePairCardLink } from '@/components/compare/compare-pair-card-link';
 import { JsonLd } from '@/components/json-ld';
 import { Card } from '@/components/ui/card';
+import { comparisonPairHref, comparisonScenarioForModel } from '@/lib/compare-agentx';
 import { getComparablePairsByModelSlug } from '@/lib/compare-availability';
 import { type ComparePair, COMPARE_MODEL_SLUGS, type CompareModelSlug } from '@/lib/compare-slug';
 import { bucketComparePairsByVendor, formatModelList } from '@/lib/compare-ssr';
@@ -18,14 +19,14 @@ import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 
 export const dynamic = 'force-dynamic';
 
-const DESCRIPTION = `哪款 Chip 每美元推理性能最高？InferenceX 是 SemiAnalysis 推出的独立开源基准测试平台，提供经过验证的、可复现的测试结果。${SUPPORTERS_LINE_ZH}横向对比 DeepSeek V4 Pro、DeepSeek R1、Kimi K2、MiniMax M3、GLM 5、Qwen 3.5 等模型基于云服务商 TCO 归一化的每百万 token 成本。`;
+const DESCRIPTION = `哪款芯片每美元推理性能最高？InferenceX 是 SemiAnalysis 推出的独立开源基准测试平台，提供经过验证的、可复现的测试结果。${SUPPORTERS_LINE_ZH}横向对比 DeepSeek V4 Pro、DeepSeek R1、Kimi K2、MiniMax M3、GLM 5、Qwen 3.5 等模型基于云服务商 TCO 归一化的每百万 token 成本。`;
 
 export const metadata: Metadata = {
-  title: 'Chip 每美元性能',
+  title: '芯片每美元性能',
   description: DESCRIPTION,
   alternates: zhAlternates('/compare-per-dollar'),
   openGraph: {
-    title: `Chip 每美元性能 | ${SITE_NAME}`,
+    title: `芯片每美元性能 | ${SITE_NAME}`,
     description: DESCRIPTION,
     url: `${SITE_URL}/zh/compare-per-dollar`,
     type: 'website',
@@ -33,7 +34,7 @@ export const metadata: Metadata = {
   },
   twitter: {
     card: 'summary_large_image',
-    title: `Chip 每美元性能 | ${SITE_NAME}`,
+    title: `芯片每美元性能 | ${SITE_NAME}`,
     description: DESCRIPTION,
   },
 };
@@ -77,7 +78,7 @@ function groupPairsByVendorForModel(
 const jsonLd = {
   '@context': 'https://schema.org',
   '@type': 'CollectionPage',
-  name: `Chip 每美元性能 | ${SITE_NAME}`,
+  name: `芯片每美元性能 | ${SITE_NAME}`,
   description: DESCRIPTION,
   url: `${SITE_URL}/zh/compare-per-dollar`,
   inLanguage: 'zh-CN',
@@ -95,12 +96,16 @@ export default async function ComparePerDollarIndexPageZh() {
       <JsonLd data={jsonLd} />
       <section>
         <Card>
-          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">Chip 每美元性能</h1>
+          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">芯片每美元性能</h1>
           <p className="mt-3 text-base lg:text-lg text-muted-foreground max-w-3xl">
             {totalUrls.toLocaleString()} 组每百万 token 成本的正面对比，涵盖{' '}
             {formatModelList(modelsWithPairs)}
             。性能按所属云服务商 TCO 归一化——每个页面展示每 token 成本图表及插值美元/百万 token
-            对比表格，帮助您在任意目标交互性水平下选出更经济的 Chip。
+            对比表格，帮助您在任意目标交互性水平下选出更经济的芯片。
+          </p>
+          <p className="mt-3 text-base lg:text-lg text-muted-foreground max-w-3xl">
+            具备 AgentX 数据的模型会打开长上下文、多轮 trace 回放结果；尚未纳入 AgentX
+            的模型则打开受控的 8K→1K 负载。每张卡片都会标明其对应场景。
           </p>
           <div className="mt-6 flex flex-wrap gap-3">
             <Link
@@ -108,7 +113,7 @@ export default async function ComparePerDollarIndexPageZh() {
               href="/zh/compare"
               className="inline-flex items-center gap-2 rounded-md bg-brand px-5 py-3 text-base lg:text-lg font-semibold text-primary-foreground shadow-sm transition-colors hover:bg-brand/90"
             >
-              Chip 延迟 + 吞吐量对比
+              芯片延迟 + 吞吐量对比
               <span aria-hidden="true" className="text-lg lg:text-xl">
                 →
               </span>
@@ -140,13 +145,16 @@ export default async function ComparePerDollarIndexPageZh() {
       {modelsWithPairs.map((model) => {
         const pairs = comparablePairsByModel.get(model.slug) ?? [];
         const groups = groupPairsByVendorForModel(model, pairs);
+        // Same scenario split /compare uses: models with AgentX data open the
+        // agentic trace replay, the rest open the fixed 8K→1K workload.
+        const scenario = comparisonScenarioForModel(model);
         return (
           <section key={model.slug} id={model.slug}>
             <Card className="flex flex-col gap-4">
               <div>
                 <h2 className="text-xl lg:text-2xl font-bold tracking-tight">{model.label}</h2>
                 <p className="text-sm text-muted-foreground mt-1">
-                  {pairs.length} 组 Chip 对比具有 {model.label} 的每 token 成本基准测试数据。
+                  {pairs.length} 组芯片对比具有 {model.label} 的每 token 成本基准测试数据。
                 </p>
               </div>
               {groups.map((group) => (
@@ -163,10 +171,11 @@ export default async function ComparePerDollarIndexPageZh() {
                       return (
                         <ComparePairCardLink
                           key={slug}
-                          href={`/zh/compare-per-dollar/${slug}`}
+                          href={comparisonPairHref('zh', slug, model, 'compare-per-dollar')}
                           slug={slug}
                           label={label}
                           archLine={archLine}
+                          scenarioLabel={scenario.label}
                         />
                       );
                     })}
diff --git a/packages/app/src/app/zh/compare-precision/[slug]/[scenario]/page.tsx b/packages/app/src/app/zh/compare-precision/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..6235ada8b
--- /dev/null
+++ b/packages/app/src/app/zh/compare-precision/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildPrecisionMetadataZh, renderPrecisionPageZh } from '../page';
+
+/**
+ * `/zh/compare-precision/<slug>/<scenario>` — the same comparison with its workload pinned by
+ * the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments rather than silently falling back to the pair's
+ *  default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildPrecisionMetadataZh(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function PrecisionScenarioPageZh({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderPrecisionPageZh(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/zh/compare-precision/[slug]/opengraph-image.tsx b/packages/app/src/app/zh/compare-precision/[slug]/opengraph-image.tsx
index 97ce9fce1..647e8a827 100644
--- a/packages/app/src/app/zh/compare-precision/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/zh/compare-precision/[slug]/opengraph-image.tsx
@@ -1,7 +1,49 @@
-export {
-  default,
-  alt,
-  size,
-  contentType,
-  generateStaticParams,
-} from '../../../compare-precision/[slug]/opengraph-image';
+import { notFound } from 'next/navigation';
+
+import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
+
+import { renderCompareOg } from '@/lib/compare-og';
+import { getAllComparablePrecisionSlugs } from '@/lib/compare-variant-availability';
+import {
+  canonicalPrecisionCompareSlug,
+  parsePrecisionCompareSlug,
+  precisionDisplayLabel,
+} from '@/lib/compare-variant-slug';
+
+export const alt = '芯片推理精度基准测试对比';
+export const size = { width: 1200, height: 630 };
+export const contentType = 'image/png';
+
+export async function generateStaticParams() {
+  const slugs = await getAllComparablePrecisionSlugs();
+  return slugs.map(({ modelSlug, gpu, precA, precB }) => ({
+    slug: canonicalPrecisionCompareSlug(modelSlug, gpu, precA, precB),
+  }));
+}
+
+export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
+  const { slug } = await params;
+  const parsed = parsePrecisionCompareSlug(slug);
+  if (
+    !parsed ||
+    canonicalPrecisionCompareSlug(parsed.model.slug, parsed.gpu, parsed.precA, parsed.precB) !==
+      slug.toLowerCase()
+  ) {
+    notFound();
+  }
+
+  const gpuLabel = HW_REGISTRY[parsed.gpu]?.label ?? parsed.gpu.toUpperCase();
+  const aLabel = precisionDisplayLabel(parsed.precA);
+  const bLabel = precisionDisplayLabel(parsed.precB);
+  const title = `${gpuLabel}：${aLabel} 对比 ${bLabel}`;
+  const titleSize = title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
+
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · 精度对比`,
+    title,
+    titleSize,
+    footer: 'AI 推理基准测试 · 精度对比',
+    language: 'zh',
+    size,
+  });
+}
diff --git a/packages/app/src/app/zh/compare-precision/[slug]/page.tsx b/packages/app/src/app/zh/compare-precision/[slug]/page.tsx
index c127130f6..b08f724f6 100644
--- a/packages/app/src/app/zh/compare-precision/[slug]/page.tsx
+++ b/packages/app/src/app/zh/compare-precision/[slug]/page.tsx
@@ -9,6 +9,12 @@ import {
 } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
+import { AGENTIC_SCENARIO_INTRO_ZH } from '@/lib/compare-ssr-zh';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
 import { getCachedBenchmarks, KNOWN_SEQUENCES, pickString } from '@/lib/compare-ssr';
 import {
   canonicalPrecisionCompareSlug,
@@ -32,6 +38,30 @@ import ComparePrecisionPageClient from '../../../compare-precision/[slug]/page-c
 
 export const dynamic = 'force-dynamic';
 
+/**
+ * `/zh/compare-precision/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than duplicated
+ * in a parallel route file so the body — redirects, JSON-LD, metadata, client
+ * props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
+/** English twin of `scenarioPath` — hreflang pairs are keyed off the
+ *  English route, so the segment has to survive the locale swap. */
+function enScenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/compare-precision/${canonical}/${scenarioSegment}`
+    : `/compare-precision/${canonical}`;
+}
+
+function scenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/zh/compare-precision/${canonical}/${scenarioSegment}`
+    : `/zh/compare-precision/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
@@ -39,6 +69,13 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildPrecisionMetadataZh(slug, {});
+}
+
+export function buildPrecisionMetadataZh(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Metadata {
   const parsed = parsePrecisionCompareSlug(slug);
   if (!parsed) return {};
   const gpuMeta = HW_REGISTRY[parsed.gpu];
@@ -51,12 +88,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     parsed.precA,
     parsed.precB,
   );
-  const url = `${SITE_URL}/zh/compare-precision/${canonical}`;
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const url = `${SITE_URL}${routePath}`;
   const description = `${gpuLabel} 上 ${aLabel} 与 ${bLabel} 精度对比（${parsed.model.label}）：来自 InferenceX（SemiAnalysis 推出的独立开源基准测试平台）的经验证、可复现的结果。${SUPPORTERS_LINE_ZH}查看哪种量化精度在各交互性水平下吞吐量和成本更优。`;
   return {
     title: `${parsed.model.label} — ${gpuLabel} ${aLabel} vs ${bLabel} — 精度对比`,
     description,
-    alternates: zhAlternates(`/compare-precision/${canonical}`),
+    alternates: zhAlternates(enScenarioPath(canonical)),
     openGraph: {
       title: `${parsed.model.label} — ${gpuLabel} ${aLabel} vs ${bLabel} — 精度对比 | ${SITE_NAME}`,
       description,
@@ -74,11 +116,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function ComparePrecisionPageZh({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderPrecisionPageZh(slug, await searchParams, {});
+}
+
+export async function renderPrecisionPageZh(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parsePrecisionCompareSlug(slug);
   if (!parsed) notFound();
 
-  const sp = await searchParams;
-
   const canonical = canonicalPrecisionCompareSlug(
     parsed.model.slug,
     parsed.gpu,
@@ -94,7 +142,9 @@ export default async function ComparePrecisionPageZh({ params, searchParams }: P
       })
       .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
       .join('&');
-    permanentRedirect(`/zh/compare-precision/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // the reader to the pair's default workload instead.
+    permanentRedirect(`${scenarioPath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
   const rows = await getCachedBenchmarks(parsed.model.dbKeys);
@@ -110,7 +160,11 @@ export default async function ComparePrecisionPageZh({ params, searchParams }: P
   );
 
   const urlSeq = pickString(sp.i_seq);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence;
+  // Path beats query beats the pair's default: a scenario segment is an
+  // explicit address for one workload, so it outranks a stale `?i_seq=`.
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const effectiveSequence =
+    pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence);
 
   const { defaultTargets, ssrRows, interactivityRange } = computeVariantCompareTableData(
     rows,
@@ -128,7 +182,7 @@ export default async function ComparePrecisionPageZh({ params, searchParams }: P
   const aLabel = precisionDisplayLabel(parsed.precA);
   const bLabel = precisionDisplayLabel(parsed.precB);
 
-  const url = `${SITE_URL}/zh/compare-precision/${canonical}`;
+  const url = `${SITE_URL}${scenarioPath(canonical, scenarioSegment)}`;
   // The PNG route exists only under the EN tree; zh JSON-LD references it there.
   const imageUrl = `${SITE_URL}/compare-precision/${canonical}/precision-comparison.png`;
   const { oldest, newest } = dateRangeForVariantPair(rows, parsed.gpu, sideA, sideB);
@@ -172,6 +226,7 @@ export default async function ComparePrecisionPageZh({ params, searchParams }: P
         precB={parsed.precB}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO_ZH : null}
         gpuLabel={gpuLabel}
         gpuVendor={gpuMeta?.vendor ?? ''}
         gpuArch={gpuMeta?.arch ?? ''}
diff --git a/packages/app/src/app/zh/compare-precision/layout.tsx b/packages/app/src/app/zh/compare-precision/layout.tsx
index 24e3fff3c..0706bbeb3 100644
--- a/packages/app/src/app/zh/compare-precision/layout.tsx
+++ b/packages/app/src/app/zh/compare-precision/layout.tsx
@@ -1,13 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
 export default function ComparePrecisionLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/zh/compare-precision/page.tsx b/packages/app/src/app/zh/compare-precision/page.tsx
index 066eb5054..df52e001c 100644
--- a/packages/app/src/app/zh/compare-precision/page.tsx
+++ b/packages/app/src/app/zh/compare-precision/page.tsx
@@ -19,14 +19,14 @@ import { canonicalPrecisionCompareSlug, precisionDisplayLabel } from '@/lib/comp
 
 export const dynamic = 'force-dynamic';
 
-const DESCRIPTION = `精度如何影响 Chip 推理性能？InferenceX 是 SemiAnalysis 推出的独立开源基准测试平台，提供经过验证的、可复现的测试结果。${SUPPORTERS_LINE_ZH}在同一 Chip 上横向对比 FP4、FP8、BF16、INT4 等多种量化精度对 DeepSeek V4 Pro、DeepSeek R1、Kimi K2、MiniMax M3、GLM 5、Qwen 3.5 等模型的影响。`;
+const DESCRIPTION = `精度如何影响芯片推理性能？InferenceX 是 SemiAnalysis 推出的独立开源基准测试平台，提供经过验证的、可复现的测试结果。${SUPPORTERS_LINE_ZH}在同一芯片上横向对比 FP4、FP8、BF16、INT4 等多种量化精度对 DeepSeek V4 Pro、DeepSeek R1、Kimi K2、MiniMax M3、GLM 5、Qwen 3.5 等模型的影响。`;
 
 export const metadata: Metadata = {
-  title: 'Chip 精度对比',
+  title: '芯片精度对比',
   description: DESCRIPTION,
   alternates: zhAlternates('/compare-precision'),
   openGraph: {
-    title: `Chip 精度对比 | ${SITE_NAME}`,
+    title: `芯片精度对比 | ${SITE_NAME}`,
     description: DESCRIPTION,
     url: `${SITE_URL}/zh/compare-precision`,
     type: 'website',
@@ -34,7 +34,7 @@ export const metadata: Metadata = {
   },
   twitter: {
     card: 'summary_large_image',
-    title: `Chip 精度对比 | ${SITE_NAME}`,
+    title: `芯片精度对比 | ${SITE_NAME}`,
     description: DESCRIPTION,
   },
 };
@@ -42,7 +42,7 @@ export const metadata: Metadata = {
 const jsonLd = {
   '@context': 'https://schema.org',
   '@type': 'CollectionPage',
-  name: `Chip 精度对比 | ${SITE_NAME}`,
+  name: `芯片精度对比 | ${SITE_NAME}`,
   description: DESCRIPTION,
   url: `${SITE_URL}/zh/compare-precision`,
   inLanguage: 'zh-CN',
@@ -60,10 +60,10 @@ export default async function ComparePrecisionIndexPageZh() {
       <JsonLd data={jsonLd} />
       <section>
         <Card>
-          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">Chip 精度对比</h1>
+          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">芯片精度对比</h1>
           <p className="mt-3 text-base lg:text-lg text-muted-foreground max-w-3xl">
             {totalUrls.toLocaleString()} 组精度正面对比，涵盖 {formatModelList(modelsWithPairs)}
-            。查看 FP4、FP8、BF16、INT4 等多种量化精度如何影响同一 Chip
+            。查看 FP4、FP8、BF16、INT4 等多种量化精度如何影响同一芯片
             上的吞吐量、成本和交互性——每个页面展示推理图表及插值对比表格。
           </p>
           <div className="mt-6 flex flex-wrap gap-3" data-testid="compare-precision-index-links">
@@ -71,7 +71,7 @@ export default async function ComparePrecisionIndexPageZh() {
               href="/zh/compare"
               className="inline-flex items-center gap-2 rounded-md border border-border px-5 py-3 text-base lg:text-lg font-semibold text-foreground shadow-sm transition-colors hover:bg-muted"
             >
-              Chip 对比
+              芯片对比
               <span aria-hidden="true" className="text-lg lg:text-xl">
                 →
               </span>
diff --git a/packages/app/src/app/zh/compare-spec-decode/[slug]/[scenario]/page.tsx b/packages/app/src/app/zh/compare-spec-decode/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..59832eef4
--- /dev/null
+++ b/packages/app/src/app/zh/compare-spec-decode/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildSpecDecodeMetadataZh, renderSpecDecodePageZh } from '../page';
+
+/**
+ * `/zh/compare-spec-decode/<slug>/<scenario>` — the same comparison with its workload pinned by
+ * the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments rather than silently falling back to the pair's
+ *  default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildSpecDecodeMetadataZh(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function SpecDecodeScenarioPageZh({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderSpecDecodePageZh(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/zh/compare-spec-decode/[slug]/opengraph-image.tsx b/packages/app/src/app/zh/compare-spec-decode/[slug]/opengraph-image.tsx
index 5c21e7e0f..400db8d69 100644
--- a/packages/app/src/app/zh/compare-spec-decode/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/zh/compare-spec-decode/[slug]/opengraph-image.tsx
@@ -1,7 +1,54 @@
-export {
-  default,
-  alt,
-  size,
-  contentType,
-  generateStaticParams,
-} from '../../../compare-spec-decode/[slug]/opengraph-image';
+import { notFound } from 'next/navigation';
+
+import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
+
+import { renderCompareOg } from '@/lib/compare-og';
+import { getAllComparableSpecDecodeSlugs } from '@/lib/compare-variant-availability';
+import {
+  canonicalSpecDecodeCompareSlug,
+  parseSpecDecodeCompareSlug,
+  precisionDisplayLabel,
+  specMethodDisplayLabel,
+} from '@/lib/compare-variant-slug';
+
+export const alt = '芯片推理投机解码基准测试对比';
+export const size = { width: 1200, height: 630 };
+export const contentType = 'image/png';
+
+export async function generateStaticParams() {
+  const slugs = await getAllComparableSpecDecodeSlugs();
+  return slugs.map(({ modelSlug, gpu, precision, method }) => ({
+    slug: canonicalSpecDecodeCompareSlug(modelSlug, gpu, precision, method),
+  }));
+}
+
+export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
+  const { slug } = await params;
+  const parsed = parseSpecDecodeCompareSlug(slug);
+  if (
+    !parsed ||
+    canonicalSpecDecodeCompareSlug(
+      parsed.model.slug,
+      parsed.gpu,
+      parsed.precision,
+      parsed.method,
+    ) !== slug.toLowerCase()
+  ) {
+    notFound();
+  }
+
+  const gpuLabel = HW_REGISTRY[parsed.gpu]?.label ?? parsed.gpu.toUpperCase();
+  const precisionLabel = precisionDisplayLabel(parsed.precision);
+  const methodLabel = specMethodDisplayLabel(parsed.model.displayName, parsed.method);
+  const title = `${gpuLabel} ${precisionLabel}：${methodLabel} 对比关闭`;
+  const titleSize = title.length > 34 ? 72 : title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
+
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · 投机解码对比`,
+    title,
+    titleSize,
+    footer: 'AI 推理基准测试 · 投机解码对比',
+    language: 'zh',
+    size,
+  });
+}
diff --git a/packages/app/src/app/zh/compare-spec-decode/[slug]/page.tsx b/packages/app/src/app/zh/compare-spec-decode/[slug]/page.tsx
index f1257b79e..6886887fb 100644
--- a/packages/app/src/app/zh/compare-spec-decode/[slug]/page.tsx
+++ b/packages/app/src/app/zh/compare-spec-decode/[slug]/page.tsx
@@ -9,6 +9,12 @@ import {
 } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
+import { AGENTIC_SCENARIO_INTRO_ZH } from '@/lib/compare-ssr-zh';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
 import { getCachedBenchmarks, KNOWN_SEQUENCES, pickString } from '@/lib/compare-ssr';
 import {
   canonicalSpecDecodeCompareSlug,
@@ -34,6 +40,30 @@ import CompareSpecDecodePageClient from '../../../compare-spec-decode/[slug]/pag
 
 export const dynamic = 'force-dynamic';
 
+/**
+ * `/zh/compare-spec-decode/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than duplicated
+ * in a parallel route file so the body — redirects, JSON-LD, metadata, client
+ * props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
+/** English twin of `scenarioPath` — hreflang pairs are keyed off the
+ *  English route, so the segment has to survive the locale swap. */
+function enScenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/compare-spec-decode/${canonical}/${scenarioSegment}`
+    : `/compare-spec-decode/${canonical}`;
+}
+
+function scenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/zh/compare-spec-decode/${canonical}/${scenarioSegment}`
+    : `/zh/compare-spec-decode/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
@@ -41,6 +71,13 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildSpecDecodeMetadataZh(slug, {});
+}
+
+export function buildSpecDecodeMetadataZh(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Metadata {
   const parsed = parseSpecDecodeCompareSlug(slug);
   if (!parsed) return {};
   const gpuMeta = HW_REGISTRY[parsed.gpu];
@@ -53,13 +90,18 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
     parsed.precision,
     parsed.method,
   );
-  const url = `${SITE_URL}/zh/compare-spec-decode/${canonical}`;
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const url = `${SITE_URL}${routePath}`;
   const description = `${parsed.model.label} 在 ${gpuLabel} ${precLabel} 上的 ${aLabel} vs Off 投机解码对比：来自 InferenceX（SemiAnalysis 推出的独立开源基准测试平台）的经验证、可复现结果。${SUPPORTERS_LINE_ZH}查看投机解码是否在各交互性水平下提升吞吐量和降低成本。`;
   const title = `${parsed.model.label} — ${gpuLabel} ${precLabel}: ${aLabel} vs Off — 投机解码对比`;
   return {
     title,
     description,
-    alternates: zhAlternates(`/compare-spec-decode/${canonical}`),
+    alternates: zhAlternates(enScenarioPath(canonical)),
     openGraph: {
       title: `${title} | ${SITE_NAME}`,
       description,
@@ -77,11 +119,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function CompareSpecDecodePageZh({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderSpecDecodePageZh(slug, await searchParams, {});
+}
+
+export async function renderSpecDecodePageZh(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parseSpecDecodeCompareSlug(slug);
   if (!parsed) notFound();
 
-  const sp = await searchParams;
-
   const canonical = canonicalSpecDecodeCompareSlug(
     parsed.model.slug,
     parsed.gpu,
@@ -97,7 +145,9 @@ export default async function CompareSpecDecodePageZh({ params, searchParams }:
       })
       .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
       .join('&');
-    permanentRedirect(`/zh/compare-spec-decode/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // the reader to the pair's default workload instead.
+    permanentRedirect(`${scenarioPath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
   const rows = await getCachedBenchmarks(parsed.model.dbKeys);
@@ -113,7 +163,9 @@ export default async function CompareSpecDecodePageZh({ params, searchParams }:
   const defaults = pickVariantPairDefaults('spec-decode', rows, parsed.gpu, sideA, sideB);
 
   const urlSeq = pickString(sp.i_seq);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : defaults.sequence;
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const effectiveSequence =
+    pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : defaults.sequence);
   const effectivePrecision = parsed.precision;
 
   const sideAFull: VariantCompareSide = {
@@ -137,7 +189,7 @@ export default async function CompareSpecDecodePageZh({ params, searchParams }:
   const summaryB = summarizeVariantSide(rows, parsed.gpu, sideBFull);
   const { oldest, newest } = dateRangeForVariantPair(rows, parsed.gpu, sideAFull, sideBFull);
 
-  const url = `${SITE_URL}/zh/compare-spec-decode/${canonical}`;
+  const url = `${SITE_URL}${scenarioPath(canonical, scenarioSegment)}`;
   // The PNG route exists only under the EN tree; zh JSON-LD references it there.
   const imageUrl = `${SITE_URL}/compare-spec-decode/${canonical}/spec-decode-comparison.png`;
 
@@ -184,6 +236,7 @@ export default async function CompareSpecDecodePageZh({ params, searchParams }:
         defaultPrecision={effectivePrecision}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO_ZH : null}
         gpuLabel={gpuLabel}
         precisionLabel={precLabel}
         gpuArch={gpuMeta?.arch ?? ''}
diff --git a/packages/app/src/app/zh/compare-spec-decode/layout.tsx b/packages/app/src/app/zh/compare-spec-decode/layout.tsx
index 67de7611b..f0a2fdb0d 100644
--- a/packages/app/src/app/zh/compare-spec-decode/layout.tsx
+++ b/packages/app/src/app/zh/compare-spec-decode/layout.tsx
@@ -1,13 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
 export default function CompareSpecDecodeLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/zh/compare-spec-decode/page.tsx b/packages/app/src/app/zh/compare-spec-decode/page.tsx
index 6e21d234e..53f509eed 100644
--- a/packages/app/src/app/zh/compare-spec-decode/page.tsx
+++ b/packages/app/src/app/zh/compare-spec-decode/page.tsx
@@ -26,14 +26,14 @@ import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 
 export const dynamic = 'force-dynamic';
 
-const DESCRIPTION = `投机解码（MTP 多 token 预测、MiniMax M3 的 EAGLE 等模型专用方法）是否能提升推理吞吐量和降低成本？InferenceX 是 SemiAnalysis 推出的独立开源基准测试平台，提供经过验证的、可复现的测试结果。${SUPPORTERS_LINE_ZH}每个页面对比同一模型和 Chip 上投机解码开启与关闭的性能差异。`;
+const DESCRIPTION = `投机解码（MTP 多 token 预测、MiniMax M3 的 EAGLE 等模型专用方法）是否能提升推理吞吐量和降低成本？InferenceX 是 SemiAnalysis 推出的独立开源基准测试平台，提供经过验证的、可复现的测试结果。${SUPPORTERS_LINE_ZH}每个页面对比同一模型和芯片上投机解码开启与关闭的性能差异。`;
 
 export const metadata: Metadata = {
-  title: 'Chip 投机解码对比',
+  title: '芯片投机解码对比',
   description: DESCRIPTION,
   alternates: zhAlternates('/compare-spec-decode'),
   openGraph: {
-    title: `Chip 投机解码对比 | ${SITE_NAME}`,
+    title: `芯片投机解码对比 | ${SITE_NAME}`,
     description: DESCRIPTION,
     url: `${SITE_URL}/zh/compare-spec-decode`,
     type: 'website',
@@ -41,7 +41,7 @@ export const metadata: Metadata = {
   },
   twitter: {
     card: 'summary_large_image',
-    title: `Chip 投机解码对比 | ${SITE_NAME}`,
+    title: `芯片投机解码对比 | ${SITE_NAME}`,
     description: DESCRIPTION,
   },
 };
@@ -49,7 +49,7 @@ export const metadata: Metadata = {
 const jsonLd = {
   '@context': 'https://schema.org',
   '@type': 'CollectionPage',
-  name: `Chip 投机解码对比 | ${SITE_NAME}`,
+  name: `芯片投机解码对比 | ${SITE_NAME}`,
   description: DESCRIPTION,
   url: `${SITE_URL}/zh/compare-spec-decode`,
   inLanguage: 'zh-CN',
@@ -84,10 +84,10 @@ export default async function CompareSpecDecodeIndexPageZh() {
       <JsonLd data={jsonLd} />
       <section>
         <Card>
-          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">Chip 投机解码对比</h1>
+          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">芯片投机解码对比</h1>
           <p className="mt-3 text-base lg:text-lg text-muted-foreground max-w-3xl">
             {totalUrls.toLocaleString()} 组投机解码对比，涵盖 {formatModelList(modelsWithPairs)}
-            。每个页面对比同一模型和 Chip 上投机解码方法（MTP、EAGLE
+            。每个页面对比同一模型和芯片上投机解码方法（MTP、EAGLE
             等）开启与关闭的推理性能——在相同交互性水平下的吞吐量、成本和交互性。
           </p>
           <div className="mt-6 flex flex-wrap gap-3" data-testid="compare-spec-decode-index-links">
@@ -95,7 +95,7 @@ export default async function CompareSpecDecodeIndexPageZh() {
               href="/zh/compare"
               className="inline-flex items-center gap-2 rounded-md border border-border px-5 py-3 text-base lg:text-lg font-semibold text-foreground shadow-sm transition-colors hover:bg-muted"
             >
-              Chip 对比
+              芯片对比
               <span aria-hidden="true" className="text-lg lg:text-xl">
                 →
               </span>
diff --git a/packages/app/src/app/zh/compare/[slug]/[scenario]/page.tsx b/packages/app/src/app/zh/compare/[slug]/[scenario]/page.tsx
new file mode 100644
index 000000000..1e5848569
--- /dev/null
+++ b/packages/app/src/app/zh/compare/[slug]/[scenario]/page.tsx
@@ -0,0 +1,37 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { type ScenarioSegment, sequenceForScenarioSegment } from '@/lib/compare-scenario-route';
+
+import { buildCompareMetadataZh, renderComparePageZh } from '../page';
+
+/**
+ * `/zh/compare/<slug>/<scenario>` — the same comparison with its workload pinned by
+ * the path instead of `?i_seq=`. Both routes render through the parent's
+ * exported body, so the two URLs cannot drift.
+ */
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string; scenario: string }>;
+  searchParams: Promise<Record<string, string | string[] | undefined>>;
+}
+
+/** Reject unknown segments rather than silently falling back to the pair's
+ *  default, which would serve 200s for arbitrary junk paths. */
+function assertScenario(scenario: string): ScenarioSegment {
+  if (!sequenceForScenarioSegment(scenario)) notFound();
+  return scenario.toLowerCase() as ScenarioSegment;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug, scenario } = await params;
+  return buildCompareMetadataZh(slug, { scenarioSegment: assertScenario(scenario) });
+}
+
+export default async function CompareScenarioPageZh({ params, searchParams }: Props) {
+  const { slug, scenario } = await params;
+  return renderComparePageZh(slug, await searchParams, {
+    scenarioSegment: assertScenario(scenario),
+  });
+}
diff --git a/packages/app/src/app/zh/compare/[slug]/opengraph-image.tsx b/packages/app/src/app/zh/compare/[slug]/opengraph-image.tsx
index 5f3d0707b..91b6adfeb 100644
--- a/packages/app/src/app/zh/compare/[slug]/opengraph-image.tsx
+++ b/packages/app/src/app/zh/compare/[slug]/opengraph-image.tsx
@@ -1,7 +1,36 @@
-export {
-  default,
-  alt,
-  size,
-  contentType,
-  generateStaticParams,
-} from '../../../compare/[slug]/opengraph-image';
+import { notFound } from 'next/navigation';
+
+import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
+
+import { getAllComparableCompareSlugs } from '@/lib/compare-availability';
+import { renderCompareOg } from '@/lib/compare-og';
+import { canonicalCompareSlug, parseCompareSlug } from '@/lib/compare-slug';
+
+export const alt = '芯片推理基准测试对比';
+export const size = { width: 1200, height: 630 };
+export const contentType = 'image/png';
+
+export async function generateStaticParams() {
+  const slugs = await getAllComparableCompareSlugs();
+  return slugs.map(({ modelSlug, a, b }) => ({ slug: canonicalCompareSlug(modelSlug, a, b) }));
+}
+
+export default async function OgImage({ params }: { params: Promise<{ slug: string }> }) {
+  const { slug } = await params;
+  const parsed = parseCompareSlug(slug);
+  if (!parsed) notFound();
+
+  const aLabel = HW_REGISTRY[parsed.a]?.label ?? parsed.a.toUpperCase();
+  const bLabel = HW_REGISTRY[parsed.b]?.label ?? parsed.b.toUpperCase();
+  const title = `${aLabel} 对比 ${bLabel}`;
+  const titleSize = title.length > 26 ? 80 : title.length > 18 ? 96 : 112;
+
+  return renderCompareOg({
+    eyebrow: `${parsed.model.label} · 芯片基准测试对比`,
+    title,
+    titleSize,
+    footer: 'AI 推理基准测试 · 延迟、吞吐量与成本',
+    language: 'zh',
+    size,
+  });
+}
diff --git a/packages/app/src/app/zh/compare/[slug]/page.tsx b/packages/app/src/app/zh/compare/[slug]/page.tsx
index 2dae12574..6daf7f781 100644
--- a/packages/app/src/app/zh/compare/[slug]/page.tsx
+++ b/packages/app/src/app/zh/compare/[slug]/page.tsx
@@ -4,7 +4,12 @@ import { notFound, permanentRedirect } from 'next/navigation';
 import { HW_REGISTRY, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
 import { JsonLd } from '@/components/json-ld';
-import { pickPairDefaults } from '@/lib/compare-pair-defaults';
+import {
+  isAgenticSequence,
+  type ScenarioSegment,
+  sequenceForScenarioSegment,
+} from '@/lib/compare-scenario-route';
+import { comparisonScenarioForModel } from '@/lib/compare-agentx';
 import {
   canonicalCompareSlug,
   compareDisplayLabel,
@@ -12,17 +17,13 @@ import {
   compareModelSeoName,
   parseCompareSlug,
 } from '@/lib/compare-slug';
+import { KNOWN_MODELS, KNOWN_PRECISIONS, KNOWN_SEQUENCES, pickString } from '@/lib/compare-ssr';
 import {
-  computeCompareTableData,
-  dateRangeForPair,
-  getCachedBenchmarks,
-  KNOWN_MODELS,
-  KNOWN_PRECISIONS,
-  KNOWN_SEQUENCES,
-  pickString,
-  summarize,
-} from '@/lib/compare-ssr';
+  getComparePageDerivedData,
+  initialCompareBenchmarkRows,
+} from '@/lib/compare-page-data.server';
 import {
+  AGENTIC_SCENARIO_INTRO_ZH,
   buildBreadcrumbJsonLdZh,
   buildJsonLdZh,
   compareMetaDescriptionZh,
@@ -34,6 +35,28 @@ import ComparePageClient from '../../../compare/[slug]/page-client';
 
 export const dynamic = 'force-dynamic';
 
+/**
+ * `/zh/compare/<slug>/<scenario>` renders this same page with the workload
+ * pinned by the path. The segment is threaded through rather than duplicated
+ * in a parallel route file so the body — redirects, JSON-LD, metadata, client
+ * props — is written once and cannot drift between the two URLs.
+ */
+export interface ScenarioOptions {
+  scenarioSegment?: ScenarioSegment;
+}
+
+/** English twin of `scenarioPath` — hreflang pairs are keyed off the
+ *  English route, so the segment has to survive the locale swap. */
+function enScenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment ? `/compare/${canonical}/${scenarioSegment}` : `/compare/${canonical}`;
+}
+
+function scenarioPath(canonical: string, scenarioSegment?: ScenarioSegment): string {
+  return scenarioSegment
+    ? `/zh/compare/${canonical}/${scenarioSegment}`
+    : `/zh/compare/${canonical}`;
+}
+
 interface Props {
   params: Promise<{ slug: string }>;
   searchParams: Promise<Record<string, string | string[] | undefined>>;
@@ -41,13 +64,25 @@ interface Props {
 
 export async function generateMetadata({ params }: Props): Promise<Metadata> {
   const { slug } = await params;
+  return buildCompareMetadataZh(slug, {});
+}
+
+export async function buildCompareMetadataZh(
+  slug: string,
+  { scenarioSegment }: ScenarioOptions,
+): Promise<Metadata> {
   const parsed = parseCompareSlug(slug);
   if (!parsed) return {};
   const fullLabel = compareModelDisplayLabel(parsed.model, parsed.a, parsed.b);
   const gpuLabel = compareDisplayLabel(parsed.a, parsed.b);
   const modelSeoName = compareModelSeoName(parsed.model);
   const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
-  const url = `${SITE_URL}/zh/compare/${canonical}`;
+  // The scenario segments are views of one comparison, so the bare slug URL
+  // stays the indexable representative and every segment canonicalizes to it.
+  // Without this, `/…/<slug>/<default-scenario>` and `/…/<slug>` would serve
+  // byte-identical pages, each claiming to be canonical.
+  const routePath = scenarioPath(canonical, scenarioSegment);
+  const url = `${SITE_URL}${routePath}`;
 
   // GPU-pair-first title (mirrors the English page) via `absolute` so the long
   // root template doesn't bury the search phrase. Model name / SKUs stay English.
@@ -55,15 +90,20 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
   // Stat-led Chinese meta description from the interpolated head-to-head numbers
   // at the slug's default operating point (falls back to boilerplate when thin).
-  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
-  const { sequence, precision } = pickPairDefaults(rows, parsed.a, parsed.b);
-  const { ssrRows } = computeCompareTableData(rows, parsed.a, parsed.b, sequence, precision);
+  const { ssrRows } = await getComparePageDerivedData(
+    parsed.model.dbKeys,
+    parsed.a,
+    parsed.b,
+    null,
+    null,
+    comparisonScenarioForModel(parsed.model).sequence,
+  );
   const description = compareMetaDescriptionZh(parsed.model, parsed.a, parsed.b, ssrRows);
 
   return {
     title: { absolute: title },
     description,
-    alternates: zhAlternates(`/compare/${canonical}`),
+    alternates: zhAlternates(enScenarioPath(canonical)),
     openGraph: {
       title: `${fullLabel} | ${SITE_NAME}`,
       description,
@@ -81,11 +121,17 @@ export async function generateMetadata({ params }: Props): Promise<Metadata> {
 
 export default async function ComparePageZh({ params, searchParams }: Props) {
   const { slug } = await params;
+  return renderComparePageZh(slug, await searchParams, {});
+}
+
+export async function renderComparePageZh(
+  slug: string,
+  sp: Record<string, string | string[] | undefined>,
+  { scenarioSegment }: ScenarioOptions,
+) {
   const parsed = parseCompareSlug(slug);
   if (!parsed) notFound();
 
-  const sp = await searchParams;
-
   const canonical = canonicalCompareSlug(parsed.model.slug, parsed.a, parsed.b);
   if (canonical !== slug.toLowerCase()) {
     const qs = Object.entries(sp)
@@ -96,36 +142,44 @@ export default async function ComparePageZh({ params, searchParams }: Props) {
       })
       .map(([k, v]) => `${encodeURIComponent(k)}=${encodeURIComponent(v)}`)
       .join('&');
-    permanentRedirect(`/zh/compare/${canonical}${qs ? `?${qs}` : ''}`);
+    // Keep the scenario segment across the redirect — dropping it would send
+    // the reader to the pair's default workload instead.
+    permanentRedirect(`${scenarioPath(canonical, scenarioSegment)}${qs ? `?${qs}` : ''}`);
   }
 
-  const rows = await getCachedBenchmarks(parsed.model.dbKeys);
-  const summaryA = summarize(rows, parsed.a);
-  const summaryB = summarize(rows, parsed.b);
-  const { sequence: pickedSequence, precision: pickedPrecision } = pickPairDefaults(
-    rows,
-    parsed.a,
-    parsed.b,
-  );
+  const fallbackSequence = comparisonScenarioForModel(parsed.model).sequence;
 
   const urlSeq = pickString(sp.i_seq);
   const urlPrec = pickString(sp.i_prec);
   const urlModel = pickString(sp.g_model);
-  const effectiveSequence = urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : pickedSequence;
-  const effectivePrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : pickedPrecision;
   const effectiveModel =
     urlModel && KNOWN_MODELS.has(urlModel) ? urlModel : parsed.model.displayName;
-
-  const { defaultTargets, ssrRows, interactivityRange } = computeCompareTableData(
-    rows,
+  // Path beats query beats the pair's default: a scenario segment is an
+  // explicit address for one workload, so it outranks a stale `?i_seq=`.
+  const pathSequence = scenarioSegment ? sequenceForScenarioSegment(scenarioSegment) : null;
+  const requestedSequence = pathSequence ?? (urlSeq && KNOWN_SEQUENCES.has(urlSeq) ? urlSeq : null);
+  const requestedPrecision = urlPrec && KNOWN_PRECISIONS.has(urlPrec) ? urlPrec : null;
+  const {
+    sequence: effectiveSequence,
+    precision: effectivePrecision,
+    summaryA,
+    summaryB,
+    defaultTargets,
+    ssrRows,
+    interactivityRange,
+    oldest,
+    newest,
+    initialPairBenchmarkRows,
+  } = await getComparePageDerivedData(
+    parsed.model.dbKeys,
     parsed.a,
     parsed.b,
-    effectiveSequence,
-    effectivePrecision,
+    requestedSequence,
+    requestedPrecision,
+    fallbackSequence,
   );
 
-  const url = `${SITE_URL}/zh/compare/${canonical}`;
-  const { oldest, newest } = dateRangeForPair(rows, parsed.a, parsed.b);
+  const url = `${SITE_URL}${scenarioPath(canonical, scenarioSegment)}`;
   const jsonLd = buildJsonLdZh(
     'full',
     parsed.model,
@@ -172,8 +226,14 @@ export default async function ComparePageZh({ params, searchParams }: Props) {
         defaultModel={effectiveModel}
         defaultSequence={effectiveSequence}
         defaultPrecision={effectivePrecision}
+        initialBenchmarkRows={initialCompareBenchmarkRows(
+          parsed.model.displayName,
+          effectiveModel,
+          initialPairBenchmarkRows,
+        )}
         ssrTableData={{ defaultTargets, ssrRows, interactivityRange }}
         narrative={narrative}
+        agenticIntro={isAgenticSequence(effectiveSequence) ? AGENTIC_SCENARIO_INTRO_ZH : null}
         aLabel={aLabel}
         bLabel={bLabel}
         aVendor={aMeta?.vendor ?? ''}
diff --git a/packages/app/src/app/zh/compare/layout.tsx b/packages/app/src/app/zh/compare/layout.tsx
index 9ccb8b2dc..76217427d 100644
--- a/packages/app/src/app/zh/compare/layout.tsx
+++ b/packages/app/src/app/zh/compare/layout.tsx
@@ -1,13 +1,5 @@
-import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { ComparePageLayout } from '@/components/compare/compare-page-layout';
 
 export default function CompareLayout({ children }: { children: React.ReactNode }) {
-  return (
-    <UnofficialRunProvider>
-      <main className="relative">
-        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
-          {children}
-        </div>
-      </main>
-    </UnofficialRunProvider>
-  );
+  return <ComparePageLayout>{children}</ComparePageLayout>;
 }
diff --git a/packages/app/src/app/zh/compare/page.tsx b/packages/app/src/app/zh/compare/page.tsx
index ee5efc30b..ca0276ca8 100644
--- a/packages/app/src/app/zh/compare/page.tsx
+++ b/packages/app/src/app/zh/compare/page.tsx
@@ -1,31 +1,31 @@
 import type { Metadata } from 'next';
 import Link from 'next/link';
 
-import {
-  HW_REGISTRY,
-  SITE_NAME,
-  SITE_URL,
-  SUPPORTERS_LINE_ZH,
-} from '@semianalysisai/inferencex-constants';
+import { HW_REGISTRY, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 
+import { AgentXCompareHero } from '@/components/compare/agentx-compare-hero';
 import { ComparePairCardLink } from '@/components/compare/compare-pair-card-link';
 import { JsonLd } from '@/components/json-ld';
 import { Card } from '@/components/ui/card';
+import { ModelLogo } from '@/components/ui/model-logo';
+import { comparisonPairHref, comparisonScenarioForModel } from '@/lib/compare-agentx';
 import { getComparablePairsByModelSlug } from '@/lib/compare-availability';
 import { type ComparePair, COMPARE_MODEL_SLUGS, type CompareModelSlug } from '@/lib/compare-slug';
 import { bucketComparePairsByVendor, formatModelList } from '@/lib/compare-ssr';
+import { type Model } from '@/lib/data-mappings';
 import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 
 export const dynamic = 'force-dynamic';
 
-const DESCRIPTION = `InferenceX 是 SemiAnalysis 推出的独立开源 Chip 推理基准测试平台，提供经过验证的、可复现的每夜测试结果。${SUPPORTERS_LINE_ZH}横向对比 DeepSeek V4 Pro、DeepSeek R1、Kimi K2、MiniMax M3、GLM 5、Qwen 3.5 等模型的延迟、吞吐量与成本。`;
+const DESCRIPTION =
+  '对比 Kimi K3、DeepSeek V4 Pro、MiniMax M3、Qwen 3.5 与 GLM 5.3 的 AgentX 智能体推理结果，并浏览定长序列芯片对比。';
 
 export const metadata: Metadata = {
-  title: 'Chip 对比',
+  title: 'AgentX 智能体推理对比',
   description: DESCRIPTION,
   alternates: zhAlternates('/compare'),
   openGraph: {
-    title: `Chip 对比 | ${SITE_NAME}`,
+    title: `AgentX 智能体推理对比 | ${SITE_NAME}`,
     description: DESCRIPTION,
     url: `${SITE_URL}/zh/compare`,
     type: 'website',
@@ -33,7 +33,7 @@ export const metadata: Metadata = {
   },
   twitter: {
     card: 'summary_large_image',
-    title: `Chip 对比 | ${SITE_NAME}`,
+    title: `AgentX 智能体推理对比 | ${SITE_NAME}`,
     description: DESCRIPTION,
   },
 };
@@ -77,7 +77,7 @@ function groupPairsByVendorForModel(
 const jsonLd = {
   '@context': 'https://schema.org',
   '@type': 'CollectionPage',
-  name: `Chip 对比 | ${SITE_NAME}`,
+  name: `AgentX 智能体推理对比 | ${SITE_NAME}`,
   description: DESCRIPTION,
   url: `${SITE_URL}/zh/compare`,
   inLanguage: 'zh-CN',
@@ -93,21 +93,29 @@ export default async function CompareIndexPageZh() {
   return (
     <>
       <JsonLd data={jsonLd} />
-      <section>
+      <AgentXCompareHero locale="zh" />
+
+      <section id="model-comparisons" data-testid="compare-model-catalog">
         <Card>
-          <h1 className="text-2xl lg:text-4xl font-bold tracking-tight">Chip 对比</h1>
+          <p className="font-mono text-xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
+            对比结果目录
+          </p>
+          <h2 className="mt-2 text-2xl font-bold tracking-tight lg:text-3xl">
+            AgentX 与 8K→1K 结果
+          </h2>
           <p className="mt-3 text-base lg:text-lg text-muted-foreground max-w-3xl">
             {totalUrls.toLocaleString()} 组推理基准测试的正面对比，涵盖{' '}
             {formatModelList(modelsWithPairs)}
-            。每个页面均包含延迟、吞吐量和成本指标的交互式图表，以及插值对比表格。
+            。已有 AgentX 数据的模型默认打开长上下文、多轮 trace replay 结果；尚未纳入 AgentX
+            的模型默认打开受控的 8K→1K 工作负载。每张卡片均标明对应场景。
           </p>
           <div className="mt-6 flex flex-wrap gap-3">
             <Link
               data-testid="compare-index-per-dollar-link-zh"
               href="/zh/compare-per-dollar"
-              className="inline-flex items-center gap-2 rounded-md bg-brand px-5 py-3 text-base lg:text-lg font-semibold text-primary-foreground shadow-sm transition-colors hover:bg-brand/90"
+              className="inline-flex items-center gap-2 rounded-md border border-border px-5 py-3 text-base lg:text-lg font-semibold text-foreground shadow-sm transition-colors hover:bg-muted"
             >
-              Chip 每美元性能对比
+              芯片每美元性能对比
               <span aria-hidden="true" className="text-lg lg:text-xl">
                 →
               </span>
@@ -139,13 +147,21 @@ export default async function CompareIndexPageZh() {
       {modelsWithPairs.map((model) => {
         const pairs = comparablePairsByModel.get(model.slug) ?? [];
         const groups = groupPairsByVendorForModel(model, pairs);
+        const scenario = comparisonScenarioForModel(model);
         return (
           <section key={model.slug} id={model.slug}>
             <Card className="flex flex-col gap-4">
               <div>
-                <h2 className="text-xl lg:text-2xl font-bold tracking-tight">{model.label}</h2>
+                {/* `displayName` 按约定即 Model 枚举值（见 compare-slug.ts 中的
+                    CompareModelSlug），因此共享的 ModelLogo 可从 MODEL_CONFIG
+                    解析各区块的品牌标识（DeepSeek、MiniMax、Kimi 等）；没有
+                    配置 logo 的模型则不渲染任何内容。 */}
+                <h2 className="flex items-center gap-2.5 text-xl lg:text-2xl font-bold tracking-tight">
+                  <ModelLogo model={model.displayName as Model} className="size-6 lg:size-7" />
+                  {model.label}
+                </h2>
                 <p className="text-sm text-muted-foreground mt-1">
-                  {pairs.length} 组 Chip 对比具有 {model.label} 的基准测试数据。
+                  {pairs.length} 组芯片对比具有 {model.label} 的基准测试数据。
                 </p>
               </div>
               {groups.map((group) => (
@@ -162,10 +178,19 @@ export default async function CompareIndexPageZh() {
                       return (
                         <ComparePairCardLink
                           key={slug}
-                          href={`/zh/compare/${slug}`}
+                          href={comparisonPairHref('zh', slug, model)}
                           slug={slug}
                           label={label}
                           archLine={archLine}
+                          scenarioLabel={scenario.label}
+                          hardwareA={{
+                            label: aMeta?.label ?? a.toUpperCase(),
+                            vendor: aMeta?.vendor,
+                          }}
+                          hardwareB={{
+                            label: bMeta?.label ?? b.toUpperCase(),
+                            vendor: bMeta?.vendor,
+                          }}
                         />
                       );
                     })}
diff --git a/packages/app/src/app/zh/datasets/page.tsx b/packages/app/src/app/zh/datasets/page.tsx
deleted file mode 100644
index e224ff871..000000000
--- a/packages/app/src/app/zh/datasets/page.tsx
+++ /dev/null
@@ -1,93 +0,0 @@
-import type { Metadata } from 'next';
-
-import { Card } from '@/components/ui/card';
-import { JsonLd } from '@/components/json-ld';
-import { DatasetList } from '@/components/datasets/dataset-list';
-import { zhAlternates, ZH_OG_LOCALE, ZH_LANG_TAG } from '@/lib/i18n';
-import { SITE_URL } from '@semianalysisai/inferencex-constants';
-
-const DESCRIPTION =
-  'InferenceX agentic 基准测试所回放的真实 Claude Code 对话 trace——方法论、分布及逐对话火焰图。';
-
-export const metadata: Metadata = {
-  title: 'Agentic 数据集',
-  description: DESCRIPTION,
-  alternates: zhAlternates('/datasets'),
-  openGraph: {
-    title: 'Agentic 数据集 | InferenceX',
-    description: DESCRIPTION,
-    url: `${SITE_URL}/zh/datasets`,
-    locale: ZH_OG_LOCALE,
-  },
-  twitter: { title: 'Agentic 数据集 | InferenceX', description: DESCRIPTION },
-};
-
-const jsonLd = {
-  '@context': 'https://schema.org',
-  '@type': 'CollectionPage',
-  name: 'InferenceX Agentic 数据集',
-  description: DESCRIPTION,
-  url: `${SITE_URL}/zh/datasets`,
-  inLanguage: ZH_LANG_TAG,
-};
-
-export default function DatasetsPageZh() {
-  return (
-    <main className="relative">
-      <JsonLd data={jsonLd} />
-      <div className="container mx-auto flex flex-col gap-6 px-4 pb-8 lg:px-8">
-        <section>
-          <Card>
-            <h1 className="mb-2 text-xl font-semibold text-foreground">Agentic 基准测试数据集</h1>
-            <p className="mb-3 text-sm text-muted-foreground">
-              InferenceX 的 agentic 基准测试并非回放合成 prompt——而是回放真实的 Claude Code
-              编码会话，以<strong>对话 trace</strong>
-              的形式捕获。每条 trace 是一次完整的多轮会话：包括主 agent 的各轮对话及其调用的所有
-              subagent，附带每轮的 input/output token 数以及重建 prefix-cache 复用所需的 64-token
-              KV-cache block hash。这些 trace 在 HuggingFace 上以{' '}
-              <code>semianalysisai/cc-traces-weka-*</code> 公开发布（apache-2.0 协议）。
-            </p>
-
-            <h2 className="mb-1.5 mt-4 text-sm font-semibold text-foreground">Trace 的采集方式</h2>
-            <p className="mb-3 text-sm text-muted-foreground">
-              生产环境中的 Claude Code 会话通过日志代理录制，该代理捕获每个 API 请求的 input 和
-              output token 数、使用的模型、时间指标（TTFT、token 间延迟），以及一组{' '}
-              <code>hash_ids</code>（每个对应请求 input 的一个 64-token KV block）。Subagent
-              调用被归组到其父轮次下。不存储任何 prompt 或 completion 文本——仅保存 token 计数和
-              block hash，因此语料库可共享，同时仍然是忠实的工作负载回放。
-            </p>
-
-            <h2 className="mb-1.5 mt-4 text-sm font-semibold text-foreground">
-              缓存前缀与未缓存后缀
-            </h2>
-            <p className="mb-3 text-sm text-muted-foreground">
-              Agentic 工作负载以 prefix 复用为主：每轮都会重新发送不断增长的对话，因此大部分 input
-              已在前几轮的 KV cache 中。我们精确重建了这一过程。在理想化的无限 cache
-              下按顺序遍历对话，某一轮的<strong>缓存前缀</strong>是其 <code>hash_ids</code>{' '}
-              中已出现过的最长前导序列；其余部分是需要（重新）计算的<strong>未缓存后缀</strong>
-              。每个 block 为 64 个 token；拆分时会限制使缓存 + 未缓存等于该轮的有效
-              input，即使最后一个 block 不完整。Subagent 在 spawn 时针对父 cache
-              的快照运行（其上下文独立，不会合并回父级）。
-            </p>
-
-            <h2 className="mb-1.5 mt-4 text-sm font-semibold text-foreground">数据集变体</h2>
-            <ul className="mb-1 list-disc space-y-1 pl-5 text-sm text-muted-foreground">
-              <li>
-                <strong>full</strong> — 所有捕获的请求，不做修改。
-              </li>
-              <li>
-                <strong>256k</strong> — 丢弃 input + output 超过 256,000 token 的请求，确保每轮都在
-                256k 上下文窗口内（用于在配置 256k 最大上下文的引擎上进行基准测试）。
-              </li>
-            </ul>
-          </Card>
-        </section>
-
-        <section className="flex flex-col gap-3">
-          <h2 className="text-lg font-semibold text-foreground">数据集</h2>
-          <DatasetList />
-        </section>
-      </div>
-    </main>
-  );
-}
diff --git a/packages/app/src/app/zh/glossary/[slug]/page.tsx b/packages/app/src/app/zh/glossary/[slug]/page.tsx
index d66c1e247..3cc736db0 100644
--- a/packages/app/src/app/zh/glossary/[slug]/page.tsx
+++ b/packages/app/src/app/zh/glossary/[slug]/page.tsx
@@ -146,16 +146,16 @@ export default async function ZhGlossaryTermPage({ params }: Props) {
                 AI 推理术语表
               </Link>
               <div className="mt-8 flex flex-wrap items-center gap-2">
-                <span className="rounded-full border border-brand/25 bg-brand/8 px-3 py-1 text-xs font-semibold tracking-[0.14em] text-brand uppercase">
+                <span className="rounded-full border border-brand/25 bg-brand/8 px-3 py-1 text-xs font-semibold tracking-eyebrow text-brand uppercase">
                   {GLOSSARY_CATEGORY_LABELS_ZH[entry.category]}
                 </span>
                 {entry.abbreviation && (
-                  <span className="font-mono text-xs tracking-[0.16em] text-muted-foreground uppercase">
+                  <span className="font-mono text-xs tracking-eyebrow text-muted-foreground uppercase">
                     {entry.abbreviation}
                   </span>
                 )}
               </div>
-              <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-[-0.035em] text-balance md:text-6xl">
+              <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-6xl">
                 {entry.term}
               </h1>
               {entry.aliases && entry.aliases.length > 0 && (
@@ -173,7 +173,7 @@ export default async function ZhGlossaryTermPage({ params }: Props) {
                 >
                   <p
                     id="plain-language-heading"
-                    className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase"
+                    className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase"
                   >
                     先用大白话
                   </p>
@@ -194,7 +194,7 @@ export default async function ZhGlossaryTermPage({ params }: Props) {
 
                 {entry.measurement && (
                   <div className="mt-8 border-l-2 border-brand bg-brand/6 px-5 py-4">
-                    <p className="text-xs font-semibold tracking-[0.14em] text-muted-foreground uppercase">
+                    <p className="text-xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
                       {entry.measurement.label}
                     </p>
                     <p className="mt-1 font-mono text-sm leading-relaxed text-foreground">
@@ -226,7 +226,7 @@ export default async function ZhGlossaryTermPage({ params }: Props) {
               </div>
 
               <aside className="border-t border-border/50 bg-muted/10 px-5 py-8 lg:border-t-0 lg:border-l lg:px-6 lg:py-12">
-                <p className="font-mono text-xs font-semibold tracking-[0.16em] text-muted-foreground uppercase">
+                <p className="font-mono text-xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
                   相关术语
                 </p>
                 <nav aria-label="相关术语" className="mt-4 flex flex-col">
@@ -257,7 +257,7 @@ export default async function ZhGlossaryTermPage({ params }: Props) {
               <Card>
                 <div className="flex flex-col gap-2 border-b border-border/50 pb-5 md:flex-row md:items-end md:justify-between">
                   <div>
-                    <p className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase">
+                    <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
                       参考资料
                     </p>
                     <h2 id="further-reading" className="mt-2 text-2xl font-semibold tracking-tight">
@@ -302,7 +302,7 @@ export default async function ZhGlossaryTermPage({ params }: Props) {
                 href={`/zh/glossary/${adjacent.previous.slug}`}
                 className="rounded-xl border border-border/40 bg-background/20 p-5 backdrop-blur-[2px] transition-colors hover:border-brand/40 hover:bg-brand/5"
               >
-                <span className="text-xs tracking-[0.14em] text-muted-foreground uppercase">
+                <span className="text-xs tracking-eyebrow text-muted-foreground uppercase">
                   ← 上一条
                 </span>
                 <span className="mt-2 block font-semibold">{adjacent.previous.term}</span>
@@ -315,7 +315,7 @@ export default async function ZhGlossaryTermPage({ params }: Props) {
                 href={`/zh/glossary/${adjacent.next.slug}`}
                 className="rounded-xl border border-border/40 bg-background/20 p-5 text-right backdrop-blur-[2px] transition-colors hover:border-brand/40 hover:bg-brand/5"
               >
-                <span className="text-xs tracking-[0.14em] text-muted-foreground uppercase">
+                <span className="text-xs tracking-eyebrow text-muted-foreground uppercase">
                   下一条 →
                 </span>
                 <span className="mt-2 block font-semibold">{adjacent.next.term}</span>
diff --git a/packages/app/src/app/zh/glossary/page.tsx b/packages/app/src/app/zh/glossary/page.tsx
index b3d772d2e..38495d622 100644
--- a/packages/app/src/app/zh/glossary/page.tsx
+++ b/packages/app/src/app/zh/glossary/page.tsx
@@ -20,7 +20,7 @@ import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-con
 
 const title = 'AI 推理术语表';
 const description =
-  '清晰、技术严谨的 LLM 推理基准、服务指标、分布式并行、数值精度、Chip 硬件与推理软件术语定义。';
+  '清晰、技术严谨的智能体推理、LLM 基准测试、服务指标、分布式并行、数值精度、芯片硬件与推理软件术语定义。';
 const browserLabels: GlossaryBrowserLabels = {
   searchLabel: '搜索 AI 推理术语表',
   searchPlaceholder: '搜索 MTP、延迟、FP4…',
@@ -40,7 +40,15 @@ const browserLabels: GlossaryBrowserLabels = {
 export const metadata: Metadata = {
   title,
   description,
-  keywords: ['AI 推理术语表', 'LLM 推理术语', 'GPU 基准术语', '分布式推理', 'LLM 性能指标'],
+  keywords: [
+    'AI 推理术语表',
+    '智能体推理术语',
+    'AgentX 基准测试术语',
+    'LLM 推理术语',
+    '芯片基准术语',
+    '分布式推理',
+    'LLM 性能指标',
+  ],
   alternates: zhAlternates('/glossary'),
   openGraph: {
     title: `${title} | ${SITE_NAME}`,
@@ -113,27 +121,27 @@ export default function ZhGlossaryPage() {
             />
             <div className="grid gap-10 lg:grid-cols-[minmax(0,1fr)_18rem] lg:items-end">
               <div>
-                <p className="font-mono text-xs font-semibold tracking-[0.2em] text-brand uppercase">
+                <p className="font-mono text-xs font-semibold tracking-eyebrow-wide text-brand uppercase">
                   技术指南 / AI 基础设施
                 </p>
-                <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-[-0.045em] text-balance md:text-6xl lg:text-7xl">
+                <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-6xl lg:text-7xl">
                   读懂推理曲线背后的语言。
                 </h1>
                 <p className="mt-6 max-w-3xl text-base leading-7 text-muted-foreground md:text-lg md:leading-8">
-                  解释 InferenceX
-                  使用的性能指标、服务技术、数值格式和分布式系统概念。所有定义都来自实测行为，而非厂商峰值规格。
+                  解释 InferenceX 使用的智能体推理、基准指标、服务技术、数值格式与分布式系统概念。
+                  各条目以实测行为和已发布的基准测试方案为依据。
                 </p>
               </div>
 
               <dl className="grid grid-cols-3 gap-px overflow-hidden rounded-lg border border-border/50 bg-border/50 lg:grid-cols-1">
                 <div className="bg-background/70 p-4">
-                  <dt className="font-mono text-[0.65rem] tracking-[0.16em] text-muted-foreground uppercase">
+                  <dt className="font-mono text-3xs tracking-eyebrow text-muted-foreground uppercase">
                     术语
                   </dt>
                   <dd className="mt-1 text-2xl font-semibold tabular-nums">{entries.length}</dd>
                 </div>
                 <div className="bg-background/70 p-4">
-                  <dt className="font-mono text-[0.65rem] tracking-[0.16em] text-muted-foreground uppercase">
+                  <dt className="font-mono text-3xs tracking-eyebrow text-muted-foreground uppercase">
                     类别
                   </dt>
                   <dd className="mt-1 text-2xl font-semibold tabular-nums">
@@ -141,7 +149,7 @@ export default function ZhGlossaryPage() {
                   </dd>
                 </div>
                 <div className="bg-background/70 p-4">
-                  <dt className="font-mono text-[0.65rem] tracking-[0.16em] text-muted-foreground uppercase">
+                  <dt className="font-mono text-3xs tracking-eyebrow text-muted-foreground uppercase">
                     参考文章
                   </dt>
                   <dd className="mt-1 text-2xl font-semibold tabular-nums">{articleCount}</dd>
@@ -162,7 +170,7 @@ export default function ZhGlossaryPage() {
 
         <section className="mt-4 grid gap-4 md:grid-cols-2">
           <Card>
-            <p className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase">
+            <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
               阅读基准曲线
             </p>
             <h2 className="mt-3 text-2xl font-semibold tracking-tight">完整曲线更能说明问题。</h2>
@@ -184,7 +192,7 @@ export default function ZhGlossaryPage() {
           </Card>
 
           <Card>
-            <p className="font-mono text-xs font-semibold tracking-[0.18em] text-brand uppercase">
+            <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
               基于实测数据
             </p>
             <h2 className="mt-3 text-2xl font-semibold tracking-tight">每个定义都连接真实方案。</h2>
@@ -193,6 +201,9 @@ export default function ZhGlossaryPage() {
               Wide EP 扩展，以及硬件不变时的软件性能提升。
             </p>
             <div className="mt-5 flex flex-wrap gap-3 text-sm font-medium">
+              <Link href="/zh/agentx/methodology" className="text-brand hover:underline">
+                AgentX 方法说明 →
+              </Link>
               <Link href="/zh/blog" className="text-brand hover:underline">
                 浏览技术文章 →
               </Link>
diff --git a/packages/app/src/app/zh/land-acknowledgement/page.tsx b/packages/app/src/app/zh/land-acknowledgement/page.tsx
index f92117001..ddfcffc27 100644
--- a/packages/app/src/app/zh/land-acknowledgement/page.tsx
+++ b/packages/app/src/app/zh/land-acknowledgement/page.tsx
@@ -53,7 +53,7 @@ export default function LandAcknowledgementPageZh() {
             <p className="mb-3 text-xs font-semibold uppercase tracking-[0.32em] text-brand">
               土地致谢
             </p>
-            <h1 className="text-4xl font-semibold tracking-[-0.04em] text-foreground md:text-5xl">
+            <h1 className="text-4xl font-semibold tracking-heading text-foreground md:text-5xl">
               我们致敬与我们美国基础设施所在土地相关的原住民家园。
             </h1>
             <p className="mt-4 text-sm leading-6 text-muted-foreground md:text-base">
@@ -79,7 +79,7 @@ export default function LandAcknowledgementPageZh() {
                 <p className="text-xs font-semibold uppercase tracking-[0.28em] text-muted-foreground">
                   {entry.region}
                 </p>
-                <h2 className="mt-3 text-xl font-semibold tracking-[-0.04em] text-foreground">
+                <h2 className="mt-3 text-xl font-semibold tracking-heading text-foreground">
                   {entry.peoples}
                 </h2>
                 <p className="mt-4 text-sm leading-6 text-muted-foreground">
diff --git a/packages/app/src/app/zh/overview/page.tsx b/packages/app/src/app/zh/overview/page.tsx
index 524d15db1..3afa91d07 100644
--- a/packages/app/src/app/zh/overview/page.tsx
+++ b/packages/app/src/app/zh/overview/page.tsx
@@ -1,53 +1,13 @@
-import type { Metadata } from 'next';
-
-import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
-
-import { OverviewPageContent } from '@/components/overview/overview-page';
-import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 import {
-  resolveOverviewComparisonMode,
-  resolveOverviewEngineScope,
-  resolveOverviewModelScope,
-  resolveOverviewReferenceHardware,
-  resolveOverviewTier,
-} from '@/lib/overview-data';
-import { getOverviewPageData } from '@/lib/overview-data.server';
+  buildOverviewMetadata,
+  type OverviewRoutePageProps,
+  renderOverviewPage,
+} from '@/lib/overview-route.server';
 
 export const dynamic = 'force-dynamic';
 
-const DESCRIPTION =
-  '按各活跃模型标注的场景，对比 MI355X、B200、B300、GB200 与 GB300 的每百万总 token 超大规模云成本。';
-
-export const metadata: Metadata = {
-  title: '推理成本总览',
-  description: DESCRIPTION,
-  alternates: zhAlternates('/overview'),
-  openGraph: {
-    title: `推理成本总览 | ${SITE_NAME}`,
-    description: DESCRIPTION,
-    url: `${SITE_URL}/zh/overview`,
-    type: 'website',
-    locale: ZH_OG_LOCALE,
-  },
-  twitter: {
-    card: 'summary_large_image',
-    title: `推理成本总览 | ${SITE_NAME}`,
-    description: DESCRIPTION,
-  },
-};
-
-interface Props {
-  searchParams: Promise<Record<string, string | string[] | undefined>>;
-}
+export const metadata = buildOverviewMetadata('zh');
 
-export default async function ZhOverviewPage({ searchParams }: Props) {
-  const sp = await searchParams;
-  const data = await getOverviewPageData(
-    resolveOverviewTier(sp.tier),
-    resolveOverviewEngineScope(sp.engine),
-    resolveOverviewComparisonMode(sp.compare),
-    resolveOverviewReferenceHardware(sp.ref),
-    resolveOverviewModelScope(sp.models),
-  );
-  return <OverviewPageContent data={data} locale="zh" />;
+export default function ZhOverviewPage({ searchParams }: OverviewRoutePageProps) {
+  return renderOverviewPage({ locale: 'zh', searchParams });
 }
diff --git a/packages/app/src/app/zh/quotes/page.tsx b/packages/app/src/app/zh/quotes/page.tsx
index b13e3dab2..8f0916ac7 100644
--- a/packages/app/src/app/zh/quotes/page.tsx
+++ b/packages/app/src/app/zh/quotes/page.tsx
@@ -5,14 +5,14 @@ import { ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
 import { SITE_URL } from '@semianalysisai/inferencex-constants';
 
 export const metadata: Metadata = {
-  title: '支持者',
+  title: '业界与社区评价',
   description:
-    'InferenceX 计划获得众多主要算力买家与 ML 社区知名成员的支持，包括来自 MiniMax、Moonshot Kimi、阿里巴巴 Qwen、OpenAI、Microsoft、vLLM、PyTorch 基金会、Oracle 等机构的支持者。',
+    'InferenceX 获得了众多大型算力采购方和 ML 社区知名人士的认可。以下评价来自 MiniMax、Moonshot Kimi、Alibaba Qwen、OpenAI、Microsoft、vLLM、PyTorch Foundation 和 Oracle 等机构的相关人士。',
   alternates: zhAlternates('/quotes'),
   openGraph: {
-    title: '支持者 | InferenceX by SemiAnalysis',
+    title: '业界与社区评价 | InferenceX by SemiAnalysis',
     description:
-      '获得 MiniMax、Moonshot Kimi、阿里巴巴 Qwen、OpenAI、Microsoft、vLLM、PyTorch 基金会、Oracle 及 ML 社区知名成员的支持。',
+      'InferenceX 获得了众多大型算力采购方和 ML 社区知名人士的认可。以下评价来自 MiniMax、Moonshot Kimi、Alibaba Qwen、OpenAI、Microsoft、vLLM、PyTorch Foundation 和 Oracle 等机构的相关人士。',
     url: `${SITE_URL}/zh/quotes`,
     locale: ZH_OG_LOCALE,
   },
diff --git a/packages/app/src/app/zh/rankings/[slug]/page.tsx b/packages/app/src/app/zh/rankings/[slug]/page.tsx
new file mode 100644
index 000000000..4c843f09a
--- /dev/null
+++ b/packages/app/src/app/zh/rankings/[slug]/page.tsx
@@ -0,0 +1,203 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { fmtCostPerMtok, fmtThroughput } from '@/components/live-seo/format';
+import {
+  RankingsDetailContent,
+  type RankingsStrings,
+} from '@/components/live-seo/rankings-page-sections';
+import { JsonLd } from '@/components/json-ld';
+import { ZH_LANG_TAG, ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { getRankingPageEntry, scenarioLabel, type RankingPageEntry } from '@/lib/rankings';
+import {
+  rankingPageDescriptionZh,
+  rankingPageHeadingZh,
+  rankingPageKeywordsZh,
+  rankingPageTitleZh,
+} from '@/lib/rankings-zh';
+import { getRankingPageData, type RankingPageData } from '@/lib/run-rankings-data.server';
+
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string }>;
+}
+
+function statLedDescriptionZh(entry: RankingPageEntry, data: RankingPageData): string {
+  const [first] = data.rows;
+  if (!first) return rankingPageDescriptionZh(entry);
+  if (entry.kind === 'fastest-gpu' && first.throughputPerGpu !== null) {
+    return `${first.hardwareLabel} 目前以单 GPU 每秒 ${fmtThroughput(first.throughputPerGpu)} token 领跑 ${entry.model.seoName} 推理。共 ${data.rows.length} 个平台实时排名，数据持续更新。`;
+  }
+  if (entry.kind === 'cheapest-gpu' && first.costPerMtok !== null) {
+    return `${first.hardwareLabel} 目前以每百万 token ${fmtCostPerMtok(first.costPerMtok)} 的成本领跑 ${entry.model.seoName} 推理。共 ${data.rows.length} 个平台实时排名，数据持续更新。`;
+  }
+  return rankingPageDescriptionZh(entry);
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug } = await params;
+  const entry = getRankingPageEntry(slug);
+  if (!entry) return {};
+  const data = await getRankingPageData(entry.slug);
+  const title = rankingPageTitleZh(entry);
+  const description = data ? statLedDescriptionZh(entry, data) : rankingPageDescriptionZh(entry);
+  const url = `${SITE_URL}/zh/rankings/${entry.slug}`;
+  return {
+    title: { absolute: `${title} | ${SITE_NAME}` },
+    description,
+    keywords: rankingPageKeywordsZh(entry),
+    authors: [{ name: AUTHOR_NAME }],
+    alternates: zhAlternates(`/rankings/${entry.slug}`),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+      locale: ZH_OG_LOCALE,
+      type: 'article',
+    },
+    twitter: { card: 'summary_large_image', title: `${title} | ${SITE_NAME}`, description },
+  };
+}
+
+function buildFaqZh(
+  entry: RankingPageEntry,
+  data: RankingPageData,
+): { question: string; answer: string }[] {
+  const model = entry.model.seoName;
+  const [first] = data.rows;
+  const workload = scenarioLabel(data.scenario, 'zh');
+  const fallbackLeader = `排行数据来自 InferenceX 持续运行的基准测试，最新领先者见上方表格。`;
+  const kindFaq =
+    entry.kind === 'fastest-gpu'
+      ? [
+          {
+            question: `${model} 推理最快的 GPU 是哪款？`,
+            answer:
+              first && typeof first.throughputPerGpu === 'number'
+                ? `按最新基准测试结果，${first.hardwareLabel} 在${workload}下以单 GPU 每秒 ${fmtThroughput(first.throughputPerGpu)} token 领先，交互速度统一设定为每用户每秒 ${data.tier} token。`
+                : fallbackLeader,
+          },
+          {
+            question: `“最快”是如何衡量的？`,
+            answer: `所有平台都在相同的交互速度档位（每用户每秒 ${data.tier} token）和${workload}下读取，再按实测单 GPU 吞吐排名。推导方式与 InferenceX 总览排行榜完全一致，因此本页排名不会与仪表盘出现分歧。`,
+          },
+        ]
+      : [
+          {
+            question: `运行 ${model} 最省钱的 GPU 是哪款？`,
+            answer:
+              first && typeof first.costPerMtok === 'number'
+                ? `按最新基准测试结果，${first.hardwareLabel} 在${workload}下成本最低，每百万 token（输入加输出）仅 ${fmtCostPerMtok(first.costPerMtok)}，按超大规模云 $/GPU/小时 价格计算，交互速度统一设定为每用户每秒 ${data.tier} token。`
+                : fallbackLeader,
+          },
+          {
+            question: `每百万 token 成本是如何计算的？`,
+            answer: `将每用户每秒 ${data.tier} token 档位下的实测单 GPU 吞吐，按 SemiAnalysis AI Cloud TCO 模型中的超大规模云 $/GPU/小时 价格换算为每百万 token（输入加输出）成本。更慢的交互档位或更便宜的租用价格会改变绝对数值，但很少改变排名顺序。`,
+          },
+        ];
+  return [
+    ...kindFaq,
+    {
+      question: `排行多久更新一次？`,
+      answer: data.newest
+        ? `基准测试在 InferenceX 集群上持续运行，本排行最新一条结果落在 ${data.newest}。页面始终渲染最新数据。`
+        : `基准测试在 InferenceX 集群上持续运行，页面始终渲染最新数据。`,
+    },
+  ];
+}
+
+export default async function ZhRankingPage({ params }: Props) {
+  const { slug } = await params;
+  const entry = getRankingPageEntry(slug);
+  if (!entry) notFound();
+  const data = await getRankingPageData(entry.slug);
+  if (!data) notFound();
+
+  const url = `${SITE_URL}/zh/rankings/${entry.slug}`;
+  const heading = rankingPageHeadingZh(entry);
+  const faq = buildFaqZh(entry, data);
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@graph': [
+      {
+        '@type': 'ItemList',
+        name: rankingPageTitleZh(entry),
+        inLanguage: ZH_LANG_TAG,
+        itemListOrder: 'https://schema.org/ItemListOrderAscending',
+        numberOfItems: data.rows.length,
+        itemListElement: data.rows.map((row) => ({
+          '@type': 'ListItem',
+          position: row.rank,
+          name: row.hardwareLabel,
+          ...(row.chip && { url: `${SITE_URL}/zh/chips/${row.chip.slug}` }),
+        })),
+      },
+      {
+        '@type': 'FAQPage',
+        inLanguage: ZH_LANG_TAG,
+        mainEntity: faq.map((item) => ({
+          '@type': 'Question',
+          name: item.question,
+          acceptedAnswer: { '@type': 'Answer', text: item.answer },
+        })),
+      },
+      {
+        '@type': 'BreadcrumbList',
+        itemListElement: [
+          {
+            '@type': 'ListItem',
+            position: 1,
+            name: 'GPU 排行榜',
+            item: `${SITE_URL}/zh/rankings`,
+          },
+          { '@type': 'ListItem', position: 2, name: heading, item: url },
+        ],
+      },
+    ],
+  };
+
+  const t: RankingsStrings = {
+    backLabel: 'GPU 排行榜',
+    heading,
+    scenarioNote: `测试基于${scenarioLabel(data.scenario, 'zh')}，交互速度统一设定为每用户每秒 ${data.tier} token。在该工作点没有实测数据的硬件不参与排名。`,
+    tableCaption: `${heading}：实时基准排行`,
+    colRank: '排名',
+    colGpu: 'GPU',
+    colThroughput: '单 GPU 每秒 token 数',
+    colCost: '每百万 token 成本',
+    colPrecision: '精度',
+    colEngine: '推理引擎',
+    emptyState: `${entry.model.seoName} 在该工作点暂无实测数据。InferenceX 集群持续运行基准测试，新结果落地后本页会自动填充。`,
+    siblingLead: entry.kind === 'fastest-gpu' ? '更关注成本而不是速度？' : '更关注速度而不是成本？',
+    siblingLabel:
+      entry.kind === 'fastest-gpu'
+        ? `查看运行 ${entry.model.seoName} 最省钱的 GPU`
+        : `查看 ${entry.model.seoName} 推理最快的 GPU`,
+    methodologyHeading: '测试方法',
+    methodologyBody: [
+      `本页所有数字都来自实测，而非纸面规格估算。InferenceX 集群使用社区推理引擎在真实硬件上部署 ${entry.model.seoName}，通过扫描并发数绘制每个平台的吞吐与交互速度前沿曲线。`,
+      `所有平台在同一工作点（每用户每秒 ${data.tier} token）读取，保证对比是等交互速度的：任何 GPU 都无法靠牺牲单用户速度换取纸面吞吐来取胜。成本按 SemiAnalysis AI Cloud TCO 模型的 $/GPU/小时 价格，把实测吞吐换算为每百万 token（输入加输出）成本。`,
+      `推导逻辑与 InferenceX 总览排行榜共用，基准测试持续重跑，新引擎版本和新配置落地后排行会自动更新。`,
+    ],
+    faqHeading: '常见问题',
+    faq,
+    exploreHeading: '继续探索数据',
+    exploreLinks: [
+      { href: '/zh/inference', label: '交互式推理仪表盘' },
+      { href: '/zh/overview', label: '跨模型推理排行榜' },
+      { href: '/zh/compare', label: 'GPU 一对一对比' },
+      { href: '/zh/chips', label: 'GPU 规格与价格' },
+    ],
+  };
+
+  return (
+    <>
+      <JsonLd data={jsonLd} />
+      <RankingsDetailContent entry={entry} data={data} t={t} locale="zh" />
+    </>
+  );
+}
diff --git a/packages/app/src/app/zh/rankings/page.tsx b/packages/app/src/app/zh/rankings/page.tsx
new file mode 100644
index 000000000..00e91df05
--- /dev/null
+++ b/packages/app/src/app/zh/rankings/page.tsx
@@ -0,0 +1,102 @@
+import type { Metadata } from 'next';
+import Link from 'next/link';
+
+import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { JsonLd } from '@/components/json-ld';
+import { Card } from '@/components/ui/card';
+import { ZH_LANG_TAG, ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
+import { getAllRankingPageEntries } from '@/lib/rankings';
+import { rankingPageHeadingZh } from '@/lib/rankings-zh';
+
+const title = 'LLM 推理 GPU 排行榜';
+const description =
+  'DeepSeek、Kimi、GLM、MiniMax、Qwen 等开源大模型推理的最快与最省钱 GPU 实时排行。按实测单 GPU 每秒 token 数与每百万 token 成本排名，数据持续更新。';
+
+export const metadata: Metadata = {
+  title,
+  description,
+  keywords: [
+    'LLM 推理最快 GPU',
+    'LLM 推理最便宜 GPU',
+    'GPU 推理排行榜',
+    'AI 推理最佳 GPU',
+    'LLM 推理 token 成本',
+    'H100 B200 MI355X 对比',
+    '单 GPU 每秒 token 数',
+    'GPU 推理排名',
+  ],
+  alternates: zhAlternates('/rankings'),
+  openGraph: {
+    title: `${title} | ${SITE_NAME}`,
+    description,
+    url: `${SITE_URL}/zh/rankings`,
+    locale: ZH_OG_LOCALE,
+    type: 'website',
+  },
+  twitter: { card: 'summary_large_image', title, description },
+};
+
+export default function ZhRankingsIndexPage() {
+  const entries = getAllRankingPageEntries();
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'CollectionPage',
+    name: title,
+    description,
+    url: `${SITE_URL}/zh/rankings`,
+    inLanguage: ZH_LANG_TAG,
+    hasPart: entries.map((entry) => ({
+      '@type': 'WebPage',
+      name: rankingPageHeadingZh(entry),
+      url: `${SITE_URL}/zh/rankings/${entry.slug}`,
+    })),
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8">
+        <div className="mx-auto max-w-5xl">
+          <header className="pt-8 md:pt-12">
+            <h1 className="max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-5xl">
+              {title}
+            </h1>
+            <p className="mt-4 max-w-3xl text-lg leading-relaxed text-muted-foreground">
+              哪款 GPU
+              最适合部署你的模型？以下每个排行都来自匹配交互速度下的实测基准数据，而非纸面规格，并随新结果自动更新。
+            </p>
+          </header>
+
+          <section className="mt-10 mb-16 grid gap-4 sm:grid-cols-2">
+            {INFERENCE_MODEL_SLUGS.map((model) => (
+              <Card key={model.slug} className="p-5">
+                <h2 className="text-lg font-semibold tracking-tight">{model.seoName}</h2>
+                <ul className="mt-3 space-y-1">
+                  <li>
+                    <Link
+                      href={`/zh/rankings/fastest-gpu-for-${model.slug}`}
+                      className="text-sm font-medium text-brand hover:underline"
+                    >
+                      {model.seoName} 最快 GPU 排行
+                    </Link>
+                  </li>
+                  <li>
+                    <Link
+                      href={`/zh/rankings/cheapest-gpu-for-${model.slug}`}
+                      className="text-sm font-medium text-brand hover:underline"
+                    >
+                      {model.seoName} 最低成本 GPU 排行
+                    </Link>
+                  </li>
+                </ul>
+              </Card>
+            ))}
+          </section>
+        </div>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/app/zh/run/[slug]/page.tsx b/packages/app/src/app/zh/run/[slug]/page.tsx
new file mode 100644
index 000000000..f45e9c419
--- /dev/null
+++ b/packages/app/src/app/zh/run/[slug]/page.tsx
@@ -0,0 +1,221 @@
+import type { Metadata } from 'next';
+import { notFound } from 'next/navigation';
+
+import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { fmtCostPerMtok, fmtThroughput } from '@/components/live-seo/format';
+import {
+  latencyQuote,
+  RunDetailContent,
+  type RunStrings,
+} from '@/components/live-seo/run-page-sections';
+import { JsonLd } from '@/components/json-ld';
+import { ZH_LANG_TAG, ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { scenarioLabel } from '@/lib/rankings';
+import { getRunPageEntry, type RunPageEntry } from '@/lib/run-pages';
+import {
+  runPageDescriptionZh,
+  runPageFaqQuestionsZh,
+  runPageHeadingZh,
+  runPageTitleZh,
+  runPageKeywordsZh,
+} from '@/lib/run-pages-zh';
+import { getRunPageData, type RunPageData } from '@/lib/run-rankings-data.server';
+
+export const dynamic = 'force-dynamic';
+
+interface Props {
+  params: Promise<{ slug: string }>;
+}
+
+function primaryRead(data: RunPageData) {
+  return data.tierLadder.find((read) => read.tier === data.primaryTier);
+}
+
+function statLedDescriptionZh(entry: RunPageEntry, data: RunPageData): string {
+  const read = primaryRead(data);
+  if (!data.hasData || !read?.throughputPerGpu) return runPageDescriptionZh(entry);
+  const cost =
+    read.costPerMtok === null
+      ? ''
+      : `，按超大规模云价格每百万 token 成本 ${fmtCostPerMtok(read.costPerMtok)}`;
+  const chipLabel = entry.chip.label;
+  return `实测数据：${entry.model.seoName} 在 ${chipLabel} 上、每用户每秒 ${data.primaryTier} token 档位下单 GPU 可持续输出 ${fmtThroughput(read.throughputPerGpu)} token/s${cost}。共 ${data.configCount} 组实测配置，数据持续更新。`;
+}
+
+export async function generateMetadata({ params }: Props): Promise<Metadata> {
+  const { slug } = await params;
+  const entry = getRunPageEntry(slug);
+  if (!entry) return {};
+  const data = await getRunPageData(entry.slug);
+  const title = runPageTitleZh(entry);
+  const description = data ? statLedDescriptionZh(entry, data) : runPageDescriptionZh(entry);
+  const url = `${SITE_URL}/zh/run/${entry.slug}`;
+  return {
+    title: { absolute: `${title} | ${SITE_NAME}` },
+    description,
+    keywords: runPageKeywordsZh(entry),
+    authors: [{ name: AUTHOR_NAME }],
+    // Valid slugs without benchmark data render an empty state; keep those
+    // out of the index (the sitemap already excludes them).
+    ...(data?.hasData ? {} : { robots: { index: false, follow: true } }),
+    alternates: zhAlternates(`/run/${entry.slug}`),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+      locale: ZH_OG_LOCALE,
+      type: 'article',
+    },
+    twitter: { card: 'summary_large_image', title: `${title} | ${SITE_NAME}`, description },
+  };
+}
+
+function buildFaqZh(
+  entry: RunPageEntry,
+  data: RunPageData,
+): { question: string; answer: string }[] {
+  const questions = runPageFaqQuestionsZh(entry);
+  const model = entry.model.seoName;
+  const chipLabel = entry.chip.label;
+  const read = primaryRead(data);
+  const workload = scenarioLabel(data.scenario, 'zh');
+
+  const peak =
+    data.bestThroughputPerGpu === null ? '暂无' : fmtThroughput(data.bestThroughputPerGpu);
+  const throughputAnswer =
+    typeof read?.throughputPerGpu === 'number'
+      ? `在${workload}、每用户每秒 ${data.primaryTier} token 的交互档位下，${chipLabel} 部署 ${model} 单 GPU 可持续输出 ${fmtThroughput(read.throughputPerGpu)} token/s${read.framework ? `，推理引擎为 ${read.framework}` : ''}${read.precision ? `，精度 ${read.precision.toUpperCase()}` : ''}。全部配置中的实测峰值吞吐为单 GPU ${peak} token/s。`
+      : `InferenceX 集群已为该组合完成 ${data.configCount} 组实测配置，各交互档位的实测结果见上方阶梯表。`;
+
+  const costAnswer =
+    typeof read?.costPerMtok === 'number'
+      ? `按超大规模云 $/GPU/小时 价格、每用户每秒 ${data.primaryTier} token 档位计算，每百万 token（输入加输出）成本为 ${fmtCostPerMtok(read.costPerMtok)}。新兴云与零售租用价格见上方表格；更慢的交互档位成本更低。`
+      : `每百万 token 成本由实测吞吐和 SemiAnalysis AI Cloud TCO 模型的 $/GPU/小时 价格换算得出，待该组合达到主交互档位后即会显示。`;
+
+  const servingAnswer = `本页数据来自 ${data.frameworks.join('、')}，精度覆盖 ${data.precisions.map((p) => p.toUpperCase()).join('、')}${data.hasDisagg ? '，包含 prefill 分离部署' : ''}${data.hasMultinode ? '，并包含多节点部署' : ''}。推理引擎持续重新构建并重跑基准，最优配置可能随时变化。`;
+
+  const methodologyAnswer = `所有数字均由 InferenceX 集群在真实 ${chipLabel} 硬件上实测，通过在${workload}下扫描并发数绘制吞吐与交互速度前沿曲线${data.newest ? `；最新一次运行落在 ${data.newest}` : ''}。推导方式与 InferenceX 总览排行榜完全一致。`;
+
+  return [
+    { question: questions.throughput, answer: throughputAnswer },
+    { question: questions.cost, answer: costAnswer },
+    { question: questions.serving, answer: servingAnswer },
+    { question: questions.methodology, answer: methodologyAnswer },
+  ];
+}
+
+export default async function ZhRunPage({ params }: Props) {
+  const { slug } = await params;
+  const entry = getRunPageEntry(slug);
+  if (!entry) notFound();
+  const data = await getRunPageData(entry.slug);
+  if (!data) notFound();
+
+  const url = `${SITE_URL}/zh/run/${entry.slug}`;
+  const heading = runPageHeadingZh(entry);
+  const faq = data.hasData ? buildFaqZh(entry, data) : [];
+  const read = primaryRead(data);
+  const model = entry.model.seoName;
+  const chipLabel = entry.chip.label;
+  const chipTitle = entry.chip.title;
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@graph': [
+      {
+        '@type': 'WebPage',
+        name: runPageTitleZh(entry),
+        url,
+        inLanguage: ZH_LANG_TAG,
+        ...(data.newest && { dateModified: data.newest }),
+        ...(data.oldest && { datePublished: data.oldest }),
+      },
+      ...(faq.length > 0
+        ? [
+            {
+              '@type': 'FAQPage',
+              inLanguage: ZH_LANG_TAG,
+              mainEntity: faq.map((item) => ({
+                '@type': 'Question',
+                name: item.question,
+                acceptedAnswer: { '@type': 'Answer', text: item.answer },
+              })),
+            },
+          ]
+        : []),
+      {
+        '@type': 'BreadcrumbList',
+        itemListElement: [
+          { '@type': 'ListItem', position: 1, name: '模型与 GPU 组合', item: `${SITE_URL}/zh/run` },
+          { '@type': 'ListItem', position: 2, name: heading, item: url },
+        ],
+      },
+    ],
+  };
+
+  const quickAnswerCost =
+    typeof read?.costPerMtok === 'number'
+      ? `，按超大规模云价格折合每百万 token ${fmtCostPerMtok(read.costPerMtok)}`
+      : '';
+  const latency = latencyQuote(data);
+  const quickAnswerLatency = latency
+    ? `全部实测配置中最快 TTFT 为 ${latency.ttft}，最快 TPOT 为 ${latency.tpot}（两者各自取最优，可能来自不同配置）。`
+    : '';
+  const quickAnswer =
+    typeof read?.throughputPerGpu === 'number'
+      ? `${model} 在 ${chipLabel} 上、每用户每秒 ${data.primaryTier} token 档位下单 GPU 可持续输出 ${fmtThroughput(read.throughputPerGpu)} token/s${quickAnswerCost}${read.framework ? `，推理引擎为 ${read.framework}` : ''}。${quickAnswerLatency}`
+      : `${model} 可在 ${chipLabel} 上运行：目前已有 ${data.configCount} 组实测配置，各档位实测结果见下方阶梯表。`;
+
+  const t: RunStrings = {
+    backHref: '/zh/run',
+    backLabel: '全部模型与 GPU 组合',
+    heading,
+    quickAnswerLabel: '快速结论',
+    quickAnswer,
+    statConfigs: '实测配置数',
+    statEngines: '推理引擎',
+    statPrecisions: '精度',
+    statFreshness: '运行日期',
+    ladderHeading: '各交互档位下的吞吐表现',
+    ladderIntro: `推理部署是一种权衡：单 GPU 上并发用户越多，每个用户收到 token 的速度就越慢。下表在${scenarioLabel(data.scenario, 'zh')}下按每个单用户速度目标读取实测前沿，取该工作点上最优的引擎与精度。`,
+    colTier: '单用户速度目标',
+    colThroughput: '单 GPU 每秒 token 数',
+    colCost: '每百万 token 成本',
+    colEngine: '推理引擎',
+    colPrecision: '精度',
+    costHeading: '实际部署成本',
+    costIntro: `将每用户每秒 ${data.primaryTier} token 工作点的实测吞吐，按 SemiAnalysis AI Cloud TCO 模型的各档租用价格换算为每百万 token（输入加输出）成本。`,
+    colPriceTier: '价格档位',
+    colGpuHour: '$/GPU/小时',
+    colCostPerMtok: '每百万 token 成本',
+    priceTierLabels: {
+      hyperscaler: '超大规模云',
+      neocloud: '新兴云',
+      retail: '零售租用',
+    },
+    emptyState: `InferenceX 集群尚未发布 ${model} 在 ${chipLabel} 上的基准测试结果。基准测试持续运行，新结果落地后本页会自动填充。`,
+    faqHeading: '常见问题',
+    faq,
+    exploreHeading: '继续探索数据',
+    exploreLinks: [
+      {
+        href: `/zh/rankings/fastest-gpu-for-${entry.model.slug}`,
+        label: `${entry.model.seoName} 推理最快 GPU 完整排行`,
+      },
+      {
+        href: `/zh/rankings/cheapest-gpu-for-${entry.model.slug}`,
+        label: `运行 ${entry.model.seoName} 最省钱 GPU 完整排行`,
+      },
+      { href: `/zh/chips/${entry.chip.slug}`, label: `${chipTitle} 规格与价格` },
+      { href: '/zh/inference', label: '交互式推理仪表盘' },
+    ],
+  };
+
+  return (
+    <>
+      <JsonLd data={jsonLd} />
+      <RunDetailContent entry={entry} data={data} t={t} />
+    </>
+  );
+}
diff --git a/packages/app/src/app/zh/run/page.tsx b/packages/app/src/app/zh/run/page.tsx
new file mode 100644
index 000000000..6a52c02b4
--- /dev/null
+++ b/packages/app/src/app/zh/run/page.tsx
@@ -0,0 +1,111 @@
+import type { Metadata } from 'next';
+import Link from 'next/link';
+
+import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { JsonLd } from '@/components/json-ld';
+import { Card } from '@/components/ui/card';
+import { ZH_LANG_TAG, ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
+import type { RunPageEntry } from '@/lib/run-pages';
+import { runPageHeadingZh } from '@/lib/run-pages-zh';
+import { getAvailableRunEntries } from '@/lib/run-rankings-data.server';
+
+export const dynamic = 'force-dynamic';
+
+const title = '任意模型在任意 GPU 上的实测结果';
+const description =
+  'InferenceX 集群实测的每一组开源大模型与 GPU 组合的吞吐、延迟与成本数据：DeepSeek、Kimi、GLM、MiniMax、Qwen 运行在 H100、H200、B200、GB200 NVL72、MI300X、MI325X、MI355X 等硬件上。';
+
+export function generateMetadata(): Metadata {
+  return {
+    title: { absolute: `${title} | ${SITE_NAME}` },
+    description,
+    keywords: [
+      'LLM GPU 基准测试',
+      '大模型 GPU 吞吐量',
+      '模型 GPU 推理性能',
+      'DeepSeek GPU 基准测试',
+      'Kimi GPU 基准测试',
+      'LLM 推理成本',
+      'H100 大模型基准测试',
+      'MI355X 大模型基准测试',
+    ],
+    alternates: zhAlternates('/run'),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url: `${SITE_URL}/zh/run`,
+      locale: ZH_OG_LOCALE,
+      type: 'website',
+    },
+    twitter: { card: 'summary_large_image', title, description },
+  };
+}
+
+export default async function ZhRunIndexPage() {
+  const entries = await getAvailableRunEntries();
+  const byModel = new Map<string, RunPageEntry[]>();
+  for (const entry of entries) {
+    const list = byModel.get(entry.model.slug) ?? [];
+    list.push(entry);
+    byModel.set(entry.model.slug, list);
+  }
+  const models = INFERENCE_MODEL_SLUGS.filter((model) => byModel.has(model.slug));
+
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'CollectionPage',
+    name: title,
+    description,
+    url: `${SITE_URL}/zh/run`,
+    inLanguage: ZH_LANG_TAG,
+    hasPart: entries.map((entry) => ({
+      '@type': 'WebPage',
+      name: runPageHeadingZh(entry),
+      url: `${SITE_URL}/zh/run/${entry.slug}`,
+    })),
+  };
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8">
+        <div className="mx-auto max-w-5xl">
+          <header className="pt-8 md:pt-12">
+            <h1 className="max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-5xl">
+              {title}
+            </h1>
+            <p className="mt-4 max-w-3xl text-lg leading-relaxed text-muted-foreground">
+              选一个模型和一款 GPU：下面每个页面都会回答它能跑多快、每百万 token
+              成本多少、以及数字出自哪个推理引擎，数据来自真实硬件上持续重跑的基准测试。
+            </p>
+          </header>
+
+          <section className="mt-10 mb-16 space-y-8">
+            {models.map((model) => (
+              <Card key={model.slug} className="p-5">
+                <h2 className="text-lg font-semibold tracking-tight">{model.seoName}</h2>
+                <ul className="mt-3 grid gap-1 sm:grid-cols-2">
+                  {(byModel.get(model.slug) ?? []).map((entry) => {
+                    const chipLabel = entry.chip.label;
+                    return (
+                      <li key={entry.slug}>
+                        <Link
+                          href={`/zh/run/${entry.slug}`}
+                          className="text-sm font-medium text-brand hover:underline"
+                        >
+                          {model.seoName} 运行在 {chipLabel}
+                        </Link>
+                      </li>
+                    );
+                  })}
+                </ul>
+              </Card>
+            ))}
+          </section>
+        </div>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/components/GlobalFilterContext.provider.test.tsx b/packages/app/src/components/GlobalFilterContext.provider.test.tsx
new file mode 100644
index 000000000..0dbbafb94
--- /dev/null
+++ b/packages/app/src/components/GlobalFilterContext.provider.test.tsx
@@ -0,0 +1,184 @@
+// @vitest-environment jsdom
+import { act, memo, useState } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { Model, Precision, Sequence } from '@/lib/data-mappings';
+
+const mocks = vi.hoisted(() => ({
+  availability: {
+    data: [] as unknown[],
+    error: null as Error | null,
+    isSuccess: true,
+  },
+  workflow: {
+    data: undefined as
+      | {
+          runs: Record<string, unknown>[];
+          changelogs: Record<string, unknown>[];
+          configs: unknown[];
+          runConfigs: unknown[];
+        }
+      | undefined,
+    isLoading: false,
+    error: null as Error | null,
+  },
+  unofficialAvailable: [],
+  getUrlParam: vi.fn(() => undefined),
+  setUrlParams: vi.fn(),
+}));
+
+vi.mock('@/hooks/api/use-availability', () => ({
+  useAvailability: () => mocks.availability,
+}));
+vi.mock('@/hooks/api/use-workflow-info', () => ({
+  useWorkflowInfo: () => mocks.workflow,
+}));
+vi.mock('@/hooks/useUrlState', () => ({
+  useUrlState: () => ({ getUrlParam: mocks.getUrlParam, setUrlParams: mocks.setUrlParams }),
+}));
+vi.mock('@/components/unofficial-run-provider', () => ({
+  useUnofficialRun: () => ({ availableModelsAndSequences: mocks.unofficialAvailable }),
+}));
+
+import {
+  GlobalFilterProvider,
+  useGlobalFilterActions,
+  useGlobalFilterSelection,
+  useGlobalFilterWorkflow,
+  type GlobalFilterActionsContextType,
+} from './GlobalFilterContext';
+
+let root: Root | undefined;
+let container: HTMLDivElement | undefined;
+let rerenderProvider: (() => void) | undefined;
+let selectModel: ((model: Model) => void) | undefined;
+let selectionRenders = 0;
+let actionRenders = 0;
+let workflowRenders = 0;
+const actionSnapshots: GlobalFilterActionsContextType[] = [];
+
+const SelectionProbe = memo(() => {
+  useGlobalFilterSelection();
+  selectionRenders += 1;
+  return null;
+});
+
+const ActionProbe = memo(() => {
+  const actions = useGlobalFilterActions();
+  selectModel = actions.setSelectedModel;
+  actionRenders += 1;
+  return null;
+});
+
+const WorkflowProbe = memo(() => {
+  useGlobalFilterWorkflow();
+  workflowRenders += 1;
+  return null;
+});
+
+const ActionIdentityProbe = memo(() => {
+  useGlobalFilterSelection();
+  actionSnapshots.push(useGlobalFilterActions());
+  return null;
+});
+
+function ProviderHarness() {
+  const [, setRevision] = useState(0);
+  rerenderProvider = () => setRevision((revision) => revision + 1);
+  return (
+    <GlobalFilterProvider
+      initialModel={Model.DeepSeek_V4_Pro}
+      initialSequence={Sequence.EightK_OneK}
+      initialPrecisions={[Precision.FP4]}
+    >
+      <SelectionProbe />
+      <ActionProbe />
+      <WorkflowProbe />
+      <ActionIdentityProbe />
+    </GlobalFilterProvider>
+  );
+}
+
+function mountProvider(): void {
+  container = document.createElement('div');
+  root = createRoot(container);
+  act(() => root?.render(<ProviderHarness />));
+}
+
+beforeEach(() => {
+  mocks.availability.data = [];
+  mocks.workflow.data = undefined;
+  mocks.workflow.isLoading = false;
+  mocks.workflow.error = null;
+  mocks.getUrlParam.mockReset();
+  mocks.getUrlParam.mockReturnValue(undefined);
+  mocks.setUrlParams.mockReset();
+  rerenderProvider = undefined;
+  selectModel = undefined;
+  selectionRenders = 0;
+  actionRenders = 0;
+  workflowRenders = 0;
+  actionSnapshots.length = 0;
+});
+
+afterEach(() => {
+  if (root) act(() => root?.unmount());
+  container?.remove();
+  root = undefined;
+  container = undefined;
+});
+
+describe('GlobalFilterProvider context isolation', () => {
+  it('does not notify selection or action consumers when workflow data changes', () => {
+    mountProvider();
+
+    act(() => {
+      mocks.workflow.data = {
+        runs: [
+          {
+            github_run_id: 42,
+            name: 'run',
+            conclusion: 'success',
+            run_attempt: 1,
+            html_url: 'https://example.test/runs/42',
+            created_at: '2026-08-20T00:00:00Z',
+            date: '2026-08-20',
+          },
+        ],
+        changelogs: [],
+        configs: [],
+        runConfigs: [],
+      };
+      rerenderProvider?.();
+    });
+
+    expect(workflowRenders).toBe(2);
+    expect(selectionRenders).toBe(1);
+    expect(actionRenders).toBe(1);
+  });
+
+  it('does not notify workflow-only consumers when selection changes', () => {
+    mountProvider();
+
+    act(() => selectModel?.(Model.DeepSeek_R1));
+
+    expect(selectionRenders).toBe(2);
+    expect(workflowRenders).toBe(1);
+  });
+
+  it('keeps every action identity stable across selection updates', () => {
+    mountProvider();
+    const initialActions = actionSnapshots[0];
+
+    act(() => selectModel?.(Model.DeepSeek_R1));
+
+    expect(actionSnapshots).toHaveLength(2);
+    expect(actionSnapshots[1]).toBe(initialActions);
+    expect(actionSnapshots[1].setSelectedModel).toBe(initialActions.setSelectedModel);
+    expect(actionSnapshots[1].setSelectedSequence).toBe(initialActions.setSelectedSequence);
+    expect(actionSnapshots[1].setSelectedPrecisions).toBe(initialActions.setSelectedPrecisions);
+    expect(actionSnapshots[1].setSelectedRunDate).toBe(initialActions.setSelectedRunDate);
+    expect(actionSnapshots[1].setSelectedRunId).toBe(initialActions.setSelectedRunId);
+  });
+});
diff --git a/packages/app/src/components/GlobalFilterContext.stale-url.test.tsx b/packages/app/src/components/GlobalFilterContext.stale-url.test.tsx
new file mode 100644
index 000000000..1a160f065
--- /dev/null
+++ b/packages/app/src/components/GlobalFilterContext.stale-url.test.tsx
@@ -0,0 +1,121 @@
+// @vitest-environment jsdom
+import { act, memo } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { Model, Sequence } from '@/lib/data-mappings';
+import type * as UrlStateModule from '@/lib/url-state';
+
+/**
+ * Regression coverage for the stale-`i_seq` snapshot bug behind the paramless
+ * AgentX hero links (see `agentxDashboardHref`).
+ *
+ * The module-level URL snapshot retains self-written params across soft
+ * navigations: every dashboard visit pins its resolved sequence via the
+ * URL-sync effect, and `refreshUrlParams` does not evict keys absent from the
+ * live URL. `getUrlParam('i_seq')` can therefore return the PREVIOUS page's
+ * sequence on a navigation whose URL carries no `i_seq` at all. Applying that
+ * value flips `sequenceExplicit`, which blocks the availability-driven AgentX
+ * default — a bare `/inference/kimi-k3` hero link would open on 8K/1K.
+ *
+ * The provider must consult `hasExplicitUrlParam('i_seq')` (which IS refreshed
+ * per navigation) before honoring the snapshot value.
+ */
+
+const mocks = vi.hoisted(() => ({
+  availability: {
+    data: [] as unknown[],
+    error: null as Error | null,
+    isSuccess: true,
+  },
+  workflow: {
+    data: undefined,
+    isLoading: false,
+    error: null as Error | null,
+  },
+  getUrlParam: vi.fn((key: string): string | undefined => (key === 'i_seq' ? '8k/1k' : undefined)),
+  setUrlParams: vi.fn(),
+  hasExplicitUrlParam: vi.fn((_key: string) => false),
+}));
+
+vi.mock('@/hooks/api/use-availability', () => ({
+  useAvailability: () => mocks.availability,
+}));
+vi.mock('@/hooks/api/use-workflow-info', () => ({
+  useWorkflowInfo: () => mocks.workflow,
+}));
+vi.mock('@/hooks/useUrlState', () => ({
+  useUrlState: () => ({ getUrlParam: mocks.getUrlParam, setUrlParams: mocks.setUrlParams }),
+}));
+vi.mock('@/lib/url-state', async (importOriginal) => ({
+  ...(await importOriginal<typeof UrlStateModule>()),
+  hasExplicitUrlParam: (key: string) => mocks.hasExplicitUrlParam(key),
+  refreshUrlParams: () => ({}),
+}));
+vi.mock('@/components/unofficial-run-provider', () => ({
+  useUnofficialRun: () => ({ availableModelsAndSequences: [] }),
+}));
+
+import { GlobalFilterProvider, useGlobalFilterSelection } from './GlobalFilterContext';
+
+/** Kimi-K3 availability: both the Agentic and the 8K/1K scenario exist. */
+const KIMI_K3_ROWS = [
+  { model: 'kimik3', isl: null, osl: null, benchmark_type: 'agentic_traces' },
+  { model: 'kimik3', isl: 8192, osl: 1024, benchmark_type: 'single_turn' },
+];
+
+let root: Root | undefined;
+let container: HTMLDivElement | undefined;
+let observedSequence: Sequence | undefined;
+let observedResolved: boolean | undefined;
+
+const SequenceProbe = memo(() => {
+  const selection = useGlobalFilterSelection();
+  observedSequence = selection.effectiveSequence;
+  observedResolved = selection.sequenceResolved;
+  return null;
+});
+
+function mountProvider(): void {
+  container = document.createElement('div');
+  root = createRoot(container);
+  act(() =>
+    root?.render(
+      <GlobalFilterProvider initialModel={Model.Kimi_K3}>
+        <SequenceProbe />
+      </GlobalFilterProvider>,
+    ),
+  );
+}
+
+beforeEach(() => {
+  mocks.availability.data = KIMI_K3_ROWS;
+  mocks.getUrlParam.mockClear();
+  mocks.setUrlParams.mockClear();
+  mocks.hasExplicitUrlParam.mockClear();
+  observedSequence = undefined;
+  observedResolved = undefined;
+});
+
+afterEach(() => {
+  if (root) act(() => root?.unmount());
+  container?.remove();
+  root = undefined;
+  container = undefined;
+});
+
+describe('GlobalFilterProvider stale i_seq snapshot', () => {
+  it('ignores a retained i_seq the current URL does not carry — AgentX default wins', () => {
+    mocks.hasExplicitUrlParam.mockReturnValue(false);
+    mountProvider();
+    expect(observedResolved).toBe(true);
+    expect(observedSequence).toBe(Sequence.AgenticTraces);
+  });
+
+  it('still honors an i_seq explicitly present in the current URL', () => {
+    mocks.hasExplicitUrlParam.mockImplementation((key: string) => key === 'i_seq');
+    mountProvider();
+    expect(observedResolved).toBe(true);
+    expect(observedSequence).toBe(Sequence.EightK_OneK);
+  });
+});
diff --git a/packages/app/src/components/GlobalFilterContext.test.ts b/packages/app/src/components/GlobalFilterContext.test.ts
new file mode 100644
index 000000000..40096cbc3
--- /dev/null
+++ b/packages/app/src/components/GlobalFilterContext.test.ts
@@ -0,0 +1,162 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  buildRunInfo,
+  getRequestedRunUrlParams,
+  resolveEffectiveRunDate,
+  resolveEffectiveRunId,
+} from '@/components/GlobalFilterContext';
+import workflowFixture from '../../cypress/fixtures/api/workflow-info.json';
+import type { WorkflowInfoResponse } from '@/lib/api';
+import type { RunInfo } from '@/components/inference/types';
+
+function run(runId: string): RunInfo {
+  return {
+    runId,
+    runDate: '2026-08-20',
+    runUrl: `https://example.test/${runId}`,
+    conclusion: 'success',
+  };
+}
+
+describe('buildRunInfo', () => {
+  it('matches the workflow fixture while preserving run and changelog order', () => {
+    const fixture = workflowFixture as unknown as WorkflowInfoResponse;
+    const result = buildRunInfo(fixture);
+
+    expect(Object.keys(result)).toEqual(
+      fixture.runs.map((fixtureRun) => String(fixtureRun.github_run_id)),
+    );
+    for (const fixtureRun of fixture.runs) {
+      const runId = String(fixtureRun.github_run_id);
+      const fixtureChangelogs = fixture.changelogs.filter(
+        (changelog) => String(changelog.workflow_run_id) === runId,
+      );
+      expect(result[runId]).toEqual({
+        runId,
+        runDate: fixtureRun.created_at,
+        runUrl: fixtureRun.html_url
+          ? `${fixtureRun.html_url}/attempts/${fixtureRun.run_attempt}`
+          : '',
+        conclusion: fixtureRun.conclusion,
+        changelog: {
+          entries: fixtureChangelogs.map((changelog) => ({
+            config_keys: changelog.config_keys,
+            description: changelog.description,
+            pr_link: changelog.pr_link,
+            head_ref: changelog.head_ref,
+            append_only: changelog.append_only,
+          })),
+        },
+      });
+    }
+  });
+
+  it('keeps changelog order and the empty-URL/no-changelog fallbacks', () => {
+    const data = {
+      runs: [
+        {
+          github_run_id: 7,
+          name: 'run seven',
+          conclusion: null,
+          run_attempt: 2,
+          html_url: null,
+          created_at: '2026-08-20T00:00:00Z',
+          date: '2026-08-20',
+        },
+        {
+          github_run_id: 8,
+          name: 'run eight',
+          conclusion: 'success',
+          run_attempt: 1,
+          html_url: null,
+          created_at: '2026-08-20T01:00:00Z',
+          date: '2026-08-20',
+        },
+      ],
+      changelogs: [
+        {
+          workflow_run_id: 7,
+          date: '2026-08-20',
+          base_ref: 'base',
+          head_ref: 'first',
+          config_keys: ['first-config'],
+          description: 'first',
+          pr_link: null,
+        },
+        {
+          workflow_run_id: 7,
+          date: '2026-08-20',
+          base_ref: 'first',
+          head_ref: 'second',
+          config_keys: ['second-config'],
+          description: 'second',
+          pr_link: 'https://example.test/pull/2',
+          append_only: true,
+        },
+      ],
+      configs: [],
+      runConfigs: [],
+    } satisfies WorkflowInfoResponse;
+
+    const result = buildRunInfo(data);
+    expect(result['7']).toMatchObject({
+      runUrl: '',
+      changelog: {
+        entries: [
+          { description: 'first', head_ref: 'first' },
+          { description: 'second', head_ref: 'second', append_only: true },
+        ],
+      },
+    });
+    expect(result['8']).toEqual({
+      runId: '8',
+      runDate: '2026-08-20T01:00:00Z',
+      runUrl: '',
+      conclusion: 'success',
+    });
+  });
+});
+
+describe('global filter requested and effective selectors', () => {
+  it('keeps an explicit available date', () => {
+    expect(
+      resolveEffectiveRunDate('2026-08-19', ['2026-08-18', '2026-08-19', '2026-08-20'], true),
+    ).toBe('2026-08-19');
+  });
+
+  it('uses latest for implicit or stale date intent', () => {
+    const dates = ['2026-08-18', '2026-08-20'];
+    expect(resolveEffectiveRunDate('', dates, false)).toBe('2026-08-20');
+    expect(resolveEffectiveRunDate('2026-08-19', dates, true)).toBe('2026-08-20');
+  });
+
+  it('preserves a requested date while availability is unresolved or empty', () => {
+    expect(resolveEffectiveRunDate('2026-08-19', [], true)).toBe('2026-08-19');
+  });
+
+  it('keeps a valid run ID and otherwise selects the newest available run', () => {
+    const runs = { '100': run('100'), '102': run('102'), '101': run('101') };
+    expect(resolveEffectiveRunId('101', runs)).toBe('101');
+    expect(resolveEffectiveRunId('missing', runs)).toBe('102');
+    expect(resolveEffectiveRunId('', runs)).toBe('102');
+  });
+
+  it('clears the effective run ID for a settled empty run map', () => {
+    expect(resolveEffectiveRunId('101', {})).toBe('');
+  });
+
+  it('serializes requested run state instead of availability fallbacks', () => {
+    const requestedDate = '2026-08-19';
+    const requestedRunId = 'missing';
+    const dates = ['2026-08-18', '2026-08-20'];
+    const runs = { '100': run('100'), '102': run('102') };
+
+    expect(resolveEffectiveRunDate(requestedDate, dates, true)).toBe('2026-08-20');
+    expect(resolveEffectiveRunId(requestedRunId, runs)).toBe('102');
+    expect(getRequestedRunUrlParams(requestedDate, requestedRunId)).toEqual({
+      g_rundate: requestedDate,
+      g_runid: requestedRunId,
+    });
+  });
+});
diff --git a/packages/app/src/components/GlobalFilterContext.tsx b/packages/app/src/components/GlobalFilterContext.tsx
index a2eec74f6..00d0102ae 100644
--- a/packages/app/src/components/GlobalFilterContext.tsx
+++ b/packages/app/src/components/GlobalFilterContext.tsx
@@ -1,5 +1,7 @@
 'use client';
 
+import { usePathname } from 'next/navigation';
+
 import {
   type ReactNode,
   createContext,
@@ -24,7 +26,11 @@ function isEnumValue<T extends Record<string, string>>(e: T, v: string): v is T[
 import { useAvailability } from '@/hooks/api/use-availability';
 import { useWorkflowInfo } from '@/hooks/api/use-workflow-info';
 import { useUrlState } from '@/hooks/useUrlState';
+import { hasExplicitUrlParam, refreshUrlParams, type UrlStateParams } from '@/lib/url-state';
+import { modelRoutePathnameRewrite, routeModelForPathname } from '@/lib/model-routes';
+import { replaceClientPathname } from '@/lib/client-navigation';
 import { useUnofficialRun } from '@/components/unofficial-run-provider';
+import type { RunInfo } from '@/components/inference/types';
 import {
   Model,
   MODEL_OPTIONS,
@@ -33,95 +39,102 @@ import {
   Sequence,
   SEQUENCE_OPTIONS,
 } from '@/lib/data-mappings';
+import { inferenceModelForPathname } from '@/lib/inference-model-slug';
 import { computeAutoSwitchDecision } from '@/lib/unofficial-run-auto-switch';
 import { countCurvesByPrecision, resolveEffectivePrecisions } from '@/lib/default-precisions';
 import { resolveEffectiveSequence } from '@/lib/default-sequence';
 import type { AvailabilityRow, WorkflowInfoResponse } from '@/lib/api';
-
 const RUNDATE_RE = /^\d{4}-\d{2}-\d{2}$/u;
 const RUNID_RE = /^[A-Za-z0-9_-]{1,64}$/u;
 
 // Placeholder for the public (non-null) `effectiveSequence` during the window
-// before availability has loaded. It must be a fixed-seq scenario — never
-// AgenticTraces — so the scenario selector doesn't flash "Agentic Traces" for a
-// fixed-seq-only model while the chart shows its loading skeleton. `8k/1k` is
-// the app default scenario. Consumers that must not act on an unresolved
-// sequence gate on `sequenceResolved` instead.
+// before availability has loaded. It stays on the fixed-sequence app default
+// until availability confirms whether the selected model has AgentX data, so a
+// fixed-seq-only model never flashes an agentic label. Consumers that must not
+// act on an unresolved sequence gate on `sequenceResolved` instead.
 // (Declared after the import block so it never references `Sequence` above its import.)
-const PRE_AVAILABILITY_SEQUENCE = Sequence.EightK_OneK;
-
-interface RunInfo {
-  runId: string;
-  runDate: string;
-  runUrl: string;
-  conclusion: string | null;
-  changelog?: {
-    entries: {
-      config_keys: string[];
-      description: string;
-      pr_link: string | null;
-      head_ref: string;
-    }[];
-  };
-}
+const APP_DEFAULT_SEQUENCE = Sequence.EightK_OneK;
 
-export interface GlobalFilterContextType {
-  // Shared filter state
+export interface GlobalFilterSelectionContextType {
   selectedModel: Model;
-  setSelectedModel: (model: Model) => void;
   selectedSequence: Sequence;
-  setSelectedSequence: (sequence: Sequence) => void;
   selectedPrecisions: string[];
-  setSelectedPrecisions: (precisions: string[]) => void;
-
-  // Effective (validated) values
   effectiveSequence: Sequence;
   /**
    * Whether `effectiveSequence` reflects the selected model's real availability
-   * (DB or unofficial run) rather than the pre-load placeholder. False during
-   * the brief window before availability loads. Consumers that trigger data
-   * fetches or render sequence-dependent labels should gate on this so a
-   * fixed-seq-only model never fires an agentic fetch or flashes "Agentic
-   * Traces" before availability settles.
+   * rather than the pre-load placeholder.
    */
   sequenceResolved: boolean;
   effectivePrecisions: string[];
+}
 
-  // Run date & run ID
-  selectedRunDate: string;
+export interface GlobalFilterActionsContextType {
+  setSelectedModel: (model: Model) => void;
+  setSelectedSequence: (sequence: Sequence) => void;
+  setSelectedPrecisions: (precisions: string[]) => void;
   setSelectedRunDate: (date: string) => void;
+  setSelectedRunId: (id: string) => void;
+}
+
+export interface GlobalFilterRunContextType {
+  selectedRunDate: string;
   selectedRunDateRev: number;
   selectedRunId: string;
-  setSelectedRunId: (id: string) => void;
+  effectiveRunDate: string;
+}
 
-  // Derived availability
+export interface GlobalFilterAvailabilityContextType {
   availableModels: Model[];
   availableSequences: Sequence[];
   availablePrecisions: string[];
   availableDates: string[];
-  effectiveRunDate: string;
-
-  // Raw availability rows (shared with inference for GPU filtering)
   availabilityRows: AvailabilityRow[] | undefined;
+  /** True once the database availability request has either succeeded or failed. */
+  availabilitySettled: boolean;
+  availabilityError: string | null;
+}
 
-  // Workflow info
-  workflowInfo: { runInfoBySequence: Record<string, RunInfo> }[] | null;
+export interface GlobalFilterWorkflowContextType {
   availableRuns: Record<string, RunInfo>;
   workflowLoading: boolean;
   workflowError: string | null;
 }
 
-/** @internal Exported for test provider wrapping only. */
-export const GlobalFilterContext = createContext<GlobalFilterContextType | undefined>(undefined);
+/** @internal Exported for focused provider tests and Cypress provider wrapping. */
+export const GlobalFilterSelectionContext = createContext<
+  GlobalFilterSelectionContextType | undefined
+>(undefined);
+/** @internal Exported for focused provider tests and Cypress provider wrapping. */
+export const GlobalFilterActionsContext = createContext<GlobalFilterActionsContextType | undefined>(
+  undefined,
+);
+/** @internal Exported for focused provider tests and Cypress provider wrapping. */
+export const GlobalFilterRunContext = createContext<GlobalFilterRunContextType | undefined>(
+  undefined,
+);
+/** @internal Exported for focused provider tests and Cypress provider wrapping. */
+export const GlobalFilterAvailabilityContext = createContext<
+  GlobalFilterAvailabilityContextType | undefined
+>(undefined);
+/** @internal Exported for focused provider tests and Cypress provider wrapping. */
+export const GlobalFilterWorkflowContext = createContext<
+  GlobalFilterWorkflowContextType | undefined
+>(undefined);
+
+/** Transform API response into the run map consumed by selectors and displays. */
+export function buildRunInfo(data: WorkflowInfoResponse): Record<string, RunInfo> {
+  const changelogsByRunId = new Map<string, (typeof data.changelogs)[number][]>();
+  for (const changelog of data.changelogs) {
+    const runId = String(changelog.workflow_run_id);
+    const indexed = changelogsByRunId.get(runId);
+    if (indexed) indexed.push(changelog);
+    else changelogsByRunId.set(runId, [changelog]);
+  }
 
-/** Transform API response into the shape the app expects. */
-function buildRunInfo(data: WorkflowInfoResponse): Record<string, RunInfo> {
   const runs: Record<string, RunInfo> = {};
   for (const run of data.runs) {
     const runId = String(run.github_run_id);
-    const runChangelogs = data.changelogs.filter(
-      (c) => String(c.workflow_run_id) === String(run.github_run_id),
-    );
+    const runChangelogs = changelogsByRunId.get(runId) ?? [];
     runs[runId] = {
       runId,
       runDate: run.created_at,
@@ -129,11 +142,12 @@ function buildRunInfo(data: WorkflowInfoResponse): Record<string, RunInfo> {
       conclusion: run.conclusion,
       ...(runChangelogs.length > 0 && {
         changelog: {
-          entries: runChangelogs.map((c) => ({
-            config_keys: c.config_keys,
-            description: c.description,
-            pr_link: c.pr_link,
-            head_ref: c.head_ref,
+          entries: runChangelogs.map((changelog) => ({
+            config_keys: changelog.config_keys,
+            description: changelog.description,
+            pr_link: changelog.pr_link,
+            head_ref: changelog.head_ref,
+            append_only: changelog.append_only,
           })),
         },
       }),
@@ -142,38 +156,107 @@ function buildRunInfo(data: WorkflowInfoResponse): Record<string, RunInfo> {
   return runs;
 }
 
+export function resolveEffectiveRunDate(
+  requestedDate: string,
+  availableDates: readonly string[],
+  requestedExplicitly: boolean,
+): string {
+  if (availableDates.length === 0) return requestedDate;
+  if (requestedExplicitly && requestedDate && availableDates.includes(requestedDate)) {
+    return requestedDate;
+  }
+  return availableDates.at(-1)!;
+}
+
+export function resolveEffectiveRunId(
+  requestedRunId: string,
+  availableRuns: Readonly<Record<string, RunInfo>>,
+): string {
+  const runIds = Object.keys(availableRuns);
+  if (runIds.length === 0) return '';
+  if (requestedRunId && Object.hasOwn(availableRuns, requestedRunId)) return requestedRunId;
+  return runIds.reduce((latest, runId) => (runId > latest ? runId : latest), runIds[0]);
+}
+
+export function getRequestedRunUrlParams(
+  requestedRunDate: string,
+  requestedRunId: string,
+): Pick<UrlStateParams, 'g_rundate' | 'g_runid'> {
+  return {
+    g_rundate: requestedRunDate,
+    g_runid: requestedRunId,
+  };
+}
+
 export function GlobalFilterProvider({
   children,
   initialModel,
   initialSequence,
   initialPrecisions,
+  initialRunDate,
+  initialRunId,
 }: {
   children: ReactNode;
   /**
    * Initial values used when no URL params are present. Lets per-route entry
    * points (e.g. `/compare/[a]-vs-[b]`) seed sensible defaults derived from
-   * actual data — without these, every page falls back to FP4/8K-1K which
-   * has no data for older GPUs (Hopper, CDNA 3).
+   * actual data, without forcing a second provider during server rendering.
    */
   initialModel?: Model;
   initialSequence?: Sequence;
   initialPrecisions?: string[];
+  initialRunDate?: string;
+  initialRunId?: string;
 }) {
-  const { hasUrlParam, getUrlParam, setUrlParams } = useUrlState();
+  const { getUrlParam, setUrlParams } = useUrlState();
+
+  // Owned by the Next router. Seeds per-model routes below and keys the
+  // URL-param layout effect further down (see that effect for the full
+  // usePathname-vs-useSearchParams rationale).
+  const pathname = usePathname();
 
   // ── Core filter state ─────────────────────────────────────────────────────
+  // Per-model routes (/historical/<slug>) imply a model at first render so the
+  // server response and hydration already show the routed model — no client
+  // effect flicker. An explicit `initialModel` prop (calculator URL seed) wins.
   const [selectedModel, setSelectedModel] = useState<Model>(
-    () => initialModel ?? Model.DeepSeek_V4_Pro,
+    () => initialModel ?? routeModelForPathname(pathname) ?? Model.DeepSeek_V4_Pro,
   );
 
-  const [selectedSequence, setSelectedSequence] = useState<Sequence>(() => {
+  const [selectedSequence, setSelectedSequenceRaw] = useState<Sequence>(() => {
     if (initialSequence) return initialSequence;
-    const urlSeq = getUrlParam('i_seq');
+    // Only honor `i_seq` when the CURRENT navigation's URL carries it. The
+    // module-level snapshot retains self-written params across soft
+    // navigations (the URL-sync effect pins the resolved sequence on every
+    // dashboard visit), so an unguarded read would revive the previous page's
+    // sequence — e.g. a paramless AgentX hero link opening on a stale 8k/1k.
+    const urlSeq = hasExplicitUrlParam('i_seq') ? getUrlParam('i_seq') : undefined;
     if (urlSeq && Object.values(Sequence).includes(urlSeq as Sequence)) return urlSeq as Sequence;
-    // Default to the 8K/1K fixed-seq scenario; the effectiveSequence fallback
-    // below handles models that lack it (e.g. agentic-only models).
+    // Default to the 8K/1K fixed-seq scenario; the effectiveSequence resolution
+    // below prefers the Agentic scenario when availability confirms the model
+    // has corresponding data, and handles models that lack 8K/1K entirely.
     return Sequence.EightK_OneK;
   });
+  // Whether the scenario was chosen explicitly (seeded `initialSequence` prop,
+  // URL `i_seq`, or a manual pick). Until then the availability-driven AgentX
+  // default applies —
+  // without this flag the initial `8k/1k` state is indistinguishable from a
+  // deliberate 8K/1K selection, and the Agentic scenario could never win.
+  //
+  // Deliberately NOT seeded from `i_seq` here: this flag feeds the scenario
+  // label rendered during the pre-availability window, and `getUrlParam` sees
+  // the query string on the client but not on the server. Reading it in the
+  // initializer would render a different label on each side and fail
+  // hydration. The layout effect below applies `i_seq` (flag included) after
+  // the first commit and before paint, which is how every other URL param
+  // lands.
+  const [sequenceExplicit, setSequenceExplicit] = useState<boolean>(
+    () => initialSequence !== undefined,
+  );
+  const setSelectedSequence = useCallback((sequence: Sequence) => {
+    setSelectedSequenceRaw(sequence);
+    setSequenceExplicit(true);
+  }, []);
 
   const initialValidPrecisions = useMemo(
     () =>
@@ -195,17 +278,35 @@ export function GlobalFilterProvider({
   }, []);
 
   // ── Run date / run ID ─────────────────────────────────────────────────────
-  const [selectedRunDate, setSelectedRunDateBase] = useState<string>('');
+  const [requestedRunDate, setRequestedRunDate] = useState<string>(() => initialRunDate ?? '');
   const [selectedRunDateRev, setSelectedRunDateRev] = useState(0);
+  const requestedRunDateExplicitRef = useRef(
+    initialRunDate !== undefined || hasExplicitUrlParam('g_rundate'),
+  );
 
-  const [selectedRunId, setSelectedRunId] = useState<string>('');
+  const [requestedRunId, setRequestedRunId] = useState<string>(() => initialRunId ?? '');
 
   // Apply URL param overrides synchronously after the first commit. Runs only
   // on the client (useEffect on server is a no-op). Updates state before paint
   // so users with shareable URLs (?i_seq=…&g_model=…) see their values without
   // flicker, and SSR/client hydration agree because initial state came from
   // props/defaults on both sides.
+  // Soft navigations do not remount this provider, and `useUrlState` caches the
+  // URL in a ref at first render — so without this, landing on
+  // `/inference?g_model=Qwen-3.5-397B-A17B` via a <Link> kept whatever model
+  // the previous page had (the default, DeepSeek). Every AgentX card on the
+  // landing page pointed at the right URL and still opened the wrong model.
+  // `usePathname` rather than `useSearchParams`: the latter forces every
+  // statically-prerendered page that mounts this provider behind a Suspense
+  // boundary, which fails the build on /ai-chart. Pathname changes on the
+  // navigations that matter here — the AgentX cards and every share link live
+  // on a different route from /inference. (`pathname` itself is declared with
+  // the core filter state above, where it also seeds per-model routes.)
   useIsomorphicLayoutEffect(() => {
+    // Pull the live URL into the shared snapshot first: `getUrlParam` below and
+    // the auto-switch guard further down both read from it, and on a soft
+    // navigation it still holds the previous page's params.
+    refreshUrlParams();
     const applyIfEnum = <T extends Record<string, string>>(
       key: 'g_model' | 'i_seq',
       enumType: T,
@@ -223,8 +324,33 @@ export function GlobalFilterProvider({
       if (value !== undefined && pattern.test(value)) apply(value);
     };
 
-    applyIfEnum('g_model', Model, setSelectedModel);
-    applyIfEnum('i_seq', Sequence, setSelectedSequence);
+    // Per-model routes imply a model on soft navigation too (the useState
+    // seed above only covers the mount; soft navigations between model pages
+    // do not remount this provider). Two route families pin the model:
+    // `/calculator/<model>` + `/historical/<model>` (routeModelForPathname)
+    // and `/inference/<model>` (inferenceModelForPathname).
+    const routeModel = routeModelForPathname(pathname);
+    if (routeModel) setSelectedModel(routeModel);
+    const pathModel = inferenceModelForPathname(pathname);
+    if (pathModel !== null) setSelectedModel(pathModel);
+    // On calculator/historical pages an explicit legacy ?g_model= still wins
+    // (applied after the route pin; the URL-sync effect below then
+    // canonicalizes the pathname to whatever model actually got applied). On
+    // an `/inference/<model>` page, `g_model` applies only when present in
+    // the CURRENT URL: the snapshot retains params from the previous
+    // navigation (`refreshUrlParams` only overwrites keys present in the live
+    // URL), and letting that stale value through would override the model the
+    // path just asked for.
+    if (pathModel === null || hasExplicitUrlParam('g_model')) {
+      applyIfEnum('g_model', Model, setSelectedModel);
+    }
+    // Same guard as the initializer above: `i_seq` applies (and flips
+    // `sequenceExplicit`, which blocks the availability-driven AgentX default)
+    // only when the current URL actually carries it — never from the snapshot's
+    // retained self-writes of a previously visited page.
+    if (hasExplicitUrlParam('i_seq')) {
+      applyIfEnum('i_seq', Sequence, setSelectedSequence);
+    }
     const urlPrec = getUrlParam('i_prec');
     if (urlPrec) {
       const precs = urlPrec
@@ -235,13 +361,28 @@ export function GlobalFilterProvider({
         setPrecisionExplicit(true);
       }
     }
-    applyIfMatches('g_rundate', RUNDATE_RE, setSelectedRunDateBase);
-    applyIfMatches('g_runid', RUNID_RE, setSelectedRunId);
+    applyIfMatches('g_rundate', RUNDATE_RE, (date) => {
+      requestedRunDateExplicitRef.current = true;
+      setRequestedRunDate(date);
+    });
+    applyIfMatches('g_runid', RUNID_RE, setRequestedRunId);
+    // Re-runs on client-side navigation as well as on mount. Keyed on the
+    // pathname, which the Next router owns: the provider's own share-link
+    // writes go through `history.replaceState` and never change it, so this
+    // cannot fight a user changing filters in place. A param-only navigation
+    // within /inference is therefore not picked up — no in-app link does that
+    // today, and covering it would mean the Suspense bailout above.
     // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, []);
+  }, [pathname]);
 
   // ── Availability data ─────────────────────────────────────────────────────
-  const { data: availabilityRows } = useAvailability();
+  const {
+    data: availabilityRows,
+    error: availabilityQueryError,
+    isPending: availabilityPending,
+  } = useAvailability();
+  const availabilitySettled = !availabilityPending;
+  const availabilityError = availabilityQueryError ? availabilityQueryError.message : null;
   const { availableModelsAndSequences: unofficialAvailable } = useUnofficialRun();
 
   const dbModelKeys = useMemo<string[]>(
@@ -272,19 +413,15 @@ export function GlobalFilterProvider({
   // leaves the user staring at a chart with no overlay points — they'd have
   // to know to open the dropdown and pick the run's model themselves.
   //
-  // Precedence on first load: the `if (urlModel)` early-bail in
-  // `computeAutoSwitchDecision` is the primary guard for explicit `g_model`
-  // intent. The dedupe ref is a secondary guard for the narrow window after
-  // an auto-switch fires but before the URL-sync effect (below) writes
-  // `g_model` back to the URL — once that runs, `urlModel` is set on every
-  // subsequent render and the ref check is effectively redundant. The ref
-  // still matters across navigations between unofficial runs because it is
-  // reset whenever the overlay set goes empty.
+  // Explicit URL intent is tracked separately from the mutable state mirrored
+  // for provider remounts. Automatic model changes serialize like any other
+  // filter, but must not become an explicit `g_model` that blocks the next
+  // unofficial-run auto-switch.
   const lastAutoSwitchKeyRef = useRef<string>('');
   useEffect(() => {
     const decision = computeAutoSwitchDecision(
       unofficialAvailable,
-      getUrlParam('g_model'),
+      hasExplicitUrlParam('g_model') ? getUrlParam('g_model') : undefined,
       selectedModel,
       lastAutoSwitchKeyRef.current,
     );
@@ -324,23 +461,24 @@ export function GlobalFilterProvider({
   // Synchronously validated sequence.
   //
   // `resolveEffectiveSequence` returns null while availability is still loading
-  // — we surface that as `sequenceResolved` so InferenceContext can gate the
+  // so InferenceProvider can gate the
   // benchmark fetch until the real sequence is known (no agentic fetch fires for
   // a fixed-seq-only model). For the non-null public `effectiveSequence` value
-  // we substitute a fixed-seq scenario (never AgenticTraces) during that window
-  // so the scenario selector never flashes "Agentic Traces"; the chart shows its
-  // normal loading skeleton until `sequenceResolved` flips true.
+  // we retain the fixed-sequence placeholder until availability confirms the
+  // AgentX scenario exists; the chart shows its normal loading skeleton until
+  // `sequenceResolved` flips true.
   const resolvedSequence = useMemo(
     () =>
       resolveEffectiveSequence({
         selectedSequence,
         availableSequences,
         availabilityLoaded,
+        sequenceExplicit,
       }),
-    [selectedSequence, availableSequences, availabilityLoaded],
+    [selectedSequence, availableSequences, availabilityLoaded, sequenceExplicit],
   );
   const sequenceResolved = resolvedSequence !== null;
-  const effectiveSequence = resolvedSequence ?? PRE_AVAILABILITY_SEQUENCE;
+  const effectiveSequence = resolvedSequence ?? APP_DEFAULT_SEQUENCE;
 
   // Precisions available for the selected model + sequence (DB ∪ unofficial run)
   const availablePrecisions = useMemo(() => {
@@ -403,31 +541,17 @@ export function GlobalFilterProvider({
     return [...new Set(rows.map((r) => r.date))].toSorted();
   }, [availabilityRows, modelRows, effectiveSequence, effectivePrecisions]);
 
-  // When true, keep the user's date if available; otherwise always use latest
-  const userPickedDateRef = useRef(Boolean(getUrlParam('g_rundate')));
-
   const setSelectedRunDateManual = useCallback((date: string) => {
-    userPickedDateRef.current = true;
-    setSelectedRunDateBase(date);
-    setSelectedRunDateRev((v) => v + 1);
+    requestedRunDateExplicitRef.current = true;
+    setRequestedRunDate(date);
+    setSelectedRunDateRev((revision) => revision + 1);
   }, []);
 
-  const effectiveRunDate = useMemo(() => {
-    if (availableDates.length === 0) return selectedRunDate;
-    const latest = availableDates.at(-1)!;
-    if (userPickedDateRef.current && selectedRunDate && availableDates.includes(selectedRunDate)) {
-      return selectedRunDate;
-    }
-    return latest;
-  }, [availableDates, selectedRunDate]);
-
-  // Sync selectedRunDate state when effectiveRunDate changes
-  useEffect(() => {
-    if (availableDates.length > 0 && effectiveRunDate !== selectedRunDate) {
-      setSelectedRunDateBase(effectiveRunDate);
-      setSelectedRunDateRev((v) => v + 1);
-    }
-  }, [effectiveRunDate, availableDates]);
+  const effectiveRunDate = resolveEffectiveRunDate(
+    requestedRunDate,
+    availableDates,
+    requestedRunDateExplicitRef.current,
+  );
 
   // ── Workflow info ─────────────────────────────────────────────────────────
   const {
@@ -443,38 +567,14 @@ export function GlobalFilterProvider({
     [workflowData],
   );
 
-  const workflowInfo = useMemo(
-    () => (Object.keys(availableRuns).length > 0 ? [{ runInfoBySequence: availableRuns }] : null),
-    [availableRuns],
+  const effectiveRunId = useMemo(
+    () => resolveEffectiveRunId(requestedRunId, availableRuns),
+    [requestedRunId, availableRuns],
   );
 
-  // Auto-select latest run ID when availableRuns change
-  const urlInitRef = useRef({ runIdApplied: false });
-
-  useEffect(() => {
-    if (availableRuns && Object.keys(availableRuns).length > 0) {
-      if (!urlInitRef.current.runIdApplied && hasUrlParam('g_runid')) {
-        const urlRunId = getUrlParam('g_runid')!;
-        urlInitRef.current.runIdApplied = true;
-        if (Object.keys(availableRuns).includes(urlRunId)) {
-          setSelectedRunId(urlRunId);
-          return;
-        }
-      }
-      urlInitRef.current.runIdApplied = true;
-
-      if (
-        !selectedRunId ||
-        (selectedRunId && !Object.keys(availableRuns).includes(selectedRunId))
-      ) {
-        const runIds = Object.keys(availableRuns);
-        const maxRunId = runIds.reduce((max, id) => (id > max ? id : max), runIds[0]);
-        setSelectedRunId(maxRunId);
-      }
-    } else if (selectedRunId !== '') {
-      setSelectedRunId('');
-    }
-  }, [availableRuns, selectedRunId]);
+  const setSelectedRunId = useCallback((runId: string) => {
+    setRequestedRunId(runId);
+  }, []);
 
   // ── URL sync ──────────────────────────────────────────────────────────────
   const isMountedRef = useRef(false);
@@ -485,8 +585,7 @@ export function GlobalFilterProvider({
     }
     setUrlParams({
       g_model: selectedModel,
-      g_rundate: selectedRunDate,
-      g_runid: selectedRunId,
+      ...getRequestedRunUrlParams(requestedRunDate, requestedRunId),
       // Don't pin the sequence to the URL until it's resolved from real
       // availability — writing the pre-load placeholder (8k/1k) would clobber a
       // shared `?i_seq=agentic-traces` link before the model's availability
@@ -496,10 +595,17 @@ export function GlobalFilterProvider({
       // leave it out so the link keeps following the per-model densest default.
       i_prec: precisionExplicit ? effectivePrecisions.join(',') : undefined,
     });
+    // Keep per-model pathnames (/calculator/<slug>, /historical/<slug>) in
+    // sync with the selected model — a plain history.replaceState, so switching
+    // models updates the address bar without an App Router navigation, RSC
+    // fetch, or component remount. Null when no rewrite is needed (bare tab
+    // path on the default model stays bare).
+    const pathnameRewrite = modelRoutePathnameRewrite(window.location.pathname, selectedModel);
+    if (pathnameRewrite) replaceClientPathname(pathnameRewrite);
   }, [
     selectedModel,
-    selectedRunDate,
-    selectedRunId,
+    requestedRunDate,
+    requestedRunId,
     effectiveSequence,
     sequenceResolved,
     effectivePrecisions,
@@ -507,32 +613,14 @@ export function GlobalFilterProvider({
     setUrlParams,
   ]);
 
-  const contextValue = useMemo<GlobalFilterContextType>(
+  const selectionValue = useMemo<GlobalFilterSelectionContextType>(
     () => ({
       selectedModel,
-      setSelectedModel,
       selectedSequence,
-      setSelectedSequence,
       selectedPrecisions,
-      setSelectedPrecisions,
       effectiveSequence,
       sequenceResolved,
       effectivePrecisions,
-      selectedRunDate: effectiveRunDate,
-      setSelectedRunDate: setSelectedRunDateManual,
-      selectedRunDateRev,
-      selectedRunId,
-      setSelectedRunId,
-      availableModels,
-      availableSequences,
-      availablePrecisions,
-      availableDates,
-      effectiveRunDate,
-      availabilityRows,
-      workflowInfo,
-      availableRuns,
-      workflowLoading,
-      workflowError,
     }),
     [
       selectedModel,
@@ -541,31 +629,113 @@ export function GlobalFilterProvider({
       effectiveSequence,
       sequenceResolved,
       effectivePrecisions,
-      effectiveRunDate,
+    ],
+  );
+
+  const actionsValue = useMemo<GlobalFilterActionsContextType>(
+    () => ({
+      setSelectedModel,
+      setSelectedSequence,
+      setSelectedPrecisions,
+      setSelectedRunDate: setSelectedRunDateManual,
+      setSelectedRunId,
+    }),
+    [
+      setSelectedModel,
+      setSelectedSequence,
+      setSelectedPrecisions,
       setSelectedRunDateManual,
+      setSelectedRunId,
+    ],
+  );
+
+  const runValue = useMemo<GlobalFilterRunContextType>(
+    () => ({
+      selectedRunDate: effectiveRunDate,
       selectedRunDateRev,
-      selectedRunId,
+      selectedRunId: effectiveRunId,
+      effectiveRunDate,
+    }),
+    [effectiveRunDate, selectedRunDateRev, effectiveRunId],
+  );
+
+  const availabilityValue = useMemo<GlobalFilterAvailabilityContextType>(
+    () => ({
+      availableModels,
+      availableSequences,
+      availablePrecisions,
+      availableDates,
+      availabilityRows,
+      availabilitySettled,
+      availabilityError,
+    }),
+    [
       availableModels,
       availableSequences,
       availablePrecisions,
       availableDates,
       availabilityRows,
-      workflowInfo,
-      availableRuns,
-      workflowLoading,
-      workflowError,
+      availabilitySettled,
+      availabilityError,
     ],
   );
 
+  const workflowValue = useMemo<GlobalFilterWorkflowContextType>(
+    () => ({ availableRuns, workflowLoading, workflowError }),
+    [availableRuns, workflowLoading, workflowError],
+  );
+
   return (
-    <GlobalFilterContext.Provider value={contextValue}>{children}</GlobalFilterContext.Provider>
+    <GlobalFilterActionsContext.Provider value={actionsValue}>
+      <GlobalFilterSelectionContext.Provider value={selectionValue}>
+        <GlobalFilterRunContext.Provider value={runValue}>
+          <GlobalFilterAvailabilityContext.Provider value={availabilityValue}>
+            <GlobalFilterWorkflowContext.Provider value={workflowValue}>
+              {children}
+            </GlobalFilterWorkflowContext.Provider>
+          </GlobalFilterAvailabilityContext.Provider>
+        </GlobalFilterRunContext.Provider>
+      </GlobalFilterSelectionContext.Provider>
+    </GlobalFilterActionsContext.Provider>
   );
 }
 
-export function useGlobalFilters() {
-  const context = useContext(GlobalFilterContext);
+export function useGlobalFilterSelection() {
+  const context = useContext(GlobalFilterSelectionContext);
+  if (context === undefined) {
+    throw new Error('useGlobalFilterSelection must be used within a GlobalFilterProvider');
+  }
+  return context;
+}
+
+export function useGlobalFilterActions() {
+  const context = useContext(GlobalFilterActionsContext);
+  if (context === undefined) {
+    throw new Error('useGlobalFilterActions must be used within a GlobalFilterProvider');
+  }
+  return context;
+}
+
+export function useGlobalFilterRun() {
+  const context = useContext(GlobalFilterRunContext);
+  if (context === undefined) {
+    throw new Error('useGlobalFilterRun must be used within a GlobalFilterProvider');
+  }
+  return context;
+}
+
+export function useGlobalFilterAvailability() {
+  const context = useContext(GlobalFilterAvailabilityContext);
+  if (context === undefined) {
+    throw new Error('useGlobalFilterAvailability must be used within a GlobalFilterProvider');
+  }
+  return context;
+}
+
+export function useGlobalFilterWorkflow() {
+  const context = useContext(GlobalFilterWorkflowContext);
   if (context === undefined) {
-    throw new Error('useGlobalFilters must be used within a GlobalFilterProvider');
+    throw new Error('useGlobalFilterWorkflow must be used within a GlobalFilterProvider');
   }
   return context;
 }
diff --git a/packages/app/src/components/about/agentx-faq.tsx b/packages/app/src/components/about/agentx-faq.tsx
index f878293e0..2220a3b63 100644
--- a/packages/app/src/components/about/agentx-faq.tsx
+++ b/packages/app/src/components/about/agentx-faq.tsx
@@ -199,8 +199,8 @@ export AIPERF_SERVICE_PROFILE_CONFIGURE_TIMEOUT=1800`,
           '保留原始轮次间延迟；仅将整个系统的全局空闲时间限制在 10 秒以内。',
           'first_turn_prefix cache busting 为每次回放的会话加入唯一首轮标记，避免虚假的跨会话 cache hit，同时保留会话内部的 cache 复用。',
           '有效测试至少运行 900 秒，默认运行 1,800 秒。',
-          '正式测量前通过 warmup 预热深层前缀。',
-          '每次运行都有随机种子，并记录在产物中。',
+          '正式测量前通过 warmup 建立深层 prefix。',
+          '每次运行都会使用 seed，并将其记录在产物中。',
         ],
       },
       {
@@ -233,7 +233,7 @@ export AIPERF_SERVICE_PROFILE_CONFIGURE_TIMEOUT=1800`,
       {
         title: '为什么第一次运行可能很久？',
         paragraphs: [
-          '发送流量前，AIPerf 会下载语料，并将其重建为完成 token 化且带有 cache 结构的会话树。结果会存入 memory-mapped 磁盘 cache。只要语料、tokenizer、重建配置、条目数量和随机种子相同，后续测试通常可以在数秒内恢复准备好的数据集。',
+          '发送流量前，AIPerf 会下载语料，并将其重建为完成 token 化且带有 cache 结构的会话树。结果会存入 memory-mapped 磁盘 cache。只要语料、tokenizer、重建配置、条目数量和 seed 相同，后续测试通常可以在数秒内恢复准备好的数据集。',
           '冷启动时应同时提高两个配置超时。随后依次执行数据集配置、warmup、定时 profiling 和 drain/export。Benchmark duration 仅覆盖 profiling，因此总耗时会更长。',
         ],
         code: `export AIPERF_DATASET_CONFIGURATION_TIMEOUT=1800
@@ -243,7 +243,7 @@ export AIPERF_SERVICE_PROFILE_CONFIGURE_TIMEOUT=1800`,
         title: '可以先跑一个短 smoke test 吗？',
         paragraphs: [
           '有效测试不能短于 900 秒。低成本有效检查可以降低 concurrency，但仍需保留 900 秒下限。--num-dataset-entries 能减少重建工作量，但会改变工作负载，不应将这种结果用于正式比较或提交。',
-          '如需在几分钟内检查连通性，可配合短 duration 使用 --unsafe-override，并接受 invalid 标记。保留完整语料并固定随机种子，可让后续有效测试复用重建后的数据集 cache。',
+          '如需在几分钟内检查连通性，可配合短 duration 使用 --unsafe-override，并接受 invalid 标记。保留完整语料并固定 seed，可让后续有效测试复用重建后的数据集 cache。',
         ],
       },
       {
@@ -282,7 +282,7 @@ export AIPERF_SERVICE_PROFILE_CONFIGURE_TIMEOUT=1800`,
         title: '实用原则',
         bullets: [
           '让语料与服务器上下文窗口匹配；约 256K 的服务器优先使用 _256k 语料。',
-          '正式比较使用带日期的固定语料和固定随机种子。',
+          '正式比较使用带日期的固定语料和固定 seed。',
           '保持 cache busting 开启；关闭后，循环使用轨迹会夸大 cache-hit 结果。',
           '保留记录的时序；压缩单条轨迹的延迟会改变 cache TTL 行为和会话重叠程度。',
           '在判断结果受服务器限制前，先监控客户端 CPU 和连接池等待时间。',
diff --git a/packages/app/src/components/about/faq-data-zh.ts b/packages/app/src/components/about/faq-data-zh.ts
index 2d161e70d..8b247b103 100644
--- a/packages/app/src/components/about/faq-data-zh.ts
+++ b/packages/app/src/components/about/faq-data-zh.ts
@@ -1,92 +1,77 @@
-import {
-  GPU_KEYS,
-  GPU_VENDORS,
-  DB_MODEL_TO_DISPLAY,
-  PRECISION_KEYS,
-  GITHUB_OWNER,
-  GITHUB_REPO,
-  FRAMEWORK_LABELS,
-} from '@semianalysisai/inferencex-constants';
-import { CAROUSEL_ORGS, CAROUSEL_LABELS } from '@/components/quotes/quotes-data';
+import { GITHUB_OWNER, GITHUB_REPO } from '@semianalysisai/inferencex-constants';
 
-import type { FaqItem } from './faq-data';
-
-/* ---------- Dynamic lists from constants ---------- */
-
-const gpusByVendor = [...GPU_KEYS].reduce<Record<string, string[]>>((acc, key) => {
-  const vendor = GPU_VENDORS[key] ?? 'Other';
-  (acc[vendor] ??= []).push(key.toUpperCase());
-  return acc;
-}, {});
-
-const modelNames = Object.values({
-  ...DB_MODEL_TO_DISPLAY,
-  'kimik2.6': 'Kimi-K2.6',
-  'kimik2.7-code': 'Kimi-K2.7-Code',
-  'minimaxm2.7': 'MiniMax-M2.7',
-  'glm5.1': 'GLM-5.1',
-});
-
-const frameworkNames = [...new Set(Object.values(FRAMEWORK_LABELS))].map((n) =>
-  n.replace(/[¹²³⁴⁵⁶⁷⁸⁹⁰]+$/u, ''),
-);
-
-const supporterOrgs = CAROUSEL_ORGS.map((org) => CAROUSEL_LABELS[org] ?? org);
+import { GENERATED_FAQ_DATA, type FaqItem } from '@/components/about/faq';
 
 /* ---------- FAQ content (Simplified Chinese) ---------- */
 
-export const FAQ_ITEMS_ZH: FaqItem[] = [
+export const FAQ_ITEMS_ZH: readonly FaqItem[] = [
   {
+    id: 'faq-what-is-inferencex',
     question: '什么是 InferenceX？',
     answer:
-      'InferenceX（原名 InferenceMAX）是一个开源、厂商中立的基准测试（benchmark）平台，持续衡量各类 Chip 和软件栈的 AI 推理性能。每当配置发生变化时，基准测试会重新运行，确保结果始终跟随模型和框架的演进保持最新。',
+      'InferenceX（原名 InferenceMAX）持续衡量各类芯片和软件栈的智能体推理与固定序列推理性能。AgentX 是其长上下文多轮编码场景。配置发生变化时，基准测试会重新运行。',
   },
   {
+    id: 'faq-who-builds-inferencex',
     question: 'InferenceX 由谁开发？',
-    answer: `InferenceX 由独立半导体与 AI 研究机构 SemiAnalysis 构建，受到 ${supporterOrgs.join('、')} 的支持与信赖。基准测试代码、数据和仪表板均在 GitHub 上开源。`,
+    answer: `InferenceX 由独立半导体与 AI 研究机构 SemiAnalysis 构建，受到 ${GENERATED_FAQ_DATA.supporterOrgs.join('、')} 的支持与信赖。基准测试代码、数据和仪表板均在 GitHub 上开源。`,
   },
   {
-    question: 'InferenceX 测试了哪些 Chip？',
+    id: 'faq-chips',
+    question: 'InferenceX 测试了哪些芯片？',
     answer: '我们会在新加速器可用时持续添加。',
-    list: Object.entries(gpusByVendor).map(([vendor, gpus]) => `${vendor}: ${gpus.join(', ')}`),
+    list: GENERATED_FAQ_DATA.gpuGroups,
   },
   {
+    id: 'faq-models',
     question: '测试了哪些 AI 模型？',
-    answer: '每个模型均在多种序列长度配置（1k/1k、1k/8k、8k/1k tokens）和并发级别下进行测试。',
-    list: modelNames,
+    answer:
+      '各模型会在其已有数据所覆盖的固定序列配置（1k/1k、1k/8k、8k/1k tokens）与多个并发级别下进行测试。具备对应数据的模型还包含 AgentX 长上下文多轮智能体编码运行。',
+    list: GENERATED_FAQ_DATA.modelNames,
   },
   {
+    id: 'faq-frameworks-configurations',
     question: '测试了哪些推理框架和配置？',
     answer: '',
     list: [
-      `框架：${frameworkNames.join(', ')}`,
-      `精度：${[...PRECISION_KEYS].map((p) => p.toUpperCase()).join(', ')}`,
+      `框架：${GENERATED_FAQ_DATA.frameworkNames.join(', ')}`,
+      `精度：${GENERATED_FAQ_DATA.precisionNames.join(', ')}`,
       '运行时：CUDA、ROCm',
-      '分离式推理（Disaggregated serving，独立的 prefill/decode Chip 池）',
+      '分离式推理（Disaggregated serving，独立的 prefill/decode 芯片池）',
       '多 token 预测（MTP）',
       '面向 MoE 模型的宽专家并行（Wide Expert Parallelism）',
     ],
   },
   {
+    id: 'faq-metrics',
     question: 'InferenceX 测量哪些指标？',
     answer: '',
     list: [
       '交互性（tok/s/user）',
-      '每 Chip token 吞吐量（tok/s/chip）',
-      '每 Chip 输入和输出吞吐量',
+      '每芯片 token 吞吐量（tok/s/chip）',
+      '每芯片输入和输出吞吐量',
       '每兆瓦 token 吞吐量（tok/s/MW）',
       'P99 首 token 延迟（TTFT）',
+      'AgentX 场景的端到端延迟、token 间延迟（ITL）、输出吞吐量、prefix cache 行为以及会话与 subagent 执行情况',
       '每百万 token 成本（总计、输入、输出）——涵盖超大规模云、NeoCoud 和裸机租赁定价',
       '每 token 能耗（焦耳，总计、输入、输出）',
       '用户自定义成本和功耗计算',
     ],
   },
   {
+    id: 'faq-normalized-interactivity',
+    question: '端到端归一化交互性与交互性有何区别？',
+    answer:
+      '交互性衡量生成开始后 token 的流式输出速度，近似为 token 间延迟（ITL）的倒数。端到端归一化交互性衡量整个请求期间的有效 token 速率，其中包含首 token 延迟（TTFT）：输出 token 数 / 端到端延迟，近似为 1 /（ITL + TTFT / 输出 token 数）。两者的单位均为 tok/s/user。TTFT 越长，归一化指标越低，短回答受到的影响尤其明显。这里的“归一化”是指按输出长度归一，并非 0–1 分数，也不是相对于其他系统的比较值。',
+  },
+  {
+    id: 'faq-benchmark-frequency',
     question: '基准测试多久运行一次？',
     answer:
       '基准测试最初按每日计划运行，但随着硬件/框架/模型组合数量的增长，这种方式已不再可行。现在，当配置发生变化（例如新软件发布、驱动更新或模型添加）时重新运行。仪表板中保留了历史数据。',
   },
   {
+    id: 'faq-open-source',
     question: 'InferenceX 是开源的吗？',
     answer: '是的。代码、数据和仪表板均为开源。',
     link: {
@@ -95,33 +80,39 @@ export const FAQ_ITEMS_ZH: FaqItem[] = [
     },
   },
   {
+    id: 'faq-benchmark-differences',
     question: 'InferenceX 与其他 AI 基准测试有何不同？',
     answer:
-      '大多数 AI 基准测试是静态的、单时间点测量，参与者提交的是专为基准测试定制的镜像，无法反映真实的线上推理性能。InferenceX 在真实硬件上持续运行，采用完全可复现的配置。所有测试脚本均提交至代码仓库，基准测试日志在 GitHub Actions 上公开可见，结果端到端可审计。',
+      'InferenceX 在真实硬件上运行固定序列工作负载与 AgentX 长上下文多轮编码场景。测试配方保存在代码仓库中，每项结果均链接至对应的 GitHub Actions 运行。',
   },
   {
+    id: 'faq-reproducibility',
     question: '结果如何实现可复现？',
     answer:
-      '仪表板上的每一个数据点均由公开的 GitHub Actions 工作流运行产生。测试配方（模型、框架、精度、并行度、序列长度、并发数）已提交至仓库，在目标硬件上实际执行，产物（日志、指标、Chip 追踪数据）上传至运行页面。用户可从任何图表的 tooltip 直接点击链接，跳转到生成该数据点的 GitHub Actions 运行。',
+      '仪表板上的每一个数据点均由公开的 GitHub Actions 工作流运行产生。测试配方（模型、框架、精度、并行度、序列长度、并发数）已提交至仓库，在目标硬件上实际执行，产物（日志、指标、芯片追踪数据）上传至运行页面。用户可从任何图表的 tooltip 直接点击链接，跳转到生成该数据点的 GitHub Actions 运行。',
   },
   {
+    id: 'faq-raw-logs',
     question: '在哪里可以查看原始基准测试日志？',
     answer:
-      '在图表上点击任意数据点即可打开 tooltip。其中的"GitHub Actions Run"链接将直接跳转到生成该数据点的工作流运行。在那里您可以查看完整的任务日志、框架和驱动版本、命令行参数，以及下载原始产物（包括请求延迟、token 计数和 Chip 功耗遥测数据）。',
+      '在图表上点击任意数据点即可打开 tooltip。其中的"GitHub Actions Run"链接将直接跳转到生成该数据点的工作流运行。在那里您可以查看完整的任务日志、框架和驱动版本、命令行参数，以及下载原始产物（包括请求延迟、token 计数和芯片功耗遥测数据）。',
   },
   {
+    id: 'faq-rerun-benchmark',
     question: '我可以自己重新运行基准测试吗？',
     answer:
       '可以。基准测试配方位于代码仓库的 /benchmarks 目录中，以独立的 shell 脚本形式存在。如果您拥有相同的硬件，可以 fork 仓库并直接运行脚本，或触发相同的 GitHub Actions 工作流来复现结果。',
   },
   {
+    id: 'faq-old-runs',
     question: '历史运行记录是否保留？',
     answer:
       '是的。GitHub Actions 保留工作流运行日志和产物 90 天。为了更长期的可审计性，我们还会每周发布完整基准测试数据库的快照作为公开的 GitHub Release，任何人都可以下载历史数据集并复现或重新分析仪表板中的任何图表。',
   },
   {
+    id: 'faq-data-use',
     question: '我可以使用 InferenceX 的数据进行自己的分析吗？',
     answer:
-      '可以。所有数据均可自由获取。仪表板支持按 Chip、模型、框架和日期范围筛选，您也可以直接从任何图表导出原始 CSV 数据。',
+      '可以。所有数据均可自由获取。仪表板支持按芯片、模型、框架和日期范围筛选，您也可以直接从任何图表导出原始 CSV 数据。',
   },
 ];
diff --git a/packages/app/src/components/about/faq-data.ts b/packages/app/src/components/about/faq-data.ts
index 6cd77e186..8438f5a25 100644
--- a/packages/app/src/components/about/faq-data.ts
+++ b/packages/app/src/components/about/faq-data.ts
@@ -1,82 +1,41 @@
-import {
-  GPU_KEYS,
-  GPU_VENDORS,
-  DB_MODEL_TO_DISPLAY,
-  PRECISION_KEYS,
-  GITHUB_OWNER,
-  GITHUB_REPO,
-  FRAMEWORK_LABELS,
-} from '@semianalysisai/inferencex-constants';
-import { CAROUSEL_ORGS, CAROUSEL_LABELS } from '@/components/quotes/quotes-data';
+import { GITHUB_OWNER, GITHUB_REPO } from '@semianalysisai/inferencex-constants';
 
-export interface FaqLink {
-  text: string;
-  href: string;
-}
-
-export interface FaqItem {
-  question: string;
-  /** Intro text shown before any list. */
-  answer: string;
-  /** Optional link rendered inline after the answer text. */
-  link?: FaqLink;
-  /** Optional bullet list rendered below the answer text. */
-  list?: string[];
-}
-
-/* ---------- Dynamic lists from constants ---------- */
-
-const gpusByVendor = [...GPU_KEYS].reduce<Record<string, string[]>>((acc, key) => {
-  const vendor = GPU_VENDORS[key] ?? 'Other';
-  (acc[vendor] ??= []).push(key.toUpperCase());
-  return acc;
-}, {});
-// /about lists each DB bucket as its own bullet, so point releases that
-// elsewhere collapse under one display name (see DB_MODEL_TO_DISPLAY) are
-// expanded back out here.
-const modelNames = Object.values({
-  ...DB_MODEL_TO_DISPLAY,
-  'kimik2.6': 'Kimi-K2.6',
-  'kimik2.7-code': 'Kimi-K2.7-Code',
-  'minimaxm2.7': 'MiniMax-M2.7',
-  'glm5.1': 'GLM-5.1',
-});
-
-const frameworkNames = [...new Set(Object.values(FRAMEWORK_LABELS))].map((n) =>
-  n.replace(/[¹²³⁴⁵⁶⁷⁸⁹⁰]+$/u, ''),
-);
-
-const supporterOrgs = CAROUSEL_ORGS.map((org) => CAROUSEL_LABELS[org] ?? org);
+import { GENERATED_FAQ_DATA, type FaqItem } from '@/components/about/faq';
 
 /* ---------- FAQ content ---------- */
 
-export const FAQ_ITEMS: FaqItem[] = [
+export const FAQ_ITEMS: readonly FaqItem[] = [
   {
+    id: 'faq-what-is-inferencex',
     question: 'What is InferenceX?',
     answer:
-      'InferenceX (formerly InferenceMAX) is an open-source, vendor-neutral benchmark that continuously measures AI inference performance across chips and software stacks. Benchmarks re-run whenever a configuration changes, so results stay current as models and frameworks evolve.',
+      'InferenceX (formerly InferenceMAX) continuously measures agentic and fixed-sequence inference performance across chips and software stacks. AgentX is its long-context, multi-turn coding scenario. Runs repeat whenever a configuration changes.',
   },
   {
+    id: 'faq-who-builds-inferencex',
     question: 'Who is behind InferenceX?',
-    answer: `InferenceX is built by SemiAnalysis, an independent semiconductor and AI research firm. It is supported and trusted by ${supporterOrgs.join(', ')}. The benchmark code, data, and dashboard are all open-source on GitHub.`,
+    answer: `InferenceX is built by SemiAnalysis, an independent semiconductor and AI research firm. It is supported and trusted by ${GENERATED_FAQ_DATA.supporterOrgs.join(', ')}. The benchmark code, data, and dashboard are all open-source on GitHub.`,
   },
   {
+    id: 'faq-chips',
     question: 'Which chips does InferenceX benchmark?',
     answer: 'New accelerators are added as they become available.',
-    list: Object.entries(gpusByVendor).map(([vendor, gpus]) => `${vendor}: ${gpus.join(', ')}`),
+    list: GENERATED_FAQ_DATA.gpuGroups,
   },
   {
+    id: 'faq-models',
     question: 'Which AI models are tested?',
     answer:
-      'Each model is tested across multiple sequence length configurations (1k/1k, 1k/8k, 8k/1k tokens) and concurrency levels.',
-    list: modelNames,
+      'Models are tested across the fixed-sequence configurations available for them (1k/1k, 1k/8k, and 8k/1k tokens) and multiple concurrency levels. Supported models with corresponding data also include AgentX long-context, multi-turn agentic coding runs.',
+    list: GENERATED_FAQ_DATA.modelNames,
   },
   {
+    id: 'faq-frameworks-configurations',
     question: 'Which inference frameworks and configurations are tested?',
     answer: '',
     list: [
-      `Frameworks: ${frameworkNames.join(', ')}`,
-      `Precisions: ${[...PRECISION_KEYS].map((p) => p.toUpperCase()).join(', ')}`,
+      `Frameworks: ${GENERATED_FAQ_DATA.frameworkNames.join(', ')}`,
+      `Precisions: ${GENERATED_FAQ_DATA.precisionNames.join(', ')}`,
       'Runtimes: CUDA, ROCm',
       'Disaggregated serving (separate prefill/decode chip pools)',
       'Multi-token prediction (MTP)',
@@ -84,6 +43,7 @@ export const FAQ_ITEMS: FaqItem[] = [
     ],
   },
   {
+    id: 'faq-metrics',
     question: 'What metrics does InferenceX measure?',
     answer: '',
     list: [
@@ -92,17 +52,26 @@ export const FAQ_ITEMS: FaqItem[] = [
       'Input and output throughput per chip',
       'Token throughput per MW (tok/s/MW)',
       'P99 time to first token (TTFT)',
+      'For AgentX: end-to-end latency, ITL, output throughput, prefix-cache behavior, and session/subagent execution',
       'Cost per million tokens (total, input, output) across hyperscaler, neocloud, and rental pricing',
       'Joules per token (total, input, output)',
       'Custom user-defined cost and power calculations',
     ],
   },
   {
+    id: 'faq-normalized-interactivity',
+    question: 'What is the difference between E2E Normalized Interactivity and Interactivity?',
+    answer:
+      'Interactivity measures how quickly tokens stream after generation begins: approximately 1 / inter-token latency (ITL). E2E Normalized Interactivity measures the effective token rate over the whole request, including time to first token (TTFT): output tokens / end-to-end latency, or approximately 1 / (ITL + TTFT / output tokens). Both use tok/s/user. The normalized value penalizes slow TTFT, especially for short responses. Here, "normalized" means normalized by output length; it is not a 0–1 score or a comparison with another system.',
+  },
+  {
+    id: 'faq-benchmark-frequency',
     question: 'How often are benchmarks run?',
     answer:
       'Benchmarks originally ran on a nightly schedule, but the number of hardware/framework/model combinations grew too large for that to be practical. Now they re-run when a configuration changes, e.g. a new software release, driver update, or model addition. Historical data is available in the dashboard.',
   },
   {
+    id: 'faq-open-source',
     question: 'Is InferenceX open source?',
     answer: 'Yes. Code, data, and dashboard are all open-source.',
     link: {
@@ -111,31 +80,37 @@ export const FAQ_ITEMS: FaqItem[] = [
     },
   },
   {
+    id: 'faq-benchmark-differences',
     question: 'How is InferenceX different from other AI benchmarks?',
     answer:
-      'Most AI benchmarks are static, point-in-time measurements where participants submit purpose-built images that do not reflect real-world serving performance. InferenceX runs continuously on real hardware with fully reproducible configurations. Every recipe is in the repo, benchmark logs are visible on GitHub Actions, and all results are auditable end-to-end.',
+      'InferenceX runs fixed-sequence workloads and the AgentX long-context, multi-turn coding scenario on real hardware. Test recipes are in the repository, and each result links to its GitHub Actions run.',
   },
   {
+    id: 'faq-reproducibility',
     question: 'How are results reproducible?',
     answer:
       'Every data point on the dashboard is produced by a public GitHub Actions workflow run. The recipe (model, framework, precision, parallelism, sequence length, concurrency) is committed to the repo, the run executes on the actual target hardware, and the resulting artifacts (logs, metrics, chip traces) are uploaded to the run page. Anyone can click through from a tooltip in any chart to the exact GitHub Actions run that produced that point.',
   },
   {
+    id: 'faq-raw-logs',
     question: 'Where can I see the raw benchmark logs?',
     answer:
       'Click any data point on a chart to open its tooltip. The "GitHub Actions Run" link goes directly to the workflow run that produced it. From there you can inspect the full job logs, the exact framework and driver versions, command line arguments, and download the raw artifacts including request latencies, token counts, and chip power telemetry.',
   },
   {
+    id: 'faq-rerun-benchmark',
     question: 'Can I rerun a benchmark myself?',
     answer:
       'Yes. The benchmark recipes live in the /benchmarks directory of the repo as standalone shell scripts. If you have access to the same hardware, you can fork the repo and run the script directly, or trigger the same GitHub Actions workflow to reproduce a result.',
   },
   {
+    id: 'faq-old-runs',
     question: 'Are old runs preserved?',
     answer:
       'Yes. GitHub Actions retains workflow run logs and artifacts for 90 days. For longer auditability, we also publish a weekly snapshot of the full benchmark database as a public GitHub Release, so anyone can download the historical dataset and reproduce or reanalyze any chart in the dashboard.',
   },
   {
+    id: 'faq-data-use',
     question: 'Can I use InferenceX data for my own analysis?',
     answer:
       'Yes. All data is freely available. The dashboard lets you filter by chip, model, framework, and date range, and you can export raw CSV data directly from any chart.',
diff --git a/packages/app/src/components/about/faq-question-link.tsx b/packages/app/src/components/about/faq-question-link.tsx
new file mode 100644
index 000000000..7195f56aa
--- /dev/null
+++ b/packages/app/src/components/about/faq-question-link.tsx
@@ -0,0 +1,81 @@
+'use client';
+
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { Check, LinkIcon } from 'lucide-react';
+
+import { track } from '@/lib/analytics';
+import { useLocale } from '@/lib/use-locale';
+
+const STRINGS = {
+  en: {
+    copyLink: 'Copy link',
+    copied: 'Copied',
+  },
+  zh: {
+    copyLink: '复制链接',
+    copied: '已复制',
+  },
+} as const;
+
+async function copyText(text: string) {
+  try {
+    await navigator.clipboard.writeText(text);
+  } catch {
+    const textArea = document.createElement('textarea');
+    textArea.value = text;
+    textArea.style.position = 'fixed';
+    textArea.style.opacity = '0';
+    document.body.append(textArea);
+    textArea.select();
+    document.execCommand('copy');
+    textArea.remove();
+  }
+}
+
+export function FaqQuestionLink({ id, question }: { id: string; question: string }) {
+  const locale = useLocale();
+  const strings = STRINGS[locale];
+  const [copied, setCopied] = useState(false);
+  const timerRef = useRef<ReturnType<typeof setTimeout>>(undefined);
+
+  useEffect(() => () => clearTimeout(timerRef.current), []);
+
+  const handleCopy = useCallback(async () => {
+    const url = new URL(window.location.href);
+    url.hash = id;
+    await copyText(url.toString());
+    track('about_faq_link_copied', { id });
+    setCopied(true);
+    clearTimeout(timerRef.current);
+    timerRef.current = setTimeout(() => setCopied(false), 2000);
+  }, [id]);
+
+  return (
+    <span className="flex items-start justify-between gap-3">
+      <a
+        href={`#${id}`}
+        onClick={() => track('about_faq_link_clicked', { id })}
+        className="min-w-0 flex-1 rounded-sm transition-colors hover:text-brand focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-brand focus-visible:ring-offset-2"
+      >
+        {question}
+      </a>
+      <button
+        type="button"
+        data-testid={`faq-copy-link-${id}`}
+        aria-label={`${copied ? strings.copied : strings.copyLink}: ${question}`}
+        title={copied ? strings.copied : strings.copyLink}
+        onClick={() => void handleCopy()}
+        className="inline-flex size-7 shrink-0 items-center justify-center rounded-md border border-border bg-background text-muted-foreground shadow-xs transition-colors hover:border-brand/50 hover:bg-muted hover:text-brand focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-brand focus-visible:ring-offset-2"
+      >
+        {copied ? (
+          <Check aria-hidden="true" className="size-4" />
+        ) : (
+          <LinkIcon aria-hidden="true" className="size-4" />
+        )}
+      </button>
+      <span aria-live="polite" className="sr-only">
+        {copied ? `${strings.copied}: ${question}` : ''}
+      </span>
+    </span>
+  );
+}
diff --git a/packages/app/src/components/about/faq.test.ts b/packages/app/src/components/about/faq.test.ts
new file mode 100644
index 000000000..465a3ac6b
--- /dev/null
+++ b/packages/app/src/components/about/faq.test.ts
@@ -0,0 +1,18 @@
+import { describe, expect, it } from 'vitest';
+
+import { FAQ_ITEMS } from '@/components/about/faq-data';
+import { FAQ_ITEMS_ZH } from '@/components/about/faq-data-zh';
+
+describe('FAQ anchors', () => {
+  it('gives every item a unique anchor', () => {
+    const ids = FAQ_ITEMS.map((item) => item.id);
+
+    expect(ids).toHaveLength(FAQ_ITEMS.length);
+    expect(new Set(ids).size).toBe(ids.length);
+    expect(ids.every((id) => id.startsWith('faq-'))).toBe(true);
+  });
+
+  it('keeps English and Chinese anchors in exact parity', () => {
+    expect(FAQ_ITEMS_ZH.map((item) => item.id)).toEqual(FAQ_ITEMS.map((item) => item.id));
+  });
+});
diff --git a/packages/app/src/components/about/faq.tsx b/packages/app/src/components/about/faq.tsx
new file mode 100644
index 000000000..d37d596d0
--- /dev/null
+++ b/packages/app/src/components/about/faq.tsx
@@ -0,0 +1,115 @@
+import {
+  DB_MODEL_TO_DISPLAY,
+  FRAMEWORK_LABELS,
+  GPU_KEYS,
+  GPU_VENDORS,
+  PRECISION_KEYS,
+} from '@semianalysisai/inferencex-constants';
+
+import { CAROUSEL_LABELS, CAROUSEL_ORGS } from '@/components/quotes/quotes-data';
+import { Card } from '@/components/ui/card';
+
+import { FaqQuestionLink } from './faq-question-link';
+
+export interface FaqLink {
+  readonly text: string;
+  readonly href: string;
+}
+
+export interface FaqItem {
+  /** Stable, locale-independent fragment identifier for direct links. */
+  readonly id: string;
+  readonly question: string;
+  /** Intro text shown before any list. */
+  readonly answer: string;
+  /** Optional link rendered inline after the answer text. */
+  readonly link?: FaqLink;
+  /** Optional bullet list rendered below the answer text. */
+  readonly list?: readonly string[];
+}
+
+const gpusByVendor = [...GPU_KEYS].reduce<Record<string, string[]>>((groups, key) => {
+  const vendor = GPU_VENDORS[key] ?? 'Other';
+  (groups[vendor] ??= []).push(key.toUpperCase());
+  return groups;
+}, {});
+
+export const GENERATED_FAQ_DATA = {
+  gpuGroups: Object.entries(gpusByVendor).map(([vendor, gpus]) => `${vendor}: ${gpus.join(', ')}`),
+  // /about lists each DB bucket separately, including point releases that other
+  // surfaces collapse under one display name.
+  modelNames: Object.values({
+    ...DB_MODEL_TO_DISPLAY,
+    'kimik2.6': 'Kimi-K2.6',
+    'kimik2.7-code': 'Kimi-K2.7-Code',
+    'minimaxm2.7': 'MiniMax-M2.7',
+    'glm5.1': 'GLM-5.1',
+  }),
+  frameworkNames: [...new Set(Object.values(FRAMEWORK_LABELS))].map((name) =>
+    name.replace(/[¹²³⁴⁵⁶⁷⁸⁹⁰]+$/u, ''),
+  ),
+  precisionNames: [...PRECISION_KEYS].map((precision) => precision.toUpperCase()),
+  supporterOrgs: CAROUSEL_ORGS.map((org) => CAROUSEL_LABELS[org] ?? org),
+} as const;
+
+export function buildFaqJsonLd(items: readonly FaqItem[], inLanguage?: string) {
+  return {
+    '@context': 'https://schema.org',
+    '@type': 'FAQPage',
+    ...(inLanguage && { inLanguage }),
+    mainEntity: items.map((item) => ({
+      '@type': 'Question',
+      name: item.question,
+      acceptedAnswer: {
+        '@type': 'Answer',
+        text: [item.answer, item.link?.text, ...(item.list ?? [])].filter(Boolean).join(' '),
+      },
+    })),
+  };
+}
+
+export function FaqList({ title, items }: { title: string; items: readonly FaqItem[] }) {
+  return (
+    <section>
+      <Card>
+        <h2 className="text-lg font-semibold mb-4">{title}</h2>
+        <dl className="divide-y divide-border">
+          {items.map((item) => (
+            <div id={item.id} key={item.id} className="scroll-mt-24 py-4 first:pt-0 last:pb-0">
+              <dt className="font-medium mb-2">
+                <FaqQuestionLink id={item.id} question={item.question} />
+              </dt>
+              <dd className="text-muted-foreground text-sm">
+                {item.answer && (
+                  <p>
+                    {item.answer}
+                    {item.link && (
+                      <>
+                        {' '}
+                        <a
+                          href={item.link.href}
+                          target="_blank"
+                          rel="noopener noreferrer"
+                          className="text-brand hover:underline font-medium"
+                        >
+                          {item.link.text}
+                        </a>
+                      </>
+                    )}
+                  </p>
+                )}
+                {item.list && (
+                  <ul className="mt-1.5 ml-8 list-disc space-y-0.5">
+                    {item.list.map((listItem) => (
+                      <li key={listItem}>{listItem}</li>
+                    ))}
+                  </ul>
+                )}
+              </dd>
+            </div>
+          ))}
+        </dl>
+      </Card>
+    </section>
+  );
+}
diff --git a/packages/app/src/components/about/promo-video.test.tsx b/packages/app/src/components/about/promo-video.test.tsx
new file mode 100644
index 000000000..b45b17b3d
--- /dev/null
+++ b/packages/app/src/components/about/promo-video.test.tsx
@@ -0,0 +1,104 @@
+// @vitest-environment jsdom
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const track = vi.fn();
+vi.mock('@/lib/analytics', () => ({ track: (...args: unknown[]) => track(...args) }));
+
+let locale: 'en' | 'zh' = 'en';
+vi.mock('@/lib/use-locale', () => ({ useLocale: () => locale }));
+
+import { PromoVideo } from '@/components/about/promo-video';
+
+let container: HTMLDivElement;
+let root: Root;
+let play: ReturnType<typeof vi.fn<() => Promise<void>>>;
+
+function renderUi(ui: React.ReactNode) {
+  act(() => root.render(ui));
+}
+
+function video(): HTMLVideoElement {
+  const el = container.querySelector('video');
+  if (!el) throw new Error('video element not rendered');
+  return el;
+}
+
+beforeEach(() => {
+  locale = 'en';
+  track.mockClear();
+  // jsdom has no media stack, so HTMLMediaElement.play is undefined by default.
+  play = vi.fn<() => Promise<void>>(() => Promise.resolve());
+  HTMLMediaElement.prototype.play = play;
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+});
+
+describe('PromoVideo', () => {
+  it('defers the download until the visitor asks for it', () => {
+    renderUi(<PromoVideo />);
+    // preload="none" is the whole point: a ~30 MB asset must not land on the
+    // critical path of /about for the majority who never press play.
+    expect(video().getAttribute('preload')).toBe('none');
+    expect(video().getAttribute('poster')).toBe('/promo-poster.webp');
+  });
+
+  it('renders an mp4 source and no controls before playback starts', () => {
+    renderUi(<PromoVideo />);
+    const source = container.querySelector('source');
+    expect(source?.getAttribute('type')).toBe('video/mp4');
+    expect(source?.getAttribute('src')).toMatch(/^https:\/\//u);
+    expect(video().hasAttribute('controls')).toBe(false);
+  });
+
+  it('plays, reveals native controls, and hides the overlay on click', () => {
+    renderUi(<PromoVideo />);
+    const overlay = container.querySelector('button');
+    expect(overlay).not.toBeNull();
+
+    act(() => overlay!.click());
+
+    expect(play).toHaveBeenCalledTimes(1);
+    expect(video().hasAttribute('controls')).toBe(true);
+    expect(container.querySelector('button')).toBeNull();
+  });
+
+  it('tracks the play interaction', () => {
+    renderUi(<PromoVideo />);
+    act(() => container.querySelector('button')!.click());
+    expect(track).toHaveBeenCalledWith('about_promo_video_play');
+  });
+
+  it('hides the overlay when playback starts without the overlay button', () => {
+    // Native controls, keyboard, or the OS media keys can start playback
+    // without routing through our overlay; the overlay must still get out of
+    // the way rather than covering the running video.
+    renderUi(<PromoVideo />);
+    act(() => {
+      video().dispatchEvent(new Event('play'));
+    });
+    expect(container.querySelector('button')).toBeNull();
+  });
+
+  it('labels the player in English on /about', () => {
+    renderUi(<PromoVideo />);
+    expect(video().getAttribute('aria-label')).toBe('Play the InferenceX overview video');
+    expect(container.querySelector('button')?.getAttribute('aria-label')).toBe(
+      'Play the InferenceX overview video',
+    );
+  });
+
+  it('labels the player in Chinese on /zh/about', () => {
+    locale = 'zh';
+    renderUi(<PromoVideo />);
+    expect(video().getAttribute('aria-label')).toBe('播放 InferenceX 概览视频');
+    expect(container.textContent).toContain('您的浏览器无法播放此视频。');
+  });
+});
diff --git a/packages/app/src/components/about/promo-video.tsx b/packages/app/src/components/about/promo-video.tsx
new file mode 100644
index 000000000..c15f34222
--- /dev/null
+++ b/packages/app/src/components/about/promo-video.tsx
@@ -0,0 +1,92 @@
+'use client';
+
+import { Play } from 'lucide-react';
+import { useRef, useState } from 'react';
+
+import { track } from '@/lib/analytics';
+import { useLocale } from '@/lib/use-locale';
+
+/**
+ * Public Vercel Blob URL for the promo reel.
+ *
+ * The encode lives in Blob rather than `public/` on purpose: at ~30 MB it would
+ * be permanent git history and dead weight in every clone and deployment
+ * bundle, for an asset that never needs to version alongside the code. Blob
+ * serves it from the same CDN with HTTP range requests, so seeking and scrubbing
+ * work exactly as they would from `public/`.
+ *
+ * The poster stays in `public/` — it is 68 KB, and inlining it avoids a
+ * cross-origin round trip on a frame that renders before any user interaction.
+ */
+const PROMO_VIDEO_SRC =
+  'https://yig6saydz8oscerh.public.blob.vercel-storage.com/media/inferencex-promo-v1-1440p-B9VFFfZ9V0Hl7tciu0F0aLrRW4WbKD.mp4';
+const PROMO_POSTER_SRC = '/promo-poster.webp';
+
+const STRINGS = {
+  en: {
+    play: 'Play the InferenceX overview video',
+    fallback: 'Your browser cannot play this video.',
+    download: 'Download the video instead',
+  },
+  zh: {
+    play: '播放 InferenceX 概览视频',
+    fallback: '您的浏览器无法播放此视频。',
+    download: '直接下载视频',
+  },
+} as const;
+
+/**
+ * Click-to-play promo reel for the About page.
+ *
+ * `preload="none"` means nothing but the poster is fetched until a visitor
+ * actually asks for the video, so the ~30 MB payload never lands on the critical
+ * path or counts against LCP for the majority who scroll straight past it.
+ */
+export function PromoVideo() {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const videoRef = useRef<HTMLVideoElement>(null);
+  const [started, setStarted] = useState(false);
+
+  function handlePlay() {
+    setStarted(true);
+    track('about_promo_video_play');
+    void videoRef.current?.play();
+  }
+
+  return (
+    <div className="relative aspect-video w-full overflow-hidden rounded-xl border border-border/40 bg-black">
+      <video
+        ref={videoRef}
+        className="size-full"
+        poster={PROMO_POSTER_SRC}
+        preload="none"
+        playsInline
+        controls={started}
+        aria-label={t.play}
+        onPlay={() => setStarted(true)}
+      >
+        <source src={PROMO_VIDEO_SRC} type="video/mp4" />
+        <p className="p-4 text-sm text-muted-foreground">
+          {t.fallback}{' '}
+          <a href={PROMO_VIDEO_SRC} className="text-brand hover:underline">
+            {t.download}
+          </a>
+        </p>
+      </video>
+
+      {!started && (
+        <button
+          type="button"
+          onClick={handlePlay}
+          aria-label={t.play}
+          className="group absolute inset-0 flex cursor-pointer items-center justify-center bg-black/20 transition-colors hover:bg-black/10 focus-visible:outline-2 focus-visible:outline-offset-2 focus-visible:outline-brand"
+        >
+          <span className="flex size-16 items-center justify-center rounded-full bg-background/80 shadow-lg backdrop-blur-sm transition-transform group-hover:scale-110 md:size-20">
+            <Play className="ml-1 size-7 fill-foreground text-foreground md:size-9" />
+          </span>
+        </button>
+      )}
+    </div>
+  );
+}
diff --git a/packages/app/src/components/ai-chart/AiChartDisplay.tsx b/packages/app/src/components/ai-chart/AiChartDisplay.tsx
index d43d6a06c..4cb5229af 100644
--- a/packages/app/src/components/ai-chart/AiChartDisplay.tsx
+++ b/packages/app/src/components/ai-chart/AiChartDisplay.tsx
@@ -1,6 +1,6 @@
 'use client';
 
-import { useCallback, useState } from 'react';
+import { useCallback, useEffect, useState } from 'react';
 import { AlertCircle, Eye, EyeOff, Sparkles } from 'lucide-react';
 
 import { track } from '@/lib/analytics';
@@ -66,10 +66,15 @@ export default function AiChartDisplay() {
   });
   const [prompt, setPrompt] = useState('');
   const [showKey, setShowKey] = useState(false);
+  const [isMac, setIsMac] = useState(false);
   const { result, isLoading, error, generate, reset } = useAiChart();
   const locale = useLocale();
   const t = STRINGS[locale];
 
+  useEffect(() => {
+    setIsMac(navigator.userAgent.includes('Mac'));
+  }, []);
+
   const apiKey = apiKeys[provider];
 
   const handleProviderChange = useCallback((value: string) => {
@@ -157,7 +162,7 @@ export default function AiChartDisplay() {
             />
             <div className="flex items-center justify-between">
               <span className="text-muted-foreground text-xs">
-                {navigator.userAgent.includes('Mac') ? '⌘' : 'Ctrl'}
+                {isMac ? '⌘' : 'Ctrl'}
                 {t.enterToGenerate}
               </span>
               <Button
diff --git a/packages/app/src/components/ai-chart/AiChartResult.tsx b/packages/app/src/components/ai-chart/AiChartResult.tsx
index b43cee2ea..f7b1f378b 100644
--- a/packages/app/src/components/ai-chart/AiChartResult.tsx
+++ b/packages/app/src/components/ai-chart/AiChartResult.tsx
@@ -233,7 +233,9 @@ function LineChart({
   const flatPoints = useMemo<LinePoint[]>(
     () =>
       Object.entries(lineData).flatMap(([hwKey, pts]) =>
-        pts.map((p) => ({ hwKey, precision: 'fp8', x: p.x, y: p.y })),
+        pts
+          .filter((point) => Number.isFinite(point.x) && Number.isFinite(point.y))
+          .map((p) => ({ hwKey, precision: 'fp8', x: p.x, y: p.y })),
       ),
     [lineData],
   );
@@ -267,6 +269,7 @@ function LineChart({
       config: {
         getColor: (key) => colorMap[key] ?? '#888',
         strokeWidth: 2.5,
+        isDefined: ({ x, y }) => Number.isFinite(x) && Number.isFinite(y),
       },
     };
     const scatterLayer: ScatterLayerConfig<LinePoint> = {
diff --git a/packages/app/src/components/ai-chart/prompt-templates.ts b/packages/app/src/components/ai-chart/prompt-templates.ts
index 59beedfea..91b2e5e91 100644
--- a/packages/app/src/components/ai-chart/prompt-templates.ts
+++ b/packages/app/src/components/ai-chart/prompt-templates.ts
@@ -71,7 +71,7 @@ Pick the chart type that best matches the user's intent. Be flexible — interpr
 - **bar** — Horizontal bar chart. Compares one metric across GPU configs at a fixed interactivity point. Best for rankings, comparisons, "which is best". DEFAULT.
 - **scatter** — XY scatter plot (x=interactivity, y=metric). Shows all data points. Good for trade-off analysis, pareto frontiers.
 - **line** — Connected lines (x=interactivity, y=metric), one line per GPU config. Good for seeing how performance changes with load, comparing curves.
-- **radar** — Spider/radar chart. Compares GPUs across MULTIPLE metrics simultaneously on normalized axes. Set radarMetrics to choose which metrics are axes (pick 3-6 that are relevant). Cost/energy metrics are auto-inverted (lower=better=further out).
+- **radar** — Spider/radar chart. Compares GPUs across MULTIPLE metrics simultaneously on normalized axes. Set radarMetrics to choose which metrics are axes (pick 3-6 that are relevant). Metric direction comes from the dashboard definition, so lower-is-better metrics render better values farther out.
 
 ## Filtering
 
@@ -82,12 +82,12 @@ Pick the chart type that best matches the user's intent. Be flexible — interpr
 
 ## Sorting & Sampling
 
-- **sortOrder** ("registry"): Sort order for bar charts. "desc" = highest value first, "asc" = lowest first, "registry" = default GPU registry order. Use "desc" when user asks for "best" or "rank by".
+- **sortOrder** ("registry"): Sort order for bar charts. "desc" = highest value first, "asc" = lowest first, "registry" = default GPU registry order. When the user asks for "best" or "rank by", use "asc" for lower-is-better metrics such as cost/energy/power and "desc" for higher-is-better metrics such as throughput.
 - **targetInteractivity** (40): The interactivity level (tok/s/user) to sample at for bar charts. Adjust if user specifies a concurrency level.
 
 ## Top-N
 
-- **topN** (null): "top 3 GPUs" → topN: 3. Ranks ALL configs by peak metric value after data loads. Don't guess — let data decide.
+- **topN** (null): "top 3 GPUs" → topN: 3. Ranks ALL configs by their best observed value in the metric's dashboard-defined direction after data loads (minimum for lower-is-better; maximum for higher-is-better). Don't guess — let data decide.
 - **topNDistinctGpus** (true): When true, topN picks the best config from each unique GPU family (so "top 2 GPUs" = 2 different GPU types). Set false when user says "top N configs" or "top N combos" (may return same GPU with different frameworks).
 
 ## Multi-Chart
diff --git a/packages/app/src/components/ai-chart/types.test.ts b/packages/app/src/components/ai-chart/types.test.ts
new file mode 100644
index 000000000..1837e1c04
--- /dev/null
+++ b/packages/app/src/components/ai-chart/types.test.ts
@@ -0,0 +1,15 @@
+import { describe, expect, it } from 'vitest';
+
+import { validateSpec } from './types';
+
+describe('validateSpec measured power axes', () => {
+  it.each([
+    'y_measuredJPerSuccessfulQuery',
+    'y_measuredWhPerSuccessfulQuery',
+    'y_measuredPowerPercentTdp',
+  ])('preserves %s as a benchmark Y-axis metric', (yAxisMetric) => {
+    const spec = validateSpec({ dataSource: 'benchmarks', yAxisMetric });
+
+    expect(spec.yAxisMetric).toBe(yAxisMetric);
+  });
+});
diff --git a/packages/app/src/components/blog/mdx-components.tsx b/packages/app/src/components/blog/mdx-components.tsx
index 68b791ef5..b684c78e4 100644
--- a/packages/app/src/components/blog/mdx-components.tsx
+++ b/packages/app/src/components/blog/mdx-components.tsx
@@ -4,6 +4,7 @@ import Link from 'next/link';
 import { slugify } from '@/lib/blog';
 import type { Locale } from '@/lib/i18n';
 import { HeadingLink } from '@/components/blog/heading-link';
+import { ThemedFigureImage } from '@/components/blog/themed-figure-image';
 import { JsonLd } from '@/components/json-ld';
 
 function childrenToText(children: ReactNode): string {
@@ -144,26 +145,13 @@ export function createMdxComponents(
       return (
         <figure className="my-6 flex flex-col items-center">
           {hasThemedVariants ? (
-            <>
-              {lightSrc && (
-                <img
-                  src={lightSrc}
-                  alt={props.alt ?? ''}
-                  loading={loading}
-                  decoding="async"
-                  className="rounded-lg w-full md:w-3/4 block dark:hidden"
-                />
-              )}
-              {darkSrc && (
-                <img
-                  src={darkSrc}
-                  alt={props.alt ?? ''}
-                  loading={loading}
-                  decoding="async"
-                  className="rounded-lg w-full md:w-3/4 hidden dark:block"
-                />
-              )}
-            </>
+            <ThemedFigureImage
+              srcLight={lightSrc}
+              srcDark={darkSrc}
+              alt={props.alt ?? ''}
+              loading={loading}
+              className="rounded-lg w-full md:w-3/4"
+            />
           ) : (
             (lightSrc || darkSrc) && (
               <img
diff --git a/packages/app/src/components/blog/themed-figure-image.test.tsx b/packages/app/src/components/blog/themed-figure-image.test.tsx
new file mode 100644
index 000000000..a2092a2dd
--- /dev/null
+++ b/packages/app/src/components/blog/themed-figure-image.test.tsx
@@ -0,0 +1,137 @@
+// @vitest-environment jsdom
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { renderToStaticMarkup } from 'react-dom/server';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const themeState = vi.hoisted(() => ({ resolvedTheme: undefined as string | undefined }));
+
+vi.mock('next-themes', () => ({
+  useTheme: () => themeState,
+}));
+
+import { createMdxComponents } from '@/components/blog/mdx-components';
+import { ThemedFigureImage } from '@/components/blog/themed-figure-image';
+
+let container: HTMLDivElement;
+let root: Root;
+
+function renderUi(ui: React.ReactNode) {
+  act(() => root.render(ui));
+}
+
+function runPrepaintScript(target = container) {
+  const script = target.querySelector('script');
+  expect(script).not.toBeNull();
+  Object.defineProperty(document, 'currentScript', {
+    configurable: true,
+    value: script,
+  });
+  try {
+    // oxlint-disable-next-line no-new-func, unicorn/new-for-builtins -- execute the emitted parser-time bootstrap verbatim.
+    Function(script?.textContent ?? '')();
+  } finally {
+    Reflect.deleteProperty(document, 'currentScript');
+  }
+}
+
+beforeEach(() => {
+  themeState.resolvedTheme = undefined;
+  document.documentElement.className = '';
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  document.documentElement.className = '';
+});
+
+describe('ThemedFigureImage', () => {
+  it.each([
+    ['light', '/chart-light.png'],
+    ['dark', '/chart-dark.png'],
+    ['minecraft', '/chart-dark.png'],
+  ])('gives an eager cold %s render exactly one active source', (theme, expectedSrc) => {
+    document.documentElement.className = theme;
+    renderUi(
+      <ThemedFigureImage
+        srcLight="/chart-light.png"
+        srcDark="/chart-dark.png"
+        alt="Throughput chart"
+        loading="eager"
+        className="chart"
+      />,
+    );
+
+    const image = container.querySelector(':scope > img');
+    expect(image?.getAttribute('src')).toBe(expectedSrc);
+    expect(container.querySelectorAll(':scope > img[src]')).toHaveLength(1);
+    expect(image?.getAttribute('loading')).toBe('eager');
+    expect(image?.getAttribute('alt')).toBe('Throughput chart');
+  });
+
+  it('selects one request-bearing source in server markup before paint', () => {
+    document.documentElement.className = 'light';
+    const serverContainer = document.createElement('div');
+    serverContainer.innerHTML = renderToStaticMarkup(
+      <ThemedFigureImage
+        srcLight="/chart-light.png"
+        srcDark="/chart-dark.png"
+        alt="Throughput chart"
+        loading="eager"
+        className="chart"
+      />,
+    );
+    const image = serverContainer.querySelector(':scope > img');
+    expect(image?.hasAttribute('src')).toBe(false);
+
+    runPrepaintScript(serverContainer);
+
+    expect(image?.getAttribute('src')).toBe('/chart-light.png');
+  });
+
+  it('updates the same image when the application theme changes', () => {
+    themeState.resolvedTheme = 'light';
+    renderUi(
+      <ThemedFigureImage
+        srcLight="/chart-light.png"
+        srcDark="/chart-dark.png"
+        alt="Throughput chart"
+        loading="lazy"
+        className="chart"
+      />,
+    );
+    const image = container.querySelector(':scope > img');
+    expect(image?.getAttribute('src')).toBe('/chart-light.png');
+
+    themeState.resolvedTheme = 'dark';
+    renderUi(
+      <ThemedFigureImage
+        srcLight="/chart-light.png"
+        srcDark="/chart-dark.png"
+        alt="Throughput chart"
+        loading="lazy"
+        className="chart"
+      />,
+    );
+
+    expect(container.querySelectorAll(':scope > img')).toHaveLength(1);
+    expect(container.querySelector(':scope > img')).toBe(image);
+    expect(image?.getAttribute('src')).toBe('/chart-dark.png');
+  });
+
+  it('leaves an unthemed MDX figure on the existing plain image path', () => {
+    const Figure = createMdxComponents().Figure;
+    renderUi(<Figure src="/plain.png" alt="Plain diagram" />);
+
+    const image = container.querySelector('img');
+    expect(container.querySelector('script')).toBeNull();
+    expect(image?.getAttribute('src')).toBe('/plain.png');
+    expect(image?.dataset.srcLight).toBeUndefined();
+    expect(image?.getAttribute('loading')).toBe('eager');
+    expect(image?.getAttribute('alt')).toBe('Plain diagram');
+  });
+});
diff --git a/packages/app/src/components/blog/themed-figure-image.tsx b/packages/app/src/components/blog/themed-figure-image.tsx
new file mode 100644
index 000000000..a17b3fbc9
--- /dev/null
+++ b/packages/app/src/components/blog/themed-figure-image.tsx
@@ -0,0 +1,72 @@
+'use client';
+
+import { useSyncExternalStore } from 'react';
+import { useTheme } from 'next-themes';
+
+const THEMED_FIGURE_BOOTSTRAP = `(()=>{const s=document.currentScript,i=s&&s.previousElementSibling;if(!(i instanceof HTMLImageElement))return;const l=document.documentElement.classList.contains('light'),v=l?i.dataset.srcLight:i.dataset.srcDark;if(v)i.setAttribute('src',v)})()`;
+
+const subscribeClient = () => () => {};
+const getClientSnapshot = () => true;
+const getServerSnapshot = () => false;
+
+interface ThemedFigureImageProps {
+  srcLight?: string;
+  srcDark?: string;
+  alt: string;
+  loading: 'eager' | 'lazy';
+  className: string;
+}
+
+/**
+ * The server leaves `src` out of the hydrated image, then a parser-time script
+ * selects the saved root theme before first paint. Client-created figures use
+ * the live theme or root class directly. The noscript image preserves article
+ * content for non-JavaScript readers and crawlers.
+ */
+export function ThemedFigureImage({
+  srcLight,
+  srcDark,
+  alt,
+  loading,
+  className,
+}: ThemedFigureImageProps) {
+  const { resolvedTheme } = useTheme();
+  const isClient = useSyncExternalStore(subscribeClient, getClientSnapshot, getServerSnapshot);
+  const clientTheme =
+    resolvedTheme ??
+    (isClient && document.documentElement.classList.contains('light') ? 'light' : 'dark');
+  const src = isClient
+    ? clientTheme === 'dark' || clientTheme === 'minecraft'
+      ? srcDark
+      : srcLight
+    : undefined;
+
+  return (
+    <>
+      <img
+        src={src}
+        data-src-light={srcLight}
+        data-src-dark={srcDark}
+        alt={alt}
+        loading={loading}
+        decoding="async"
+        className={className}
+        suppressHydrationWarning
+      />
+      <script
+        aria-hidden="true"
+        suppressHydrationWarning
+        dangerouslySetInnerHTML={{ __html: THEMED_FIGURE_BOOTSTRAP }}
+      />
+      <noscript>
+        <img
+          src={srcDark ?? srcLight}
+          alt={alt}
+          loading={loading}
+          decoding="async"
+          className={className}
+        />
+      </noscript>
+    </>
+  );
+}
diff --git a/packages/app/src/components/calculator/CalculatorTable.tsx b/packages/app/src/components/calculator/CalculatorTable.tsx
index d5f75f864..9fab792d6 100644
--- a/packages/app/src/components/calculator/CalculatorTable.tsx
+++ b/packages/app/src/components/calculator/CalculatorTable.tsx
@@ -50,7 +50,7 @@ const STRINGS = {
     costPrefix: '成本 (',
     costSuffix: ')',
     concurrency: '并发数',
-    footer: '数值基于真实 InferenceMAX 基准测试数据插值计算。仅显示在测量范围内有数据的 Chip。',
+    footer: '数值基于真实 InferenceMAX 基准测试数据插值计算。仅显示在测量范围内有数据的芯片。',
   },
 } as const;
 
diff --git a/packages/app/src/components/calculator/CostTargetPanel.tsx b/packages/app/src/components/calculator/CostTargetPanel.tsx
new file mode 100644
index 000000000..dc86001a8
--- /dev/null
+++ b/packages/app/src/components/calculator/CostTargetPanel.tsx
@@ -0,0 +1,379 @@
+'use client';
+
+import Link from 'next/link';
+import { useCallback, useMemo, useState } from 'react';
+
+import { track } from '@/lib/analytics';
+import type { HardwareConfig } from '@/components/inference/types';
+import { Card } from '@/components/ui/card';
+import { CollapsibleSection } from '@/components/ui/collapsible-section';
+import { type DataTableColumn, DataTable } from '@/components/ui/data-table';
+import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
+import { Input } from '@/components/ui/input';
+import { LabelWithTooltip } from '@/components/ui/label-with-tooltip';
+import { TooltipProvider } from '@/components/ui/tooltip';
+import { getGpuSpecs, getHardwareConfig } from '@/lib/constants';
+import { readUrlParams, writeUrlParams } from '@/lib/url-state';
+import { getDisplayLabel } from '@/lib/utils';
+import { useLocale } from '@/lib/use-locale';
+
+import { computeFleetStats, formatCompact } from './fleet';
+import { interpolateForGPU, maxInteractivityAtCost } from './interpolation';
+import { outputTokPerChip } from './lifecycle';
+import { getCostProviderLabel, getThroughputForType } from './ThroughputBarChart';
+import type { CostProvider, CostType, GPUDataPoint } from './types';
+
+/**
+ * The inverse of the target-interactivity slider: the slider fixes a speed and
+ * reads off the cost, this fixes a cost ceiling and reads off the highest speed
+ * that stays under it. That is why it sits directly beneath the slider rather
+ * than among the fleet-economics sections — it answers a question about the
+ * operating point, not about a deployment's lifetime.
+ *
+ * Deliberately independent of the slider's own value: `gpuDataByGroupKey` is the
+ * unfiltered frontier, so a chip that interpolates to nothing at the current
+ * target still gets an answer here.
+ */
+interface CostTargetPanelProps {
+  gpuDataByGroupKey: Record<string, GPUDataPoint[]>;
+  hardwareConfig: HardwareConfig;
+  costProvider: CostProvider;
+  costType: CostType;
+  /** Legend visibility by base hwKey. */
+  visibleHwKeys: Set<string>;
+  /**
+   * Facility power budget, owned by the page. Only used to turn the answer into
+   * concurrent users; the interactivity result itself does not need it.
+   */
+  mw: number | null;
+  /** Raw MW input value (page-owned so the `c_mw` URL param has one writer). */
+  mwInput: string;
+  onMwInputChange: (raw: string) => void;
+}
+
+const STRINGS = {
+  en: {
+    toggleSection: 'Expand or fold this section',
+    costCapTitle: 'Interactivity Within a Cost Target',
+    costCapDescription:
+      'Set a cost ceiling per million tokens and find the highest interactivity each chip can serve without exceeding it.',
+    costCapLabel: 'Cost Target ($/M tok)',
+    costCapTooltip:
+      'Maximum acceptable cost per million tokens (at the selected pricing tier and token type). The answer is the highest interactivity whose interpolated cost stays at or below this ceiling.',
+    costCapPlaceholder: 'e.g. 0.50',
+    mwLabel: 'Facility Power (MW)',
+    mwTooltip:
+      'Optional facility power budget in megawatts. When set, the answer is also expressed as concurrent users on a fleet sized to this budget, using all-in power per chip (host, networking, cooling) from the SemiAnalysis Datacenter Industry Model, not bare TDP.',
+    mwPlaceholder: 'e.g. 10',
+    colGpu: 'Chip',
+    colMaxInteractivity: 'Max Interactivity (tok/s/user)',
+    colTputAtIv: 'Throughput (tok/s/chip)',
+    colUsers: 'Concurrent Users',
+    notReachable: 'Not reachable',
+    costCapEmpty: 'Enter a cost target to find the serveable interactivity per chip.',
+    costCapNoGpus: 'No chips are visible to evaluate — enable hardware in the chart legend.',
+    note: 'Note:',
+    disaggFleet:
+      ' Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) report input and output throughput per prefill chip and per decode chip rather than per chip overall, so the interactivity a cost ceiling permits inherits that per-pool basis on the Input and Output token types. On Total it is per chip overall for both kinds.',
+    assumptions: (tier: string) =>
+      `Assumes 100% utilization at the answering operating point and owned-datacenter economics at the ${tier} tier. Concurrent users, when shown, is re-based onto chips overall.`,
+    source: 'Source: ',
+  },
+  zh: {
+    toggleSection: '展开或折叠此板块',
+    costCapTitle: '成本上限下的交互性',
+    costCapDescription:
+      '设定每百万 token 的成本上限，查看每款芯片在不超支前提下可提供的最高交互性。',
+    costCapLabel: '成本上限 ($/M tok)',
+    costCapTooltip:
+      '每百万 token 的最高可接受成本（按所选定价层级和 token 类型）。结果为插值成本不超过该上限的最高交互性。',
+    costCapPlaceholder: '如 0.50',
+    mwLabel: '设施功率 (MW)',
+    mwTooltip:
+      '可选的设施功率预算（兆瓦）。设置后，结果会同时换算为按该预算确定规模的集群可服务的并发用户数；芯片数量按每芯片全含功率（主机、网络、散热）计算，数据来自 SemiAnalysis Datacenter Industry Model，而非裸 TDP。',
+    mwPlaceholder: '如 10',
+    colGpu: '芯片',
+    colMaxInteractivity: '最高交互性 (tok/s/user)',
+    colTputAtIv: '吞吐量 (tok/s/chip)',
+    colUsers: '并发用户数',
+    notReachable: '无法达到',
+    costCapEmpty: '输入成本上限，以查看每款芯片可服务的交互性。',
+    costCapNoGpus: '当前无可见芯片可评估。请在图表图例中启用硬件。',
+    note: '注：',
+    disaggFleet:
+      ' 解耦推理配置（如 MoRI SGLang、Dynamo TRTLLM）的输入与输出吞吐量分别按预填充芯片与解码芯片报告，而非按芯片总数，因此在「输入」与「输出」token 类型下，成本上限所允许的交互性沿用了上述按池计的口径。在「总计」口径下，两种部署方式均按芯片总数计。',
+    assumptions: (tier: string) =>
+      `假设结果操作点下 100% 利用率，并采用 ${tier} 层级的自有数据中心经济模型。如显示并发用户数，则已改按芯片总数计。`,
+    source: '来源：',
+  },
+} as const;
+
+function getLabel(
+  r: { hwKey: string; precision?: string },
+  hardwareConfig: HardwareConfig,
+): string {
+  const config = hardwareConfig[r.hwKey] || getHardwareConfig(r.hwKey);
+  const baseName = config ? getDisplayLabel(config) : r.hwKey;
+  return r.precision ? `${baseName} (${r.precision.toUpperCase()})` : baseName;
+}
+
+function parsePositive(raw: string): number | null {
+  const parsed = parseFloat(raw);
+  return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
+}
+
+interface CostCapRow {
+  resultKey: string;
+  hwKey: string;
+  precision?: string;
+  /** null = cost target not reachable anywhere on this chip's frontier */
+  maxInteractivity: number | null;
+  tputAtIv: number | null;
+  users: number | null;
+}
+
+export default function CostTargetPanel({
+  gpuDataByGroupKey,
+  hardwareConfig,
+  costProvider,
+  costType,
+  visibleHwKeys,
+  mw,
+  mwInput,
+  onMwInputChange,
+}: CostTargetPanelProps) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+
+  const [costCapInput, setCostCapInput] = useState<string>(() => readUrlParams().c_costcap ?? '');
+  const costCap = useMemo(() => parsePositive(costCapInput), [costCapInput]);
+
+  const handleCostCapChange = useCallback((e: React.ChangeEvent<HTMLInputElement>) => {
+    setCostCapInput(e.target.value);
+    writeUrlParams({ c_costcap: parsePositive(e.target.value) ? e.target.value : '' });
+  }, []);
+
+  const handleCostCapBlur = useCallback(() => {
+    track('calculator_cost_target_set', { costCap: costCapInput });
+  }, [costCapInput]);
+
+  // Gated ONLY by the legend's hw visibility — never by results interpolated at
+  // the current slider target, which would hide a chip whose frontier is fine.
+  const visibleGroupKeys = useMemo(
+    () =>
+      new Set(
+        Object.keys(gpuDataByGroupKey).filter((groupKey) =>
+          visibleHwKeys.has(groupKey.includes('__') ? groupKey.split('__')[0] : groupKey),
+        ),
+      ),
+    [gpuDataByGroupKey, visibleHwKeys],
+  );
+
+  const hasDisagg = useMemo(
+    () =>
+      Object.entries(gpuDataByGroupKey).some(
+        ([groupKey, points]) => visibleGroupKeys.has(groupKey) && points.some((p) => p.disagg),
+      ),
+    [visibleGroupKeys, gpuDataByGroupKey],
+  );
+
+  const costCapRows = useMemo<CostCapRow[]>(() => {
+    if (!costCap) return [];
+    const rows: CostCapRow[] = [];
+    for (const [groupKey, points] of Object.entries(gpuDataByGroupKey)) {
+      if (!visibleGroupKeys.has(groupKey)) continue;
+      const hwKey = groupKey.includes('__') ? groupKey.split('__')[0] : groupKey;
+      const precision = groupKey.includes('__') ? groupKey.split('__')[1] : undefined;
+
+      const maxIv = maxInteractivityAtCost(points, costCap, costProvider, costType);
+      if (maxIv === null) {
+        rows.push({
+          resultKey: groupKey,
+          hwKey,
+          precision,
+          maxInteractivity: null,
+          tputAtIv: null,
+          users: null,
+        });
+        continue;
+      }
+
+      const atIv = interpolateForGPU(points, maxIv, 'interactivity_to_throughput', costProvider);
+      const specs = getGpuSpecs(hwKey);
+      const stats =
+        atIv && mw
+          ? computeFleetStats({
+              mw,
+              powerKwPerGpu: specs.power,
+              costPerGpuHour: specs[costProvider],
+              tputPerGpu: getThroughputForType(atIv, costType),
+              outputTputPerGpu: outputTokPerChip(
+                atIv.value,
+                atIv.inputTokenShare,
+                atIv.outputTputValue,
+              ),
+              interactivity: maxIv,
+            })
+          : null;
+
+      rows.push({
+        resultKey: groupKey,
+        hwKey,
+        precision,
+        maxInteractivity: maxIv,
+        tputAtIv: atIv ? getThroughputForType(atIv, costType) : null,
+        users: stats ? stats.concurrentUsers : null,
+      });
+    }
+    // Highest achievable interactivity first; unreachable rows last.
+    return rows.toSorted(
+      (a, b) => (b.maxInteractivity ?? -Infinity) - (a.maxInteractivity ?? -Infinity),
+    );
+  }, [costCap, visibleGroupKeys, gpuDataByGroupKey, costProvider, costType, mw]);
+
+  const costCapColumns = useMemo<DataTableColumn<CostCapRow>[]>(() => {
+    const columns: DataTableColumn<CostCapRow>[] = [
+      {
+        header: t.colGpu,
+        cell: (r) => getLabel(r, hardwareConfig),
+        sortValue: (r) => getLabel(r, hardwareConfig),
+        className: 'font-medium whitespace-nowrap',
+      },
+      {
+        header: t.colMaxInteractivity,
+        align: 'right',
+        cell: (r) =>
+          r.maxInteractivity === null ? (
+            <span className="text-muted-foreground">{t.notReachable}</span>
+          ) : (
+            r.maxInteractivity.toFixed(1)
+          ),
+        sortValue: (r) => r.maxInteractivity ?? -Infinity,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colTputAtIv,
+        align: 'right',
+        cell: (r) => (r.tputAtIv === null ? '—' : r.tputAtIv.toFixed(1)),
+        sortValue: (r) => r.tputAtIv ?? -Infinity,
+        className: 'tabular-nums',
+      },
+    ];
+    if (mw) {
+      columns.push({
+        header: t.colUsers,
+        align: 'right',
+        cell: (r) => (r.users === null ? '—' : formatCompact(r.users)),
+        sortValue: (r) => r.users ?? -Infinity,
+        className: 'tabular-nums',
+      });
+    }
+    return columns;
+  }, [hardwareConfig, t, mw]);
+
+  return (
+    <TooltipProvider delayDuration={0}>
+      <section data-testid="calculator-costcap-section">
+        <Card>
+          <CollapsibleSection
+            title={t.costCapTitle}
+            toggleLabel={t.toggleSection}
+            testId="calculator-costcap-collapse"
+            onToggle={(open) => track('calculator_section_toggled', { section: 'costcap', open })}
+          >
+            <div className="flex flex-col gap-4">
+              <p className="text-muted-foreground text-sm">{t.costCapDescription}</p>
+              <div className="flex flex-wrap gap-4">
+                <div className="flex flex-col space-y-1.5 max-w-48">
+                  <LabelWithTooltip
+                    htmlFor="calc-costcap"
+                    label={t.costCapLabel}
+                    tooltip={t.costCapTooltip}
+                  />
+                  <Input
+                    id="calc-costcap"
+                    data-testid="calc-costcap-input"
+                    type="number"
+                    min={0}
+                    step="any"
+                    placeholder={t.costCapPlaceholder}
+                    value={costCapInput}
+                    onChange={handleCostCapChange}
+                    onBlur={handleCostCapBlur}
+                    className="w-32 h-9"
+                  />
+                </div>
+                <div className="flex flex-col space-y-1.5 max-w-48">
+                  <LabelWithTooltip
+                    htmlFor="calc-costcap-mw"
+                    label={t.mwLabel}
+                    tooltip={t.mwTooltip}
+                  />
+                  <Input
+                    id="calc-costcap-mw"
+                    data-testid="calc-costcap-mw-input"
+                    type="number"
+                    min={0}
+                    step="any"
+                    placeholder={t.mwPlaceholder}
+                    value={mwInput}
+                    onChange={(e) => onMwInputChange(e.target.value)}
+                    onBlur={() => track('calculator_costcap_mw_set', { mw: mwInput })}
+                    className="w-32 h-9"
+                  />
+                </div>
+              </div>
+              {costCap && costCapRows.length > 0 ? (
+                <>
+                  <DataTable
+                    data={costCapRows}
+                    columns={costCapColumns}
+                    testId="calculator-costcap-table"
+                    analyticsPrefix="calculator_costcap_table"
+                  />
+                  {hasDisagg && (
+                    <p className="text-muted-foreground text-xs border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1">
+                      <strong>{t.note}</strong>
+                      {t.disaggFleet}
+                    </p>
+                  )}
+                  <p className="text-xs text-muted-foreground mt-3">
+                    {t.assumptions(getCostProviderLabel(costProvider))}
+                  </p>
+                  <p className="text-muted-foreground mt-1">
+                    <small>
+                      {t.source}
+                      <Link
+                        target="_blank"
+                        className="underline hover:text-foreground"
+                        href="https://semianalysis.com/datacenter-industry-model/"
+                      >
+                        SemiAnalysis Datacenter Industry Model
+                        <ExternalLinkIcon />
+                      </Link>
+                      {' & '}
+                      <Link
+                        target="_blank"
+                        className="underline hover:text-foreground"
+                        href="https://semianalysis.com/ai-cloud-tco-model/"
+                      >
+                        SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model
+                        <ExternalLinkIcon />
+                      </Link>
+                    </small>
+                  </p>
+                </>
+              ) : (
+                <p className="text-sm text-muted-foreground" data-testid="calculator-costcap-empty">
+                  {/* With a valid target set, empty means nothing is legend-visible
+                      (unreachable chips still produce rows). */}
+                  {costCap ? t.costCapNoGpus : t.costCapEmpty}
+                </p>
+              )}
+            </div>
+          </CollapsibleSection>
+        </Card>
+      </section>
+    </TooltipProvider>
+  );
+}
diff --git a/packages/app/src/components/calculator/FleetLifecycle.tsx b/packages/app/src/components/calculator/FleetLifecycle.tsx
new file mode 100644
index 000000000..f42217a49
--- /dev/null
+++ b/packages/app/src/components/calculator/FleetLifecycle.tsx
@@ -0,0 +1,1559 @@
+'use client';
+
+import { BarChart3, Table2 } from 'lucide-react';
+import Link from 'next/link';
+import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+
+import {
+  getModelReleaseDate,
+  TCO_SOURCE_TITLE,
+  TCO_SOURCE_URL,
+} from '@semianalysisai/inferencex-constants';
+
+import type { HardwareConfig } from '@/components/inference/types';
+import {
+  includesJalapenoResult,
+  includesVeraRubinResult,
+  JalapenoOfficialPreviewNotice,
+  VeraRubinOfficialPreviewNotice,
+} from '@/components/official-preview-notice';
+import { Card } from '@/components/ui/card';
+import { type DataTableColumn, DataTable } from '@/components/ui/data-table';
+import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
+import { Input } from '@/components/ui/input';
+import { LabelWithTooltip } from '@/components/ui/label-with-tooltip';
+import { ChartButtons } from '@/components/ui/chart-buttons';
+import { SegmentedToggle, type SegmentedToggleOption } from '@/components/ui/segmented-toggle';
+import { TooltipProvider } from '@/components/ui/tooltip';
+import { track } from '@/lib/analytics';
+import { exportToCsv } from '@/lib/csv-export';
+import { getGpuSpecs, getHardwareConfig } from '@/lib/constants';
+import {
+  getModelLabel,
+  getSequenceLabel,
+  Sequence,
+  type Model,
+  type Percentile,
+} from '@/lib/data-mappings';
+import { DEFAULT_LIFECYCLE_RAMP_MONTHS, readUrlParams, writeUrlParams } from '@/lib/url-state';
+import { useLocale } from '@/lib/use-locale';
+import { getDisplayLabel } from '@/lib/utils';
+
+import { computeFleetStats, formatCompact } from './fleet';
+import FleetLifecycleChart, {
+  type LifecycleChartSeries,
+  type LifecycleMetric,
+} from './FleetLifecycleChart';
+import { mergeProgressionsByChip, type ChipProgression } from './historical-best';
+import {
+  availabilityFromInterrupts,
+  outputTokPerChip,
+  splitTokenStreams,
+  breakEvenPricePerMTok,
+  computeLifecycle,
+  effectiveTokPerSec,
+  MS_PER_MONTH,
+  type LifecycleAssumptions,
+  type LifecycleSeries,
+  type ThroughputStep,
+} from './lifecycle';
+import { getCostProviderLabel, getThroughputForType } from './ThroughputBarChart';
+import type { CalculatorMode, CostProvider, CostType, InterpolatedResult } from './types';
+import { useHistoricalBest } from './useHistoricalBest';
+
+interface FleetLifecycleProps {
+  hardwareConfig: HardwareConfig;
+  costProvider: CostProvider;
+  costType: CostType;
+  /** Current target interactivity (tok/s/user) from the calculator's slider. */
+  targetValue: number;
+  mode: CalculatorMode;
+  /** Legend visibility by base hwKey — shared with the chart above. */
+  visibleHwKeys: Set<string>;
+  selectedModel: Model;
+  selectedSequence: Sequence;
+  selectedPrecisions: string[];
+  selectedPercentile?: Percentile;
+  /** Facility power budget in MW, shared with the fleet planner. */
+  /**
+   * The facility power budget, raw. Owned by the parent because the URL seed and
+   * the fleet-sizing basis are shared, but the *input* lives here: this section
+   * is the only thing that consumes it, and a control in one section feeding a
+   * table in another is what the old Fleet Projection split got wrong.
+   */
+  mwInput: string;
+  onMwInputChange: (raw: string) => void;
+  /** Resolves a series colour from the calculator's theme palette. */
+  colorResolver: (hwKey: string) => string;
+}
+
+type LifecycleView = 'chart' | 'table';
+
+const LIFECYCLE_VIEW_OPTIONS: SegmentedToggleOption<LifecycleView>[] = [
+  {
+    value: 'chart',
+    label: 'Chart',
+    icon: <BarChart3 className="size-3.5" />,
+    testId: 'calculator-lifecycle-chart-view-btn',
+  },
+  {
+    value: 'table',
+    label: 'Table',
+    icon: <Table2 className="size-3.5" />,
+    testId: 'calculator-lifecycle-table-view-btn',
+  },
+];
+
+const STRINGS = {
+  en: {
+    title: 'Fleet Lifecycle',
+    viewChart: 'Chart',
+    viewTable: 'Table',
+    viewAria: 'View mode',
+    captionTarget: (target: number) => `${target.toFixed(0)} tok/s/user`,
+    captionPrices: (input: string, output: string, cachedPct: string | null) =>
+      `$${input} in / $${output} out per M tok${cachedPct === null ? '' : `, cached ${cachedPct}%`}`,
+    description:
+      'A fixed fleet, from the day the model shipped. The chips never change. The software serving them does, so each rollout is a config that beat every config before it, climbing from what the fleet already served to its own numbers. The gap to the cost line is the return on that work.',
+    tooSmall:
+      'This power budget is too small to power a single chip of the shown hardware. Try a larger value.',
+    mwLabel: 'Facility Power (MW)',
+    mwTooltip:
+      'Total facility power budget in megawatts. Chip count uses all-in power per chip (host, networking, cooling) from the SemiAnalysis Datacenter Industry Model, not bare TDP.',
+    mwPlaceholder: 'e.g. 10',
+    colChips: 'Chips',
+    colUsers: 'Concurrent Users now',
+    needMw: 'Enter a facility power budget to project lifecycle economics.',
+    noReleaseDate:
+      'No release date is on file for this model, so the timeline is anchored to its first benchmark run instead.',
+    loading: 'Loading run history…',
+    errorPrefix: 'Could not load run history: ',
+    noneMeasured:
+      'No chip was measured at this interactivity on any run date. Move the target interactivity slider into a measured range. The ranges each chip has been measured over are listed below.',
+    priceLabel: 'Token Price: input, output ($/M tok)',
+    priceInputLabel: 'Input token price, $ per million',
+    outputPriceInputLabel: 'Output token price, $ per million',
+    priceTooltip:
+      "Sale price of input and output tokens, set separately. Providers bill output at a multiple of input, and the two streams are wildly unequal: 8 input tokens per output token on a fixed 8k/1k sequence, around 130 on agentic traces. Revenue counts both, whichever token type the cost matrix above is set to. The pair defaults to the prices that exactly zero the cheapest visible fleet's margin at its latest config. This is the competitive floor, where that chip earns nothing and every pricier chip is underwater. It starts with output priced at 4x input, roughly matching major vendors. Break-even with two prices is a line rather than a point, so a reset scales both and keeps whatever ratio you have set. Everything above that line is your assumption, not a measurement.",
+    priceReset: 'Reset to break-even',
+    mtbiLabel: 'MTBI (days)',
+    mtbiTooltip:
+      'Mean time between interruptions. Combined with recovery time this becomes an availability haircut on revenue. Leave blank to model no interruptions.',
+    recoveryLabel: 'Recovery (hours)',
+    recoveryTooltip: 'Hours to restore service after one interruption.',
+    rampLabel: 'Ramp (months)',
+    rampTooltip:
+      'Months for a config to roll out across the fleet. Every config gets one: it climbs from whatever the fleet was already serving to the new config\u2019s numbers, and the first climbs from zero. Cost runs at full rate throughout because racks bill from the moment they are energised, not from the moment they are loaded. The first rollout therefore opens at a full day\u2019s cost against no revenue, which is the deepest the margin line ever goes. Your assumption, not a measurement; set it to 0 for configs that take effect instantly.',
+    horizonLabel: 'Horizon (months from release)',
+    horizonTooltip:
+      "How far past the model's release date to project. Past the last sweep, the latest config is held flat. That is what the fleet earns if optimisation stops, not a forecast of further gains.",
+    cacheLabel: 'Cached input (% of price)',
+    cacheTooltip:
+      'What a cached input token sells for, as a percentage of the price above. Agentic traces reuse enormous prefixes: a median 133 input tokens per output token, with a median 92% served from cache on these runs. Providers bill a cache read at a fraction of a fresh token, so charging full price for all of them overstates margin by close to an order of magnitude. The cached fraction is measured per config, not assumed; only the percentage is yours. Set it to 100 to bill every token the same. Fixed sequences record no cache hits at all, which is why this input only appears for agentic traces.',
+    singleRung:
+      'Every chip here has been measured on a single run date, so each line is one config held flat. The staircase this section exists to show needs several dates per config, and agentic trace history does not go back far enough yet. The levels are measured; the absence of steps is a gap in the history, not a finding about the hardware.',
+    colChip: 'Chip',
+    colConfigNow: 'Config Now',
+    colFirst: 'First Run',
+    colLatest: 'Latest Best',
+    colSteps: 'Improvements',
+    colGain: 'Gain',
+    colTpPerMw: (tokenType: string) => `${tokenType}tok/s/MW now`,
+    colRevenue: 'Revenue $/day',
+    colCost: 'Cost $/day',
+    colMargin: 'Margin $/day',
+    colPayback: 'Payback',
+    colLifetime: 'Cumulative Margin',
+    colAvailability: 'Availability',
+    never: 'Never',
+    monthsSuffix: 'mo',
+    unmeasuredTitle: 'Not measured at this interactivity',
+    unmeasuredIntro:
+      'These chips have run history for this scenario but no single run date was measured across the target interactivity, so no honest number exists for them. The nearest interactivity each one actually reached:',
+    unmeasuredRange: (below: number | null, above: number | null, dates: number) => {
+      const dateLabel = `${dates} run ${dates === 1 ? 'date' : 'dates'}`;
+      if (below === null && above === null) return `nothing measured (${dateLabel})`;
+      if (below === null)
+        return `nothing below; nearest above ${above!.toFixed(1)} tok/s/user (${dateLabel})`;
+      if (above === null)
+        return `nearest below ${below.toFixed(1)} tok/s/user; nothing above (${dateLabel})`;
+      return `nearest ${below.toFixed(1)} below and ${above.toFixed(1)} above, tok/s/user (${dateLabel})`;
+    },
+    unplottable: (chips: string) =>
+      `No fleet could be sized for ${chips} at this power budget. The chip has no registered power figure, or its measured throughput sizes to nothing. Listed rather than dropped so the chart is never quietly missing a chip.`,
+    note: 'Note:',
+    disagg:
+      ' A step can be won by a disaggregated configuration, and the handover is a like-for-like comparison: total tok/s/chip is reported per chip overall for both kinds, and prefill and decode chips are the same silicon, so neither the sizing nor the cost line moves at the switch. Where this section needs an input- or output-token figure, to rank a step or to price the two streams apart, it derives one from that total and the run’s measured token mix rather than reading the per-chip input and output rates directly. Those two are reported per prefill and per decode chip on a disaggregated run, which is why the throughput charts elsewhere caveat them as not apples-to-apples; deriving instead keeps every comparison here on one denominator. What the fleet does change is its shape. The winning config’s prefill:decode ratio is assumed to apply across the whole fleet from the moment that config rolls out. In practice, this is a redeployment that the ramp window stands in for, not a measurement. The config named on each step says which kind won it.',
+    hybrid:
+      " One line per chip, not per software config: at any moment it follows whichever framework, precision and speculative-decoding combination was ahead, so the config serving the fleet changes along the line and each step names the one that took over. Legend entries still filter configs, which removes them from candidacy. Each step is a measured run date whose interpolated throughput at the target beat every earlier date; a sweep that failed to beat the incumbent is not a step, because the fleet kept serving the config it already had. A config does not take effect the instant a sweep finds it, so each one rolls out over the ramp window, climbing from what the fleet already served to its own numbers. Power and $/chip/hr are today's values from the TCO model, and cost is flat throughout because no config moves either term. It is the same silicon either way. Reads outside a run's measured interactivity range are excluded rather than clamped.",
+    overlayExempt:
+      ' Unofficial runs loaded via a run link are not shown here because the run-history API serves ingested official results only.',
+    metricLabel: 'Y Axis',
+    metricTooltip:
+      'What to plot. Margin is the per-day revenue minus the flat fleet cost, so the break-even rule shows which side of it a fleet is on. Margin/MW and Revenue/MW divide their daily rates by provisioned power. Revenue drops the cost term, which makes rollouts easier to compare across chips of very different cost, but a chip being higher no longer means it is more profitable. Cumulative Revenue is the running total taken in since launch, not a rate; it compounds that caveat, because the largest area under a revenue curve may still never have covered its cost.',
+    metricMargin: 'Margin',
+    metricMarginPerMw: 'Margin/MW',
+    metricRevenue: 'Revenue',
+    metricRevenuePerMw: 'Revenue/MW',
+    metricCumulativeRevenue: 'Cum. Revenue',
+    chartY: 'Margin ($/day)',
+    chartYMarginPerMw: 'Margin ($/MW/day)',
+    chartYRevenue: 'Revenue ($/day)',
+    chartYRevenuePerMw: 'Revenue ($/MW/day)',
+    chartYCumulativeRevenue: 'Cumulative Revenue ($)',
+    chartBreakEven: 'break-even',
+    tipDate: 'Measured',
+    tipConfig: 'Config',
+    tipMargin: 'Margin/day',
+    tipMarginPerMw: 'Margin/MW/day',
+    tipRevenue: 'Revenue/day',
+    tipRevenuePerMw: 'Revenue/MW/day',
+    tipCumulativeRevenue: 'Cumulative revenue',
+    tipCost: 'Cost/day',
+    tipCumulative: 'Cumulative margin',
+    tipSinceFirst: 'Since first run',
+    tipRunLink: 'Open run',
+    chartInstructions:
+      'Hover to read every chip at that date · Click to freeze the readout, click again to release · Shift+Scroll to zoom horizontally · Drag to pan · Double-click to reset',
+    assumptions: (tier: string, chips: string, release: string) =>
+      `Anchored at the ${release} release. Fleet sized by facility power at ${chips}; cost = chips × ${tier} $/chip/hr, flat for the whole window. Revenue is priced on the selected token type and reduced by the availability haircut. Price, ramp, MTBI, recovery and horizon are your assumptions. The throughput steps are not.`,
+    source: 'Source: ',
+  },
+  zh: {
+    title: '集群生命周期',
+    viewChart: '图表',
+    viewTable: '表格',
+    viewAria: '显示模式',
+    captionTarget: (target: number) => `${target.toFixed(0)} tok/s/用户`,
+    captionPrices: (input: string, output: string, cachedPct: string | null) =>
+      `输入 $${input} / 输出 $${output} 每百万 token${cachedPct === null ? '' : `，缓存 ${cachedPct}%`}`,
+    description:
+      '固定集群自模型发布之日起的表现。芯片始终不变，变化的是为其提供服务的软件。每次推广采用的配置都优于此前所有配置，并从集群当前服务水平逐步爬升至新配置水平。曲线与成本线之间的差距，就是这些优化工作的回报。',
+    tooSmall: '该功率预算不足以为所示任一芯片供电。请尝试更大的数值。',
+    mwLabel: '设施功率 (MW)',
+    mwTooltip:
+      '设施总功率预算（兆瓦）。芯片数量按每芯片全含功率（主机、网络、散热）计算，数据来自 SemiAnalysis Datacenter Industry Model，而非裸 TDP。',
+    mwPlaceholder: '如 10',
+    colChips: '芯片数',
+    colUsers: '当前并发用户数',
+    needMw: '输入设施功率预算，以测算生命周期经济性。',
+    noReleaseDate: '该模型暂无发布日期记录，时间轴改以其首次基准测试运行为起点。',
+    loading: '正在加载运行历史……',
+    errorPrefix: '无法加载运行历史：',
+    noneMeasured:
+      '在任何运行日期下都没有芯片在该交互性下被实测过。请将目标交互性滑块移入已实测区间。各芯片的实测区间见下方列表。',
+    priceLabel: 'Token 价格：输入、输出 ($/M tok)',
+    priceInputLabel: '输入 token 价格，$/百万',
+    outputPriceInputLabel: '输出 token 价格，$/百万',
+    priceTooltip:
+      'Token 售价，输入与输出分别设定。服务商对输出 token 的计价通常是输入的数倍，而两类 token 的数量极不均衡：固定 8k/1k 序列下每个输出 token 对应 8 个输入 token，agentic traces 下约为 130 个。无论上方成本矩阵选择哪种 token 类型，收入均同时计入两者。二者默认取使当前可见集群中成本最低者在其最新配置下利润恰好为零的价格，即竞争底线。该价格下这款芯片不赚不亏，而所有更贵的芯片均为亏损。初始比例为输出价格是输入价格的 4 倍，与主流厂商的定价大致相当。在双价格下保本点是一条直线而非一个点，因此重置会按你当前设定的比例同时缩放两个价格。高于该线的部分属于你的假设，而非实测值。',
+    priceReset: '重置为保本价',
+    mtbiLabel: '平均无故障间隔 (天)',
+    mtbiTooltip: '平均中断间隔时间。与恢复时间共同构成对收入的可用性折损。留空表示不建模中断。',
+    recoveryLabel: '恢复时间 (小时)',
+    recoveryTooltip: '一次中断后恢复服务所需的小时数。',
+    rampLabel: '爬坡期 (月)',
+    rampTooltip:
+      '一个配置在集群中完成推广所需的月数。每个配置都有各自的推广曲线：从集群当前已提供的水平爬升至新配置的水平，首个配置从零开始爬升。成本在整个期间按满额计入，因为机架自通电起即开始计费，而非自满载起。因此首次推广开始时即需承担一整天的成本而收入为零，这正是利润率曲线的全程最低点。这是你的假设而非实测值；设为 0 表示配置立即生效。',
+    horizonLabel: '测算期 (自发布起月数)',
+    horizonTooltip:
+      '自模型发布日期起向后测算的月数。在最后一次扫描之后，最新配置将保持不变。这代表若优化停止时集群的收益，而非对后续提升的预测。',
+    cacheLabel: '缓存输入 (占价格百分比)',
+    cacheTooltip:
+      '一个缓存输入 token 的售价，以上方价格的百分比表示。Agentic traces 会复用极长的前缀：中位数为每个输出 token 对应 133 个输入 token，其中中位数 92% 在本批运行中由缓存提供。服务商对缓存读取仅按新鲜 token 价格的一小部分计费，因此若对全部 token 按满价计费，将使利润被高估近一个数量级。缓存占比按各配置实测得出，并非假设；仅该百分比属于你的假设。设为 100 表示所有 token 同价。固定序列完全没有缓存命中记录，因此该输入项仅在 Agentic Traces 下出现。',
+    singleRung:
+      '此处每款芯片都只有单一运行日期的实测数据，因此每条线都是一个配置的水平延伸。本模块所要呈现的阶梯需要同一配置在多个日期上的数据，而 agentic traces 的历史尚不够长。图中的水平值是实测的；缺少台阶反映的是历史数据的空缺，而非关于硬件的结论。',
+    colChip: '芯片',
+    colConfigNow: '当前配置',
+    colFirst: '首次运行',
+    colLatest: '最新最佳',
+    colSteps: '提升次数',
+    colGain: '提升倍数',
+    colTpPerMw: (tokenType: string) => `当前${tokenType} tok/s/MW`,
+    colRevenue: '收入 $/天',
+    colCost: '成本 $/天',
+    colMargin: '利润 $/天',
+    colPayback: '回本时间',
+    colLifetime: '累计利润',
+    colAvailability: '可用性',
+    never: '无法回本',
+    monthsSuffix: '个月',
+    unmeasuredTitle: '该交互性下无实测数据',
+    unmeasuredIntro:
+      '以下芯片在该场景下有运行历史，但没有任何单次运行的实测范围覆盖目标交互性，因此无法给出可靠数值。各自最接近的实测交互性：',
+    unmeasuredRange: (below: number | null, above: number | null, dates: number) => {
+      const dateLabel = `共 ${dates} 个运行日期`;
+      if (below === null && above === null) return `无实测数据（${dateLabel}）`;
+      if (below === null)
+        return `下方无实测；上方最近 ${above!.toFixed(1)} tok/s/user（${dateLabel}）`;
+      if (above === null)
+        return `下方最近 ${below.toFixed(1)} tok/s/user；上方无实测（${dateLabel}）`;
+      return `最近实测为下方 ${below.toFixed(1)}、上方 ${above.toFixed(1)} tok/s/user（${dateLabel}）`;
+    },
+    unplottable: (chips: string) =>
+      `在该功率预算下无法为 ${chips} 组建集群。该芯片缺少已登记的功耗数据，或其实测吞吐无法组成任何规模。此处列出而非直接剔除，以免图表静默遗漏芯片。`,
+    note: '注意：',
+    disagg:
+      '台阶可以由解耦配置取得，且该次交接是同类比较：两种部署方式的总 tok/s/chip 均按芯片总数报告，且预填充与解码使用的是同一种硅片，因此切换时集群规模与成本线均不发生变化。当本模块需要输入或输出 token 的单独数值时（用于台阶排序，或将两类 token 分别计价），它由该总量与该次运行实测的 token 构成比推导得出，而非直接读取每芯片的输入/输出速率。在解耦部署下，这两个速率分别按预填充芯片与解码芯片计算，这也正是别处吞吐图表标注其不可直接比较的原因；改为推导可使本模块的所有比较统一在同一分母上。真正改变的是集群形态。自该配置开始推广起，其预填充:解码比例被假定应用于整个集群。实际部署中，这相当于一次重新部署，由爬坡期窗口代为体现，并非实测值。每一级台阶标注的配置即说明其类型。',
+    hybrid:
+      '每款芯片一条曲线，而非每个软件配置一条：曲线在任一时刻都跟随当时领先的框架、精度与投机解码组合，因此服务集群的配置会沿曲线变化，每一级台阶都标注接管的配置。图例项仍可筛选配置，被隐藏的配置将不参与竞争。每一级台阶都是一个实测运行日期，其在目标交互性下的插值吞吐量优于此前所有日期；未能超越现有配置的扫描不构成台阶，因为集群仍在运行原有配置。配置不会在扫描发现的瞬间生效，因此每个配置都会在推广期内从集群当前已提供的水平爬升至其自身水平。功率与 $/chip/hr 为 TCO 模型的当前值，成本在整个期间保持水平，因为任何配置都不会改变这两项。两种情况下都是同一款芯片。超出某次运行实测交互性区间的结果会被排除而非钳制。',
+    overlayExempt:
+      '通过运行链接加载的非官方运行不会显示在此，因为运行历史 API 仅提供已入库的官方结果。',
+    metricLabel: 'Y 轴',
+    metricTooltip:
+      '选择绘制的内容。利润为日均收入减去固定的集群成本，因此保本线可显示集群位于其哪一侧。每 MW 利润和每 MW 收入分别将对应的日均指标除以已配置功率。收入不扣除成本，便于在成本差异很大的芯片之间比较推广曲线，但此时位置更高并不代表更赚钱。累计收入是自上线以来的累计总额，而非日均指标；它会进一步放大上述注意事项，因为收入曲线下面积最大的芯片仍可能从未收回成本。',
+    metricMargin: '利润',
+    metricMarginPerMw: '每 MW 利润',
+    metricRevenue: '收入',
+    metricRevenuePerMw: '每 MW 收入',
+    metricCumulativeRevenue: '累计收入',
+    chartY: '利润 ($/天)',
+    chartYMarginPerMw: '每 MW 利润 ($/MW/天)',
+    chartYRevenue: '收入 ($/天)',
+    chartYRevenuePerMw: '每 MW 收入 ($/MW/天)',
+    chartYCumulativeRevenue: '累计收入 ($)',
+    chartBreakEven: '保本线',
+    tipDate: '实测于',
+    tipConfig: '配置',
+    tipMargin: '每日利润',
+    tipMarginPerMw: '每 MW 每日利润',
+    tipRevenue: '每日收入',
+    tipRevenuePerMw: '每 MW 每日收入',
+    tipCumulativeRevenue: '累计收入',
+    tipCost: '每日成本',
+    tipCumulative: '累计利润',
+    tipSinceFirst: '相比首次运行',
+    tipRunLink: '查看运行',
+    chartInstructions:
+      '悬停可读取该日期下所有芯片的数值 · 点击可冻结读数，再次点击解除 · Shift+滚轮 横向缩放 · 拖动平移 · 双击重置',
+    assumptions: (tier: string, chips: string, release: string) =>
+      `以 ${release} 发布日期为起点。集群规模按 ${chips} 的设施功率测算；成本 = 芯片数 × ${tier} $/chip/hr，在整个测算期内保持不变。收入按所选 token 类型计价，并扣除可用性折损。价格、爬坡期、平均无故障间隔、恢复时间与测算期为你的假设。吞吐量台阶不是假设。`,
+    source: '来源：',
+  },
+} as const;
+
+const DEFAULTS = {
+  mtbiDays: 24,
+  recoveryHours: 12,
+  // A nominal half-month to bring a fleet to full load. Purely an assumption,
+  // and labelled as one — no measurement in this repo speaks to it.
+  rampMonths: Number(DEFAULT_LIFECYCLE_RAMP_MONTHS),
+  /**
+   * A cached input token sells for a tenth of a fresh one — the ratio DeepSeek
+   * and Anthropic both publish, and the order of magnitude the others sit at.
+   * An assumption like the rest; the cached *fraction* it applies to is measured.
+   */
+  cachedInputPct: 10,
+  /**
+   * An output token sells for four times an input one until the user says
+   * otherwise — DeepSeek's own published API pricing is $0.27 / $1.10, and the
+   * major vendors sit between 2x and 5x. Only used to seed the pair and to hold
+   * their ratio through a reset; once both fields exist they are what is billed.
+   */
+  outputPriceMultiple: 4,
+};
+
+/**
+ * A price for a field, in as few digits as say it.
+ *
+ * `toFixed(4)` alone is wrong at the bottom of the range: a fleet cheap enough,
+ * or a ratio small enough, produces a real price below 0.00005 and rounds it to
+ * "0.0000" — a field that reads as "these tokens are free" when they are not.
+ * Small values fall back to significant digits instead.
+ */
+function formatPrice(value: number): string {
+  if (!Number.isFinite(value) || value <= 0) return '0';
+  return value < 0.0001 ? value.toPrecision(3) : value.toFixed(4);
+}
+
+/**
+ * Same price, for reading rather than editing.
+ *
+ * The input fields are padded to four decimals so a seeded break-even figure
+ * lines up as you type into it; a caption is prose, and `$12.0000` there reads
+ * as spurious precision. Trailing zeros are dropped — `12`, `12.5`, `0.1234` —
+ * which keeps sub-cent prices intact, since a break-even figure on a large fleet
+ * routinely lands below a cent and rounding it to one would print `$0`.
+ */
+function formatCaptionPrice(value: number): string {
+  const fixed = formatPrice(value);
+  return fixed.includes('.') ? fixed.replace(/\.?0+$/u, '') : fixed;
+}
+
+function parseNonNegative(raw: string): number | null {
+  const parsed = parseFloat(raw);
+  return Number.isFinite(parsed) && parsed >= 0 ? parsed : null;
+}
+
+/**
+ * Resolve the two-field price seed from a share link.
+ *
+ * `c_price` predates the input/output split, so an old link supplies only the
+ * input side. Preserve that explicit value and derive its missing peer at the
+ * default ratio; otherwise marking the pair as URL-owned leaves output blank and
+ * silently bills every output token at $0. The symmetric case makes a manually
+ * authored output-only link complete for the same reason.
+ */
+function initialPriceInputs(params: ReturnType<typeof readUrlParams>): {
+  input: string;
+  output: string;
+  edited: boolean;
+} {
+  let input = params.c_price ?? '';
+  let output = params.c_oprice ?? '';
+  const parsedInput = parseNonNegative(input);
+  const parsedOutput = parseNonNegative(output);
+
+  if (input && !output && parsedInput !== null) {
+    output = formatPrice(parsedInput * DEFAULTS.outputPriceMultiple);
+  } else if (!input && output && parsedOutput !== null) {
+    input = formatPrice(parsedOutput / DEFAULTS.outputPriceMultiple);
+  }
+
+  return { input, output, edited: Boolean(input || output) };
+}
+
+function getLabel(hwKey: string, hardwareConfig: HardwareConfig): string {
+  const config = hardwareConfig[hwKey] || getHardwareConfig(hwKey);
+  return config ? getDisplayLabel(config) : hwKey;
+}
+
+/**
+ * The software half of a hwKey's label — what distinguishes it from bare silicon.
+ * Display labels put it in parentheses ("B200 (SGLang, MTP)"); when they don't,
+ * fall back to the hwKey's own suffix so the cell is never blank for a config
+ * that genuinely differs.
+ */
+function configSuffix(hwKey: string, baseGpu: string, hardwareConfig: HardwareConfig): string {
+  if (hwKey === baseGpu) return '—';
+  const label = getLabel(hwKey, hardwareConfig);
+  const parenthetical = /\((?<inner>[^)]*)\)/u.exec(label);
+  if (parenthetical?.groups?.inner) return parenthetical.groups.inner;
+  const suffix = hwKey.slice(baseGpu.length).replace(/^_/u, '').replaceAll('_', ', ');
+  return suffix || '—';
+}
+
+/**
+ * Signed money. Carries a trillions step: a cumulative margin at a
+ * hyperscaler-sized power budget genuinely reaches it, and `$1532.50B` is
+ * harder to read than `$1.53T`.
+ */
+function formatMoney(value: number): string {
+  const abs = Math.abs(value);
+  const sign = value < 0 ? '-' : '';
+  if (abs >= 1e12) return `${sign}$${(abs / 1e12).toFixed(2)}T`;
+  if (abs >= 1e9) return `${sign}$${(abs / 1e9).toFixed(2)}B`;
+  if (abs >= 1e6) return `${sign}$${(abs / 1e6).toFixed(2)}M`;
+  if (abs >= 1e3) return `${sign}$${(abs / 1e3).toFixed(1)}k`;
+  return `${sign}$${abs.toFixed(0)}`;
+}
+
+/** Short label for the config behind a step, for the tooltip. */
+function configLabel(result: InterpolatedResult): string {
+  const point = result.nearestPoints[0];
+  if (!point) return '';
+  const parts = [point.precision?.toUpperCase()].filter(Boolean) as string[];
+  if (point.disagg) parts.push('disagg');
+  if (result.concurrency > 0) parts.push(`conc ${result.concurrency}`);
+  return parts.join(' · ');
+}
+
+/** A run date, linked to its workflow run when one is recorded. */
+function runLink(date: string, runUrls: string[]) {
+  if (runUrls.length === 0) return <>{date}</>;
+  return (
+    <Link href={runUrls[0]!} target="_blank" className="underline hover:text-foreground">
+      {date}
+      <ExternalLinkIcon />
+    </Link>
+  );
+}
+
+interface LifecycleRow {
+  progression: ChipProgression;
+  label: string;
+  /** The config the chip is running at the latest rung, e.g. "SGLang, MTP". */
+  configNow: string;
+  /** A real hwKey to resolve the series colour from. */
+  colorKey: string;
+  tpPerMw: number;
+  /** Chips the power budget provisions — flat across the whole lifecycle. */
+  gpus: number;
+  /** Streams the fleet serves at the target interactivity on its latest config. */
+  concurrentUsersNow: number;
+  disagg: boolean;
+  series: LifecycleSeries;
+}
+
+export default function FleetLifecycle({
+  hardwareConfig,
+  costProvider,
+  costType,
+  targetValue,
+  mode,
+  visibleHwKeys,
+  selectedModel,
+  selectedSequence,
+  selectedPrecisions,
+  selectedPercentile,
+  mwInput,
+  onMwInputChange,
+  colorResolver,
+}: FleetLifecycleProps) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  /** A zero or blank budget sizes no fleet, so it is treated as unset. */
+  const mw = useMemo(() => {
+    const parsed = parseNonNegative(mwInput);
+    return parsed !== null && parsed > 0 ? parsed : null;
+  }, [mwInput]);
+
+  const historical = useHistoricalBest({
+    model: selectedModel,
+    sequence: selectedSequence,
+    precisions: selectedPrecisions,
+    targetValue,
+    mode,
+    costProvider,
+    costType,
+    percentile: selectedPercentile,
+    // The history payload is several MB — do not fetch it until the section can
+    // actually produce something.
+    enabled: Boolean(mw),
+  });
+
+  const [mtbiInput, setMtbiInput] = useState(
+    () => readUrlParams().c_mtbi ?? String(DEFAULTS.mtbiDays),
+  );
+  const [recoveryInput, setRecoveryInput] = useState(
+    () => readUrlParams().c_rec ?? String(DEFAULTS.recoveryHours),
+  );
+  const [rampInput, setRampInput] = useState(
+    () => readUrlParams().c_ramp ?? String(DEFAULTS.rampMonths),
+  );
+  const [cacheInput, setCacheInput] = useState(
+    () => readUrlParams().c_cache ?? String(DEFAULTS.cachedInputPct),
+  );
+  /**
+   * Only agentic runs record a cache hit rate, so only they can be discounted.
+   * The control is hidden elsewhere rather than shown as a no-op — a knob that
+   * moves nothing is worse than no knob.
+   */
+  const isAgentic = selectedSequence === Sequence.AgenticTraces;
+  const cacheReadRatio = isAgentic
+    ? Math.min(1, (parseNonNegative(cacheInput) ?? DEFAULTS.cachedInputPct) / 100)
+    : 1;
+
+  const [horizonInput, setHorizonInput] = useState(() => readUrlParams().c_life ?? '');
+  // Like the price, the horizon is seeded from the data until the user takes it
+  // over: a fixed 60-month default spends most of the axis on a flat tail
+  // projecting the last config forward, which carries no information.
+  const horizonEdited = useRef(Boolean(readUrlParams().c_life));
+  const [yMetric, setYMetric] = useState<LifecycleMetric>(() => {
+    const seeded = readUrlParams().c_ly;
+    // Allowlisted rather than cast: the param is user-editable, and an unknown
+    // value must fall back to the default instead of reaching `metricValue`.
+    return seeded === 'revenue' ||
+      seeded === 'revenuePerMw' ||
+      seeded === 'cumulativeRevenue' ||
+      seeded === 'marginPerMw'
+      ? seeded
+      : 'margin';
+  });
+  const [initialPrices] = useState(() => initialPriceInputs(readUrlParams()));
+  const [priceInput, setPriceInput] = useState(initialPrices.input);
+  const [outputPriceInput, setOutputPriceInput] = useState(initialPrices.output);
+  // A price arriving from the URL is the user's, so it must not be overwritten
+  // by the break-even default. Either field counts: they are seeded as a pair,
+  // so re-seeding one after the user set the other would silently move the ratio
+  // they chose.
+  const priceEdited = useRef(initialPrices.edited);
+  /**
+   * Ratio the pair is seeded at, and restored to by a reset.
+   *
+   * State rather than derived from the two fields: deriving it would make the
+   * seed depend on the values the seed just wrote, and the re-seed that follows a
+   * reset would read the ratio it had itself produced. Only a user edit moves it,
+   * so a reader who set 3x gets 3x back rather than the default.
+   */
+  const [seedMultiple, setSeedMultiple] = useState(() => {
+    const input = parseNonNegative(initialPrices.input);
+    const output = parseNonNegative(initialPrices.output);
+    if (input !== null && input > 0 && output !== null && output > 0) return output / input;
+    return DEFAULTS.outputPriceMultiple;
+  });
+
+  /**
+   * Anchor for the timeline. The model's release date is the honest zero — the
+   * question is how far the software has come since the weights existed. Models
+   * without a sourced release date fall back to their first measured run.
+   */
+  const releaseDate = getModelReleaseDate(selectedModel);
+  /**
+   * One series per chip, not per hwKey. The legend still filters hwKeys, so
+   * hiding a framework removes it from candidacy for its chip's envelope — the
+   * legend stays the control it always was, one level below the lines.
+   */
+  const visibleHistoricalProgressions = useMemo(
+    () => historical.progressions.filter((progression) => visibleHwKeys.has(progression.hwKey)),
+    [historical.progressions, visibleHwKeys],
+  );
+  const visibleProgressions = useMemo(
+    () => mergeProgressionsByChip(visibleHistoricalProgressions),
+    [visibleHistoricalProgressions],
+  );
+  const showsJalapenoPreview = includesJalapenoResult(
+    visibleHistoricalProgressions.map((progression) => progression.hwKey),
+  );
+  const showsVeraRubinPreview = includesVeraRubinResult(
+    visibleHistoricalProgressions.map((progression) => progression.hwKey),
+  );
+
+  const anchorDate = useMemo(() => {
+    if (releaseDate) return releaseDate;
+    let earliest: string | null = null;
+    for (const p of visibleProgressions) {
+      const first = p.steps[0]?.date;
+      if (first && (earliest === null || first < earliest)) earliest = first;
+    }
+    return earliest;
+  }, [releaseDate, visibleProgressions]);
+
+  const anchorMs = useMemo(
+    () => (anchorDate ? Date.parse(`${anchorDate}T00:00:00Z`) : Number.NaN),
+    [anchorDate],
+  );
+
+  const visibleUnmeasured = useMemo(
+    () => historical.unmeasured.filter((entry) => visibleHwKeys.has(entry.hwKey)),
+    [historical.unmeasured, visibleHwKeys],
+  );
+
+  /** Months from the anchor to the last measured sweep, across visible chips. */
+  const measuredMonths = useMemo(() => {
+    if (!Number.isFinite(anchorMs)) return null;
+    let latest = -Infinity;
+    for (const p of visibleProgressions) {
+      const last = p.steps.at(-1)?.date;
+      if (!last) continue;
+      latest = Math.max(latest, (Date.parse(`${last}T00:00:00Z`) - anchorMs) / MS_PER_MONTH);
+    }
+    return Number.isFinite(latest) ? latest : null;
+  }, [anchorMs, visibleProgressions]);
+
+  useEffect(() => {
+    if (horizonEdited.current || measuredMonths === null) return;
+    // A short tail past the last sweep so the final step is readable rather than
+    // pinned to the right edge.
+    setHorizonInput(String(Math.max(1, Math.ceil(measuredMonths + 2))));
+  }, [measuredMonths]);
+
+  const horizonMonths = parseNonNegative(horizonInput) ?? 0;
+
+  /**
+   * Per chip: the step schedule (one rung per measured improvement) plus the flat
+   * fleet cost. Chip count and $/chip/hr do not move when a config improves, so
+   * cost is computed once from the opening rung.
+   */
+  const { fleets, unplottable } = useMemo(() => {
+    if (!mw || !Number.isFinite(anchorMs)) return { fleets: [], unplottable: [] };
+    // Chips that survived selection but cannot be turned into a fleet — no
+    // registered power figure for the base GPU, or a throughput that sizes to
+    // nothing. They are named below rather than dropped: a chip that is in the
+    // legend and in no row, with no explanation, reads as a bug in the data.
+    const absent: string[] = [];
+    const sized = visibleProgressions.flatMap((progression) => {
+      // Power and $/chip/hr come from the base GPU, so they are identical across
+      // the hwKeys pooled into this line — which is what keeps cost flat even
+      // though the winning config changes.
+      const specs = getGpuSpecs(progression.baseGpu);
+      const steps: ThroughputStep[] = [];
+      let costPerHour: number | null = null;
+      // Chip count is mw / all-in power, so it is the same at every rung. Users
+      // is not: the fleet streams more of them as throughput improves, so it is
+      // the *latest* rung's figure, matching the `tok/s/MW now` column.
+      let gpus: number | null = null;
+      let concurrentUsersNow: number | null = null;
+      // Power the fleet actually occupies, not the budget: chip counts are whole,
+      // so the last fraction of a chip's worth of the budget is never provisioned.
+      let provisionedMw: number | null = null;
+
+      for (const step of progression.steps) {
+        const stats = computeFleetStats({
+          mw,
+          powerKwPerGpu: specs.power,
+          costPerGpuHour: specs[costProvider],
+          tputPerGpu: getThroughputForType(step.result, costType),
+          // Through the accessor even though this one is always the output rate:
+          // the cost-matrix rule exists so every throughput read goes through one
+          // chokepoint, and a direct field read silently diverges if it gains logic.
+          outputTputPerGpu: outputTokPerChip(
+            getThroughputForType(step.result, 'total'),
+            step.result.inputTokenShare,
+            getThroughputForType(step.result, 'output'),
+          ),
+          interactivity: targetValue,
+        });
+        if (!stats) continue;
+        costPerHour ??= stats.costPerHour;
+        provisionedMw ??= (stats.gpus * specs.power) / 1000;
+        gpus ??= stats.gpus;
+        concurrentUsersNow = stats.concurrentUsers;
+        steps.push({
+          month: (Date.parse(`${step.date}T00:00:00Z`) - anchorMs) / MS_PER_MONTH,
+          // Always the total-token rate, never the cost-type one: the fleet sells
+          // everything it produces, whichever token type the cost matrix above is
+          // expressed in. Split by the measured mix so both streams stay on the
+          // per-chip denominator the fleet was sized and costed on.
+          ...splitTokenStreams(
+            stats.gpus * getThroughputForType(step.result, 'total'),
+            step.result.inputTokenShare,
+            step.result.cacheHitRate,
+            cacheReadRatio,
+          ),
+        });
+      }
+
+      if (
+        steps.length === 0 ||
+        costPerHour === null ||
+        provisionedMw === null ||
+        gpus === null ||
+        concurrentUsersNow === null
+      ) {
+        absent.push(progression.baseGpu);
+        return [];
+      }
+      return [{ progression, steps, costPerHour, provisionedMw, gpus, concurrentUsersNow }];
+    });
+    return { fleets: sized, unplottable: absent };
+  }, [mw, anchorMs, visibleProgressions, costProvider, costType, targetValue, cacheReadRatio]);
+
+  // Interrupts sell fewer tokens off the same racks, so they raise break-even.
+  // The seeded price has to carry the same haircut the plotted margin does, or
+  // the default lands slightly below break-even instead of on it.
+  const availability = useMemo(
+    () =>
+      availabilityFromInterrupts(
+        parseNonNegative(mtbiInput) ?? 0,
+        parseNonNegative(recoveryInput) ?? 0,
+      ),
+    [mtbiInput, recoveryInput],
+  );
+
+  /**
+   * Break-even of the cheapest visible fleet at its latest config — the
+   * competitive floor as it stands today, not as it stood at release.
+   */
+
+  const breakEven = useMemo(() => {
+    let cheapest: number | null = null;
+    for (const { steps, costPerHour } of fleets) {
+      const latest = steps.at(-1);
+      if (!latest) continue;
+      // Two prices make break-even a line, not a point. Fixing the ratio picks
+      // one solution off it: solve the input price against a rate that already
+      // counts each output token as `multiple` input tokens' worth of revenue.
+      const price = breakEvenPricePerMTok(
+        costPerHour,
+        effectiveTokPerSec(latest.billableInputTokPerSec, latest.outputTokPerSec, seedMultiple),
+        availability,
+      );
+      if (price === null) continue;
+      if (cheapest === null || price < cheapest) cheapest = price;
+    }
+    return cheapest;
+  }, [fleets, availability, seedMultiple]);
+
+  // Seed and re-seed both prices from break-even until the user takes them over.
+  useEffect(() => {
+    if (priceEdited.current || breakEven === null) return;
+    setPriceInput(formatPrice(breakEven));
+    setOutputPriceInput(formatPrice(breakEven * seedMultiple));
+  }, [breakEven, seedMultiple]);
+
+  const assumptions = useMemo<LifecycleAssumptions>(
+    () => ({
+      mtbiDays: parseNonNegative(mtbiInput) ?? 0,
+      recoveryHours: parseNonNegative(recoveryInput) ?? 0,
+      inputPricePerMTok: parseNonNegative(priceInput) ?? 0,
+      outputPricePerMTok: parseNonNegative(outputPriceInput) ?? 0,
+      rampMonths: parseNonNegative(rampInput) ?? 0,
+    }),
+    [mtbiInput, recoveryInput, priceInput, outputPriceInput, rampInput],
+  );
+
+  const rows = useMemo<LifecycleRow[]>(
+    () =>
+      fleets.flatMap(
+        ({ progression, steps, costPerHour, provisionedMw, gpus, concurrentUsersNow }) => {
+          const series = computeLifecycle({
+            steps,
+            costPerHour,
+            provisionedMw,
+            horizonMonths,
+            assumptions,
+          });
+          if (!series) return [];
+          const latest = progression.steps.at(-1)!;
+          const latestHwKey = latest.result.hwKey ?? progression.baseGpu;
+          return [
+            {
+              progression,
+              // The chip, not the config: the config is what changes along the line.
+              label: getLabel(progression.baseGpu, hardwareConfig),
+              configNow: configSuffix(latestHwKey, progression.baseGpu, hardwareConfig),
+              // The palette is built over the *active* hwKeys, so a bare base key
+              // resolves to the fallback grey. Colour the line by the config it is
+              // running now, which is both a real key and the legend entry a reader
+              // would look for.
+              colorKey: latestHwKey,
+              tpPerMw: latest.rankValue,
+              gpus,
+              concurrentUsersNow,
+              disagg: progression.disagg,
+              series,
+            },
+          ];
+        },
+      ),
+    [fleets, horizonMonths, assumptions, hardwareConfig],
+  );
+
+  const hasDisagg = useMemo(() => rows.some((r) => r.disagg), [rows]);
+
+  /**
+   * Every plotted chip is a single measured date, so no line has a step in it.
+   * Worth saying out loud: a flat line here means the history is one date deep,
+   * not that the software stopped improving.
+   */
+  const allSingleRung = useMemo(
+    () => rows.length > 0 && rows.every((r) => r.progression.steps.length < 2),
+    [rows],
+  );
+
+  const [view, setView] = useState<LifecycleView>('chart');
+
+  const viewOptions = useMemo<SegmentedToggleOption<LifecycleView>[]>(
+    () =>
+      LIFECYCLE_VIEW_OPTIONS.map((option) => ({
+        ...option,
+        label: option.value === 'chart' ? t.viewChart : t.viewTable,
+      })),
+    [t],
+  );
+
+  /**
+   * The same options without test ids. The toggle is rendered twice — once in the
+   * button row, once in the caption for narrow screens — and two elements sharing
+   * a test id makes every `cy.click` on it ambiguous.
+   */
+  const mobileViewOptions = useMemo<SegmentedToggleOption<LifecycleView>[]>(
+    () => viewOptions.map(({ testId: _testId, ...option }) => option),
+    [viewOptions],
+  );
+
+  const handleViewChange = useCallback((value: LifecycleView) => {
+    setView(value);
+    track('calculator_lifecycle_view_changed', { view: value });
+  }, []);
+  const chartData = useMemo<LifecycleChartSeries[]>(
+    () =>
+      rows.map((r) => {
+        // Step risers are keyed by month so the tooltip can name the run behind
+        // each one; months come from the same arithmetic the schedule used.
+        const stepInfo = new Map<
+          number,
+          { date: string; config: string; factor: number; runUrls: string[] }
+        >();
+        for (const step of r.progression.steps) {
+          const month = (Date.parse(`${step.date}T00:00:00Z`) - anchorMs) / MS_PER_MONTH;
+          // The config is now the thing that changes along the line, so name the
+          // framework that took over here, not just its precision.
+          const hwKey = step.result.hwKey ?? r.progression.baseGpu;
+          const software = configSuffix(hwKey, r.progression.baseGpu, hardwareConfig);
+          const detail = configLabel(step.result);
+          stepInfo.set(month, {
+            date: step.date,
+            config: [software === '—' ? '' : software, detail].filter(Boolean).join(' · '),
+            factor: step.factorOverFirst,
+            runUrls: step.runUrls,
+          });
+        }
+        return {
+          key: r.progression.key,
+          label: r.label,
+          color: colorResolver(r.colorKey),
+          series: r.series,
+          stepInfo,
+        };
+      }),
+    [rows, colorResolver, anchorMs],
+  );
+
+  const tokenTypeLabel = costType === 'input' ? 'input ' : costType === 'output' ? 'output ' : '';
+
+  const handleAssumption = useCallback(
+    (
+      setter: (v: string) => void,
+      param: 'c_mtbi' | 'c_rec' | 'c_life' | 'c_ramp' | 'c_cache',
+      event: string,
+    ) =>
+      (e: React.ChangeEvent<HTMLInputElement>) => {
+        const raw = e.target.value;
+        setter(raw);
+        writeUrlParams({ [param]: parseNonNegative(raw) === null ? '' : raw });
+        track(event, { value: raw });
+      },
+    [],
+  );
+
+  const handleMetricChange = useCallback((value: LifecycleMetric) => {
+    setYMetric(value);
+    writeUrlParams({ c_ly: value === 'margin' ? '' : value });
+    track('calculator_lifecycle_metric_set', { value });
+  }, []);
+
+  const handlePriceChange = useCallback(
+    (e: React.ChangeEvent<HTMLInputElement>) => {
+      const raw = e.target.value;
+      priceEdited.current = true;
+      setPriceInput(raw);
+      const input = parseNonNegative(raw);
+      const output = parseNonNegative(outputPriceInput);
+      if (input !== null && input > 0 && output !== null && output > 0) {
+        setSeedMultiple(output / input);
+      }
+      writeUrlParams({ c_price: parseNonNegative(raw) === null ? '' : raw });
+      track('calculator_lifecycle_price_set', { value: raw });
+    },
+    [outputPriceInput],
+  );
+
+  const handleOutputPriceChange = useCallback(
+    (e: React.ChangeEvent<HTMLInputElement>) => {
+      const raw = e.target.value;
+      priceEdited.current = true;
+      setOutputPriceInput(raw);
+      const input = parseNonNegative(priceInput);
+      const output = parseNonNegative(raw);
+      if (input !== null && input > 0 && output !== null && output > 0) {
+        setSeedMultiple(output / input);
+      }
+      writeUrlParams({ c_oprice: parseNonNegative(raw) === null ? '' : raw });
+      track('calculator_lifecycle_output_price_set', { value: raw });
+    },
+    [priceInput],
+  );
+
+  const handlePriceReset = useCallback(() => {
+    priceEdited.current = false;
+    if (breakEven !== null) {
+      // Scale both, keeping whatever ratio the user had set — the pair is what
+      // break-even is solved for, so resetting one alone would land off the line.
+      setPriceInput(formatPrice(breakEven));
+      setOutputPriceInput(formatPrice(breakEven * seedMultiple));
+    }
+    writeUrlParams({ c_price: '', c_oprice: '' });
+    track('calculator_lifecycle_price_reset', {});
+  }, [breakEven, seedMultiple]);
+
+  /**
+   * The table's own numbers, not the chart's samples: one row per chip, which is
+   * what a reader who asked for this section is comparing.
+   */
+  const handleExportCsv = useCallback(() => {
+    const headers = [
+      t.colChip,
+      t.colConfigNow,
+      t.colFirst,
+      t.colLatest,
+      t.colSteps,
+      t.colGain,
+      t.colChips,
+      t.colTpPerMw(tokenTypeLabel),
+      t.colUsers,
+      t.colRevenue,
+      t.colCost,
+      t.colMargin,
+      t.colPayback,
+      t.colLifetime,
+      t.colAvailability,
+    ];
+    const body = rows.map((r) => [
+      r.label,
+      r.configNow,
+      r.progression.steps[0]?.date ?? '',
+      r.progression.steps.at(-1)?.date ?? '',
+      r.series.improvementCount,
+      r.series.improvementFactor,
+      r.gpus,
+      r.tpPerMw,
+      r.concurrentUsersNow,
+      r.series.revenuePerDay,
+      r.series.costPerDay,
+      r.series.marginPerDay,
+      r.series.paybackMonth ?? '',
+      r.series.lifetimeMargin,
+      r.series.availability,
+    ]);
+    exportToCsv(`InferenceX_fleet_lifecycle_${selectedModel}`, headers, body, [
+      // The assumptions are not in the rows, and a CSV read six months later
+      // cannot be reconstructed without them.
+      `Assumptions: input $${priceInput}/M tok, output $${outputPriceInput}/M tok, ramp ${rampInput} mo, MTBI ${mtbiInput} d, recovery ${recoveryInput} h, horizon ${horizonInput} mo, power ${mw ?? ''} MW`,
+    ]);
+  }, [
+    rows,
+    t,
+    tokenTypeLabel,
+    selectedModel,
+    priceInput,
+    outputPriceInput,
+    rampInput,
+    mtbiInput,
+    recoveryInput,
+    horizonInput,
+    mw,
+  ]);
+
+  /**
+   * The two token prices, for the caption. Revenue — and so every margin plotted
+   * here — is linear in these, and on the margin metrics they are *seeded* from
+   * break-even rather than typed, so a reader comparing two screenshots needs to
+   * see them next to the MW figure to know whether they are looking at the same
+   * scenario.
+   *
+   * Null until the seed lands: the fields start empty and a break-even effect
+   * fills them, and `$/M tok` with nothing in front of it is worse than silence.
+   * The cached tier is quoted only for agentic, which is the only sequence that
+   * measures a hit rate for it to apply to — and there it is load-bearing, since
+   * most input tokens bill at it rather than at the input price beside it.
+   */
+  const captionPrices = useMemo(() => {
+    const input = parseNonNegative(priceInput);
+    const output = parseNonNegative(outputPriceInput);
+    if (input === null || output === null) return null;
+    return t.captionPrices(
+      formatCaptionPrice(input),
+      formatCaptionPrice(output),
+      isAgentic ? String(Math.round(cacheReadRatio * 100)) : null,
+    );
+  }, [priceInput, outputPriceInput, isAgentic, cacheReadRatio, t]);
+
+  /**
+   * The figure's own header, in the same shape the bar chart above uses: title,
+   * then the provenance a reader needs to know what they are looking at. The
+   * chart renders it as its caption; the table view renders it as a figcaption,
+   * so switching tabs never loses the heading.
+   */
+  const caption = (
+    <>
+      <div className="flex items-start justify-between gap-4">
+        <h2 className="text-lg font-semibold">{t.title}</h2>
+        <SegmentedToggle
+          value={view}
+          options={mobileViewOptions}
+          onValueChange={handleViewChange}
+          ariaLabel={t.viewAria}
+          className="md:hidden shrink-0"
+        />
+      </div>
+      <p className="text-sm text-muted-foreground mb-2">
+        {getModelLabel(selectedModel)} • {getSequenceLabel(selectedSequence, locale)} •{' '}
+        {t.captionTarget(targetValue)}
+        {mw ? ` • ${mw} MW` : ''}
+        {captionPrices === null ? '' : ` • ${captionPrices}`} • {t.source}SemiAnalysis
+      </p>
+    </>
+  );
+
+  const columns = useMemo<DataTableColumn<LifecycleRow>[]>(
+    () => [
+      {
+        header: t.colChip,
+        cell: (r) => r.label,
+        sortValue: (r) => r.label,
+        className: 'font-medium whitespace-nowrap',
+      },
+      {
+        header: t.colConfigNow,
+        // The line is one chip across many configs, so which config it ended on
+        // is part of reading the row.
+        cell: (r) => <span className="whitespace-nowrap">{r.configNow}</span>,
+        sortValue: (r) => r.configNow,
+        className: 'text-muted-foreground',
+      },
+      {
+        header: t.colFirst,
+        // Provenance is not optional: every rung of the line is a real sweep.
+        cell: (r) => {
+          const first = r.progression.steps[0]!;
+          return <span className="whitespace-nowrap">{runLink(first.date, first.runUrls)}</span>;
+        },
+        sortValue: (r) => r.progression.steps[0]!.date,
+      },
+      {
+        header: t.colLatest,
+        cell: (r) => {
+          const latest = r.progression.steps.at(-1)!;
+          return <span className="whitespace-nowrap">{runLink(latest.date, latest.runUrls)}</span>;
+        },
+        sortValue: (r) => r.progression.steps.at(-1)!.date,
+      },
+      {
+        header: t.colSteps,
+        align: 'right',
+        cell: (r) => String(r.series.improvementCount),
+        sortValue: (r) => r.series.improvementCount,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colGain,
+        align: 'right',
+        cell: (r) => `${r.series.improvementFactor.toFixed(2)}×`,
+        sortValue: (r) => r.series.improvementFactor,
+        className: 'tabular-nums',
+      },
+      {
+        // The physical sizing the whole projection rests on. It used to live in a
+        // separate Fleet Projection section, which meant the budget that produced
+        // it and the economics that consume it were three sections apart.
+        header: t.colChips,
+        align: 'right',
+        cell: (r) => formatCompact(r.gpus, 0),
+        sortValue: (r) => r.gpus,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colTpPerMw(tokenTypeLabel),
+        align: 'right',
+        cell: (r) => formatCompact(r.tpPerMw),
+        sortValue: (r) => r.tpPerMw,
+        className: 'tabular-nums',
+      },
+      {
+        // Streams, not tokens: the same fleet throughput read as demand served.
+        header: t.colUsers,
+        align: 'right',
+        cell: (r) => formatCompact(r.concurrentUsersNow, 0),
+        sortValue: (r) => r.concurrentUsersNow,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colRevenue,
+        align: 'right',
+        cell: (r) => formatMoney(r.series.revenuePerDay),
+        sortValue: (r) => r.series.revenuePerDay,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colCost,
+        align: 'right',
+        cell: (r) => formatMoney(r.series.costPerDay),
+        sortValue: (r) => r.series.costPerDay,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colMargin,
+        align: 'right',
+        cell: (r) => formatMoney(r.series.marginPerDay),
+        sortValue: (r) => r.series.marginPerDay,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colPayback,
+        align: 'right',
+        cell: (r) =>
+          r.series.paybackMonth === null ? (
+            <span className="text-muted-foreground">{t.never}</span>
+          ) : (
+            `${r.series.paybackMonth.toFixed(1)} ${t.monthsSuffix}`
+          ),
+        sortValue: (r) => r.series.paybackMonth ?? Infinity,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colLifetime,
+        align: 'right',
+        cell: (r) => formatMoney(r.series.lifetimeMargin),
+        sortValue: (r) => r.series.lifetimeMargin,
+        className: 'tabular-nums',
+      },
+      {
+        header: t.colAvailability,
+        align: 'right',
+        cell: (r) => `${(r.series.availability * 100).toFixed(2)}%`,
+        sortValue: (r) => r.series.availability,
+        className: 'tabular-nums',
+      },
+    ],
+    [t, tokenTypeLabel],
+  );
+
+  const assumptionInputs = [
+    {
+      id: 'calc-lifecycle-horizon',
+      label: t.horizonLabel,
+      tooltip: t.horizonTooltip,
+      value: horizonInput,
+      onChange: (e: React.ChangeEvent<HTMLInputElement>) => {
+        horizonEdited.current = true;
+        handleAssumption(setHorizonInput, 'c_life', 'calculator_lifecycle_horizon_set')(e);
+      },
+    },
+    {
+      id: 'calc-lifecycle-ramp',
+      label: t.rampLabel,
+      tooltip: t.rampTooltip,
+      value: rampInput,
+      onChange: handleAssumption(setRampInput, 'c_ramp', 'calculator_lifecycle_ramp_set'),
+    },
+    ...(isAgentic
+      ? [
+          {
+            id: 'calc-lifecycle-cache',
+            label: t.cacheLabel,
+            tooltip: t.cacheTooltip,
+            value: cacheInput,
+            onChange: handleAssumption(setCacheInput, 'c_cache', 'calculator_lifecycle_cache_set'),
+          },
+        ]
+      : []),
+    {
+      id: 'calc-lifecycle-mtbi',
+      label: t.mtbiLabel,
+      tooltip: t.mtbiTooltip,
+      value: mtbiInput,
+      onChange: handleAssumption(setMtbiInput, 'c_mtbi', 'calculator_lifecycle_mtbi_set'),
+    },
+    {
+      id: 'calc-lifecycle-recovery',
+      label: t.recoveryLabel,
+      tooltip: t.recoveryTooltip,
+      value: recoveryInput,
+      onChange: handleAssumption(setRecoveryInput, 'c_rec', 'calculator_lifecycle_recovery_set'),
+    },
+  ];
+
+  const body = () => {
+    if (!mw) {
+      return (
+        <p className="text-sm text-muted-foreground" data-testid="calculator-lifecycle-empty">
+          {t.needMw}
+        </p>
+      );
+    }
+    if (historical.error) {
+      return (
+        <p className="text-sm text-muted-foreground" data-testid="calculator-lifecycle-error">
+          {t.errorPrefix}
+          {historical.error}
+        </p>
+      );
+    }
+    if (historical.loading) {
+      return (
+        <p className="text-sm text-muted-foreground" data-testid="calculator-lifecycle-loading">
+          {t.loading}
+        </p>
+      );
+    }
+
+    return (
+      <>
+        <div className="flex flex-wrap gap-4">
+          <div className="flex flex-col space-y-1.5">
+            <LabelWithTooltip
+              htmlFor="calc-lifecycle-price"
+              label={t.priceLabel}
+              tooltip={t.priceTooltip}
+            />
+            <div className="flex items-center gap-2">
+              <Input
+                id="calc-lifecycle-price"
+                data-testid="calc-lifecycle-price-input"
+                type="number"
+                min={0}
+                step="any"
+                value={priceInput}
+                onChange={handlePriceChange}
+                aria-label={t.priceInputLabel}
+                className="w-24 h-9"
+              />
+              <Input
+                id="calc-lifecycle-output-price"
+                data-testid="calc-lifecycle-output-price-input"
+                type="number"
+                min={0}
+                step="any"
+                value={outputPriceInput}
+                onChange={handleOutputPriceChange}
+                aria-label={t.outputPriceInputLabel}
+                className="w-24 h-9"
+              />
+              {priceEdited.current && breakEven !== null && (
+                <button
+                  type="button"
+                  onClick={handlePriceReset}
+                  data-testid="calc-lifecycle-price-reset"
+                  className="text-xs text-muted-foreground hover:text-foreground underline shrink-0"
+                >
+                  {t.priceReset}
+                </button>
+              )}
+            </div>
+          </div>
+          <div className="flex flex-col space-y-1.5">
+            <LabelWithTooltip label={t.metricLabel} tooltip={t.metricTooltip} />
+            <SegmentedToggle<LifecycleMetric>
+              value={yMetric}
+              onValueChange={handleMetricChange}
+              ariaLabel={t.metricLabel}
+              testId="calc-lifecycle-metric"
+              options={[
+                { value: 'margin', label: t.metricMargin, testId: 'calc-lifecycle-metric-margin' },
+                {
+                  value: 'marginPerMw',
+                  label: t.metricMarginPerMw,
+                  testId: 'calc-lifecycle-metric-margin-per-mw',
+                },
+                {
+                  value: 'revenue',
+                  label: t.metricRevenue,
+                  testId: 'calc-lifecycle-metric-revenue',
+                },
+                {
+                  value: 'revenuePerMw',
+                  label: t.metricRevenuePerMw,
+                  testId: 'calc-lifecycle-metric-revenue-per-mw',
+                },
+                {
+                  value: 'cumulativeRevenue',
+                  label: t.metricCumulativeRevenue,
+                  testId: 'calc-lifecycle-metric-cumulative-revenue',
+                },
+              ]}
+            />
+          </div>
+          {assumptionInputs.map((input) => (
+            <div key={input.id} className="flex flex-col space-y-1.5">
+              <LabelWithTooltip htmlFor={input.id} label={input.label} tooltip={input.tooltip} />
+              <Input
+                id={input.id}
+                data-testid={`${input.id}-input`}
+                type="number"
+                min={0}
+                step={input.id === 'calc-lifecycle-ramp' ? 0.25 : 'any'}
+                value={input.value}
+                onChange={input.onChange}
+                className="w-32 h-9"
+              />
+            </div>
+          ))}
+        </div>
+
+        {rows.length > 0 && Number.isFinite(anchorMs) ? (
+          <figure data-testid="calculator-lifecycle-figure" className="relative rounded-lg">
+            <ChartButtons
+              chartId="fleet-lifecycle"
+              analyticsPrefix="calculator_lifecycle"
+              zoomResetEvent="d3chart_zoom_reset_fleet-lifecycle"
+              onExportCsv={handleExportCsv}
+              exportFileName={`InferenceX_fleet_lifecycle_${selectedModel}`}
+              // A PNG of a paginated HTML table is not a useful artefact; the CSV
+              // is the export for that view.
+              hideImageExport={view === 'table'}
+              leadingControls={
+                <SegmentedToggle
+                  value={view}
+                  options={viewOptions}
+                  onValueChange={handleViewChange}
+                  ariaLabel={t.viewAria}
+                  testId="calculator-lifecycle-view-toggle"
+                  className="shrink-0"
+                />
+              }
+            />
+            {showsJalapenoPreview && <JalapenoOfficialPreviewNotice />}
+            {showsVeraRubinPreview && <VeraRubinOfficialPreviewNotice />}
+            {view === 'table' ? (
+              <>
+                <figcaption>{caption}</figcaption>
+                <DataTable
+                  data={rows}
+                  columns={columns}
+                  testId="calculator-lifecycle-table"
+                  analyticsPrefix="calculator_lifecycle_table"
+                  // One row per chip, every one of them named in the legend and on
+                  // the chart. A search box over five rows is furniture.
+                  searchable={false}
+                />
+              </>
+            ) : (
+              <FleetLifecycleChart
+                data={chartData}
+                metric={yMetric}
+                anchorMs={anchorMs}
+                yLabel={
+                  yMetric === 'revenue'
+                    ? t.chartYRevenue
+                    : yMetric === 'revenuePerMw'
+                      ? t.chartYRevenuePerMw
+                      : yMetric === 'cumulativeRevenue'
+                        ? t.chartYCumulativeRevenue
+                        : yMetric === 'marginPerMw'
+                          ? t.chartYMarginPerMw
+                          : t.chartY
+                }
+                caption={caption}
+                breakEvenLabel={t.chartBreakEven}
+                instructions={t.chartInstructions}
+                labels={{
+                  date: t.tipDate,
+                  config: t.tipConfig,
+                  marginPerDay: t.tipMargin,
+                  marginPerMw: t.tipMarginPerMw,
+                  revenuePerDay: t.tipRevenue,
+                  revenuePerMw: t.tipRevenuePerMw,
+                  cumulativeRevenue: t.tipCumulativeRevenue,
+                  costPerDay: t.tipCost,
+                  cumulative: t.tipCumulative,
+                  runLink: t.tipRunLink,
+                  sinceFirst: t.tipSinceFirst,
+                }}
+              />
+            )}
+          </figure>
+        ) : (
+          /* Nothing to plot has two causes, and blaming the wrong one sends the
+             reader to the wrong control: chips were selected but none could be
+             sized (the budget), versus nothing measured at this speed (the
+             slider). `unplottable` distinguishes them. */
+          <p className="text-sm text-muted-foreground" data-testid="calculator-lifecycle-none">
+            {unplottable.length > 0 ? t.tooSmall : t.noneMeasured}
+          </p>
+        )}
+
+        {visibleUnmeasured.length > 0 && (
+          <div
+            className="text-xs text-muted-foreground"
+            data-testid="calculator-lifecycle-unmeasured"
+          >
+            <strong>{t.unmeasuredTitle}.</strong> {t.unmeasuredIntro}
+            <ul className="mt-1 list-disc pl-5">
+              {visibleUnmeasured.map((u) => (
+                <li key={u.hwKey}>
+                  {getLabel(u.hwKey, hardwareConfig)}:{' '}
+                  {t.unmeasuredRange(u.nearestBelow, u.nearestAbove, u.datesConsidered)}
+                </li>
+              ))}
+            </ul>
+          </div>
+        )}
+
+        {unplottable.length > 0 && (
+          <p
+            className="text-muted-foreground text-xs"
+            data-testid="calculator-lifecycle-unplottable"
+          >
+            <strong>{t.note}</strong>{' '}
+            {t.unplottable(unplottable.map((gpu) => getLabel(gpu, hardwareConfig)).join(', '))}
+          </p>
+        )}
+
+        {allSingleRung && (
+          <p
+            className="text-muted-foreground text-xs border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1"
+            data-testid="calculator-lifecycle-single-rung"
+          >
+            <strong>{t.note}</strong> {t.singleRung}
+          </p>
+        )}
+
+        {!releaseDate && rows.length > 0 && (
+          <p
+            className="text-muted-foreground text-xs"
+            data-testid="calculator-lifecycle-no-release"
+          >
+            <strong>{t.note}</strong> {t.noReleaseDate}
+          </p>
+        )}
+
+        {hasDisagg && (
+          <p className="text-muted-foreground text-xs border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1">
+            <strong>{t.note}</strong>
+            {t.disagg}
+          </p>
+        )}
+
+        <p className="text-muted-foreground text-xs border-l-2 border-border pl-2 py-1">
+          <strong>{t.note}</strong>
+          {t.hybrid}
+          {t.overlayExempt}
+        </p>
+
+        <div>
+          <p className="text-xs text-muted-foreground mt-1">
+            {t.assumptions(getCostProviderLabel(costProvider), `${mw} MW`, anchorDate ?? '—')}
+          </p>
+          <p className="text-muted-foreground mt-1">
+            <small>
+              {t.source}
+              <Link
+                target="_blank"
+                className="underline hover:text-foreground"
+                href="https://semianalysis.com/datacenter-industry-model/"
+              >
+                SemiAnalysis Datacenter Industry Model
+                <ExternalLinkIcon />
+              </Link>
+              {' & '}
+              <Link
+                target="_blank"
+                className="underline hover:text-foreground"
+                href={TCO_SOURCE_URL}
+              >
+                {TCO_SOURCE_TITLE}
+                <ExternalLinkIcon />
+              </Link>
+            </small>
+          </p>
+        </div>
+      </>
+    );
+  };
+
+  return (
+    <TooltipProvider delayDuration={0}>
+      <section data-testid="calculator-lifecycle-section">
+        <Card>
+          <div className="flex flex-col gap-4">
+            <div>
+              <h2 className="text-lg font-semibold mb-2">{t.title}</h2>
+              <p className="text-muted-foreground text-sm">{t.description}</p>
+            </div>
+            {/* Outside `body()` on purpose: every other control is only meaningful
+                once a fleet exists, but this is the one that brings it into being,
+                so it has to render in the empty state too. */}
+            <div className="flex flex-col space-y-1.5">
+              <LabelWithTooltip htmlFor="calc-fleet-mw" label={t.mwLabel} tooltip={t.mwTooltip} />
+              <Input
+                id="calc-fleet-mw"
+                data-testid="calc-fleet-mw-input"
+                type="number"
+                min={0}
+                step="any"
+                placeholder={t.mwPlaceholder}
+                value={mwInput}
+                onChange={(e) => onMwInputChange(e.target.value)}
+                onBlur={() => track('calculator_fleet_mw_set', { mw: mwInput })}
+                className="w-32 h-9"
+              />
+            </div>
+            {body()}
+          </div>
+        </Card>
+      </section>
+    </TooltipProvider>
+  );
+}
diff --git a/packages/app/src/components/calculator/FleetLifecycleChart.tsx b/packages/app/src/components/calculator/FleetLifecycleChart.tsx
new file mode 100644
index 000000000..d6b3e8bef
--- /dev/null
+++ b/packages/app/src/components/calculator/FleetLifecycleChart.tsx
@@ -0,0 +1,688 @@
+'use client';
+
+import * as d3 from 'd3';
+import React, { useCallback, useMemo, useRef } from 'react';
+
+import { formatLargeNumber } from '@/lib/chart-rendering';
+import { getChartWatermark } from '@/lib/data-mappings';
+import {
+  D3Chart,
+  type D3ChartHandle,
+  type LayerConfig,
+  type RenderContext,
+  type ScaleConfig,
+} from '@/lib/d3-chart/D3Chart';
+import { escapeHtml } from '@/lib/utils';
+
+import {
+  isBreakEvenAnchored,
+  metricValue,
+  MS_PER_MONTH,
+  valueAtMonth,
+  type LifecycleMetric,
+  type LifecyclePoint,
+  type LifecycleSeries,
+} from './lifecycle';
+
+// Right margin holds the end-of-line chip labels, which sit outside the plot.
+const CHART_MARGIN = { top: 20, right: 104, bottom: 50, left: 70 };
+
+/** Minimum vertical gap between end-of-line labels, px. */
+const LABEL_MIN_GAP = 13;
+
+/**
+ * Half the label's cap height, near enough. Labels are centred on their line's
+ * end value, so a series plateauing at the very top of the domain would have
+ * half its glyphs above the SVG edge without this inset.
+ */
+const LABEL_HALF_HEIGHT = 6;
+
+/** One plotted series: an hwKey's margin staircase since the model shipped. */
+export interface LifecycleChartSeries {
+  key: string;
+  label: string;
+  color: string;
+  series: LifecycleSeries;
+  /** Run date, config label and run links for each step riser, keyed by month. */
+  stepInfo: Map<number, { date: string; config: string; factor: number; runUrls: string[] }>;
+}
+
+/** A hovered step riser — one measured config improvement. */
+interface StepMarker {
+  x: number;
+  y: number;
+  precision: string;
+  seriesKey: string;
+  month: number;
+}
+
+/**
+ * Hover positions the readout can snap to: a uniform grid across the window.
+ *
+ * The alternative — snapping to the sample points themselves — reads badly here,
+ * because sample density varies by two orders of magnitude along a line. Each
+ * rollout carries 24 samples over its ramp (a day or two apart) while a plateau
+ * carries a single sample spanning months, so the cursor would crawl through a
+ * riser and then jump a year. A uniform grid moves at a constant speed, and at
+ * this count each slice is well under a pixel wide unzoomed.
+ */
+const HOVER_SLICE_COUNT = 1200;
+
+/**
+ * How near a slice must be to a step for the readout to include that step's
+ * config and run links, in slices.
+ *
+ * A step's own instant almost never coincides with a grid slice, and at this
+ * density a slice is a fraction of a pixel — so matching by equality would make
+ * the run links unreachable by pointing at the dot, which is the only place a
+ * reader would think to look for them. Four slices is about three pixels
+ * unzoomed: near enough to mean "that dot", far enough to be hittable.
+ */
+const STEP_MATCH_SLICES = 4;
+
+/** One position on the time axis the readout can rest at. */
+interface HoverSlice {
+  /** Timestamp, ms. */
+  x: number;
+  /** Present only on measured-step scatter markers. */
+  seriesKey?: string;
+}
+
+/** A measured config change, placed on the time axis. */
+interface StepAt {
+  x: number;
+  seriesKey: string;
+  month: number;
+}
+
+/** One series' state at a hovered slice. */
+interface SliceReading {
+  label: string;
+  color: string;
+  value: number;
+  cumulative: number;
+}
+
+/** Which lifecycle quantity the y axis plots. Defined in `lifecycle.ts`, which selects on it. */
+export type { LifecycleMetric } from './lifecycle';
+
+interface FleetLifecycleChartProps {
+  data: LifecycleChartSeries[];
+  metric: LifecycleMetric;
+  /** Anchor for the x axis — the model's release date, as a timestamp. */
+  anchorMs: number;
+  yLabel: string;
+  labels: {
+    date: string;
+    config: string;
+    marginPerDay: string;
+    marginPerMw: string;
+    revenuePerDay: string;
+    revenuePerMw: string;
+    cumulativeRevenue: string;
+    costPerDay: string;
+    /** Names the running total column, which is cumulative *margin* whatever the y axis plots. */
+    cumulative: string;
+    sinceFirst: string;
+    runLink: string;
+  };
+  breakEvenLabel: string;
+  /** Localised hint under the chart. */
+  instructions: string;
+  legendElement?: React.ReactNode;
+  caption?: React.ReactNode;
+}
+
+/** CSS class names derive from series keys, so strip separators as TrendChart does. */
+const safeKey = (key: string) => key.replaceAll(/[|_\s]/gu, '-');
+
+/** Signed money, negative rendered as -$X rather than $-X. */
+const money = (value: number) => `${value < 0 ? '-$' : '$'}${formatLargeNumber(Math.abs(value))}`;
+
+const FleetLifecycleChart = React.memo(
+  ({
+    data,
+    metric,
+    anchorMs,
+    yLabel,
+    labels,
+    breakEvenLabel,
+    instructions,
+    legendElement,
+    caption,
+  }: FleetLifecycleChartProps) => {
+    const chartRef = useRef<D3ChartHandle | null>(null);
+    const toMs = useCallback((month: number) => anchorMs + month * MS_PER_MONTH, [anchorMs]);
+    const valueOf = useCallback((point: LifecyclePoint) => metricValue(point, metric), [metric]);
+
+    const { lineDataRecord, markers, bySafeKey } = useMemo(() => {
+      const record: Record<string, { x: number; y: number }[]> = {};
+      const flat: StepMarker[] = [];
+      const lookup = new Map<string, LifecycleChartSeries>();
+
+      for (const entry of data) {
+        const key = safeKey(entry.key);
+        lookup.set(key, entry);
+        record[key] = entry.series.points.map((p) => ({ x: toMs(p.month), y: valueOf(p) }));
+        // Markers mark measured config changes and nothing else — no rug of
+        // synthetic samples, and every dot is a sweep the user can look up.
+        for (const p of entry.series.points) {
+          if (!p.isStep) continue;
+          flat.push({
+            x: toMs(p.month),
+            y: valueOf(p),
+            precision: 'fp4',
+            seriesKey: key,
+            month: p.month,
+          });
+        }
+      }
+      return { lineDataRecord: record, markers: flat, bySafeKey: lookup };
+    }, [data, toMs, valueOf]);
+
+    /** The grid the hover readout snaps to, plus where the steps sit on it. */
+    const { hoverSlices, stepsOnAxis, sliceSpacing } = useMemo(() => {
+      let min = Infinity;
+      let max = -Infinity;
+      for (const entry of data) {
+        min = Math.min(min, toMs(entry.series.startMonth));
+        max = Math.max(max, toMs(entry.series.endMonth));
+      }
+      if (!Number.isFinite(min) || !(max > min)) {
+        return { hoverSlices: [], stepsOnAxis: [], sliceSpacing: 0 };
+      }
+
+      const spacing = (max - min) / HOVER_SLICE_COUNT;
+      const slices: HoverSlice[] = Array.from({ length: HOVER_SLICE_COUNT + 1 }, (_, i) => ({
+        x: min + i * spacing,
+      }));
+      const steps: StepAt[] = [];
+      for (const entry of data) {
+        for (const point of entry.series.points) {
+          if (!point.isStep) continue;
+          steps.push({ x: toMs(point.month), seriesKey: safeKey(entry.key), month: point.month });
+        }
+      }
+      return { hoverSlices: slices, stepsOnAxis: steps, sliceSpacing: spacing };
+    }, [data, toMs]);
+
+    /**
+     * What the custom layers should draw *now*.
+     *
+     * The zoom behaviour installed by a render captures that render's layer
+     * callbacks, and `d3.zoom`'s double-click reset is a transition — so its
+     * trailing events can fire after a later render has already redrawn. A stale
+     * callback then re-appends what the new render just removed: switch the axis to
+     * revenue mid-reset and the break-even rule comes back and stays, because
+     * nothing renders again afterwards. Reading through a ref keeps every draw,
+     * stale closure or not, consistent with the current props.
+     */
+    const latest = useRef({ metric, lineDataRecord, bySafeKey });
+    latest.current = { metric, lineDataRecord, bySafeKey };
+
+    const xScaleConfig = useMemo<ScaleConfig>(() => {
+      let min = Infinity;
+      let max = -Infinity;
+      for (const entry of data) {
+        min = Math.min(min, toMs(entry.series.startMonth));
+        max = Math.max(max, toMs(entry.series.endMonth));
+      }
+      if (!Number.isFinite(min)) return { type: 'time', domain: [new Date(), new Date()] };
+      // nice defaults to true in scale-builders, which rounds a multi-year span
+      // out to whole years and leaves dead space before the release date and
+      // after the horizon. The domain here is meaningful on both ends.
+      return { type: 'time', domain: [new Date(min), new Date(max)], nice: false };
+    }, [data, toMs]);
+
+    const yScaleConfig = useMemo<ScaleConfig>(() => {
+      let min = 0;
+      let max = 0;
+      for (const entry of data) {
+        for (const p of entry.series.points) {
+          const value = valueOf(p);
+          if (value < min) min = value;
+          if (value > max) max = value;
+        }
+      }
+      // Always include zero. For margin that is the whole question — which side of
+      // break-even a fleet is on; for revenue it is the floor a rollout starts from.
+      const span = max - min || Math.max(Math.abs(max), 1);
+      return { type: 'linear', domain: [min - span * 0.05, max + span * 0.05], nice: true };
+    }, [data, valueOf]);
+
+    /**
+     * The break-even rule. No reference-line layer exists to reuse, and a custom
+     * layer must be idempotent: the chart re-renders into the same zoom group, so
+     * appending unconditionally leaves a stale rule behind at the previous
+     * y-scale on every data change.
+     */
+    const drawZeroRule = useCallback(
+      (
+        zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+        ctx: RenderContext,
+        yScale: d3.ScaleLinear<number, number>,
+      ) => {
+        const y = yScale(0);
+        zoomGroup.selectAll('.lifecycle-zero-rule').remove();
+        // Zero is break-even only for margin. On a revenue axis it is just the
+        // bottom of the scale, and labelling it "break-even" would be a lie: each
+        // chip breaks even at its own cost line, not at zero revenue.
+        if (!isBreakEvenAnchored(latest.current.metric)) return;
+        if (!Number.isFinite(y)) return;
+
+        // Lowered explicitly: `renderLines` keeps its paths in place across
+        // re-renders via a data join, while this layer removes and re-appends, so
+        // without this the rule climbs above the lines after the first repaint.
+        const group = zoomGroup.append('g').attr('class', 'lifecycle-zero-rule').lower();
+        group
+          .append('line')
+          .attr('x1', 0)
+          .attr('x2', ctx.width)
+          .attr('y1', y)
+          .attr('y2', y)
+          .attr('stroke', 'currentColor')
+          .attr('stroke-width', 1)
+          .attr('stroke-dasharray', '4 3')
+          .attr('opacity', 0.55);
+        group
+          .append('text')
+          .attr('x', 2)
+          .attr('y', y - 5)
+          .attr('text-anchor', 'start')
+          .attr('fill', 'currentColor')
+          .attr('opacity', 0.7)
+          .attr('font-size', 10)
+          .text(breakEvenLabel);
+      },
+      [breakEvenLabel],
+    );
+
+    const renderZeroRule = useCallback(
+      (zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>, ctx: RenderContext) =>
+        drawZeroRule(zoomGroup, ctx, ctx.yScale as d3.ScaleLinear<number, number>),
+      [drawZeroRule],
+    );
+
+    // Without this the rule stays pinned to the base scale while the lines move,
+    // so a zoomed chart shows break-even in the wrong place.
+    const zoomZeroRule = useCallback(
+      (zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>, ctx: RenderContext) =>
+        drawZeroRule(
+          zoomGroup,
+          ctx,
+          ((ctx as { newYScale?: unknown }).newYScale ?? ctx.yScale) as d3.ScaleLinear<
+            number,
+            number
+          >,
+        ),
+      [drawZeroRule],
+    );
+
+    /**
+     * Chip names at the right end of each line, so a series can be identified
+     * without crossing back to the legend.
+     *
+     * Drawn into `ctx.layout.g` rather than the zoom group: layers render into a
+     * clipped group (`clipContent` defaults to true), and these sit past the plot
+     * width, so they would be clipped away entirely.
+     */
+    const drawSeriesLabels = useCallback(
+      (
+        ctx: RenderContext,
+        xScale: d3.ScaleTime<number, number>,
+        yScale: d3.ScaleLinear<number, number>,
+      ) => {
+        const group = ctx.layout.g;
+        group.selectAll('.lifecycle-series-label').remove();
+
+        // Where the plot is cut off, in data units. Panning moves it, so the label
+        // must follow the line's *visible* end, not its last data point.
+        const edge = Number(xScale.invert(ctx.width));
+
+        const { lineDataRecord: lines, bySafeKey: lookup } = latest.current;
+        const placed = Object.entries(lines)
+          .flatMap(([key, points]) => {
+            const entry = lookup.get(key);
+            const last = points.at(-1);
+            const first = points[0];
+            if (!entry || !last || !first) return [];
+            // Panned entirely past the right edge: there is no line to name.
+            if (first.x > edge) return [];
+
+            let value = last.y;
+            let x = xScale(last.x);
+            if (last.x > edge) {
+              // The line runs off the edge. Interpolate the segment that straddles
+              // it, so the label sits level with where the line actually leaves.
+              x = ctx.width;
+              for (let i = 1; i < points.length; i += 1) {
+                const before = points[i - 1]!;
+                const after = points[i]!;
+                if (after.x < edge) continue;
+                const span = after.x - before.x;
+                const t = span > 0 ? (edge - before.x) / span : 0;
+                value = before.y + (after.y - before.y) * t;
+                break;
+              }
+            }
+
+            const y = yScale(value);
+            if (!Number.isFinite(y) || !Number.isFinite(x)) return [];
+            // Zooming the y axis can push the end value outside the visible range;
+            // pin the label to the nearest edge rather than let it escape.
+            const top = LABEL_HALF_HEIGHT;
+            const bottom = ctx.height - LABEL_HALF_HEIGHT;
+            return [
+              {
+                label: entry.label,
+                color: entry.color,
+                x: Math.min(Math.max(x, 0), ctx.width) + 6,
+                y: Math.min(Math.max(y, top), bottom),
+              },
+            ];
+          })
+          .toSorted((a, b) => a.y - b.y);
+
+        // Nudge downwards so near-identical end values stay legible.
+        for (let i = 1; i < placed.length; i += 1) {
+          const previous = placed[i - 1]!;
+          const current = placed[i]!;
+          if (current.y - previous.y < LABEL_MIN_GAP) current.y = previous.y + LABEL_MIN_GAP;
+        }
+
+        // Nudging only ever pushes down, so the stack can run off the bottom.
+        // Slide the whole block back up by the overflow, keeping the gaps.
+        const overflow = (placed.at(-1)?.y ?? 0) - (ctx.height - LABEL_HALF_HEIGHT);
+        if (overflow > 0) {
+          for (const entry of placed) entry.y = Math.max(entry.y - overflow, LABEL_HALF_HEIGHT);
+        }
+
+        for (const { label, color, x, y } of placed) {
+          group
+            .append('text')
+            .attr('class', 'lifecycle-series-label')
+            .attr('x', x)
+            .attr('y', y)
+            .attr('dominant-baseline', 'middle')
+            .attr('fill', color)
+            .attr('font-size', 11)
+            .attr('font-weight', 500)
+            .text(label);
+        }
+      },
+      [],
+    );
+
+    const renderSeriesLabels = useCallback(
+      (_zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>, ctx: RenderContext) =>
+        drawSeriesLabels(
+          ctx,
+          ctx.xScale as d3.ScaleTime<number, number>,
+          ctx.yScale as d3.ScaleLinear<number, number>,
+        ),
+      [drawSeriesLabels],
+    );
+
+    const zoomSeriesLabels = useCallback(
+      (_zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>, ctx: RenderContext) => {
+        const zoomed = ctx as { newXScale?: unknown; newYScale?: unknown };
+        drawSeriesLabels(
+          ctx,
+          (zoomed.newXScale ?? ctx.xScale) as d3.ScaleTime<number, number>,
+          (zoomed.newYScale ?? ctx.yScale) as d3.ScaleLinear<number, number>,
+        );
+      },
+      [drawSeriesLabels],
+    );
+
+    const layers = useMemo<LayerConfig<HoverSlice>[]>(
+      () => [
+        {
+          type: 'custom' as const,
+          key: 'lifecycle-zero',
+          render: renderZeroRule,
+          onZoom: zoomZeroRule,
+        },
+        {
+          type: 'line' as const,
+          key: 'lifecycle-lines',
+          lines: lineDataRecord,
+          config: {
+            getColor: (key: string) => bySafeKey.get(key)?.color ?? '#888',
+            strokeWidth: 2,
+            // Every riser is a rollout curve and every plateau is flat, and the
+            // samples already carry that shape — so interpolate linearly and draw
+            // exactly what was computed. A step curve would flatten the rollouts
+            // into stairs; a spline would overshoot on the near-vertical ones.
+            curve: d3.curveLinear,
+          },
+        },
+        {
+          type: 'scatter' as const,
+          key: 'lifecycle-steps',
+          data: markers,
+          config: {
+            getColor: (d: HoverSlice & { precision: string; y: number }) =>
+              bySafeKey.get(d.seriesKey ?? '')?.color ?? '#888',
+          },
+        },
+        {
+          type: 'custom' as const,
+          key: 'lifecycle-series-labels',
+          render: renderSeriesLabels,
+          onZoom: zoomSeriesLabels,
+        },
+      ],
+      [
+        renderZeroRule,
+        zoomZeroRule,
+        renderSeriesLabels,
+        zoomSeriesLabels,
+        lineDataRecord,
+        markers,
+        bySafeKey,
+      ],
+    );
+
+    /**
+     * A series' plotted rate and cumulative margin at an arbitrary instant.
+     *
+     * Interpolates linearly between samples, which is exactly what the line does
+     * (`curveLinear`), so the readout always agrees with the pixel under the rule.
+     * Returns null outside the series' own window — a chip first measured a year
+     * after release has no line, and so no number, before then.
+     */
+    const readingAt = useCallback(
+      (entry: LifecycleChartSeries, month: number): SliceReading | null => {
+        const points = entry.series.points;
+        const value = valueAtMonth(points, month, valueOf);
+        const cumulative = valueAtMonth(points, month, (p) => p.cumulative);
+        if (value === null || cumulative === null) return null;
+        return { label: entry.label, color: entry.color, value, cumulative };
+      },
+      [valueOf],
+    );
+
+    const sliceDate = useMemo(() => d3.timeFormat('%d %b %Y'), []);
+
+    /**
+     * Was the readout frozen when the current click started?
+     *
+     * Written by a capture-phase listener on the SVG, so it is settled before the
+     * plot overlay's own bubble-phase handler pins the new slice. Reading the
+     * pinned state afterwards would always say "frozen" and the toggle below could
+     * never tell a freeze from a release.
+     */
+    const wasFrozen = useRef(false);
+
+    const handleRender = useCallback((ctx: RenderContext) => {
+      // Named listener, so re-renders replace it rather than stacking copies.
+      ctx.layout.svg.on(
+        'click.lifecycle-freeze',
+        () => {
+          wasFrozen.current = chartRef.current?.isPinned() ?? false;
+        },
+        { capture: true },
+      );
+    }, []);
+
+    /**
+     * The hover readout: every visible chip's number at the instant under the
+     * rule, in one popup.
+     *
+     * Per-point tooltips answered "what is this dot?", which is the wrong question
+     * for this chart — the lines are only interesting against each other, and a
+     * reader comparing two chips at one date had to hover twice and hold the first
+     * number in their head. This reads every line at once. Steps keep their
+     * detail: any step nearest this slice contributes its config and, once frozen,
+     * its run links, so the per-rung audit trail survives.
+     */
+    const tooltipConfig = useMemo(
+      () => ({
+        rulerType: 'vertical' as const,
+        proximityHover: true,
+        getDataX: (d: HoverSlice) => d.x,
+        content: (d: HoverSlice, isPinned: boolean) => {
+          const month = (d.x - anchorMs) / MS_PER_MONTH;
+          const readings: SliceReading[] = [];
+          for (const entry of data) {
+            const reading = readingAt(entry, month);
+            if (reading) readings.push(reading);
+          }
+          readings.sort((a, b) => b.value - a.value);
+
+          const valueHeader =
+            metric === 'revenue'
+              ? labels.revenuePerDay
+              : metric === 'revenuePerMw'
+                ? labels.revenuePerMw
+                : metric === 'cumulativeRevenue'
+                  ? labels.cumulativeRevenue
+                  : metric === 'marginPerMw'
+                    ? labels.marginPerMw
+                    : labels.marginPerDay;
+          const rows = readings
+            .map(
+              (r) =>
+                `<tr><td class="pr-3 whitespace-nowrap" style="color: ${escapeHtml(r.color)}">${escapeHtml(r.label)}</td>
+                <td class="pr-3 text-right tabular-nums font-medium">${money(r.value)}</td>
+                <td class="text-right tabular-nums text-muted-foreground">${money(r.cumulative)}</td></tr>`,
+            )
+            .join('');
+
+          // Config detail is frozen-only. On hover the popup is a scanning tool —
+          // one line per chip, the same shape wherever the cursor is — and a step
+          // block appearing under it as the cursor crosses a dot both reflows the
+          // rows and buries the comparison the reader came for. A click says "tell
+          // me about this instant", and that is when the detail belongs.
+          const tolerance = sliceSpacing * STEP_MATCH_SLICES;
+          const steps = isPinned
+            ? stepsOnAxis
+                .filter((step) => Math.abs(step.x - d.x) <= tolerance)
+                .flatMap((step) => {
+                  const entry = bySafeKey.get(step.seriesKey);
+                  const info = entry?.stepInfo.get(step.month);
+                  if (!entry || !info) return [];
+                  const factor = info.factor > 1.005 ? ` (${info.factor.toFixed(2)}x)` : '';
+                  // Every rung links its run, not just the first and last. The
+                  // table only links the opening and closing sweeps, so without
+                  // this the intermediate rungs — exactly where an anomalous run
+                  // that was never purged would sit — have no audit trail
+                  // anywhere in the UI.
+                  const links = info.runUrls
+                    .map(
+                      (url, i) =>
+                        `<a href="${escapeHtml(url)}" target="_blank" rel="noopener noreferrer" class="text-primary underline mr-2">${escapeHtml(labels.runLink)}${
+                          info.runUrls.length > 1 ? ` ${i + 1}` : ''
+                        }</a>`,
+                    )
+                    .join('');
+                  return [
+                    `<div class="mt-1">
+                  <span style="color: ${escapeHtml(entry.color)}">${escapeHtml(entry.label)}</span>
+                  <span class="text-muted-foreground"> ${escapeHtml(labels.config)}: ${escapeHtml(info.config)}${escapeHtml(factor)} · ${escapeHtml(labels.date)} ${escapeHtml(info.date)}</span>${links ? ` ${links}` : ''}
+                </div>`,
+                  ];
+                })
+                .join('')
+            : '';
+
+          return `<div class="rounded-md border bg-background/95 px-3 py-2 text-xs shadow-md backdrop-blur-sm" style="min-width: 240px; user-select: ${isPinned ? 'text' : 'none'};">
+            <div class="font-semibold mb-1">${escapeHtml(sliceDate(new Date(d.x)))}</div>
+            ${
+              rows
+                ? `<table class="w-full"><thead><tr class="text-muted-foreground font-normal">
+                    <th></th>
+                    <th class="pr-3 text-right font-normal">${escapeHtml(valueHeader)}</th>
+                    <th class="text-right font-normal">${escapeHtml(labels.cumulative)}</th>
+                  </tr></thead><tbody>${rows}</tbody></table>`
+                : ''
+            }
+            ${steps}
+          </div>`;
+        },
+        getRulerX: (d: HoverSlice, xScale: any) => xScale(d.x),
+        onPointClick: () => {
+          // The overlay pins on every click, so a second click would just re-pin.
+          // `wasFrozen` is read in the capture phase, before that pin lands, which
+          // makes the same gesture a toggle: click to freeze, click to release.
+          if (!wasFrozen.current) return;
+          chartRef.current?.dismissTooltip();
+          chartRef.current?.hideTooltip();
+        },
+      }),
+      [anchorMs, bySafeKey, data, labels, metric, readingAt, sliceDate, sliceSpacing, stepsOnAxis],
+    );
+
+    const xAxisConfig = useMemo(
+      () => ({
+        tickFormat: d3.timeFormat('%b %Y') as any,
+        tickCount: 8,
+      }),
+      [],
+    );
+
+    const yAxisConfig = useMemo(
+      () => ({
+        label: yLabel,
+        tickFormat: (v: d3.AxisDomain) => money(v as number),
+        tickCount: 8,
+      }),
+      [yLabel],
+    );
+
+    // x only: the y axis is a money scale whose zero line is the whole point, and
+    // rescaling it would move break-even under the reader. Matches TrendChart.
+    const zoomConfig = useMemo(
+      () => ({ enabled: true, axes: 'x' as const, scaleExtent: [1, 10] as [number, number] }),
+      [],
+    );
+
+    return (
+      <D3Chart<HoverSlice>
+        ref={chartRef}
+        chartId="fleet-lifecycle"
+        data={hoverSlices}
+        height={420}
+        margin={CHART_MARGIN}
+        watermark={getChartWatermark()}
+        testId="calculator-lifecycle-chart-svg"
+        xScale={xScaleConfig}
+        yScale={yScaleConfig}
+        xAxis={xAxisConfig}
+        yAxis={yAxisConfig}
+        layers={layers}
+        zoom={zoomConfig}
+        instructions={instructions}
+        tooltip={tooltipConfig}
+        onRender={handleRender}
+        legendElement={legendElement}
+        caption={caption}
+      />
+    );
+  },
+);
+
+FleetLifecycleChart.displayName = 'FleetLifecycleChart';
+
+export default FleetLifecycleChart;
diff --git a/packages/app/src/components/calculator/FleetLifecycleDisplay.tsx b/packages/app/src/components/calculator/FleetLifecycleDisplay.tsx
new file mode 100644
index 000000000..5f54ef466
--- /dev/null
+++ b/packages/app/src/components/calculator/FleetLifecycleDisplay.tsx
@@ -0,0 +1,607 @@
+'use client';
+
+import { useCallback, useMemo, useState } from 'react';
+
+import FleetLifecycle from '@/components/calculator/FleetLifecycle';
+import {
+  resolveCalculatorTarget,
+  resolveCalculatorTargetInputValue,
+  resolveCalculatorVisibility,
+  type CalculatorVisibilityIntent,
+} from '@/components/calculator/ThroughputCalculatorDisplay';
+import type { CalculatorUrlSeed } from '@/components/calculator/url-seed';
+import {
+  GlobalFilterProvider,
+  useGlobalFilterActions,
+  useGlobalFilterAvailability,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+} from '@/components/GlobalFilterContext';
+import { Card } from '@/components/ui/card';
+import ChartLegendItem from '@/components/ui/chart-legend-item';
+import { ChartShareActions } from '@/components/ui/chart-display-helpers';
+import {
+  ModelSelector,
+  PercentileSelector,
+  PrecisionSelector,
+  ScenarioSelector,
+} from '@/components/ui/chart-selectors';
+import { Input } from '@/components/ui/input';
+import { Label } from '@/components/ui/label';
+import { LabelWithTooltip } from '@/components/ui/label-with-tooltip';
+import { MultiSelect } from '@/components/ui/multi-select';
+import { Skeleton } from '@/components/ui/skeleton';
+import { Switch } from '@/components/ui/switch';
+import { TooltipProvider } from '@/components/ui/tooltip';
+import { useThemeColors } from '@/hooks/useThemeColors';
+import { useOpenDropdown } from '@/hooks/useOpenDropdown';
+import { useUrlState } from '@/hooks/useUrlState';
+import { track } from '@/lib/analytics';
+import { getModelSortIndex } from '@/lib/constants';
+import { Percentile, Sequence, type Model } from '@/lib/data-mappings';
+import { DEFAULT_FLEET_MW, readUrlParams, writeUrlParams } from '@/lib/url-state';
+import { useFeatureGate } from '@/lib/use-feature-gate';
+import { useLocale } from '@/lib/use-locale';
+import { getDisplayLabel } from '@/lib/utils';
+
+import type { CostProvider, CostType } from './types';
+import { useThroughputData } from './useThroughputData';
+
+const COST_PROVIDER_OPTIONS: {
+  value: CostProvider;
+  label: string;
+  labelZh: string;
+}[] = [
+  { value: 'costh', label: 'Hyperscaler', labelZh: '超大规模云服务商' },
+  { value: 'costn', label: 'Neocloud', labelZh: 'Neocloud' },
+  { value: 'costr', label: '3yr Rental', labelZh: '3 年租赁' },
+];
+
+const COST_TYPE_OPTIONS: { value: CostType; label: string }[] = [
+  { value: 'total', label: 'Total Tokens' },
+  { value: 'input', label: 'Input Tokens' },
+  { value: 'output', label: 'Output Tokens' },
+];
+
+const STRINGS = {
+  en: {
+    title: 'Fleet Lifecycle',
+    description:
+      'Pick the model, workload, and target interactivity. The projection below sizes a fixed fleet of each chip against a facility power budget and reads the full run history at this operating point — see the section itself for what the lines mean.',
+    costProviderLabel: 'Cost Provider',
+    costProviderTooltip:
+      'The pricing tier used for the fleet cost line. Hyperscaler (e.g. AWS/GCP), Neocloud (e.g. CoreWeave), or 3-year rental.',
+    costProviderPlaceholder: 'Cost provider',
+    tokenTypeLabel: 'Token Type',
+    tokenTypeTooltip:
+      'Whether the per-chip figures quoted in the table use total tokens, input tokens only, or output tokens only. Revenue always counts both streams.',
+    tokenTypePlaceholder: 'Token type',
+    targetLabel: 'Target Interactivity (tok/s/user)',
+    targetTooltip:
+      'The interactivity operating point used for interpolation. Each chip\u2019s fleet serves this speed; every run date is read at it.',
+    targetAgenticLabel: (percentile: string) => `Target ${percentile} Interactivity (tok/s/user)`,
+    targetAgenticTooltip: (percentile: string) =>
+      `The ${percentile} interactivity operating point used for agentic workload interpolation. Each chip\u2019s fleet serves this speed; every run date is read at it.`,
+    errorLoading: 'Error loading data. Please try a different selection.',
+    highContrast: 'High Contrast',
+    resetFilter: 'Reset filter',
+  },
+  zh: {
+    title: '集群生命周期',
+    description:
+      '选择模型、工作负载与目标交互性。下方的测算会按设施功率预算确定各芯片固定集群的规模，并在该操作点上读取完整运行历史——各条曲线的含义见板块内说明。',
+    costProviderLabel: '成本供应商',
+    costProviderTooltip:
+      '集群成本线采用的定价层级。Hyperscaler（如 AWS/GCP）、Neocloud（如 CoreWeave）或 3 年租赁。',
+    costProviderPlaceholder: '成本供应商',
+    tokenTypeLabel: 'Token 类型',
+    tokenTypeTooltip:
+      '表格中每芯片指标按总 token、仅输入 token 还是仅输出 token 计。收入始终同时计入两类 token。',
+    tokenTypePlaceholder: 'Token 类型',
+    targetLabel: '目标交互性 (tok/s/user)',
+    targetTooltip:
+      '用于插值的交互性操作点。每款芯片的集群按此速度提供服务；每个运行日期都在该点读取。',
+    targetAgenticLabel: (percentile: string) => `目标 ${percentile} 交互性 (tok/s/user)`,
+    targetAgenticTooltip: (percentile: string) =>
+      `用于智能体工作负载插值的 ${percentile} 交互性操作点。每款芯片的集群按此速度提供服务；每个运行日期都在该点读取。`,
+    errorLoading: '加载数据出错，请尝试其他选择。',
+    highContrast: '高对比度',
+    resetFilter: '重置筛选',
+  },
+} as const;
+
+const COST_TYPE_LABELS_ZH: Record<CostType, string> = {
+  total: '总 Token',
+  input: '输入 Token',
+  output: '输出 Token',
+};
+
+export default function FleetLifecycleDisplay({ urlSeed }: { urlSeed?: CalculatorUrlSeed }) {
+  return (
+    <GlobalFilterProvider
+      initialModel={urlSeed?.model}
+      initialSequence={urlSeed?.sequence}
+      initialPrecisions={urlSeed?.precisions}
+      initialRunDate={urlSeed?.runDate}
+      initialRunId={urlSeed?.runId}
+    >
+      <FleetLifecycleInner initialPercentile={urlSeed?.percentile ?? Percentile.P90} />
+    </GlobalFilterProvider>
+  );
+}
+
+function FleetLifecycleInner({ initialPercentile }: { initialPercentile: Percentile }) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const { setUrlParam } = useUrlState();
+  const { openDropdown, handleDropdownOpenChange } = useOpenDropdown();
+
+  const {
+    selectedModel,
+    effectiveSequence: selectedSequence,
+    effectivePrecisions: selectedPrecisions,
+  } = useGlobalFilterSelection();
+  const { setSelectedModel, setSelectedSequence, setSelectedPrecisions } = useGlobalFilterActions();
+  const { selectedRunDate } = useGlobalFilterRun();
+  const { availablePrecisions, availableSequences, availableModels } =
+    useGlobalFilterAvailability();
+  const mode = 'interactivity_to_throughput' as const;
+  const [costProvider, setCostProvider] = useState<CostProvider>('costh');
+  const [costType, setCostType] = useState<CostType>('total');
+  const [requestedTargetValue, setRequestedTargetValue] = useState<number>(35);
+  const [inputValue, setInputValue] = useState<string>('35');
+  const [isTargetInputFocused, setIsTargetInputFocused] = useState(false);
+  const [selectedPercentile, setSelectedPercentile] = useState<Percentile>(initialPercentile);
+  // Owned here so the `c_mw` URL seed is read once, at the page level — the
+  // lifecycle section renders the input and is its only consumer.
+  const [mwInput, setMwInput] = useState<string>(() => readUrlParams().c_mw ?? DEFAULT_FLEET_MW);
+  const [visibilityIntent, setVisibilityIntent] = useState<CalculatorVisibilityIntent | null>(null);
+  const [highContrast, setHighContrast] = useState(false);
+
+  const { hardwareConfig, ranges, loading, error, hasData, availableHwKeys } = useThroughputData(
+    selectedModel,
+    selectedSequence,
+    selectedPrecisions,
+    selectedRunDate,
+    undefined,
+    selectedPercentile,
+    undefined,
+    true,
+    costType,
+  );
+
+  const isAgenticSequence = selectedSequence === Sequence.AgenticTraces;
+  // AgentX publishes on P90, so the percentile control stays behind the
+  // ↑↑↓↓ feature gate, matching the calculator.
+  const featureGateUnlocked = useFeatureGate();
+  const percentileLabel = selectedPercentile.toUpperCase();
+
+  const handleMwInputChange = useCallback((raw: string) => {
+    setMwInput(raw);
+    const parsed = parseFloat(raw);
+    writeUrlParams({ c_mw: Number.isFinite(parsed) && parsed > 0 ? raw : '' });
+  }, []);
+
+  // The selection key represents user-driven scope, mirroring the calculator's
+  // rule: a data or filter transition reseeds visibility to everything.
+  const selectionKey = `${selectedModel}|${selectedSequence}|${[...selectedPrecisions]
+    .toSorted()
+    .join(',')}|${selectedRunDate}|${[...availableHwKeys].toSorted().join(',')}`;
+
+  const visibleHwKeys = useMemo(
+    () => resolveCalculatorVisibility(visibilityIntent, selectionKey, availableHwKeys),
+    [visibilityIntent, selectionKey, availableHwKeys],
+  );
+  const visibleKeysArray = useMemo(() => [...visibleHwKeys], [visibleHwKeys]);
+  const { resolveColor } = useThemeColors({
+    highContrast,
+    activeKeys: visibleKeysArray,
+  });
+
+  const targetValue = useMemo(
+    () => resolveCalculatorTarget(requestedTargetValue, hasData, ranges.interactivity),
+    [hasData, ranges.interactivity, requestedTargetValue],
+  );
+  const currentRange = ranges.interactivity;
+
+  const handleSliderChange = useCallback((event: React.ChangeEvent<HTMLInputElement>) => {
+    const value = Number(event.target.value);
+    setRequestedTargetValue(value);
+    setInputValue(String(value));
+  }, []);
+
+  const handleInputChange = useCallback((e: React.ChangeEvent<HTMLInputElement>) => {
+    setInputValue(e.target.value);
+    const parsed = parseFloat(e.target.value);
+    if (!isNaN(parsed) && parsed >= 0) {
+      setRequestedTargetValue(parsed);
+    }
+  }, []);
+
+  const handleInputFocus = useCallback(() => {
+    setInputValue(String(targetValue));
+    setIsTargetInputFocused(true);
+  }, [targetValue]);
+
+  const handleInputBlur = useCallback(() => {
+    const parsed = parseFloat(inputValue);
+    if (isNaN(parsed) || parsed < 0) {
+      setInputValue(String(targetValue));
+    } else {
+      const { min, max } = ranges.interactivity;
+      const clamped = Math.max(min, Math.min(max, parsed));
+      setRequestedTargetValue(clamped);
+      setInputValue(String(clamped));
+    }
+    setIsTargetInputFocused(false);
+    track('fleet_target_set', { mode, value: targetValue });
+  }, [inputValue, targetValue, mode, ranges.interactivity]);
+
+  const handleModelChange = useCallback(
+    (value: string) => {
+      setVisibilityIntent(null);
+      setSelectedModel(value as Model);
+      track('fleet_model_selected', { model: value });
+    },
+    [setSelectedModel],
+  );
+
+  const handleSequenceChange = useCallback(
+    (value: string) => {
+      setVisibilityIntent(null);
+      setSelectedSequence(value as Sequence);
+      track('fleet_sequence_selected', { sequence: value });
+    },
+    [setSelectedSequence],
+  );
+
+  const handlePrecisionChange = useCallback(
+    (value: string[]) => {
+      setVisibilityIntent(null);
+      setSelectedPrecisions(value);
+      track('fleet_precision_selected', { precision: value.join(',') });
+    },
+    [setSelectedPrecisions],
+  );
+
+  const handlePercentileChange = useCallback(
+    (value: Percentile) => {
+      setSelectedPercentile(value);
+      setUrlParam('i_pctl', value);
+      track('fleet_percentile_selected', { percentile: value });
+    },
+    [setUrlParam],
+  );
+
+  const toggleGpuVisibility = useCallback(
+    (hwKey: string) => {
+      const visibleLegendKeys = availableHwKeys.filter((key) => visibleHwKeys.has(key));
+      const allVisible = visibleLegendKeys.length === availableHwKeys.length;
+      const isVisible = visibleHwKeys.has(hwKey);
+      let next: Set<string>;
+      if (isVisible && allVisible) {
+        next = new Set([hwKey]);
+      } else if (isVisible && visibleLegendKeys.length === 1) {
+        next = new Set(availableHwKeys);
+      } else {
+        next = new Set(visibleHwKeys);
+        if (isVisible) next.delete(hwKey);
+        else next.add(hwKey);
+      }
+      setVisibilityIntent({
+        scopeKey: selectionKey,
+        visible: next,
+        known: new Set(availableHwKeys),
+      });
+      track('fleet_gpu_toggled', { gpu: hwKey });
+    },
+    [availableHwKeys, visibleHwKeys, selectionKey],
+  );
+
+  const handleResetGpus = useCallback(() => {
+    setVisibilityIntent({
+      scopeKey: selectionKey,
+      visible: new Set(availableHwKeys),
+      known: new Set(availableHwKeys),
+    });
+    track('fleet_gpu_reset', { gpuCount: availableHwKeys.length });
+  }, [availableHwKeys, selectionKey]);
+
+  // The legend filters which chips are candidates for the lifecycle lines,
+  // exactly as the calculator's chart legend did when the section lived there.
+  // It renders in the controls card rather than inside the chart so it stays
+  // reachable from the table view and — critically — after isolating a config
+  // that is unplottable at the current target, when the chart itself unmounts.
+  const legendItems = useMemo(
+    () =>
+      Object.entries(hardwareConfig)
+        .filter(([key]) => availableHwKeys.includes(key))
+        .toSorted(([a], [b]) => getModelSortIndex(a) - getModelSortIndex(b) || a.localeCompare(b))
+        .map(([key, config]) => ({
+          name: config.name,
+          label: getDisplayLabel(config),
+          color: resolveColor(key),
+          title: config.gpu,
+          hw: key,
+          isActive: visibleHwKeys.has(key),
+          onClick: () => toggleGpuVisibility(key),
+        })),
+    [hardwareConfig, availableHwKeys, visibleHwKeys, resolveColor, toggleGpuVisibility],
+  );
+
+  const costTypeLabels: Record<CostType, string> = useMemo(
+    () =>
+      locale === 'zh'
+        ? COST_TYPE_LABELS_ZH
+        : { total: 'Total Tokens', input: 'Input Tokens', output: 'Output Tokens' },
+    [locale],
+  );
+
+  if (!loading && error) {
+    console.error(error);
+    return (
+      <Card>
+        <div className="flex items-center justify-center h-64 text-muted-foreground">
+          {t.errorLoading}
+        </div>
+      </Card>
+    );
+  }
+
+  return (
+    <div className="flex flex-col gap-4">
+      <section data-testid="fleet-controls">
+        <Card className="relative z-30">
+          <div className="flex flex-col gap-4">
+            <div className="flex items-start justify-between">
+              <div>
+                <h2 className="text-lg font-semibold mb-2">{t.title}</h2>
+                <p className="text-muted-foreground text-sm mb-4">{t.description}</p>
+              </div>
+              <ChartShareActions />
+            </div>
+
+            <TooltipProvider delayDuration={0}>
+              <div
+                className={`grid grid-cols-1 md:grid-cols-2 gap-4 ${
+                  isAgenticSequence ? 'lg:grid-cols-6' : 'lg:grid-cols-5'
+                }`}
+              >
+                <ModelSelector
+                  id="fleet-model"
+                  data-testid="fleet-model-selector"
+                  value={selectedModel}
+                  onChange={handleModelChange}
+                  open={openDropdown === 'model'}
+                  onOpenChange={handleDropdownOpenChange('model')}
+                  availableModels={availableModels}
+                />
+                <ScenarioSelector
+                  id="fleet-sequence"
+                  data-testid="fleet-sequence-selector"
+                  value={selectedSequence}
+                  onChange={handleSequenceChange}
+                  open={openDropdown === 'sequence'}
+                  onOpenChange={handleDropdownOpenChange('sequence')}
+                  availableSequences={availableSequences}
+                />
+                {isAgenticSequence && featureGateUnlocked && (
+                  <PercentileSelector
+                    id="fleet-percentile"
+                    data-testid="fleet-percentile-selector"
+                    value={selectedPercentile}
+                    onChange={handlePercentileChange}
+                  />
+                )}
+                <PrecisionSelector
+                  id="fleet-precision"
+                  data-testid="fleet-precision-selector"
+                  value={selectedPrecisions}
+                  onChange={handlePrecisionChange}
+                  open={openDropdown === 'precision'}
+                  onOpenChange={handleDropdownOpenChange('precision')}
+                  availablePrecisions={availablePrecisions}
+                />
+
+                <div className="flex flex-col space-y-1.5 lg:col-span-1">
+                  <LabelWithTooltip
+                    htmlFor="fleet-cost"
+                    label={t.costProviderLabel}
+                    tooltip={t.costProviderTooltip}
+                  />
+                  <div id="fleet-cost" data-testid="fleet-cost-selector">
+                    <MultiSelect
+                      options={COST_PROVIDER_OPTIONS.map((provider) => ({
+                        value: provider.value,
+                        label: locale === 'zh' ? provider.labelZh : provider.label,
+                      }))}
+                      value={[costProvider]}
+                      onChange={(values) => {
+                        const next = values[0];
+                        if (!next) return;
+                        setCostProvider(next as CostProvider);
+                        track('fleet_cost_provider_changed', { provider: next });
+                      }}
+                      open={openDropdown === 'costProvider'}
+                      onOpenChange={handleDropdownOpenChange('costProvider')}
+                      placeholder={t.costProviderPlaceholder}
+                      minSelections={1}
+                      maxSelections={1}
+                      showClearAll={false}
+                      searchable={false}
+                      plainSelectedText
+                      showSelectionSummary={false}
+                    />
+                  </div>
+                </div>
+
+                <div className="flex flex-col space-y-1.5 lg:col-span-1">
+                  <LabelWithTooltip
+                    htmlFor="fleet-cost-type"
+                    label={t.tokenTypeLabel}
+                    tooltip={t.tokenTypeTooltip}
+                  />
+                  <div id="fleet-cost-type" data-testid="fleet-cost-type-selector">
+                    <MultiSelect
+                      options={COST_TYPE_OPTIONS.map((ct) => ({
+                        value: ct.value,
+                        label: costTypeLabels[ct.value],
+                      }))}
+                      value={[costType]}
+                      onChange={(values) => {
+                        const next = values[0];
+                        if (!next) return;
+                        setCostType(next as CostType);
+                        track('fleet_cost_type_changed', { costType: next });
+                      }}
+                      open={openDropdown === 'costType'}
+                      onOpenChange={handleDropdownOpenChange('costType')}
+                      placeholder={t.tokenTypePlaceholder}
+                      minSelections={1}
+                      maxSelections={1}
+                      showClearAll={false}
+                      searchable={false}
+                      plainSelectedText
+                      showSelectionSummary={false}
+                    />
+                  </div>
+                </div>
+              </div>
+
+              {/* Target value slider + input */}
+              {!loading && hasData && (
+                <div className="space-y-2">
+                  <LabelWithTooltip
+                    htmlFor="fleet-target"
+                    label={
+                      isAgenticSequence ? t.targetAgenticLabel(percentileLabel) : t.targetLabel
+                    }
+                    tooltip={
+                      isAgenticSequence ? t.targetAgenticTooltip(percentileLabel) : t.targetTooltip
+                    }
+                  />
+                  <div className="flex items-center gap-4">
+                    <div className="flex-1">
+                      <input
+                        id="fleet-target"
+                        type="range"
+                        min={currentRange.min}
+                        max={currentRange.max}
+                        step={1}
+                        value={targetValue}
+                        onChange={handleSliderChange}
+                        onPointerUp={() =>
+                          track('fleet_target_slider_set', { mode, value: targetValue })
+                        }
+                        className="w-full h-2 appearance-none rounded-full bg-secondary cursor-pointer
+                        [&::-webkit-slider-thumb]:appearance-none [&::-webkit-slider-thumb]:w-4
+                        [&::-webkit-slider-thumb]:h-4 [&::-webkit-slider-thumb]:rounded-full
+                        [&::-webkit-slider-thumb]:bg-primary [&::-webkit-slider-thumb]:cursor-pointer
+                        [&::-moz-range-thumb]:w-4 [&::-moz-range-thumb]:h-4
+                        [&::-moz-range-thumb]:rounded-full [&::-moz-range-thumb]:bg-primary
+                        [&::-moz-range-thumb]:cursor-pointer [&::-moz-range-thumb]:border-0"
+                      />
+                      <div
+                        className="relative h-4 text-xs text-muted-foreground"
+                        style={{ marginLeft: 8, marginRight: 8 }}
+                      >
+                        {Array.from({ length: 6 }, (_, i) => (
+                          <span
+                            key={i}
+                            className="absolute -translate-x-1/2"
+                            style={{ left: `${(i / 5) * 100}%` }}
+                          >
+                            {Math.round(
+                              currentRange.min + (currentRange.max - currentRange.min) * (i / 5),
+                            )}
+                          </span>
+                        ))}
+                      </div>
+                    </div>
+                    <Input
+                      type="number"
+                      value={resolveCalculatorTargetInputValue(
+                        inputValue,
+                        targetValue,
+                        isTargetInputFocused,
+                      )}
+                      onFocus={handleInputFocus}
+                      onChange={handleInputChange}
+                      onBlur={handleInputBlur}
+                      className="w-24 h-9"
+                      min={0}
+                    />
+                  </div>
+                </div>
+              )}
+            </TooltipProvider>
+
+            {!loading && legendItems.length > 0 && (
+              <div
+                data-testid="fleet-legend"
+                className="flex flex-wrap items-center gap-x-6 gap-y-2 border-t border-border pt-3"
+              >
+                <ul className="flex flex-wrap items-center gap-x-4 gap-y-1">
+                  {legendItems.map((item) => (
+                    <ChartLegendItem key={item.hw} {...item} sidebarMode />
+                  ))}
+                </ul>
+                {visibleHwKeys.size < availableHwKeys.length && (
+                  <button
+                    type="button"
+                    data-testid="fleet-reset-filter"
+                    onClick={handleResetGpus}
+                    className="text-xs text-muted-foreground hover:text-foreground underline cursor-pointer"
+                  >
+                    {t.resetFilter}
+                  </button>
+                )}
+                <div className="flex items-center gap-2">
+                  <Switch
+                    id="fleet-high-contrast"
+                    data-testid="fleet-high-contrast"
+                    checked={highContrast}
+                    onCheckedChange={(checked: boolean) => {
+                      setHighContrast(checked);
+                      track('fleet_high_contrast_toggled', { enabled: checked });
+                    }}
+                  />
+                  <Label
+                    htmlFor="fleet-high-contrast"
+                    className="text-xs text-muted-foreground hover:text-foreground cursor-pointer"
+                  >
+                    {t.highContrast}
+                  </Label>
+                </div>
+              </div>
+            )}
+          </div>
+        </Card>
+      </section>
+
+      {loading && (
+        <Card>
+          <Skeleton className="h-64 w-full" />
+        </Card>
+      )}
+
+      {!loading && hasData && (
+        <FleetLifecycle
+          hardwareConfig={hardwareConfig}
+          costProvider={costProvider}
+          costType={costType}
+          targetValue={targetValue}
+          mode={mode}
+          visibleHwKeys={visibleHwKeys}
+          selectedModel={selectedModel}
+          selectedSequence={selectedSequence}
+          selectedPrecisions={selectedPrecisions}
+          selectedPercentile={selectedPercentile}
+          mwInput={mwInput}
+          onMwInputChange={handleMwInputChange}
+          colorResolver={resolveColor}
+        />
+      )}
+    </div>
+  );
+}
diff --git a/packages/app/src/components/calculator/FleetPlanner.tsx b/packages/app/src/components/calculator/FleetPlanner.tsx
deleted file mode 100644
index 431b5e83d..000000000
--- a/packages/app/src/components/calculator/FleetPlanner.tsx
+++ /dev/null
@@ -1,505 +0,0 @@
-'use client';
-
-import Link from 'next/link';
-import { useCallback, useMemo, useState } from 'react';
-
-import { track } from '@/lib/analytics';
-import type { HardwareConfig } from '@/components/inference/types';
-import { Card } from '@/components/ui/card';
-import { type DataTableColumn, DataTable } from '@/components/ui/data-table';
-import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
-import { Input } from '@/components/ui/input';
-import { LabelWithTooltip } from '@/components/ui/label-with-tooltip';
-import { TooltipProvider } from '@/components/ui/tooltip';
-import { getGpuSpecs, getHardwareConfig } from '@/lib/constants';
-import { readUrlParams, writeUrlParams } from '@/lib/url-state';
-import { getDisplayLabel } from '@/lib/utils';
-import { useLocale } from '@/lib/use-locale';
-
-import { computeFleetStats, formatCompact, type FleetStats } from './fleet';
-import { interpolateForGPU, maxInteractivityAtCost } from './interpolation';
-import { getCostProviderLabel, getThroughputForType } from './ThroughputBarChart';
-import type { CostProvider, CostType, GPUDataPoint, InterpolatedResult } from './types';
-
-interface FleetPlannerProps {
-  results: InterpolatedResult[];
-  gpuDataByGroupKey: Record<string, GPUDataPoint[]>;
-  hardwareConfig: HardwareConfig;
-  costProvider: CostProvider;
-  costType: CostType;
-  /** Current target interactivity (tok/s/user) the results were interpolated at. */
-  targetValue: number;
-  /** Legend visibility by base hwKey — the cost-cap card must not depend on
-   * `results`, which is filtered at the current slider target. */
-  visibleHwKeys: Set<string>;
-}
-
-const STRINGS = {
-  en: {
-    fleetTitle: 'Fleet Projection',
-    fleetDescription:
-      'Size a deployment by facility power: how many chips fit in your megawatt budget, and what they serve at the target interactivity above.',
-    mwLabel: 'Facility Power (MW)',
-    mwTooltip:
-      'Total facility power budget in megawatts. Chip count uses all-in power per chip (host, networking, cooling) from the SemiAnalysis Datacenter Industry Model — not bare TDP.',
-    mwPlaceholder: 'e.g. 10',
-    colGpu: 'Chip',
-    colGpus: 'Chips',
-    colFleetTput: (tokenType: string) => `Fleet ${tokenType}tok/s`,
-    colUsers: 'Concurrent Users',
-    colCostHr: 'Fleet $/hr',
-    colCostMo: 'Fleet $/mo',
-    fleetEmpty: 'Enter a facility power budget to project fleet capacity and cost.',
-    fleetTooSmall:
-      'This power budget is too small to power a single chip of the shown hardware — try a larger value.',
-    fleetNoGpus: 'No chips are visible to project — enable hardware in the chart legend.',
-    costCapTitle: 'Interactivity Within a Cost Target',
-    costCapDescription:
-      'Set a cost ceiling per million tokens and find the highest interactivity each chip can serve without exceeding it.',
-    costCapLabel: 'Cost Target ($/M tok)',
-    costCapTooltip:
-      'Maximum acceptable cost per million tokens (at the selected pricing tier and token type). The answer is the highest interactivity whose interpolated cost stays at or below this ceiling.',
-    costCapPlaceholder: 'e.g. 0.50',
-    colMaxInteractivity: 'Max Interactivity (tok/s/user)',
-    colTputAtIv: 'Throughput (tok/s/chip)',
-    notReachable: 'Not reachable',
-    costCapEmpty: 'Enter a cost target to find the serveable interactivity per chip.',
-    costCapNoGpus: 'No chips are visible to evaluate — enable hardware in the chart legend.',
-    note: 'Note:',
-    disaggFleet:
-      ' Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) report throughput per decode chip or per prefill chip, rather than per total chip count — fleet sizes and costs for those configs are not an apples-to-apples comparison with aggregated configs.',
-    assumptions: (tier: string) =>
-      `Assumes 100% utilization at this operating point and owned-datacenter economics: fleet cost = chips × ${tier} $/chip/hr, months = 730 hr. Facility power is all-in per-chip power (host, networking, cooling), not bare TDP.`,
-    source: 'Source: ',
-    tokenTypeTotal: '',
-    tokenTypeInput: 'input ',
-    tokenTypeOutput: 'output ',
-  },
-  zh: {
-    fleetTitle: '集群规模测算',
-    fleetDescription:
-      '按设施功率规划部署：在给定兆瓦预算内可容纳多少 Chip，以及在上方目标交互性下的服务能力。',
-    mwLabel: '设施功率 (MW)',
-    mwTooltip:
-      '设施总功率预算（兆瓦）。Chip 数量按每 Chip 全含功率（主机、网络、散热）计算，数据来自 SemiAnalysis Datacenter Industry Model，而非裸 TDP。',
-    mwPlaceholder: '如 10',
-    colGpu: 'Chip',
-    colGpus: 'Chip 数',
-    colFleetTput: (tokenType: string) => `集群${tokenType} tok/s`,
-    colUsers: '并发用户数',
-    colCostHr: '集群 $/hr',
-    colCostMo: '集群 $/mo',
-    fleetEmpty: '输入设施功率预算以测算集群容量与成本。',
-    fleetTooSmall: '该功率预算不足以为所示任一 Chip 供电——请尝试更大的数值。',
-    fleetNoGpus: '当前无可见 Chip 可测算——请在图表图例中启用硬件。',
-    costCapTitle: '成本上限下的交互性',
-    costCapDescription:
-      '设定每百万 token 的成本上限，查看每款 Chip 在不超支前提下可提供的最高交互性。',
-    costCapLabel: '成本上限 ($/M tok)',
-    costCapTooltip:
-      '每百万 token 的最高可接受成本（按所选定价层级和 token 类型）。结果为插值成本不超过该上限的最高交互性。',
-    costCapPlaceholder: '如 0.50',
-    colMaxInteractivity: '最高交互性 (tok/s/user)',
-    colTputAtIv: '吞吐量 (tok/s/chip)',
-    notReachable: '无法达到',
-    costCapEmpty: '输入成本上限以查看每款 Chip 可提供的交互性。',
-    costCapNoGpus: '当前无可见 Chip 可评估——请在图表图例中启用硬件。',
-    note: '注意：',
-    disaggFleet:
-      '解耦推理配置（如 MoRI SGLang、Dynamo TRTLLM）按解码 Chip 或预填充 Chip 报告吞吐量，而非按 Chip 总数——这类配置的集群规模与成本和聚合配置并非同类比较。',
-    assumptions: (tier: string) =>
-      `假设该操作点下 100% 利用率及自有数据中心经济模型：集群成本 = Chip 数 × ${tier} $/chip/hr，每月按 730 小时计。设施功率为每 Chip 全含功率（主机、网络、散热），非裸 TDP。`,
-    source: '来源：',
-    tokenTypeTotal: '总',
-    tokenTypeInput: '输入',
-    tokenTypeOutput: '输出',
-  },
-} as const;
-
-function getLabel(
-  r: { hwKey: string; precision?: string },
-  hardwareConfig: HardwareConfig,
-): string {
-  const config = hardwareConfig[r.hwKey] || getHardwareConfig(r.hwKey);
-  const baseName = config ? getDisplayLabel(config) : r.hwKey;
-  return r.precision ? `${baseName} (${r.precision.toUpperCase()})` : baseName;
-}
-
-function parsePositive(raw: string): number | null {
-  const parsed = parseFloat(raw);
-  return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
-}
-
-interface FleetRow {
-  result: InterpolatedResult;
-  stats: FleetStats;
-}
-
-interface CostCapRow {
-  resultKey: string;
-  hwKey: string;
-  precision?: string;
-  /** null = cost target not reachable anywhere on this GPU's frontier */
-  maxInteractivity: number | null;
-  tputAtIv: number | null;
-  users: number | null;
-}
-
-export default function FleetPlanner({
-  results,
-  gpuDataByGroupKey,
-  hardwareConfig,
-  costProvider,
-  costType,
-  targetValue,
-  visibleHwKeys,
-}: FleetPlannerProps) {
-  const locale = useLocale();
-  const t = STRINGS[locale];
-
-  const [mwInput, setMwInput] = useState<string>(() => readUrlParams().c_mw ?? '');
-  const [costCapInput, setCostCapInput] = useState<string>(() => readUrlParams().c_costcap ?? '');
-
-  const mw = useMemo(() => parsePositive(mwInput), [mwInput]);
-  const costCap = useMemo(() => parsePositive(costCapInput), [costCapInput]);
-
-  const handleMwChange = useCallback((e: React.ChangeEvent<HTMLInputElement>) => {
-    setMwInput(e.target.value);
-    writeUrlParams({ c_mw: parsePositive(e.target.value) ? e.target.value : '' });
-  }, []);
-
-  const handleMwBlur = useCallback(() => {
-    track('calculator_fleet_mw_set', { mw: mwInput });
-  }, [mwInput]);
-
-  const handleCostCapChange = useCallback((e: React.ChangeEvent<HTMLInputElement>) => {
-    setCostCapInput(e.target.value);
-    writeUrlParams({ c_costcap: parsePositive(e.target.value) ? e.target.value : '' });
-  }, []);
-
-  const handleCostCapBlur = useCallback(() => {
-    track('calculator_cost_target_set', { costCap: costCapInput });
-  }, [costCapInput]);
-
-  // ---- Fleet projection rows (at the calculator's current target interactivity) ----
-  const fleetRows = useMemo<FleetRow[]>(() => {
-    if (!mw) return [];
-    const rows: FleetRow[] = [];
-    for (const result of results) {
-      const specs = getGpuSpecs(result.hwKey);
-      const stats = computeFleetStats({
-        mw,
-        powerKwPerGpu: specs.power,
-        costPerGpuHour: specs[costProvider],
-        tputPerGpu: getThroughputForType(result, costType),
-        outputTputPerGpu: result.outputTputValue,
-        interactivity: targetValue,
-      });
-      if (stats) rows.push({ result, stats });
-    }
-    return rows;
-  }, [mw, results, costProvider, costType, targetValue]);
-
-  // Groups shown in the target-independent card: gated ONLY by the legend's
-  // hw visibility — never by `results`, which is filtered at the current
-  // slider target (a GPU with zero interpolated throughput at an extreme
-  // target must still appear here, since its frontier is unchanged).
-  const visibleGroupKeys = useMemo(
-    () =>
-      new Set(
-        Object.keys(gpuDataByGroupKey).filter((groupKey) =>
-          visibleHwKeys.has(groupKey.includes('__') ? groupKey.split('__')[0] : groupKey),
-        ),
-      ),
-    [gpuDataByGroupKey, visibleHwKeys],
-  );
-
-  // Any visible group containing a disagg config taints both tables' per-GPU
-  // numbers (fleet sizing AND cost-cap interactivity), so check the raw points
-  // rather than only the fleet rows' bracketing frontier points.
-  const hasDisagg = useMemo(
-    () =>
-      Object.entries(gpuDataByGroupKey).some(
-        ([groupKey, points]) => visibleGroupKeys.has(groupKey) && points.some((p) => p.disagg),
-      ),
-    [visibleGroupKeys, gpuDataByGroupKey],
-  );
-
-  // ---- Cost-cap rows (independent of target interactivity) ----
-  const costCapRows = useMemo<CostCapRow[]>(() => {
-    if (!costCap) return [];
-    const rows: CostCapRow[] = [];
-    for (const [groupKey, points] of Object.entries(gpuDataByGroupKey)) {
-      if (!visibleGroupKeys.has(groupKey)) continue;
-      const hwKey = groupKey.includes('__') ? groupKey.split('__')[0] : groupKey;
-      const precision = groupKey.includes('__') ? groupKey.split('__')[1] : undefined;
-
-      const maxIv = maxInteractivityAtCost(points, costCap, costProvider, costType);
-      if (maxIv === null) {
-        rows.push({
-          resultKey: groupKey,
-          hwKey,
-          precision,
-          maxInteractivity: null,
-          tputAtIv: null,
-          users: null,
-        });
-        continue;
-      }
-
-      const atIv = interpolateForGPU(points, maxIv, 'interactivity_to_throughput', costProvider);
-      const specs = getGpuSpecs(hwKey);
-      const stats =
-        atIv && mw
-          ? computeFleetStats({
-              mw,
-              powerKwPerGpu: specs.power,
-              costPerGpuHour: specs[costProvider],
-              tputPerGpu: getThroughputForType(atIv, costType),
-              outputTputPerGpu: atIv.outputTputValue,
-              interactivity: maxIv,
-            })
-          : null;
-
-      rows.push({
-        resultKey: groupKey,
-        hwKey,
-        precision,
-        maxInteractivity: maxIv,
-        tputAtIv: atIv ? getThroughputForType(atIv, costType) : null,
-        users: stats ? stats.concurrentUsers : null,
-      });
-    }
-    // Highest achievable interactivity first; unreachable rows last.
-    return rows.toSorted(
-      (a, b) => (b.maxInteractivity ?? -Infinity) - (a.maxInteractivity ?? -Infinity),
-    );
-  }, [costCap, visibleGroupKeys, gpuDataByGroupKey, costProvider, costType, mw]);
-
-  const tokenTypeLabel =
-    costType === 'input'
-      ? t.tokenTypeInput
-      : costType === 'output'
-        ? t.tokenTypeOutput
-        : t.tokenTypeTotal;
-
-  const fleetColumns = useMemo<DataTableColumn<FleetRow>[]>(
-    () => [
-      {
-        header: t.colGpu,
-        cell: (r) => getLabel(r.result, hardwareConfig),
-        sortValue: (r) => getLabel(r.result, hardwareConfig),
-        className: 'font-medium whitespace-nowrap',
-      },
-      {
-        header: t.colGpus,
-        align: 'right',
-        cell: (r) => r.stats.gpus.toLocaleString(),
-        sortValue: (r) => r.stats.gpus,
-        className: 'tabular-nums',
-      },
-      {
-        header: t.colFleetTput(tokenTypeLabel),
-        align: 'right',
-        cell: (r) => formatCompact(r.stats.fleetTokPerSec),
-        sortValue: (r) => r.stats.fleetTokPerSec,
-        className: 'tabular-nums',
-      },
-      {
-        header: t.colUsers,
-        align: 'right',
-        cell: (r) => formatCompact(r.stats.concurrentUsers),
-        sortValue: (r) => r.stats.concurrentUsers,
-        className: 'tabular-nums',
-      },
-      {
-        header: t.colCostHr,
-        align: 'right',
-        cell: (r) => `$${formatCompact(r.stats.costPerHour)}`,
-        sortValue: (r) => r.stats.costPerHour,
-        className: 'tabular-nums',
-      },
-      {
-        header: t.colCostMo,
-        align: 'right',
-        cell: (r) => `$${formatCompact(r.stats.costPerMonth)}`,
-        sortValue: (r) => r.stats.costPerMonth,
-        className: 'tabular-nums',
-      },
-    ],
-    [hardwareConfig, t, tokenTypeLabel],
-  );
-
-  const costCapColumns = useMemo<DataTableColumn<CostCapRow>[]>(() => {
-    const columns: DataTableColumn<CostCapRow>[] = [
-      {
-        header: t.colGpu,
-        cell: (r) => getLabel(r, hardwareConfig),
-        sortValue: (r) => getLabel(r, hardwareConfig),
-        className: 'font-medium whitespace-nowrap',
-      },
-      {
-        header: t.colMaxInteractivity,
-        align: 'right',
-        cell: (r) =>
-          r.maxInteractivity === null ? (
-            <span className="text-muted-foreground">{t.notReachable}</span>
-          ) : (
-            r.maxInteractivity.toFixed(1)
-          ),
-        sortValue: (r) => r.maxInteractivity ?? -Infinity,
-        className: 'tabular-nums',
-      },
-      {
-        header: t.colTputAtIv,
-        align: 'right',
-        cell: (r) => (r.tputAtIv === null ? '—' : r.tputAtIv.toFixed(1)),
-        sortValue: (r) => r.tputAtIv ?? -Infinity,
-        className: 'tabular-nums',
-      },
-    ];
-    if (mw) {
-      columns.push({
-        header: t.colUsers,
-        align: 'right',
-        cell: (r) => (r.users === null ? '—' : formatCompact(r.users)),
-        sortValue: (r) => r.users ?? -Infinity,
-        className: 'tabular-nums',
-      });
-    }
-    return columns;
-  }, [hardwareConfig, t, mw]);
-
-  const disaggBanner = hasDisagg && (
-    <p className="text-muted-foreground text-xs border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1">
-      <strong>{t.note}</strong>
-      {t.disaggFleet}
-    </p>
-  );
-
-  const assumptionsFooter = (
-    <>
-      <p className="text-xs text-muted-foreground mt-3">
-        {t.assumptions(getCostProviderLabel(costProvider))}
-      </p>
-      <p className="text-muted-foreground mt-1">
-        <small>
-          {t.source}
-          <Link
-            target="_blank"
-            className="underline hover:text-foreground"
-            href="https://semianalysis.com/datacenter-industry-model/"
-          >
-            SemiAnalysis Datacenter Industry Model
-            <ExternalLinkIcon />
-          </Link>
-          {' & '}
-          <Link
-            target="_blank"
-            className="underline hover:text-foreground"
-            href="https://semianalysis.com/ai-cloud-tco-model/"
-          >
-            SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model
-            <ExternalLinkIcon />
-          </Link>
-        </small>
-      </p>
-    </>
-  );
-
-  return (
-    <TooltipProvider delayDuration={0}>
-      <section data-testid="calculator-fleet-section">
-        <Card>
-          <div className="flex flex-col gap-4">
-            <div>
-              <h2 className="text-lg font-semibold mb-2">{t.fleetTitle}</h2>
-              <p className="text-muted-foreground text-sm">{t.fleetDescription}</p>
-            </div>
-            <div className="flex flex-col space-y-1.5 max-w-48">
-              <LabelWithTooltip htmlFor="calc-fleet-mw" label={t.mwLabel} tooltip={t.mwTooltip} />
-              <Input
-                id="calc-fleet-mw"
-                data-testid="calc-fleet-mw-input"
-                type="number"
-                min={0}
-                step="any"
-                placeholder={t.mwPlaceholder}
-                value={mwInput}
-                onChange={handleMwChange}
-                onBlur={handleMwBlur}
-                className="w-32 h-9"
-              />
-            </div>
-            {mw && fleetRows.length > 0 ? (
-              <>
-                <DataTable
-                  data={fleetRows}
-                  columns={fleetColumns}
-                  testId="calculator-fleet-table"
-                  analyticsPrefix="calculator_fleet_table"
-                />
-                {disaggBanner}
-                {assumptionsFooter}
-              </>
-            ) : (
-              <p className="text-sm text-muted-foreground" data-testid="calculator-fleet-empty">
-                {/* Empty with MW set has two causes: nothing visible to project
-                    (legend/slider filtered everything) vs. a budget below one GPU. */}
-                {mw ? (results.length > 0 ? t.fleetTooSmall : t.fleetNoGpus) : t.fleetEmpty}
-              </p>
-            )}
-          </div>
-        </Card>
-      </section>
-
-      <section data-testid="calculator-costcap-section">
-        <Card>
-          <div className="flex flex-col gap-4">
-            <div>
-              <h2 className="text-lg font-semibold mb-2">{t.costCapTitle}</h2>
-              <p className="text-muted-foreground text-sm">{t.costCapDescription}</p>
-            </div>
-            <div className="flex flex-col space-y-1.5 max-w-48">
-              <LabelWithTooltip
-                htmlFor="calc-costcap"
-                label={t.costCapLabel}
-                tooltip={t.costCapTooltip}
-              />
-              <Input
-                id="calc-costcap"
-                data-testid="calc-costcap-input"
-                type="number"
-                min={0}
-                step="any"
-                placeholder={t.costCapPlaceholder}
-                value={costCapInput}
-                onChange={handleCostCapChange}
-                onBlur={handleCostCapBlur}
-                className="w-32 h-9"
-              />
-            </div>
-            {costCap && costCapRows.length > 0 ? (
-              <>
-                <DataTable
-                  data={costCapRows}
-                  columns={costCapColumns}
-                  testId="calculator-costcap-table"
-                  analyticsPrefix="calculator_costcap_table"
-                />
-                {disaggBanner}
-                {assumptionsFooter}
-              </>
-            ) : (
-              <p className="text-sm text-muted-foreground" data-testid="calculator-costcap-empty">
-                {/* With a valid target set, empty means nothing is legend-visible
-                    (unreachable GPUs still produce rows). */}
-                {costCap ? t.costCapNoGpus : t.costCapEmpty}
-              </p>
-            )}
-          </div>
-        </Card>
-      </section>
-    </TooltipProvider>
-  );
-}
diff --git a/packages/app/src/components/calculator/ThroughputBarChart.test.ts b/packages/app/src/components/calculator/ThroughputBarChart.test.ts
index ace9bbcc6..62b58b41a 100644
--- a/packages/app/src/components/calculator/ThroughputBarChart.test.ts
+++ b/packages/app/src/components/calculator/ThroughputBarChart.test.ts
@@ -14,6 +14,7 @@ import {
   getMetricValue,
   getSortedResults,
   getThroughputForType,
+  getComparableTpPerMwForType,
   getTpPerMwForType,
   getValueLabel,
 } from './ThroughputBarChart';
@@ -623,14 +624,89 @@ describe('generateTooltipHTML overlay treatment', () => {
       false,
       {
         unofficialRun: '非官方运行',
-        branch: '分支',
+        branch: '分支：',
+        dismiss: '点击其他区域关闭',
         viewRun: '查看工作流运行',
         clamped: '超出实测范围——显示最接近的数据点',
+        parallelism: '并行策略：',
+        cost: '成本：',
+        concurrency: '并发数：',
+        precision: '精度：',
+        disaggregated: '分离式：',
+        yes: '是',
       },
     );
     expect(html).toContain('非官方运行');
-    expect(html).toContain('分支');
+    expect(html).toContain('分支：');
     expect(html).toContain('查看工作流运行');
     expect(html).not.toContain('UNOFFICIAL RUN');
   });
 });
+
+/** An interpolated read, aggregated by default: 800 + 100 = 900 = the total. */
+const rankResult = (over: Partial<InterpolatedResult> = {}): InterpolatedResult =>
+  ({
+    hwKey: 'mi355x',
+    resultKey: 'mi355x',
+    value: 900,
+    inputTputValue: 800,
+    outputTputValue: 100,
+    cost: 1,
+    costInput: 1,
+    costOutput: 1,
+    tpPerMw: 9000,
+    inputTpPerMw: 8000,
+    outputTpPerMw: 1000,
+    concurrency: 8,
+    nearestPoints: [],
+    ...over,
+  }) as InterpolatedResult;
+
+describe('getComparableTpPerMwForType', () => {
+  it('leaves the total basis alone — it is already per total chip', () => {
+    const r = rankResult({ inputTokenShare: 0.5 });
+    expect(getComparableTpPerMwForType(r, 'total')).toBe(r.tpPerMw);
+    expect(getComparableTpPerMwForType(r, 'total')).toBe(getTpPerMwForType(r, 'total'));
+  });
+
+  it('is an identity for an aggregated config, whose rates already sum to the total', () => {
+    // 800 + 100 = 900, so the share is 8/9 and tpPerMw x share is inputTpPerMw.
+    const r = rankResult({ inputTokenShare: 800 / 900 });
+    expect(getComparableTpPerMwForType(r, 'input')).toBeCloseTo(8000, 6);
+    expect(getComparableTpPerMwForType(r, 'output')).toBeCloseTo(1000, 6);
+  });
+
+  it('deflates a disaggregated config back onto the per-total-chip basis', () => {
+    // 16 prefill + 8 decode chips serving 8192:1024. The reported rates sum to
+    // 1.667x the total, so ranking on them raw reads high.
+    const r = rankResult({
+      value: 300,
+      inputTputValue: 400,
+      outputTputValue: 100,
+      tpPerMw: 3000,
+      inputTpPerMw: 4000,
+      outputTpPerMw: 1000,
+      inputTokenShare: 8192 / 9216,
+    });
+    expect(getTpPerMwForType(r, 'output')).toBe(1000);
+    // Output is a ninth of the tokens, on 3000 tok/s/MW overall.
+    expect(getComparableTpPerMwForType(r, 'output')).toBeCloseTo(3000 / 9, 6);
+    expect(getComparableTpPerMwForType(r, 'output')).toBeLessThan(getTpPerMwForType(r, 'output'));
+    // And the two token types partition the total rather than exceeding it.
+    expect(
+      getComparableTpPerMwForType(r, 'input') + getComparableTpPerMwForType(r, 'output'),
+    ).toBeCloseTo(3000, 6);
+    expect(getTpPerMwForType(r, 'input') + getTpPerMwForType(r, 'output')).toBeGreaterThan(3000);
+  });
+
+  it('falls back to the raw figure when no share was recovered', () => {
+    const r = rankResult();
+    expect(getComparableTpPerMwForType(r, 'input')).toBe(r.inputTpPerMw);
+    expect(getComparableTpPerMwForType(r, 'output')).toBe(r.outputTpPerMw);
+  });
+
+  it('clamps a share outside [0,1]', () => {
+    expect(getComparableTpPerMwForType(rankResult({ inputTokenShare: 1.4 }), 'output')).toBe(0);
+    expect(getComparableTpPerMwForType(rankResult({ inputTokenShare: -1 }), 'input')).toBe(0);
+  });
+});
diff --git a/packages/app/src/components/calculator/ThroughputBarChart.tsx b/packages/app/src/components/calculator/ThroughputBarChart.tsx
index ddc138cb7..ba8f88fcf 100644
--- a/packages/app/src/components/calculator/ThroughputBarChart.tsx
+++ b/packages/app/src/components/calculator/ThroughputBarChart.tsx
@@ -2,7 +2,7 @@
 
 import { track } from '@/lib/analytics';
 import * as d3 from 'd3';
-import { useEffect, useMemo, useRef } from 'react';
+import { useCallback, useEffect, useMemo, useRef } from 'react';
 import { useLocale } from '@/lib/use-locale';
 
 import type { HardwareConfig } from '@/components/inference/types';
@@ -37,15 +37,29 @@ import type {
 const OVERLAY_STRINGS = {
   en: {
     unofficialRun: 'UNOFFICIAL RUN',
-    branch: 'Branch',
+    branch: 'Branch:',
+    dismiss: 'Click elsewhere to dismiss',
     viewRun: 'View workflow run',
     clamped: 'Outside measured range — showing nearest data point',
+    parallelism: 'Parallelism:',
+    cost: 'Cost:',
+    concurrency: 'Concurrency:',
+    precision: 'Precision:',
+    disaggregated: 'Disaggregated:',
+    yes: 'Yes',
   },
   zh: {
     unofficialRun: '非官方运行',
-    branch: '分支',
+    branch: '分支：',
+    dismiss: '点击其他区域关闭',
     viewRun: '查看工作流运行',
     clamped: '超出实测范围——显示最接近的数据点',
+    parallelism: '并行策略：',
+    cost: '成本：',
+    concurrency: '并发数：',
+    precision: '精度：',
+    disaggregated: '分离式：',
+    yes: '是',
   },
 } as const;
 
@@ -81,6 +95,37 @@ export function getTpPerMwForType(d: InterpolatedResult, costType: CostType): nu
   return d.tpPerMw; // total
 }
 
+/**
+ * The same figure on a per-total-chip basis, so configs can be ranked against
+ * each other.
+ *
+ * `inputTpPerMw` and `outputTpPerMw` are derived from `input_tput_per_gpu` and
+ * `output_tput_per_gpu`, which a disaggregated run reports per *prefill* and per
+ * *decode* chip. Divided by fewer chips, they read high — which is exactly the
+ * caveat the throughput charts carry ("not an apples-to-apples comparison"). That
+ * is fine for a bar chart the reader is told to interpret, and not fine for a
+ * ranking that silently picks a winner: measured against production history, a
+ * raw output ranking hands 5 of 46 chip-winners to a disaggregated config that a
+ * comparable basis rejects, and changes 10 of 46 outright (17 of 46 on input).
+ *
+ * `tpPerMw` is already per chip overall — the total basis needs no correction and
+ * is unchanged by this (46 of 46 winners identical) — so the token-type figures
+ * are recovered by splitting it with the same measured share revenue uses. For an
+ * aggregated config that is an identity: the rates already sum to the total, so
+ * `tpPerMw x share` *is* `inputTpPerMw`.
+ *
+ * Falls back to the raw figure when no share was recovered, which leaves
+ * behaviour exactly as it was rather than ranking on a zero.
+ */
+export function getComparableTpPerMwForType(d: InterpolatedResult, costType: CostType): number {
+  if (costType === 'total') return d.tpPerMw;
+  if (typeof d.inputTokenShare !== 'number' || !Number.isFinite(d.inputTokenShare)) {
+    return getTpPerMwForType(d, costType);
+  }
+  const share = Math.max(0, Math.min(1, d.inputTokenShare));
+  return d.tpPerMw * (costType === 'input' ? share : 1 - share);
+}
+
 export function getMetricValue(
   d: InterpolatedResult,
   barMetric: BarMetric,
@@ -290,7 +335,7 @@ export function generateTooltipHTML(
 
   let parallelismHtml: string;
   if (ep !== null && ep !== undefined && ep > 1 && tp === ep) {
-    parallelismHtml = `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>Parallelism:</strong> ${dpAttn ? 'DEP' : 'TEP'}${tp}</div>`;
+    parallelismHtml = `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${overlayStrings.parallelism}</strong> ${dpAttn ? 'DEP' : 'TEP'}${tp}</div>`;
   } else if (ep !== null && ep !== undefined && ep > 1) {
     parallelismHtml = `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>TP:</strong> ${tp}, <strong>EP:</strong> ${ep}${dpAttn ? ', <strong>DPA:</strong> True' : ''}</div>`;
   } else {
@@ -320,7 +365,7 @@ export function generateTooltipHTML(
 
   const overlayBranchHtml =
     d.isOverlay && d.runLabel
-      ? `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${overlayStrings.branch}:</strong> ${escapeHtml(d.runLabel)}</div>`
+      ? `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${overlayStrings.branch}</strong> ${escapeHtml(d.runLabel)}</div>`
       : '';
   const overlayHeaderHtml = d.isOverlay
     ? `<div style="color: var(--destructive, #ef4444); font-size: 11px; font-weight: 600; margin-bottom: 4px;">${overlayStrings.unofficialRun}</div>${overlayBranchHtml}`
@@ -328,7 +373,7 @@ export function generateTooltipHTML(
 
   return `
     <div style="background: var(--popover); border: 1px solid var(--border); border-radius: 8px; padding: 12px; box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1); max-width: 320px; pointer-events: auto; user-select: ${isPinned ? 'text' : 'none'};">
-      ${isPinned ? '<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">Click elsewhere to dismiss</div>' : ''}
+      ${isPinned ? `<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">${overlayStrings.dismiss}</div>` : ''}
       ${overlayHeaderHtml}
       <div style="color: var(--foreground); font-size: 13px; font-weight: 600; margin-bottom: 8px;">
         ${label}
@@ -338,17 +383,17 @@ export function generateTooltipHTML(
         <strong>${metricName}:</strong> ${metricDisplay}
       </div>
       <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Cost:</strong> $${costValue.toFixed(3)}${costLabel}
+        <strong>${overlayStrings.cost}</strong> $${costValue.toFixed(3)}${costLabel}
       </div>
       <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
         <strong>tok/s/MW:</strong> ${getTpPerMwForType(d, costType).toFixed(0)}
       </div>
       <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Concurrency:</strong> ~${d.concurrency}
+        <strong>${overlayStrings.concurrency}</strong> ~${d.concurrency}
       </div>
-      ${precision ? `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>Precision:</strong> ${escapeHtml(precision.toUpperCase())}</div>` : ''}
+      ${precision ? `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${overlayStrings.precision}</strong> ${escapeHtml(precision.toUpperCase())}</div>` : ''}
       ${parallelismHtml}
-      ${disagg ? '<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>Disaggregated:</strong> Yes</div>' : ''}
+      ${disagg ? `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${overlayStrings.disaggregated}</strong> ${overlayStrings.yes}</div>` : ''}
       ${runLinkHtml}
     </div>
   `;
@@ -424,12 +469,15 @@ export default function ThroughputBarChart({
   // Color resolution: unofficial-run overlay bars take the run's palette color
   // (so they match the banner + legend swatch — see lib/overlay-run-style.ts);
   // official bars prefer the dynamic colorResolver, falling back to static config.
-  const resolveBarColor = (d: InterpolatedResult) =>
-    d.isOverlay
-      ? overlayRunColor(d.runIndex ?? 0)
-      : colorResolver
-        ? colorResolver(d.hwKey)
-        : getColor();
+  const resolveBarColor = useCallback(
+    (datum: InterpolatedResult) =>
+      datum.isOverlay
+        ? overlayRunColor(datum.runIndex ?? 0)
+        : colorResolver
+          ? colorResolver(datum.hwKey)
+          : getColor(),
+    [colorResolver],
+  );
 
   // Stable refs to avoid re-running the D3 effect
   const hoveredBarXRef = useRef(0);
@@ -474,6 +522,60 @@ export default function ThroughputBarChart({
     [sortedResults],
   );
 
+  const dataIdentity = useMemo(
+    () => JSON.stringify(sortedResults.map((datum) => datum.resultKey).toSorted()),
+    [sortedResults],
+  );
+  const metricIdentity = useMemo(
+    () =>
+      JSON.stringify({
+        bars: sortedResults.map((datum) =>
+          JSON.stringify([
+            datum.resultKey,
+            getMetricValue(datum, barMetric, costType),
+            getValueLabel(datum, barMetric, mode, costType),
+            getCostForType(datum, costType),
+            getThroughputForType(datum, costType),
+            getLabel(datum, hardwareConfig),
+          ]),
+        ),
+        barMetric,
+        costType,
+        mode,
+        maxBarValue,
+        yDomain,
+      }),
+    [barMetric, costType, hardwareConfig, maxBarValue, mode, sortedResults, yDomain],
+  );
+  const paletteIdentity = useMemo(
+    () =>
+      JSON.stringify(
+        sortedResults
+          .map((datum) => JSON.stringify([datum.resultKey, resolveBarColor(datum)]))
+          .toSorted(),
+      ),
+    [resolveBarColor, sortedResults],
+  );
+  const xScaleConfig = useMemo(
+    () => ({ type: 'linear' as const, domain: [0, maxBarValue] as [number, number] }),
+    [maxBarValue],
+  );
+  const yScaleConfig = useMemo(
+    () => ({ type: 'band' as const, domain: yDomain, padding: 0.3 }),
+    [yDomain],
+  );
+  const zoomConfig = useMemo(
+    () => ({
+      enabled: true,
+      axes: 'x' as const,
+      scaleExtent: [0.1, 1] as [number, number],
+      rescaleX: (xScale: ContinuousScale, transform: d3.ZoomTransform) =>
+        xScale.copy().domain([0, maxBarValue / transform.k]) as ContinuousScale,
+      customTransformStorage: (transform: d3.ZoomTransform) => d3.zoomIdentity.scale(transform.k),
+    }),
+    [maxBarValue],
+  );
+
   // ── Layers ──
 
   const layers = useMemo(() => {
@@ -495,6 +597,7 @@ export default function ThroughputBarChart({
     const labelLayer: CustomLayerConfig = {
       type: 'custom',
       key: 'bar-labels',
+      displayIdentity: paletteIdentity,
       render: (zoomGroup, ctx) => {
         const xScale = ctx.xScale as d3.ScaleLinear<number, number>;
         const yScale = ctx.yScale as d3.ScaleBand<string>;
@@ -539,6 +642,20 @@ export default function ThroughputBarChart({
         const barColor = (d: InterpolatedResult) => resolveBarColor(d);
         positionLabelPairs(zoomGroup, xScale, ctx.width, barMetric, costType, barColor);
       },
+      onDisplayUpdate: (zoomGroup, ctx) => {
+        const baseXScale = ctx.xScale as d3.ScaleLinear<number, number>;
+        const svgNode = ctx.layout.svg.node();
+        const transform = svgNode ? d3.zoomTransform(svgNode) : d3.zoomIdentity;
+        const currentXScale = baseXScale.copy().domain([0, maxBarValue / transform.k]);
+        positionLabelPairs(
+          zoomGroup,
+          currentXScale,
+          ctx.width,
+          barMetric,
+          costType,
+          resolveBarColor,
+        );
+      },
       onZoom: (zoomGroup, ctx) => {
         const newXScale = ctx.newXScale as d3.ScaleLinear<number, number>;
         const barColor = (d: InterpolatedResult) => resolveBarColor(d);
@@ -547,46 +664,58 @@ export default function ThroughputBarChart({
     };
 
     return [barLayer, labelLayer];
-  }, [sortedResults, barMetric, costType, hardwareConfig, mode, colorResolver]);
+  }, [sortedResults, barMetric, costType, mode, maxBarValue, paletteIdentity, resolveBarColor]);
 
   // ── Tooltip ──
+  const tooltipStateRef = useRef({ hardwareConfig, mode, barMetric, costType, runUrl, locale });
+  tooltipStateRef.current = { hardwareConfig, mode, barMetric, costType, runUrl, locale };
 
   const tooltip = useMemo(
     () => ({
       rulerType: 'vertical' as const,
-      content: (d: InterpolatedResult, isPinned: boolean) =>
-        generateTooltipHTML(
-          d,
-          hardwareConfig,
-          mode,
-          barMetric,
-          costType,
-          runUrl,
+      content: (datum: InterpolatedResult, isPinned: boolean) => {
+        const state = tooltipStateRef.current;
+        return generateTooltipHTML(
+          datum,
+          state.hardwareConfig,
+          state.mode,
+          state.barMetric,
+          state.costType,
+          state.runUrl,
           isPinned,
-          OVERLAY_STRINGS[locale],
-        ),
+          OVERLAY_STRINGS[state.locale],
+        );
+      },
       getRulerX: () => hoveredBarXRef.current,
-      onHoverStart: (sel: d3.Selection<any, InterpolatedResult, any, any>) => {
-        hoveredBarXRef.current = parseFloat(sel.attr('width') || '0');
+      onHoverStart: (
+        selection: d3.Selection<SVGRectElement, InterpolatedResult, SVGGElement, unknown>,
+      ) => {
+        hoveredBarXRef.current = Number.parseFloat(selection.attr('width') || '0');
         const hasSelection = selectedBarsRef.current.size > 0;
         if (!hasSelection) {
-          sel.attr('opacity', 1).attr('stroke', 'var(--foreground)').attr('stroke-width', 1.5);
+          selection
+            .attr('opacity', 1)
+            .attr('stroke', 'var(--foreground)')
+            .attr('stroke-width', 1.5);
         }
       },
-      onHoverEnd: (sel: d3.Selection<any, InterpolatedResult, any, any>, d: InterpolatedResult) => {
+      onHoverEnd: (
+        selection: d3.Selection<SVGRectElement, InterpolatedResult, SVGGElement, unknown>,
+        datum: InterpolatedResult,
+      ) => {
         const hasSelection = selectedBarsRef.current.size > 0;
-        const isSelected = selectedBarsRef.current.has(d.resultKey);
-        sel
+        const isSelected = selectedBarsRef.current.has(datum.resultKey);
+        selection
           .attr('opacity', hasSelection ? (isSelected ? 0.95 : 0.15) : 0.85)
           .attr('stroke', 'none');
       },
-      onPointClick: (d: InterpolatedResult) => {
-        onBarSelectRef.current(d.resultKey);
-        track('calculator_bar_selected', { gpu: d.hwKey, precision: d.precision });
+      onPointClick: (datum: InterpolatedResult) => {
+        onBarSelectRef.current(datum.resultKey);
+        track('calculator_bar_selected', { gpu: datum.hwKey, precision: datum.precision });
       },
       attachToLayer: 0,
     }),
-    [hardwareConfig, mode, barMetric, costType, runUrl, locale],
+    [],
   );
 
   // ── Y axis customize: map resultKey → display label, then split into two-line GPU labels ──
@@ -652,25 +781,21 @@ export default function ThroughputBarChart({
       ref={chartRef}
       chartId="calculator-chart"
       data={sortedResults}
+      dataIdentity={dataIdentity}
+      metricIdentity={metricIdentity}
+      displayIdentity={paletteIdentity}
       height={dynamicHeight}
       margin={dynamicMargin}
       watermark={getChartWatermark()}
       testId="calculator-bar-chart"
       grabCursor
       clipContent={false}
-      xScale={{ type: 'linear', domain: [0, maxBarValue] }}
-      yScale={{ type: 'band', domain: yDomain, padding: 0.3 }}
+      xScale={xScaleConfig}
+      yScale={yScaleConfig}
       xAxis={xAxisConfig}
       yAxis={yAxisConfig}
       layers={layers}
-      zoom={{
-        enabled: true,
-        axes: 'x',
-        scaleExtent: [0.1, 1],
-        rescaleX: (xScale, transform) =>
-          xScale.copy().domain([0, maxBarValue / transform.k]) as ContinuousScale,
-        customTransformStorage: (transform) => d3.zoomIdentity.scale(transform.k),
-      }}
+      zoom={zoomConfig}
       instructions="Shift+Scroll to zoom horizontally · Drag to pan · Double-click to reset · Click a bar to select"
       tooltip={tooltip}
       onRender={onRender}
diff --git a/packages/app/src/components/calculator/ThroughputCalculatorDisplay.test.ts b/packages/app/src/components/calculator/ThroughputCalculatorDisplay.test.ts
new file mode 100644
index 000000000..40911f23d
--- /dev/null
+++ b/packages/app/src/components/calculator/ThroughputCalculatorDisplay.test.ts
@@ -0,0 +1,60 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  resolveCalculatorBarSelection,
+  resolveCalculatorTarget,
+  resolveCalculatorTargetInputValue,
+  resolveCalculatorVisibility,
+} from '@/components/calculator/ThroughputCalculatorDisplay';
+
+describe('calculator effective state selectors', () => {
+  it('starts a new data scope with every available GPU visible', () => {
+    expect(resolveCalculatorVisibility(null, 'scope-b', ['h100', 'b200'])).toEqual(
+      new Set(['h100', 'b200']),
+    );
+    expect(
+      resolveCalculatorVisibility(
+        { scopeKey: 'scope-a', visible: new Set(['h100']), known: new Set(['h100']) },
+        'scope-b',
+        ['b200'],
+      ),
+    ).toEqual(new Set(['b200']));
+  });
+
+  it('adds late overlay GPUs while retaining visibility intent for known GPUs', () => {
+    const intent = {
+      scopeKey: 'scope',
+      visible: new Set(['h100']),
+      known: new Set(['h100', 'b200']),
+    };
+    expect(resolveCalculatorVisibility(intent, 'scope', ['h100', 'b200', 'overlay'])).toEqual(
+      new Set(['h100', 'overlay']),
+    );
+  });
+
+  it('drops departed GPUs and never leaves an available chart empty', () => {
+    const intent = {
+      scopeKey: 'scope',
+      visible: new Set(['departed']),
+      known: new Set(['departed', 'h100']),
+    };
+    expect(resolveCalculatorVisibility(intent, 'scope', ['h100'])).toEqual(new Set(['h100']));
+  });
+
+  it('keeps requested target intent while deriving a clamped effective target', () => {
+    expect(resolveCalculatorTarget(35, false, { min: 50, max: 100 })).toBe(35);
+    expect(resolveCalculatorTarget(35, true, { min: 50, max: 100 })).toBe(50);
+    expect(resolveCalculatorTarget(120, true, { min: 50, max: 100 })).toBe(100);
+  });
+
+  it('shows the effective clamp when idle while preserving the active editing buffer', () => {
+    expect(resolveCalculatorTargetInputValue('100', 80, false)).toBe('80');
+    expect(resolveCalculatorTargetInputValue('100', 80, true)).toBe('100');
+  });
+
+  it('clears stale bar selections when results change and prunes missing IDs', () => {
+    const intent = { resultsKey: 'old', selected: new Set(['a', 'b']) };
+    expect(resolveCalculatorBarSelection(intent, 'new', new Set(['a', 'b']))).toEqual(new Set());
+    expect(resolveCalculatorBarSelection(intent, 'old', new Set(['b']))).toEqual(new Set(['b']));
+  });
+});
diff --git a/packages/app/src/components/calculator/ThroughputCalculatorDisplay.tsx b/packages/app/src/components/calculator/ThroughputCalculatorDisplay.tsx
index c6a7234ae..1dc21286b 100644
--- a/packages/app/src/components/calculator/ThroughputCalculatorDisplay.tsx
+++ b/packages/app/src/components/calculator/ThroughputCalculatorDisplay.tsx
@@ -5,14 +5,28 @@ import Link from 'next/link';
 import { BarChart3, Table2 } from 'lucide-react';
 import { useFeatureGate } from '@/lib/use-feature-gate';
 import { useLocale } from '@/lib/use-locale';
-import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { useCallback, useMemo, useState } from 'react';
+
+import {
+  HW_REGISTRY,
+  TCO_SOURCE_TITLE,
+  TCO_SOURCE_URL,
+} from '@semianalysisai/inferencex-constants';
 
 import CalculatorTable from '@/components/calculator/CalculatorTable';
-import FleetPlanner from '@/components/calculator/FleetPlanner';
+import CostTargetPanel from '@/components/calculator/CostTargetPanel';
 import type { CalculatorUrlSeed } from '@/components/calculator/url-seed';
-import { GlobalFilterProvider, useGlobalFilters } from '@/components/GlobalFilterContext';
+import {
+  GlobalFilterProvider,
+  useGlobalFilterActions,
+  useGlobalFilterAvailability,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+  useGlobalFilterWorkflow,
+} from '@/components/GlobalFilterContext';
 import { Badge } from '@/components/ui/badge';
 import { Card } from '@/components/ui/card';
+import { CollapsibleSection } from '@/components/ui/collapsible-section';
 import { ChartButtons } from '@/components/ui/chart-buttons';
 import ChartLegend from '@/components/ui/chart-legend';
 import { ChartShareActions } from '@/components/ui/chart-display-helpers';
@@ -26,8 +40,16 @@ import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
 import { Input } from '@/components/ui/input';
 import { LabelWithTooltip } from '@/components/ui/label-with-tooltip';
 import { UnofficialDomainNotice } from '@/components/ui/unofficial-domain-notice';
+import {
+  includesJalapenoResult,
+  includesVeraRubinResult,
+  JalapenoOfficialPreviewNotice,
+  VeraRubinOfficialPreviewNotice,
+} from '@/components/official-preview-notice';
 import { useUnofficialRun } from '@/components/unofficial-run-provider';
 import { overlayRunColor } from '@/lib/overlay-run-style';
+import { localePath } from '@/lib/i18n';
+import { DEFAULT_FLEET_MW, readUrlParams, writeUrlParams } from '@/lib/url-state';
 import { Switch } from '@/components/ui/switch';
 import { TooltipProvider } from '@/components/ui/tooltip';
 import { MultiSelect } from '@/components/ui/multi-select';
@@ -42,10 +64,10 @@ import {
   getPrecisionLabel,
   getSequenceLabel,
 } from '@/lib/data-mappings';
-import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
 import { getHardwareConfig, getModelSortIndex } from '@/lib/constants';
 import { useThemeColors } from '@/hooks/useThemeColors';
 import { useUrlState } from '@/hooks/useUrlState';
+import { useOpenDropdown } from '@/hooks/useOpenDropdown';
 
 import { getDisplayLabel } from '@/lib/utils';
 import { exportToCsv } from '@/lib/csv-export';
@@ -67,10 +89,14 @@ import type {
 } from './types';
 import { useThroughputData } from './useThroughputData';
 
-const COST_PROVIDER_OPTIONS: { value: CostProvider; label: string }[] = [
-  { value: 'costh', label: 'Hyperscaler' },
-  { value: 'costn', label: 'Neocloud' },
-  { value: 'costr', label: '3yr Rental' },
+const COST_PROVIDER_OPTIONS: {
+  value: CostProvider;
+  label: string;
+  labelZh: string;
+}[] = [
+  { value: 'costh', label: 'Hyperscaler', labelZh: '超大规模云服务商' },
+  { value: 'costn', label: 'Neocloud', labelZh: 'Neocloud' },
+  { value: 'costr', label: '3yr Rental', labelZh: '3 年租赁' },
 ];
 
 const COST_TYPE_OPTIONS: { value: CostType; label: string }[] = [
@@ -85,6 +111,51 @@ const BAR_METRIC_OPTIONS: { value: BarMetric; label: string }[] = [
   { value: 'cost', label: 'Cost' },
 ];
 
+export interface CalculatorVisibilityIntent {
+  scopeKey: string;
+  visible: Set<string>;
+  known: Set<string>;
+}
+
+export function resolveCalculatorVisibility(
+  intent: CalculatorVisibilityIntent | null,
+  scopeKey: string,
+  legendHwKeys: readonly string[],
+): Set<string> {
+  if (!intent || intent.scopeKey !== scopeKey) return new Set(legendHwKeys);
+  const available = new Set(legendHwKeys);
+  const next = new Set([...intent.visible].filter((key) => available.has(key)));
+  for (const key of legendHwKeys) {
+    if (!intent.known.has(key)) next.add(key);
+  }
+  return next.size === 0 && legendHwKeys.length > 0 ? available : next;
+}
+
+export function resolveCalculatorTarget(
+  requestedTarget: number,
+  hasData: boolean,
+  range: { min: number; max: number },
+): number {
+  if (!hasData) return requestedTarget;
+  return Math.max(range.min, Math.min(range.max, requestedTarget));
+}
+
+export function resolveCalculatorTargetInputValue(
+  editingValue: string,
+  effectiveTarget: number,
+  isEditing: boolean,
+): string {
+  return isEditing ? editingValue : String(effectiveTarget);
+}
+
+export function resolveCalculatorBarSelection(
+  intent: { resultsKey: string; selected: Set<string> } | null,
+  resultsKey: string,
+  availableResultKeys: ReadonlySet<string>,
+): Set<string> {
+  if (!intent || intent.resultsKey !== resultsKey) return new Set();
+  return new Set([...intent.selected].filter((resultKey) => availableResultKeys.has(resultKey)));
+}
 type CalculatorViewMode = 'chart' | 'table';
 
 const CALCULATOR_VIEW_MODE_OPTIONS: SegmentedToggleOption<CalculatorViewMode>[] = [
@@ -126,12 +197,13 @@ const STRINGS = {
       'The interactivity operating point used for interpolation. Adjust the slider to compare chip throughput, cost, and power efficiency at different interactivity levels.',
     targetAgenticLabel: (percentile: string) => `Target ${percentile} Interactivity (tok/s/user)`,
     targetAgenticTooltip: (percentile: string) =>
-      `The ${percentile} interactivity operating point used for agentic trace interpolation. Adjust the slider to compare chip throughput, cost, and power efficiency.`,
+      `The ${percentile} interactivity operating point used for agentic workload interpolation. Adjust the slider to compare chip throughput, cost, and power efficiency.`,
     metricThroughput: 'Throughput',
     metricCost: 'Cost',
     viewChart: 'Chart',
     viewTable: 'Table',
     viewModeAria: 'View mode',
+    toggleSection: 'Expand or fold this section',
     errorLoading: 'Error loading data. Please try a different selection.',
     clickToCompare: 'selected. Click another bar to compare.',
     clearSelection: 'Clear selection',
@@ -146,9 +218,12 @@ const STRINGS = {
     updated: ' • Updated: ',
     note: 'Note:',
     disaggCost:
-      ' Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate cost per decode chip or per prefill chip, rather than per total chip count. This makes direct cost comparison with aggregated configs not an apples-to-apples comparison.',
+      ' Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) report input and output throughput per prefill chip and per decode chip rather than per chip overall, and $/M tok is derived from those rates — so on the Input and Output token types a disaggregated config reads cheaper than it is, by a median 2× and up to 18× across run history. Total-token cost is unaffected: it comes from throughput per chip overall, which both kinds report on the same basis. The Fleet Lifecycle page deliberately differs — it derives its per-token-type figures from that total, so everything there is on one denominator.',
     disaggThroughput:
-      ' Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) calculate throughput per decode chip or per prefill chip, rather than per total chip count. This makes direct throughput comparison with aggregated configs not an apples-to-apples comparison.',
+      ' Disaggregated inference configurations (e.g., MoRI SGLang, Dynamo TRTLLM) report input and output throughput per prefill chip and per decode chip rather than per chip overall — divided by fewer chips, so on the Input and Output token types a disaggregated config reads faster per chip than it is, by a median 2× and up to 18× on input and 7× on output. Total throughput is unaffected: both kinds report it per chip overall. The Fleet Lifecycle page deliberately differs — it derives its per-token-type figures from that total, so everything there is on one denominator.',
+    fleetMoved:
+      'Fleet Lifecycle has its own page: a fixed fleet sized against a power budget, projected across its life with every measured config improvement.',
+    fleetMovedLink: 'Open Fleet Lifecycle',
     compMetricThroughput: 'throughput',
     compMetricCost: 'cost efficiency',
     compMetricPower: 'tok/s/MW',
@@ -162,7 +237,7 @@ const STRINGS = {
   zh: {
     title: 'TCO 计算器',
     description:
-      '设定目标交互性（tokens/sec/user），比较所有 Chip 的吞吐量和成本。数值基于真实基准测试数据插值计算。',
+      '设定目标交互性（tokens/sec/user），比较所有芯片的吞吐量和成本。数值基于真实基准测试数据插值计算。',
     costProviderLabel: '成本供应商',
     costProviderTooltip:
       '用于计算每百万 token 成本的定价层级。Hyperscaler（如 AWS/GCP）、Neocloud（如 CoreWeave）或 3 年租赁。',
@@ -175,15 +250,16 @@ const STRINGS = {
       '图表中显示的比较指标。吞吐量（tok/s/chip）、能效（tok/s/MW）或每百万 token 成本。',
     targetLabel: '目标交互性 (tok/s/user)',
     targetTooltip:
-      '用于插值的交互性操作点。调整滑块以比较不同交互性级别下 Chip 的吞吐量、成本和能效。',
+      '用于插值的交互性操作点。调整滑块以比较不同交互性级别下芯片的吞吐量、成本和能效。',
     targetAgenticLabel: (percentile: string) => `目标 ${percentile} 交互性 (tok/s/user)`,
     targetAgenticTooltip: (percentile: string) =>
-      `用于智能体轨迹插值的 ${percentile} 交互性操作点。调整滑块以比较 Chip 的吞吐量、成本和能效。`,
+      `用于智能体工作负载插值的 ${percentile} 交互性操作点。调整滑块以比较芯片的吞吐量、成本和能效。`,
     metricThroughput: '吞吐量',
     metricCost: '成本',
     viewChart: '图表',
     viewTable: '表格',
     viewModeAria: '显示模式',
+    toggleSection: '展开或折叠此板块',
     errorLoading: '加载数据出错，请尝试其他选择。',
     clickToCompare: '已选中。点击另一个柱状图进行对比。',
     clearSelection: '清除选择',
@@ -192,15 +268,18 @@ const STRINGS = {
     totalTokens: '总 Token',
     inputTokens: '输入 Token',
     outputTokens: '输出 Token',
-    allInPower: '全含功率/Chip：',
+    allInPower: '全含功率/芯片：',
     tcoPerHr: 'TCO $/chip/hr：',
     source: '来源：',
     updated: ' • 更新于：',
     note: '注意：',
     disaggCost:
-      '解耦推理配置（如 MoRI SGLang、Dynamo TRTLLM）按解码 Chip 或预填充 Chip 计算成本，而非按 Chip 总数。因此与聚合配置的直接成本对比并非同类比较。',
+      '解耦推理配置（如 MoRI SGLang、Dynamo TRTLLM）的输入与输出吞吐量分别按预填充芯片与解码芯片报告，而非按芯片总数，而 $/M tok 由这些速率推导。因此在「输入」与「输出」token 类型下，解耦配置显示的成本低于实际，中位数偏低 2 倍，在运行历史中最高达 18 倍。总计口径的成本不受影响：它来自按芯片总数计的吞吐量，两种部署方式在该口径上一致。「集群生命周期」页面与此有意不同：它由该总量推导各 token 类型的数值，因此那里的所有数字都在同一分母上。',
     disaggThroughput:
-      '解耦推理配置（如 MoRI SGLang、Dynamo TRTLLM）按解码 Chip 或预填充 Chip 计算吞吐量，而非按 Chip 总数。因此与聚合配置的直接吞吐量对比并非同类比较。',
+      '解耦推理配置（如 MoRI SGLang、Dynamo TRTLLM）的输入与输出吞吐量分别按预填充芯片与解码芯片报告，而非按芯片总数。由于除以的芯片数更少，在「输入」与「输出」token 类型下，解耦配置显示的每芯片吞吐量高于实际，中位数偏高 2 倍，输入最高达 18 倍、输出最高达 7 倍。总计口径的吞吐量不受影响：两种部署方式均按芯片总数报告。「集群生命周期」页面与此有意不同：它由该总量推导各 token 类型的数值，因此那里的所有数字都在同一分母上。',
+    fleetMoved:
+      '「集群生命周期」已移至独立页面：按功率预算确定固定集群的规模，结合每次实测配置改进，测算其整个生命周期的经济性。',
+    fleetMovedLink: '打开集群生命周期',
     compMetricThroughput: '吞吐量',
     compMetricCost: '成本效率',
     compMetricPower: 'tok/s/MW',
@@ -239,66 +318,58 @@ function getChartTitleZh(
     }
     default: {
       return mode === 'interactivity_to_throughput'
-        ? `${targetLabel}下每 Chip ${tokenTypeLabel} token 吞吐量`
+        ? `${targetLabel}下每芯片 ${tokenTypeLabel} token 吞吐量`
         : `${targetLabel}下的交互性`;
     }
   }
 }
 
 export default function ThroughputCalculatorDisplay({ urlSeed }: { urlSeed?: CalculatorUrlSeed }) {
-  const inner = (
-    <ThroughputCalculatorInner initialPercentile={urlSeed?.percentile ?? Percentile.P90} />
+  return (
+    <GlobalFilterProvider
+      initialModel={urlSeed?.model}
+      initialSequence={urlSeed?.sequence}
+      initialPrecisions={urlSeed?.precisions}
+      initialRunDate={urlSeed?.runDate}
+      initialRunId={urlSeed?.runId}
+    >
+      <ThroughputCalculatorInner initialPercentile={urlSeed?.percentile ?? Percentile.P90} />
+    </GlobalFilterProvider>
   );
-  if (urlSeed && (urlSeed.model || urlSeed.sequence || urlSeed.precisions)) {
-    return (
-      <GlobalFilterProvider
-        initialModel={urlSeed.model}
-        initialSequence={urlSeed.sequence}
-        initialPrecisions={urlSeed.precisions}
-      >
-        {inner}
-      </GlobalFilterProvider>
-    );
-  }
-  return inner;
 }
 
 function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: Percentile }) {
   const locale = useLocale();
   const t = STRINGS[locale];
   const { setUrlParam } = useUrlState();
-  const [openDropdown, setOpenDropdown] = useState<string | null>(null);
-  const handleDropdownOpenChange = (dropdownKey: string) => (isOpen: boolean) => {
-    if (isOpen) {
-      setOpenDropdown(dropdownKey);
-      return;
-    }
-    setOpenDropdown((current) => (current === dropdownKey ? null : current));
-  };
+  const { openDropdown, handleDropdownOpenChange } = useOpenDropdown();
 
   const {
     selectedModel,
-    setSelectedModel,
-    selectedRunDate,
-    workflowInfo,
     effectiveSequence: selectedSequence,
-    setSelectedSequence,
     effectivePrecisions: selectedPrecisions,
-    setSelectedPrecisions,
-    availablePrecisions,
-    availableSequences,
-    availableModels,
-  } = useGlobalFilters();
-
+  } = useGlobalFilterSelection();
+  const { setSelectedModel, setSelectedSequence, setSelectedPrecisions } = useGlobalFilterActions();
+  const { selectedRunDate, selectedRunId } = useGlobalFilterRun();
+  const { availablePrecisions, availableSequences, availableModels } =
+    useGlobalFilterAvailability();
+  const { availableRuns } = useGlobalFilterWorkflow();
   const mode = 'interactivity_to_throughput' as const;
   const [costProvider, setCostProvider] = useState<CostProvider>('costh');
   const [costType, setCostType] = useState<CostType>('total');
-  const [targetValue, setTargetValue] = useState<number>(35);
+  const [requestedTargetValue, setRequestedTargetValue] = useState<number>(35);
   const [inputValue, setInputValue] = useState<string>('35');
+  const [isTargetInputFocused, setIsTargetInputFocused] = useState(false);
   const [barMetric, setBarMetric] = useState<BarMetric>('throughput');
   const [selectedPercentile, setSelectedPercentile] = useState<Percentile>(initialPercentile);
-  const [visibleHwKeys, setVisibleHwKeys] = useState<Set<string>>(new Set());
-  const [selectedBars, setSelectedBars] = useState<Set<string>>(new Set());
+  // The `c_mw` URL seed is read once here; the cost-target panel renders the
+  // input and consumes the budget. The Fleet Lifecycle page shares the param.
+  const [mwInput, setMwInput] = useState<string>(() => readUrlParams().c_mw ?? DEFAULT_FLEET_MW);
+  const [visibilityIntent, setVisibilityIntent] = useState<CalculatorVisibilityIntent | null>(null);
+  const [barSelectionIntent, setBarSelectionIntent] = useState<{
+    resultsKey: string;
+    selected: Set<string>;
+  } | null>(null);
   const [isLegendExpanded, setIsLegendExpanded] = useState(true);
   const [highContrast, setHighContrast] = useState(false);
   const [viewMode, setViewMode] = useState<CalculatorViewMode>('chart');
@@ -340,7 +411,7 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
     getResults,
     getOverlayResults,
     loading,
-    error,
+    error: throughputError,
     hasData,
     hasOverlayData,
     availableHwKeys,
@@ -352,7 +423,11 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
     selectedRunDate,
     overlayInput,
     selectedPercentile,
+    undefined,
+    true,
+    costType,
   );
+  const error = throughputError;
 
   const isAgenticSequence = selectedSequence === Sequence.AgenticTraces;
   // AgentX publishes on P90, so the percentile control is an insider affordance
@@ -387,84 +462,38 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
     return [...new Set([...availableHwKeys, ...overlayAvailableHwKeys])];
   }, [isUnofficialRun, availableHwKeys, overlayAvailableHwKeys]);
 
-  // Dynamic vendor-aware colors for visible GPUs
+  const handleMwInputChange = useCallback((raw: string) => {
+    setMwInput(raw);
+    const parsed = parseFloat(raw);
+    writeUrlParams({ c_mw: Number.isFinite(parsed) && parsed > 0 ? raw : '' });
+  }, []);
+
+  const mw = useMemo(() => {
+    const parsed = parseFloat(mwInput);
+    return Number.isFinite(parsed) && parsed > 0 ? parsed : null;
+  }, [mwInput]);
+
+  // The selection key represents user-driven scope. Overlay hardware is left
+  // out on purpose so late overlay arrivals augment, rather than reset, intent.
+  const selectionKey = `${selectedModel}|${selectedSequence}|${[...selectedPrecisions]
+    .toSorted()
+    .join(',')}|${selectedRunDate}|${[...availableHwKeys].toSorted().join(',')}`;
+
+  const visibleHwKeys = useMemo(
+    () => resolveCalculatorVisibility(visibilityIntent, selectionKey, legendHwKeys),
+    [visibilityIntent, selectionKey, legendHwKeys],
+  );
   const visibleKeysArray = useMemo(() => [...visibleHwKeys], [visibleHwKeys]);
   const { resolveColor } = useThemeColors({
     highContrast,
     activeKeys: visibleKeysArray,
   });
 
-  // Track previous available keys to detect when the GPU set changes
-  const prevAvailableKeyRef = useRef<string>('');
-  const prevOverlayKeyRef = useRef<string>('');
-
-  // Reset visible GPUs on a user-driven selection change. The key is the
-  // selection itself PLUS the official hardware list — the selection so an
-  // overlay-only model/sequence (where the official list is empty and stays
-  // empty) still reseeds, the official list so anything else that changes which
-  // GPUs have data still reseeds. Percentile is deliberately excluded: it
-  // recalculates agentic values, but should preserve the user's GPU filters
-  // whenever the available hardware set is unchanged. Also deliberately NOT
-  // keyed on the merged list: an unofficial run is fetched separately and
-  // usually lands after the benchmarks, so a late arrival — or a run dismissal
-  // — would otherwise wipe GPU filters the user had already set.
-  const selectionKey = `${selectedModel}|${selectedSequence}|${[...selectedPrecisions]
-    .toSorted()
-    .join(',')}|${selectedRunDate}|${[...availableHwKeys].toSorted().join(',')}`;
-  useEffect(() => {
-    // Nothing to seed from yet (first load, before either source has resolved).
-    // Guards on the MERGED list: an empty official list is a real state, not
-    // just a loading one, and bailing on it would leave stale official keys in
-    // `visibleHwKeys` and throw off the solo/show-all arithmetic below.
-    if (legendHwKeys.length === 0) return;
-    if (selectionKey !== prevAvailableKeyRef.current) {
-      prevAvailableKeyRef.current = selectionKey;
-      setVisibleHwKeys(new Set(legendHwKeys));
-    }
-  }, [selectionKey, legendHwKeys]);
-
-  // Overlay hardware arriving or leaving is additive: newly available overlay
-  // GPUs start visible, ones that are gone stop being tracked, and every other
-  // entry keeps whatever the user set.
-  useEffect(() => {
-    const key = overlayAvailableHwKeys.join(',');
-    if (key === prevOverlayKeyRef.current) return;
-    const prev = prevOverlayKeyRef.current ? prevOverlayKeyRef.current.split(',') : [];
-    prevOverlayKeyRef.current = key;
-
-    const added = overlayAvailableHwKeys.filter((k) => !prev.includes(k));
-    // Only drop hardware that has no official data either — otherwise dismissing
-    // a run would hide a GPU whose official bar is still on the chart.
-    const removed = prev.filter(
-      (k) => !overlayAvailableHwKeys.includes(k) && !availableHwKeys.includes(k),
-    );
-    if (added.length === 0 && removed.length === 0) return;
-
-    setVisibleHwKeys((cur) => {
-      const next = new Set(cur);
-      added.forEach((k) => next.add(k));
-      removed.forEach((k) => next.delete(k));
-      // Never strand the user with an empty chart. Falls back to everything
-      // that still has data, official AND overlay — on an overlay-only
-      // selection the official list is empty, so falling back to it would blank
-      // the chart while overlay bars were still available.
-      if (next.size === 0) return new Set([...availableHwKeys, ...overlayAvailableHwKeys]);
-      return next;
-    });
-  }, [overlayAvailableHwKeys, availableHwKeys]);
-
   const hasAnyData = hasData || hasOverlayData;
-
-  // Clamp target into range when data changes
-  useEffect(() => {
-    if (!hasAnyData) return;
-    const { min, max } = ranges.interactivity;
-    if (targetValue < min || targetValue > max) {
-      const clamped = Math.max(min, Math.min(max, targetValue));
-      setTargetValue(clamped);
-      setInputValue(String(clamped));
-    }
-  }, [hasAnyData, ranges]);
+  const targetValue = useMemo(
+    () => resolveCalculatorTarget(requestedTargetValue, hasAnyData, ranges.interactivity),
+    [hasAnyData, ranges.interactivity, requestedTargetValue],
+  );
 
   const results: InterpolatedResult[] = useMemo(() => {
     if (!hasData) return [];
@@ -479,7 +508,7 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
     hideSkuAboveConfigLimit,
   ]);
 
-  /** Branch + URL per run index, stamped onto overlay results for labels/tooltips. */
+  /** Branch and URL per run index, stamped onto overlay results for labels and tooltips. */
   const runInfoByIndex = useMemo(() => {
     const map: Record<number, { branch: string; url: string }> = {};
     unofficialRunInfos.forEach((info, idx) => {
@@ -509,33 +538,59 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
     hideSkuAboveConfigLimit,
   ]);
 
-  /**
-   * Bars drawn in the chart: official + overlay. Deliberately NOT used by the
-   * table, the CSV export, or the fleet planner — those stay official-only, so
-   * an exported sheet or a fleet projection never silently mixes in numbers
-   * from an unmerged branch.
-   */
   const barResults = useMemo(
     () => (overlayResults.length > 0 ? [...results, ...overlayResults] : results),
     [results, overlayResults],
   );
+  const showsJalapenoPreview = useMemo(
+    () => includesJalapenoResult(results.map((result) => result.hwKey)),
+    [results],
+  );
+  const showsVeraRubinPreview = useMemo(
+    () => includesVeraRubinResult(results.map((result) => result.hwKey)),
+    [results],
+  );
+  const barResultsKey = useMemo(
+    () =>
+      barResults
+        .map(
+          (result) =>
+            `${result.resultKey}:${result.value}:${result.cost}:${result.tpPerMw}:${result.concurrency}`,
+        )
+        .join('|'),
+    [barResults],
+  );
+  const selectedBars = useMemo(
+    () =>
+      resolveCalculatorBarSelection(
+        barSelectionIntent,
+        barResultsKey,
+        new Set(barResults.map((result) => result.resultKey)),
+      ),
+    [barSelectionIntent, barResultsKey, barResults],
+  );
 
-  const currentRange = useMemo(() => ranges.interactivity, [ranges]);
+  const currentRange = ranges.interactivity;
 
-  const handleSliderChange = useCallback((e: React.ChangeEvent<HTMLInputElement>) => {
-    const val = Number(e.target.value);
-    setTargetValue(val);
-    setInputValue(String(val));
+  const handleSliderChange = useCallback((event: React.ChangeEvent<HTMLInputElement>) => {
+    const value = Number(event.target.value);
+    setRequestedTargetValue(value);
+    setInputValue(String(value));
   }, []);
 
   const handleInputChange = useCallback((e: React.ChangeEvent<HTMLInputElement>) => {
     setInputValue(e.target.value);
     const parsed = parseFloat(e.target.value);
     if (!isNaN(parsed) && parsed >= 0) {
-      setTargetValue(parsed);
+      setRequestedTargetValue(parsed);
     }
   }, []);
 
+  const handleInputFocus = useCallback(() => {
+    setInputValue(String(targetValue));
+    setIsTargetInputFocused(true);
+  }, [targetValue]);
+
   const handleInputBlur = useCallback(() => {
     const parsed = parseFloat(inputValue);
     if (isNaN(parsed) || parsed < 0) {
@@ -543,11 +598,12 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
     } else {
       const { min, max } = ranges.interactivity;
       const clamped = Math.max(min, Math.min(max, parsed));
-      setTargetValue(clamped);
+      setRequestedTargetValue(clamped);
       setInputValue(String(clamped));
     }
+    setIsTargetInputFocused(false);
     track('calculator_target_set', { mode, value: targetValue });
-  }, [inputValue, targetValue, mode, ranges]);
+  }, [inputValue, targetValue, mode, ranges.interactivity]);
 
   const handleCostProviderChange = useCallback((value: string) => {
     setCostProvider(value as CostProvider);
@@ -561,6 +617,7 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
 
   const handleModelChange = useCallback(
     (value: string) => {
+      setVisibilityIntent(null);
       setSelectedModel(value as Model);
       track('calculator_model_selected', { model: value });
     },
@@ -569,6 +626,7 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
 
   const handleSequenceChange = useCallback(
     (value: string) => {
+      setVisibilityIntent(null);
       setSelectedSequence(value as Sequence);
       track('calculator_sequence_selected', { sequence: value });
     },
@@ -577,6 +635,7 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
 
   const handlePrecisionChange = useCallback(
     (value: string[]) => {
+      setVisibilityIntent(null);
       setSelectedPrecisions(value);
       track('calculator_precision_selected', { precision: value.join(',') });
     },
@@ -599,42 +658,41 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
 
   const toggleGpuVisibility = useCallback(
     (hwKey: string) => {
-      setVisibleHwKeys((prev) => {
-        // Count against the legend rather than the raw set size, so an entry
-        // that is no longer in the legend can never skew solo/show-all.
-        const visibleLegendKeys = legendHwKeys.filter((k) => prev.has(k));
-        const allVisible = visibleLegendKeys.length === legendHwKeys.length;
-        const isVisible = prev.has(hwKey);
-
-        if (isVisible) {
-          if (allVisible) {
-            // If all visible and clicking one, solo it
-            return new Set([hwKey]);
-          } else if (visibleLegendKeys.length === 1) {
-            // If only one visible and clicking it, show all
-            return new Set(legendHwKeys);
-          }
-          // Remove it
-          const next = new Set(prev);
-          next.delete(hwKey);
-          return next;
-        }
-        // Add it
-        const next = new Set([...prev, hwKey]);
-        return next;
+      const visibleLegendKeys = legendHwKeys.filter((key) => visibleHwKeys.has(key));
+      const allVisible = visibleLegendKeys.length === legendHwKeys.length;
+      const isVisible = visibleHwKeys.has(hwKey);
+      let next: Set<string>;
+      if (isVisible && allVisible) {
+        next = new Set([hwKey]);
+      } else if (isVisible && visibleLegendKeys.length === 1) {
+        next = new Set(legendHwKeys);
+      } else {
+        next = new Set(visibleHwKeys);
+        if (isVisible) next.delete(hwKey);
+        else next.add(hwKey);
+      }
+      setVisibilityIntent({
+        scopeKey: selectionKey,
+        visible: next,
+        known: new Set(legendHwKeys),
       });
       track('calculator_gpu_toggled', { gpu: hwKey });
     },
-    [legendHwKeys],
+    [legendHwKeys, visibleHwKeys, selectionKey],
   );
 
-  const removeGpu = useCallback((hwKey: string) => {
-    setVisibleHwKeys((prev) => {
-      const next = new Set(prev);
+  const removeGpu = useCallback(
+    (hwKey: string) => {
+      const next = new Set(visibleHwKeys);
       next.delete(hwKey);
-      return next;
-    });
-  }, []);
+      setVisibilityIntent({
+        scopeKey: selectionKey,
+        visible: next,
+        known: new Set(legendHwKeys),
+      });
+    },
+    [visibleHwKeys, selectionKey, legendHwKeys],
+  );
 
   const handleExportCsv = useCallback(() => {
     const { headers, rows } = calculatorChartToCsv(results, targetValue, (hwKey) => {
@@ -655,21 +713,22 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
   }, []);
 
   const handleResetGpus = useCallback(() => {
-    setVisibleHwKeys(new Set(legendHwKeys));
+    setVisibilityIntent({
+      scopeKey: selectionKey,
+      visible: new Set(legendHwKeys),
+      known: new Set(legendHwKeys),
+    });
     track('calculator_gpu_reset', { gpuCount: legendHwKeys.length });
-  }, [legendHwKeys]);
-
-  // Derive runUrl from workflowInfo for the selected sequence
-  const runUrl = useMemo(() => {
-    if (!Array.isArray(workflowInfo) || workflowInfo.length === 0) return undefined;
-    const wf = workflowInfo[0];
-    return wf?.runInfoBySequence?.[selectedSequence]?.runUrl;
-  }, [workflowInfo, selectedSequence]);
-
-  // Handle bar selection: click to toggle (uses resultKey for unique identification)
-  const handleBarSelect = useCallback((resultKey: string) => {
-    setSelectedBars((prev) => {
-      const next = new Set(prev);
+  }, [legendHwKeys, selectionKey]);
+
+  const runUrl = availableRuns[selectedRunId]?.runUrl;
+
+  // Selection intent is tied to the current result values. Any data or filter
+  // transition changes the key, so stale comparisons disappear without a
+  // post-render clearing effect.
+  const handleBarSelect = useCallback(
+    (resultKey: string) => {
+      const next = new Set(selectedBars);
       if (next.has(resultKey)) {
         next.delete(resultKey);
         track('calculator_bar_deselected', { resultKey });
@@ -677,14 +736,10 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
         next.add(resultKey);
         track('calculator_bar_selected', { resultKey, totalSelected: next.size });
       }
-      return next;
-    });
-  }, []);
-
-  // Clear bar selection when results change (data/filter changes)
-  useEffect(() => {
-    setSelectedBars(new Set());
-  }, [barResults]);
+      setBarSelectionIntent({ resultsKey: barResultsKey, selected: next });
+    },
+    [selectedBars, barResultsKey],
+  );
 
   // Generate comparison text when 2+ bars are selected. Overlay bars are
   // selectable too, so this reads the combined chart list.
@@ -822,6 +877,28 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
     resolveColor,
   ]);
 
+  // Hoisted out of the caption because the fold header needs the same words: a
+  // folded section shows only its title, and it has to be this chart's title
+  // rather than a static one, since the title carries the metric and target.
+  const chartTitle =
+    locale === 'zh'
+      ? getChartTitleZh(
+          barMetric,
+          mode,
+          targetValue,
+          costType,
+          costProvider,
+          isAgenticSequence ? selectedPercentile : undefined,
+        )
+      : getChartTitle(
+          barMetric,
+          mode,
+          targetValue,
+          costType,
+          costProvider,
+          isAgenticSequence ? selectedPercentile : undefined,
+        );
+
   if (!loading && error) {
     console.error(error);
     return (
@@ -897,9 +974,9 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
                   />
                   <div id="calc-cost" data-testid="calc-cost-selector">
                     <MultiSelect
-                      options={COST_PROVIDER_OPTIONS.map((c) => ({
-                        value: c.value,
-                        label: c.label,
+                      options={COST_PROVIDER_OPTIONS.map((provider) => ({
+                        value: provider.value,
+                        label: locale === 'zh' ? provider.labelZh : provider.label,
                       }))}
                       value={[costProvider]}
                       onChange={(values) => {
@@ -1053,7 +1130,12 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
                     </div>
                     <Input
                       type="number"
-                      value={inputValue}
+                      value={resolveCalculatorTargetInputValue(
+                        inputValue,
+                        targetValue,
+                        isTargetInputFocused,
+                      )}
+                      onFocus={handleInputFocus}
                       onChange={handleInputChange}
                       onBlur={handleInputBlur}
                       className="w-24 h-9"
@@ -1067,242 +1149,251 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
         </Card>
       </section>
 
+      {/* The inverse of the target-interactivity slider directly above: fix a cost
+          ceiling, read off the fastest speed that stays under it. */}
+      {!loading && hasData && (
+        <CostTargetPanel
+          gpuDataByGroupKey={gpuDataByGroupKey}
+          hardwareConfig={hardwareConfig}
+          costProvider={costProvider}
+          costType={costType}
+          visibleHwKeys={visibleHwKeys}
+          mw={mw}
+          mwInput={mwInput}
+          onMwInputChange={handleMwInputChange}
+        />
+      )}
+
       {/* Chart / Table */}
       <section data-testid="calculator-chart-section">
-        <figure data-testid="calculator-figure" className="relative rounded-lg">
-          <ChartButtons
-            chartId="calculator-chart"
-            analyticsPrefix="calculator"
-            zoomResetEvent="d3chart_zoom_reset_calculator-chart"
-            onExportCsv={handleExportCsv}
-            setIsLegendExpanded={setIsLegendExpanded}
-            exportFileName={`InferenceX_calculator_${selectedModel}`}
-            leadingControls={
-              <SegmentedToggle
-                value={viewMode}
-                options={viewModeOptions}
-                onValueChange={handleViewModeChange}
-                ariaLabel={t.viewModeAria}
-                testId="calculator-view-toggle"
-                className="shrink-0"
-              />
-            }
-          />
-          <Card>
-            {loading ? (
-              <Skeleton className="h-125 w-full" />
-            ) : (
-              <>
-                {(() => {
-                  const captionContent = (
-                    <>
-                      <div className="flex items-start justify-between gap-4">
-                        <h2 className="text-lg font-semibold">
-                          {locale === 'zh'
-                            ? getChartTitleZh(
-                                barMetric,
-                                mode,
-                                targetValue,
-                                costType,
-                                costProvider,
-                                isAgenticSequence ? selectedPercentile : undefined,
-                              )
-                            : getChartTitle(
-                                barMetric,
-                                mode,
-                                targetValue,
-                                costType,
-                                costProvider,
-                                isAgenticSequence ? selectedPercentile : undefined,
-                              )}
-                        </h2>
-                        <SegmentedToggle
-                          value={viewMode}
-                          options={mobileViewModeOptions}
-                          onValueChange={handleViewModeChange}
-                          ariaLabel={t.viewModeAria}
-                          className="md:hidden shrink-0"
-                        />
-                      </div>
-                      <p className="text-sm text-muted-foreground mb-2">
-                        {getModelLabel(selectedModel)} •{' '}
-                        {selectedPrecisions
-                          .map((p) => getPrecisionLabel(p as Precision))
-                          .join(', ')}{' '}
-                        • {getSequenceLabel(selectedSequence, locale)} • {t.source}SemiAnalysis
-                        InferenceX™
-                        {selectedRunDate && (
+        <CollapsibleSection
+          title={chartTitle}
+          // The chart's own caption already carries this title, so the header
+          // shows it only while folded.
+          titleWhenOpen={false}
+          toggleLabel={t.toggleSection}
+          testId="calculator-chart-collapse"
+          onToggle={(open) => track('calculator_section_toggled', { section: 'chart', open })}
+        >
+          <figure data-testid="calculator-figure" className="relative rounded-lg">
+            <ChartButtons
+              chartId="calculator-chart"
+              analyticsPrefix="calculator"
+              zoomResetEvent="d3chart_zoom_reset_calculator-chart"
+              onExportCsv={handleExportCsv}
+              setIsLegendExpanded={setIsLegendExpanded}
+              exportFileName={`InferenceX_calculator_${selectedModel}`}
+              leadingControls={
+                <SegmentedToggle
+                  value={viewMode}
+                  options={viewModeOptions}
+                  onValueChange={handleViewModeChange}
+                  ariaLabel={t.viewModeAria}
+                  testId="calculator-view-toggle"
+                  className="shrink-0"
+                />
+              }
+            />
+            <Card>
+              {loading ? (
+                <Skeleton className="h-125 w-full" />
+              ) : (
+                <>
+                  {(() => {
+                    const captionContent = (
+                      <>
+                        <div className="flex items-start justify-between gap-4">
+                          <h2 className="text-lg font-semibold">{chartTitle}</h2>
+                          <SegmentedToggle
+                            value={viewMode}
+                            options={mobileViewModeOptions}
+                            onValueChange={handleViewModeChange}
+                            ariaLabel={t.viewModeAria}
+                            className="md:hidden shrink-0"
+                          />
+                        </div>
+                        <p className="text-sm text-muted-foreground mb-2">
+                          {getModelLabel(selectedModel)} •{' '}
+                          {selectedPrecisions
+                            .map((p) => getPrecisionLabel(p as Precision))
+                            .join(', ')}{' '}
+                          • {getSequenceLabel(selectedSequence, locale)} • {t.source}SemiAnalysis
+                          InferenceX™
+                          {selectedRunDate && (
+                            <>
+                              {t.updated}
+                              {selectedRunDate}
+                            </>
+                          )}
+                        </p>
+                        {showsJalapenoPreview && <JalapenoOfficialPreviewNotice />}
+                        {showsVeraRubinPreview && <VeraRubinOfficialPreviewNotice />}
+                        {barMetric === 'power' && barResults.length > 0 && (
                           <>
-                            {t.updated}
-                            {selectedRunDate}
+                            <p
+                              className="text-muted-foreground mb-2 flex flex-wrap gap-2 items-center"
+                              data-testid="calculator-cost-badges"
+                            >
+                              {t.allInPower}
+                              {Object.entries(HW_REGISTRY).map(([base, specs]) => (
+                                <Badge key={base} variant="outline">
+                                  {specs.badgeLabel ?? base.toUpperCase()}: {specs.power}kW
+                                </Badge>
+                              ))}
+                            </p>
+                            <p className="text-muted-foreground">
+                              <small>
+                                {t.source}
+                                <Link
+                                  target="_blank"
+                                  className="underline hover:text-foreground"
+                                  href="https://semianalysis.com/datacenter-industry-model/"
+                                >
+                                  SemiAnalysis Datacenter Industry Model
+                                  <ExternalLinkIcon />
+                                </Link>
+                              </small>
+                            </p>
                           </>
                         )}
-                      </p>
-                      {barMetric === 'power' && barResults.length > 0 && (
-                        <>
-                          <p
-                            className="text-muted-foreground mb-2 flex flex-wrap gap-2 items-center"
-                            data-testid="calculator-cost-badges"
-                          >
-                            {t.allInPower}
-                            {Object.entries(HW_REGISTRY).map(([base, specs]) => (
-                              <Badge key={base} variant="outline">
-                                {base.toUpperCase()}: {specs.power}kW
-                              </Badge>
-                            ))}
-                          </p>
-                          <p className="text-muted-foreground">
-                            <small>
-                              {t.source}
-                              <Link
-                                target="_blank"
-                                className="underline hover:text-foreground"
-                                href="https://semianalysis.com/datacenter-industry-model/"
-                              >
-                                SemiAnalysis Datacenter Industry Model
-                                <ExternalLinkIcon />
-                              </Link>
-                            </small>
-                          </p>
-                        </>
-                      )}
-                      {barMetric === 'cost' && barResults.length > 0 && (
-                        <>
-                          <p
-                            className="text-muted-foreground mb-2 flex flex-wrap gap-2 items-center"
-                            data-testid="calculator-cost-badges"
-                          >
-                            {t.tcoPerHr}
-                            {Object.entries(HW_REGISTRY).map(([base, specs]) => (
-                              <Badge key={base} variant="outline">
-                                {base.toUpperCase()}: $
-                                {(costProvider === 'costh'
-                                  ? specs.costh
-                                  : costProvider === 'costn'
-                                    ? specs.costn
-                                    : specs.costr
-                                ).toFixed(2)}
-                                /hr
-                              </Badge>
-                            ))}
-                          </p>
-                          <p className="text-muted-foreground">
-                            <small>
-                              {t.source}
-                              <Link
-                                target="_blank"
-                                className="underline hover:text-foreground"
-                                href="https://semianalysis.com/ai-cloud-tco-model/"
-                              >
-                                SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model
-                                <ExternalLinkIcon />
-                              </Link>
-                            </small>
-                          </p>
-                        </>
-                      )}
-                      {/* Per-token-type cost only: the input- and output-token
+                        {barMetric === 'cost' && barResults.length > 0 && (
+                          <>
+                            <p
+                              className="text-muted-foreground mb-2 flex flex-wrap gap-2 items-center"
+                              data-testid="calculator-cost-badges"
+                            >
+                              {t.tcoPerHr}
+                              {Object.entries(HW_REGISTRY).map(([base, specs]) => (
+                                <Badge key={base} variant="outline">
+                                  {specs.badgeLabel ?? base.toUpperCase()}: $
+                                  {(costProvider === 'costh'
+                                    ? specs.costh
+                                    : costProvider === 'costn'
+                                      ? specs.costn
+                                      : specs.costr
+                                  ).toFixed(2)}
+                                  /hr
+                                </Badge>
+                              ))}
+                            </p>
+                            <p className="text-muted-foreground">
+                              <small>
+                                {t.source}
+                                <Link
+                                  target="_blank"
+                                  className="underline hover:text-foreground"
+                                  href={TCO_SOURCE_URL}
+                                >
+                                  {TCO_SOURCE_TITLE}
+                                  <ExternalLinkIcon />
+                                </Link>
+                              </small>
+                            </p>
+                          </>
+                        )}
+                        {/* Per-token-type cost only: the input- and output-token
                           costs are attributed to one side of a disagg config's
                           prefill/decode split, while the total-token cost uses
                           the whole chip count — the same denominator an
                           aggregated config uses — so it needs no caveat. */}
-                      <div
-                        className={`overflow-hidden transition-all duration-200 ease-in-out ${
-                          barMetric === 'cost' && costType !== 'total'
-                            ? 'max-h-20 opacity-100'
-                            : 'max-h-0 opacity-0'
-                        }`}
-                      >
-                        <p
-                          data-testid="calculator-disagg-cost-note"
-                          className="text-muted-foreground text-xs mt-2 border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1"
+                        <div
+                          className={`overflow-hidden transition-all duration-200 ease-in-out ${
+                            barMetric === 'cost' && costType !== 'total'
+                              ? 'max-h-20 opacity-100'
+                              : 'max-h-0 opacity-0'
+                          }`}
                         >
-                          <strong>{t.note}</strong>
-                          {t.disaggCost}
-                        </p>
-                      </div>
-                      <div
-                        className={`overflow-hidden transition-all duration-200 ease-in-out ${
-                          barMetric === 'throughput' || barMetric === 'power'
-                            ? 'max-h-20 opacity-100'
-                            : 'max-h-0 opacity-0'
-                        }`}
-                      >
-                        <p className="text-muted-foreground text-xs mt-2 border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1">
-                          <strong>{t.note}</strong>
-                          {t.disaggThroughput}
-                        </p>
-                      </div>
-                      <UnofficialDomainNotice />
-                    </>
-                  );
-
-                  return viewMode === 'chart' ? (
-                    <ThroughputBarChart
-                      caption={captionContent}
-                      results={barResults}
-                      hardwareConfig={hardwareConfig}
-                      mode={mode}
-                      targetValue={targetValue}
-                      barMetric={barMetric}
-                      costType={costType}
-                      runUrl={runUrl}
-                      selectedBars={selectedBars}
-                      onBarSelect={handleBarSelect}
-                      colorResolver={resolveColor}
-                      legendElement={
-                        legendHwKeys.length > 0 ? (
-                          <ChartLegend
-                            variant="sidebar"
-                            legendItems={legendItems}
-                            onItemRemove={removeGpu}
-                            isLegendExpanded={isLegendExpanded}
-                            onExpandedChange={(expanded) => {
-                              setIsLegendExpanded(expanded);
-                              track('calculator_legend_expanded', { expanded });
-                            }}
-                            switches={[
-                              {
-                                id: 'calc-high-contrast',
-                                label: t.highContrast,
-                                checked: highContrast,
-                                onCheckedChange: (checked: boolean) => {
-                                  setHighContrast(checked);
-                                  track('calculator_high_contrast_toggled', { enabled: checked });
-                                },
-                              },
-                            ]}
-                            actions={
-                              visibleHwKeys.size < legendHwKeys.length
-                                ? [
-                                    {
-                                      id: 'calc-reset-filter',
-                                      label: t.resetFilter,
-                                      onClick: handleResetGpus,
-                                    },
-                                  ]
-                                : []
-                            }
-                            enableTooltips={true}
-                          />
-                        ) : undefined
-                      }
-                    />
-                  ) : (
-                    <>
-                      <figcaption>{captionContent}</figcaption>
-                      <CalculatorTable
-                        results={results}
-                        costType={costType}
+                          <p
+                            data-testid="calculator-disagg-cost-note"
+                            className="text-muted-foreground text-xs mt-2 border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1"
+                          >
+                            <strong>{t.note}</strong>
+                            {t.disaggCost}
+                          </p>
+                        </div>
+                        <div
+                          className={`overflow-hidden transition-all duration-200 ease-in-out ${
+                            barMetric === 'throughput' || barMetric === 'power'
+                              ? 'max-h-20 opacity-100'
+                              : 'max-h-0 opacity-0'
+                          }`}
+                        >
+                          <p className="text-muted-foreground text-xs mt-2 border-l-2 border-amber-500 pl-2 bg-amber-500/5 py-1">
+                            <strong>{t.note}</strong>
+                            {t.disaggThroughput}
+                          </p>
+                        </div>
+                        <UnofficialDomainNotice />
+                      </>
+                    );
+
+                    return viewMode === 'chart' ? (
+                      <ThroughputBarChart
+                        caption={captionContent}
+                        results={barResults}
                         hardwareConfig={hardwareConfig}
+                        mode={mode}
+                        targetValue={targetValue}
+                        barMetric={barMetric}
+                        costType={costType}
+                        runUrl={runUrl}
+                        selectedBars={selectedBars}
+                        onBarSelect={handleBarSelect}
+                        colorResolver={resolveColor}
+                        legendElement={
+                          legendHwKeys.length > 0 ? (
+                            <ChartLegend
+                              variant="sidebar"
+                              legendItems={legendItems}
+                              onItemRemove={removeGpu}
+                              isLegendExpanded={isLegendExpanded}
+                              onExpandedChange={(expanded) => {
+                                setIsLegendExpanded(expanded);
+                                track('calculator_legend_expanded', { expanded });
+                              }}
+                              switches={[
+                                {
+                                  id: 'calc-high-contrast',
+                                  label: t.highContrast,
+                                  checked: highContrast,
+                                  onCheckedChange: (checked: boolean) => {
+                                    setHighContrast(checked);
+                                    track('calculator_high_contrast_toggled', { enabled: checked });
+                                  },
+                                },
+                              ]}
+                              actions={
+                                visibleHwKeys.size < legendHwKeys.length
+                                  ? [
+                                      {
+                                        id: 'calc-reset-filter',
+                                        label: t.resetFilter,
+                                        onClick: handleResetGpus,
+                                      },
+                                    ]
+                                  : []
+                              }
+                              enableTooltips={true}
+                            />
+                          ) : undefined
+                        }
                       />
-                    </>
-                  );
-                })()}
-              </>
-            )}
-          </Card>
-        </figure>
+                    ) : (
+                      <>
+                        <figcaption>{captionContent}</figcaption>
+                        <CalculatorTable
+                          results={results}
+                          costType={costType}
+                          hardwareConfig={hardwareConfig}
+                        />
+                      </>
+                    );
+                  })()}
+                </>
+              )}
+            </Card>
+          </figure>
+        </CollapsibleSection>
       </section>
 
       {/* Comparison banner — only shown in chart view */}
@@ -1336,7 +1427,7 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
                 type="button"
                 onClick={() => {
                   track('calculator_selection_cleared', { clearedCount: selectedBars.size });
-                  setSelectedBars(new Set());
+                  setBarSelectionIntent({ resultsKey: barResultsKey, selected: new Set() });
                 }}
                 className="text-xs text-muted-foreground hover:text-foreground underline shrink-0"
               >
@@ -1347,17 +1438,23 @@ function ThroughputCalculatorInner({ initialPercentile }: { initialPercentile: P
         </section>
       )}
 
-      {/* Fleet planner: MW-budget projection + cost-target inverse lookup */}
+      {/* Fleet lifecycle now lives on its own page (/fleet); leave a pointer
+          so calculator users can find where the section moved. */}
       {!loading && hasData && (
-        <FleetPlanner
-          results={results}
-          gpuDataByGroupKey={gpuDataByGroupKey}
-          hardwareConfig={hardwareConfig}
-          costProvider={costProvider}
-          costType={costType}
-          targetValue={targetValue}
-          visibleHwKeys={visibleHwKeys}
-        />
+        <section data-testid="calculator-fleet-pointer">
+          <Card>
+            <p className="text-sm text-muted-foreground">
+              {t.fleetMoved}{' '}
+              <Link
+                href={localePath('/fleet', locale)}
+                className="underline hover:text-foreground"
+                data-testid="calculator-fleet-pointer-link"
+              >
+                {t.fleetMovedLink}
+              </Link>
+            </p>
+          </Card>
+        </section>
       )}
     </div>
   );
diff --git a/packages/app/src/components/calculator/historical-best.test.ts b/packages/app/src/components/calculator/historical-best.test.ts
new file mode 100644
index 000000000..9829bf461
--- /dev/null
+++ b/packages/app/src/components/calculator/historical-best.test.ts
@@ -0,0 +1,710 @@
+import { describe, expect, it } from 'vitest';
+
+import type { BenchmarkRow } from '@/lib/api';
+import { Sequence } from '@/lib/data-mappings';
+
+import {
+  bestSoFarProgression,
+  groupHistoryByHwKeyAndDate,
+  mergeProgressionsByChip,
+  selectBestFromGroups,
+  selectHistoricalBest,
+  type GroupHistoryOptions,
+  type SelectBestOptions,
+} from './historical-best';
+import type { InterpolatedResult } from './types';
+
+/**
+ * Rows go through the real `buildGpuGroups`, so `hardware` must exist in the
+ * hardware registry and isl/osl must match the selected sequence.
+ */
+function makeRow(overrides: Partial<BenchmarkRow> = {}): BenchmarkRow {
+  return {
+    id: 1,
+    hardware: 'b300',
+    framework: 'sglang',
+    model: 'dsv4',
+    precision: 'fp4',
+    spec_method: 'none',
+    disagg: false,
+    is_multinode: false,
+    prefill_tp: 8,
+    prefill_ep: 8,
+    prefill_dp_attention: false,
+    prefill_num_workers: 1,
+    decode_tp: 8,
+    decode_ep: 8,
+    decode_dp_attention: false,
+    decode_num_workers: 1,
+    num_prefill_gpu: 8,
+    num_decode_gpu: 8,
+    benchmark_type: 'single_turn',
+    isl: 1024,
+    osl: 1024,
+    conc: 8,
+    offload_mode: 'off',
+    image: 'sglang:test',
+    metrics: {
+      median_intvty: 50,
+      tput_per_gpu: 900,
+      output_tput_per_gpu: 300,
+      input_tput_per_gpu: 600,
+    },
+    date: '2026-07-19',
+    run_url: null,
+    ...overrides,
+  };
+}
+
+/** A sweep point: one row at a given interactivity / throughput on a given date. */
+function sweep(
+  date: string,
+  interactivity: number,
+  tputPerGpu: number,
+  overrides: Partial<BenchmarkRow> = {},
+): BenchmarkRow {
+  return makeRow({
+    id: Math.round(interactivity * 1000 + tputPerGpu),
+    date,
+    conc: Math.round(tputPerGpu / interactivity),
+    metrics: {
+      median_intvty: interactivity,
+      tput_per_gpu: tputPerGpu,
+      output_tput_per_gpu: tputPerGpu * 0.3,
+      input_tput_per_gpu: tputPerGpu * 0.7,
+    },
+    ...overrides,
+  });
+}
+
+/** Ranks by total-token throughput — stands in for the injected cost-matrix accessor. */
+const rankByThroughput = (r: InterpolatedResult) => r.value;
+
+function options(
+  rows: BenchmarkRow[],
+  over: Partial<GroupHistoryOptions & SelectBestOptions> = {},
+) {
+  return {
+    rows,
+    sequence: Sequence.OneK_OneK,
+    precisions: ['fp4'],
+    targetValue: 50,
+    mode: 'interactivity_to_throughput' as const,
+    costProvider: 'costh' as const,
+    rank: rankByThroughput,
+    ...over,
+  };
+}
+
+describe('selectHistoricalBest', () => {
+  it('returns empty results for empty input', () => {
+    expect(selectHistoricalBest(options([]))).toEqual({ best: [], unmeasured: [], datesSeen: 0 });
+    expect(selectHistoricalBest(options([makeRow()], { precisions: [] })).best).toEqual([]);
+  });
+
+  it('promotes an earlier date when it beat the latest at the target', () => {
+    // June sweep is stronger at the target; July explored elsewhere and regressed.
+    const rows = [
+      sweep('2026-06-01', 20, 1200),
+      sweep('2026-06-01', 50, 900),
+      sweep('2026-06-01', 80, 400),
+      sweep('2026-07-01', 20, 1000),
+      sweep('2026-07-01', 50, 600),
+      sweep('2026-07-01', 80, 300),
+    ];
+
+    const { best, datesSeen } = selectHistoricalBest(options(rows));
+    expect(datesSeen).toBe(2);
+    expect(best).toHaveLength(1);
+    const entry = best[0]!;
+    expect(entry.date).toBe('2026-06-01');
+    expect(entry.latestDate).toBe('2026-07-01');
+    expect(entry.supersededLatest).toBe(true);
+    expect(entry.datesConsidered).toBe(2);
+    expect(entry.datesMeasured).toBe(2);
+    expect(entry.result.value).toBeCloseTo(900, 6);
+    // The latest date is still reported, so the UI can show what was given up.
+    expect(entry.latestRankValue).toBeCloseTo(600, 6);
+  });
+
+  it('keeps the latest date when it is genuinely the best', () => {
+    const rows = [
+      sweep('2026-06-01', 20, 900),
+      sweep('2026-06-01', 50, 600),
+      sweep('2026-07-01', 20, 1400),
+      sweep('2026-07-01', 50, 1100),
+    ];
+
+    const entry = selectHistoricalBest(options(rows)).best[0]!;
+    expect(entry.date).toBe('2026-07-01');
+    expect(entry.supersededLatest).toBe(false);
+    expect(entry.latestRankValue).toBeCloseTo(entry.rankValue, 6);
+  });
+
+  it('never lets a clamped read win', () => {
+    // The June sweep tops out at 30 tok/s/user but is very high throughput
+    // there. Clamping would credit 5000 tok/s/gpu at a target of 50, which
+    // June never measured — that is the artifact this rule exists to kill.
+    const rows = [
+      sweep('2026-06-01', 10, 6000),
+      sweep('2026-06-01', 30, 5000),
+      sweep('2026-07-01', 40, 800),
+      sweep('2026-07-01', 60, 500),
+    ];
+
+    const entry = selectHistoricalBest(options(rows)).best[0]!;
+    expect(entry.date).toBe('2026-07-01');
+    expect(entry.datesMeasured).toBe(1);
+    expect(entry.datesConsidered).toBe(2);
+    expect(entry.result.clamped).toBeFalsy();
+    // Nowhere near June's clamped 5000.
+    expect(entry.result.value).toBeLessThan(1000);
+  });
+
+  it('reports an hwKey measured nowhere near the target instead of dropping it', () => {
+    const rows = [sweep('2026-06-01', 60, 900), sweep('2026-06-01', 90, 400)];
+
+    const { best, unmeasured } = selectHistoricalBest(options(rows, { targetValue: 20 }));
+    expect(best).toEqual([]);
+    expect(unmeasured).toHaveLength(1);
+    // The nearest measured point lets the UI explain *why* there is no number.
+    // Everything this chip has is above the target, so there is no value below.
+    expect(unmeasured[0]!.nearestBelow).toBeNull();
+    expect(unmeasured[0]!.nearestAbove).toBe(60);
+    expect(unmeasured[0]!.datesConsidered).toBe(1);
+  });
+
+  it('never reports a neighbourhood that brackets the target it just excluded', () => {
+    // Two dates, each a single point, on opposite sides of the target. The old
+    // disclosure took the union of raw points across dates and reported
+    // "measured 20.0-80.0" for a target of 50 — an interval no sweep covered,
+    // which reads as a bug in the exclusion rather than an explanation of it.
+    // Neither date can be read at 50: a one-point frontier is only readable at
+    // its own interactivity.
+    const rows = [sweep('2026-06-01', 20, 1000), sweep('2026-07-01', 80, 400)];
+
+    const { best, unmeasured } = selectHistoricalBest(options(rows, { targetValue: 50 }));
+    expect(best).toEqual([]);
+    expect(unmeasured).toHaveLength(1);
+    const entry = unmeasured[0]!;
+    expect(entry.nearestBelow).toBe(20);
+    expect(entry.nearestAbove).toBe(80);
+    expect(entry.datesConsidered).toBe(2);
+    // The claim the pair makes is "nothing at 50", not "measured 20 through 80".
+    // Whatever is reported, the target must not sit strictly inside a span any
+    // single date covered — no such date exists, or there would be a winner.
+    expect(entry.nearestBelow).toBeLessThan(50);
+    expect(entry.nearestAbove).toBeGreaterThan(50);
+  });
+
+  it('ranks by the injected accessor, not by throughput', () => {
+    const rows = [
+      sweep('2026-06-01', 40, 1000),
+      sweep('2026-06-01', 60, 900),
+      sweep('2026-07-01', 40, 800),
+      sweep('2026-07-01', 60, 700),
+    ];
+
+    // Inverting the ranking must invert the winner — proof the caller's
+    // cost-matrix accessor is what decides, not a hardcoded field.
+    const highest = selectHistoricalBest(options(rows)).best[0]!;
+    const lowest = selectHistoricalBest(options(rows, { rank: (r) => -r.value })).best[0]!;
+    expect(highest.date).toBe('2026-06-01');
+    expect(lowest.date).toBe('2026-07-01');
+  });
+
+  it('ignores a rank that is not a finite number', () => {
+    const rows = [sweep('2026-06-01', 40, 1000), sweep('2026-06-01', 60, 900)];
+    const { best, unmeasured } = selectHistoricalBest(options(rows, { rank: () => NaN }));
+    expect(best).toEqual([]);
+    expect(unmeasured.map((u) => u.hwKey)).toHaveLength(1);
+  });
+
+  it('keeps each date on its own frontier', () => {
+    // If dates were pooled, the June high-throughput point and the July
+    // high-interactivity point would form one frontier and interpolate to a
+    // config no single sweep ever produced.
+    const rows = [
+      sweep('2026-06-01', 30, 2000),
+      sweep('2026-06-01', 40, 1800),
+      sweep('2026-07-01', 60, 900),
+      sweep('2026-07-01', 80, 400),
+    ];
+
+    const { best, unmeasured } = selectHistoricalBest(options(rows, { targetValue: 50 }));
+    // Neither date measured 50, and pooling is the only way to get a read.
+    expect(best).toEqual([]);
+    expect(unmeasured).toHaveLength(1);
+  });
+
+  it('tracks separate hardware separately and ranks the winners', () => {
+    const rows = [
+      sweep('2026-06-01', 40, 1000, { hardware: 'b300' }),
+      sweep('2026-06-01', 60, 800, { hardware: 'b300' }),
+      sweep('2026-06-01', 40, 400, { hardware: 'h200', precision: 'fp8' }),
+      sweep('2026-06-01', 60, 300, { hardware: 'h200', precision: 'fp8' }),
+    ];
+
+    const { best } = selectHistoricalBest(options(rows, { precisions: ['fp4', 'fp8'] }));
+    expect(best).toHaveLength(2);
+    // Ranked best first.
+    expect(best[0]!.rankValue).toBeGreaterThan(best[1]!.rankValue);
+    expect(best[0]!.hwKey).toContain('b300');
+    expect(best[1]!.hwKey).toContain('h200');
+    // One series per hwKey, keyed for the legend.
+    for (const entry of best) expect(entry.result.resultKey).toBe(entry.hwKey);
+  });
+
+  it('pools precisions into one frontier per hwKey', () => {
+    // Precision is part of the config, and the question is what the chip's best
+    // config does — so fp8 winning on one date must be eligible.
+    const rows = [
+      sweep('2026-06-01', 40, 700, { precision: 'fp4' }),
+      sweep('2026-06-01', 60, 500, { precision: 'fp4' }),
+      sweep('2026-06-01', 40, 1500, { precision: 'fp8' }),
+      sweep('2026-06-01', 60, 1200, { precision: 'fp8' }),
+    ];
+
+    const { best } = selectHistoricalBest(options(rows, { precisions: ['fp4', 'fp8'] }));
+    expect(best).toHaveLength(1);
+    // The fp8 frontier dominates, so the read reflects it.
+    expect(best[0]!.result.value).toBeGreaterThan(1000);
+  });
+
+  it('restricts the search to the visible legend keys', () => {
+    const rows = [
+      sweep('2026-06-01', 40, 1000, { hardware: 'b300' }),
+      sweep('2026-06-01', 60, 800, { hardware: 'b300' }),
+      sweep('2026-06-01', 40, 400, { hardware: 'h200', precision: 'fp8' }),
+      sweep('2026-06-01', 60, 300, { hardware: 'h200', precision: 'fp8' }),
+    ];
+
+    const all = selectHistoricalBest(options(rows, { precisions: ['fp4', 'fp8'] }));
+    const visible = new Set([all.best[0]!.hwKey]);
+    const filtered = selectHistoricalBest(
+      options(rows, { precisions: ['fp4', 'fp8'], visibleHwKeys: visible }),
+    );
+    expect(filtered.best).toHaveLength(1);
+    expect(filtered.best[0]!.hwKey).toBe(all.best[0]!.hwKey);
+    expect(filtered.unmeasured).toEqual([]);
+  });
+
+  it('carries the winning date run URLs for auditability', () => {
+    const june = 'https://github.com/org/repo/actions/runs/111';
+    const july = 'https://github.com/org/repo/actions/runs/222';
+    const rows = [
+      sweep('2026-06-01', 40, 1200, { run_url: june }),
+      sweep('2026-06-01', 60, 1000, { run_url: june }),
+      sweep('2026-07-01', 40, 500, { run_url: july }),
+      sweep('2026-07-01', 60, 400, { run_url: july }),
+    ];
+
+    const entry = selectHistoricalBest(options(rows)).best[0]!;
+    expect(entry.date).toBe('2026-06-01');
+    // Only the winning date's run — the number has to be traceable to it.
+    expect(entry.runUrls).toEqual([june]);
+  });
+
+  it('reports every run pooled into a winning date', () => {
+    // getAllBenchmarksForHistory has no DISTINCT ON, so a same-day re-sweep
+    // lands in one bucket. Surfacing both URLs makes that visible.
+    const a = 'https://github.com/org/repo/actions/runs/111';
+    const b = 'https://github.com/org/repo/actions/runs/222';
+    const rows = [
+      sweep('2026-06-01', 40, 1200, { run_url: a }),
+      sweep('2026-06-01', 60, 1000, { run_url: b }),
+    ];
+
+    const entry = selectHistoricalBest(options(rows)).best[0]!;
+    expect(entry.runUrls.toSorted()).toEqual([a, b]);
+  });
+
+  it('still selects when runs have no URL', () => {
+    const rows = [sweep('2026-06-01', 40, 1200), sweep('2026-06-01', 60, 1000)];
+    const entry = selectHistoricalBest(options(rows)).best[0]!;
+    expect(entry.runUrls).toEqual([]);
+  });
+
+  it('works in throughput_to_interactivity mode', () => {
+    const rows = [
+      sweep('2026-06-01', 30, 400),
+      sweep('2026-06-01', 10, 900),
+      sweep('2026-07-01', 20, 400),
+      sweep('2026-07-01', 8, 900),
+    ];
+
+    const { best } = selectHistoricalBest(
+      options(rows, { mode: 'throughput_to_interactivity', targetValue: 600 }),
+    );
+    expect(best).toHaveLength(1);
+    // Reads an interactivity, and June's curve is the higher one.
+    expect(best[0]!.date).toBe('2026-06-01');
+    expect(best[0]!.result.value).toBeGreaterThan(10);
+    expect(best[0]!.result.value).toBeLessThan(30);
+  });
+
+  it('reuses one grouping across several targets', () => {
+    // The two stages exist so moving the slider re-reads the frontiers without
+    // rebuilding them; the staged path must agree with the composed one.
+    const rows = [
+      sweep('2026-06-01', 20, 1200),
+      sweep('2026-06-01', 50, 900),
+      sweep('2026-06-01', 80, 400),
+      sweep('2026-07-01', 20, 1000),
+      sweep('2026-07-01', 50, 600),
+      sweep('2026-07-01', 80, 300),
+    ];
+
+    const groups = groupHistoryByHwKeyAndDate(options(rows));
+    expect(groups.datesSeen).toBe(2);
+    expect([...groups.byHwKey.values()][0]).toHaveLength(2);
+
+    for (const targetValue of [25, 50, 75]) {
+      const staged = selectBestFromGroups(groups, { ...options(rows), targetValue });
+      const composed = selectHistoricalBest(options(rows, { targetValue }));
+      expect(staged.best.map((e) => [e.hwKey, e.date, e.rankValue])).toEqual(
+        composed.best.map((e) => [e.hwKey, e.date, e.rankValue]),
+      );
+    }
+  });
+
+  it('drops rows whose sequence does not match the selection', () => {
+    const rows = [
+      sweep('2026-06-01', 40, 1000, { isl: 8192, osl: 1024 }),
+      sweep('2026-06-01', 60, 800, { isl: 8192, osl: 1024 }),
+    ];
+    expect(selectHistoricalBest(options(rows)).best).toEqual([]);
+    expect(selectHistoricalBest(options(rows)).datesSeen).toBe(0);
+  });
+});
+
+describe('bestSoFarProgression', () => {
+  it('keeps only the dates that beat every earlier date', () => {
+    // The MI355X shape: a baseline, a big jump, a regression that must not
+    // register, then another jump.
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 400 : 200)),
+      ...[30, 60].map((iv) => sweep('2026-05-02', iv, iv === 30 ? 900 : 500)),
+      // A sweep exploring elsewhere that came out worse at the target — the
+      // fleet kept serving the config it already had, so this is not a step.
+      ...[30, 60].map((iv) => sweep('2026-05-10', iv, iv === 30 ? 700 : 400)),
+      ...[30, 60].map((iv) => sweep('2026-05-27', iv, iv === 30 ? 1600 : 800)),
+    ];
+
+    const groups = groupHistoryByHwKeyAndDate(options(rows));
+    const progressions = bestSoFarProgression(groups, { ...options(rows), targetValue: 45 });
+
+    expect(progressions).toHaveLength(1);
+    const p = progressions[0]!;
+    expect(p.steps.map((s) => s.date)).toEqual(['2026-04-25', '2026-05-02', '2026-05-27']);
+    expect(p.datesMeasured).toBe(4);
+    expect(p.datesConsidered).toBe(4);
+  });
+
+  it('is strictly increasing in rank value', () => {
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 400 : 200)),
+      ...[30, 60].map((iv) => sweep('2026-05-02', iv, iv === 30 ? 900 : 500)),
+      ...[30, 60].map((iv) => sweep('2026-05-27', iv, iv === 30 ? 1600 : 800)),
+    ];
+    const groups = groupHistoryByHwKeyAndDate(options(rows));
+    const p = bestSoFarProgression(groups, { ...options(rows), targetValue: 45 }).at(0)!;
+    for (let i = 1; i < p.steps.length; i += 1) {
+      expect(p.steps[i]!.rankValue).toBeGreaterThan(p.steps[i - 1]!.rankValue);
+    }
+  });
+
+  it('reports each step gain over the opening config, with the first at 1', () => {
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 400 : 200)),
+      ...[30, 60].map((iv) => sweep('2026-05-27', iv, iv === 30 ? 1600 : 800)),
+    ];
+    const groups = groupHistoryByHwKeyAndDate(options(rows));
+    const p = bestSoFarProgression(groups, { ...options(rows), targetValue: 45 }).at(0)!;
+    expect(p.steps[0]!.factorOverFirst).toBe(1);
+    expect(p.steps.at(-1)!.factorOverFirst).toBeCloseTo(
+      p.steps.at(-1)!.rankValue / p.steps[0]!.rankValue,
+      9,
+    );
+    expect(p.steps.at(-1)!.factorOverFirst).toBeGreaterThan(1);
+  });
+
+  it('never lets a clamped read become a step', () => {
+    // April tops out at 30 tok/s/user; at a target of 45 it must not contribute,
+    // so the progression opens at the May sweep.
+    const rows = [
+      ...[10, 30].map((iv) => sweep('2026-04-25', iv, iv === 10 ? 6000 : 5000)),
+      ...[40, 60].map((iv) => sweep('2026-05-02', iv, iv === 40 ? 800 : 500)),
+    ];
+    const groups = groupHistoryByHwKeyAndDate(options(rows));
+    const p = bestSoFarProgression(groups, { ...options(rows), targetValue: 45 }).at(0)!;
+    expect(p.steps.map((s) => s.date)).toEqual(['2026-05-02']);
+    expect(p.steps[0]!.result.clamped).toBeFalsy();
+  });
+
+  it('carries the run URLs for every step', () => {
+    const runA = 'https://github.com/org/repo/actions/runs/111';
+    const runB = 'https://github.com/org/repo/actions/runs/222';
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 400 : 200, { run_url: runA })),
+      ...[30, 60].map((iv) => sweep('2026-05-27', iv, iv === 30 ? 1600 : 800, { run_url: runB })),
+    ];
+    const groups = groupHistoryByHwKeyAndDate(options(rows));
+    const p = bestSoFarProgression(groups, { ...options(rows), targetValue: 45 }).at(0)!;
+    expect(p.steps.map((s) => s.runUrls)).toEqual([[runA], [runB]]);
+  });
+
+  it('ends at the same config the all-time best selection reports', () => {
+    // The staircase's last rung and the headline figure must agree, or the table
+    // would contradict the chart.
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 400 : 200)),
+      ...[30, 60].map((iv) => sweep('2026-05-27', iv, iv === 30 ? 1600 : 800)),
+      ...[30, 60].map((iv) => sweep('2026-06-10', iv, iv === 30 ? 900 : 500)),
+    ];
+    const opts = { ...options(rows), targetValue: 45 };
+    const groups = groupHistoryByHwKeyAndDate(opts);
+    const best = selectBestFromGroups(groups, opts).best[0]!;
+    const p = bestSoFarProgression(groups, opts).at(0)!;
+    expect(p.steps.at(-1)!.date).toBe(best.date);
+    expect(p.steps.at(-1)!.rankValue).toBeCloseTo(best.rankValue, 9);
+  });
+
+  it('tracks each chip separately, ranked by its latest best', () => {
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 1400 : 900)),
+      ...[30, 60].map((iv) =>
+        sweep('2026-04-25', iv, iv === 30 ? 400 : 200, { hardware: 'h200', precision: 'fp8' }),
+      ),
+    ];
+    const opts = { ...options(rows), precisions: ['fp4', 'fp8'], targetValue: 45 };
+    const progressions = bestSoFarProgression(groupHistoryByHwKeyAndDate(opts), opts);
+    expect(progressions).toHaveLength(2);
+    expect(progressions[0]!.steps.at(-1)!.rankValue).toBeGreaterThan(
+      progressions[1]!.steps.at(-1)!.rankValue,
+    );
+  });
+
+  it('omits chips with no unclamped read at all', () => {
+    const rows = [70, 90].map((iv) => sweep('2026-04-25', iv, iv === 70 ? 400 : 200));
+    const opts = { ...options(rows), targetValue: 20 };
+    expect(bestSoFarProgression(groupHistoryByHwKeyAndDate(opts), opts)).toEqual([]);
+  });
+});
+
+describe('mergeProgressionsByChip', () => {
+  /** Two frameworks on one B300, each improving on its own dates. */
+  const twoFrameworks = [
+    // sglang: opens strong, then a modest gain.
+    ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 900 : 500)),
+    ...[30, 60].map((iv) => sweep('2026-06-01', iv, iv === 30 ? 1000 : 600)),
+    // trtllm: starts behind, then overtakes.
+    ...[30, 60].map((iv) =>
+      sweep('2026-05-02', iv, iv === 30 ? 600 : 300, { framework: 'trtllm' }),
+    ),
+    ...[30, 60].map((iv) =>
+      sweep('2026-07-04', iv, iv === 30 ? 1600 : 800, { framework: 'trtllm' }),
+    ),
+  ];
+
+  const mergeAt = (rows: BenchmarkRow[], targetValue = 45) => {
+    const opts = { ...options(rows), targetValue };
+    return mergeProgressionsByChip(bestSoFarProgression(groupHistoryByHwKeyAndDate(opts), opts));
+  };
+
+  it('draws one line per chip, not one per software config', () => {
+    // Four hwKey-level progressions' worth of history, one piece of silicon.
+    const chips = mergeAt(twoFrameworks);
+    expect(chips).toHaveLength(1);
+    expect(chips[0]!.baseGpu).toBe('b300');
+    expect(chips[0]!.key).toBe('b300');
+  });
+
+  it('follows whichever config is ahead at each moment', () => {
+    const chips = mergeAt(twoFrameworks);
+    const steps = chips[0]!.steps;
+    // trtllm's 2026-05-02 read (600) loses to sglang's incumbent 900, so it is
+    // not a rung; its 2026-07-04 read (1600) takes over and is.
+    expect(steps.map((s) => s.date)).toEqual(['2026-04-25', '2026-06-01', '2026-07-04']);
+    expect(steps.map((s) => s.result.hwKey)).toEqual(['b300_sglang', 'b300_sglang', 'b300_trt']);
+    expect(chips[0]!.hwKeysUsed).toEqual(['b300_sglang', 'b300_trt']);
+  });
+
+  it('is monotonic, so a losing config never pulls the line down', () => {
+    const steps = mergeAt(twoFrameworks)[0]!.steps;
+    for (let i = 1; i < steps.length; i += 1) {
+      expect(steps[i]!.rankValue).toBeGreaterThan(steps[i - 1]!.rankValue);
+    }
+  });
+
+  it('measures gain against the merged opening rung, not the winner own first run', () => {
+    const steps = mergeAt(twoFrameworks)[0]!.steps;
+    expect(steps[0]!.factorOverFirst).toBe(1);
+    // The last rung is trtllm's, but the gain is over sglang's opening read —
+    // that is what the fleet actually improved from.
+    expect(steps.at(-1)!.factorOverFirst).toBeCloseTo(
+      steps.at(-1)!.rankValue / steps[0]!.rankValue,
+      9,
+    );
+    expect(steps.at(-1)!.factorOverFirst).toBeGreaterThan(1);
+  });
+
+  it('lets disagg compete for the same line as aggregated configs', () => {
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 900 : 500)),
+      // Same silicon, run disaggregated, and stronger at the target.
+      ...[30, 60].map((iv) =>
+        sweep('2026-05-02', iv, iv === 30 ? 1600 : 800, {
+          framework: 'sglang-disagg',
+          disagg: true,
+        }),
+      ),
+    ];
+    const chips = mergeAt(rows);
+    expect(chips).toHaveLength(1);
+    const chip = chips[0]!;
+    expect(chip.baseGpu).toBe('b300');
+    // Both rungs are on the one line, and the disagg run took over.
+    expect(chip.steps.map((s) => s.date)).toEqual(['2026-04-25', '2026-05-02']);
+    // The chip is flagged so the section can carry the basis caveat.
+    expect(chip.disagg).toBe(true);
+  });
+
+  it('does not flag a chip as disagg when none of its rungs are', () => {
+    expect(mergeAt(twoFrameworks)[0]!.disagg).toBe(false);
+  });
+
+  it('separates distinct silicon', () => {
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 900 : 500)),
+      ...[30, 60].map((iv) =>
+        sweep('2026-05-02', iv, iv === 30 ? 1600 : 800, { hardware: 'b200' }),
+      ),
+    ];
+    const chips = mergeAt(rows);
+    expect(chips.map((c) => c.baseGpu).toSorted()).toEqual(['b200', 'b300']);
+  });
+
+  it('ranks chips by where they ended up', () => {
+    const rows = [
+      ...[30, 60].map((iv) => sweep('2026-04-25', iv, iv === 30 ? 900 : 500)),
+      ...[30, 60].map((iv) =>
+        sweep('2026-05-02', iv, iv === 30 ? 1600 : 800, { hardware: 'b200' }),
+      ),
+    ];
+    const chips = mergeAt(rows);
+    expect(chips[0]!.baseGpu).toBe('b200');
+    expect(chips[0]!.steps.at(-1)!.rankValue).toBeGreaterThan(chips[1]!.steps.at(-1)!.rankValue);
+  });
+
+  it('returns nothing for no progressions', () => {
+    expect(mergeProgressionsByChip([])).toEqual([]);
+  });
+});
+
+describe('agentic traces', () => {
+  /** An agentic sweep point: no ISL/OSL, interactivity read at the percentile. */
+  function agenticSweep(
+    date: string,
+    interactivity: number,
+    tputPerGpu: number,
+    cacheHitRate?: number,
+  ): BenchmarkRow {
+    return makeRow({
+      id: Math.round(interactivity * 1000 + tputPerGpu),
+      date,
+      benchmark_type: 'agentic_traces',
+      isl: null,
+      osl: null,
+      conc: Math.round(tputPerGpu / interactivity),
+      metrics: {
+        // Agentic interactivity is always derived as 1/itl — `*_intvty` straight
+        // off the artifact is dropped, the harness definition having drifted.
+        p90_itl: 1 / interactivity,
+        p90_e2el: 10,
+        tput_per_gpu: tputPerGpu,
+        // The agentic mix: input tokens dominate by two orders of magnitude.
+        output_tput_per_gpu: tputPerGpu * 0.008,
+        input_tput_per_gpu: tputPerGpu * 0.992,
+        ...(cacheHitRate === undefined ? {} : { server_gpu_cache_hit_rate: cacheHitRate }),
+      },
+    });
+  }
+
+  const agenticRows = [
+    agenticSweep('2026-08-07', 40, 800, 0.9),
+    agenticSweep('2026-08-07', 80, 500, 0.95),
+    agenticSweep('2026-08-14', 40, 1000, 0.88),
+    agenticSweep('2026-08-14', 80, 700, 0.93),
+  ];
+
+  it('groups rows that have no ISL/OSL, keyed on the agentic sequence', () => {
+    // The section refused agentic traces on the premise that history could not be
+    // keyed without ISL/OSL. It can: `rowToSequence` keys on benchmark_type.
+    const groups = groupHistoryByHwKeyAndDate({
+      rows: agenticRows,
+      sequence: Sequence.AgenticTraces,
+      precisions: ['fp4'],
+    });
+    expect(groups.datesSeen).toBe(2);
+    const dated = [...groups.byHwKey.values()][0]!;
+    expect(dated).toHaveLength(2);
+    expect(dated[0]!.points).toHaveLength(2);
+    // And the interactivity came from the percentile metric, not median_intvty.
+    expect(dated[0]!.points.map((p) => p.interactivity).toSorted((a, b) => a - b)).toEqual([
+      40, 80,
+    ]);
+  });
+
+  it('reads a cached fraction at the operating point, and none for fixed sequences', () => {
+    const agentic = selectBestFromGroups(
+      groupHistoryByHwKeyAndDate({
+        rows: agenticRows,
+        sequence: Sequence.AgenticTraces,
+        precisions: ['fp4'],
+      }),
+      {
+        targetValue: 60,
+        mode: 'interactivity_to_throughput',
+        costProvider: 'costh',
+        rank: rankByThroughput,
+      },
+    );
+    expect(agentic.best).toHaveLength(1);
+    const rate = agentic.best[0]!.result.cacheHitRate!;
+    expect(rate).toBeGreaterThan(0.88);
+    expect(rate).toBeLessThan(0.95);
+
+    // The same pipeline on fixed-sequence rows reports nothing to discount.
+    const fixed = selectBestFromGroups(
+      groupHistoryByHwKeyAndDate(
+        options([sweep('2026-07-19', 40, 800), sweep('2026-07-19', 80, 500)]),
+      ),
+      {
+        targetValue: 60,
+        mode: 'interactivity_to_throughput',
+        costProvider: 'costh',
+        rank: rankByThroughput,
+      },
+    );
+    expect(fixed.best[0]!.result.cacheHitRate).toBeUndefined();
+  });
+
+  it('still builds a best-so-far staircase when a later date wins', () => {
+    const progressions = bestSoFarProgression(
+      groupHistoryByHwKeyAndDate({
+        rows: agenticRows,
+        sequence: Sequence.AgenticTraces,
+        precisions: ['fp4'],
+      }),
+      {
+        targetValue: 60,
+        mode: 'interactivity_to_throughput',
+        costProvider: 'costh',
+        rank: rankByThroughput,
+      },
+    );
+    expect(progressions).toHaveLength(1);
+    expect(progressions[0]!.steps.map((s) => s.date)).toEqual(['2026-08-07', '2026-08-14']);
+  });
+});
diff --git a/packages/app/src/components/calculator/historical-best.ts b/packages/app/src/components/calculator/historical-best.ts
new file mode 100644
index 000000000..794d8b7c1
--- /dev/null
+++ b/packages/app/src/components/calculator/historical-best.ts
@@ -0,0 +1,484 @@
+/**
+ * All-time best operating point per hwKey — pure, no React.
+ *
+ * The calculator reads one run date: whatever the latest sweep measured is what
+ * a chip is credited with. That understates hardware whose best config was
+ * found earlier and has since been superseded by a sweep exploring a different
+ * part of the space. Measured against production history, ~37% of configs have
+ * their best read at the target on an earlier date, worth up to ~3.6× — so for
+ * a fleet-lifecycle projection, where the question is "what can this chip do?"
+ * rather than "what did last week's sweep do?", the all-time best is the honest
+ * number.
+ *
+ * Two rules make that defensible rather than merely flattering:
+ *
+ * 1. **Clamped reads are discarded.** `interpolateForGPU` always returns a
+ *    value, clamping the target into each frontier's measured range, which at a
+ *    low target credits a sweep's peak throughput at an interactivity it never
+ *    served. Searching every date multiplies the chances of picking up such an
+ *    edge read — at target 20 tok/s/user, 63% of naive winners are clamped — so
+ *    this module applies the no-extrapolation rule instead (the same rule
+ *    `useInterpolatedTrendData` uses) and reports the hwKeys it thereby has no
+ *    read for, rather than silently dropping them.
+ * 2. **Provenance travels with the number.** Every winner carries its date and
+ *    run URL, because the run date stamped above the calculator no longer
+ *    describes it.
+ *
+ * Precisions are pooled into one frontier per hwKey: the question is what the
+ * chip's best config achieves, and precision is part of the config.
+ */
+
+import type { BenchmarkRow } from '@/lib/api';
+import { Percentile, type Sequence } from '@/lib/data-mappings';
+
+import { interpolateForGPU, paretoFrontUpperLeft } from './interpolation';
+import { buildGpuGroups, type GroupMeta } from './useThroughputData';
+import type {
+  CalculatorMode,
+  CostProvider,
+  CostType,
+  GPUDataPoint,
+  InterpolatedResult,
+} from './types';
+
+/** Separator for the `hwKey|date` group key. Dates never contain a pipe. */
+const KEY_SEP = '|';
+
+interface DatedGroupMeta extends GroupMeta {
+  date: string;
+}
+
+export interface HistoricalBestEntry {
+  hwKey: string;
+  /** Run date the winning frontier was measured on. */
+  date: string;
+  /** Run URLs contributing to that date's frontier — usually one. */
+  runUrls: string[];
+  /** The winning read, unclamped by construction. */
+  result: InterpolatedResult;
+  /** Rank value the winner was chosen by, from the caller's `rank`. */
+  rankValue: number;
+  /** Dates with rows for this hwKey. */
+  datesConsidered: number;
+  /** Dates that produced an unclamped read at the target. */
+  datesMeasured: number;
+  /** Latest date with rows, whether or not it won. */
+  latestDate: string;
+  /** True when an earlier date beat the latest one — the feature doing its job. */
+  supersededLatest: boolean;
+  /** Rank value of the latest date's read, when it had an unclamped one. */
+  latestRankValue: number | null;
+}
+
+export interface HistoricalUnmeasured {
+  hwKey: string;
+  /**
+   * Nearest value on each side of the target that some date's Pareto frontier
+   * actually reached, or null where nothing was measured on that side. Not a
+   * range: no single sweep spans the target — that is why this hwKey is here —
+   * so quoting one interval would claim coverage that does not exist.
+   */
+  nearestBelow: number | null;
+  nearestAbove: number | null;
+  datesConsidered: number;
+}
+
+export interface HistoricalBestOutcome {
+  /** Winners, ranked best first. */
+  best: HistoricalBestEntry[];
+  /** hwKeys with data that was never measured at the target. Never silently dropped. */
+  unmeasured: HistoricalUnmeasured[];
+  /** Distinct run dates present in the input. */
+  datesSeen: number;
+}
+
+/** One run date's sweep for one hwKey, ready to interpolate. */
+export interface DatedSweep {
+  date: string;
+  points: GPUDataPoint[];
+  /** Run URLs pooled into this date. */
+  runUrls: string[];
+}
+
+/** History grouped per hwKey per date — the expensive half, independent of the target. */
+export interface HistoryGroups {
+  byHwKey: Map<string, DatedSweep[]>;
+  datesSeen: number;
+}
+
+export interface GroupHistoryOptions {
+  rows: BenchmarkRow[];
+  sequence: Sequence;
+  precisions: string[];
+  /** Agentic percentile; ignored for fixed sequences. */
+  percentile?: Percentile;
+  tokenType?: CostType;
+  /**
+   * Restricts grouping. Callers driving a legend should leave this unset and
+   * filter for display instead, so toggling a legend entry does not rebuild
+   * every frontier.
+   */
+  visibleHwKeys?: Set<string>;
+}
+
+/**
+ * Stage one: bucket raw history rows into one sweep per (hwKey, date).
+ *
+ * Split from selection so a caller can memoize this on the rows and re-run only
+ * the read when the target interactivity moves.
+ */
+export function groupHistoryByHwKeyAndDate(options: GroupHistoryOptions): HistoryGroups {
+  const {
+    rows,
+    sequence,
+    precisions,
+    percentile = Percentile.P90,
+    tokenType = 'total',
+    visibleHwKeys,
+  } = options;
+  if (rows.length === 0 || precisions.length === 0) {
+    return { byHwKey: new Map(), datesSeen: 0 };
+  }
+
+  // Collected during classification so it covers exactly the rows that survive
+  // buildGpuGroups' sequence/precision/hardware filters.
+  const runUrlsByGroup = new Map<string, Set<string>>();
+
+  const { grouped, groupMeta } = buildGpuGroups<DatedGroupMeta>(rows, {
+    sequence,
+    precisions,
+    percentile,
+    tokenType,
+    classify: (hwKey, row) => {
+      if (visibleHwKeys && !visibleHwKeys.has(hwKey)) return null;
+      if (!row.date) return null;
+      const key = `${hwKey}${KEY_SEP}${row.date}`;
+      if (row.run_url) {
+        const urls = runUrlsByGroup.get(key) ?? new Set<string>();
+        urls.add(row.run_url);
+        runUrlsByGroup.set(key, urls);
+      }
+      return { key, meta: { hwKey, date: row.date } };
+    },
+  });
+
+  const byHwKey = new Map<string, DatedSweep[]>();
+  const allDates = new Set<string>();
+  for (const [groupKey, points] of Object.entries(grouped)) {
+    const meta = groupMeta[groupKey];
+    if (!meta || points.length === 0) continue;
+    allDates.add(meta.date);
+    const list = byHwKey.get(meta.hwKey) ?? [];
+    list.push({
+      date: meta.date,
+      points,
+      runUrls: [...(runUrlsByGroup.get(groupKey) ?? [])],
+    });
+    byHwKey.set(meta.hwKey, list);
+  }
+
+  return { byHwKey, datesSeen: allDates.size };
+}
+
+export interface SelectBestOptions {
+  targetValue: number;
+  mode: CalculatorMode;
+  costProvider: CostProvider;
+  /**
+   * Ranks candidate reads. Callers pass the cost-matrix accessor for the
+   * selected token type — `(r) => getTpPerMwForType(r, costType)` — so the
+   * winner is chosen on the same basis the fleet is sized by.
+   */
+  rank: (result: InterpolatedResult) => number;
+}
+
+/**
+ * Stage two: read every date's frontier at the target and keep each hwKey's
+ * best unclamped result.
+ */
+export function selectBestFromGroups(
+  groups: HistoryGroups,
+  options: SelectBestOptions,
+): HistoricalBestOutcome {
+  const { targetValue, mode, costProvider, rank } = options;
+  const { byHwKey, datesSeen } = groups;
+
+  const getInputValue = (p: GPUDataPoint) =>
+    mode === 'interactivity_to_throughput' ? p.interactivity : p.throughput;
+  // The frontier's other axis. Only the unmeasured disclosure below needs it,
+  // to rebuild the same Pareto front `interpolateForGPU` read from.
+  const getOutputValue = (p: GPUDataPoint) =>
+    mode === 'interactivity_to_throughput' ? p.throughput : p.interactivity;
+
+  const best: HistoricalBestEntry[] = [];
+  const unmeasured: HistoricalUnmeasured[] = [];
+
+  for (const [hwKey, dated] of byHwKey) {
+    const latestDate = dated.reduce((a, b) => (b.date > a ? b.date : a), dated[0]!.date);
+
+    let winner: { sweep: DatedSweep; result: InterpolatedResult; rankValue: number } | null = null;
+    let latestRankValue: number | null = null;
+    let datesMeasured = 0;
+
+    for (const sweep of dated) {
+      const result = interpolateForGPU(sweep.points, targetValue, mode, costProvider);
+      // The no-extrapolation rule: a clamped read is the frontier's nearest
+      // edge, not a measurement at the target, and must not win.
+      if (!result || result.clamped || !(result.value > 0)) continue;
+
+      datesMeasured += 1;
+      const rankValue = rank(result);
+      if (!Number.isFinite(rankValue)) continue;
+      if (sweep.date === latestDate) latestRankValue = rankValue;
+      if (!winner || rankValue > winner.rankValue) winner = { sweep, result, rankValue };
+    }
+
+    if (!winner) {
+      // What rejected this hwKey was per-date and per-frontier: no single sweep's
+      // Pareto front spans the target. So the disclosure has to be per-date and
+      // per-frontier too. Taking the union of every raw point across every date
+      // — which this did — yields an interval no sweep ever covered, and one that
+      // can contain the very target the chip was just excluded at: a reader sees
+      // "measured 6.7-18.5" next to a target of 8 and concludes the exclusion is
+      // broken. Two dates bracketing a gap is enough to produce that, and so is a
+      // frontier narrower than its own raw points.
+      //
+      // Reported instead: the nearest frontier value on each side of the target.
+      // That is the honest shape of the miss — how far the target sits from
+      // anything measured — and it cannot imply coverage, because a frontier
+      // value on both sides of the target within one date would have made the
+      // read succeed.
+      let nearestBelow: number | null = null;
+      let nearestAbove: number | null = null;
+      for (const { points } of dated) {
+        const front = paretoFrontUpperLeft([...points], getInputValue, getOutputValue);
+        for (const p of front) {
+          const v = getInputValue(p);
+          if (!Number.isFinite(v)) continue;
+          if (v <= targetValue && (nearestBelow === null || v > nearestBelow)) nearestBelow = v;
+          if (v >= targetValue && (nearestAbove === null || v < nearestAbove)) nearestAbove = v;
+        }
+      }
+      unmeasured.push({
+        hwKey,
+        nearestBelow,
+        nearestAbove,
+        datesConsidered: dated.length,
+      });
+      continue;
+    }
+
+    best.push({
+      hwKey,
+      date: winner.sweep.date,
+      runUrls: winner.sweep.runUrls,
+      // The result's own resultKey is the dated group key; the caller keys on
+      // hwKey, so restate it here.
+      result: { ...winner.result, hwKey, resultKey: hwKey },
+      rankValue: winner.rankValue,
+      datesConsidered: dated.length,
+      datesMeasured,
+      latestDate,
+      supersededLatest: winner.sweep.date !== latestDate,
+      latestRankValue,
+    });
+  }
+
+  best.sort((a, b) => b.rankValue - a.rankValue || a.hwKey.localeCompare(b.hwKey));
+  unmeasured.sort((a, b) => a.hwKey.localeCompare(b.hwKey));
+
+  return { best, unmeasured, datesSeen };
+}
+
+/** Convenience composition of both stages, for callers with no memoization needs. */
+export function selectHistoricalBest(
+  options: GroupHistoryOptions & SelectBestOptions,
+): HistoricalBestOutcome {
+  return selectBestFromGroups(groupHistoryByHwKeyAndDate(options), options);
+}
+
+/** One rung of a chip's best-so-far progression. */
+export interface ProgressionStep {
+  /** Run date this config was measured on. */
+  date: string;
+  /** Run URLs pooled into that date. */
+  runUrls: string[];
+  /** The read at this date — the new best-so-far. */
+  result: InterpolatedResult;
+  /** Rank value of this read. Strictly greater than the previous rung's. */
+  rankValue: number;
+  /** rankValue ÷ the first rung's, i.e. gain over the opening config. */
+  factorOverFirst: number;
+}
+
+export interface HistoricalProgression {
+  hwKey: string;
+  /** Chronological, strictly improving. First rung is the opening measurement. */
+  steps: ProgressionStep[];
+  /** Dates that produced an unclamped read, improving or not. */
+  datesMeasured: number;
+  datesConsidered: number;
+}
+
+/**
+ * Each hwKey's best-so-far progression at the target, in calendar order.
+ *
+ * `selectBestFromGroups` answers "what is the best this chip has ever done?".
+ * This answers "how did it get there?" — the running maximum over run dates,
+ * keeping only the dates that improved on everything before them. That staircase
+ * is what a fixed fleet's revenue actually followed: the chips never changed,
+ * the software serving them did.
+ *
+ * Same two rules as the best-of selection: clamped reads never count, and every
+ * rung keeps its date and run URLs so the step can be traced to the sweep that
+ * produced it.
+ */
+export function bestSoFarProgression(
+  groups: HistoryGroups,
+  options: SelectBestOptions,
+): HistoricalProgression[] {
+  const { targetValue, mode, costProvider, rank } = options;
+  const progressions: HistoricalProgression[] = [];
+
+  for (const [hwKey, dated] of groups.byHwKey) {
+    const chronological = [...dated].toSorted((a, b) => a.date.localeCompare(b.date));
+    const steps: ProgressionStep[] = [];
+    let best = -Infinity;
+    let datesMeasured = 0;
+
+    for (const sweep of chronological) {
+      const result = interpolateForGPU(sweep.points, targetValue, mode, costProvider);
+      if (!result || result.clamped || !(result.value > 0)) continue;
+      const rankValue = rank(result);
+      if (!Number.isFinite(rankValue) || rankValue <= 0) continue;
+
+      datesMeasured += 1;
+      // Only rungs: a sweep that failed to beat the incumbent leaves the fleet
+      // serving the config it already had, so it is not a step in the line.
+      if (rankValue <= best) continue;
+      best = rankValue;
+      steps.push({
+        date: sweep.date,
+        runUrls: sweep.runUrls,
+        result: { ...result, hwKey, resultKey: hwKey },
+        rankValue,
+        // Filled in below, once the opening rung is known.
+        factorOverFirst: 1,
+      });
+    }
+
+    if (steps.length === 0) continue;
+    const firstRank = steps[0]!.rankValue;
+    for (const step of steps) step.factorOverFirst = step.rankValue / firstRank;
+
+    progressions.push({
+      hwKey,
+      steps,
+      datesMeasured,
+      datesConsidered: dated.length,
+    });
+  }
+
+  progressions.sort(
+    (a, b) =>
+      (b.steps.at(-1)?.rankValue ?? 0) - (a.steps.at(-1)?.rankValue ?? 0) ||
+      a.hwKey.localeCompare(b.hwKey),
+  );
+  return progressions;
+}
+
+/** True when a read came from disaggregated points, whose fleet math differs. */
+function resultIsDisagg(result: InterpolatedResult): boolean {
+  return result.nearestPoints.some((p) => p.disagg);
+}
+
+export interface ChipProgression {
+  /** Group key: the base GPU. Stable across steps. */
+  key: string;
+  /** Base GPU registry key, e.g. `b200` — the silicon the fleet is built from. */
+  baseGpu: string;
+  /** True when any rung came from a disaggregated run. */
+  disagg: boolean;
+  /**
+   * The pooled running maximum. Each rung carries its own `result`, so the hwKey
+   * serving the fleet can differ from one rung to the next.
+   */
+  steps: ProgressionStep[];
+  /** Every hwKey that contributed a rung, in the order they took over. */
+  hwKeysUsed: string[];
+}
+
+/**
+ * Collapse per-hwKey progressions into one per chip — the best way to run that
+ * silicon at any given time.
+ *
+ * A single piece of hardware appears in the history under many hwKeys, because
+ * hwKey encodes the software: `b200_trtllm`, `b200_sglang`, `b200_sglang_mtp`
+ * and so on are all one B200. Drawing a line each says a fleet operator ran
+ * seven fleets, when they ran one and kept re-deploying it onto whichever config
+ * was ahead. So the fleet's line is the upper envelope over that chip's hwKeys,
+ * and a rung's winning config is part of what the rung reports.
+ *
+ * Because each input progression is already a running maximum, the pointwise
+ * maximum over them is the running maximum of their union — so this merges by
+ * walking the rungs in date order and keeping the ones that beat the incumbent.
+ *
+ * **Disagg competes alongside aggregated configs.** Disaggregated runs report
+ * throughput per decode or per prefill chip rather than per total chip, so a rung
+ * won by a disagg config is not sized on quite the same basis as one won by an
+ * aggregated config. They are pooled anyway — the operator's question is what the
+ * silicon can be made to do — and the caveat travels with the rung that carries
+ * it: `disagg` flags a chip whose progression involves any, and the config named
+ * on each step says which kind won it.
+ */
+export function mergeProgressionsByChip(
+  progressions: readonly HistoricalProgression[],
+): ChipProgression[] {
+  const byChip = new Map<string, { baseGpu: string; steps: ProgressionStep[] }>();
+
+  for (const progression of progressions) {
+    const baseGpu = progression.hwKey.split('_')[0] ?? progression.hwKey;
+    const entry = byChip.get(baseGpu) ?? { baseGpu, steps: [] };
+    entry.steps.push(...progression.steps);
+    byChip.set(baseGpu, entry);
+  }
+
+  const chips: ChipProgression[] = [];
+  for (const [key, { baseGpu, steps: pooled }] of byChip) {
+    // Date order, and within a date the stronger read first so it is the one
+    // that becomes the rung.
+    const candidates = [...pooled].toSorted(
+      (a, b) => a.date.localeCompare(b.date) || b.rankValue - a.rankValue,
+    );
+
+    const steps: ProgressionStep[] = [];
+    const hwKeysUsed: string[] = [];
+    let best = -Infinity;
+    for (const candidate of candidates) {
+      if (candidate.rankValue <= best) continue;
+      best = candidate.rankValue;
+      steps.push(candidate);
+      const hwKey = candidate.result.hwKey;
+      if (hwKey && hwKeysUsed.at(-1) !== hwKey) hwKeysUsed.push(hwKey);
+    }
+
+    if (steps.length === 0) continue;
+    const firstRank = steps[0]!.rankValue;
+    // factorOverFirst is relative to the merged opening rung, which is not
+    // necessarily the opening rung of the hwKey that rung came from.
+    chips.push({
+      key,
+      baseGpu,
+      disagg: steps.some((s) => resultIsDisagg(s.result)),
+      steps: steps.map((s) => ({ ...s, factorOverFirst: s.rankValue / firstRank })),
+      hwKeysUsed,
+    });
+  }
+
+  chips.sort(
+    (a, b) =>
+      (b.steps.at(-1)?.rankValue ?? 0) - (a.steps.at(-1)?.rankValue ?? 0) ||
+      a.key.localeCompare(b.key),
+  );
+  return chips;
+}
diff --git a/packages/app/src/components/calculator/interpolation.ts b/packages/app/src/components/calculator/interpolation.ts
index 4a408450a..f0e4459cb 100644
--- a/packages/app/src/components/calculator/interpolation.ts
+++ b/packages/app/src/components/calculator/interpolation.ts
@@ -393,6 +393,8 @@ export function interpolateForGPU(
       tpPerMw: sorted[0].tpPerMw,
       inputTpPerMw: sorted[0].inputTpPerMw,
       outputTpPerMw: sorted[0].outputTpPerMw,
+      cacheHitRate: sorted[0].cacheHitRate,
+      inputTokenShare: sorted[0].inputTokenShare,
       concurrency: sorted[0].concurrency,
       nearestPoints: [sorted[0]],
       clamped,
@@ -448,6 +450,22 @@ export function interpolateForGPU(
   const outputTpPerMw = buildMetric((p) => p.outputTpPerMw);
   const concurrency = Math.round(buildMetric((p) => p.concurrency));
 
+  // Only splined when every frontier point carries a measured rate. All absent
+  // is the fixed-sequence case, where there is no cache to discount. A *mixed*
+  // frontier is the interesting one: substituting 0 for the missing points
+  // would invent a dip in the cached fraction, so the whole sweep opts out and
+  // its input tokens bill at full price — understating margin rather than
+  // overstating it, which is the safe direction for this number to be wrong in.
+  const cacheHitRate = sorted.every((p) => typeof p.cacheHitRate === 'number')
+    ? buildMetric((p) => p.cacheHitRate!)
+    : undefined;
+
+  // Same all-or-nothing rule as the cached fraction: a frontier only partly
+  // pinned down opts out rather than having a guessed share splined into it.
+  const inputTokenShare = sorted.every((p) => typeof p.inputTokenShare === 'number')
+    ? buildMetric((p) => p.inputTokenShare!)
+    : undefined;
+
   let nearestPoints: GPUDataPoint[];
   if (clampedTarget <= minInput) {
     nearestPoints = [sorted[0]];
@@ -474,6 +492,8 @@ export function interpolateForGPU(
     tpPerMw,
     inputTpPerMw,
     outputTpPerMw,
+    cacheHitRate,
+    inputTokenShare,
     concurrency,
     nearestPoints,
     clamped,
diff --git a/packages/app/src/components/calculator/lifecycle.test.ts b/packages/app/src/components/calculator/lifecycle.test.ts
new file mode 100644
index 000000000..b88e53323
--- /dev/null
+++ b/packages/app/src/components/calculator/lifecycle.test.ts
@@ -0,0 +1,918 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  availabilityFromInterrupts,
+  billableInputRate,
+  effectiveTokPerSec,
+  isBreakEvenAnchored,
+  outputTokPerChip,
+  splitTokenStreams,
+  metricValue,
+  breakEvenPricePerMTok,
+  computeLifecycle,
+  valueAtMonth,
+  type LifecycleAssumptions,
+  type LifecyclePoint,
+  type ThroughputStep,
+} from './lifecycle';
+
+const HOURS_PER_DAY = 24;
+const DAYS_PER_MONTH = 365.25 / 12;
+
+/** A 10 MW fleet, so $/MW/day is the daily margin over ten. */
+const PROVISIONED_MW = 10;
+
+/** Hand-computed revenue $/day at the fixture's price and availability. */
+const rev = (tput: number) => ((tput * 86_400) / 1e6) * 10 * (24 / 24.5);
+
+const assumptions: LifecycleAssumptions = {
+  mtbiDays: 24,
+  recoveryHours: 12,
+  // Comfortably above the fixture's break-even, so the base case pays back.
+  inputPricePerMTok: 10,
+  outputPricePerMTok: 0,
+  // The step behaviour is what most of these tests are about, so they start at
+  // full load. The ramp has its own describe block below.
+  rampMonths: 0,
+};
+
+describe('availabilityFromInterrupts', () => {
+  it('applies the mtbi / (mtbi + recovery) haircut', () => {
+    expect(availabilityFromInterrupts(24, 12)).toBeCloseTo(24 / 24.5, 10);
+  });
+
+  it('is monotonic in both directions', () => {
+    expect(availabilityFromInterrupts(48, 12)).toBeGreaterThan(availabilityFromInterrupts(24, 12));
+    expect(availabilityFromInterrupts(24, 6)).toBeGreaterThan(availabilityFromInterrupts(24, 12));
+  });
+
+  it('treats a blank or nonsensical input as "no interruptions modelled"', () => {
+    // A missing MTBI must not read as "always down" — it is an optional refinement.
+    expect(availabilityFromInterrupts(0, 12)).toBe(1);
+    expect(availabilityFromInterrupts(-1, 12)).toBe(1);
+    expect(availabilityFromInterrupts(NaN, 12)).toBe(1);
+    expect(availabilityFromInterrupts(24, 0)).toBe(1);
+  });
+});
+
+describe('breakEvenPricePerMTok', () => {
+  it('returns the price at which revenue exactly covers cost', () => {
+    // 1000 tok/s = 3.6M tok/hr; $36/hr → $10 per M tok
+    expect(breakEvenPricePerMTok(36, 1000)).toBeCloseTo(10, 10);
+  });
+
+  it('is the price that zeroes the margin at that throughput', () => {
+    const tokPerSec = 1_000_000;
+    const costPerHour = 12_345;
+    const price = breakEvenPricePerMTok(costPerHour, tokPerSec)!;
+    const series = computeLifecycle({
+      steps: [{ month: 0, billableInputTokPerSec: tokPerSec, outputTokPerSec: 0 }],
+      costPerHour,
+      provisionedMw: PROVISIONED_MW,
+      horizonMonths: 24,
+      assumptions: { ...assumptions, mtbiDays: 0, inputPricePerMTok: price, outputPricePerMTok: 0 },
+    })!;
+    expect(series.marginPerDay).toBeCloseTo(0, 6);
+  });
+
+  it('rises by the interrupt haircut, and still zeroes the margin', () => {
+    // The same racks sell fewer tokens when interrupted, so break-even is higher.
+    // Without the availability argument the seeded "break-even" price leaves the
+    // cheapest fleet a few percent under water.
+    const tokPerSec = 1_000_000;
+    const costPerHour = 12_345;
+    const availability = availabilityFromInterrupts(24, 12);
+    const flat = breakEvenPricePerMTok(costPerHour, tokPerSec)!;
+    const haircut = breakEvenPricePerMTok(costPerHour, tokPerSec, availability)!;
+    expect(haircut).toBeCloseTo(flat / availability, 10);
+
+    const series = computeLifecycle({
+      steps: [{ month: 0, billableInputTokPerSec: tokPerSec, outputTokPerSec: 0 }],
+      costPerHour,
+      provisionedMw: PROVISIONED_MW,
+      horizonMonths: 24,
+      assumptions: { ...assumptions, inputPricePerMTok: haircut, outputPricePerMTok: 0 },
+    })!;
+    expect(series.availability).toBeCloseTo(availability, 10);
+    expect(series.marginPerDay).toBeCloseTo(0, 6);
+
+    // The unadjusted price is the defect being pinned: it reads as break-even
+    // but plots a loss.
+    const naive = computeLifecycle({
+      steps: [{ month: 0, billableInputTokPerSec: tokPerSec, outputTokPerSec: 0 }],
+      costPerHour,
+      provisionedMw: PROVISIONED_MW,
+      horizonMonths: 24,
+      assumptions: { ...assumptions, inputPricePerMTok: flat, outputPricePerMTok: 0 },
+    })!;
+    expect(naive.marginPerDay).toBeLessThan(0);
+  });
+
+  it('returns null when no tokens are produced', () => {
+    expect(breakEvenPricePerMTok(100, 0)).toBeNull();
+    expect(breakEvenPricePerMTok(100, NaN)).toBeNull();
+  });
+
+  it('returns null for an impossible availability', () => {
+    expect(breakEvenPricePerMTok(100, 1000, 0)).toBeNull();
+    expect(breakEvenPricePerMTok(100, 1000, NaN)).toBeNull();
+  });
+});
+
+describe('computeLifecycle', () => {
+  const costPerHour = 10_000;
+  // Three measured configs: the opening sweep, then two improvements — the
+  // shape MI355X showed on DeepSeek-V4-Pro.
+  const steps: ThroughputStep[] = [
+    { month: 0, billableInputTokPerSec: 400_000, outputTokPerSec: 0 },
+    { month: 3, billableInputTokPerSec: 900_000, outputTokPerSec: 0 },
+    { month: 6, billableInputTokPerSec: 1_600_000, outputTokPerSec: 0 },
+  ];
+  const base = {
+    steps,
+    costPerHour,
+    provisionedMw: PROVISIONED_MW,
+    horizonMonths: 24,
+    assumptions,
+  };
+
+  it('holds each config flat until the next one lands', () => {
+    const series = computeLifecycle(base)!;
+    // One point per step plus a closing point at the horizon: with curveStepAfter
+    // a value holds until the next x, so that is the whole staircase.
+    const risers = series.points.filter((p) => p.isStep);
+    expect(risers.map((p) => p.month)).toEqual([0, 3, 6]);
+    expect(risers[1]!.revenue).toBeGreaterThan(risers[0]!.revenue);
+    expect(risers[2]!.revenue).toBeGreaterThan(risers[1]!.revenue);
+    // The closing point carries the last config forward to the horizon.
+    const last = series.points.at(-1)!;
+    expect(last.isStep).toBe(false);
+    expect(last.month).toBe(24);
+    expect(last.revenue).toBeCloseTo(risers[2]!.revenue, 6);
+  });
+
+  it('keeps cost flat — a better config does not change chips or their price', () => {
+    const series = computeLifecycle(base)!;
+    for (const p of series.points) {
+      expect(p.cost).toBeCloseTo(costPerHour * HOURS_PER_DAY, 6);
+    }
+  });
+
+  it('reports the gain the measured improvements were worth', () => {
+    const series = computeLifecycle(base)!;
+    expect(series.improvementFactor).toBeCloseTo(1_600_000 / 400_000, 9);
+    expect(series.improvementCount).toBe(2);
+    expect(series.revenuePerDay).toBeGreaterThan(series.firstRevenuePerDay);
+    expect(series.marginPerDay).toBeGreaterThan(series.firstMarginPerDay);
+  });
+
+  it('scales revenue with throughput, price and availability', () => {
+    const series = computeLifecycle(base)!;
+    const expected = ((1_600_000 * 86_400) / 1e6) * assumptions.inputPricePerMTok * (24 / 24.5);
+    expect(series.revenuePerDay).toBeCloseTo(expected, 6);
+    expect(series.availability).toBeCloseTo(24 / 24.5, 10);
+  });
+
+  it('holds the latest config flat past the last sweep, out to the horizon', () => {
+    const series = computeLifecycle(base)!;
+    expect(series.endMonth).toBe(24);
+    const last = series.points.at(-1)!;
+    expect(last.month).toBe(24);
+    expect(last.revenue).toBeCloseTo(series.revenuePerDay, 6);
+  });
+
+  it('starts at the first measured run, not at the anchor', () => {
+    // A chip first measured three months after release has no line before then.
+    const late = computeLifecycle({
+      ...base,
+      steps: [{ month: 3, billableInputTokPerSec: 400_000, outputTokPerSec: 0 }],
+    })!;
+    expect(late.startMonth).toBe(3);
+    expect(late.points[0]!.month).toBe(3);
+  });
+
+  it('integrates cumulative margin across the steps', () => {
+    const series = computeLifecycle(base)!;
+    // Hand-integrate: three stretches at their own margins.
+    const cost = costPerHour * HOURS_PER_DAY;
+    const expected =
+      (rev(400_000) - cost) * 3 * DAYS_PER_MONTH +
+      (rev(900_000) - cost) * 3 * DAYS_PER_MONTH +
+      (rev(1_600_000) - cost) * 18 * DAYS_PER_MONTH;
+    expect(series.lifetimeMargin).toBeCloseTo(expected, 4);
+  });
+
+  it('reports payback within the stretch where cumulative crosses zero', () => {
+    const series = computeLifecycle(base)!;
+    expect(series.paybackMonth).not.toBeNull();
+    expect(series.paybackMonth!).toBeGreaterThanOrEqual(0);
+    expect(series.paybackMonth!).toBeLessThanOrEqual(24);
+  });
+
+  it('reports no payback for a fleet that never covers its cost', () => {
+    const series = computeLifecycle({
+      ...base,
+      assumptions: { ...assumptions, inputPricePerMTok: 0, outputPricePerMTok: 0 },
+    })!;
+    expect(series.paybackMonth).toBeNull();
+    expect(series.lifetimeMargin).toBeLessThan(0);
+  });
+
+  it('can start underwater and end profitable — the point of the chart', () => {
+    // Priced so the opening config loses money and the last one makes it.
+    const price = breakEvenPricePerMTok(costPerHour, 700_000)!;
+    const series = computeLifecycle({
+      ...base,
+      assumptions: { ...assumptions, mtbiDays: 0, inputPricePerMTok: price, outputPricePerMTok: 0 },
+    })!;
+    const risers = series.points.filter((p) => p.isStep);
+    expect(risers[0]!.margin).toBeLessThan(0);
+    expect(risers.at(-1)!.margin).toBeGreaterThan(0);
+  });
+
+  it('ignores steps beyond the horizon', () => {
+    const series = computeLifecycle({ ...base, horizonMonths: 4 })!;
+    // Only the first two configs land inside a four-month window.
+    expect(series.points.filter((p) => p.isStep).map((p) => p.month)).toEqual([0, 3]);
+    expect(series.improvementCount).toBe(1);
+    expect(series.endMonth).toBe(4);
+  });
+
+  it('sorts and cleans the schedule it is given', () => {
+    const series = computeLifecycle({
+      ...base,
+      steps: [
+        { month: 6, billableInputTokPerSec: 1_600_000, outputTokPerSec: 0 },
+        { month: 0, billableInputTokPerSec: 400_000, outputTokPerSec: 0 },
+        { month: 3, billableInputTokPerSec: 0, outputTokPerSec: 0 }, // no throughput — dropped
+        { month: NaN, billableInputTokPerSec: 500_000, outputTokPerSec: 0 }, // unusable month — dropped
+      ],
+    })!;
+    expect(series.points.filter((p) => p.isStep).map((p) => p.month)).toEqual([0, 6]);
+  });
+
+  describe('ramp', () => {
+    const ramped = { ...base, assumptions: { ...assumptions, rampMonths: 3 } };
+
+    it('starts the first config at zero revenue and maxes out at its numbers', () => {
+      const series = computeLifecycle(ramped)!;
+      expect(series.rampEndMonth).toBe(3);
+      const first = series.points[0]!;
+      // Nothing is being served yet, but the racks are already billing.
+      expect(first.revenue).toBeCloseTo(0, 6);
+      expect(first.margin).toBeCloseTo(-costPerHour * HOURS_PER_DAY, 6);
+      // At the end of its rollout the first config is serving its full rate.
+      const atRampEnd = series.points.find((p) => p.month === 3)!;
+      expect(atRampEnd.revenue).toBeCloseTo(rev(400_000), 6);
+    });
+
+    it('gives every config its own rollout, from the incumbent level to its own', () => {
+      const series = computeLifecycle(ramped)!;
+      const risers = series.points.filter((p) => p.isStep);
+      expect(risers.map((p) => p.month)).toEqual([0, 3, 6]);
+      // The second config starts from the first's level, not from zero.
+      expect(risers[1]!.revenue).toBeCloseTo(rev(400_000), 6);
+      // ...and maxes out at its own by the end of its window.
+      const atSix = series.points.find((p) => p.month === 6)!;
+      expect(atSix.revenue).toBeCloseTo(rev(900_000), 6);
+      const atNine = series.points.find((p) => p.month === 9)!;
+      expect(atNine.revenue).toBeCloseTo(rev(1_600_000), 6);
+    });
+
+    it('rises monotonically, with no jump at a rollout boundary', () => {
+      const series = computeLifecycle(ramped)!;
+      for (let i = 1; i < series.points.length; i += 1) {
+        expect(series.points[i]!.revenue).toBeGreaterThanOrEqual(
+          series.points[i - 1]!.revenue - 1e-6,
+        );
+      }
+    });
+
+    it('starts a rollout from where the previous one actually got to', () => {
+      // A 12-month ramp cannot finish before the next config lands at month 3, so
+      // the second rollout must pick up mid-climb rather than from 400k.
+      const series = computeLifecycle({
+        ...base,
+        assumptions: { ...assumptions, rampMonths: 12 },
+      })!;
+      const risers = series.points.filter((p) => p.isStep);
+      const second = risers[1]!;
+      expect(second.month).toBe(3);
+      expect(second.revenue).toBeGreaterThan(0);
+      expect(second.revenue).toBeLessThan(rev(400_000));
+    });
+
+    it('keeps cost flat through every rollout — racks bill once energised', () => {
+      const series = computeLifecycle(ramped)!;
+      const full = costPerHour * HOURS_PER_DAY;
+      expect(series.points.length).toBeGreaterThan(20);
+      for (const p of series.points) expect(p.cost).toBeCloseTo(full, 6);
+    });
+
+    it('delays payback, because the first rollout is paid for while it earns its way up', () => {
+      const flat = computeLifecycle(base)!;
+      const series = computeLifecycle(ramped)!;
+      expect(series.paybackMonth).not.toBeNull();
+      expect(series.paybackMonth!).toBeGreaterThan(flat.paybackMonth!);
+      expect(series.lifetimeMargin).toBeLessThan(flat.lifetimeMargin);
+    });
+
+    it('treats a zero or nonsensical ramp as "already at full load"', () => {
+      for (const rampMonths of [0, -1, NaN]) {
+        const series = computeLifecycle({
+          ...base,
+          assumptions: { ...assumptions, rampMonths },
+        })!;
+        expect(series.rampEndMonth).toBe(series.startMonth);
+        expect(series.points[0]!.revenue).toBeCloseTo(rev(400_000), 6);
+        expect(series.points.some((p) => p.isRamp)).toBe(false);
+      }
+    });
+
+    it('reports the end-of-window rate, which is past every rollout', () => {
+      const series = computeLifecycle(ramped)!;
+      expect(series.revenuePerDay).toBeCloseTo(rev(1_600_000), 6);
+    });
+
+    it('integrates a rollout to the analytic area under the smoothstep', () => {
+      // A full smoothstep is odd-symmetric about its midpoint, so its mean is
+      // exactly half the end level. Pins the trapezoid path, which every other
+      // integration test dodges by using rampMonths: 0.
+      const series = computeLifecycle({
+        steps: [{ month: 0, billableInputTokPerSec: 1_600_000, outputTokPerSec: 0 }],
+        costPerHour,
+        provisionedMw: PROVISIONED_MW,
+        horizonMonths: 6,
+        assumptions: { ...assumptions, rampMonths: 6 },
+      })!;
+      const cost = costPerHour * HOURS_PER_DAY;
+      const expected = (rev(1_600_000) * 0.5 - cost) * 6 * DAYS_PER_MONTH;
+      expect(series.lifetimeMargin).toBeCloseTo(expected, 4);
+    });
+
+    it('still flags a rollout as ramping when the next config cuts it short', () => {
+      // The last sample of a pre-empted rollout is mid-climb, so it must not
+      // read as "at full load".
+      const series = computeLifecycle({
+        steps: [
+          { month: 0, billableInputTokPerSec: 400_000, outputTokPerSec: 0 },
+          { month: 3, billableInputTokPerSec: 1_600_000, outputTokPerSec: 0 },
+        ],
+        costPerHour,
+        provisionedMw: PROVISIONED_MW,
+        horizonMonths: 24,
+        assumptions: { ...assumptions, rampMonths: 12 },
+      })!;
+      const cutShort = series.points.filter((p) => p.month === 3 && !p.isStep);
+      expect(cutShort.length).toBeGreaterThan(0);
+      for (const point of cutShort) {
+        expect(point.revenue).toBeLessThan(rev(400_000));
+        expect(point.isRamp).toBe(true);
+      }
+    });
+
+    it('collapses configs landing in the same month, keeping the best', () => {
+      // Same-day rungs share a month. The superseded one must not get a
+      // zero-width rollout, which would spike the line to a rate never served.
+      const series = computeLifecycle({
+        steps: [
+          { month: 0, billableInputTokPerSec: 400_000, outputTokPerSec: 0 },
+          { month: 6, billableInputTokPerSec: 900_000, outputTokPerSec: 0 },
+          { month: 6, billableInputTokPerSec: 1_600_000, outputTokPerSec: 0 },
+        ],
+        costPerHour,
+        provisionedMw: PROVISIONED_MW,
+        horizonMonths: 24,
+        assumptions: { ...assumptions, rampMonths: 3 },
+      })!;
+      // One riser per month, not two.
+      expect(series.points.filter((p) => p.isStep).map((p) => p.month)).toEqual([0, 6]);
+      expect(series.improvementCount).toBe(1);
+      // The superseded config's full rate is never plotted, and the level only
+      // ever climbs.
+      for (const point of series.points) {
+        expect(point.revenue).toBeLessThanOrEqual(rev(1_600_000) + 1e-6);
+      }
+      for (let i = 1; i < series.points.length; i += 1) {
+        expect(series.points[i]!.revenue).toBeGreaterThanOrEqual(
+          series.points[i - 1]!.revenue - 1e-6,
+        );
+      }
+    });
+  });
+
+  it('interpolates payback inside the stretch that crosses zero', () => {
+    // Priced so the opening config loses money for three months and the next one
+    // digs the fleet out. The crossing is then closed-form, which pins the
+    // interpolation rather than just bracketing it.
+    const price = 3;
+    const series = computeLifecycle({
+      ...base,
+      steps: [
+        { month: 0, billableInputTokPerSec: 400_000, outputTokPerSec: 0 },
+        { month: 3, billableInputTokPerSec: 1_600_000, outputTokPerSec: 0 },
+      ],
+      assumptions: { ...assumptions, mtbiDays: 0, inputPricePerMTok: price, outputPricePerMTok: 0 },
+    })!;
+    const cost = costPerHour * HOURS_PER_DAY;
+    const perDay = (tput: number) => ((tput * 86_400) / 1e6) * price;
+    const first = perDay(400_000) - cost;
+    const second = perDay(1_600_000) - cost;
+    expect(first).toBeLessThan(0);
+    expect(second).toBeGreaterThan(0);
+    // Three months of losses repaid at the second config's rate.
+    expect(series.paybackMonth!).toBeCloseTo(3 + (-first * 3) / second, 6);
+  });
+
+  describe('cumulative revenue', () => {
+    it('is the running area under the revenue curve, cost never subtracted', () => {
+      const series = computeLifecycle(base)!;
+      const first = series.points[0]!;
+      const last = series.points.at(-1)!;
+      // Nothing has been earned at the first sample, and it only ever grows —
+      // unlike cumulative margin, which starts underwater whenever price is low.
+      expect(first.cumulativeRevenue).toBe(0);
+      for (const [i, point] of series.points.entries()) {
+        if (i === 0) continue;
+        expect(point.cumulativeRevenue).toBeGreaterThanOrEqual(
+          series.points[i - 1]!.cumulativeRevenue,
+        );
+      }
+      expect(last.cumulativeRevenue).toBeGreaterThan(0);
+    });
+
+    it('stays exactly one flat cost above cumulative margin', () => {
+      // The invariant that makes the two comparable on one chart: both are
+      // trapezoided over the same intervals, so their difference is cost x elapsed
+      // and nothing else. A separate accumulator that drifted would fail here.
+      const series = computeLifecycle(base)!;
+      const costPerDay = costPerHour * HOURS_PER_DAY;
+      for (const point of series.points) {
+        const elapsedDays = (point.month - series.startMonth) * DAYS_PER_MONTH;
+        expect(point.cumulativeRevenue - point.cumulative).toBeCloseTo(costPerDay * elapsedDays, 4);
+      }
+    });
+
+    it('is zero at every sample when the price is zero', () => {
+      const series = computeLifecycle({
+        ...base,
+        assumptions: { ...assumptions, inputPricePerMTok: 0, outputPricePerMTok: 0 },
+      })!;
+      for (const point of series.points) expect(point.cumulativeRevenue).toBe(0);
+      // Cumulative margin is then pure accumulated cost, which is the check that
+      // the two accumulators are not accidentally the same variable.
+      expect(series.points.at(-1)!.cumulative).toBeLessThan(0);
+    });
+  });
+
+  it('returns null without steps, a cost basis or a horizon past the first run', () => {
+    expect(computeLifecycle({ ...base, steps: [] })).toBeNull();
+    expect(computeLifecycle({ ...base, costPerHour: NaN })).toBeNull();
+    expect(computeLifecycle({ ...base, horizonMonths: 0 })).toBeNull();
+    expect(
+      computeLifecycle({
+        ...base,
+        steps: [{ month: 30, billableInputTokPerSec: 1, outputTokPerSec: 0 }],
+        horizonMonths: 24,
+      }),
+    ).toBeNull();
+  });
+});
+
+const point = (month: number, margin: number): LifecyclePoint => ({
+  month,
+  revenue: margin + 100,
+  cost: 100,
+  margin,
+  revenuePerMw: (margin + 100) / PROVISIONED_MW,
+  marginPerMw: margin / PROVISIONED_MW,
+  cumulative: margin * 10,
+  cumulativeRevenue: (margin + 100) * 10,
+  isStep: false,
+  isRamp: false,
+});
+
+const margin = (p: LifecyclePoint) => p.margin;
+
+describe('valueAtMonth', () => {
+  const points = [point(2, 0), point(4, 200), point(9, 200)];
+
+  it('interpolates linearly between samples, matching what the line draws', () => {
+    expect(valueAtMonth(points, 3, margin)).toBeCloseTo(100, 9);
+    expect(valueAtMonth(points, 2, margin)).toBeCloseTo(0, 9);
+    expect(valueAtMonth(points, 4, margin)).toBeCloseTo(200, 9);
+    expect(valueAtMonth(points, 6.5, margin)).toBeCloseTo(200, 9);
+  });
+
+  it('returns null outside the series window instead of clamping to its ends', () => {
+    // A chip first measured a year after release has no line before then, so it
+    // has no number either — the readout and the surface both depend on this.
+    expect(valueAtMonth(points, 1.9, margin)).toBeNull();
+    expect(valueAtMonth(points, 9.1, margin)).toBeNull();
+    expect(valueAtMonth([], 3, margin)).toBeNull();
+  });
+
+  it('reads whichever field the caller picks', () => {
+    expect(valueAtMonth(points, 3, (p) => p.cumulative)).toBeCloseTo(1000, 9);
+    expect(valueAtMonth(points, 3, (p) => p.cost)).toBeCloseTo(100, 9);
+  });
+
+  it('does not divide by zero on samples that share a month', () => {
+    // Forced verticals emit two samples at one instant, which is how a config
+    // landing with no ramp window draws as a riser. Reading exactly at the riser
+    // gives the level the fleet serves *from* that instant — the top of the step,
+    // matching what `isStep` means — rather than the level it just left.
+    const vertical = [point(1, 0), point(3, 0), point(3, 500), point(6, 500)];
+    expect(valueAtMonth(vertical, 3, margin)).toBeCloseTo(500, 9);
+    expect(valueAtMonth(vertical, 4.5, margin)).toBeCloseTo(500, 9);
+  });
+});
+
+describe('ramp sampling is anchored, not proportional', () => {
+  // A rung that no sweep measured at the selected interactivity is dropped from
+  // the timeline, so the *same* governing config can be followed by a different
+  // next rung — or by none — depending on the target. That changes where the
+  // governing rollout's ramp ends.
+  //
+  // Nothing about the underlying curve changes: `levelWithin` is a function of
+  // the rollout's own start, from, to and the user's rampMonths. Only the
+  // reconstruction can move, and it must not, or the same config reconstructs
+  // differently depending on what else happened to be measured beside it.
+  const ramped: LifecycleAssumptions = { ...assumptions, rampMonths: 3 };
+
+  it('reconstructs a governing rollout identically whether or not it is cut short', () => {
+    const alone = computeLifecycle({
+      steps: [{ month: 0, billableInputTokPerSec: 1000, outputTokPerSec: 0 }],
+      costPerHour: 100,
+      provisionedMw: PROVISIONED_MW,
+      horizonMonths: 12,
+      assumptions: ramped,
+    })!;
+    // Same first rung; a second lands at 1.5, halfway through the first ramp, and
+    // truncates it. Before 1.5 the fleet is running the identical rollout.
+    const truncated = computeLifecycle({
+      steps: [
+        { month: 0, billableInputTokPerSec: 1000, outputTokPerSec: 0 },
+        { month: 1.5, billableInputTokPerSec: 3000, outputTokPerSec: 0 },
+      ],
+      costPerHour: 100,
+      provisionedMw: PROVISIONED_MW,
+      horizonMonths: 12,
+      assumptions: ramped,
+    })!;
+
+    for (const month of [0.1, 0.2, 0.35, 0.5, 0.75, 1, 1.2, 1.4]) {
+      const a = valueAtMonth(alone.points, month, (p) => p.revenue);
+      const b = valueAtMonth(truncated.points, month, (p) => p.revenue);
+      // Guard the guard: a mistyped accessor yields NaN, and `expect(NaN).toBe(NaN)`
+      // passes, so without this the comparison below would hold no matter what
+      // the sampler did.
+      expect(Number.isFinite(a), `month ${month} readable`).toBe(true);
+      // Bit-equal, not merely close: the sample months are the same months, so
+      // the interpolation is the same arithmetic on the same numbers. Dividing
+      // the ramp window into a fixed number of pieces instead would put the
+      // samples of the truncated run at different months and break this.
+      expect(b, `month ${month}`).toBe(a);
+    }
+  });
+
+  it('spaces ramp samples by the assumption, so a shorter window gets fewer', () => {
+    // The flip side of the guarantee above: resolution is fixed in months, so a
+    // rollout cut short carries proportionally fewer samples rather than the
+    // same count squeezed into less time. Equal spacing is the whole mechanism.
+    const series = computeLifecycle({
+      steps: [{ month: 0, billableInputTokPerSec: 1000, outputTokPerSec: 0 }],
+      costPerHour: 100,
+      provisionedMw: PROVISIONED_MW,
+      horizonMonths: 12,
+      assumptions: ramped,
+    })!;
+    const rampMonthsSampled = series.points.filter((p) => p.month <= 3).map((p) => p.month);
+    const gaps = rampMonthsSampled.slice(1).map((m, i) => m - rampMonthsSampled[i]!);
+    for (const gap of gaps) expect(gap).toBeCloseTo(3 / 24, 9);
+  });
+});
+
+describe('billableInputRate', () => {
+  // The agentic shape: input tokens dominate and most are served from cache.
+  const input = 13_525;
+
+  it('is an exact identity where nothing was measured, which is every fixed sequence', () => {
+    // Not "close to": fixed-sequence rows carry no cache metric on any row, so
+    // this path must return the caller's own number bit-for-bit. Anything else
+    // would move 8k/1k margins as a side effect of an agentic feature.
+    expect(billableInputRate(input, undefined, 0.1)).toBe(input);
+  });
+
+  it('is an exact identity at a 100% cached price, whatever the hit rate', () => {
+    expect(billableInputRate(input, 0.92, 1)).toBe(input);
+    expect(billableInputRate(input, 1, 1)).toBe(input);
+  });
+
+  it('bills the cached share at the cached fraction of the price', () => {
+    // 92% of the stream sells at 10%, the other 8% at full price.
+    expect(billableInputRate(input, 0.92, 0.1)).toBeCloseTo(input * (1 - 0.92 * 0.9), 9);
+    // At this mix the billable input rate collapses to a sixth of the raw one,
+    // which is the size of the correction the feature exists for.
+    expect(billableInputRate(input, 0.92, 0.1) / input).toBeLessThan(0.2);
+  });
+
+  it('clamps a hit rate the data reports above 1 rather than billing negative tokens', () => {
+    // Real rows report up to 1.185. Left unclamped that would discount more
+    // input tokens than were served.
+    expect(billableInputRate(input, 1.185, 0)).toBe(billableInputRate(input, 1, 0));
+    expect(billableInputRate(input, 1.185, 0)).toBeGreaterThanOrEqual(0);
+  });
+
+  it('is zero for an unusable input rate rather than NaN', () => {
+    expect(billableInputRate(Number.NaN, 0.9, 0.1)).toBe(0);
+    expect(billableInputRate(0, 0.9, 0.1)).toBe(0);
+    expect(billableInputRate(-5, 0.9, 0.1)).toBe(0);
+  });
+});
+
+describe('input and output prices', () => {
+  const costPerHour = 1000;
+  const IN = 800_000;
+  const OUT = 100_000;
+  const twoStream = {
+    steps: [{ month: 0, billableInputTokPerSec: IN, outputTokPerSec: OUT }],
+    costPerHour,
+    provisionedMw: PROVISIONED_MW,
+    horizonMonths: 12,
+    assumptions: { ...assumptions, inputPricePerMTok: 0.27, outputPricePerMTok: 1.1 },
+  };
+
+  it('charges each stream its own price', () => {
+    const series = computeLifecycle(twoStream)!;
+    const expected = ((IN * 0.27 + OUT * 1.1) / 1e6) * 86_400 * availabilityFromInterrupts(24, 12);
+    expect(series.revenuePerDay).toBeCloseTo(expected, 6);
+  });
+
+  it('is not a blended single price — the split changes the answer', () => {
+    // The same 900k tok/s at one blended price cannot reproduce it: output is a
+    // ninth of the tokens and a third of the revenue at these prices.
+    const blended = computeLifecycle({
+      ...twoStream,
+      assumptions: { ...assumptions, inputPricePerMTok: 0.27, outputPricePerMTok: 0.27 },
+    })!;
+    expect(blended.revenuePerDay).toBeLessThan(computeLifecycle(twoStream)!.revenuePerDay);
+    const outputShare = (OUT * 1.1) / (IN * 0.27 + OUT * 1.1);
+    expect(outputShare).toBeGreaterThan(0.3);
+  });
+
+  it('raising only the output price raises revenue, leaving cost alone', () => {
+    const dearer = computeLifecycle({
+      ...twoStream,
+      assumptions: { ...assumptions, inputPricePerMTok: 0.27, outputPricePerMTok: 2.2 },
+    })!;
+    const base = computeLifecycle(twoStream)!;
+    expect(dearer.revenuePerDay).toBeGreaterThan(base.revenuePerDay);
+    expect(dearer.costPerDay).toBe(base.costPerDay);
+    // Exactly the output term doubled, nothing else.
+    expect(dearer.revenuePerDay - base.revenuePerDay).toBeCloseTo(
+      ((OUT * 1.1) / 1e6) * 86_400 * availabilityFromInterrupts(24, 12),
+      6,
+    );
+  });
+
+  it('reports a gain that neither price can move, counted over both streams', () => {
+    // "Gain" is a statement about the hardware, so it counts tokens, not dollars.
+    // The streams grow disproportionately on purpose: input is flat and only
+    // output improves, so a gain counted on the input stream alone would read
+    // 1.0x and a gain counted on both reads 1.222x. A proportional fixture
+    // cannot tell those apart.
+    const stepped = {
+      ...twoStream,
+      steps: [
+        { month: 0, billableInputTokPerSec: IN, outputTokPerSec: OUT },
+        { month: 3, billableInputTokPerSec: IN, outputTokPerSec: OUT * 3 },
+      ],
+    };
+    const cheap = computeLifecycle(stepped)!;
+    const dear = computeLifecycle({
+      ...stepped,
+      assumptions: { ...assumptions, inputPricePerMTok: 50, outputPricePerMTok: 1 },
+    })!;
+    expect(cheap.improvementFactor).toBeCloseTo((IN + OUT * 3) / (IN + OUT), 9);
+    expect(cheap.improvementFactor).toBeGreaterThan(1.2);
+    // Same tokens, wildly different prices, same gain.
+    expect(dear.improvementFactor).toBeCloseTo(cheap.improvementFactor, 9);
+  });
+
+  it('keeps a step whose tokens are real even when both prices are zero', () => {
+    // A zero price is a legitimate what-if, not a reason to drop the config from
+    // the timeline — the staircase is measured, the prices are assumptions.
+    const free = computeLifecycle({
+      ...twoStream,
+      assumptions: { ...assumptions, inputPricePerMTok: 0, outputPricePerMTok: 0 },
+    })!;
+    expect(free.points.length).toBeGreaterThan(0);
+    expect(free.revenuePerDay).toBe(0);
+    expect(free.marginPerDay).toBeCloseTo(-free.costPerDay, 6);
+  });
+});
+
+describe('effectiveTokPerSec', () => {
+  it('weights output tokens by how much more they sell for', () => {
+    expect(effectiveTokPerSec(800_000, 100_000, 4)).toBe(800_000 + 400_000);
+  });
+
+  it('is what makes a two-price break-even a single solve', () => {
+    // Break-even on the effective rate returns an input price which, paired with
+    // `multiple x` that output price, zeroes the margin exactly.
+    const costPerHour = 1000;
+    const IN = 800_000;
+    const OUT = 100_000;
+    const multiple = 4;
+    const inputPrice = breakEvenPricePerMTok(
+      costPerHour,
+      effectiveTokPerSec(IN, OUT, multiple),
+      1,
+    )!;
+    const series = computeLifecycle({
+      steps: [{ month: 0, billableInputTokPerSec: IN, outputTokPerSec: OUT }],
+      costPerHour,
+      provisionedMw: PROVISIONED_MW,
+      horizonMonths: 12,
+      assumptions: {
+        mtbiDays: 0,
+        recoveryHours: 0,
+        rampMonths: 0,
+        inputPricePerMTok: inputPrice,
+        outputPricePerMTok: inputPrice * multiple,
+      },
+    })!;
+    expect(series.marginPerDay).toBeCloseTo(0, 6);
+  });
+
+  it('treats an unusable multiple as output earning nothing', () => {
+    expect(effectiveTokPerSec(100, 50, Number.NaN)).toBe(100);
+    expect(effectiveTokPerSec(100, 50, -1)).toBe(100);
+  });
+});
+
+describe('per-megawatt lifecycle metrics', () => {
+  const costPerHour = 10_000;
+  const steps: ThroughputStep[] = [
+    { month: 0, billableInputTokPerSec: 400_000, outputTokPerSec: 0 },
+    { month: 6, billableInputTokPerSec: 1_600_000, outputTokPerSec: 0 },
+  ];
+  const base = {
+    steps,
+    costPerHour,
+    provisionedMw: PROVISIONED_MW,
+    horizonMonths: 24,
+    assumptions,
+  };
+
+  it('divides daily revenue and margin by the power actually provisioned', () => {
+    const series = computeLifecycle(base)!;
+    for (const p of series.points) {
+      expect(p.revenuePerMw).toBeCloseTo(p.revenue / PROVISIONED_MW, 9);
+      expect(p.marginPerMw).toBeCloseTo(p.margin / PROVISIONED_MW, 9);
+    }
+  });
+
+  it('divides by provisioned power, so a denser fleet on the same budget reads higher', () => {
+    // Two fleets earning the same $/day, one occupying half the power. Per MW
+    // they are a factor of two apart even though `margin` cannot tell them apart.
+    const wide = computeLifecycle(base)!;
+    const dense = computeLifecycle({ ...base, provisionedMw: PROVISIONED_MW / 2 })!;
+    expect(dense.points.at(-1)!.margin).toBeCloseTo(wide.points.at(-1)!.margin, 9);
+    expect(dense.points.at(-1)!.marginPerMw).toBeCloseTo(wide.points.at(-1)!.marginPerMw * 2, 9);
+  });
+
+  it('crosses zero at the same instant margin does', () => {
+    // The rescale is by a positive constant, which is what licenses the chart's
+    // break-even rule and the surface's plane to stay at y = 0 on this metric.
+    const series = computeLifecycle(base)!;
+    for (const p of series.points) {
+      expect(Math.sign(p.marginPerMw)).toBe(Math.sign(p.margin));
+    }
+    expect(isBreakEvenAnchored('marginPerMw')).toBe(true);
+    expect(isBreakEvenAnchored('revenue')).toBe(false);
+    expect(isBreakEvenAnchored('revenuePerMw')).toBe(false);
+    expect(isBreakEvenAnchored('cumulativeRevenue')).toBe(false);
+  });
+
+  it('stays at zero rather than dividing by an unusable power figure', () => {
+    for (const provisionedMw of [0, -1, Number.NaN]) {
+      const series = computeLifecycle({ ...base, provisionedMw })!;
+      for (const p of series.points) {
+        expect(Number.isFinite(p.revenuePerMw)).toBe(true);
+        expect(p.revenuePerMw).toBe(0);
+        expect(Number.isFinite(p.marginPerMw)).toBe(true);
+        expect(p.marginPerMw).toBe(0);
+      }
+    }
+  });
+
+  it('is selected by metricValue', () => {
+    const series = computeLifecycle(base)!;
+    const p = series.points.at(-1)!;
+    expect(metricValue(p, 'revenuePerMw')).toBe(p.revenuePerMw);
+    expect(metricValue(p, 'marginPerMw')).toBe(p.marginPerMw);
+    expect(metricValue(p, 'margin')).toBe(p.margin);
+  });
+});
+
+describe('splitTokenStreams', () => {
+  // 8k/1k: every request is 8192 in / 1024 out, so the share is the sequence shape.
+  const SHARE = 8192 / 9216;
+
+  it('splits the per-chip total by the measured mix', () => {
+    const { billableInputTokPerSec, outputTokPerSec } = splitTokenStreams(
+      9000,
+      SHARE,
+      undefined,
+      1,
+    );
+    expect(billableInputTokPerSec).toBeCloseTo(9000 * SHARE, 9);
+    expect(outputTokPerSec).toBeCloseTo(9000 * (1 - SHARE), 9);
+    // The whole point: the two streams add back up to the rate the fleet was
+    // sized and costed on. Reading the per-GPU rates directly did not.
+    expect(billableInputTokPerSec + outputTokPerSec).toBeCloseTo(9000, 6);
+  });
+
+  it('cannot bill more tokens than the chips served, whatever the reported rates', () => {
+    // A disaggregated run reports input per prefill chip and output per decode
+    // chip, so those two summed reach 16x the per-chip total in production. Any
+    // share is still a share: the streams sum to the total.
+    for (const share of [0, 0.5, SHARE, 1]) {
+      const s = splitTokenStreams(1000, share, undefined, 1);
+      expect(s.billableInputTokPerSec + s.outputTokPerSec).toBeCloseTo(1000, 6);
+    }
+  });
+
+  it('clamps a share outside [0,1] rather than producing a negative stream', () => {
+    expect(splitTokenStreams(1000, 1.5, undefined, 1).outputTokPerSec).toBe(0);
+    expect(splitTokenStreams(1000, -0.5, undefined, 1).billableInputTokPerSec).toBe(0);
+  });
+
+  it('charges nothing as input when the mix is unknown', () => {
+    // Understating revenue beats inventing a mix.
+    const s = splitTokenStreams(1000, undefined, undefined, 1);
+    expect(s.billableInputTokPerSec).toBe(0);
+    expect(s.outputTokPerSec).toBe(1000);
+  });
+
+  it('applies the cached discount to the input stream only', () => {
+    const plain = splitTokenStreams(1000, 0.9, undefined, 0.1);
+    const cached = splitTokenStreams(1000, 0.9, 0.5, 0.1);
+    expect(cached.outputTokPerSec).toBeCloseTo(plain.outputTokPerSec, 9);
+    expect(cached.billableInputTokPerSec).toBeLessThan(plain.billableInputTokPerSec);
+    // Half the input stream at a tenth of the price.
+    expect(cached.billableInputTokPerSec).toBeCloseTo(900 * (1 - 0.5 * 0.9), 9);
+  });
+
+  it('is zero for an unusable total', () => {
+    for (const total of [0, -1, Number.NaN]) {
+      const s = splitTokenStreams(total, 0.9, undefined, 1);
+      expect(s.billableInputTokPerSec).toBe(0);
+      expect(s.outputTokPerSec).toBe(0);
+    }
+  });
+
+  it('keeps revenue monotone in the total when the mix holds — the dip regression', () => {
+    // The MI355X dip: 2026-07-22 mori-sglang (disaggregated) handed over to
+    // 2026-07-28 atom_mtp (aggregated) with a HIGHER total tok/s/MW, yet the
+    // plotted margin fell, because the disagg run's input/output rates were on a
+    // prefill/decode denominator and billed up to 16x the tokens its chips served.
+    // Split from the total instead and a rank that rose cannot make revenue fall.
+    const revenue = (total: number) => {
+      const s = splitTokenStreams(total, SHARE, undefined, 1);
+      return s.billableInputTokPerSec * 0.27 + s.outputTokPerSec * 1.1;
+    };
+    // The two real per-chip totals from that handover, in order.
+    expect(revenue(2336)).toBeLessThan(revenue(2618));
+  });
+});
+
+describe('outputTokPerChip', () => {
+  it('is the reported rate when nothing pins the mix down', () => {
+    // Every aggregated run: unchanged by this helper, bit for bit.
+    expect(outputTokPerChip(1000, undefined, 123)).toBe(123);
+  });
+
+  it('re-bases a per-decode-chip rate onto chips overall', () => {
+    // The real shape: 8 prefill + 8 decode on 8k/1k. Output is reported per decode
+    // chip, so it reads 2x what the fleet's chips overall deliver. Concurrent users
+    // are a count of output streams, so that 2x lands straight on the figure.
+    const share = 8192 / 9216;
+    const perDecodeChip = 6.857;
+    const perChipOverall = outputTokPerChip(30.689, share, perDecodeChip);
+    expect(perChipOverall).toBeCloseTo(30.689 * (1 - share), 6);
+    expect(perDecodeChip / perChipOverall).toBeCloseTo(2, 1);
+  });
+
+  it('clamps a nonsensical share instead of returning a negative rate', () => {
+    expect(outputTokPerChip(1000, 1.5, 50)).toBe(0);
+    expect(outputTokPerChip(1000, -1, 50)).toBe(1000);
+  });
+
+  it('is zero for an unusable total', () => {
+    expect(outputTokPerChip(0, 0.5, 50)).toBe(0);
+    expect(outputTokPerChip(Number.NaN, 0.5, 50)).toBe(0);
+  });
+});
diff --git a/packages/app/src/components/calculator/lifecycle.ts b/packages/app/src/components/calculator/lifecycle.ts
new file mode 100644
index 000000000..703b9d0af
--- /dev/null
+++ b/packages/app/src/components/calculator/lifecycle.ts
@@ -0,0 +1,732 @@
+/**
+ * Pure fleet lifecycle math — no React, no 'use client'.
+ *
+ * `fleet.ts` answers "what does this fleet earn and cost per hour, right now?".
+ * This answers "what has it earned and cost since the model shipped?" — and the
+ * shape of that answer is measured, not assumed.
+ *
+ * A fixed fleet's revenue is not flat. The chips do not change, but the software
+ * serving them does: over the months after a model's release, sweeps find better
+ * configs, and each one raises the tokens the same hardware delivers at the same
+ * interactivity. MI355X on DeepSeek-V4-Pro is the clearest example — an FP8
+ * baseline, then FP4 with graphs and FlashMLA, then AITER GEMMs and Triton SWA,
+ * then fused kernels — each step lifting the frontier.
+ *
+ *   revenue │              ,──── each config rolls out to its own numbers
+ *           │         ,───╯
+ *           │    ,───╯
+ *   ────────┼──╱──────────────────────────────  cost: flat, the racks
+ *           │ ╱ the first config rolls out from zero
+ *           └──────────────────────────────────▶  months since model release
+ *
+ * A config does not take effect the instant a sweep finds it: it is rolled out.
+ * So every config gets its own ramp, climbing from whatever the fleet was already
+ * serving to that config's numbers. The first config climbs from zero — nothing is
+ * being served before it lands.
+ *
+ * Cost is constant throughout, so the gap between the lines is the return on
+ * software progress, which is the thing no single benchmark date can show.
+ *
+ * Three conventions worth stating:
+ *
+ * 1. **Cost is flat, including through every rollout.** It is `chips x $/chip/hr`,
+ *    and neither term moves when a config rolls out — racks bill from the moment
+ *    they are energised, not from the moment they are loaded. So the first rollout
+ *    is paid for in full while it earns its way up from zero, which is what puts
+ *    the opening months below the rule and gives payback its meaning.
+ * 2. **The ramp is an assumption; the steps are not.** Ramp length is a user
+ *    input, defaulting to half a month, and it applies to every rollout.
+ *    Which configs exist, when, and how fast they ran are all measured.
+ * 3. **Interrupts are an availability haircut, not drawn events.** A 24-day MTBI
+ *    over a multi-year window is thousands of events, each far under a pixel, so
+ *    drawing them is aliasing noise. They scale revenue instead.
+ *
+ * The lifecycle assumptions here (price, ramp, MTBI, recovery, horizon) are user
+ * inputs, and callers must present them as such. The throughput steps are not:
+ * they are measured.
+ */
+
+const HOURS_PER_DAY = 24;
+
+/**
+ * Calendar month, in days — 365.25 ÷ 12.
+ *
+ * Deliberately not `HOURS_PER_MONTH / 24`. That constant is 730 (a flat 365-day
+ * year), which is the right convention for `costPerMonth` in fleet.ts, where it
+ * is a billing figure. It is the wrong one here: the x axis converts dates to
+ * months with 365.25 ÷ 12 (`MS_PER_MONTH` below), and this constant converts
+ * those months back to days to integrate. Two calendars across one round trip
+ * understated every interval by 0.068% — small, but it is the only conversion
+ * factor in the model that existed twice with two values, and the integral it
+ * feeds is the section's headline number.
+ */
+export const DAYS_PER_MONTH = 365.25 / 12;
+
+/** The same month, in milliseconds. Shared so the axis and the integral agree. */
+export const MS_PER_MONTH = DAYS_PER_MONTH * HOURS_PER_DAY * 3600 * 1000;
+
+const TOKENS_PER_MILLION = 1e6;
+const SECONDS_PER_DAY = 86_400;
+
+/**
+ * Input tokens an operator can charge the fresh-token price for, tok/s.
+ *
+ * Agentic traces run at a median 133:1 input:output token ratio with a median
+ * 92% prefix-cache hit rate, and providers bill a cache read at a fraction of a
+ * fresh input token. Charging one blended price against the raw token rate
+ * therefore overstates agentic margin by close to an order of magnitude. Fixed
+ * sequences carry no cache metric on any row, so `cacheHitRate` is undefined
+ * there and this returns `base` untouched.
+ *
+ * Expressed as a *subtraction* from the caller's throughput rather than as
+ * `output + billableInput`, so the result is exactly `base` whenever there is
+ * nothing to discount — no rounding drift against the independently splined
+ * total, and no behaviour change outside agentic traces by construction rather
+ * than by luck.
+ *
+ * @param base           tok/s for the selected token type — the undiscounted rate.
+ * @param inputTput      tok/s of input tokens, the only stream a cache can serve.
+ * @param cacheHitRate   cached fraction of those input tokens, or undefined when unmeasured.
+ * @param cacheReadRatio price of a cached token as a fraction of a fresh one (1 = no discount).
+ */
+export function billableInputRate(
+  inputTput: number,
+  cacheHitRate: number | undefined,
+  cacheReadRatio: number,
+): number {
+  if (!Number.isFinite(inputTput) || inputTput <= 0) return 0;
+  if (typeof cacheHitRate !== 'number' || !Number.isFinite(cacheHitRate)) return inputTput;
+  const hit = Math.max(0, Math.min(1, cacheHitRate));
+  const ratio = Math.max(0, Math.min(1, cacheReadRatio));
+  // A cache read still sells, just for less: `hit` of the stream at `ratio` of
+  // the price is the same revenue as `hit * ratio` of the stream at full price.
+  return inputTput * (1 - hit * (1 - ratio));
+}
+
+export interface LifecycleAssumptions {
+  /** Mean time between interruptions, in days. Zero means none modelled. */
+  mtbiDays: number;
+  /** Hours to recover from one interruption. */
+  recoveryHours: number;
+  /**
+   * Sale price of input tokens, $/M tok. Defaults to break-even upstream.
+   *
+   * Cached input tokens are already discounted into the step's billable input
+   * rate — see {@link billableInputRate} — so this is the fresh-token price.
+   */
+  inputPricePerMTok: number;
+  /**
+   * Sale price of output tokens, $/M tok. Separate because providers bill them
+   * at a multiple of input, and the two streams are wildly unequal: 8:1 on a
+   * fixed 8k/1k sequence and ~130:1 on agentic traces, so a single blended price
+   * hides which side of the workload the money is actually on.
+   */
+  outputPricePerMTok: number;
+  /**
+   * Months for a config to roll out across the fleet. Applies to every config,
+   * including the first, which rolls out from zero. Zero means each config takes
+   * effect instantly. An assumption, not a measurement.
+   */
+  rampMonths: number;
+}
+
+/**
+ * A measured config improvement: from `month` onwards the fleet rolls out to
+ * `billableTokPerSec`, and holds it until the next step.
+ */
+export interface ThroughputStep {
+  /** Months since the anchor date (the model's release), >= 0. */
+  month: number;
+  /**
+   * Fleet input tokens once this config lands, tok/s, already reduced for the
+   * share served from cache — see {@link billableInputRate}.
+   *
+   * Deliberately not named `fleetTokPerSec` like {@link FleetStats}: that one is
+   * the physical rate the racks deliver, and the two diverge whenever a workload
+   * serves input tokens from cache. `FleetStats.fleetTokPerSec` is the physical
+   * rate; revenue here is charged on these two.
+   */
+  billableInputTokPerSec: number;
+  /** Fleet output tokens once this config lands, tok/s. Never cached. */
+  outputTokPerSec: number;
+}
+
+/**
+ * The token rate a break-even solve should be handed, given the price ratio the
+ * user has set: input tokens plus output tokens weighted by how much more an
+ * output token sells for. Feeding this to {@link breakEvenPricePerMTok} returns
+ * the input price that zeroes the margin while preserving that ratio.
+ */
+export function effectiveTokPerSec(
+  inputTokPerSec: number,
+  outputTokPerSec: number,
+  outputPriceMultiple: number,
+): number {
+  const multiple =
+    Number.isFinite(outputPriceMultiple) && outputPriceMultiple > 0 ? outputPriceMultiple : 0;
+  return inputTokPerSec + outputTokPerSec * multiple;
+}
+
+/**
+ * Split a per-chip total token rate into the two streams revenue is charged on.
+ *
+ * Takes a *share* rather than two rates because the two rates are not always on
+ * the same denominator: a disaggregated run reports input per prefill chip and
+ * output per decode chip, while the total — the figure the fleet is sized and
+ * costed on — is per chip overall. Reading the rates directly billed a
+ * disaggregated config for up to 16x the tokens its chips actually served, which
+ * put a visible dip in the margin line the moment a later, genuinely better
+ * aggregated config took over. See `inputTokenShare` in `useThroughputData.ts`.
+ *
+ * With no share to hand, nothing is charged as input: understating revenue beats
+ * inventing a mix.
+ */
+export function splitTokenStreams(
+  totalTokPerSec: number,
+  inputTokenShare: number | undefined,
+  cacheHitRate: number | undefined,
+  cacheReadRatio: number,
+): { billableInputTokPerSec: number; outputTokPerSec: number } {
+  if (!Number.isFinite(totalTokPerSec) || totalTokPerSec <= 0) {
+    return { billableInputTokPerSec: 0, outputTokPerSec: 0 };
+  }
+  const share =
+    typeof inputTokenShare === 'number' && Number.isFinite(inputTokenShare)
+      ? Math.max(0, Math.min(1, inputTokenShare))
+      : 0;
+  return {
+    billableInputTokPerSec: billableInputRate(totalTokPerSec * share, cacheHitRate, cacheReadRatio),
+    outputTokPerSec: totalTokPerSec * (1 - share),
+  };
+}
+
+/**
+ * Output tokens per chip **overall**, for a fleet sized on chips overall.
+ *
+ * `outputThroughput` is per *decode* chip on a disaggregated run, so dividing it
+ * into a whole-fleet chip count overstates the output rate by
+ * `(prefill + decode) / decode` — a median of 2x and up to 7x in production
+ * history. Concurrent users are a count of output streams, so that error lands
+ * directly on the figure a reader is most likely to quote.
+ *
+ * Falls back to the reported rate where no share is known, which is the same
+ * number as before for every aggregated run.
+ */
+export function outputTokPerChip(
+  totalTokPerChip: number,
+  inputTokenShare: number | undefined,
+  reportedOutputTokPerChip: number,
+): number {
+  if (typeof inputTokenShare !== 'number' || !Number.isFinite(inputTokenShare)) {
+    return reportedOutputTokPerChip;
+  }
+  if (!Number.isFinite(totalTokPerChip) || totalTokPerChip <= 0) return 0;
+  return totalTokPerChip * (1 - Math.max(0, Math.min(1, inputTokenShare)));
+}
+
+/** A price is only a price when it is finite and positive; anything else is free. */
+function nonNegativePrice(value: number): number {
+  return Number.isFinite(value) && value > 0 ? value : 0;
+}
+
+/** Tokens a step sells per second, both streams. Price-independent. */
+function stepTokPerSec(step: ThroughputStep): number {
+  return step.billableInputTokPerSec + step.outputTokPerSec;
+}
+
+export interface LifecycleInputs {
+  /**
+   * Chronological, non-empty. The first step's month is when this hardware was
+   * first measured on the model — before that there is no data, so no line.
+   */
+  steps: readonly ThroughputStep[];
+  /** Fleet TCO for the selected tier ($/hr). Constant: configs don't change it. */
+  costPerHour: number;
+  /**
+   * Facility power this fleet actually occupies, MW — `chips x kW/chip / 1000`,
+   * **not** the budget the user typed. Chip counts are whole, so a fleet fills
+   * its budget to within one chip and the remainder is stranded power nobody is
+   * paying for capacity in. Dividing by the nominal budget would credit each
+   * chip with power it never provisioned.
+   *
+   * Only used to express the per-MW metrics. Zero or negative leaves those
+   * metrics at zero rather than dividing by it.
+   */
+  provisionedMw: number;
+  /** End of the modelled window, months since the anchor. */
+  horizonMonths: number;
+  assumptions: LifecycleAssumptions;
+}
+
+export interface LifecyclePoint {
+  /** Months since the anchor date. */
+  month: number;
+  /** Revenue rate, $/day. */
+  revenue: number;
+  /** Cost rate, $/day. */
+  cost: number;
+  /** revenue − cost, $/day. */
+  margin: number;
+  /** Revenue per megawatt provisioned, $/MW/day. */
+  revenuePerMw: number;
+  /**
+   * The same margin per megawatt provisioned, $/MW/day.
+   *
+   * Worth being clear about what this does and does not add. Every chip in the
+   * section is sized to the same power budget, so this is very nearly `margin`
+   * divided by a constant and it re-ranks almost nothing — the only spread comes
+   * from how completely each chip's power density fills the budget. What it buys
+   * is a figure that does not move when the budget does, which is the unit a
+   * power-constrained plan is actually written in.
+   */
+  marginPerMw: number;
+  /** Cumulative margin from the first step to here, $. */
+  cumulative: number;
+  /**
+   * Cumulative *revenue* from the first step to here, $ — the area under the
+   * revenue curve, with no cost subtracted.
+   *
+   * Separate from `cumulative` rather than derived from it, because
+   * `cumulative + cost x elapsed` only reproduces it when the two are integrated
+   * over identical intervals, and a caller reading at an arbitrary month has no
+   * elapsed term to hand. Accumulated by the same trapezoid rule so the two agree
+   * exactly at every sample.
+   */
+  cumulativeRevenue: number;
+  /** True at the instant a new config takes effect — the riser of the step. */
+  isStep: boolean;
+  /** True while a rollout is still climbing, i.e. not yet at its config's rate. */
+  isRamp: boolean;
+}
+
+export interface LifecycleSeries {
+  /**
+   * Densely sampled across every rollout, plus the flat stretch each config holds
+   * and a closing point at the horizon. Interpolate linearly to draw them: the
+   * shape is already in the samples.
+   */
+  points: LifecyclePoint[];
+  /**
+   * Revenue at the end of the window, $/day — the latest config at whatever level
+   * its rollout has reached by then. Equal to that config's full rate only if the
+   * rollout finished before the horizon and no later config pre-empted it.
+   */
+  revenuePerDay: number;
+  /**
+   * Revenue at the first measured config's full rate, $/day. A rollout still
+   * climbing when the next config lands never attains it; it is the rate that
+   * config was measured at, which is what makes it comparable across chips.
+   */
+  firstRevenuePerDay: number;
+  /** Flat cost rate, $/day. */
+  costPerDay: number;
+  /** `revenuePerDay` − cost, $/day. */
+  marginPerDay: number;
+  /** `firstRevenuePerDay` − cost, $/day. */
+  firstMarginPerDay: number;
+  /**
+   * Latest config's throughput ÷ the first's — what software progress has been
+   * worth. A ratio of measured rates, deliberately independent of the rollout
+   * assumption, so it does not move when the ramp input changes — and of both
+   * prices, so it stays a statement about the hardware and not about billing.
+   * Counted over both token streams.
+   */
+  improvementFactor: number;
+  availability: number;
+  /** Months since the anchor at which cumulative margin first turns positive. */
+  paybackMonth: number | null;
+  /** Cumulative margin over the whole window, $. */
+  lifetimeMargin: number;
+  /** Month of the first measured config. */
+  startMonth: number;
+  endMonth: number;
+  /**
+   * Month the first rollout completes — when the opening config is fully deployed.
+   * Equals `startMonth` when no ramp is modelled. Later rollouts have their own
+   * windows; this one is called out because it is where revenue leaves zero.
+   */
+  rampEndMonth: number;
+  /** Number of measured improvements, i.e. steps after the first. */
+  improvementCount: number;
+}
+
+/**
+ * What a view plots. `margin` and `revenue` are rates in $/day; their `PerMw`
+ * variants divide the same rate by provisioned power. `cumulativeRevenue` is the
+ * area under the revenue curve since the fleet's first config, in $.
+ *
+ * Lives here rather than in a chart component because the selection is pure and
+ * the chart is not the only thing that reads it.
+ *
+ * Note the unit change: a caller that formats an axis or a tooltip must ask
+ * `isCumulative` rather than assume $/day, and anything anchored to zero as
+ * break-even (the dashed rule) applies to margin metrics alone.
+ */
+export type LifecycleMetric =
+  | 'margin'
+  | 'marginPerMw'
+  | 'revenue'
+  | 'revenuePerMw'
+  | 'cumulativeRevenue';
+
+/** True when the metric is a running total in $ rather than a rate in $/day. */
+export function isCumulative(metric: LifecycleMetric): boolean {
+  return metric === 'cumulativeRevenue';
+}
+
+/**
+ * True when zero on this metric means break-even, so a view may draw a rule
+ * there. Per-MW margin is still revenue − cost, only rescaled by
+ * a positive number, so its zero crossing is the same instant as `margin`'s.
+ */
+export function isBreakEvenAnchored(metric: LifecycleMetric): boolean {
+  return metric === 'margin' || metric === 'marginPerMw';
+}
+
+/** The quantity a metric names, for reading off a sampled point. */
+export function metricValue(point: LifecyclePoint, metric: LifecycleMetric): number {
+  if (metric === 'revenue') return point.revenue;
+  if (metric === 'revenuePerMw') return point.revenuePerMw;
+  if (metric === 'cumulativeRevenue') return point.cumulativeRevenue;
+  if (metric === 'marginPerMw') return point.marginPerMw;
+  return point.margin;
+}
+
+/**
+ * A sampled series' value at an arbitrary month, or null outside its own window.
+ *
+ * Linear between samples, which is exactly what a `curveLinear` line draws through
+ * them — so a readout built on this always agrees with the pixel under the cursor.
+ * Null rather than a clamp at the ends: a chip first measured a year after release
+ * has no line before then, and so no number either.
+ */
+export function valueAtMonth(
+  points: readonly LifecyclePoint[],
+  month: number,
+  pick: (point: LifecyclePoint) => number,
+): number | null {
+  const first = points[0];
+  const last = points.at(-1);
+  if (!first || !last) return null;
+  if (month < first.month || month > last.month) return null;
+
+  let lo = 0;
+  let hi = points.length - 1;
+  while (lo < hi) {
+    const mid = (lo + hi + 1) >> 1;
+    if (points[mid]!.month <= month) lo = mid;
+    else hi = mid - 1;
+  }
+  const before = points[lo]!;
+  const after = points[Math.min(lo + 1, points.length - 1)]!;
+  const span = after.month - before.month;
+  const t = span > 0 ? (month - before.month) / span : 0;
+  const a = pick(before);
+  return a + (pick(after) - a) * t;
+}
+
+/**
+ * Fraction of wall-clock time the fleet is serving traffic.
+ *
+ *   availability = mtbi / (mtbi + recovery)
+ *
+ * with both terms in days. A zero or negative MTBI means "no interruptions
+ * modelled" (1), not "always down" — the input is an optional refinement, and
+ * treating a blank field as total failure would be a hostile default.
+ */
+export function availabilityFromInterrupts(mtbiDays: number, recoveryHours: number): number {
+  if (!Number.isFinite(mtbiDays) || mtbiDays <= 0) return 1;
+  if (!Number.isFinite(recoveryHours) || recoveryHours <= 0) return 1;
+  return mtbiDays / (mtbiDays + recoveryHours / HOURS_PER_DAY);
+}
+
+/**
+ * Price at which revenue exactly covers cost at a given throughput, $/M tok.
+ *
+ * Derived from the fleet's own cost and throughput, so the margin this module
+ * plots is exactly zero at this price — which is why `availability` belongs here.
+ * Interrupts cost the same racks fewer billable tokens, so the break-even price
+ * rises by exactly the haircut; ignoring it returns a price at which this
+ * module's own margin is negative.
+ *
+ * With two prices, break-even is a line rather than a point: any (input, output)
+ * pair on it zeroes the margin. Callers resolve that by fixing the ratio between
+ * the two and passing the **effective** rate — `billableInput + ratio x output` —
+ * which makes the answer the input price, and the output price `ratio x` it. See
+ * {@link effectiveTokPerSec}.
+ */
+export function breakEvenPricePerMTok(
+  costPerHour: number,
+  tokPerSec: number,
+  availability = 1,
+): number | null {
+  if (!(tokPerSec > 0) || !Number.isFinite(costPerHour)) return null;
+  if (!(availability > 0) || !Number.isFinite(availability)) return null;
+  return (costPerHour / (tokPerSec * 3600 * availability)) * TOKENS_PER_MILLION;
+}
+
+/** Revenue rate in $/day for a throughput, price and availability. */
+function revenuePerDayFor(revenuePerSec: number, availability: number): number {
+  return revenuePerSec * SECONDS_PER_DAY * availability;
+}
+
+/**
+ * What a step sells per second, in dollars — the quantity the rollout ramps.
+ *
+ * Both streams belong to one config and roll out together, so ramping their
+ * priced sum is exactly ramping each and pricing after. Carrying one scalar
+ * through the integral keeps the trapezoid rule and the payback search unchanged.
+ */
+function revenuePerSecOf(
+  step: ThroughputStep,
+  inputPricePerMTok: number,
+  outputPricePerMTok: number,
+): number {
+  return (
+    (step.billableInputTokPerSec * inputPricePerMTok + step.outputTokPerSec * outputPricePerMTok) /
+    TOKENS_PER_MILLION
+  );
+}
+
+/**
+ * Samples per full ramp window. Enough that a smoothstep reads as a curve rather
+ * than stairs. A rollout cut short by the next config gets proportionally fewer,
+ * because the cadence is fixed — see the sampling loop.
+ */
+const RAMP_SAMPLES = 24;
+
+/**
+ * Month tolerance for "this sample is already the ramp's end". Guards against
+ * emitting a duplicate at `rampEnd` when the cadence happens to land on it.
+ */
+const EPSILON_MONTHS = 1e-9;
+
+/** Smoothstep: eases in and out, so the ramp has no kink at either end. */
+function smoothstep(t: number): number {
+  if (t <= 0) return 0;
+  if (t >= 1) return 1;
+  return t * t * (3 - 2 * t);
+}
+
+/**
+ * Progress through a ramp that began at `startMonth`, in [0, 1].
+ *
+ * Used twice: once per config rollout, and once for the capacity buildout that
+ * rides along with the first rollout. Ramps are timed from the event that starts
+ * them — a config lands and then rolls out — never from the model's release.
+ */
+export function rampFractionAt(month: number, startMonth: number, rampMonths: number): number {
+  if (!Number.isFinite(rampMonths) || rampMonths <= 0) return 1;
+  return smoothstep((month - startMonth) / rampMonths);
+}
+
+/**
+ * Returns null when the projection is meaningless: no steps, no cost basis, or a
+ * horizon that ends before the first measured config.
+ */
+export function computeLifecycle(inputs: LifecycleInputs): LifecycleSeries | null {
+  const { steps, costPerHour, provisionedMw, horizonMonths, assumptions } = inputs;
+  // Guarded rather than trusted: an unregistered chip or a budget too small for
+  // one chip must leave the per-MW metric flat at zero, not NaN or Infinity.
+  const perMw = Number.isFinite(provisionedMw) && provisionedMw > 0 ? 1 / provisionedMw : 0;
+  if (steps.length === 0 || !Number.isFinite(costPerHour)) return null;
+
+  const sorted = [...steps]
+    .filter((s) => Number.isFinite(s.month) && stepTokPerSec(s) > 0)
+    .toSorted((a, b) => a.month - b.month);
+  if (sorted.length === 0) return null;
+
+  // Two configs can land in the same month — the progression feeding this keeps
+  // every rung that beat the incumbent, and same-day rungs share a month. Only
+  // the best of them ever serves traffic, so collapse them. Emitting both would
+  // give the superseded config a zero-width rollout, which spikes the line to a
+  // full rate the fleet never reached and marks two risers at one instant.
+  const ordered: ThroughputStep[] = [];
+  for (const step of sorted) {
+    const previous = ordered.at(-1);
+    if (previous && previous.month === step.month) {
+      if (stepTokPerSec(step) > stepTokPerSec(previous)) ordered[ordered.length - 1] = step;
+      continue;
+    }
+    ordered.push(step);
+  }
+
+  const startMonth = ordered[0]!.month;
+  const endMonth = Math.max(horizonMonths, startMonth);
+  if (!(endMonth > startMonth)) return null;
+
+  const availability = availabilityFromInterrupts(assumptions.mtbiDays, assumptions.recoveryHours);
+  const inputPrice = nonNegativePrice(assumptions.inputPricePerMTok);
+  const outputPrice = nonNegativePrice(assumptions.outputPricePerMTok);
+  const costPerDay = costPerHour * HOURS_PER_DAY;
+  const rampMonths =
+    Number.isFinite(assumptions.rampMonths) && assumptions.rampMonths > 0
+      ? assumptions.rampMonths
+      : 0;
+  const rampEndMonth = Math.min(startMonth + rampMonths, endMonth);
+
+  const applied = ordered.filter((s) => s.month < endMonth);
+  if (applied.length === 0) return null;
+
+  /**
+   * One rollout per config: when it lands, the fleet climbs from whatever it was
+   * already serving to that config's numbers over the ramp window. `from` is the
+   * level actually reached by the previous rollout, so a config landing before the
+   * previous one finished rolling out starts from there rather than jumping.
+   */
+  const rollouts: { start: number; end: number; from: number; to: number }[] = [];
+  for (const step of applied) {
+    const previous = rollouts.at(-1);
+    const from = previous ? levelWithin(previous, step.month) : 0;
+    rollouts.push({
+      start: step.month,
+      end: Math.min(step.month + rampMonths, endMonth),
+      from,
+      to: revenuePerSecOf(step, inputPrice, outputPrice),
+    });
+  }
+
+  /** Throughput at `month` within one rollout, clamped to its end level. */
+  function levelWithin(
+    rollout: { start: number; end: number; from: number; to: number },
+    month: number,
+  ): number {
+    const fraction = rampFractionAt(month, rollout.start, rampMonths);
+    return rollout.from + (rollout.to - rollout.from) * fraction;
+  }
+
+  // Build the samples in order: each rollout contributes its ramp, then the flat
+  // stretch it holds until the next one lands.
+  const samples: { month: number; level: number; isStep: boolean; isRamp: boolean }[] = [];
+  for (let i = 0; i < rollouts.length; i += 1) {
+    const rollout = rollouts[i]!;
+    const next = rollouts[i + 1];
+    const holdUntil = next ? Math.min(next.start, endMonth) : endMonth;
+    const rampEnd = Math.min(rollout.end, holdUntil);
+
+    if (rampEnd > rollout.start) {
+      // A fixed cadence anchored at the rollout's start — deliberately not the
+      // window divided into a fixed count.
+      //
+      // `levelWithin` is a smoothstep, which is convex over the first half of its
+      // window, so reading between samples with a straight line overestimates by
+      // an amount that grows with the spacing. Dividing the window into
+      // RAMP_SAMPLES pieces makes that spacing a function of `rampEnd` — and
+      // `rampEnd` depends on whether the *next* config is present. One line is
+      // self-consistent either way, but two lifecycles built from the same runs
+      // at different interactivities would sample the identical governing config
+      // on different grids and disagree. A cadence fixed by `rampMonths` — a user
+      // assumption, independent of which configs happen to be readable — puts the
+      // samples at the same months regardless.
+      const spacing = rampMonths / RAMP_SAMPLES;
+      let s = 0;
+      for (; rollout.start + spacing * s < rampEnd - EPSILON_MONTHS; s += 1) {
+        const month = rollout.start + spacing * s;
+        samples.push({
+          month,
+          level: levelWithin(rollout, month),
+          isStep: s === 0,
+          isRamp: true,
+        });
+      }
+      samples.push({
+        month: rampEnd,
+        level: levelWithin(rollout, rampEnd),
+        isStep: s === 0,
+        // A rollout cut short by the next config is still climbing at its last
+        // sample, so the flag follows the level, not the sample index.
+        isRamp: rampEnd < rollout.end,
+      });
+    } else {
+      // No ramp window at all: emit the incumbent level at this instant first so
+      // the line rises vertically instead of sloping into the new config.
+      if (i > 0)
+        samples.push({ month: rollout.start, level: rollout.from, isStep: false, isRamp: false });
+      samples.push({ month: rollout.start, level: rollout.to, isStep: true, isRamp: false });
+    }
+
+    if (holdUntil > rampEnd) {
+      samples.push({
+        month: holdUntil,
+        level: levelWithin(rollout, rampEnd),
+        isStep: false,
+        isRamp: false,
+      });
+    }
+  }
+
+  const points: LifecyclePoint[] = [];
+  let cumulative = 0;
+  let cumulativeRevenue = 0;
+  let paybackMonth: number | null = null;
+
+  // `level` is already dollars per second — see `revenuePerSecOf`.
+  const revenueOf = (sample: { level: number }) => revenuePerDayFor(sample.level, availability);
+
+  for (let i = 0; i < samples.length; i += 1) {
+    const here = samples[i]!;
+    const revenue = revenueOf(here);
+    const cost = costPerDay;
+    if (i > 0) {
+      const prev = samples[i - 1]!;
+      // Revenue varies continuously between samples while cost does not, so
+      // trapezoid the revenue and subtract the flat cost. Sample density is what
+      // makes this accurate; a forced vertical shares its month with the previous
+      // sample and so contributes nothing.
+      const meanRevenue = (revenueOf(prev) + revenue) / 2;
+      const margin = meanRevenue - costPerDay;
+      const days = (here.month - prev.month) * DAYS_PER_MONTH;
+      const before = cumulative;
+      cumulative += margin * days;
+      // Same interval, same trapezoid: the two running totals stay exactly one
+      // flat cost apart, which is what makes them comparable on one chart.
+      cumulativeRevenue += meanRevenue * days;
+      if (paybackMonth === null && before <= 0 && cumulative > 0 && margin > 0) {
+        // -before / margin is the days still owed at `prev`; convert to months.
+        paybackMonth = prev.month + -before / margin / DAYS_PER_MONTH;
+      }
+    }
+    points.push({
+      month: here.month,
+      revenue,
+      cost,
+      margin: revenue - cost,
+      revenuePerMw: revenue * perMw,
+      marginPerMw: (revenue - cost) * perMw,
+      cumulative,
+      cumulativeRevenue,
+      isStep: here.isStep,
+      isRamp: here.isRamp,
+    });
+  }
+
+  if (points.length === 0) return null;
+
+  const firstTput = stepTokPerSec(applied[0]!);
+  const lastApplied = applied.at(-1)!;
+  // Reported rates are what the fleet earns at the end of the window — which is
+  // the latest config at whatever load it has reached by then.
+  const last = points.at(-1)!;
+  const revenuePerDay = last.revenue;
+  const firstRevenuePerDay = revenuePerDayFor(
+    revenuePerSecOf(applied[0]!, inputPrice, outputPrice),
+    availability,
+  );
+
+  return {
+    points,
+    revenuePerDay,
+    firstRevenuePerDay,
+    costPerDay,
+    marginPerDay: revenuePerDay - costPerDay,
+    firstMarginPerDay: firstRevenuePerDay - costPerDay,
+    improvementFactor: firstTput > 0 ? stepTokPerSec(lastApplied) / firstTput : 1,
+    availability,
+    paybackMonth,
+    lifetimeMargin: cumulative,
+    startMonth,
+    endMonth,
+    rampEndMonth,
+    improvementCount: Math.max(0, applied.length - 1),
+  };
+}
diff --git a/packages/app/src/components/calculator/types.ts b/packages/app/src/components/calculator/types.ts
index de5a7ff76..797915f1f 100644
--- a/packages/app/src/components/calculator/types.ts
+++ b/packages/app/src/components/calculator/types.ts
@@ -20,6 +20,32 @@ export interface GPUDataPoint {
   throughput: number; // tokens/sec/gpu total (tput_per_gpu = y in interactivity chart)
   outputThroughput: number; // output tokens/sec/gpu
   inputThroughput: number; // input tokens/sec/gpu
+  /**
+   * Fraction of this config's input tokens served from cache rather than
+   * prefilled — `server_gpu_cache_hit_rate + server_external_cache_hit_rate`,
+   * clamped to [0,1].
+   *
+   * Summed because the two are disjoint in the measured data: across the
+   * production rows carrying both, the sum never exceeds 1 and never exceeds
+   * `theoretical_cache_hit_rate`. Clamped anyway — the GPU figure alone reaches
+   * 1.185 on some rows.
+   *
+   * Undefined for fixed sequences, which carry neither metric on any row. That
+   * absence is what makes cached-input billing a no-op outside agentic traces.
+   */
+  cacheHitRate?: number;
+  /**
+   * Fraction of the tokens this config serves that are input tokens.
+   *
+   * Not simply `inputThroughput / (inputThroughput + outputThroughput)`: on a
+   * disaggregated run those two are per *prefill* and per *decode* chip while
+   * `throughput` is per chip overall, so they sum to as much as 16x the total.
+   * Revenue is charged on the fleet's chips, so it has to use the same
+   * denominator the fleet is sized and costed on — hence a share applied to
+   * `throughput` rather than the two rates read directly. See `inputTokenShare`
+   * in `useThroughputData.ts` for how the share is recovered.
+   */
+  inputTokenShare?: number;
   concurrency: number;
   tp: number;
   precision: string;
@@ -53,6 +79,24 @@ export interface InterpolatedResult {
   tpPerMw: number; // total throughput per megawatt at that operating point
   inputTpPerMw: number; // input throughput per megawatt at that operating point
   outputTpPerMw: number; // output throughput per megawatt at that operating point
+  /**
+   * Cached fraction of input tokens at that operating point, or undefined when
+   * the frontier did not carry a measured rate on every point (which is every
+   * fixed-sequence frontier). See {@link GPUDataPoint.cacheHitRate}.
+   */
+  cacheHitRate?: number;
+  /**
+   * Fraction of the tokens this config serves that are input tokens.
+   *
+   * Not simply `inputThroughput / (inputThroughput + outputThroughput)`: on a
+   * disaggregated run those two are per *prefill* and per *decode* chip while
+   * `throughput` is per chip overall, so they sum to as much as 16x the total.
+   * Revenue is charged on the fleet's chips, so it has to use the same
+   * denominator the fleet is sized and costed on — hence a share applied to
+   * `throughput` rather than the two rates read directly. See `inputTokenShare`
+   * in `useThroughputData.ts` for how the share is recovered.
+   */
+  inputTokenShare?: number;
   concurrency: number; // concurrency at that operating point
   nearestPoints: GPUDataPoint[]; // the data points used for interpolation
   /**
diff --git a/packages/app/src/components/calculator/url-seed.test.ts b/packages/app/src/components/calculator/url-seed.test.ts
index e9082ce94..f67c733ed 100644
--- a/packages/app/src/components/calculator/url-seed.test.ts
+++ b/packages/app/src/components/calculator/url-seed.test.ts
@@ -56,6 +56,27 @@ describe('resolveCalculatorUrlSeed', () => {
     });
   });
 
+  it('includes valid run date and run ID for the single server-seeded provider', () => {
+    expect(
+      resolveCalculatorUrlSeed({
+        g_rundate: '2026-08-20',
+        g_runid: 'run_123',
+      }),
+    ).toEqual({
+      runDate: '2026-08-20',
+      runId: 'run_123',
+    });
+  });
+
+  it('ignores invalid run date and run ID seeds', () => {
+    expect(
+      resolveCalculatorUrlSeed({
+        g_rundate: '08/20/2026',
+        g_runid: 'contains spaces',
+      }),
+    ).toEqual({});
+  });
+
   it('picks the first value when a param is repeated as an array', () => {
     expect(resolveCalculatorUrlSeed({ g_model: ['DeepSeek-V4-Pro', 'GLM-5'] })).toEqual({
       model: Model.DeepSeek_V4_Pro,
diff --git a/packages/app/src/components/calculator/url-seed.ts b/packages/app/src/components/calculator/url-seed.ts
index c7dc02afe..f8741ad48 100644
--- a/packages/app/src/components/calculator/url-seed.ts
+++ b/packages/app/src/components/calculator/url-seed.ts
@@ -14,6 +14,8 @@ export interface CalculatorUrlSeed {
   sequence?: Sequence;
   precisions?: string[];
   percentile?: Percentile;
+  runDate?: string;
+  runId?: string;
 }
 
 function pickString(value: string | string[] | undefined): string | undefined {
@@ -57,6 +59,16 @@ export function resolveCalculatorUrlSeed(
     seed.percentile = percentileParam as Percentile;
   }
 
+  const runDateParam = pickString(sp.g_rundate);
+  if (runDateParam && /^\d{4}-\d{2}-\d{2}$/u.test(runDateParam)) {
+    seed.runDate = runDateParam;
+  }
+
+  const runIdParam = pickString(sp.g_runid);
+  if (runIdParam && /^[A-Za-z0-9_-]{1,64}$/u.test(runIdParam)) {
+    seed.runId = runIdParam;
+  }
+
   return seed;
 }
 
diff --git a/packages/app/src/components/calculator/useHistoricalBest.ts b/packages/app/src/components/calculator/useHistoricalBest.ts
new file mode 100644
index 000000000..f3df57c7a
--- /dev/null
+++ b/packages/app/src/components/calculator/useHistoricalBest.ts
@@ -0,0 +1,130 @@
+'use client';
+
+import { useMemo } from 'react';
+
+import { sequenceToIslOsl } from '@semianalysisai/inferencex-constants';
+
+import { useBenchmarkHistory } from '@/hooks/api/use-benchmark-history';
+import { useStableValue } from '@/hooks/useStableValue';
+import { Percentile, type Model, type Sequence } from '@/lib/data-mappings';
+
+import {
+  bestSoFarProgression,
+  groupHistoryByHwKeyAndDate,
+  selectBestFromGroups,
+  type HistoricalBestOutcome,
+  type HistoricalProgression,
+} from './historical-best';
+import { getComparableTpPerMwForType } from './ThroughputBarChart';
+import type { CalculatorMode, CostProvider, CostType, InterpolatedResult } from './types';
+
+const EMPTY: HistoricalBestOutcome = { best: [], unmeasured: [], datesSeen: 0 };
+
+export interface UseHistoricalBestOptions {
+  model: Model;
+  sequence: Sequence;
+  precisions: string[];
+  targetValue: number;
+  mode: CalculatorMode;
+  costProvider: CostProvider;
+  costType: CostType;
+  percentile?: Percentile;
+  /**
+   * Gates the fetch. This response is several MB, so it must stay false until
+   * the section is actually being used.
+   */
+  enabled: boolean;
+}
+
+export interface UseHistoricalBestResult extends HistoricalBestOutcome {
+  /** Each chip's best-so-far staircase over run dates, for the lifecycle chart. */
+  progressions: HistoricalProgression[];
+  loading: boolean;
+  error: string | null;
+}
+
+/**
+ * Each hwKey's all-time best operating point at the target, from the full run
+ * history rather than the latest run date.
+ *
+ * Deliberately computes every hwKey, not just the visible ones: legend
+ * filtering is applied by the consumer for display, so toggling a series never
+ * rebuilds a frontier.
+ */
+export function useHistoricalBest(options: UseHistoricalBestOptions): UseHistoricalBestResult {
+  const {
+    model,
+    sequence,
+    targetValue,
+    mode,
+    costProvider,
+    costType,
+    percentile = Percentile.P90,
+    enabled,
+  } = options;
+
+  // Callers commonly derive this array inline, so a fresh identity every render
+  // would rebuild every frontier on every render.
+  const precisions = useStableValue(
+    options.precisions,
+    (prev, next) => prev.length === next.length && prev.every((p, i) => p === next[i]),
+  );
+
+  // Agentic traces have no ISL/OSL to key history by; the endpoint takes a
+  // `benchmarkType` instead and drops the sequence filter server-side. The
+  // `view` trim is a no-op there for the same reason — no sequence to key the
+  // metric allowlist on — so those rows come back whole, which is affordable
+  // because the agentic payload is a fraction of a fixed-sequence one.
+  const islOsl = sequenceToIslOsl(sequence);
+
+  const {
+    data: rows,
+    isLoading,
+    error,
+  } = useBenchmarkHistory(model, islOsl?.isl ?? 0, islOsl?.osl ?? 0, {
+    ...(islOsl === null ? { benchmarkType: 'agentic_traces' as const } : {}),
+    view: 'calculator',
+    enabled,
+  });
+
+  // Stage one — the expensive half. Independent of the target, so moving the
+  // interactivity slider does not rebuild every date's frontier.
+  const groups = useMemo(() => {
+    if (!rows) return null;
+    return groupHistoryByHwKeyAndDate({
+      rows,
+      sequence,
+      precisions,
+      percentile,
+      tokenType: costType,
+    });
+  }, [rows, sequence, precisions, percentile, costType]);
+
+  // Stage two — re-read the frontiers at the current operating point. Both the
+  // all-time best and the progression share one selection basis, so the table's
+  // headline figure is always the last rung of the plotted staircase.
+  const selection = useMemo(() => {
+    if (!groups) return { outcome: EMPTY, progressions: [] as HistoricalProgression[] };
+    const selectOptions = {
+      targetValue,
+      mode,
+      costProvider,
+      // The cost matrix's own accessor decides the winner, so the ranking basis
+      // always matches the selected token type — but on a per-total-chip basis,
+      // because a ranking has to be apples-to-apples even where a bar chart can
+      // get away with a footnote. See `getComparableTpPerMwForType`.
+      rank: (result: InterpolatedResult) => getComparableTpPerMwForType(result, costType),
+    };
+    return {
+      outcome: selectBestFromGroups(groups, selectOptions),
+      progressions: bestSoFarProgression(groups, selectOptions),
+    };
+  }, [groups, targetValue, mode, costProvider, costType]);
+
+  return {
+    ...selection.outcome,
+    progressions: selection.progressions,
+    loading: enabled && (isLoading || !rows),
+    error: error ? error.message : null,
+  };
+}
diff --git a/packages/app/src/components/calculator/useThroughputData.test.ts b/packages/app/src/components/calculator/useThroughputData.test.ts
index eff301846..fb5be46bc 100644
--- a/packages/app/src/components/calculator/useThroughputData.test.ts
+++ b/packages/app/src/components/calculator/useThroughputData.test.ts
@@ -6,6 +6,7 @@ import { describe, expect, it } from 'vitest';
 import type { BenchmarkRow } from '@/lib/api';
 import { Percentile, Sequence } from '@/lib/data-mappings';
 import { overlayRunIndex } from '@/lib/overlay-run-style';
+import { SUPPLEMENTAL_BENCHMARK_ROWS } from '@/lib/supplemental-benchmarks';
 
 import type { GPUDataPoint } from './types';
 import {
@@ -26,6 +27,8 @@ const PYTHON_INTERPOLATION_HELPER = resolve(
   '.claude/skills/write-inferencex-blog/iso_interactivity.py',
 );
 
+const classifyByHardware = (hwKey: string) => ({ key: hwKey, meta: { hwKey } });
+
 function interpolateWithPython(request: Record<string, unknown>): number | null {
   const result = spawnSync('python3', [PYTHON_INTERPOLATION_HELPER], {
     input: JSON.stringify(request),
@@ -82,6 +85,31 @@ describe('sign', () => {
   });
 });
 
+describe('snapshot token-metric capabilities', () => {
+  const julyVrRows = SUPPLEMENTAL_BENCHMARK_ROWS.filter((row) => row.hardware === 'vr200');
+  const options = {
+    sequence: Sequence.EightK_OneK,
+    precisions: ['fp4'],
+    classify: classifyByHardware,
+  };
+
+  it('keeps July VR200 for output calculator metrics and hides total/input', () => {
+    expect(
+      Object.keys(buildGpuGroups(julyVrRows, { ...options, tokenType: 'output' }).grouped),
+    ).toEqual(['vr200_rubin-july']);
+    expect(buildGpuGroups(julyVrRows, { ...options, tokenType: 'total' }).grouped).toEqual({});
+    expect(buildGpuGroups(julyVrRows, { ...options, tokenType: 'input' }).grouped).toEqual({});
+  });
+
+  it('applies the same restriction to unofficial overlay rows', () => {
+    const overlayRows = julyVrRows.map((row) => ({
+      ...row,
+      run_url: 'https://github.com/org/repo/actions/runs/1',
+    }));
+    expect(buildGpuGroups(overlayRows, { ...options, tokenType: 'total' }).grouped).toEqual({});
+  });
+});
+
 // =========================================================================
 // getCostField()
 // =========================================================================
@@ -389,6 +417,73 @@ describe('interpolateForGPU', () => {
     expect(result!.nearestPoints).toHaveLength(1);
   });
 
+  describe('cached-input fraction', () => {
+    // 300 → 500 tok/s across the range so the frontier is well formed; the cache
+    // rate is what these assertions are about.
+    const withCache = (rates: (number | undefined)[]) =>
+      rates.map((cacheHitRate, i) =>
+        makePoint({
+          interactivity: 20 + i * 20,
+          throughput: 500 - i * 100,
+          inputThroughput: 400 - i * 80,
+          ...(cacheHitRate === undefined ? {} : { cacheHitRate }),
+        }),
+      );
+
+    it('interpolates the rate between measured points', () => {
+      const result = interpolateForGPU(
+        withCache([0.4, 0.8]),
+        30,
+        'interactivity_to_throughput',
+        'costh',
+      );
+      // Halfway along the frontier, so between the two measured rates and
+      // strictly inside them — not pinned to either end.
+      expect(result!.cacheHitRate).toBeGreaterThan(0.4);
+      expect(result!.cacheHitRate).toBeLessThan(0.8);
+    });
+
+    it('stays inside the frontier range rather than overshooting', () => {
+      const result = interpolateForGPU(
+        withCache([0.1, 0.9, 0.2]),
+        45,
+        'interactivity_to_throughput',
+        'costh',
+      );
+      expect(result!.cacheHitRate).toBeGreaterThanOrEqual(0.1);
+      expect(result!.cacheHitRate).toBeLessThanOrEqual(0.9);
+    });
+
+    it('is undefined when no point carries a rate — every fixed sequence', () => {
+      const result = interpolateForGPU(
+        withCache([undefined, undefined]),
+        30,
+        'interactivity_to_throughput',
+        'costh',
+      );
+      expect(result!.cacheHitRate).toBeUndefined();
+    });
+
+    it('opts the whole frontier out when only some points carry a rate', () => {
+      // Splining a 0 in for the unmeasured point would invent a dip in the
+      // cached fraction and overstate the billable rate there. Opting out bills
+      // every input token at full price instead — wrong in the safe direction.
+      const result = interpolateForGPU(
+        withCache([0.9, undefined]),
+        30,
+        'interactivity_to_throughput',
+        'costh',
+      );
+      expect(result!.cacheHitRate).toBeUndefined();
+    });
+
+    it('carries the rate through the single-point path too', () => {
+      const points = [makePoint({ interactivity: 30, throughput: 500, cacheHitRate: 0.77 })];
+      const result = interpolateForGPU(points, 25, 'interactivity_to_throughput', 'costh');
+      expect(result!.cacheHitRate).toBe(0.77);
+    });
+  });
+
   it('single GPU clamps any target to the lone pareto-front point', () => {
     const points = [makePoint({ interactivity: 30, throughput: 500 })];
     const result = interpolateForGPU(points, 25, 'interactivity_to_throughput', 'costh');
@@ -1082,6 +1177,204 @@ describe('buildGpuGroups', () => {
     expect(first.tpPerMw).toBeGreaterThan(0);
   });
 
+  describe('cached-input fraction', () => {
+    const withMetrics = (extra: Record<string, number>) =>
+      makeRow({
+        metrics: {
+          median_intvty: 50,
+          tput_per_gpu: 900,
+          output_tput_per_gpu: 300,
+          input_tput_per_gpu: 600,
+          ...extra,
+        },
+      });
+    const only = (row: BenchmarkRow) => {
+      const { grouped } = buildGpuGroups([row], {
+        ...shared,
+        classify: singlePrecisionClassify,
+      });
+      return Object.values(grouped)[0][0];
+    };
+
+    it('is absent when the row records no cache metric — every fixed-sequence row', () => {
+      // Not zero: absent. Zero would mean "measured, and nothing was cached",
+      // which is a claim the data does not make.
+      expect(only(makeRow()).cacheHitRate).toBeUndefined();
+    });
+
+    it('sums the GPU and external hit rates', () => {
+      // Disjoint in the measured data: across production rows carrying both, the
+      // sum never exceeds 1 nor the theoretical ceiling.
+      expect(
+        only(withMetrics({ server_gpu_cache_hit_rate: 0.66, server_external_cache_hit_rate: 0.25 }))
+          .cacheHitRate,
+      ).toBeCloseTo(0.91, 10);
+    });
+
+    // The three tiers do not all stack. These three cases are the shapes the
+    // production data actually comes in; see `cacheHitRateOf` for the counts.
+    it('ignores the CPU tier when an external rate is reported — external already counts it', () => {
+      // The shape of 106 of 132 production rows with a non-zero CPU rate. Adding
+      // CPU here is what breached `theoretical_cache_hit_rate` on 56 of them.
+      expect(
+        only(
+          withMetrics({
+            server_gpu_cache_hit_rate: 0.819,
+            server_external_cache_hit_rate: 0.06,
+            server_cpu_cache_hit_rate: 0.067,
+          }),
+        ).cacheHitRate,
+      ).toBeCloseTo(0.879, 10);
+    });
+
+    it('adds the CPU tier when no external rate is reported', () => {
+      // The shape of the 26 offload-on rows with no external figure, where
+      // `gpu + cpu` breaches the ceiling 0 times. Dropping CPU understated the
+      // cached share by ~5.5pp on these, which overstates revenue.
+      expect(
+        only(withMetrics({ server_gpu_cache_hit_rate: 0.771, server_cpu_cache_hit_rate: 0.055 }))
+          .cacheHitRate,
+      ).toBeCloseTo(0.826, 10);
+    });
+
+    it('reads the CPU tier alone, and an external rate of zero still suppresses it', () => {
+      expect(only(withMetrics({ server_cpu_cache_hit_rate: 0.5 })).cacheHitRate).toBeCloseTo(
+        0.5,
+        10,
+      );
+      // A reported zero is a measurement, not an absence: the router saw no
+      // external hits, and it is still the figure that accounts for the offload
+      // tier. Reading 0.5 here would double-count. No production row has this
+      // shape today, so this pins a deliberate choice about unobserved data —
+      // it is not a regression test for something measured.
+      expect(
+        only(withMetrics({ server_external_cache_hit_rate: 0, server_cpu_cache_hit_rate: 0.5 }))
+          .cacheHitRate,
+      ).toBe(0);
+    });
+
+    it('reads either metric alone', () => {
+      expect(only(withMetrics({ server_gpu_cache_hit_rate: 0.4 })).cacheHitRate).toBeCloseTo(
+        0.4,
+        10,
+      );
+      expect(only(withMetrics({ server_external_cache_hit_rate: 0.3 })).cacheHitRate).toBeCloseTo(
+        0.3,
+        10,
+      );
+    });
+
+    it('clamps into [0,1] — real rows report a GPU rate as high as 1.185', () => {
+      expect(only(withMetrics({ server_gpu_cache_hit_rate: 1.185 })).cacheHitRate).toBe(1);
+      expect(only(withMetrics({ server_gpu_cache_hit_rate: -0.2 })).cacheHitRate).toBe(0);
+    });
+  });
+
+  describe('input token share', () => {
+    const only = (row: BenchmarkRow) => {
+      const { grouped } = buildGpuGroups([row], { ...shared, classify: singlePrecisionClassify });
+      return Object.values(grouped)[0][0];
+    };
+
+    it('takes the share from the rates when they agree with the total', () => {
+      // Self-consistent (700 + 300 = 1000), so the measured split is the split —
+      // even where it disagrees with the sequence shape.
+      const point = only(
+        makeRow({
+          metrics: {
+            median_intvty: 50,
+            tput_per_gpu: 1000,
+            input_tput_per_gpu: 700,
+            output_tput_per_gpu: 300,
+          },
+        }),
+      );
+      expect(point.inputTokenShare).toBeCloseTo(0.7, 9);
+    });
+
+    it('falls back to ISL:OSL when the rates are on a different denominator', () => {
+      // The disaggregated shape, with the real arithmetic: 16 prefill + 8 decode
+      // chips serving 8192:1024. Input is per prefill chip (6400/16 = 400),
+      // output per decode chip (800/8 = 100), total per chip overall
+      // (7200/24 = 300). The rates sum to 1.667x the total, and the split they
+      // imply — 0.8 — is not the split the workload has, which is 8192/9216.
+      const point = only(
+        makeRow({
+          disagg: true,
+          isl: 1024,
+          osl: 1024,
+          metrics: {
+            median_intvty: 50,
+            tput_per_gpu: 300,
+            input_tput_per_gpu: 400,
+            output_tput_per_gpu: 100,
+          },
+        }),
+      );
+      // 1k/1k here, because that is the sequence this describe block selects.
+      expect(point.inputTokenShare).toBeCloseTo(0.5, 9);
+      expect(point.inputTokenShare).not.toBeCloseTo(0.8, 2);
+    });
+
+    it('falls back to the run token counts when there is no fixed sequence', () => {
+      // Agentic traces have no ISL/OSL to fall back to, so the run's own
+      // prompt:generation counts pin the mix instead.
+      const { grouped } = buildGpuGroups(
+        [
+          makeRow({
+            disagg: true,
+            benchmark_type: 'agentic_traces',
+            isl: null,
+            osl: null,
+            metrics: {
+              p90_itl: 1 / 50,
+              tput_per_gpu: 300,
+              input_tput_per_gpu: 400,
+              output_tput_per_gpu: 100,
+              total_prompt_tokens: 9000,
+              total_generation_tokens: 1000,
+            },
+          }),
+        ],
+        {
+          sequence: Sequence.AgenticTraces,
+          precisions: ['fp4'],
+          classify: singlePrecisionClassify,
+        },
+      );
+      expect(Object.values(grouped)[0][0].inputTokenShare).toBeCloseTo(0.9, 9);
+    });
+
+    it('leaves the share unknown when disaggregated rates have no trustworthy mix', () => {
+      const { grouped } = buildGpuGroups(
+        [
+          makeRow({
+            disagg: true,
+            benchmark_type: 'agentic_traces',
+            isl: null,
+            osl: null,
+            metrics: {
+              p90_itl: 1 / 50,
+              tput_per_gpu: 300,
+              input_tput_per_gpu: 400,
+              output_tput_per_gpu: 100,
+            },
+          }),
+        ],
+        {
+          sequence: Sequence.AgenticTraces,
+          precisions: ['fp4'],
+          classify: singlePrecisionClassify,
+        },
+      );
+
+      // 400 / (400 + 100) is a ratio of per-pool rates, not a fleet-wide
+      // token share. Leaving it absent makes lifecycle revenue charge no input
+      // tokens instead of inventing an 80% input mix.
+      expect(Object.values(grouped)[0][0].inputTokenShare).toBeUndefined();
+    });
+  });
+
   it('drops rows whose isl/osl do not match the selected sequence', () => {
     const { grouped } = buildGpuGroups(
       [makeRow({ isl: 8192, osl: 1024 }), makeRow({ isl: null, osl: null })],
@@ -1216,7 +1509,13 @@ describe('buildGpuGroups', () => {
     });
   });
 
-  it('restricts agentic interpolation to the same date-scoped e2e Pareto winners as the chart', () => {
+  it('keeps an e2e-dominated agentic point, because grouping no longer gates on the e2e frontier', () => {
+    // Until #736 this dropped every point that lost on E2E normalized
+    // interactivity, so grouping matched the chart's canonical frontier. That
+    // gating was removed deliberately: a frontier is now computed from the axes
+    // the reader actually selected, so grouping keeps every measured point and
+    // eligibility is decided downstream. Concurrency 2 is dominated on e2e by
+    // concurrency 1 and must survive anyway.
     const agenticRow = (
       conc: number,
       interactivity: number,
@@ -1245,7 +1544,6 @@ describe('buildGpuGroups', () => {
         agenticRow(1, 100, 20, 900),
         agenticRow(2, 80, 30, 800), // e2e-dominated by concurrency 1
         agenticRow(4, 60, 40, 1200),
-        // A different date gets its own e2e frontier and must survive.
         agenticRow(8, 40, 50, 700, '2026-07-20'),
       ],
       {
@@ -1257,7 +1555,7 @@ describe('buildGpuGroups', () => {
     );
 
     const points = Object.values(grouped)[0];
-    expect(points.map((point) => point.concurrency).toSorted()).toEqual([1, 4, 8]);
+    expect(points.map((point) => point.concurrency).toSorted()).toEqual([1, 2, 4, 8]);
   });
 
   it('keeps agentic overlay frontiers isolated per unofficial run', () => {
diff --git a/packages/app/src/components/calculator/useThroughputData.ts b/packages/app/src/components/calculator/useThroughputData.ts
index fe2c25493..a2b44cba2 100644
--- a/packages/app/src/components/calculator/useThroughputData.ts
+++ b/packages/app/src/components/calculator/useThroughputData.ts
@@ -12,6 +12,7 @@ import { getHardwareKey } from '@/lib/chart-utils';
 import { getModelSortIndex, getHardwareConfig, getGpuSpecs } from '@/lib/constants';
 import { Percentile, Sequence, type Model } from '@/lib/data-mappings';
 import { overlayRunIndex } from '@/lib/overlay-run-style';
+import { supportsTokenMetric } from '@/lib/supplemental-benchmarks';
 
 import {
   getCostField,
@@ -24,8 +25,7 @@ import {
   recoverReciprocalNumerator,
   sign,
 } from './interpolation';
-import { restrictAgenticPointsToE2eFrontier } from '@/lib/agentic-frontier';
-import type { CostProvider, GPUDataPoint, InterpolatedResult } from './types';
+import type { CostProvider, CostType, GPUDataPoint, InterpolatedResult } from './types';
 
 // Re-export pure functions so existing imports from this module keep working.
 export {
@@ -66,6 +66,92 @@ function getAgenticMetric(
   return typeof value === 'number' ? value : 0;
 }
 
+/**
+ * Fraction of a row's input tokens served from cache, or null when the row
+ * carries no cache metric at all — which is every fixed-sequence row.
+ *
+ * Three tiers are reported and they do not all stack. Measured across 326
+ * production agentic rows, checked against each row's own
+ * `theoretical_cache_hit_rate` ceiling:
+ *
+ * - **GPU + external are disjoint.** Their sum never exceeds 1 nor the ceiling
+ *   on any row carrying both, so they add.
+ * - **The CPU-offload tier double-counts external where both are reported.**
+ *   `gpu + external + cpu` breaches the ceiling on 56 of the 132 rows with a
+ *   non-zero CPU rate, and *every* breach is a row that also reports an
+ *   external rate — the router-side external figure already contains the
+ *   offload tier there.
+ * - **Where no external rate is reported, the CPU tier is real and disjoint.**
+ *   On the 26 offload-on rows with no external figure, `gpu + cpu` breaches the
+ *   ceiling 0 times, median CPU rate 0.055.
+ *
+ * Hence the conditional: external when present, CPU only in its absence. Adding
+ * CPU unconditionally would overstate the cached share; dropping it entirely —
+ * which this did before — understated it by a median 5.54pp on those 26 rows,
+ * and an understated cached share bills more input at the fresh-token price and
+ * so *overstates* revenue (by a median 27% on the input leg there). That is the
+ * direction worth spending a branch to avoid.
+ *
+ * A reported external `0` suppresses CPU rather than counting as an absence: a
+ * router reporting no external hits has still accounted for the offload tier. No
+ * production row currently has that shape, so that arm is a deliberate choice
+ * about unobserved data, pinned by a test rather than measured.
+ *
+ * The clamp is not decorative — the GPU figure alone reaches 1.185 on some rows.
+ */
+function cacheHitRateOf(m: Record<string, number>): number | null {
+  const gpu = m.server_gpu_cache_hit_rate;
+  const external = m.server_external_cache_hit_rate;
+  const cpu = m.server_cpu_cache_hit_rate;
+  const hasExternal = typeof external === 'number';
+  // Only the tier that is not already counted by `external` is added.
+  const secondary = hasExternal ? external : typeof cpu === 'number' ? cpu : undefined;
+  if (typeof gpu !== 'number' && secondary === undefined) return null;
+  const sum = (typeof gpu === 'number' ? gpu : 0) + (secondary ?? 0);
+  return Math.max(0, Math.min(1, sum));
+}
+
+/**
+ * Fraction of a config's tokens that are input tokens.
+ *
+ * The obvious formula — `input / (input + output)` off the per-GPU rates — is
+ * wrong for disaggregated runs, and wrong by a lot. Those rows report
+ * `input_tput_per_gpu` per *prefill* chip and `output_tput_per_gpu` per *decode*
+ * chip, while `tput_per_gpu` is per chip overall. Across production history the
+ * two rates sum to between 1.0x and 16.1x the total, exactly tracking
+ * `(prefill + decode) x (isl/prefill + osl/decode) / (isl + osl)`. Aggregated
+ * rows sum to 1.0000x on all 937 of them.
+ *
+ * So: trust the measured rates when they are self-consistent, and fall back to
+ * the structural mix when they are not. For a fixed sequence the structural mix
+ * is exactly ISL:OSL — every request has that shape, so no config can change it.
+ * For agentic traces it is the run's own prompt:generation token counts.
+ *
+ * Returns null when nothing in the row pins the mix down.
+ */
+function inputTokenShare(row: BenchmarkRow, inputTput: number, outputTput: number): number | null {
+  const tput = row.metrics.tput_per_gpu ?? 0;
+  const sum = inputTput + outputTput;
+  // Self-consistent: the rates share the denominator `tput_per_gpu` uses, so the
+  // split they imply is the measured one. 1% covers float noise, not a
+  // prefill/decode mismatch (the smallest of those in production is 1.63x).
+  if (tput > 0 && sum > 0 && Math.abs(sum / tput - 1) <= 0.01) return inputTput / sum;
+
+  const { isl, osl } = row;
+  if (typeof isl === 'number' && typeof osl === 'number' && isl + osl > 0) {
+    return isl / (isl + osl);
+  }
+  const prompt = row.metrics.total_prompt_tokens;
+  const generated = row.metrics.total_generation_tokens;
+  if (typeof prompt === 'number' && typeof generated === 'number' && prompt + generated > 0) {
+    return prompt / (prompt + generated);
+  }
+  // The remaining rates are known to be on incompatible denominators. Treat
+  // the mix as unknown rather than turning a per-prefill/per-decode ratio into
+  // a fleet-wide token share and billing input volume the fleet did not serve.
+  return null;
+}
+
 /**
  * Build `GPUDataPoint` groups from raw benchmark rows.
  *
@@ -80,6 +166,8 @@ export function buildGpuGroups<M extends GroupMeta>(
     precisions: string[];
     /** Agentic x/e2e latency percentile. Fixed-sequence rows keep the median. */
     percentile?: Percentile;
+    /** Token basis selected by the consumer; applies to official and overlay rows. */
+    tokenType?: CostType;
     /** Derive a row's group key + metadata. Return null to drop the row. */
     classify: (hwKey: string, row: BenchmarkRow) => { key: string; meta: M } | null;
   },
@@ -88,7 +176,13 @@ export function buildGpuGroups<M extends GroupMeta>(
   groupMeta: Record<string, M>;
   hwConfigMap: HardwareConfig;
 } {
-  const { sequence, precisions, percentile = Percentile.P90, classify } = options;
+  const {
+    sequence,
+    precisions,
+    percentile = Percentile.P90,
+    tokenType = 'total',
+    classify,
+  } = options;
   const grouped: Record<string, GPUDataPoint[]> = {};
   const groupMeta: Record<string, M> = {};
   const hwConfigMap: HardwareConfig = {};
@@ -96,6 +190,7 @@ export function buildGpuGroups<M extends GroupMeta>(
   for (const row of rows) {
     if (rowToSequence(row) !== sequence) continue;
     if (!precisions.includes(row.precision)) continue;
+    if (!supportsTokenMetric(row, tokenType)) continue;
 
     const entry = rowToAggDataEntry(row);
     const hwKey = getHardwareKey(entry);
@@ -112,6 +207,8 @@ export function buildGpuGroups<M extends GroupMeta>(
     const tput = m.tput_per_gpu ?? 0;
     const outputTput = m.output_tput_per_gpu ?? tput;
     const inputTput = m.input_tput_per_gpu ?? 0;
+    const cacheHitRate = cacheHitRateOf(m);
+    const tokenShare = inputTokenShare(row, inputTput, outputTput);
     const specs = getGpuSpecs(hwKey);
     const power = specs.power;
 
@@ -133,6 +230,8 @@ export function buildGpuGroups<M extends GroupMeta>(
       throughput: tput,
       outputThroughput: outputTput,
       inputThroughput: inputTput,
+      ...(cacheHitRate === null ? {} : { cacheHitRate }),
+      ...(tokenShare === null ? {} : { inputTokenShare: tokenShare }),
       concurrency: row.conc,
       tp: row.decode_tp,
       precision: row.precision,
@@ -154,18 +253,6 @@ export function buildGpuGroups<M extends GroupMeta>(
     });
   }
 
-  if (sequence === Sequence.AgenticTraces) {
-    for (const groupKey of Object.keys(grouped)) {
-      const restricted = restrictAgenticPointsToE2eFrontier(grouped[groupKey]);
-      if (restricted.length === 0) {
-        delete grouped[groupKey];
-        delete groupMeta[groupKey];
-      } else {
-        grouped[groupKey] = restricted;
-      }
-    }
-  }
-
   return { grouped, groupMeta, hwConfigMap };
 }
 
@@ -188,16 +275,35 @@ export function useThroughputData(
   selectedRunDate: string,
   overlay?: OverlayInput,
   selectedPercentile: Percentile = Percentile.P90,
+  initialRows?: BenchmarkRow[],
+  enabled = true,
+  selectedTokenType: CostType = 'total',
 ) {
+  const initialCacheScope = useMemo(
+    () =>
+      initialRows
+        ? `compare-initial:${[...new Set(initialRows.map((row) => row.hardware))].toSorted().join(',')}`
+        : undefined,
+    [initialRows],
+  );
   // Reuse the same API + React Query cache as the inference charts
   const {
     data: allRows,
     isLoading: queryLoading,
     error: queryError,
-  } = useBenchmarks(selectedModel, selectedRunDate, true, undefined, undefined, {
-    type: 'calculator',
-    sequence: selectedSequence,
-  });
+  } = useBenchmarks(
+    selectedModel,
+    selectedRunDate,
+    enabled,
+    undefined,
+    undefined,
+    {
+      type: 'calculator',
+      sequence: selectedSequence,
+      ...(initialCacheScope ? { cacheScope: initialCacheScope } : {}),
+    },
+    initialRows,
+  );
 
   const loading = queryLoading || !allRows;
   const error = queryError ? queryError.message : null;
@@ -232,6 +338,7 @@ export function useThroughputData(
       sequence: selectedSequence,
       precisions: selectedPrecisions,
       percentile: selectedPercentile,
+      tokenType: selectedTokenType,
     };
 
     const official = buildGpuGroups<GroupMeta>(allRows, {
@@ -285,6 +392,7 @@ export function useThroughputData(
     selectedSequence,
     selectedPrecisions,
     selectedPercentile,
+    selectedTokenType,
     overlayRows,
     runIndexByUrl,
   ]);
diff --git a/packages/app/src/components/chips/chip-page-sections.tsx b/packages/app/src/components/chips/chip-page-sections.tsx
new file mode 100644
index 000000000..b4795b17e
--- /dev/null
+++ b/packages/app/src/components/chips/chip-page-sections.tsx
@@ -0,0 +1,668 @@
+/**
+ * Server-rendered sections shared by the /chips and /zh/chips pages.
+ *
+ * Locale is passed as a prop per AGENTS.md rule 5 (server components take an
+ * optional `locale`); all prose comes from `chip-pages.ts` (en) and
+ * `chip-pages-zh.ts` (zh), and every number is derived from GPU_SPECS and
+ * HW_REGISTRY so the pages can never drift from the dashboard.
+ */
+import Link from 'next/link';
+
+import { JsonLd } from '@/components/json-ld';
+import { Card } from '@/components/ui/card';
+import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
+import { getPostBySlug } from '@/lib/blog';
+import {
+  CHIP_VS_HIGHLIGHT_LABELS_EN,
+  buildChipFaq,
+  buildChipVsFaq,
+  buildChipVsHighlights,
+  type ChipFaqItem,
+  type ChipPageEntry,
+  type ChipVsPage,
+  getAllChipPages,
+  getAllChipVsPages,
+  getChipHw,
+  getChipSpec,
+} from '@/lib/chip-pages';
+import {
+  CHIP_VS_HIGHLIGHT_LABELS_ZH,
+  buildZhChipFaq,
+  buildZhChipVsFaq,
+  getZhChipTranslation,
+  localizeVsHighlightValueZh,
+} from '@/lib/chip-pages-zh';
+import { getGlossaryEntry } from '@/lib/glossary';
+import { getZhGlossaryEntry } from '@/lib/glossary-zh';
+import { type Locale, localePath, ZH_LANG_TAG } from '@/lib/i18n';
+import {
+  ACCELERATOR_MODEL_TITLE,
+  ACCELERATOR_MODEL_URL,
+  SITE_URL,
+  TCO_MODEL_TITLE,
+  TCO_SOURCE_URL,
+} from '@semianalysisai/inferencex-constants';
+
+const STRINGS = {
+  en: {
+    breadcrumbRoot: 'AI Chips',
+    backToIndex: 'All AI inference chips',
+    indexTitle: 'AI Chips for LLM Inference',
+    indexIntro:
+      'Specs, cloud pricing and continuously measured inference benchmarks for every chip InferenceX covers. Each page joins the hardware data used by the live dashboard with the SemiAnalysis AI Cloud TCO rates.',
+    chipPages: 'Chip pages',
+    vsPages: 'Head-to-head comparisons',
+    specsHeading: 'Specifications',
+    pricingHeading: 'Power & cloud pricing',
+    faqHeading: 'Frequently asked questions',
+    relatedArticles: 'Benchmark analyses',
+    relatedTerms: 'Glossary',
+    relatedChips: 'Related chips',
+    liveCta: 'See live benchmark results',
+    liveCtaBody:
+      'Every number above is static hardware data. Delivered tokens per second, cost per million tokens and energy per token are measured continuously on the dashboard:',
+    liveLinkInference: 'Live inference dashboard',
+    liveLinkCompare: 'Chip-vs-chip model comparisons',
+    liveLinkPerDollar: 'Performance per dollar',
+    liveLinkOverview: 'Benchmark methodology & overview',
+    modelsHeading: 'Go deeper with the SemiAnalysis models',
+    modelsBody:
+      'InferenceX measures delivered inference performance. The SemiAnalysis institutional models cover the market behind these chips: who ships them, who buys them, and what they cost to own.',
+    modelsAccDesc:
+      'SKU-level AI accelerator shipments, pricing and specifications, from foundry wafer starts and HBM supply through customer-level installed base, quarterly with multi-year forecasts.',
+    modelsTcoDesc:
+      'The source of the hourly rates on this page: all-in GPU cost of ownership built up from server capex, power, colocation and cost of capital, with rental price scenarios and a full cluster finance suite.',
+    pricingSource: 'Source: $/chip/hr rate tiers from the SemiAnalysis',
+    specLabels: {
+      vendor: 'Vendor',
+      arch: 'Architecture',
+      memory: 'Memory (usable)',
+      memoryBandwidth: 'Memory bandwidth',
+      fp4: 'FP4 dense TFLOP/s',
+      fp8: 'FP8 dense TFLOP/s',
+      bf16: 'BF16 dense TFLOP/s',
+      scaleUp: 'Scale-up interconnect',
+      scaleUpBandwidth: 'Scale-up bandwidth per chip',
+      worldSize: 'Scale-up world size',
+      scaleUpTopology: 'Scale-up topology',
+      scaleOut: 'Scale-out network',
+      nic: 'NIC',
+      tdp: 'TDP per chip',
+      allInPower: 'All-in power per chip',
+      costHyperscaler: 'Hyperscaler $/chip/hr',
+      costNeocloud: 'Neocloud $/chip/hr',
+      costRetail: 'Retail $/chip/hr',
+    },
+    notSupported: 'Not supported',
+    none: 'N/A (NVLink domain only)',
+    vsRatioNote: 'Ratios are spec-sheet values; see the live compare pages for measured deltas.',
+    vsRatioNa: 'n/a',
+    vsSpecTableCaption: 'Spec-sheet comparison',
+    overviewHeading: 'Overview',
+    benchmarkHeading: 'How InferenceX benchmarks it',
+    vsIntroPrefix: 'Spec-sheet and pricing comparison of',
+    vsIntroAnd: 'and',
+    vsIntroSuffix:
+      'with links to continuously measured LLM inference benchmarks on identical workloads.',
+  },
+  zh: {
+    breadcrumbRoot: 'AI 芯片',
+    backToIndex: '全部 AI 推理芯片',
+    indexTitle: '面向 LLM 推理的 AI 芯片',
+    indexIntro:
+      'InferenceX 覆盖的每一款芯片的规格、云端价格与持续测量的推理基准测试。每个页面都将实时仪表板使用的硬件数据与 SemiAnalysis AI Cloud TCO 费率相结合。',
+    chipPages: '芯片页面',
+    vsPages: '芯片对比',
+    specsHeading: '规格参数',
+    pricingHeading: '功耗与云端价格',
+    faqHeading: '常见问题',
+    relatedArticles: '基准测试分析',
+    relatedTerms: '术语表',
+    relatedChips: '相关芯片',
+    liveCta: '查看实时基准测试结果',
+    liveCtaBody:
+      '以上都是静态硬件数据。实际交付的 token 吞吐量、每百万 token 成本和每 token 能耗在仪表板上持续测量：',
+    liveLinkInference: '实时推理仪表板',
+    liveLinkCompare: '芯片对芯片模型对比',
+    liveLinkPerDollar: '每美元性能',
+    liveLinkOverview: '基准测试方法与总览',
+    modelsHeading: '通过 SemiAnalysis 行业模型深入研究',
+    modelsBody:
+      'InferenceX 测量实际交付的推理性能，而 SemiAnalysis 机构级行业模型覆盖这些芯片背后的市场：谁在出货、谁在采购、拥有成本是多少。',
+    modelsAccDesc:
+      '按 SKU 跟踪 AI 加速器的出货量、价格与规格，覆盖从晶圆代工投片、HBM 供应到客户级装机量的完整链路，按季度更新并包含多年预测。',
+    modelsTcoDesc:
+      '本页每小时费率的数据来源：从服务器资本开支、电力、托管与资金成本自下而上构建 GPU 全含拥有成本，并提供租赁价格情景与完整的集群财务模型。',
+    pricingSource: '来源：$/芯片/小时 费率取自 SemiAnalysis',
+    specLabels: {
+      vendor: '厂商',
+      arch: '架构',
+      memory: '显存（可用）',
+      memoryBandwidth: '内存带宽',
+      fp4: 'FP4 稠密 TFLOP/s',
+      fp8: 'FP8 稠密 TFLOP/s',
+      bf16: 'BF16 稠密 TFLOP/s',
+      scaleUp: 'Scale-up 互联',
+      scaleUpBandwidth: '单芯片 scale-up 带宽',
+      worldSize: 'Scale-up 域规模',
+      scaleUpTopology: 'Scale-up 拓扑',
+      scaleOut: 'Scale-out 网络',
+      nic: '网卡',
+      tdp: '单芯片 TDP',
+      allInPower: '单芯片整机功耗',
+      costHyperscaler: '超大规模云 $/芯片/小时',
+      costNeocloud: 'Neocloud $/芯片/小时',
+      costRetail: '零售档 $/芯片/小时',
+    },
+    notSupported: '不支持',
+    none: '无（仅 NVLink 域内互联）',
+    vsRatioNote: '倍数为纸面规格对比；实测差距请见实时对比页面。',
+    vsRatioNa: '不适用',
+    vsSpecTableCaption: '纸面规格对比',
+    overviewHeading: '概览',
+    benchmarkHeading: 'InferenceX 如何测试这款芯片',
+    vsIntroPrefix: '本页对比',
+    vsIntroAnd: '与',
+    vsIntroSuffix: '的纸面规格与价格，并链接到相同工作负载上持续测量的 LLM 推理基准测试。',
+  },
+} as const;
+
+function chipFaq(entry: ChipPageEntry, locale: Locale): readonly ChipFaqItem[] {
+  return locale === 'zh' ? buildZhChipFaq(entry) : buildChipFaq(entry);
+}
+
+function vsFaq(page: ChipVsPage, locale: Locale): readonly ChipFaqItem[] {
+  return locale === 'zh' ? buildZhChipVsFaq(page) : buildChipVsFaq(page);
+}
+
+function faqJsonLd(faq: readonly ChipFaqItem[], url: string, locale: Locale) {
+  return {
+    '@type': 'FAQPage',
+    '@id': `${url}#faq`,
+    ...(locale === 'zh' ? { inLanguage: ZH_LANG_TAG } : {}),
+    mainEntity: faq.map((item) => ({
+      '@type': 'Question',
+      name: item.question,
+      acceptedAnswer: { '@type': 'Answer', text: item.answer },
+    })),
+  };
+}
+
+function breadcrumbJsonLd(locale: Locale, leafName: string, leafPath: string) {
+  const t = STRINGS[locale];
+  return {
+    '@type': 'BreadcrumbList',
+    itemListElement: [
+      {
+        '@type': 'ListItem',
+        position: 1,
+        name: t.breadcrumbRoot,
+        item: `${SITE_URL}${localePath('/chips', locale)}`,
+      },
+      {
+        '@type': 'ListItem',
+        position: 2,
+        name: leafName,
+        item: `${SITE_URL}${localePath(leafPath, locale)}`,
+      },
+    ],
+  };
+}
+
+const FaqSection = ({ faq, locale }: { faq: readonly ChipFaqItem[]; locale: Locale }) => (
+  <section aria-labelledby="chip-faq" className="mt-10 border-t border-border/50 pt-10">
+    <h2 id="chip-faq" className="text-xl font-semibold tracking-tight">
+      {STRINGS[locale].faqHeading}
+    </h2>
+    <dl className="mt-4 space-y-6">
+      {faq.map((item) => (
+        <div key={item.question}>
+          <dt className="font-medium">{item.question}</dt>
+          <dd className="mt-2 leading-7 text-muted-foreground">{item.answer}</dd>
+        </div>
+      ))}
+    </dl>
+  </section>
+);
+
+const LiveResultsSection = ({ locale }: { locale: Locale }) => {
+  const t = STRINGS[locale];
+  const links = [
+    { href: localePath('/', locale), label: t.liveLinkInference },
+    { href: localePath('/compare', locale), label: t.liveLinkCompare },
+    { href: localePath('/compare-per-dollar', locale), label: t.liveLinkPerDollar },
+    { href: localePath('/overview', locale), label: t.liveLinkOverview },
+  ];
+  return (
+    <section
+      aria-labelledby="chip-live-results"
+      className="mt-10 rounded-xl border border-brand/20 bg-brand/6 p-5 md:p-6"
+    >
+      <h2 id="chip-live-results" className="text-xl font-semibold tracking-tight">
+        {t.liveCta}
+      </h2>
+      <p className="mt-2 leading-7 text-muted-foreground">{t.liveCtaBody}</p>
+      <ul className="mt-3 space-y-1">
+        {links.map((link) => (
+          <li key={link.href}>
+            <Link href={link.href} className="text-sm font-medium text-brand hover:underline">
+              {link.label} →
+            </Link>
+          </li>
+        ))}
+      </ul>
+    </section>
+  );
+};
+
+const SEMIANALYSIS_MODELS = [
+  {
+    href: ACCELERATOR_MODEL_URL,
+    title: ACCELERATOR_MODEL_TITLE,
+    descKey: 'modelsAccDesc',
+  },
+  {
+    href: TCO_SOURCE_URL,
+    title: TCO_MODEL_TITLE,
+    descKey: 'modelsTcoDesc',
+  },
+] as const;
+
+const ModelsSection = ({ locale }: { locale: Locale }) => {
+  const t = STRINGS[locale];
+  return (
+    <section aria-labelledby="chip-models" className="mt-10 border-t border-border/50 pt-10">
+      <h2 id="chip-models" className="text-xl font-semibold tracking-tight">
+        {t.modelsHeading}
+      </h2>
+      <p className="mt-2 leading-7 text-muted-foreground">{t.modelsBody}</p>
+      <div className="mt-4 grid gap-4 sm:grid-cols-2">
+        {SEMIANALYSIS_MODELS.map((model) => (
+          <a
+            key={model.href}
+            href={model.href}
+            target="_blank"
+            rel="noreferrer"
+            className="group block h-full"
+          >
+            <Card className="h-full p-5 transition-colors hover:border-brand/50">
+              <h3 className="font-semibold text-brand">
+                SemiAnalysis {model.title}
+                <ExternalLinkIcon />
+              </h3>
+              <p className="mt-2 text-sm leading-6 text-muted-foreground">{t[model.descKey]}</p>
+            </Card>
+          </a>
+        ))}
+      </div>
+    </section>
+  );
+};
+
+const SpecTable = ({ entry, locale }: { entry: ChipPageEntry; locale: Locale }) => {
+  const t = STRINGS[locale];
+  const spec = getChipSpec(entry);
+  const hw = getChipHw(entry);
+  const rows: readonly [string, string][] = [
+    [t.specLabels.vendor, hw.vendor],
+    [t.specLabels.arch, hw.arch],
+    [t.specLabels.memory, `${spec.memory} ${spec.memoryType}`],
+    [t.specLabels.memoryBandwidth, spec.memoryBandwidth],
+    [t.specLabels.fp4, spec.fp4 ? spec.fp4.toLocaleString('en-US') : t.notSupported],
+    [t.specLabels.fp8, spec.fp8.toLocaleString('en-US')],
+    [t.specLabels.bf16, spec.bf16.toLocaleString('en-US')],
+    [t.specLabels.scaleUp, spec.scaleUpTech],
+    [t.specLabels.scaleUpBandwidth, spec.scaleUpBandwidth],
+    [t.specLabels.worldSize, String(spec.scaleUpWorldSize)],
+    [t.specLabels.scaleUpTopology, spec.scaleUpTopology],
+    [t.specLabels.scaleOut, spec.scaleOutTech ?? t.none],
+    [t.specLabels.nic, spec.nic ?? t.none],
+    [t.specLabels.tdp, `${hw.tdp.toLocaleString('en-US')} W`],
+    [t.specLabels.allInPower, `${hw.power} kW`],
+    [t.specLabels.costHyperscaler, `$${hw.costh.toFixed(2)}`],
+    [t.specLabels.costNeocloud, `$${hw.costn.toFixed(2)}`],
+    [t.specLabels.costRetail, `$${hw.costr.toFixed(2)}`],
+  ];
+  return (
+    <div className="mt-4 overflow-x-auto">
+      <table className="w-full text-sm">
+        <tbody>
+          {rows.map(([label, value]) => (
+            <tr key={label} className="border-b border-border/40">
+              <th scope="row" className="py-2 pr-4 text-left font-medium text-muted-foreground">
+                {label}
+              </th>
+              <td className="py-2 font-mono">{value}</td>
+            </tr>
+          ))}
+        </tbody>
+      </table>
+      <p className="mt-3 text-sm text-muted-foreground">
+        {t.pricingSource}{' '}
+        <a
+          href={TCO_SOURCE_URL}
+          target="_blank"
+          rel="noreferrer"
+          className="group underline hover:text-foreground"
+        >
+          {TCO_MODEL_TITLE}
+          <ExternalLinkIcon />
+        </a>
+      </p>
+    </div>
+  );
+};
+
+const RelatedLinks = ({ entry, locale }: { entry: ChipPageEntry; locale: Locale }) => {
+  const t = STRINGS[locale];
+  const posts = entry.relatedBlogSlugs.flatMap((slug) => {
+    const post = getPostBySlug(slug, locale === 'zh' ? 'zh' : undefined) ?? getPostBySlug(slug);
+    return post ? [{ slug, title: post.meta.title }] : [];
+  });
+  const terms = entry.relatedGlossarySlugs.flatMap((slug) => {
+    const term = locale === 'zh' ? getZhGlossaryEntry(slug) : getGlossaryEntry(slug);
+    return term ? [{ slug, term: term.term }] : [];
+  });
+  const chips = entry.relatedChipSlugs.flatMap((slug) => {
+    const chip = getAllChipPages().find((page) => page.slug === slug);
+    return chip ? [chip] : [];
+  });
+  return (
+    <div className="mt-10 grid gap-8 border-t border-border/50 pt-10 md:grid-cols-3">
+      <section aria-label={t.relatedArticles}>
+        <h2 className="text-sm font-semibold tracking-eyebrow text-muted-foreground uppercase">
+          {t.relatedArticles}
+        </h2>
+        <ul className="mt-3 space-y-2">
+          {posts.map((post) => (
+            <li key={post.slug}>
+              <Link
+                href={localePath(`/blog/${post.slug}`, locale)}
+                className="text-sm text-brand hover:underline"
+              >
+                {post.title}
+              </Link>
+            </li>
+          ))}
+        </ul>
+      </section>
+      <section aria-label={t.relatedTerms}>
+        <h2 className="text-sm font-semibold tracking-eyebrow text-muted-foreground uppercase">
+          {t.relatedTerms}
+        </h2>
+        <ul className="mt-3 space-y-2">
+          {terms.map((term) => (
+            <li key={term.slug}>
+              <Link
+                href={localePath(`/glossary/${term.slug}`, locale)}
+                className="text-sm text-brand hover:underline"
+              >
+                {term.term}
+              </Link>
+            </li>
+          ))}
+        </ul>
+      </section>
+      <section aria-label={t.relatedChips}>
+        <h2 className="text-sm font-semibold tracking-eyebrow text-muted-foreground uppercase">
+          {t.relatedChips}
+        </h2>
+        <ul className="mt-3 space-y-2">
+          {chips.map((chip) => (
+            <li key={chip.slug}>
+              <Link
+                href={localePath(`/chips/${chip.slug}`, locale)}
+                className="text-sm text-brand hover:underline"
+              >
+                {chip.title}
+              </Link>
+            </li>
+          ))}
+        </ul>
+      </section>
+    </div>
+  );
+};
+
+export const ChipDetailContent = ({ entry, locale }: { entry: ChipPageEntry; locale: Locale }) => {
+  const t = STRINGS[locale];
+  const translation = locale === 'zh' ? getZhChipTranslation(entry.slug) : undefined;
+  const overview = translation?.overview ?? entry.overview;
+  const benchmarkContext = translation?.benchmarkContext ?? entry.benchmarkContext;
+  const faq = chipFaq(entry, locale);
+  const url = `${SITE_URL}${localePath(`/chips/${entry.slug}`, locale)}`;
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@graph': [
+      breadcrumbJsonLd(locale, entry.title, `/chips/${entry.slug}`),
+      faqJsonLd(faq, url, locale),
+    ],
+  };
+  const hw = getChipHw(entry);
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8">
+        <article className="mx-auto max-w-5xl">
+          <Card className="overflow-hidden p-0">
+            <header className="relative border-b border-border/50 px-5 py-8 md:px-10 md:py-12">
+              <Link
+                href={localePath('/chips', locale)}
+                className="group inline-flex items-center gap-2 text-sm text-muted-foreground transition-colors hover:text-brand"
+              >
+                <span aria-hidden="true">←</span>
+                {t.backToIndex}
+              </Link>
+              <div className="mt-8 flex flex-wrap items-center gap-2">
+                <span className="rounded-full border border-brand/25 bg-brand/8 px-3 py-1 text-xs font-semibold tracking-eyebrow text-brand uppercase">
+                  {hw.vendor} · {hw.arch}
+                </span>
+              </div>
+              <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-6xl">
+                {entry.title}
+              </h1>
+            </header>
+            <div className="px-5 py-8 md:px-10 md:py-12">
+              <section aria-labelledby="chip-overview">
+                <h2 id="chip-overview" className="text-xl font-semibold tracking-tight">
+                  {t.overviewHeading}
+                </h2>
+                {overview.map((paragraph) => (
+                  <p key={paragraph.slice(0, 32)} className="mt-3 leading-7 text-muted-foreground">
+                    {paragraph}
+                  </p>
+                ))}
+              </section>
+              <section
+                aria-labelledby="chip-specs"
+                className="mt-10 border-t border-border/50 pt-10"
+              >
+                <h2 id="chip-specs" className="text-xl font-semibold tracking-tight">
+                  {t.specsHeading}
+                </h2>
+                <SpecTable entry={entry} locale={locale} />
+              </section>
+              <section
+                aria-labelledby="chip-benchmarks"
+                className="mt-10 border-t border-border/50 pt-10"
+              >
+                <h2 id="chip-benchmarks" className="text-xl font-semibold tracking-tight">
+                  {t.benchmarkHeading}
+                </h2>
+                <p className="mt-3 leading-7 text-muted-foreground">{benchmarkContext}</p>
+              </section>
+              <FaqSection faq={faq} locale={locale} />
+              <LiveResultsSection locale={locale} />
+              <ModelsSection locale={locale} />
+              <RelatedLinks entry={entry} locale={locale} />
+            </div>
+          </Card>
+        </article>
+      </div>
+    </main>
+  );
+};
+
+export const ChipVsContent = ({ page, locale }: { page: ChipVsPage; locale: Locale }) => {
+  const t = STRINGS[locale];
+  const highlights = buildChipVsHighlights(page);
+  const labels = locale === 'zh' ? CHIP_VS_HIGHLIGHT_LABELS_ZH : CHIP_VS_HIGHLIGHT_LABELS_EN;
+  const faq = vsFaq(page, locale);
+  const title = `${page.a.label} vs ${page.b.label}`;
+  const url = `${SITE_URL}${localePath(`/chips/${page.slug}`, locale)}`;
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@graph': [breadcrumbJsonLd(locale, title, `/chips/${page.slug}`), faqJsonLd(faq, url, locale)],
+  };
+  const localizeValue = locale === 'zh' ? localizeVsHighlightValueZh : (value: string) => value;
+
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8">
+        <article className="mx-auto max-w-5xl">
+          <Card className="overflow-hidden p-0">
+            <header className="relative border-b border-border/50 px-5 py-8 md:px-10 md:py-12">
+              <Link
+                href={localePath('/chips', locale)}
+                className="group inline-flex items-center gap-2 text-sm text-muted-foreground transition-colors hover:text-brand"
+              >
+                <span aria-hidden="true">←</span>
+                {t.backToIndex}
+              </Link>
+              <h1 className="mt-8 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-6xl">
+                {title}
+              </h1>
+              <p className="mt-4 max-w-3xl leading-7 text-muted-foreground">
+                {t.vsIntroPrefix}{' '}
+                <Link
+                  href={localePath(`/chips/${page.a.slug}`, locale)}
+                  className="text-brand hover:underline"
+                >
+                  {page.a.title}
+                </Link>{' '}
+                {t.vsIntroAnd}{' '}
+                <Link
+                  href={localePath(`/chips/${page.b.slug}`, locale)}
+                  className="text-brand hover:underline"
+                >
+                  {page.b.title}
+                </Link>{' '}
+                {t.vsIntroSuffix}
+              </p>
+            </header>
+            <div className="px-5 py-8 md:px-10 md:py-12">
+              <section aria-labelledby="vs-table">
+                <h2 id="vs-table" className="text-xl font-semibold tracking-tight">
+                  {t.vsSpecTableCaption}
+                </h2>
+                <div className="mt-4 overflow-x-auto">
+                  <table className="w-full text-sm">
+                    <thead>
+                      <tr className="border-b border-border/60 text-left">
+                        <th className="py-2 pr-4 font-medium text-muted-foreground"> </th>
+                        <th className="py-2 pr-4 font-semibold">{page.a.label}</th>
+                        <th className="py-2 pr-4 font-semibold">{page.b.label}</th>
+                        <th className="py-2 font-medium text-muted-foreground">
+                          {page.a.label} / {page.b.label}
+                        </th>
+                      </tr>
+                    </thead>
+                    <tbody>
+                      {highlights.map((row) => (
+                        <tr key={row.key} className="border-b border-border/40">
+                          <th
+                            scope="row"
+                            className="py-2 pr-4 text-left font-medium text-muted-foreground"
+                          >
+                            {labels[row.key]}
+                          </th>
+                          <td className="py-2 pr-4 font-mono">{localizeValue(row.aValue)}</td>
+                          <td className="py-2 pr-4 font-mono">{localizeValue(row.bValue)}</td>
+                          <td className="py-2 font-mono">{row.ratio ?? t.vsRatioNa}</td>
+                        </tr>
+                      ))}
+                    </tbody>
+                  </table>
+                </div>
+                <p className="mt-3 text-sm text-muted-foreground">{t.vsRatioNote}</p>
+              </section>
+              <FaqSection faq={faq} locale={locale} />
+              <LiveResultsSection locale={locale} />
+              <ModelsSection locale={locale} />
+            </div>
+          </Card>
+        </article>
+      </div>
+    </main>
+  );
+};
+
+export const ChipsIndexContent = ({ locale }: { locale: Locale }) => {
+  const t = STRINGS[locale];
+  const chips = getAllChipPages();
+  const vsPages = getAllChipVsPages();
+  const jsonLd = {
+    '@context': 'https://schema.org',
+    '@type': 'ItemList',
+    name: t.indexTitle,
+    itemListElement: chips.map((chip, index) => ({
+      '@type': 'ListItem',
+      position: index + 1,
+      name: chip.title,
+      url: `${SITE_URL}${localePath(`/chips/${chip.slug}`, locale)}`,
+    })),
+  };
+  return (
+    <main className="relative">
+      <JsonLd data={jsonLd} />
+      <div className="container mx-auto px-4 lg:px-8">
+        <div className="mx-auto max-w-5xl">
+          <header className="py-8 md:py-12">
+            <h1 className="text-4xl font-bold tracking-heading md:text-5xl">{t.indexTitle}</h1>
+            <p className="mt-4 max-w-3xl leading-7 text-muted-foreground">{t.indexIntro}</p>
+          </header>
+          <section aria-labelledby="chips-list">
+            <h2 id="chips-list" className="text-xl font-semibold tracking-tight">
+              {t.chipPages}
+            </h2>
+            <div className="mt-4 grid gap-4 sm:grid-cols-2 lg:grid-cols-3">
+              {chips.map((chip) => {
+                const translation = locale === 'zh' ? getZhChipTranslation(chip.slug) : undefined;
+                const summary = translation?.summary ?? chip.summary;
+                return (
+                  <Link key={chip.slug} href={localePath(`/chips/${chip.slug}`, locale)}>
+                    <Card className="h-full p-5 transition-colors hover:border-brand/50">
+                      <h3 className="font-semibold">{chip.title}</h3>
+                      <p className="mt-2 line-clamp-4 text-sm leading-6 text-muted-foreground">
+                        {summary}
+                      </p>
+                    </Card>
+                  </Link>
+                );
+              })}
+            </div>
+          </section>
+          <section aria-labelledby="vs-list" className="mt-12 pb-12">
+            <h2 id="vs-list" className="text-xl font-semibold tracking-tight">
+              {t.vsPages}
+            </h2>
+            <ul className="mt-4 grid gap-2 sm:grid-cols-2 lg:grid-cols-3">
+              {vsPages.map((page) => (
+                <li key={page.slug}>
+                  <Link
+                    href={localePath(`/chips/${page.slug}`, locale)}
+                    className="text-sm font-medium text-brand hover:underline"
+                  >
+                    {page.a.label} vs {page.b.label} →
+                  </Link>
+                </li>
+              ))}
+            </ul>
+          </section>
+        </div>
+      </div>
+    </main>
+  );
+};
diff --git a/packages/app/src/components/collectivex/CollectiveXDisplay.tsx b/packages/app/src/components/collectivex/CollectiveXDisplay.tsx
index 892cdc553..881e9d648 100644
--- a/packages/app/src/components/collectivex/CollectiveXDisplay.tsx
+++ b/packages/app/src/components/collectivex/CollectiveXDisplay.tsx
@@ -25,7 +25,6 @@ import { track } from '@/lib/analytics';
 import { useLocale } from '@/lib/use-locale';
 
 import { CollectiveXChart } from './CollectiveXChart';
-import { CollectiveXInventory } from './CollectiveXInventory';
 import { CollectiveXKvSection } from './CollectiveXKvSection';
 import { CollectiveXRunsTable } from './CollectiveXRunsTable';
 import {
@@ -34,6 +33,7 @@ import {
   collectiveXRunDasharray,
   collectiveXSeriesForRun,
   collectiveXTopologyLabel,
+  normalizeCollectiveXSku,
   seriesMatchesSelection,
   type CollectiveXSeriesSelection,
 } from './data';
@@ -56,6 +56,8 @@ interface SelectOption<T extends string | number> {
   label: string;
 }
 
+type CollectiveXSuiteFilter = 'all' | 'ep' | 'kv';
+
 const PERCENTILE_OPTIONS: SegmentedToggleOption<CollectiveXPercentile>[] = [
   { value: 'p50', label: 'p50' },
   { value: 'p90', label: 'p90' },
@@ -102,6 +104,10 @@ const STRINGS = {
     runsDescription:
       'Every stored run for the selected benchmark version. Check one or more runs to compare them in the explorer.',
     runsShown: 'Runs shown',
+    suiteControl: 'Suites',
+    suiteAria: 'Filter CollectiveX runs by suite',
+    allSuites: 'All',
+    noSuiteRuns: 'No runs contain the selected suite.',
     selectRuns: 'Select one or more runs from the table to show their data.',
     selectedRunsFailed: 'One or more selected runs failed to load.',
     runControl: 'Run',
@@ -165,7 +171,7 @@ const STRINGS = {
     yAxis: {
       latency: '延迟',
       'tokens-per-second': '所选延迟分位点的 token 速率',
-      'payload-rate': '所选延迟分位点的载荷带宽（每 Chip）',
+      'payload-rate': '所选延迟分位点的载荷带宽（每芯片）',
     },
     mode: { normal: '常规', 'low-latency': '低延迟' },
     fabricScope: { all: '全部', 'scale-up': '域内', 'scale-out': '跨域' },
@@ -176,7 +182,6 @@ const STRINGS = {
       'gate-weighted': '门控加权合并',
     },
     tabs: {
-      inventory: 'Matrix case inventory',
       case: 'Selected matrix case',
       evidence: '证据',
     },
@@ -206,6 +211,10 @@ const STRINGS = {
     runsDescription:
       'Every stored run for the selected benchmark version. Check one or more runs to compare them in the explorer.',
     runsShown: 'Runs shown',
+    suiteControl: '测试套件',
+    suiteAria: '按测试套件筛选 CollectiveX 运行',
+    allSuites: '全部',
+    noSuiteRuns: '没有包含所选测试套件的运行。',
     selectRuns: 'Select one or more runs from the table to show their data.',
     selectedRunsFailed: 'One or more selected runs failed to load.',
     runControl: 'Run',
@@ -239,7 +248,7 @@ const STRINGS = {
     resetFilter: '重置筛选',
     stableOrdering: '排名顺序稳定性已通过',
     samplingContract: (trials: number, iterations: number, samples: number, warmups: number) =>
-      `${trials}×${iterations} = 每个分项 ${samples} 个样本 · ${warmups} 次同步预热`,
+      `${trials}×${iterations} = 每个分项 ${samples} 个样本 · 同步 warmup ${warmups} 次`,
     selectedFactorsDiffer: '所选配置存在差异',
     differenceLabels: {
       model: '模型',
@@ -270,7 +279,7 @@ const STRINGS = {
     isolatedNote: '分项之和为派生值，不用于计算吞吐量。',
     payloadNote: '逻辑载荷速率按所选延迟分位点派生，不代表物理链路带宽。',
     payloadBandwidthNote:
-      '载荷带宽为完整逻辑载荷（含 FP8 缩放字节）÷ 延迟（每 Chip），是基于逻辑字节的派生速率，不代表物理链路带宽。工具提示中的 β/α 为延迟对字节在整个梯度上的最小二乘拟合（β = 每 Chip 带宽项，α = 固定开销）。',
+      '载荷带宽为完整逻辑载荷（含 FP8 缩放字节）÷ 延迟（每芯片），是基于逻辑字节的派生速率，不代表物理链路带宽。工具提示中的 β/α 为延迟对字节在整个梯度上的最小二乘拟合（β = 每芯片带宽项，α = 固定开销）。',
     provenance: '发布数据溯源',
     runLabel: 'Run',
     attemptLabel: 'Attempt',
@@ -302,7 +311,7 @@ const ADMIN_TOKEN_STORAGE_KEY = 'collectivex-admin-token';
 function ControlGroup({ label, children }: { label: string; children: React.ReactNode }) {
   return (
     <div className="min-w-0 space-y-1.5">
-      <Label className="text-[11px] uppercase tracking-wide text-muted-foreground">{label}</Label>
+      <Label className="text-2xs uppercase tracking-wide text-muted-foreground">{label}</Label>
       {children}
     </div>
   );
@@ -323,11 +332,21 @@ export default function CollectiveXDisplay() {
   const locale = useLocale();
   const t = STRINGS[locale];
   const [version, setVersion] = useState<CollectiveXVersion>(COLLECTIVEX_DEFAULT_VERSION);
+  const [suiteFilter, setSuiteFilter] = useState<CollectiveXSuiteFilter>('all');
   const [visibleRunIds, setVisibleRunIds] = useState<Set<string>>(new Set());
   const [bulkDeletingRunIds, setBulkDeletingRunIds] = useState<Set<string>>(new Set());
   const initializedVersionRef = useRef<CollectiveXVersion | null>(null);
   const runsQuery = useCollectiveXRuns(version);
   const runList = runsQuery.data?.runs ?? [];
+  const filteredRunList = useMemo(
+    () =>
+      runList.filter((run) => {
+        if (suiteFilter === 'all') return true;
+        const kvRequested = run.kv_cases?.requested ?? 0;
+        return suiteFilter === 'kv' ? kvRequested > 0 : run.requested_cases - kvRequested > 0;
+      }),
+    [runList, suiteFilter],
+  );
   const orderedVisibleRunIds = useMemo(() => {
     const liveRunIds = new Set(runList.map((run) => run.run_id));
     return [...visibleRunIds].filter((runId) => liveRunIds.has(runId));
@@ -366,9 +385,9 @@ export default function CollectiveXDisplay() {
   const [epSize, setEpSize] = useState(8);
   const [operation, setOperation] = useState<CollectiveXOperation>('roundtrip');
   const [phase, setPhase] = useState<CollectiveXPhase>('decode');
-  // Normal (throughput) kernels are the baseline; the availability effect
-  // below falls back when a slice only measured low-latency kernels.
-  const [modes, setModes] = useState<CollectiveXMode[]>(['normal']);
+  // Compare throughput-oriented and low-latency kernels by default. The
+  // availability effect below falls back when a slice has neither preference.
+  const [modes, setModes] = useState<CollectiveXMode[]>(['normal', 'low-latency']);
   // Prefer FP8 when the run measured it; the availability effect below falls
   // back to bf16 for runs (or EP/phase slices) without FP8 series.
   const [precision, setPrecision] = useState<CollectiveXPrecision>('fp8');
@@ -508,7 +527,10 @@ export default function CollectiveXDisplay() {
     [combinedSeries, seriesSelection],
   );
   const skuOptions = useMemo(
-    () => ['all', ...new Set(matchedSeries.map((item) => item.system.sku))],
+    () => [
+      'all',
+      ...new Set(matchedSeries.map((item) => normalizeCollectiveXSku(item.system.sku))),
+    ],
     [matchedSeries],
   );
   const backendOptions = useMemo(
@@ -516,7 +538,7 @@ export default function CollectiveXDisplay() {
       'all',
       ...new Set(
         matchedSeries
-          .filter((item) => sku === 'all' || item.system.sku === sku)
+          .filter((item) => sku === 'all' || normalizeCollectiveXSku(item.system.sku) === sku)
           .map((item) => item.backend),
       ),
     ],
@@ -530,7 +552,7 @@ export default function CollectiveXDisplay() {
     () =>
       matchedSeries.filter(
         (item) =>
-          (sku === 'all' || item.system.sku === sku) &&
+          (sku === 'all' || normalizeCollectiveXSku(item.system.sku) === sku) &&
           (backend === 'all' || item.backend === backend),
       ),
     [backend, matchedSeries, sku],
@@ -729,7 +751,7 @@ export default function CollectiveXDisplay() {
               <h1 className="text-xl font-semibold">CollectiveX</h1>
               <span
                 data-testid="collectivex-run-conclusion"
-                className={`rounded-md border px-2 py-0.5 text-[11px] font-medium ${
+                className={`rounded-md border px-2 py-0.5 text-2xs font-medium ${
                   singleDataset ? singleConclusionClass : CONCLUSION_FALLBACK_CLASS
                 }`}
               >
@@ -800,6 +822,22 @@ export default function CollectiveXDisplay() {
             <p className="mt-1 text-sm text-muted-foreground">{t.runsDescription}</p>
           </div>
           <div className="flex w-full flex-col gap-2 sm:flex-row sm:items-end md:w-auto">
+            <ControlGroup label={t.suiteControl}>
+              <SegmentedToggle
+                value={suiteFilter}
+                options={[
+                  { value: 'all', label: t.allSuites },
+                  { value: 'ep', label: 'EP' },
+                  { value: 'kv', label: 'KV' },
+                ]}
+                onValueChange={(suite) => {
+                  setSuiteFilter(suite);
+                  track('collectivex_suite_filter_changed', { suite });
+                }}
+                ariaLabel={t.suiteAria}
+                testId="collectivex-suite-filter"
+              />
+            </ControlGroup>
             <Button
               type="button"
               variant="destructive"
@@ -830,13 +868,14 @@ export default function CollectiveXDisplay() {
           </div>
         </div>
         <CollectiveXRunsTable
-          runs={runList}
+          runs={filteredRunList}
           selectedRunIndexById={selectedRunIndexById}
           visibleRunIds={visibleRunIds}
           loadingRunIds={loadingRunIds}
           deletingRunIds={deletingRunIds}
           onVisibleChange={handleVisibleRunChange}
           onDelete={(runId) => void handleDeleteRun(runId)}
+          emptyMessage={suiteFilter === 'all' ? undefined : t.noSuiteRuns}
         />
       </Card>
 
@@ -1084,10 +1123,6 @@ export default function CollectiveXDisplay() {
               />
             </div>
           </Card>
-          <CollectiveXInventory
-            key={`${version}-${datasets.map((dataset) => `${dataset.run.run_id}:${dataset.run.run_attempt}`).join(',')}`}
-            datasets={datasets}
-          />
         </>
       )}
     </section>
diff --git a/packages/app/src/components/collectivex/CollectiveXInventory.tsx b/packages/app/src/components/collectivex/CollectiveXInventory.tsx
deleted file mode 100644
index 11c25d671..000000000
--- a/packages/app/src/components/collectivex/CollectiveXInventory.tsx
+++ /dev/null
@@ -1,169 +0,0 @@
-'use client';
-
-import { useMemo } from 'react';
-
-import { Badge } from '@/components/ui/badge';
-import { Card } from '@/components/ui/card';
-import { type DataTableColumn, DataTable } from '@/components/ui/data-table';
-
-import { collectiveXTopologyLabel } from './data';
-import type { CollectiveXCoverage, CollectiveXDataset, CollectiveXTerminalStatus } from './types';
-
-type CollectiveXRunCoverage = CollectiveXCoverage & { run_id: string };
-
-const TERMINAL_ORDER: CollectiveXTerminalStatus[] = [
-  'measured',
-  'unsupported',
-  'failed',
-  'invalid',
-  'diagnostic',
-  'pending',
-];
-
-const STATUS_CLASS: Record<CollectiveXTerminalStatus, string> = {
-  measured: 'border-emerald-600/40 bg-emerald-500/10 text-emerald-700 dark:text-emerald-300',
-  unsupported: 'border-zinc-500/40 bg-zinc-500/10 text-zinc-700 dark:text-zinc-300',
-  failed: 'border-red-700/50 bg-red-700/10 text-red-800 dark:text-red-300',
-  invalid: 'border-red-600/40 bg-red-500/10 text-red-700 dark:text-red-300',
-  diagnostic: 'border-amber-600/40 bg-amber-500/10 text-amber-700 dark:text-amber-300',
-  pending: 'border-zinc-500/40 bg-zinc-500/5 text-muted-foreground',
-};
-
-function terminalCounts(item: CollectiveXCoverage): Record<CollectiveXTerminalStatus, number> {
-  const counts = Object.fromEntries(TERMINAL_ORDER.map((status) => [status, 0])) as Record<
-    CollectiveXTerminalStatus,
-    number
-  >;
-  for (const point of item.points) counts[point.terminal_status] += 1;
-  return counts;
-}
-
-function TerminalBadges({ item }: { item: CollectiveXCoverage }) {
-  const counts = terminalCounts(item);
-  const reasons = [...new Set(item.points.flatMap((point) => point.reason ?? []))];
-  return (
-    <div className="min-w-44">
-      <div className="flex flex-wrap gap-1">
-        {TERMINAL_ORDER.filter((status) => counts[status] > 0).map((status) => (
-          <Badge key={status} variant="outline" className={STATUS_CLASS[status]}>
-            {status} {counts[status]}
-          </Badge>
-        ))}
-      </div>
-      {reasons.length > 0 && (
-        <p className="mt-1 text-xs text-muted-foreground">{reasons.join(', ')}</p>
-      )}
-    </div>
-  );
-}
-
-export function CollectiveXInventory({ datasets }: { datasets: CollectiveXDataset[] }) {
-  const rows = useMemo<CollectiveXRunCoverage[]>(
-    () =>
-      datasets.flatMap((dataset) =>
-        dataset.coverage.map((item) => ({ ...item, run_id: dataset.run.run_id })),
-      ),
-    [datasets],
-  );
-  const columns = useMemo<DataTableColumn<CollectiveXRunCoverage>[]>(
-    () => [
-      {
-        header: 'Run',
-        cell: (row) => <span className="font-mono text-xs">#{row.run_id}</span>,
-        sortValue: (row) => Number(row.run_id),
-        className: 'whitespace-nowrap',
-      },
-      {
-        header: 'Case',
-        cell: (row) => (
-          <div className="min-w-56">
-            <p className="font-medium">{row.label}</p>
-            <p className="mt-0.5 font-mono text-[11px] text-muted-foreground">{row.case_id}</p>
-          </div>
-        ),
-        sortValue: (row) => `${row.label} ${row.case_id}`,
-      },
-      { header: 'SKU', cell: (row) => row.sku.toUpperCase(), sortValue: (row) => row.sku },
-      {
-        header: 'Backend',
-        cell: (row) => row.backend,
-        sortValue: (row) => row.backend,
-        className: 'whitespace-nowrap',
-      },
-      {
-        header: 'EP',
-        cell: (row) => `EP${row.topology.ep_size}`,
-        sortValue: (row) => row.topology.ep_size,
-      },
-      {
-        header: 'Phase',
-        cell: (row) => row.phase,
-        sortValue: (row) => row.phase,
-      },
-      {
-        header: 'Mode',
-        cell: (row) => row.mode,
-        sortValue: (row) => row.mode,
-      },
-      {
-        header: 'Precision',
-        cell: (row) => row.precision,
-        sortValue: (row) => row.precision,
-      },
-      {
-        header: 'Topology',
-        cell: (row) => collectiveXTopologyLabel(row.topology),
-        sortValue: (row) => collectiveXTopologyLabel(row.topology),
-        className: 'whitespace-nowrap',
-      },
-      {
-        header: 'Disposition',
-        cell: (row) => (
-          <div className="min-w-48">
-            <p>
-              {row.disposition} · {row.outcome}
-            </p>
-            {(row.detail || row.reason) && (
-              <p className="text-xs text-muted-foreground">{row.detail ?? row.reason}</p>
-            )}
-          </div>
-        ),
-        sortValue: (row) =>
-          `${row.disposition} ${row.outcome} ${row.reason ?? ''} ${row.detail ?? ''}`,
-      },
-      {
-        header: 'Point status',
-        cell: (row) => <TerminalBadges item={row} />,
-        sortValue: (row) =>
-          `${TERMINAL_ORDER.map((status) => `${status}:${terminalCounts(row)[status]}`).join(' ')} ${row.points.map((point) => point.reason ?? '').join(' ')}`,
-      },
-    ],
-    [],
-  );
-  const points = rows.flatMap((item) => item.points);
-  const measured = points.filter((point) => point.terminal_status === 'measured').length;
-  const unsupported = points.filter((point) => point.terminal_status === 'unsupported').length;
-  // Counted from this table's own EP coverage rows: the run-level totals also
-  // include kv-transfer cases, which live in their own card, not here.
-  const measuredCases = rows.filter((row) => row.outcome === 'success').length;
-  const unsupportedCases = rows.filter((row) => row.outcome === 'unsupported').length;
-  const terminalPoints = datasets.reduce((sum, dataset) => sum + dataset.run.terminal_points, 0);
-  const requestedPoints = datasets.reduce((sum, dataset) => sum + dataset.run.requested_points, 0);
-
-  return (
-    <Card data-testid="collectivex-inventory" className="min-w-0 w-full max-w-full overflow-hidden">
-      <h2 className="text-lg font-semibold">Matrix case inventory</h2>
-      <p className="mt-1 text-sm text-muted-foreground">
-        {datasets.length} runs · {rows.length} cases · {measuredCases} measured · {unsupportedCases}{' '}
-        unsupported · {terminalPoints}/{requestedPoints} terminal points · {measured} measured ·{' '}
-        {unsupported} unsupported
-      </p>
-      <DataTable
-        data={rows}
-        columns={columns}
-        testId="collectivex-inventory-table"
-        analyticsPrefix="collectivex_inventory"
-      />
-    </Card>
-  );
-}
diff --git a/packages/app/src/components/collectivex/CollectiveXKvFrontierChart.tsx b/packages/app/src/components/collectivex/CollectiveXKvFrontierChart.tsx
new file mode 100644
index 000000000..3857d6bdc
--- /dev/null
+++ b/packages/app/src/components/collectivex/CollectiveXKvFrontierChart.tsx
@@ -0,0 +1,229 @@
+'use client';
+
+import * as d3 from 'd3';
+import { useMemo } from 'react';
+
+import { D3Chart } from '@/lib/d3-chart/D3Chart';
+import { useLocale } from '@/lib/use-locale';
+
+import {
+  type CollectiveXKvFrontierPoint,
+  type CollectiveXKvFrontierSelection,
+  type CollectiveXKvRunCase,
+  collectiveXKvFrontierPoints,
+  collectiveXRunDasharray,
+} from './data';
+
+interface CollectiveXKvFrontierChartProps {
+  chartId: string;
+  cases: CollectiveXKvRunCase[];
+  colors: Record<string, string>;
+  selection: CollectiveXKvFrontierSelection;
+  caption?: React.ReactNode;
+  legendElement?: React.ReactNode;
+  testId?: string;
+}
+
+const STRINGS = {
+  en: {
+    noData: 'No measured KV rows match the selected page size and direction.',
+    instructions:
+      'Shift+Scroll to zoom · Drag to pan · Double-click to reset · Click a point to pin tooltip',
+    xAxis: (op: CollectiveXKvFrontierSelection['op']) =>
+      `Aggregate ${op} bandwidth at p50 (GB/s, log)`,
+    yAxis: 'Burst p95 latency per in-flight request (ms, log)',
+    dismiss: 'Click elsewhere to dismiss',
+    skuFrontier: 'SKU-wide frontier',
+    backendFrontier: 'backend frontier',
+    dominated: 'dominated',
+    pointContext: (row: CollectiveXKvFrontierPoint['row'], tier: string) =>
+      `${row.op} · page ${row.page_tokens} · batch ${row.batch} · ISL ${row.isl.toLocaleString('en-US')} · <strong>${tier}</strong>`,
+    pointMetrics: (point: CollectiveXKvFrontierPoint) =>
+      `Aggregate ${point.x.toFixed(point.x >= 100 ? 0 : 2)} GB/s · p95 ÷ in-flight ${point.y.toFixed(point.y >= 100 ? 0 : 1)} ms`,
+    latency: (point: CollectiveXKvFrontierPoint) =>
+      `Burst latency p50 / p95: ${point.row.latency_ms.p50.toFixed(1)} / ${point.row.latency_ms.p95.toFixed(1)} ms · ${point.row.descs.toLocaleString('en-US')} descriptors/request`,
+    verify: (passed: boolean) => `verify: ${passed ? 'passed' : 'FAILED'}`,
+  },
+  zh: {
+    noData: '没有与所选页大小和传输方向匹配的 KV 实测数据。',
+    instructions: 'Shift+滚轮缩放 · 拖动平移 · 双击重置 · 点击数据点固定提示框',
+    xAxis: (op: CollectiveXKvFrontierSelection['op']) => `p50 聚合 ${op} 带宽（GB/s，对数）`,
+    yAxis: '每个在途请求的突发 p95 延迟（ms，对数）',
+    dismiss: '点击其他位置关闭',
+    skuFrontier: 'SKU 级帕累托前沿',
+    backendFrontier: '后端帕累托前沿',
+    dominated: '被支配',
+    pointContext: (row: CollectiveXKvFrontierPoint['row'], tier: string) =>
+      `${row.op} · 页大小 ${row.page_tokens} · batch ${row.batch} · ISL ${row.isl.toLocaleString('en-US')} · <strong>${tier}</strong>`,
+    pointMetrics: (point: CollectiveXKvFrontierPoint) =>
+      `聚合带宽 ${point.x.toFixed(point.x >= 100 ? 0 : 2)} GB/s · p95 ÷ 在途请求数 ${point.y.toFixed(point.y >= 100 ? 0 : 1)} ms`,
+    latency: (point: CollectiveXKvFrontierPoint) =>
+      `突发延迟 p50 / p95：${point.row.latency_ms.p50.toFixed(1)} / ${point.row.latency_ms.p95.toFixed(1)} ms · ${point.row.descs.toLocaleString('en-US')} 个描述符/请求`,
+    verify: (passed: boolean) => `校验：${passed ? '通过' : '失败'}`,
+  },
+} as const;
+
+function paddedDomain(values: number[]): [number, number] {
+  if (values.length === 0) return [1, 10];
+  const min = d3.min(values) ?? 1;
+  const max = d3.max(values) ?? 1;
+  return min === max ? [min / 2, max * 2] : [min / 1.15, max * 1.15];
+}
+
+function formatCompact(value: number): string {
+  if (value >= 1e3) return `${(value / 1e3).toFixed(value < 1e4 ? 1 : 0)}k`;
+  if (value >= 10) return value.toFixed(0);
+  if (value >= 1) return value.toFixed(1);
+  return value.toFixed(2);
+}
+
+function escapeHtml(value: string): string {
+  return value
+    .replaceAll('&', '&amp;')
+    .replaceAll('<', '&lt;')
+    .replaceAll('>', '&gt;')
+    .replaceAll('"', '&quot;')
+    .replaceAll("'", '&#039;');
+}
+
+export function CollectiveXKvFrontierChart({
+  chartId,
+  cases,
+  colors,
+  selection,
+  caption,
+  legendElement,
+  testId,
+}: CollectiveXKvFrontierChartProps) {
+  const locale = useLocale();
+  const strings = STRINGS[locale === 'zh' ? 'zh' : 'en'];
+  const points = useMemo(() => collectiveXKvFrontierPoints(cases, selection), [cases, selection]);
+  const runIndexBySeries = useMemo(
+    () => new Map(cases.map((kase) => [`${kase.run_id}:${kase.case_id}`, kase.run_index])),
+    [cases],
+  );
+  // Roofline per series in the /inference style, drawn through the full batch
+  // ladder: because raising the batch improves both axes until the backend
+  // saturates, the strict Pareto set is usually a single point, so the ladder
+  // itself is the achievable curve and its endpoint the frontier.
+  const lines = useMemo(() => {
+    const bySeries: Record<string, CollectiveXKvFrontierPoint[]> = {};
+    for (const point of points) {
+      (bySeries[point.seriesId] ??= []).push(point);
+    }
+    const result: Record<string, { x: number; y: number }[]> = {};
+    for (const [seriesId, seriesPoints] of Object.entries(bySeries)) {
+      result[seriesId] = seriesPoints
+        .toSorted((a, b) => a.row.batch - b.row.batch)
+        .map((point) => ({ x: point.x, y: point.y }));
+    }
+    return result;
+  }, [points]);
+  const colorBySeries = useMemo(
+    () => new Map(points.map((point) => [point.seriesId, point.colorKey])),
+    [points],
+  );
+
+  const xDomain = useMemo(() => paddedDomain(points.map((point) => point.x)), [points]);
+  const yDomain = useMemo(() => paddedDomain(points.map((point) => point.y)), [points]);
+
+  const noDataOverlay =
+    points.length === 0 ? (
+      <div className="absolute inset-0 flex items-center justify-center pointer-events-none">
+        <p className="text-sm text-muted-foreground">{strings.noData}</p>
+      </div>
+    ) : undefined;
+
+  return (
+    <D3Chart<CollectiveXKvFrontierPoint>
+      chartId={chartId}
+      data={points}
+      height={420}
+      margin={{ top: 24, right: 20, bottom: 62, left: 78 }}
+      watermark="logo"
+      testId={testId}
+      grabCursor
+      instructions={strings.instructions}
+      xScale={{ type: 'log', domain: xDomain, nice: false }}
+      yScale={{ type: 'log', domain: yDomain, nice: false }}
+      xAxis={{
+        label: strings.xAxis(selection.op),
+        tickCount: 6,
+        tickFormat: (value) => formatCompact(Number(value)),
+      }}
+      yAxis={{
+        label: strings.yAxis,
+        tickCount: 5,
+        tickFormat: (value) => formatCompact(Number(value)),
+      }}
+      layers={[
+        {
+          type: 'line',
+          key: 'collectivex-kv-frontier-lines',
+          lines,
+          config: {
+            getColor: (key) => colors[colorBySeries.get(key) ?? ''] ?? '#888',
+            getStrokeDasharray: (key) => collectiveXRunDasharray(runIndexBySeries.get(key) ?? 0),
+            strokeWidth: 2.5,
+            curve: d3.curveMonotoneX,
+          },
+        },
+        {
+          type: 'point',
+          key: 'collectivex-kv-frontier-points',
+          data: points,
+          config: {
+            getCx: () => 0,
+            getCy: () => 0,
+            getX: (point) => point.x,
+            getY: (point) => point.y,
+            getColor: (point) => colors[point.colorKey] ?? '#888',
+            getRadius: () => 3.5,
+            keyFn: (point) => `${point.seriesId}-${point.row.batch}`,
+            maxPoints: Infinity,
+          },
+        },
+      ]}
+      zoom={{
+        enabled: true,
+        axes: 'both',
+        scaleExtent: [1, 20],
+        resetEventName: `collectivex_zoom_reset_${chartId}`,
+      }}
+      tooltip={{
+        rulerType: 'crosshair',
+        attachToLayer: 1,
+        content: (point, isPinned) => {
+          const color = colors[point.colorKey] ?? '#888';
+          const { row } = point;
+          const tier = point.onSkuFrontier
+            ? strings.skuFrontier
+            : point.onSeriesFrontier
+              ? strings.backendFrontier
+              : strings.dominated;
+          return `<div class="rounded-md border bg-background/95 px-3 py-2 text-xs shadow-md backdrop-blur-sm" style="min-width: 230px; max-width: 380px; user-select: ${isPinned ? 'text' : 'none'}">
+            ${isPinned ? `<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">${strings.dismiss}</div>` : ''}
+            <div class="font-semibold mb-1" style="color: ${color}">${escapeHtml(point.seriesLabel)}</div>
+            <div>${strings.pointContext(row, tier)}</div>
+            <div>${strings.pointMetrics(point)}</div>
+            <div class="text-muted-foreground">${strings.latency(point)}</div>
+            <div class="text-muted-foreground">${strings.verify(row.verify_passed)}</div>
+          </div>`;
+        },
+        getRulerX: (point, scale) =>
+          (scale as d3.ScaleLinear<number, number> | d3.ScaleLogarithmic<number, number>)(point.x),
+        getRulerY: (point, scale) => scale(point.y),
+        onHoverStart: (selectionEl) => {
+          selectionEl.attr('r', 6);
+        },
+        onHoverEnd: (selectionEl) => {
+          selectionEl.attr('r', 3.5);
+        },
+      }}
+      transitionDuration={200}
+      legendElement={legendElement}
+      noDataOverlay={noDataOverlay}
+      caption={caption}
+    />
+  );
+}
diff --git a/packages/app/src/components/collectivex/CollectiveXKvSection.tsx b/packages/app/src/components/collectivex/CollectiveXKvSection.tsx
index 3c5175e68..9da42fe7f 100644
--- a/packages/app/src/components/collectivex/CollectiveXKvSection.tsx
+++ b/packages/app/src/components/collectivex/CollectiveXKvSection.tsx
@@ -13,6 +13,7 @@ import { track } from '@/lib/analytics';
 import { useLocale } from '@/lib/use-locale';
 
 import { CollectiveXKvChart } from './CollectiveXKvChart';
+import { CollectiveXKvFrontierChart } from './CollectiveXKvFrontierChart';
 import {
   type CollectiveXKvChartSelection,
   type CollectiveXKvRunCase,
@@ -20,6 +21,7 @@ import {
   collectiveXKvColorKey,
   collectiveXKvLegendLabel,
   collectiveXRunDasharray,
+  collectiveXSkuLabel,
 } from './data';
 import type { CollectiveXDataset, CollectiveXOutcome } from './types';
 
@@ -33,10 +35,19 @@ const STRINGS = {
       'b1/bmax are requests posted per burst.',
     batchCaption: 'at the largest measured ISL',
     islCaption: 'at batch 1',
+    frontierCaption:
+      'each line walks the batch ladder at the largest measured ISL; down-right is better. ' +
+      'A backend that serializes requests collapses to a single point; hover a point for its ' +
+      'Pareto status.',
+    frontierOption: 'Frontier',
     yControl: 'Metric',
     xControl: 'X axis',
     pageControl: 'Page size',
     opControl: 'Direction',
+    metricAriaLabel: 'CollectiveX KV metric',
+    xAriaLabel: 'CollectiveX KV X axis',
+    pageAriaLabel: 'CollectiveX KV page size',
+    opAriaLabel: 'CollectiveX KV direction',
   },
   zh: {
     heading: 'KV 缓存传输',
@@ -46,10 +57,18 @@ const STRINGS = {
       'GB/s 为最大 ISL 处按突发聚合的 pull 带宽；b1/bmax 表示每次突发提交的请求数。',
     batchCaption: '取最大实测 ISL',
     islCaption: '取批大小 1',
+    frontierCaption:
+      '每条线连接最大实测 ISL 下的批大小阶梯，越靠右下越优。' +
+      '串行处理请求的后端会收缩为一个点；悬停可查看各点的帕累托状态。',
+    frontierOption: '帕累托前沿',
     yControl: '指标',
     xControl: 'X 轴',
     pageControl: '页大小',
     opControl: '方向',
+    metricAriaLabel: 'CollectiveX KV 指标',
+    xAriaLabel: 'CollectiveX KV X 轴',
+    pageAriaLabel: 'CollectiveX KV 页大小',
+    opAriaLabel: 'CollectiveX KV 传输方向',
   },
 } as const;
 
@@ -87,7 +106,7 @@ export function CollectiveXKvSection({
   const locale = useLocale();
   const strings = STRINGS[locale === 'zh' ? 'zh' : 'en'];
   const [yAxis, setYAxis] = useState<CollectiveXKvChartSelection['y']>('bandwidth');
-  const [xAxis, setXAxis] = useState<CollectiveXKvChartSelection['x']>('batch');
+  const [xAxis, setXAxis] = useState<CollectiveXKvChartSelection['x'] | 'frontier'>('batch');
   const [pageTokens, setPageTokens] = useState<'64' | '16'>('64');
   const [op, setOp] = useState<CollectiveXKvChartSelection['op']>('pull');
   // Legend toggles are keyed to the current series set: when checked runs
@@ -97,7 +116,9 @@ export function CollectiveXKvSection({
     ids: Set<string>;
     signature: string;
   } | null>(null);
-  const [legendExpanded, setLegendExpanded] = useState(false);
+  // Default open: with the closable sidebar panel, `false` now means fully
+  // hidden (reopen button only), so the legend must start expanded.
+  const [legendExpanded, setLegendExpanded] = useState(true);
 
   const rows = useMemo<CollectiveXKvRunCase[]>(
     () =>
@@ -177,7 +198,7 @@ export function CollectiveXKvSection({
         sortValue: (row) => Number(row.run_id),
         className: 'whitespace-nowrap',
       },
-      { header: 'SKU', cell: (row) => row.sku.toUpperCase(), sortValue: (row) => row.sku },
+      { header: 'SKU', cell: (row) => collectiveXSkuLabel(row.sku), sortValue: (row) => row.sku },
       {
         header: 'Backend',
         cell: (row) => row.backend,
@@ -245,7 +266,7 @@ export function CollectiveXKvSection({
   if (rows.length === 0) return null;
   const measured = rows.filter((row) => row.outcome === 'success').length;
   const selection: CollectiveXKvChartSelection = {
-    x: xAxis,
+    x: xAxis === 'frontier' ? 'batch' : xAxis,
     y: yAxis,
     op,
     pageTokens: Number(pageTokens),
@@ -259,22 +280,24 @@ export function CollectiveXKvSection({
       {measuredCases.length > 0 && (
         <>
           <div className="mt-3 flex flex-wrap items-end gap-x-5 gap-y-3">
-            <div className="grid gap-1.5">
-              <Label className="text-xs text-muted-foreground">{strings.yControl}</Label>
-              <SegmentedToggle
-                value={yAxis}
-                onValueChange={(value) => {
-                  setYAxis(value);
-                  track('collectivex_kv_metric_changed', { metric: value });
-                }}
-                ariaLabel="CollectiveX kv metric"
-                testId="collectivex-kv-metric-toggle"
-                options={[
-                  { value: 'bandwidth', label: 'GB/s' },
-                  { value: 'latency', label: 'ms' },
-                ]}
-              />
-            </div>
+            {xAxis !== 'frontier' && (
+              <div className="grid gap-1.5">
+                <Label className="text-xs text-muted-foreground">{strings.yControl}</Label>
+                <SegmentedToggle
+                  value={yAxis}
+                  onValueChange={(value) => {
+                    setYAxis(value);
+                    track('collectivex_kv_metric_changed', { metric: value });
+                  }}
+                  ariaLabel={strings.metricAriaLabel}
+                  testId="collectivex-kv-metric-toggle"
+                  options={[
+                    { value: 'bandwidth', label: 'GB/s' },
+                    { value: 'latency', label: 'ms' },
+                  ]}
+                />
+              </div>
+            )}
             <div className="grid gap-1.5">
               <Label className="text-xs text-muted-foreground">{strings.xControl}</Label>
               <SegmentedToggle
@@ -283,11 +306,12 @@ export function CollectiveXKvSection({
                   setXAxis(value);
                   track('collectivex_kv_xaxis_changed', { axis: value });
                 }}
-                ariaLabel="CollectiveX kv x axis"
+                ariaLabel={strings.xAriaLabel}
                 testId="collectivex-kv-xaxis-toggle"
                 options={[
                   { value: 'batch', label: 'Batch' },
                   { value: 'isl', label: 'ISL' },
+                  { value: 'frontier', label: strings.frontierOption },
                 ]}
               />
             </div>
@@ -296,7 +320,7 @@ export function CollectiveXKvSection({
               <SegmentedToggle
                 value={pageTokens}
                 onValueChange={setPageTokens}
-                ariaLabel="CollectiveX kv page size"
+                ariaLabel={strings.pageAriaLabel}
                 testId="collectivex-kv-page-toggle"
                 options={[
                   { value: '64', label: '64' },
@@ -309,7 +333,7 @@ export function CollectiveXKvSection({
               <SegmentedToggle
                 value={op}
                 onValueChange={setOp}
-                ariaLabel="CollectiveX kv direction"
+                ariaLabel={strings.opAriaLabel}
                 testId="collectivex-kv-op-toggle"
                 options={[
                   { value: 'pull', label: 'pull' },
@@ -319,28 +343,52 @@ export function CollectiveXKvSection({
             </div>
           </div>
           <div className="relative mt-3">
-            <CollectiveXKvChart
-              chartId="collectivex-kv"
-              testId="collectivex-kv-chart"
-              cases={activeCases}
-              colors={colors}
-              selection={selection}
-              caption={
-                <p className="text-sm text-muted-foreground">
-                  {op} · page {pageTokens} ·{' '}
-                  {xAxis === 'batch' ? strings.batchCaption : strings.islCaption}
-                </p>
-              }
-              legendElement={
-                <ChartLegend
-                  variant="sidebar"
-                  legendItems={legendItems}
-                  disableActiveSort
-                  isLegendExpanded={legendExpanded}
-                  onExpandedChange={setLegendExpanded}
-                />
-              }
-            />
+            {xAxis === 'frontier' ? (
+              <CollectiveXKvFrontierChart
+                chartId="collectivex-kv-frontier"
+                testId="collectivex-kv-frontier-chart"
+                cases={activeCases}
+                colors={colors}
+                selection={{ op, pageTokens: Number(pageTokens) }}
+                caption={
+                  <p className="text-sm text-muted-foreground">
+                    {op} · page {pageTokens} · {strings.frontierCaption}
+                  </p>
+                }
+                legendElement={
+                  <ChartLegend
+                    variant="sidebar"
+                    legendItems={legendItems}
+                    disableActiveSort
+                    isLegendExpanded={legendExpanded}
+                    onExpandedChange={setLegendExpanded}
+                  />
+                }
+              />
+            ) : (
+              <CollectiveXKvChart
+                chartId="collectivex-kv"
+                testId="collectivex-kv-chart"
+                cases={activeCases}
+                colors={colors}
+                selection={selection}
+                caption={
+                  <p className="text-sm text-muted-foreground">
+                    {op} · page {pageTokens} ·{' '}
+                    {xAxis === 'batch' ? strings.batchCaption : strings.islCaption}
+                  </p>
+                }
+                legendElement={
+                  <ChartLegend
+                    variant="sidebar"
+                    legendItems={legendItems}
+                    disableActiveSort
+                    isLegendExpanded={legendExpanded}
+                    onExpandedChange={setLegendExpanded}
+                  />
+                }
+              />
+            )}
           </div>
         </>
       )}
diff --git a/packages/app/src/components/collectivex/CollectiveXRunsTable.tsx b/packages/app/src/components/collectivex/CollectiveXRunsTable.tsx
index 54271549e..e4009f8ac 100644
--- a/packages/app/src/components/collectivex/CollectiveXRunsTable.tsx
+++ b/packages/app/src/components/collectivex/CollectiveXRunsTable.tsx
@@ -7,7 +7,7 @@ import { track } from '@/lib/analytics';
 import { useLocale } from '@/lib/use-locale';
 import { cn } from '@/lib/utils';
 
-import { collectiveXRunDasharray } from './data';
+import { collectiveXRunDasharray, collectiveXSkuLabel } from './data';
 import type { CollectiveXRunSummary } from './types';
 
 interface CollectiveXRunsTableProps {
@@ -18,6 +18,7 @@ interface CollectiveXRunsTableProps {
   deletingRunIds: ReadonlySet<string>;
   onVisibleChange: (runId: string, visible: boolean) => void;
   onDelete: (runId: string) => void;
+  emptyMessage?: string;
 }
 
 const STRINGS = {
@@ -43,7 +44,7 @@ const STRINGS = {
     shown: 'Shown',
     run: 'Run',
     result: 'Result',
-    suites: 'Suites',
+    suites: '测试套件',
     cases: 'Measured cases',
     points: 'Terminal points',
     skus: 'SKUs',
@@ -86,12 +87,15 @@ export function CollectiveXRunsTable({
   deletingRunIds,
   onVisibleChange,
   onDelete,
+  emptyMessage,
 }: CollectiveXRunsTableProps) {
   const locale = useLocale();
   const t = STRINGS[locale];
 
   if (runs.length === 0) {
-    return <p className="py-6 text-center text-sm text-muted-foreground">{t.empty}</p>;
+    return (
+      <p className="py-6 text-center text-sm text-muted-foreground">{emptyMessage ?? t.empty}</p>
+    );
   }
 
   return (
@@ -197,7 +201,7 @@ export function CollectiveXRunsTable({
                 <td className="px-3 py-1.5">
                   <span
                     className={cn(
-                      'inline-flex rounded-md border px-2 py-0.5 text-[11px] font-medium',
+                      'inline-flex rounded-md border px-2 py-0.5 text-2xs font-medium',
                       CONCLUSION_CLASSES[conclusion] ?? CONCLUSION_FALLBACK_CLASS,
                     )}
                   >
@@ -211,7 +215,7 @@ export function CollectiveXRunsTable({
                         title={`EP: ${epMeasured}/${epRequested} measured`}
                         data-testid={`collectivex-run-suite-ep-${run.run_id}`}
                         className={cn(
-                          'inline-flex rounded-md border px-2 py-0.5 text-[11px] font-medium',
+                          'inline-flex rounded-md border px-2 py-0.5 text-2xs font-medium',
                           SUITE_BADGE_CLASSES.ep,
                         )}
                       >
@@ -223,7 +227,7 @@ export function CollectiveXRunsTable({
                         title={`KV transfer: ${kvMeasured}/${kvRequested} measured`}
                         data-testid={`collectivex-run-suite-kv-${run.run_id}`}
                         className={cn(
-                          'inline-flex rounded-md border px-2 py-0.5 text-[11px] font-medium',
+                          'inline-flex rounded-md border px-2 py-0.5 text-2xs font-medium',
                           SUITE_BADGE_CLASSES.kv,
                         )}
                       >
@@ -240,7 +244,7 @@ export function CollectiveXRunsTable({
                 </td>
                 <td className="px-3 py-1.5">
                   {run.covered_skus.length > 0
-                    ? run.covered_skus.map((sku) => sku.toUpperCase()).join(', ')
+                    ? [...new Set(run.covered_skus.map(collectiveXSkuLabel))].join(', ')
                     : '—'}
                 </td>
                 <td className="whitespace-nowrap px-3 py-1.5 text-xs text-muted-foreground">
diff --git a/packages/app/src/components/collectivex/CollectiveXTables.tsx b/packages/app/src/components/collectivex/CollectiveXTables.tsx
deleted file mode 100644
index 14c5843ea..000000000
--- a/packages/app/src/components/collectivex/CollectiveXTables.tsx
+++ /dev/null
@@ -1,267 +0,0 @@
-'use client';
-
-import { useMemo } from 'react';
-
-import { Badge } from '@/components/ui/badge';
-import { Card } from '@/components/ui/card';
-import { type DataTableColumn, DataTable } from '@/components/ui/data-table';
-import { useLocale } from '@/lib/use-locale';
-
-import { collectiveXTopologyLabel } from './data';
-
-import type { CollectiveXCoverage, CollectiveXOutcome } from './types';
-
-const OUTCOME_CLASSES = {
-  success: 'border-emerald-600/40 bg-emerald-500/15 text-emerald-700 dark:text-emerald-300',
-  unsupported: 'border-zinc-500/40 bg-zinc-500/15 text-zinc-700 dark:text-zinc-300',
-  failed: 'border-red-700/50 bg-red-700/15 text-red-800 dark:text-red-300',
-  invalid: 'border-red-600/40 bg-red-500/15 text-red-700 dark:text-red-300',
-  diagnostic: 'border-amber-600/40 bg-amber-500/15 text-amber-700 dark:text-amber-300',
-  pending: 'border-zinc-500/40 bg-zinc-500/5 text-muted-foreground',
-} satisfies Record<CollectiveXOutcome, string>;
-
-const STRINGS = {
-  en: {
-    outcome: {
-      success: 'success',
-      unsupported: 'unsupported',
-      failed: 'failed',
-      invalid: 'invalid',
-      diagnostic: 'diagnostic',
-      pending: 'pending',
-    },
-    phase: { decode: 'decode', prefill: 'prefill' },
-    mode: { normal: 'Normal', 'low-latency': 'Low latency' },
-    scope: { 'scale-up': 'Scale-up', 'scale-out': 'Scale-out' },
-    disposition: { runnable: 'runnable', unsupported: 'unsupported' },
-    case: 'Case',
-    sku: 'SKU',
-    backend: 'Backend',
-    phaseHeader: 'Phase',
-    precisionHeader: 'Precision',
-    modeHeader: 'Mode',
-    epHeader: 'EP',
-    scopeHeader: 'Fabric scope',
-    topologyHeader: 'Topology',
-    dispositionHeader: 'Disposition',
-    outcomeHeader: 'Outcome',
-    attempts: 'Attempts',
-    selected: 'Selected',
-    caseId: 'Case ID',
-    attemptId: 'Attempt ID',
-    failureMode: 'Failure mode',
-    reason: 'Reason',
-    terminalCoverage: 'Terminal coverage',
-    allocation: 'Allocation',
-    run: 'Run',
-    attempt: 'Try',
-    role: 'Role',
-    terminalRole: 'terminal selection',
-    allocationRole: 'allocation selection',
-    retainedRole: 'retained',
-    evidence: 'Evidence',
-    retainedAttempts: 'Retained attempts',
-  },
-  zh: {
-    outcome: {
-      success: '成功',
-      unsupported: '不支持',
-      failed: '失败',
-      invalid: '无效',
-      diagnostic: '诊断',
-      pending: '待运行',
-    },
-    phase: { decode: '解码', prefill: '预填充' },
-    mode: { normal: '常规', 'low-latency': '低延迟' },
-    scope: { 'scale-up': '域内（scale-up）', 'scale-out': '跨域（scale-out）' },
-    disposition: { runnable: '可运行', unsupported: '不支持' },
-    case: '用例',
-    sku: 'SKU',
-    backend: '后端',
-    phaseHeader: '阶段',
-    precisionHeader: '精度',
-    modeHeader: '模式',
-    epHeader: 'EP',
-    scopeHeader: '互联范围',
-    topologyHeader: '拓扑',
-    dispositionHeader: '计划状态',
-    outcomeHeader: '结果',
-    attempts: '尝试次数',
-    selected: '已选尝试',
-    caseId: '用例 ID',
-    attemptId: '尝试 ID',
-    failureMode: '失败类型',
-    reason: '原因',
-    terminalCoverage: '终结状态覆盖',
-    allocation: '独立分配',
-    run: '运行',
-    attempt: '尝试序号',
-    role: '用途',
-    terminalRole: '终结状态选择',
-    allocationRole: '独立分配选择',
-    retainedRole: '保留',
-    evidence: '证据数',
-    retainedAttempts: '保留的全部尝试',
-  },
-} as const;
-
-const REASON_LABELS = {
-  zh: {
-    'artifact-validation-failed': '产物校验失败',
-    'backend-platform-unsupported': '后端不支持该平台',
-    'backend-token-capacity': '后端 token 容量不足',
-    'launcher-setup-failed': '启动器初始化失败',
-    'repository-staging-failed': '代码仓库暂存失败',
-    'container-registry-verification-failed': '容器镜像仓库校验失败',
-    'scheduler-allocation-failed': '调度资源分配失败',
-    'container-image-preparation-failed': '容器镜像准备失败',
-    'container-image-identity-failed': '容器镜像身份校验失败',
-    'container-runtime-launch-failed': '容器运行时启动失败',
-    'backend-setup-failed': '后端初始化失败',
-    'artifact-collection-failed': '产物收集失败',
-    'runtime-identity-mismatch': '运行时身份不匹配',
-    'execution-timeout': '执行超时',
-    'execution-deadlock': '执行死锁',
-    'distributed-command-failed': '分布式命令执行失败',
-    'post-emit-distributed-command-failed': '结果写出后的分布式命令失败',
-    'unsupported-capability': '能力不支持',
-    'execution-failed': '执行失败',
-    'validation-failed': '校验失败',
-    'diagnostic-evidence': '诊断证据',
-    capability: '能力限制',
-    setup: '初始化',
-    'repository-stage': '代码仓库暂存',
-    'registry-verification': '镜像仓库校验',
-    'scheduler-allocation': '调度资源分配',
-    'container-import': '容器镜像导入',
-    'container-hash': '容器镜像哈希校验',
-    'container-launch': '容器启动',
-    'backend-setup': '后端初始化',
-    'artifact-collection': '产物收集',
-    'runtime-identity': '运行时身份',
-    timeout: '超时',
-    deadlock: '死锁',
-    execution: '执行',
-    'insufficient-allocations': '独立分配不足',
-    'incomplete-repeat-coverage': '重复运行覆盖不完整',
-    'correctness-failed': '正确性校验失败',
-    'missing-measured-roundtrip-p99': '缺少实测往返 p99',
-    'unstable-ordering': '排名顺序不稳定',
-    'incomplete-provenance': '来源与运行溯源不完整',
-    'noncanonical-workload': '工作负载不符合规范',
-    'unresolved-anomaly': '异常尚未解释',
-    'semantic-correctness-failed': '语义正确性校验失败',
-    'measurement-nonconformant': '测量协议不符合要求',
-    'expert-oracle-incomplete': '专家路由正确性校验不完整',
-    'incomplete-aligned-repeats': '对齐的重复运行不完整',
-    'missing-uniform-baseline': '缺少 uniform 基线',
-    'incomplete-routing-anchors': '路由基准锚点不完整',
-    'implementation-config-mismatch': '实现配置不一致',
-    'unmatched-token-coverage': 'token 点位覆盖不一致',
-    'awaiting-v1-runs': '等待 CollectiveX v1 运行结果',
-  },
-} as const;
-
-export function collectiveXReasonLabel(value: string, locale: 'en' | 'zh'): string {
-  if (locale === 'en') return value;
-  return REASON_LABELS.zh[value as keyof typeof REASON_LABELS.zh] ?? value;
-}
-
-function OutcomeBadge({ outcome }: { outcome: CollectiveXOutcome }) {
-  const t = STRINGS[useLocale()];
-  return (
-    <Badge variant="outline" className={OUTCOME_CLASSES[outcome]}>
-      {t.outcome[outcome]}
-    </Badge>
-  );
-}
-
-function shortId(value: string | null): string {
-  if (value === null) return '-';
-  const suffix = value.lastIndexOf('-');
-  return suffix === -1 ? value : value.slice(suffix + 1).slice(-8);
-}
-
-export function CollectiveXCoverageTable({ coverage }: { coverage: CollectiveXCoverage[] }) {
-  const locale = useLocale();
-  const t = STRINGS[locale];
-  const columns = useMemo<DataTableColumn<CollectiveXCoverage>[]>(
-    () => [
-      {
-        header: t.case,
-        cell: (row) => (
-          <div title={`${t.caseId}: ${row.case_id}`}>
-            <p className="font-medium whitespace-nowrap">{row.label}</p>
-            <p className="font-mono text-[11px] text-muted-foreground">{shortId(row.case_id)}</p>
-          </div>
-        ),
-        sortValue: (row) => row.label,
-      },
-      {
-        header: t.sku,
-        cell: (row) => row.sku.toUpperCase(),
-        sortValue: (row) => row.sku,
-      },
-      {
-        header: t.backend,
-        cell: (row) => row.backend,
-        sortValue: (row) => row.backend,
-      },
-      {
-        header: t.phaseHeader,
-        cell: (row) => t.phase[row.phase],
-        sortValue: (row) => t.phase[row.phase],
-      },
-      {
-        header: t.modeHeader,
-        cell: (row) => row.mode,
-        sortValue: (row) => row.mode,
-      },
-      {
-        header: t.precisionHeader,
-        cell: (row) => row.precision,
-        sortValue: (row) => row.precision,
-      },
-      {
-        header: t.epHeader,
-        cell: (row) => `EP${row.topology.ep_size}`,
-        sortValue: (row) => row.topology.ep_size,
-      },
-      {
-        header: t.topologyHeader,
-        cell: (row) => collectiveXTopologyLabel(row.topology),
-        sortValue: (row) => collectiveXTopologyLabel(row.topology),
-        className: 'whitespace-nowrap',
-      },
-      {
-        header: t.dispositionHeader,
-        cell: (row) => t.disposition[row.disposition],
-        sortValue: (row) => t.disposition[row.disposition],
-      },
-      {
-        header: t.outcomeHeader,
-        cell: (row) => <OutcomeBadge outcome={row.outcome} />,
-        sortValue: (row) => t.outcome[row.outcome],
-      },
-      {
-        header: t.reason,
-        cell: (row) => (row.reason ? collectiveXReasonLabel(row.reason, locale) : '-'),
-        sortValue: (row) =>
-          row.reason ? `${collectiveXReasonLabel(row.reason, locale)} ${row.reason}` : '',
-      },
-    ],
-    [locale, t],
-  );
-
-  return (
-    <Card data-testid="collectivex-coverage">
-      <h2 className="text-lg font-semibold">{t.terminalCoverage}</h2>
-      <DataTable
-        data={coverage}
-        columns={columns}
-        testId="collectivex-coverage-table"
-        analyticsPrefix="collectivex_coverage_table"
-      />
-    </Card>
-  );
-}
diff --git a/packages/app/src/components/collectivex/data.test.ts b/packages/app/src/components/collectivex/data.test.ts
index 5e7a7c80a..5d050c88c 100644
--- a/packages/app/src/components/collectivex/data.test.ts
+++ b/packages/app/src/components/collectivex/data.test.ts
@@ -2,17 +2,21 @@ import { describe, expect, it } from 'vitest';
 
 import {
   chartPoints,
+  collectiveXCaseLabel,
   collectiveXColorKey,
   collectiveXLegendLabel,
   collectiveXRunDasharray,
   collectiveXSeriesForRun,
   collectiveXSeriesLabel,
+  collectiveXSkuLabel,
   collectiveXTopologyLabel,
   fitAlphaBeta,
   metricValue,
+  normalizeCollectiveXSku,
   seriesMatchesSelection,
   type CollectiveXSeriesSelection,
   collectiveXKvChartPoints,
+  collectiveXKvFrontierPoints,
   type CollectiveXKvRunCase,
   collectiveXKvCell,
 } from './data';
@@ -39,15 +43,21 @@ describe('collectiveXTopologyLabel', () => {
 });
 
 describe('collectiveXSeriesLabel', () => {
+  it('normalizes runner-pool suffixes in known SKU labels', () => {
+    expect(normalizeCollectiveXSku('B200-nscale')).toBe('b200');
+    expect(normalizeCollectiveXSku('H100-DGXC')).toBe('h100');
+    expect(collectiveXSkuLabel('B200-nscale')).toBe('B200');
+    expect(collectiveXCaseLabel('h100-dgxc · deepep-v2', 'h100-dgxc')).toBe('H100 · deepep-v2');
+    expect(normalizeCollectiveXSku('future-accelerator')).toBe('future-accelerator');
+  });
+
   it('renders the varying identity axes of a series', () => {
-    expect(collectiveXSeriesLabel(scaleUp)).toBe(
-      'H200-DGXC · deepep-v2 · EP8 · normal · decode · bf16',
-    );
+    expect(collectiveXSeriesLabel(scaleUp)).toBe('H200 · deepep-v2 · EP8 · normal · decode · bf16');
   });
 
   it('distinguishes the dispatch precision', () => {
     expect(collectiveXSeriesLabel(makeCollectiveXSeries({ precision: 'fp8' }))).toBe(
-      'H200-DGXC · deepep-v2 · EP8 · normal · decode · fp8',
+      'H200 · deepep-v2 · EP8 · normal · decode · fp8',
     );
   });
 
@@ -58,14 +68,14 @@ describe('collectiveXSeriesLabel', () => {
   it('includes the run id for namespaced multi-run series', () => {
     const [runSeries] = collectiveXSeriesForRun([scaleUp], '30165164821');
     expect(collectiveXSeriesLabel(runSeries)).toBe(
-      '#30165164821 · H200-DGXC · deepep-v2 · EP8 · normal · decode · bf16',
+      '#30165164821 · H200 · deepep-v2 · EP8 · normal · decode · bf16',
     );
   });
 
   it('keeps run ids out of the configuration label used by the legend', () => {
     const [runSeries] = collectiveXSeriesForRun([scaleUp], '30165164821');
     expect(collectiveXLegendLabel(runSeries)).toBe(
-      'H200-DGXC · deepep-v2 · EP8 · normal · decode · bf16',
+      'H200 · deepep-v2 · EP8 · normal · decode · bf16',
     );
   });
 });
@@ -109,6 +119,12 @@ describe('collectiveXColorKey', () => {
     const [second] = collectiveXSeriesForRun([scaleUp], '30177021271');
     expect(collectiveXColorKey(first)).toBe(collectiveXColorKey(second));
   });
+
+  it('assigns the same hardware color across runner-pool SKU suffixes', () => {
+    const nscale = makeCollectiveXSeries({ sku: 'b200-nscale' });
+    const dgxc = makeCollectiveXSeries({ sku: 'b200-dgxc' });
+    expect(collectiveXColorKey(nscale)).toBe(collectiveXColorKey(dgxc));
+  });
 });
 
 describe('collectiveXSeriesForRun', () => {
@@ -292,6 +308,8 @@ describe('chartPoints', () => {
   });
 });
 
+const p95 = (value: number) => ({ p50: value, p95: value, min: value, max: value, n: 24 });
+
 const kvRow = ({
   latency_p50,
   ...overrides
@@ -423,4 +441,71 @@ describe('collectiveXKvChartPoints', () => {
     });
     expect(points).toEqual([]);
   });
+
+  describe('collectiveXKvFrontierPoints', () => {
+    const selection = { op: 'pull', pageTokens: 64 } as const;
+
+    it('plots aggregate GB/s against p95 per in-flight request at the largest ISL', () => {
+      const points = collectiveXKvFrontierPoints(
+        [
+          kase([
+            { isl: 4096, batch: 1, gbps_p50: 5, latency_ms: p95(2) },
+            { isl: 32768, batch: 1, gbps_p50: 7.5, latency_ms: p95(24) },
+            { isl: 32768, batch: 16, gbps_p50: 30, latency_ms: p95(96) },
+            { isl: 32768, batch: 16, op: 'push', gbps_p50: 99, latency_ms: p95(1) },
+          ]),
+        ],
+        selection,
+      );
+      expect(points.map((point) => [point.x, point.y])).toEqual([
+        [7.5, 24],
+        [30, 6],
+      ]);
+      expect(points[0].sku).toBe('gb200');
+    });
+
+    it('fades within-series dominated points and keeps the ladder frontier', () => {
+      const points = collectiveXKvFrontierPoints(
+        [
+          kase([
+            // batch 4 beats batch 1 on both axes; batch 16 trades latency for rate.
+            { batch: 1, gbps_p50: 7.5, latency_ms: p95(26) },
+            { batch: 4, gbps_p50: 20, latency_ms: p95(40) },
+            { batch: 16, gbps_p50: 30, latency_ms: p95(320) },
+          ]),
+        ],
+        selection,
+      );
+      expect(points.map((point) => point.onSeriesFrontier)).toEqual([false, true, true]);
+    });
+
+    it('rings only the SKU-wide frontier across backends of the same SKU', () => {
+      const points = collectiveXKvFrontierPoints(
+        [
+          kase([{ batch: 1, gbps_p50: 20, latency_ms: p95(10) }]),
+          kase([{ batch: 1, gbps_p50: 8, latency_ms: p95(30) }], {
+            case_id: 'gb200-mooncake-kv-dsv4-rdma-xfer-ep2-paged-fp8',
+            backend: 'mooncake',
+          }),
+        ],
+        selection,
+      );
+      expect(points.map((point) => point.onSeriesFrontier)).toEqual([true, true]);
+      expect(points.map((point) => point.onSkuFrontier)).toEqual([true, false]);
+    });
+
+    it('never lets one SKU dominate another', () => {
+      const points = collectiveXKvFrontierPoints(
+        [
+          kase([{ batch: 1, gbps_p50: 20, latency_ms: p95(10) }]),
+          kase([{ batch: 1, gbps_p50: 8, latency_ms: p95(30) }], {
+            case_id: 'h200-nixl-kv-dsv4-rdma-xfer-ep2-paged-fp8',
+            sku: 'h200',
+          }),
+        ],
+        selection,
+      );
+      expect(points.map((point) => point.onSkuFrontier)).toEqual([true, true]);
+    });
+  });
 });
diff --git a/packages/app/src/components/collectivex/data.ts b/packages/app/src/components/collectivex/data.ts
index 857a884eb..90f0b93bf 100644
--- a/packages/app/src/components/collectivex/data.ts
+++ b/packages/app/src/components/collectivex/data.ts
@@ -1,3 +1,5 @@
+import { GPU_KEYS } from '@semianalysisai/inferencex-constants';
+
 import type {
   CollectiveXChartPoint,
   CollectiveXComponent,
@@ -23,6 +25,25 @@ export interface CollectiveXSeriesSelection {
 
 const BASE_RUN_DASHARRAYS = ['none', '9 4', '3 3', '10 3 2 3', '2 3', '12 3 2 3'] as const;
 
+/**
+ * CollectiveX artifacts identify runner pools in the SKU (for example,
+ * `b200-nscale` or `h100-dgxc`). Collapse known hardware identifiers to the
+ * canonical GPU key while leaving unknown SKU structure intact.
+ */
+export function normalizeCollectiveXSku(sku: string): string {
+  const normalized = sku.trim().toLowerCase();
+  const base = normalized.split(':').at(-1)?.split('-')[0] ?? normalized;
+  return GPU_KEYS.has(base) ? base : normalized;
+}
+
+export function collectiveXSkuLabel(sku: string): string {
+  return normalizeCollectiveXSku(sku).toUpperCase();
+}
+
+export function collectiveXCaseLabel(label: string, sku: string): string {
+  return label.startsWith(sku) ? `${collectiveXSkuLabel(sku)}${label.slice(sku.length)}` : label;
+}
+
 /**
  * Stable within the newest-first run table. Preserve the six simple patterns,
  * then encode higher indexes as a unique dash/gap/accent tuple so two listed
@@ -57,7 +78,7 @@ export function collectiveXTopologyLabel(
 }
 
 export function collectiveXLegendLabel(series: CollectiveXSeries): string {
-  return `${series.system.sku.toUpperCase()} · ${series.backend} · EP${series.system.ep_size} · ${series.mode} · ${series.phase} · ${series.precision}`;
+  return `${collectiveXSkuLabel(series.system.sku)} · ${series.backend} · EP${series.system.ep_size} · ${series.mode} · ${series.phase} · ${series.precision}`;
 }
 
 export function collectiveXSeriesLabel(series: CollectiveXSeries | CollectiveXRunSeries): string {
@@ -66,7 +87,7 @@ export function collectiveXSeriesLabel(series: CollectiveXSeries | CollectiveXRu
 }
 
 export function collectiveXColorKey(series: CollectiveXSeries | CollectiveXRunSeries): string {
-  return `${series.system.vendor}_${series.system.sku}_${series.backend}_ep${series.system.ep_size}_${series.mode}_${series.phase}_${series.precision}`;
+  return `${series.system.vendor}_${normalizeCollectiveXSku(series.system.sku)}_${series.backend}_ep${series.system.ep_size}_${series.mode}_${series.phase}_${series.precision}`;
 }
 
 /** Namespace series ids and attach the run's current selection-order style index. */
@@ -244,11 +265,11 @@ export interface CollectiveXKvChartPoint {
 }
 
 export function collectiveXKvColorKey(kase: CollectiveXKvCase): string {
-  return `${kase.vendor ?? 'unknown'}_${kase.sku}_${kase.backend}_${kase.fabric}_${kase.precision}`;
+  return `${kase.vendor ?? 'unknown'}_${normalizeCollectiveXSku(kase.sku)}_${kase.backend}_${kase.fabric}_${kase.precision}`;
 }
 
 export function collectiveXKvLegendLabel(kase: CollectiveXKvCase): string {
-  return `${kase.sku.toUpperCase()} · ${kase.backend} · ${kase.fabric} · ${kase.precision}`;
+  return `${collectiveXSkuLabel(kase.sku)} · ${kase.backend} · ${kase.fabric} · ${kase.precision}`;
 }
 
 /**
@@ -257,6 +278,86 @@ export function collectiveXKvLegendLabel(kase: CollectiveXKvCase): string {
  * story); ISL on the x axis reads at batch 1 (a single request's handoff).
  * Paged rows only: the single-descriptor bulk ceiling stays a table column.
  */
+export interface CollectiveXKvFrontierSelection {
+  op: 'pull' | 'push';
+  pageTokens: number;
+}
+
+export interface CollectiveXKvFrontierPoint {
+  seriesId: string;
+  seriesLabel: string;
+  colorKey: string;
+  sku: string;
+  /** Aggregate burst bandwidth at p50 (GB/s). */
+  x: number;
+  /** Burst p95 latency divided by requests in flight (ms per request). */
+  y: number;
+  onSeriesFrontier: boolean;
+  onSkuFrontier: boolean;
+  row: CollectiveXKvRow;
+}
+
+/** More throughput at no worse per-request cost, strictly better in one. */
+function frontierDominates(
+  a: Pick<CollectiveXKvFrontierPoint, 'x' | 'y'>,
+  b: Pick<CollectiveXKvFrontierPoint, 'x' | 'y'>,
+): boolean {
+  return a.x >= b.x && a.y <= b.y && (a.x > b.x || a.y < b.y);
+}
+
+/**
+ * Concurrency-frontier points: each case's batch ladder at the largest
+ * measured ISL (the same slice the batch view plots), re-expressed as
+ * aggregate GB/s against p95 latency per in-flight request. A backend that
+ * overlaps requests traces a curve toward the lower right; a serializing
+ * backend collapses to a point (constant aggregate rate, constant
+ * per-request cost). Two dominance tiers: within a series (its own best
+ * operating points) and across every series of the same SKU (which backend
+ * to deploy on that machine).
+ */
+export function collectiveXKvFrontierPoints(
+  cases: readonly CollectiveXKvRunCase[],
+  selection: CollectiveXKvFrontierSelection,
+): CollectiveXKvFrontierPoint[] {
+  const points = cases.flatMap((kase) => {
+    const matching = kase.rows.filter(
+      (row) =>
+        row.kind === 'paged' && row.op === selection.op && row.page_tokens === selection.pageTokens,
+    );
+    if (matching.length === 0) return [];
+    const isl = Math.max(...matching.map((row) => row.isl));
+    return matching
+      .filter((row) => row.isl === isl)
+      .flatMap((row) => {
+        const x = row.gbps_p50;
+        const y = row.latency_ms.p95 / row.batch;
+        if (!Number.isFinite(x) || x <= 0 || !Number.isFinite(y) || y <= 0) return [];
+        return [
+          {
+            seriesId: `${kase.run_id}:${kase.case_id}`,
+            seriesLabel: `#${kase.run_id} · ${collectiveXKvLegendLabel(kase)}`,
+            colorKey: collectiveXKvColorKey(kase),
+            sku: normalizeCollectiveXSku(kase.sku),
+            x,
+            y,
+            onSeriesFrontier: false,
+            onSkuFrontier: false,
+            row,
+          },
+        ];
+      });
+  });
+  for (const point of points) {
+    point.onSeriesFrontier = !points.some(
+      (other) => other.seriesId === point.seriesId && frontierDominates(other, point),
+    );
+    point.onSkuFrontier = !points.some(
+      (other) => other.sku === point.sku && frontierDominates(other, point),
+    );
+  }
+  return points;
+}
+
 export function collectiveXKvChartPoints(
   cases: readonly CollectiveXKvRunCase[],
   selection: CollectiveXKvChartSelection,
diff --git a/packages/app/src/components/compare/agentx-compare-hero.tsx b/packages/app/src/components/compare/agentx-compare-hero.tsx
new file mode 100644
index 000000000..e5d0df27e
--- /dev/null
+++ b/packages/app/src/components/compare/agentx-compare-hero.tsx
@@ -0,0 +1,166 @@
+import { ArrowRight } from 'lucide-react';
+
+import { Card } from '@/components/ui/card';
+import { MinecraftSplash } from '@/components/minecraft/minecraft-splash';
+import { NewBadge } from '@/components/ui/new-badge';
+import { agentxDashboardHref, FEATURED_AGENTX_MODELS } from '@/lib/compare-agentx';
+
+import { CompareIndexTrackedLink } from './compare-index-tracked-link';
+
+/**
+ * Full-color brand marks for the ledger rows, sharing the `*-color.svg`
+ * assets used by `/model` pages and inference chart captions. Keyed by
+ * compare slug so a featured model without a registered mark simply renders
+ * without one instead of breaking the row.
+ */
+const MODEL_LOGOS: Record<string, string> = {
+  'kimi-k3': '/logos/kimi-color.svg',
+  'deepseek-v4': '/logos/deepseek-color.svg',
+  // GLM ships under the Z.ai product brand, so the ledger shows the Z.ai
+  // mark rather than the Zhipu corporate dot cluster.
+  'glm-5-2': '/logos/zai-color.svg',
+  'minimax-m3': '/logos/minimax-color.svg',
+  'qwen-3-8-flash-next': '/logos/qwen-color.svg',
+  'qwen-3-5': '/logos/qwen-color.svg',
+};
+
+const STRINGS = {
+  en: {
+    eyebrow: 'AgentX / live results',
+    title: 'Compare Realistic Agentic Inference Perf',
+    description:
+      'Long Context Multi Turn Inference Performance. Compare Across OpenAI Jalapeño, MI355X, GB300 NVL72, GB200 NVL72, B200, H200, H100, RTX Pro, and soon TPUv7/v8 & Rubin NVL72 & MI455X UALoE72',
+    overview: 'Overview',
+    dashboard: 'Full dashboard',
+    ledgerTitle: 'Models with AgentX results',
+    modelAction: 'View results',
+    newModel: 'NEW',
+  },
+  zh: {
+    eyebrow: 'AgentX｜最新结果',
+    title: '真实智能体工作负载下的推理性能对比',
+    description:
+      '比较不同硬件平台在长上下文、多轮智能体工作负载下的推理性能，覆盖 OpenAI Jalapeño、MI355X、GB300 NVL72、GB200 NVL72、B200、H200、H100 和 RTX Pro，即将支持 TPUv7/v8、Rubin NVL72 与 MI455X UALoE72。',
+    overview: '总览',
+    dashboard: '查看完整仪表板',
+    ledgerTitle: '已发布 AgentX 结果的模型',
+    modelAction: '查看结果',
+    newModel: '新',
+  },
+} as const;
+
+/**
+ * The hero leads `/compare` and the landing page. `/compare` owns the page
+ * `h1`; the landing page renders the hero under its own section flow, so the
+ * caller picks the heading level rather than shipping a second `h1`.
+ */
+export function AgentXCompareHero({
+  locale,
+  headingLevel = 'h1',
+  surface = 'compare',
+}: {
+  locale: 'en' | 'zh';
+  headingLevel?: 'h1' | 'h2';
+  surface?: 'compare' | 'landing';
+}) {
+  const t = STRINGS[locale];
+  const prefix = locale === 'zh' ? '/zh' : '';
+  const Heading = headingLevel;
+
+  return (
+    <section data-testid="compare-agentx-primary">
+      <Card className="overflow-hidden p-0 md:p-0">
+        <div className="grid lg:grid-cols-[minmax(0,1.15fr)_minmax(19rem,0.85fr)]">
+          <div className="flex flex-col justify-center px-6 py-5 md:px-8 md:py-6 lg:px-10 lg:py-7">
+            <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
+              {t.eyebrow}
+            </p>
+            {/* `relative` anchors the splash, which positions itself absolutely
+                at the top right. Landing only: /compare is not the launch
+                surface, and the announcement belongs on the front page. */}
+            <div className="relative">
+              <Heading className="mt-3 max-w-2xl text-2xl/[1.8rem] font-semibold tracking-tight text-foreground lg:text-[2.4rem]/[2.4rem]">
+                {t.title}
+              </Heading>
+              {surface === 'landing' && <MinecraftSplash />}
+            </div>
+            <p className="mt-3 max-w-3xl text-base leading-7 text-muted-foreground lg:text-lg">
+              {t.description}
+            </p>
+            <div className="mt-5 flex flex-wrap gap-3">
+              <CompareIndexTrackedLink
+                data-testid="compare-agentx-overview-link"
+                href={`${prefix}/overview`}
+                analyticsEvent="compare_agentx_overview_clicked"
+                analyticsSurface={surface}
+                appNavigation
+                className="inline-flex min-h-11 items-center gap-2 rounded-md bg-brand px-5 py-2.5 font-semibold text-primary-foreground shadow-sm transition-colors hover:bg-brand/90"
+              >
+                {t.overview}
+                <ArrowRight aria-hidden="true" className="size-4" />
+              </CompareIndexTrackedLink>
+              <CompareIndexTrackedLink
+                data-testid="compare-agentx-dashboard-link"
+                href={agentxDashboardHref(locale, FEATURED_AGENTX_MODELS[0])}
+                analyticsEvent="compare_agentx_dashboard_clicked"
+                analyticsTarget="kimi-k3"
+                analyticsSurface={surface}
+                appNavigation
+                className="inline-flex min-h-11 items-center gap-2 rounded-md border border-border px-5 py-2.5 font-semibold text-foreground transition-colors hover:bg-muted"
+              >
+                {t.dashboard}
+                <ArrowRight aria-hidden="true" className="size-4" />
+              </CompareIndexTrackedLink>
+            </div>
+          </div>
+
+          <div className="border-t border-border/70 bg-muted/15 lg:border-t-0 lg:border-l">
+            {/* The visible ledger header is dropped; `ledgerTitle` stays as the
+                nav's accessible name so screen readers still get the label. */}
+            <nav aria-label={t.ledgerTitle} className="divide-y divide-border/70">
+              {FEATURED_AGENTX_MODELS.map((model) => (
+                <CompareIndexTrackedLink
+                  key={model.slug}
+                  data-testid={`compare-agentx-model-${model.slug}`}
+                  href={agentxDashboardHref(locale, model)}
+                  analyticsEvent="compare_agentx_model_clicked"
+                  analyticsTarget={model.slug}
+                  analyticsSurface={surface}
+                  appNavigation
+                  className="group flex min-h-14 items-center justify-between gap-4 px-5 py-2.5 transition-colors hover:bg-brand/5 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-inset focus-visible:ring-ring"
+                >
+                  <span className="flex min-w-0 items-center gap-3">
+                    {MODEL_LOGOS[model.slug] && (
+                      <img
+                        src={MODEL_LOGOS[model.slug]}
+                        alt=""
+                        aria-hidden="true"
+                        width={32}
+                        height={32}
+                        loading="lazy"
+                        className="size-8 shrink-0 object-contain"
+                      />
+                    )}
+                    <span className="min-w-0">
+                      <span className="flex items-center gap-1.5 text-sm font-semibold leading-tight text-foreground group-hover:text-brand">
+                        <span className="min-w-0">{model.label}</span>
+                        <NewBadge data-new-badge="agentx-ledger">{t.newModel}</NewBadge>
+                      </span>
+                      <span className="mt-1 block font-mono text-3xs tracking-eyebrow text-brand uppercase">
+                        AgentX
+                      </span>
+                    </span>
+                  </span>
+                  <span className="inline-flex shrink-0 items-center gap-1 text-xs font-medium text-muted-foreground group-hover:text-foreground">
+                    {t.modelAction}
+                    <ArrowRight aria-hidden="true" className="size-3.5" />
+                  </span>
+                </CompareIndexTrackedLink>
+              ))}
+            </nav>
+          </div>
+        </div>
+      </Card>
+    </section>
+  );
+}
diff --git a/packages/app/src/components/compare/compare-index-tracked-link.tsx b/packages/app/src/components/compare/compare-index-tracked-link.tsx
new file mode 100644
index 000000000..5d372dfc8
--- /dev/null
+++ b/packages/app/src/components/compare/compare-index-tracked-link.tsx
@@ -0,0 +1,51 @@
+'use client';
+
+import Link from 'next/link';
+import { useRouter } from 'next/navigation';
+
+import { track } from '@/lib/analytics';
+import { navigateInApp } from '@/lib/client-navigation';
+
+interface CompareIndexTrackedLinkProps extends React.ComponentProps<typeof Link> {
+  analyticsEvent:
+    | 'compare_agentx_overview_clicked'
+    | 'compare_agentx_dashboard_clicked'
+    | 'compare_agentx_model_clicked';
+  analyticsTarget?: string;
+  /** Which page rendered the hero, so `/compare` and `/` clicks stay separable. */
+  analyticsSurface?: string;
+  /** Route the click through `navigateInApp`, the way the header nav and the
+   *  landing card already do for `/overview` and `/inference`. The first
+   *  dashboard transition can request the route without committing the URL, so
+   *  those destinations need the retry; content routes do not. */
+  appNavigation?: boolean;
+}
+
+export function CompareIndexTrackedLink({
+  analyticsEvent,
+  analyticsTarget,
+  analyticsSurface,
+  appNavigation = false,
+  onClick,
+  ...props
+}: CompareIndexTrackedLinkProps) {
+  const router = useRouter();
+
+  return (
+    <Link
+      {...props}
+      onClick={(event) => {
+        onClick?.(event);
+        if (event.defaultPrevented) return;
+        const payload = {
+          ...(analyticsTarget ? { target: analyticsTarget } : {}),
+          ...(analyticsSurface ? { surface: analyticsSurface } : {}),
+        };
+        track(analyticsEvent, Object.keys(payload).length > 0 ? payload : undefined);
+        if (appNavigation && typeof props.href === 'string') {
+          navigateInApp(event, router, props.href);
+        }
+      }}
+    />
+  );
+}
diff --git a/packages/app/src/components/compare/compare-interpolated-table.tsx b/packages/app/src/components/compare/compare-interpolated-table.tsx
index ff8a264b2..35296e06e 100644
--- a/packages/app/src/components/compare/compare-interpolated-table.tsx
+++ b/packages/app/src/components/compare/compare-interpolated-table.tsx
@@ -1,6 +1,6 @@
 'use client';
 
-import { useCallback, useEffect, useMemo, useState } from 'react';
+import { useCallback, useMemo, useState } from 'react';
 
 import { interpolateForGPU } from '@/components/calculator/interpolation';
 import type { GPUDataPoint, InterpolatedResult } from '@/components/calculator/types';
@@ -45,9 +45,12 @@ export interface CompareInterpolatedTableProps {
   metricLabelOverrides?: Record<string, string>;
 }
 
-interface ColumnData {
+interface ColumnIntent {
   target: number;
   inputValue: string;
+}
+
+interface ColumnData extends ColumnIntent {
   a: InterpolatedResult | null;
   b: InterpolatedResult | null;
 }
@@ -105,46 +108,37 @@ export function CompareInterpolatedTable({
       ? { ...m, label: metricLabelOverrides[m.label] }
       : m,
   );
-  const [columns, setColumns] = useState<ColumnData[]>(() =>
-    defaultTargets.map((target, i) => ({
+  const [columnIntents, setColumnIntents] = useState<ColumnIntent[]>(() =>
+    defaultTargets.map((target) => ({
       target,
       inputValue: String(target),
-      a: ssrRows[i]?.a ?? null,
-      b: ssrRows[i]?.b ?? null,
     })),
   );
 
-  const hasClientDataA = gpuDataPointsA.length > 0;
-  const hasClientDataB = gpuDataPointsB.length > 0;
-  const hasClientData = hasClientDataA || hasClientDataB;
-
-  // When client-side data changes (model/sequence/precision), recompute all columns
-  useEffect(() => {
-    if (!hasClientData) return;
-    setColumns((prev) =>
-      prev.map((col) => ({
-        ...col,
-        a: hasClientDataA
-          ? interpolateForGPU(gpuDataPointsA, col.target, 'interactivity_to_throughput', 'costh')
-          : col.a,
-        b: hasClientDataB
-          ? interpolateForGPU(gpuDataPointsB, col.target, 'interactivity_to_throughput', 'costh')
-          : col.b,
+  const columns = useMemo<ColumnData[]>(
+    () =>
+      columnIntents.map((column, index) => ({
+        ...column,
+        a:
+          gpuDataPointsA.length > 0
+            ? interpolateForGPU(
+                gpuDataPointsA,
+                column.target,
+                'interactivity_to_throughput',
+                'costh',
+              )
+            : (ssrRows[index]?.a ?? null),
+        b:
+          gpuDataPointsB.length > 0
+            ? interpolateForGPU(
+                gpuDataPointsB,
+                column.target,
+                'interactivity_to_throughput',
+                'costh',
+              )
+            : (ssrRows[index]?.b ?? null),
       })),
-    );
-  }, [gpuDataPointsA, gpuDataPointsB, hasClientData, hasClientDataA, hasClientDataB]);
-
-  const reinterpolate = useCallback(
-    (target: number, prevA: InterpolatedResult | null, prevB: InterpolatedResult | null) => {
-      const resultA = hasClientDataA
-        ? interpolateForGPU(gpuDataPointsA, target, 'interactivity_to_throughput', 'costh')
-        : prevA;
-      const resultB = hasClientDataB
-        ? interpolateForGPU(gpuDataPointsB, target, 'interactivity_to_throughput', 'costh')
-        : prevB;
-      return { a: resultA, b: resultB };
-    },
-    [gpuDataPointsA, gpuDataPointsB, hasClientDataA, hasClientDataB],
+    [columnIntents, gpuDataPointsA, gpuDataPointsB, ssrRows],
   );
 
   /**
@@ -154,24 +148,22 @@ export function CompareInterpolatedTable({
   const commitColumnTarget = useCallback(
     (colIndex: number): boolean => {
       let committed = false;
-      setColumns((prev) => {
+      setColumnIntents((prev) => {
         const next = [...prev];
-        const col = next[colIndex];
-        const parsed = parseFloat(col.inputValue);
+        const column = next[colIndex];
+        const parsed = parseFloat(column.inputValue);
 
         if (isNaN(parsed) || parsed <= 0) {
-          next[colIndex] = { ...col, inputValue: String(col.target) };
+          next[colIndex] = { ...column, inputValue: String(column.target) };
           return next;
         }
 
         const clamped = Math.round(
           Math.max(interactivityRange.min, Math.min(interactivityRange.max, parsed)),
         );
-        const results = reinterpolate(clamped, col.a, col.b);
         next[colIndex] = {
           target: clamped,
           inputValue: String(clamped),
-          ...results,
         };
         committed = true;
         return next;
@@ -181,7 +173,7 @@ export function CompareInterpolatedTable({
       }
       return committed;
     },
-    [interactivityRange, reinterpolate],
+    [interactivityRange],
   );
 
   /**
@@ -190,36 +182,26 @@ export function CompareInterpolatedTable({
    */
   const handleInputChange = useCallback(
     (colIndex: number, value: string) => {
-      setColumns((prev) => {
+      setColumnIntents((prev) => {
         const next = [...prev];
-        const col = next[colIndex];
+        const column = next[colIndex];
         const parsed = parseFloat(value);
 
-        // If valid number, update both inputValue and interpolated results
         if (!isNaN(parsed) && parsed > 0) {
           const clamped = Math.round(
             Math.max(interactivityRange.min, Math.min(interactivityRange.max, parsed)),
           );
-          const results = reinterpolate(clamped, col.a, col.b);
           next[colIndex] = {
             target: clamped,
             inputValue: value,
-            ...results,
           };
         } else {
-          next[colIndex] = { ...next[colIndex], inputValue: value };
+          next[colIndex] = { ...column, inputValue: value };
         }
         return next;
       });
     },
-    [
-      hasClientData,
-      hasClientDataA,
-      hasClientDataB,
-      interactivityRange,
-      reinterpolate,
-      gpuDataPointsA,
-    ],
+    [interactivityRange],
   );
 
   const handleInputBlur = useCallback(
diff --git a/packages/app/src/components/compare/compare-page-layout.tsx b/packages/app/src/components/compare/compare-page-layout.tsx
new file mode 100644
index 000000000..9564a6cec
--- /dev/null
+++ b/packages/app/src/components/compare/compare-page-layout.tsx
@@ -0,0 +1,20 @@
+import type { ReactNode } from 'react';
+
+import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+
+interface ComparePageLayoutProps {
+  children: ReactNode;
+}
+
+/** Focused compare-page shell without the dashboard route navigation. */
+export function ComparePageLayout({ children }: ComparePageLayoutProps) {
+  return (
+    <UnofficialRunProvider>
+      <main className="relative">
+        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4 pb-8">
+          {children}
+        </div>
+      </main>
+    </UnofficialRunProvider>
+  );
+}
diff --git a/packages/app/src/components/compare/compare-pair-card-link.tsx b/packages/app/src/components/compare/compare-pair-card-link.tsx
index 0b1054932..384591163 100644
--- a/packages/app/src/components/compare/compare-pair-card-link.tsx
+++ b/packages/app/src/components/compare/compare-pair-card-link.tsx
@@ -4,31 +4,78 @@ import { ArrowRight } from 'lucide-react';
 
 import { track } from '@/lib/analytics';
 
+import { HwVendorLogo } from '@/components/ui/hw-vendor-logo';
+
+/** One side of the "A vs B" pair: display label plus vendor for the logo. */
+interface PairHardware {
+  label: string;
+  vendor?: string;
+}
+
 interface ComparePairCardLinkProps {
   href: string;
   slug: string;
   label: string;
   archLine: string;
+  scenarioLabel?: 'AgentX' | '8K→1K';
+  /** When both sides are provided, the title renders each hardware label with
+   *  its vendor logo beside it instead of the plain `label` string. */
+  hardwareA?: PairHardware;
+  hardwareB?: PairHardware;
 }
 
-export function ComparePairCardLink({ href, slug, label, archLine }: ComparePairCardLinkProps) {
+export function ComparePairCardLink({
+  href,
+  slug,
+  label,
+  archLine,
+  scenarioLabel,
+  hardwareA,
+  hardwareB,
+}: ComparePairCardLinkProps) {
   return (
     <a
       href={href}
+      data-scenario={scenarioLabel}
       className="group relative flex flex-col rounded-xl border border-border bg-background/20 backdrop-blur-[2px] p-5 transition-all duration-200 hover:border-brand/50 hover:shadow-lg hover:shadow-brand/5 hover:scale-[1.01]"
       onClick={(e) => {
         if (e.metaKey || e.ctrlKey || e.shiftKey || e.altKey || e.button !== 0) return;
         e.preventDefault();
-        track('compare_index_pair_clicked', { slug, label });
+        track('compare_index_pair_clicked', {
+          slug,
+          label,
+          ...(scenarioLabel ? { scenario: scenarioLabel } : {}),
+        });
         window.location.href = href;
       }}
     >
       <div className="absolute inset-y-3 left-0 w-0.5 rounded-full bg-brand/60 transition-all duration-200 group-hover:bg-brand group-hover:inset-y-2" />
       <div className="flex items-start justify-between gap-2">
         <div className="min-w-0 flex-1">
-          <h3 className="font-semibold text-sm leading-tight group-hover:text-brand transition-colors duration-200">
-            {label}
-          </h3>
+          <div className="flex flex-wrap items-center gap-2">
+            <h3 className="font-semibold text-sm leading-tight group-hover:text-brand transition-colors duration-200">
+              {hardwareA && hardwareB ? (
+                <>
+                  <span className="inline-flex items-center gap-1.5">
+                    <HwVendorLogo vendor={hardwareA.vendor} />
+                    {hardwareA.label}
+                  </span>{' '}
+                  <span className="font-normal text-muted-foreground">vs</span>{' '}
+                  <span className="inline-flex items-center gap-1.5">
+                    <HwVendorLogo vendor={hardwareB.vendor} />
+                    {hardwareB.label}
+                  </span>
+                </>
+              ) : (
+                label
+              )}
+            </h3>
+            {scenarioLabel && (
+              <span className="inline-flex min-h-5 items-center rounded-full border border-brand/30 bg-brand/10 px-2 font-mono text-3xs font-semibold leading-none text-brand">
+                {scenarioLabel}
+              </span>
+            )}
+          </div>
           <p className="text-xs text-muted-foreground leading-relaxed mt-1.5">{archLine}</p>
         </div>
         <ArrowRight className="size-4 shrink-0 text-muted-foreground transition-all duration-200 group-hover:translate-x-0.5 group-hover:text-brand" />
diff --git a/packages/app/src/components/compare/compare-table-renderer.tsx b/packages/app/src/components/compare/compare-table-renderer.tsx
new file mode 100644
index 000000000..4238003ad
--- /dev/null
+++ b/packages/app/src/components/compare/compare-table-renderer.tsx
@@ -0,0 +1,55 @@
+import type { GPUDataPoint, InterpolatedResult } from '@/components/calculator/types';
+import { CompareInterpolatedTable } from '@/components/compare/compare-interpolated-table';
+
+export interface CompareTableData {
+  defaultTargets: number[];
+  ssrRows: { target: number; a: InterpolatedResult | null; b: InterpolatedResult | null }[];
+  interactivityRange: { min: number; max: number };
+}
+
+interface CompareTableRendererProps {
+  aLabel: string;
+  bLabel: string;
+  ssrTableData: CompareTableData;
+  interactivityRange: { min: number; max: number };
+  gpuDataPointsA: GPUDataPoint[];
+  gpuDataPointsB: GPUDataPoint[];
+  emptyStateText: string;
+  visibleMetricLabels?: string[];
+  metricLabelOverrides?: Record<string, string>;
+}
+
+/** Pure empty/data renderer shared by every compare-page client. */
+export function CompareTableRenderer({
+  aLabel,
+  bLabel,
+  ssrTableData,
+  interactivityRange,
+  gpuDataPointsA,
+  gpuDataPointsB,
+  emptyStateText,
+  visibleMetricLabels,
+  metricLabelOverrides,
+}: CompareTableRendererProps) {
+  if (ssrTableData.defaultTargets.length === 0) {
+    return (
+      <div className="border border-border/50 rounded-md px-4 py-3 text-sm text-muted-foreground bg-muted/30">
+        {emptyStateText}
+      </div>
+    );
+  }
+
+  return (
+    <CompareInterpolatedTable
+      aLabel={aLabel}
+      bLabel={bLabel}
+      ssrRows={ssrTableData.ssrRows}
+      defaultTargets={ssrTableData.defaultTargets}
+      interactivityRange={interactivityRange}
+      gpuDataPointsA={gpuDataPointsA}
+      gpuDataPointsB={gpuDataPointsB}
+      visibleMetricLabels={visibleMetricLabels}
+      metricLabelOverrides={metricLabelOverrides}
+    />
+  );
+}
diff --git a/packages/app/src/components/dashboard-shell.tsx b/packages/app/src/components/dashboard-shell.tsx
index d44ff8432..68c98d19f 100644
--- a/packages/app/src/components/dashboard-shell.tsx
+++ b/packages/app/src/components/dashboard-shell.tsx
@@ -4,30 +4,42 @@ import { GlobalFilterProvider } from '@/components/GlobalFilterContext';
 import { NudgeEngine } from '@/components/nudge-engine';
 import { TabNav } from '@/components/tab-nav';
 import { UnofficialRunProvider } from '@/components/unofficial-run-provider';
+import { dashboardShellCapabilitiesForPathname } from '@/lib/dashboard-routes';
+import { inferenceModelForPathname } from '@/lib/inference-model-slug';
 import { usePathname } from 'next/navigation';
 
 export function DashboardShell({ children }: { children: React.ReactNode }) {
   const pathname = usePathname();
-  const content = (
-    <main className="relative">
-      <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-4">
-        <TabNav />
-        {children}
-      </div>
-    </main>
-  );
-  if (pathname === '/collectivex' || pathname === '/zh/collectivex') return content;
+  const { providers: providerCapabilities, dashboardNudge } =
+    dashboardShellCapabilitiesForPathname(pathname);
+
+  let content = children;
+  if (providerCapabilities.globalFilters) {
+    // `/inference/<model>` pages pin the model from the path. Seeding the
+    // provider (rather than only applying it in an effect) keeps the server
+    // render and first client paint on the right model — `usePathname` is
+    // available during SSR, so hydration agrees. Soft navigations between
+    // model pages don't remount this provider; those are handled by the
+    // pathname-keyed effect inside GlobalFilterProvider.
+    content = (
+      <GlobalFilterProvider initialModel={inferenceModelForPathname(pathname) ?? undefined}>
+        {content}
+      </GlobalFilterProvider>
+    );
+  }
+  if (providerCapabilities.unofficialRuns) {
+    content = <UnofficialRunProvider>{content}</UnofficialRunProvider>;
+  }
+
   return (
     <>
-      <NudgeEngine scope="dashboard" />
-      <UnofficialRunProvider>
-        <main className="relative">
-          <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-4">
-            <TabNav />
-            <GlobalFilterProvider>{children}</GlobalFilterProvider>
-          </div>
-        </main>
-      </UnofficialRunProvider>
+      {dashboardNudge && <NudgeEngine scope="dashboard" />}
+      <main className="relative">
+        <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-4">
+          <TabNav />
+          {content}
+        </div>
+      </main>
     </>
   );
 }
diff --git a/packages/app/src/components/datasets/agentx-methodology-article.tsx b/packages/app/src/components/datasets/agentx-methodology-article.tsx
new file mode 100644
index 000000000..966b037de
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-methodology-article.tsx
@@ -0,0 +1,641 @@
+import Image from 'next/image';
+
+import { AgentXMethodologyLink } from './agentx-methodology-link';
+
+type Locale = 'en' | 'zh';
+
+const ASSET_ROOT = '/images/agentx-methodology';
+
+const FIGURE_ASSETS = {
+  corpus: { src: `${ASSET_ROOT}/corpus-scale.png`, width: 4200, height: 1980 },
+  hashRatio: { src: `${ASSET_ROOT}/hash-token-ratio.png`, width: 4200, height: 3690 },
+  traceRecord: { src: `${ASSET_ROOT}/weka-trace-record.png`, width: 4200, height: 4746 },
+  requestDistributions: {
+    src: `${ASSET_ROOT}/request-distributions.png`,
+    width: 4200,
+    height: 4110,
+  },
+  requestDistributions256k: {
+    src: `${ASSET_ROOT}/request-distributions-256k.png`,
+    width: 4200,
+    height: 4110,
+  },
+  subagentDistributions: {
+    src: `${ASSET_ROOT}/subagent-distributions.png`,
+    width: 4200,
+    height: 2850,
+  },
+  subagentDistributions256k: {
+    src: `${ASSET_ROOT}/subagent-distributions-256k.png`,
+    width: 4200,
+    height: 2850,
+  },
+  replayLinear: { src: `${ASSET_ROOT}/replay-linear.png`, width: 4200, height: 2397 },
+  replaySingle: {
+    src: `${ASSET_ROOT}/replay-single-subagent.png`,
+    width: 4200,
+    height: 3546,
+  },
+  replayJoined: {
+    src: `${ASSET_ROOT}/replay-two-subagents-joined.png`,
+    width: 4200,
+    height: 4080,
+  },
+  replayTwo: {
+    src: `${ASSET_ROOT}/replay-two-subagents.png`,
+    width: 4200,
+    height: 4479,
+  },
+  replayFlatspawn: {
+    src: `${ASSET_ROOT}/replay-flatspawn.png`,
+    width: 4200,
+    height: 4134,
+  },
+  replaySidecars: {
+    src: `${ASSET_ROOT}/replay-subagent-sidecars.png`,
+    width: 4200,
+    height: 4818,
+  },
+  replayMulti: {
+    src: `${ASSET_ROOT}/replay-multi-subagent.png`,
+    width: 4200,
+    height: 5130,
+  },
+  queueDepth: { src: `${ASSET_ROOT}/request-queue-depth.png`, width: 1360, height: 612 },
+  pareto: { src: `${ASSET_ROOT}/pareto-b200-minimax-m3.png`, width: 2048, height: 1167 },
+  warmup: { src: `${ASSET_ROOT}/warmup-snapshot.png`, width: 4200, height: 2712 },
+  warmupCache: { src: `${ASSET_ROOT}/warmup-cache-sources.png`, width: 2048, height: 895 },
+  profilingCache: {
+    src: `${ASSET_ROOT}/profiling-cache-sources.png`,
+    width: 2048,
+    height: 897,
+  },
+  acceptanceControls: {
+    src: `${ASSET_ROOT}/engine-acceptance-controls.png`,
+    width: 2048,
+    height: 1376,
+  },
+  goldenAl: { src: `${ASSET_ROOT}/golden-acceptance-length.png`, width: 747, height: 662 },
+} as const;
+
+const CONTENT = {
+  en: {
+    eyebrow: 'Field guide / AI infrastructure',
+    title: 'AgentX Methodology',
+    lead: 'AgentX converts opt-in Claude Code proxy traces into deterministic AIPerf workloads. This page documents how collection metadata becomes a replay and which controls define a valid benchmark result.',
+    facts: [
+      { value: '393', label: 'published sessions' },
+      { value: '64', label: 'tokens per hash block' },
+      { value: '1 hour', label: 'profiling window' },
+      { value: '25–75%', label: 'seeded start range' },
+    ],
+    collectionTitle: 'Dataset collection',
+    collectionParagraphs: [
+      'Participants opt in to an HTTP proxy that records request arrival and completion times, input and output token counts, conversation IDs, and subagent IDs. The published corpus does not contain prompts, source code, tool arguments, or tool results.',
+      'The proxy represents each input as session-scoped chained hashes in 64-token blocks. Matching block IDs preserve shared prefixes within a session. AIPerf replaces the blocks with deterministic synthetic coding and tool-use tokens before replay.',
+      'The client cannot see provider-side chat templates, proprietary tokenizers, server tools, encrypted reasoning content, or the exact token expansion of images and documents. Model-specific padding and deterministic placeholders approximate those fields. The placeholders and padding do not contain original prompts, code, or tool payloads.',
+    ],
+    datasetTitle: 'The v1.0 dataset',
+    datasetParagraphs: [
+      'The v1.0 release contains 393 sessions built on June 21, 2026. Each selected session has at least 20 requests, uses Claude Code 2.1.139 or newer, and has no more than 10 concurrent subagents. Processing removes exact duplicates, short classifier calls used for security monitoring or title generation, and requests whose reconstructed input exceeds 990k tokens.',
+      'The full variant retains contexts up to 1M tokens. The 256k variant removes requests above its cap while retaining the relative timing and subagent overlap of the requests that remain. Both use the WEKA trace format consumed by AIPerf.',
+    ],
+    replayTitle: 'From trace to replay graph',
+    replayParagraphs: [
+      'AIPerf converts each trace into a directed acyclic graph (DAG). Main-agent requests form a linear chain. Subagent requests form separate chains that spawn after an eligible parent request and join before the next dependent main-agent request. One-off auxiliary requests can run without a join edge.',
+      'The trace contains request timestamps and observed branch IDs, not the tool-level event that caused each branch. Replay therefore preserves recorded ordering, branch overlap, and inter-turn delays without claiming a provider-internal causal history.',
+    ],
+    measurementTitle: 'Concurrency and reported metrics',
+    measurementParagraphs: [
+      'Concurrency is the number of live agent clients. It is not a fixed HTTP request batch because one client can fan out into several subagent requests. The server can have more in-flight requests than the configured client concurrency.',
+      'AgentX runs closed loop: a client submits its next eligible request after dependencies complete. Faster systems progress farther through their sampled sessions during the same hour, so the exact request mix can vary, especially at low concurrency. Results should report throughput with time to first token (TTFT) and interactivity. A single latency value does not describe the run.',
+    ],
+    warmupTitle: 'Warmup, timing, and determinism',
+    warmupParagraphs: [
+      'A fixed seed selects each replay start uniformly from 25% to 75% of the recorded session duration. A max_tokens=1 primer materializes the active main-agent and subagent prefixes. Each replay lane then completes 10 additional warmup requests before the measurement barrier opens.',
+      'Only the following one-hour profiling window contributes reported metrics. The seed fixes session sampling, start points, and synthetic payloads. A unique cache-bust marker is added to each recycled play so unrelated replays cannot accumulate a shared prefix.',
+    ],
+    acceptanceTitle: 'Synthetic payloads and speculative decoding',
+    acceptanceParagraphs: [
+      'Synthetic tokens preserve input length and prefix structure, but their draft-token acceptance does not match natural model output. AgentX therefore uses a measured acceptance length from the coding category of SPEED-Bench for each model, speculative method, draft length, and thinking mode.',
+      'The serving engines expose a forced-acceptance control, and InferenceX records the selected values in versioned golden acceptance-length files. This separates serving-system performance from acceptance variation caused by the synthetic payload. AgentX does not evaluate model answer quality.',
+    ],
+    dramTitle: 'DRAM offload rules',
+    dramParagraphs: [
+      'KV-cache offload changes the capacity available to long-running sessions. Servers without a standardized DRAM configuration are capped at 3 TB. Standard GB200 NVL72, GB300 NVL72, and TPUv7 systems may use their installed capacity.',
+      'A benchmark configuration can access only the fraction of host DRAM proportional to its GPU allocation. This prevents a small GPU partition from claiming the memory budget of the entire server.',
+    ],
+    limitsTitle: 'Scope and reproducibility',
+    limits:
+      'The replay preserves client-visible request lengths, timing, branch structure, and KV-prefix reuse. It cannot reproduce provider-hidden transformations or the semantic content of the original session. Use the published corpus and locked scenario settings when comparing systems.',
+    sourcesTitle: 'Primary sources',
+    sources: [
+      { label: 'AgentX v1.0 full dataset', target: 'dataset-full' },
+      { label: 'AgentX v1.0 256k dataset', target: 'dataset-256k' },
+      { label: 'AIPerf WEKA trace replay guide', target: 'aiperf-weka' },
+      { label: 'SPEED-Bench paper', target: 'speed-bench' },
+      { label: 'SGLang forced acceptance control', target: 'sglang-pr' },
+      { label: 'TensorRT-LLM forced acceptance control', target: 'trtllm-pr' },
+      { label: 'vLLM synthetic acceptance control', target: 'vllm-pr' },
+      { label: 'ATOM synthetic acceptance control', target: 'atom-pr' },
+      { label: 'InferenceX golden acceptance-length values', target: 'golden-al' },
+    ],
+    figures: {
+      corpus: {
+        alt: 'Corpus dashboard showing 8,271 sessions, 3.41 million requests, 613.27 billion tokens, 99% cache-hit rate, and token-source shares.',
+        caption:
+          'Corpus snapshot used during dataset selection. The displayed cost is a list-price estimate for this snapshot; it is not a benchmark output.',
+      },
+      hashRatio: {
+        alt: 'Median reconstructed-hash-token to provider-token ratio by sequence length, overall and split by model.',
+        caption:
+          'Across 135,282 requests from 393 sessions, the median reconstructed/provider token ratio is 1.004. Shaded bands report p25–p75, not a per-request bound.',
+      },
+      traceRecord: {
+        alt: 'Annotated AgentX trace JSON showing session ID, 64-token hash blocks, request timing, input and output counts, and a subagent group.',
+        caption:
+          'An abridged WEKA record. Block IDs are local to one session; repeated IDs identify the prompt prefix shared by later requests.',
+      },
+      requestDistributions: {
+        alt: 'Log-scale distributions for inter-turn latency, input sequence length, and output sequence length in the v1.0 dataset.',
+        caption:
+          'Request distributions in the published v1.0 corpus. The median request has 142,016 input tokens and 444 output tokens.',
+      },
+      requestDistributions256k: {
+        alt: 'Log-scale inter-turn latency, input length, and output length distributions for the AgentX 256k dataset variant.',
+        caption:
+          'After requests above the 256k input cap are removed, the variant retains 68,266 requests. Median input length is 88,768 tokens; median output length is 376 tokens.',
+      },
+      subagentDistributions: {
+        alt: 'Distributions of subagent wall-clock duration and number of subagent groups per session.',
+        caption:
+          'The 175 sessions with subagents contain 1,697 groups. Median group duration is 2.27 minutes, and the median participating session has four groups.',
+      },
+      subagentDistributions256k: {
+        alt: 'Subagent group duration and groups-per-session distributions for the AgentX 256k dataset variant.',
+        caption:
+          'The 256k variant retains 1,697 subagent groups. Median duration is 2.27 minutes, and p95 duration is 18.5 minutes.',
+      },
+      replayLinear: {
+        alt: 'A linear four-request trace converted from recorded timing into a replay dependency chain.',
+        caption:
+          'A session without subagents becomes a linear dependency chain with recorded inter-turn delays.',
+      },
+      replaySingle: {
+        alt: 'A main-agent trace with one two-request subagent converted into a dependency graph with a join gate.',
+        caption:
+          'One subagent branch spawns after the first main request and joins before the next dependent main request.',
+      },
+      replayJoined: {
+        alt: 'Two subagent chains branching from a main agent and rejoining at one dependency gate.',
+        caption:
+          'Two subagent chains spawn after the same main request and share one join gate before the main chain resumes.',
+      },
+      replayTwo: {
+        alt: 'Two parallel subagents joining the main agent and a one-off auxiliary request that does not rejoin.',
+        caption:
+          'Parallel subagents share a join gate. The auxiliary request runs independently and has no join edge.',
+      },
+      replayFlatspawn: {
+        alt: 'Two subagent chains and one auxiliary request running in parallel before a three-way join.',
+        caption:
+          'A flat-spawn group runs two subagent chains and one auxiliary request in parallel, then joins all three before the main chain resumes.',
+      },
+      replaySidecars: {
+        alt: 'Two subagent chains and two plain sidecar requests running in parallel before one join gate.',
+        caption:
+          'Two plain sidecar requests without duration metadata run beside two subagent chains. Replay waits for all four at one join.',
+      },
+      replayMulti: {
+        alt: 'Four parallel subagents grouped by two join points, plus an auxiliary request, shown as a replay dependency graph.',
+        caption:
+          'Multiple subagent groups retain their separate join points, while the auxiliary branch remains detached.',
+      },
+      queueDepth: {
+        alt: 'Request queue depth over about one hour, separating running, waiting, and total requests for a concurrent AgentX replay.',
+        caption:
+          'A 50-client replay produces a changing number of running and waiting HTTP requests as session graphs fan out and join.',
+      },
+      pareto: {
+        alt: 'B200 vLLM MiniMax-M3 throughput per chip versus p90 interactivity curve across client concurrency levels.',
+        caption:
+          'Each labeled point is a client-concurrency setting. Higher concurrency raises throughput while reducing per-client interactivity.',
+      },
+      warmup: {
+        alt: 'Four replay trajectories showing seeded 25–75% warmup points and primer requests for active main-agent and subagent streams.',
+        caption:
+          'The fixed seed chooses t* within the shaded 25–75% interval. Primers establish the active prefix state before profiling.',
+      },
+      warmupCache: {
+        alt: 'Warmup prompt-token share shifting from cache misses toward HBM cache hits over roughly one minute.',
+        caption:
+          'Warmup begins with cache misses, then materializes enough prefix state for most prompt tokens to hit HBM.',
+      },
+      profilingCache: {
+        alt: 'One-hour profiling prompt-token share dominated by HBM cache hits after the initial seconds.',
+        caption:
+          'The measured hour starts after warmup. HBM cache hits dominate the prompt-token source throughout the window.',
+      },
+      acceptanceControls: {
+        alt: 'Merged SGLang, TensorRT-LLM, vLLM, and ATOM pull requests adding forced speculative-decoding acceptance controls.',
+        caption:
+          'Merged engine changes expose the controls required to apply the same acceptance assumption across serving stacks.',
+      },
+      goldenAl: {
+        alt: 'InferenceX golden acceptance-length YAML for DeepSeek-V4-Pro, listing values by speculative-token count and thinking mode.',
+        caption:
+          'A versioned golden acceptance-length file records the SPEED-Bench measurement used for each draft length and thinking mode.',
+      },
+    },
+  },
+  zh: {
+    eyebrow: 'Field guide / AI infrastructure',
+    title: 'AgentX 方法论',
+    lead: 'AgentX 将自愿采集的 Claude Code 代理 trace 转换为确定性的 AIPerf 工作负载。本页说明采集元数据如何生成回放，以及有效基准测试结果所遵循的控制规则。',
+    facts: [
+      { value: '393', label: '个公开会话' },
+      { value: '64', label: '每个 hash block 的 token 数' },
+      { value: '1 小时', label: 'profiling 窗口' },
+      { value: '25–75%', label: '固定 seed 起点范围' },
+    ],
+    collectionTitle: '数据采集',
+    collectionParagraphs: [
+      '参与者主动启用 HTTP 代理后，代理会记录请求到达与完成时间、input 和 output token 数、conversation ID 与 subagent ID。公开语料不包含 prompt、源代码、tool argument 或 tool result。',
+      '代理以 64-token block 为单位，将每段 input 表示为会话内串联 hash。同一会话中相同的 block ID 会保留共享 prefix。回放前，AIPerf 会使用确定性的合成编码与 tool-use token 替换这些 block。',
+      '客户端无法看到服务端 chat template、专有 tokenizer、服务端 tool、加密的 reasoning 内容，也无法精确得知图片和文档最终展开成多少 token。AgentX 使用针对模型校准的 padding 和确定性 placeholder 处理这些字段；其中不包含原始 prompt、代码或 tool payload。',
+    ],
+    datasetTitle: 'v1.0 数据集',
+    datasetParagraphs: [
+      'v1.0 于 2026 年 6 月 21 日构建，共包含 393 个会话。每个入选会话至少有 20 个请求，Claude Code 版本不低于 2.1.139，并且同时运行的 subagent 不超过 10 个。处理流程会移除完全重复的请求、用于安全监控或标题生成的短 classifier 调用，以及重建后 input 超过 990k token 的请求。',
+      'full 变体保留最高 1M token 的上下文。256k 变体会移除超过上限的请求，同时保留其余请求的相对时间与 subagent 重叠关系。两者均采用 AIPerf 可读取的 WEKA trace 格式。',
+    ],
+    replayTitle: '从 trace 到回放图',
+    replayParagraphs: [
+      'AIPerf 将每条 trace 转换为有向无环图（DAG）。主 agent 请求形成线性链；subagent 请求形成独立链，在符合条件的父请求完成后启动，并在下一个依赖它的主 agent 请求前汇合。一次性辅助请求可以在没有 join edge 的情况下运行。',
+      'Trace 记录请求时间戳和可观测的分支 ID，但不记录触发分支的 tool 级事件。回放会保留请求顺序、分支重叠与轮次间延迟，但不会推断服务端内部因果关系。',
+    ],
+    measurementTitle: 'Concurrency 与结果指标',
+    measurementParagraphs: [
+      'Concurrency 表示同时运行的 agent 客户端数量，不是固定的 HTTP request batch。一个客户端可能展开为多个 subagent 请求，因此服务器上的瞬时请求数可以高于配置的客户端 concurrency。',
+      'AgentX 采用 closed-loop 模式：依赖满足后，客户端才提交下一个可执行请求。更快的系统会在同一小时内推进到采样会话的更后位置，因此实际请求组合可能略有不同，低并发时尤为明显。报告结果时应同时给出吞吐量、首 token 延迟（TTFT）和 interactivity；单一 latency 值无法描述完整运行。',
+    ],
+    warmupTitle: 'Warmup、计时与确定性',
+    warmupParagraphs: [
+      '固定 seed 会在每段会话记录时长的 25% 至 75% 区间内均匀选择回放起点。随后使用 max_tokens=1 的 primer 建立当前主 agent 与 subagent 的 prefix，再让每条回放 lane 完成 10 个额外 warmup 请求，最后才开启测量 barrier。',
+      '对外指标只统计随后一小时的 profiling 窗口。Seed 会固定会话采样、起点和合成 payload。每次循环使用唯一的 cache-bust 标记，避免无关回放逐步形成共享 prefix。',
+    ],
+    acceptanceTitle: '合成 payload 与 speculative decoding',
+    acceptanceParagraphs: [
+      '合成 token 会保留 input 长度和 prefix 结构，但其 draft token 接受情况与自然模型输出不同。因此，AgentX 会针对每组模型、speculative 方法、draft length 与 thinking mode，使用 SPEED-Bench 编码类别测得的 acceptance length。',
+      '推理引擎提供强制 acceptance 控制，InferenceX 则把选定值记录在带版本的 golden acceptance-length 文件中。这样可以把推理系统性能与合成 payload 引起的 acceptance 波动分开。AgentX 不评估模型回答质量。',
+    ],
+    dramTitle: 'DRAM offload 规则',
+    dramParagraphs: [
+      'KV cache offload 会改变长会话可用的容量。没有标准化 DRAM 配置的服务器上限为 3 TB；GB200 NVL72、GB300 NVL72 和 TPUv7 等标准化系统可使用实际装机容量。',
+      '每种基准测试配置只能按其 GPU 占比使用对应的 host DRAM，避免较小的 GPU 分区占用整台服务器的内存预算。',
+    ],
+    limitsTitle: '适用范围与复现',
+    limits:
+      '回放会保留客户端可见的请求长度、时间关系、分支结构和 KV prefix 复用，但无法复现服务端隐藏转换或原始会话的语义内容。比较不同系统时，应使用公开语料和锁定的场景配置。',
+    sourcesTitle: '主要来源',
+    sources: [
+      { label: 'AgentX v1.0 full 数据集', target: 'dataset-full' },
+      { label: 'AgentX v1.0 256k 数据集', target: 'dataset-256k' },
+      { label: 'AIPerf WEKA trace 回放指南', target: 'aiperf-weka' },
+      { label: 'SPEED-Bench 论文', target: 'speed-bench' },
+      { label: 'SGLang 强制 acceptance 控制', target: 'sglang-pr' },
+      { label: 'TensorRT-LLM 强制 acceptance 控制', target: 'trtllm-pr' },
+      { label: 'vLLM 合成 acceptance 控制', target: 'vllm-pr' },
+      { label: 'ATOM 合成 acceptance 控制', target: 'atom-pr' },
+      { label: 'InferenceX golden acceptance-length 数值', target: 'golden-al' },
+    ],
+    figures: {
+      corpus: {
+        alt: '语料仪表板显示 8,271 个会话、341 万个请求、6132.7 亿 token、99% cache-hit rate 及 token 来源占比。',
+        caption:
+          '筛选数据集时使用的语料快照。图中的费用按该快照和公开 API 标价估算，并非基准测试输出指标。',
+      },
+      hashRatio: {
+        alt: '按序列长度展示重建 hash token 与服务商 token 数的中位比值，并分别给出整体和各模型结果。',
+        caption:
+          '393 个会话的 135,282 个请求中，重建 token 数与服务商 token 数之比的中位数为 1.004。阴影表示 p25–p75 区间，不代表每个请求都有固定误差上限。',
+      },
+      traceRecord: {
+        alt: '带注释的 AgentX trace JSON，包含 session ID、64-token hash block、请求时间、input 和 output 数量及一个 subagent group。',
+        caption:
+          '一条删节后的 WEKA 记录。Block ID 只在单个会话内有效；重复 ID 表示后续请求共享的 prompt prefix。',
+      },
+      requestDistributions: {
+        alt: 'v1.0 数据集中轮次间延迟、input sequence length 与 output sequence length 的对数分布。',
+        caption:
+          'v1.0 公开语料的请求分布。单请求 input token 中位数为 142,016，output token 中位数为 444。',
+      },
+      requestDistributions256k: {
+        alt: 'AgentX 256k 数据集变体中轮次间延迟、input 长度与 output 长度的对数分布。',
+        caption:
+          '移除超过 256k input 上限的请求后，该变体保留 68,266 个请求。input token 中位数为 88,768，output token 中位数为 376。',
+      },
+      subagentDistributions: {
+        alt: 'Subagent wall-clock duration 以及每个会话中 subagent group 数量的分布。',
+        caption:
+          '175 个包含 subagent 的会话共有 1,697 个 group。Group 时长中位数为 2.27 分钟；这些会话的 group 数中位数为 4。',
+      },
+      subagentDistributions256k: {
+        alt: 'AgentX 256k 数据集变体中的 subagent group 时长与每会话 group 数分布。',
+        caption:
+          '256k 变体保留 1,697 个 subagent group。时长中位数为 2.27 分钟，p95 为 18.5 分钟。',
+      },
+      replayLinear: {
+        alt: '线性四请求 trace 根据记录时间转换为回放依赖链。',
+        caption: '不含 subagent 的会话会转换为线性依赖链，并保留记录中的轮次间延迟。',
+      },
+      replaySingle: {
+        alt: '一个包含双请求 subagent 的主 agent trace 被转换为带 join gate 的依赖图。',
+        caption:
+          'Subagent 分支在第一个主请求后启动，并在下一个依赖它的主请求前通过 join gate 汇合。',
+      },
+      replayJoined: {
+        alt: '两条 subagent 链从主 agent 分支，并在同一个依赖 gate 汇合。',
+        caption: '两条 subagent 链在同一个主请求完成后启动，并在主链继续前共用一个 join gate。',
+      },
+      replayTwo: {
+        alt: '两个并行 subagent 汇合到主 agent，另有一个不会汇合的一次性辅助请求。',
+        caption: '并行 subagent 共享 join gate；辅助请求独立运行，不带 join edge。',
+      },
+      replayFlatspawn: {
+        alt: '两条 subagent 链与一个辅助请求并行运行，随后通过三路 join 汇合。',
+        caption:
+          'Flat-spawn group 并行运行两条 subagent 链和一个辅助请求，三者汇合后主链才继续执行。',
+      },
+      replaySidecars: {
+        alt: '两条 subagent 链与两个普通 sidecar 请求并行运行，随后汇合到同一个 join gate。',
+        caption:
+          '两个不带 duration 元数据的普通 sidecar 请求与两条 subagent 链并行运行；回放会在同一个 join 等待四者完成。',
+      },
+      replayMulti: {
+        alt: '四个并行 subagent 按两个 join point 分组，并带有辅助请求的回放依赖图。',
+        caption: '多个 subagent group 保留各自的 join point，辅助分支仍保持独立。',
+      },
+      queueDepth: {
+        alt: '约一小时内的请求队列深度，分别显示并发 AgentX 回放中的 running、waiting 与 total request。',
+        caption:
+          '在 50 个客户端的回放中，会话图持续展开和汇合，running 与 waiting HTTP 请求数也随之变化。',
+      },
+      pareto: {
+        alt: 'B200 vLLM MiniMax-M3 在不同客户端 concurrency 下的单芯片吞吐量与 p90 interactivity 曲线。',
+        caption:
+          '每个标记点对应一种客户端 concurrency。提高 concurrency 会增加吞吐量，同时降低每个客户端的 interactivity。',
+      },
+      warmup: {
+        alt: '四条回放轨迹，显示固定 seed 的 25–75% warmup 起点，以及主 agent 和 subagent 活跃流的 primer 请求。',
+        caption:
+          '固定 seed 在阴影所示的 25–75% 区间选择 t*；primer 会在 profiling 前建立活跃 prefix 状态。',
+      },
+      warmupCache: {
+        alt: '约一分钟内，warmup prompt token 来源从 cache miss 转向 HBM cache hit。',
+        caption:
+          'Warmup 初期以 cache miss 为主，随后建立 prefix 状态，使大部分 prompt token 命中 HBM。',
+      },
+      profilingCache: {
+        alt: '一小时 profiling 中，初始几秒后 prompt token 来源以 HBM cache hit 为主。',
+        caption:
+          '测量窗口在 warmup 后开始，整个 profiling 小时内 prompt token 来源主要为 HBM cache hit。',
+      },
+      acceptanceControls: {
+        alt: 'SGLang、TensorRT-LLM、vLLM 与 ATOM 中加入强制 speculative-decoding acceptance 控制的已合并 PR。',
+        caption:
+          '这些已合并的推理引擎变更提供所需控制，使不同 serving stack 可以采用相同的 acceptance 假设。',
+      },
+      goldenAl: {
+        alt: 'InferenceX DeepSeek-V4-Pro golden acceptance-length YAML，按 speculative-token 数量和 thinking mode 列出数值。',
+        caption:
+          '带版本的 golden acceptance-length 文件记录每种 draft length 与 thinking mode 所使用的 SPEED-Bench 测量值。',
+      },
+    },
+  },
+} as const;
+
+const SOURCE_HREFS = {
+  'dataset-full': 'https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126',
+  'dataset-256k': 'https://huggingface.co/datasets/semianalysisai/cc-traces-weka-062126-256k',
+  'aiperf-weka': 'https://github.com/ai-dynamo/aiperf/blob/main/docs/tutorials/weka-trace.md',
+  'speed-bench': 'https://arxiv.org/abs/2604.09557',
+  'sglang-pr': 'https://github.com/sgl-project/sglang/pull/10771',
+  'trtllm-pr': 'https://github.com/NVIDIA/TensorRT-LLM/pull/9371',
+  'vllm-pr': 'https://github.com/vllm-project/vllm/pull/40662',
+  'atom-pr': 'https://github.com/ROCm/ATOM/pull/1850',
+  'golden-al':
+    'https://github.com/SemiAnalysisAI/InferenceX/blob/main/golden_al_distribution/dsv4_mtp.yaml',
+} as const;
+
+type FigureKey = keyof typeof FIGURE_ASSETS;
+interface FigureCopy {
+  alt: string;
+  caption: string;
+}
+
+function MethodFigure({
+  figure,
+  copy,
+  locale,
+}: {
+  figure: FigureKey;
+  copy: FigureCopy;
+  locale: Locale;
+}) {
+  const asset = FIGURE_ASSETS[figure];
+  return (
+    <AgentXMethodologyLink
+      href={asset.src}
+      analyticsEvent="agentx_methodology_figure_opened"
+      analyticsTarget={figure}
+      target="_blank"
+      rel="noopener noreferrer"
+      prefetch={false}
+      aria-label={`${copy.alt} ${
+        locale === 'zh' ? '打开原始分辨率图片。' : 'Open the full-resolution image.'
+      }`}
+      className="group my-7 block rounded-xl focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2"
+    >
+      <figure
+        className="overflow-hidden rounded-xl border border-border/70 bg-card transition-colors group-hover:border-primary/50"
+        data-testid={`agentx-methodology-figure-${figure}`}
+      >
+        <Image
+          src={asset.src}
+          alt={copy.alt}
+          width={asset.width}
+          height={asset.height}
+          sizes="(max-width: 768px) 100vw, 1152px"
+          quality={100}
+          className="h-auto w-full"
+        />
+        <figcaption className="border-t border-border/70 px-4 py-3 text-sm leading-6 text-muted-foreground">
+          <span className="block">{copy.caption}</span>
+          <span className="mt-1 inline-block font-medium text-foreground underline decoration-border underline-offset-4 transition-colors group-hover:text-primary group-hover:decoration-primary">
+            {locale === 'zh' ? '查看原始分辨率图片' : 'View full-resolution image'} ↗
+          </span>
+        </figcaption>
+      </figure>
+    </AgentXMethodologyLink>
+  );
+}
+
+function Paragraphs({ items }: { items: readonly string[] }) {
+  return (
+    <div className="space-y-4 text-base leading-7 text-muted-foreground">
+      {items.map((item) => (
+        <p key={item}>{item}</p>
+      ))}
+    </div>
+  );
+}
+
+function SectionHeading({ id, children }: { id: string; children: React.ReactNode }) {
+  return (
+    <h2 id={id} className="mb-4 scroll-mt-24 text-2xl font-semibold tracking-tight text-foreground">
+      {children}
+    </h2>
+  );
+}
+
+export function AgentXMethodologyArticle({ locale }: { locale: Locale }) {
+  const t = CONTENT[locale];
+  const prefix = locale === 'zh' ? '/zh' : '';
+
+  return (
+    <article data-testid="agentx-methodology-article">
+      <header className="border-b border-border/70 pb-8">
+        <AgentXMethodologyLink
+          href={`${prefix}/agentx`}
+          analyticsEvent="agentx_methodology_returned"
+          analyticsTarget="overview"
+          className="text-sm text-muted-foreground transition-colors hover:text-foreground"
+        >
+          {locale === 'zh' ? '← 返回 AgentX' : '← Back to AgentX'}
+        </AgentXMethodologyLink>
+        <p className="mt-8 font-mono text-xs font-semibold tracking-eyebrow-wide text-brand uppercase">
+          {t.eyebrow}
+        </p>
+        <h1 className="mt-3 text-4xl font-semibold tracking-tight text-foreground sm:text-5xl">
+          {t.title}
+        </h1>
+        <p className="mt-5 max-w-3xl text-lg leading-8 text-muted-foreground">{t.lead}</p>
+        <dl className="mt-8 grid grid-cols-2 gap-px overflow-hidden rounded-xl border border-border/70 bg-border/70 lg:grid-cols-4">
+          {t.facts.map((fact) => (
+            <div key={fact.label} className="flex flex-col bg-card px-4 py-5">
+              <dt className="order-2 mt-1 text-sm leading-5 text-muted-foreground">{fact.label}</dt>
+              <dd className="order-1 font-mono text-xl font-semibold tabular-nums text-foreground">
+                {fact.value}
+              </dd>
+            </div>
+          ))}
+        </dl>
+      </header>
+
+      <div className="py-10">
+        <section aria-labelledby="collection">
+          <SectionHeading id="collection">{t.collectionTitle}</SectionHeading>
+          <Paragraphs items={t.collectionParagraphs} />
+          <MethodFigure figure="corpus" copy={t.figures.corpus} locale={locale} />
+          <MethodFigure figure="hashRatio" copy={t.figures.hashRatio} locale={locale} />
+        </section>
+
+        <section aria-labelledby="dataset" className="mt-14">
+          <SectionHeading id="dataset">{t.datasetTitle}</SectionHeading>
+          <Paragraphs items={t.datasetParagraphs} />
+          <MethodFigure figure="traceRecord" copy={t.figures.traceRecord} locale={locale} />
+          <MethodFigure
+            figure="requestDistributions"
+            copy={t.figures.requestDistributions}
+            locale={locale}
+          />
+          <MethodFigure
+            figure="requestDistributions256k"
+            copy={t.figures.requestDistributions256k}
+            locale={locale}
+          />
+          <MethodFigure
+            figure="subagentDistributions"
+            copy={t.figures.subagentDistributions}
+            locale={locale}
+          />
+          <MethodFigure
+            figure="subagentDistributions256k"
+            copy={t.figures.subagentDistributions256k}
+            locale={locale}
+          />
+        </section>
+
+        <section aria-labelledby="replay" className="mt-14">
+          <SectionHeading id="replay">{t.replayTitle}</SectionHeading>
+          <Paragraphs items={t.replayParagraphs} />
+          <MethodFigure figure="replayLinear" copy={t.figures.replayLinear} locale={locale} />
+          <MethodFigure figure="replaySingle" copy={t.figures.replaySingle} locale={locale} />
+          <MethodFigure figure="replayJoined" copy={t.figures.replayJoined} locale={locale} />
+          <MethodFigure figure="replayTwo" copy={t.figures.replayTwo} locale={locale} />
+          <MethodFigure figure="replayFlatspawn" copy={t.figures.replayFlatspawn} locale={locale} />
+          <MethodFigure figure="replaySidecars" copy={t.figures.replaySidecars} locale={locale} />
+          <MethodFigure figure="replayMulti" copy={t.figures.replayMulti} locale={locale} />
+        </section>
+
+        <section aria-labelledby="measurement" className="mt-14">
+          <SectionHeading id="measurement">{t.measurementTitle}</SectionHeading>
+          <Paragraphs items={t.measurementParagraphs} />
+          <MethodFigure figure="queueDepth" copy={t.figures.queueDepth} locale={locale} />
+          <MethodFigure figure="pareto" copy={t.figures.pareto} locale={locale} />
+        </section>
+
+        <section aria-labelledby="warmup" className="mt-14">
+          <SectionHeading id="warmup">{t.warmupTitle}</SectionHeading>
+          <Paragraphs items={t.warmupParagraphs} />
+          <MethodFigure figure="warmup" copy={t.figures.warmup} locale={locale} />
+          <div className="grid gap-6 lg:grid-cols-2">
+            <MethodFigure figure="warmupCache" copy={t.figures.warmupCache} locale={locale} />
+            <MethodFigure figure="profilingCache" copy={t.figures.profilingCache} locale={locale} />
+          </div>
+        </section>
+
+        <section aria-labelledby="acceptance" className="mt-14">
+          <SectionHeading id="acceptance">{t.acceptanceTitle}</SectionHeading>
+          <Paragraphs items={t.acceptanceParagraphs} />
+          <MethodFigure
+            figure="acceptanceControls"
+            copy={t.figures.acceptanceControls}
+            locale={locale}
+          />
+          <div className="mx-auto max-w-xl">
+            <MethodFigure figure="goldenAl" copy={t.figures.goldenAl} locale={locale} />
+          </div>
+        </section>
+
+        <section aria-labelledby="dram" className="mt-14">
+          <SectionHeading id="dram">{t.dramTitle}</SectionHeading>
+          <Paragraphs items={t.dramParagraphs} />
+        </section>
+
+        <section aria-labelledby="scope" className="mt-14 border-t border-border/70 pt-10">
+          <SectionHeading id="scope">{t.limitsTitle}</SectionHeading>
+          <p className="text-base leading-7 text-muted-foreground">{t.limits}</p>
+          <h3 className="mt-8 text-base font-semibold text-foreground">{t.sourcesTitle}</h3>
+          <ul className="mt-3 grid gap-x-8 gap-y-2 sm:grid-cols-2">
+            {t.sources.map((source) => (
+              <li key={source.target}>
+                <AgentXMethodologyLink
+                  href={SOURCE_HREFS[source.target]}
+                  analyticsEvent="agentx_methodology_source_opened"
+                  analyticsTarget={source.target}
+                  target="_blank"
+                  rel="noopener noreferrer"
+                  className="text-sm leading-6 text-primary underline decoration-primary/40 underline-offset-4 hover:decoration-primary"
+                >
+                  {source.label} ↗
+                </AgentXMethodologyLink>
+              </li>
+            ))}
+          </ul>
+        </section>
+      </div>
+    </article>
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-methodology-link.tsx b/packages/app/src/components/datasets/agentx-methodology-link.tsx
new file mode 100644
index 000000000..a6094eb32
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-methodology-link.tsx
@@ -0,0 +1,36 @@
+'use client';
+
+import Link from 'next/link';
+
+import { track } from '@/lib/analytics';
+
+interface AgentXMethodologyLinkProps extends Omit<React.ComponentProps<typeof Link>, 'href'> {
+  href: string;
+  analyticsEvent:
+    | 'agentx_methodology_opened'
+    | 'agentx_methodology_returned'
+    | 'agentx_methodology_figure_opened'
+    | 'agentx_methodology_source_opened'
+    | 'agentx_results_opened';
+  analyticsTarget?: string;
+}
+
+export function AgentXMethodologyLink({
+  href,
+  analyticsEvent,
+  analyticsTarget,
+  onClick,
+  ...props
+}: AgentXMethodologyLinkProps) {
+  return (
+    <Link
+      {...props}
+      href={href}
+      onClick={(event) => {
+        onClick?.(event);
+        if (event.defaultPrevented) return;
+        track(analyticsEvent, analyticsTarget ? { target: analyticsTarget } : undefined);
+      }}
+    />
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-methodology.tsx b/packages/app/src/components/datasets/agentx-methodology.tsx
new file mode 100644
index 000000000..2c846afcc
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-methodology.tsx
@@ -0,0 +1,295 @@
+import { Card } from '@/components/ui/card';
+
+import { AgentXMethodologyLink } from './agentx-methodology-link';
+import { AgentXOptimizationsCallout } from './agentx-optimizations-callout';
+import { AgentXTelemetryCallout } from './agentx-telemetry-callout';
+
+type Locale = 'en' | 'zh';
+
+const CONTENT = {
+  en: {
+    eyebrow: 'AgentX v1.0 methodology',
+    title: 'AgentX Benchmark Datasets',
+    intro:
+      'AgentX derives replay workloads from opt-in Claude Code sessions. The published traces remove prompt, code, and tool payloads while retaining request lengths, prefix reuse, subagent branches, and timing.',
+    processTitle: 'How an AgentX run is built',
+    steps: [
+      {
+        title: 'Capture',
+        description:
+          'An opt-in HTTP proxy records request and response timing, token counts, conversation IDs, and subagent IDs as sessions run.',
+      },
+      {
+        title: 'Transform',
+        description:
+          'Original prompts, source code, tool arguments, and tool results are removed. Inputs become session-scoped chained hashes in 64-token blocks, preserving matching prefixes without revealing content.',
+      },
+      {
+        title: 'Reconstruct',
+        description:
+          'AIPerf fills those blocks with deterministic synthetic coding and tool-use tokens, then rebuilds each session as a directed acyclic graph (DAG) of main-agent turns, parallel subagents, auxiliary requests, and inter-turn tool time.',
+      },
+      {
+        title: 'Replay and measure',
+        description:
+          'A seeded warmup establishes cache state before each configuration is profiled for one hour across a sweep of concurrent clients. Per-replay cache-bust markers stop unrelated sessions from sharing prefixes.',
+      },
+    ],
+    datasetTitle: 'What is in the v1.0 dataset',
+    datasetIntro:
+      'AgentX v1.0 uses 393 Claude Code sessions selected from an internal, opt-in trace corpus. Eligible sessions contain at least 20 requests, use Claude Code 2.1.139 or newer, and have no more than 10 concurrent subagents. Processing removes duplicate requests, client-specific security-monitor and title-generation calls, and reconstructed inputs above 990k tokens.',
+    profileLabel: 'v1.0 trace profile',
+    stats: [
+      { value: '393', label: 'sessions' },
+      { value: '142k', label: 'median input tokens / request' },
+      { value: '444', label: 'median output tokens / request' },
+      { value: '44%', label: 'sessions with subagents' },
+    ],
+    controlsTitle: 'Replay controls',
+    controls: [
+      {
+        title: 'Steady-state start',
+        description:
+          'A fixed seed selects a point 25–75% through each conversation. Primer requests and 10 additional warmup requests per replay lane materialize KV cache before profiling begins.',
+      },
+      {
+        title: 'Deterministic replay',
+        description:
+          'The seed fixes conversation sampling, starting points, and synthetic content. Reported metrics cover only the one-hour profiling window, not warmup.',
+      },
+      {
+        title: 'Speculative decoding',
+        description:
+          'Because synthetic tokens can distort draft-token acceptance, AgentX forces an acceptance length measured with SPEED-Bench for each model, speculator, draft length, and thinking-mode combination.',
+      },
+      {
+        title: 'DRAM offload',
+        description:
+          'Servers without standardized DRAM are capped at 3 TB. Standard GB200 NVL72, GB300 NVL72, and TPUv7 systems use installed capacity, and every configuration may access only the share proportional to its GPU allocation.',
+      },
+    ],
+    readingTitle: 'How to read AgentX results',
+    reading:
+      'Concurrency means concurrent agent clients, not a fixed request batch. AgentX is closed loop, so faster configurations complete more requests and can encounter a slightly different workload mix. This variation is most visible at low concurrency. Report throughput with time to first token (TTFT) and interactivity; a single latency statistic does not describe the full run.',
+    qualityNote:
+      'AgentX measures serving-system performance. Its synthetic payloads do not support model-quality evaluation.',
+    limits:
+      'The client cannot observe provider-side chat templates, proprietary tokenizers, server tools, encrypted reasoning content, or the exact token expansion of images and documents. AgentX uses deterministic placeholders and model-specific padding for those inputs. The resulting traces reproduce request lengths, timing, conversation topology, and KV-reuse patterns. They do not contain the original conversations.',
+    variantsLabel: 'Dataset variants',
+    variants: [
+      {
+        title: 'full',
+        description: 'The complete AgentX v1.0 replay set, including contexts up to 1M tokens.',
+      },
+      {
+        title: '256k',
+        description:
+          'A context-limited variant for models and inference engines configured with a maximum context of 256k tokens.',
+      },
+    ],
+    methodologyCta: 'Read the full methodology',
+    resultsCta: 'View AgentX Performance Results',
+  },
+  zh: {
+    eyebrow: 'AgentX v1.0 测试方法',
+    title: 'AgentX 基准测试数据集',
+    intro:
+      'AgentX 根据参与者自愿提供的 Claude Code 会话构建回放工作负载。公开 trace 会移除 prompt、源代码和 tool payload，仅保留请求长度、prefix 复用关系、subagent 分支和时间信息。',
+    processTitle: 'AgentX 如何构建一次回放',
+    steps: [
+      {
+        title: '采集',
+        description:
+          '参与者自愿启用 HTTP 代理后，代理会在会话期间记录请求与响应的时间信息、token 数量、会话 ID 和 subagent ID。',
+      },
+      {
+        title: '转换',
+        description:
+          '原始 prompt、源代码、tool argument 和 tool result 均会移除。Input 按 64-token block 转换为会话内串联 hash，在不暴露内容的前提下保留相同 prefix。',
+      },
+      {
+        title: '重建',
+        description:
+          'AIPerf 用确定性生成的 coding token 和 tool-use token 填充这些 block，再将每个会话重建为有向无环图（DAG）。图中包含 main agent 轮次、并行 subagent、辅助请求，以及轮次间的 tool 执行时间。',
+      },
+      {
+        title: '回放与测量',
+        description:
+          '每种配置先用固定 seed 执行 warmup，建立 cache 状态；随后在不同并发客户端数量下进行一小时 profiling。每次回放都会加入独立的 cache-bust 标记，避免无关会话共享 prefix。',
+      },
+    ],
+    datasetTitle: 'AgentX v1.0 数据集概况',
+    datasetIntro:
+      'AgentX v1.0 数据集包含从内部自愿采集的 trace 中筛选出的 393 个 Claude Code 会话。入选会话需至少包含 20 个请求，Claude Code 版本不低于 2.1.139，同时运行的 subagent 不超过 10 个。处理过程中还会移除重复请求、客户端特有的安全监控和标题生成请求，以及重建后 input 长度超过 990k token 的请求。',
+    profileLabel: 'v1.0 trace 概况',
+    stats: [
+      { value: '393', label: '会话数' },
+      { value: '142k', label: '单请求 input token 数中位数' },
+      { value: '444', label: '单请求 output token 数中位数' },
+      { value: '44%', label: '含 subagent 的会话' },
+    ],
+    controlsTitle: '回放控制',
+    controls: [
+      {
+        title: '从稳态开始测量',
+        description:
+          '固定 seed 会从每段会话 25%–75% 的区间内选择起点。正式 profiling 前，primer 请求和每条回放通道额外执行的 10 个 warmup 请求会先建立 KV cache。',
+      },
+      {
+        title: '确定性回放',
+        description:
+          '固定 seed 决定会话采样、起点和合成内容。对外报告的指标仅覆盖一小时 profiling 窗口，不包含 warmup。',
+      },
+      {
+        title: '投机解码',
+        description:
+          '合成 token 可能影响 draft token 的接受率，因此 AgentX 会按 model、speculator、draft length 和 thinking mode 的组合，采用 SPEED-Bench 的实测结果固定 acceptance length。',
+      },
+      {
+        title: 'DRAM offload',
+        description:
+          '对于没有标准 DRAM 配置的服务器，可用 DRAM 上限为 3 TB。GB200 NVL72、GB300 NVL72 和 TPUv7 标准系统按实际装机容量计算；每种配置只能使用与其 GPU 分配比例相对应的 DRAM。',
+      },
+    ],
+    readingTitle: '如何解读 AgentX 结果',
+    reading:
+      '这里的 concurrency 指同时运行的 agent 客户端数量，而不是固定大小的 request batch。AgentX 采用 closed-loop 模式，因此速度更快的配置会完成更多请求，实际遇到的工作负载组合可能略有差异；低并发时这种差异最明显。报告吞吐量时，应同时给出首 token 延迟（TTFT）和 interactivity；单看一个 latency 指标不足以反映整次运行。',
+    qualityNote: 'AgentX 衡量推理系统性能。合成 payload 不能用于模型质量评估。',
+    limits:
+      '客户端无法观测服务端的 chat template、专有 tokenizer、服务端 tool、加密 reasoning 内容，也无法准确获知图片和文档展开后的 token 数量。AgentX 对这些输入使用确定性 placeholder 和针对不同模型的 padding。重建后的 trace 会复现请求长度、时间关系、会话拓扑和 KV 复用模式，但不包含原始会话内容。',
+    variantsLabel: '数据集变体',
+    variants: [
+      {
+        title: 'full',
+        description: '完整的 AgentX v1.0 回放集，包含最长 1M token 的上下文。',
+      },
+      {
+        title: '256k',
+        description: '适用于最大上下文配置为 256k token 的模型和推理引擎。',
+      },
+    ],
+    methodologyCta: '查看完整测试方法',
+    resultsCta: '查看 AgentX 性能结果',
+  },
+} as const;
+
+export function AgentXMethodology({ locale }: { locale: Locale }) {
+  const t = CONTENT[locale];
+
+  return (
+    <Card className="overflow-hidden p-0" data-testid="agentx-methodology">
+      <header className="border-b border-border/60 bg-muted/20 px-5 py-5 sm:px-6">
+        <p className="mb-2 font-mono text-2xs font-medium uppercase tracking-eyebrow text-primary">
+          {t.eyebrow}
+        </p>
+        <h1 className="text-xl font-semibold text-foreground">{t.title}</h1>
+        <p className="mt-3 max-w-4xl text-sm leading-6 text-muted-foreground">{t.intro}</p>
+        <AgentXMethodologyLink
+          href={locale === 'zh' ? '/zh/inference' : '/inference'}
+          analyticsEvent="agentx_results_opened"
+          analyticsTarget="methodology-header"
+          data-testid="agentx-results-cta"
+          className="mt-4 inline-flex min-h-11 items-center rounded-md bg-primary px-4 py-2 text-sm font-semibold text-primary-foreground transition-colors hover:bg-primary/90"
+        >
+          {t.resultsCta} →
+        </AgentXMethodologyLink>
+      </header>
+
+      <div className="space-y-8 px-5 py-6 sm:px-6">
+        <AgentXOptimizationsCallout locale={locale} />
+
+        <AgentXTelemetryCallout locale={locale} />
+
+        <section aria-labelledby="agentx-process-title">
+          <h2 id="agentx-process-title" className="mb-3 text-base font-semibold text-foreground">
+            {t.processTitle}
+          </h2>
+          <ol className="grid gap-px overflow-hidden rounded-lg border border-border/70 bg-border/70 sm:grid-cols-2">
+            {t.steps.map((step, index) => (
+              <li key={step.title} className="bg-card p-4" data-testid="agentx-methodology-step">
+                <div className="mb-2 flex items-baseline gap-2">
+                  <span className="font-mono text-2xs text-primary">
+                    {String(index + 1).padStart(2, '0')}
+                  </span>
+                  <h3 className="text-sm font-semibold text-foreground">{step.title}</h3>
+                </div>
+                <p className="text-sm leading-6 text-muted-foreground">{step.description}</p>
+              </li>
+            ))}
+          </ol>
+        </section>
+
+        <section aria-labelledby="agentx-dataset-title">
+          <h2 id="agentx-dataset-title" className="mb-2 text-base font-semibold text-foreground">
+            {t.datasetTitle}
+          </h2>
+          <p className="text-sm leading-6 text-muted-foreground">{t.datasetIntro}</p>
+          <div className="mt-4 border-y border-border/70 py-4">
+            <p className="mb-3 font-mono text-xs font-semibold tracking-eyebrow-wide text-brand uppercase">
+              {t.profileLabel}
+            </p>
+            <dl className="grid grid-cols-2 gap-x-5 gap-y-4 lg:grid-cols-4">
+              {t.stats.map((stat) => (
+                <div key={stat.label} className="flex flex-col">
+                  <dt className="order-2 mt-0.5 text-xs leading-5 text-muted-foreground">
+                    {stat.label}
+                  </dt>
+                  <dd className="order-1 font-mono text-lg font-semibold tabular-nums text-foreground">
+                    {stat.value}
+                  </dd>
+                </div>
+              ))}
+            </dl>
+          </div>
+          <ul
+            aria-label={t.variantsLabel}
+            className="mt-4 space-y-1.5 text-sm leading-6 text-muted-foreground"
+          >
+            {t.variants.map((variant) => (
+              <li key={variant.title}>
+                <strong className="font-semibold text-foreground">{variant.title}</strong>:{' '}
+                {variant.description}
+              </li>
+            ))}
+          </ul>
+        </section>
+
+        <section aria-labelledby="agentx-controls-title">
+          <h2 id="agentx-controls-title" className="mb-3 text-base font-semibold text-foreground">
+            {t.controlsTitle}
+          </h2>
+          <dl className="grid gap-x-8 gap-y-4 md:grid-cols-2">
+            {t.controls.map((control) => (
+              <div key={control.title} className="border-l border-border pl-3">
+                <dt className="text-sm font-semibold text-foreground">{control.title}</dt>
+                <dd className="mt-1 text-sm leading-6 text-muted-foreground">
+                  {control.description}
+                </dd>
+              </div>
+            ))}
+          </dl>
+        </section>
+
+        <section aria-labelledby="agentx-reading-title">
+          <h2 id="agentx-reading-title" className="mb-2 text-base font-semibold text-foreground">
+            {t.readingTitle}
+          </h2>
+          <p className="text-sm leading-6 text-muted-foreground">{t.reading}</p>
+          <p className="mt-3 border-l-2 border-primary bg-primary/5 px-4 py-3 text-sm font-medium leading-6 text-foreground">
+            {t.qualityNote}
+          </p>
+          <p className="mt-3 text-sm leading-6 text-muted-foreground">{t.limits}</p>
+          <AgentXMethodologyLink
+            href={locale === 'zh' ? '/zh/agentx/methodology' : '/agentx/methodology'}
+            analyticsEvent="agentx_methodology_opened"
+            analyticsTarget="compact-overview"
+            data-testid="agentx-methodology-cta"
+            className="mt-4 inline-flex min-h-11 items-center rounded-md border border-primary/40 px-4 py-2 text-sm font-semibold text-primary transition-colors hover:bg-primary/10"
+          >
+            {t.methodologyCta} →
+          </AgentXMethodologyLink>
+        </section>
+      </div>
+    </Card>
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-optimizations-article.tsx b/packages/app/src/components/datasets/agentx-optimizations-article.tsx
new file mode 100644
index 000000000..849b6b6ee
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-optimizations-article.tsx
@@ -0,0 +1,333 @@
+import Image from 'next/image';
+
+import {
+  OPTIMIZATION_FIGURES,
+  prLabel,
+  prUrl,
+  type OptimizationFramework,
+  type OptimizationHighlight,
+  type OptimizationSection,
+  type OptimizationsOverview,
+} from '@/lib/agentx-optimizations';
+import { getLocalizedFrameworks, getOptimizationsOverview } from '@/lib/agentx-optimizations-zh';
+import type { Locale } from '@/lib/i18n';
+
+import { AgentXOptimizationsLink } from './agentx-optimizations-link';
+
+function localePrefix(locale: Locale): string {
+  return locale === 'zh' ? '/zh' : '';
+}
+
+function Highlights({ items }: { items: readonly OptimizationHighlight[] }) {
+  return (
+    <dl className="mt-8 grid grid-cols-2 gap-px overflow-hidden rounded-xl border border-border/70 bg-border/70 lg:grid-cols-4">
+      {items.map((item) => (
+        <div key={item.label} className="flex flex-col bg-card px-4 py-5">
+          <dt className="order-2 mt-1 text-sm leading-5 text-muted-foreground">{item.label}</dt>
+          <dd className="order-1 font-mono text-xl font-semibold tabular-nums text-foreground">
+            {item.value}
+          </dd>
+        </div>
+      ))}
+    </dl>
+  );
+}
+
+function SectionFigure({
+  figure,
+  ui,
+}: {
+  figure: NonNullable<OptimizationSection['figure']>;
+  ui: OptimizationsOverview['ui'];
+}) {
+  const asset = OPTIMIZATION_FIGURES[figure.key];
+  return (
+    <AgentXOptimizationsLink
+      href={asset.src}
+      analyticsEvent="agentx_optimizations_figure_opened"
+      analyticsTarget={figure.key}
+      target="_blank"
+      rel="noopener noreferrer"
+      prefetch={false}
+      className="group my-7 block rounded-xl focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2"
+    >
+      <figure
+        className="overflow-hidden rounded-xl border border-border/70 bg-card transition-colors group-hover:border-primary/50"
+        data-testid={`agentx-optimizations-figure-${figure.key}`}
+      >
+        <Image
+          src={asset.src}
+          alt={figure.alt}
+          width={asset.width}
+          height={asset.height}
+          sizes="(max-width: 768px) 100vw, 1152px"
+          quality={100}
+          className="h-auto w-full bg-white"
+        />
+        <figcaption className="border-t border-border/70 px-4 py-3 text-sm leading-6 text-muted-foreground">
+          <span className="block">{figure.caption}</span>
+          <span className="mt-1 inline-block font-medium text-foreground underline decoration-border underline-offset-4 transition-colors group-hover:text-primary group-hover:decoration-primary">
+            {ui.figureCta} ↗
+          </span>
+        </figcaption>
+      </figure>
+    </AgentXOptimizationsLink>
+  );
+}
+
+function SectionBlock({
+  section,
+  ui,
+  className,
+}: {
+  section: OptimizationSection;
+  ui: OptimizationsOverview['ui'];
+  className?: string;
+}) {
+  return (
+    <section
+      aria-labelledby={section.id}
+      className={className}
+      data-testid={`agentx-optimizations-section-${section.id}`}
+    >
+      <h2
+        id={section.id}
+        className="mb-4 scroll-mt-24 text-2xl font-semibold tracking-tight text-foreground"
+      >
+        {section.heading}
+      </h2>
+      <div className="space-y-4 text-base leading-7 text-muted-foreground">
+        {section.paragraphs.map((paragraph) => (
+          <p key={paragraph}>{paragraph}</p>
+        ))}
+      </div>
+      {section.figure && <SectionFigure figure={section.figure} ui={ui} />}
+      {section.prs && section.prs.length > 0 && (
+        <div className="mt-5">
+          <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
+            {ui.prsLabel}
+          </p>
+          <ul className="mt-2 flex flex-wrap gap-2">
+            {section.prs.map((pr) => (
+              <li key={prUrl(pr)}>
+                <AgentXOptimizationsLink
+                  href={prUrl(pr)}
+                  analyticsEvent="agentx_optimizations_pr_opened"
+                  analyticsTarget={`${pr.repo}#${pr.number}`}
+                  target="_blank"
+                  rel="noopener noreferrer"
+                  prefetch={false}
+                  data-testid="agentx-optimizations-pr"
+                  className="inline-flex items-center rounded-md border border-border bg-card px-2.5 py-1 font-mono text-xs text-muted-foreground transition-colors hover:border-primary/50 hover:text-primary"
+                >
+                  {prLabel(pr)} ↗
+                </AgentXOptimizationsLink>
+              </li>
+            ))}
+          </ul>
+        </div>
+      )}
+      {section.links && section.links.length > 0 && (
+        <div className="mt-5">
+          <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
+            {ui.referencesLabel}
+          </p>
+          <ul className="mt-2 grid gap-x-8 gap-y-2 sm:grid-cols-2">
+            {section.links.map((link) => (
+              <li key={link.href}>
+                <AgentXOptimizationsLink
+                  href={link.href}
+                  analyticsEvent="agentx_optimizations_reference_opened"
+                  analyticsTarget={link.href}
+                  target="_blank"
+                  rel="noopener noreferrer"
+                  prefetch={false}
+                  className="text-sm leading-6 text-primary underline decoration-primary/40 underline-offset-4 hover:decoration-primary"
+                >
+                  {link.label} ↗
+                </AgentXOptimizationsLink>
+              </li>
+            ))}
+          </ul>
+        </div>
+      )}
+    </section>
+  );
+}
+
+function FrameworkCard({
+  framework,
+  overview,
+  locale,
+}: {
+  framework: OptimizationFramework;
+  overview: OptimizationsOverview;
+  locale: Locale;
+}) {
+  return (
+    <AgentXOptimizationsLink
+      href={`${localePrefix(locale)}/agentx/optimizations/${framework.slug}`}
+      analyticsEvent="agentx_optimizations_framework_opened"
+      analyticsTarget={framework.slug}
+      data-testid="agentx-optimizations-card"
+      data-framework={framework.slug}
+      className="group flex flex-col rounded-xl border border-border/70 bg-card p-5 transition-colors hover:border-primary/50"
+    >
+      <p className="font-mono text-2xs font-medium tracking-eyebrow text-brand uppercase">
+        {overview.layerLabels[framework.layer]}
+      </p>
+      <h3 className="mt-2 text-lg font-semibold text-foreground">{framework.name}</h3>
+      <p className="mt-2 flex-1 text-sm leading-6 text-muted-foreground">{framework.summary}</p>
+      <span className="mt-4 inline-flex items-center text-sm font-semibold text-primary">
+        {overview.ui.readMore} →
+      </span>
+    </AgentXOptimizationsLink>
+  );
+}
+
+/** The /agentx/optimizations index: why the work happened, plus every project. */
+export function AgentXOptimizationsIndex({ locale }: { locale: Locale }) {
+  const overview = getOptimizationsOverview(locale);
+  const frameworks = getLocalizedFrameworks(locale);
+  const [ecosystem, ...rest] = overview.sections;
+
+  return (
+    <article data-testid="agentx-optimizations-index">
+      <header className="border-b border-border/70 pb-8">
+        <AgentXOptimizationsLink
+          href={`${localePrefix(locale)}/agentx`}
+          analyticsEvent="agentx_optimizations_returned"
+          analyticsTarget="overview"
+          className="text-sm text-muted-foreground transition-colors hover:text-foreground"
+        >
+          {overview.ui.backToAgentX}
+        </AgentXOptimizationsLink>
+        <p className="mt-8 font-mono text-xs font-semibold tracking-eyebrow-wide text-brand uppercase">
+          {overview.eyebrow}
+        </p>
+        <h1 className="mt-3 text-4xl font-semibold tracking-tight text-foreground sm:text-5xl">
+          {overview.title}
+        </h1>
+        <p className="mt-5 max-w-3xl text-lg leading-8 text-muted-foreground">{overview.lead}</p>
+        <Highlights items={overview.highlights} />
+      </header>
+
+      <div className="py-10">
+        <div className="space-y-4 text-base leading-7 text-muted-foreground">
+          {overview.intro.map((paragraph) => (
+            <p key={paragraph}>{paragraph}</p>
+          ))}
+        </div>
+
+        <section aria-labelledby="projects" className="mt-14">
+          <h2
+            id="projects"
+            className="mb-3 scroll-mt-24 text-2xl font-semibold tracking-tight text-foreground"
+          >
+            {overview.frameworksTitle}
+          </h2>
+          <p className="max-w-3xl text-base leading-7 text-muted-foreground">
+            {overview.frameworksIntro}
+          </p>
+          <div className="mt-6 grid gap-4 sm:grid-cols-2 lg:grid-cols-3">
+            {frameworks.map((framework) => (
+              <FrameworkCard
+                key={framework.slug}
+                framework={framework}
+                overview={overview}
+                locale={locale}
+              />
+            ))}
+          </div>
+        </section>
+
+        {ecosystem && <SectionBlock section={ecosystem} ui={overview.ui} className="mt-14" />}
+        {rest.map((section) => (
+          <SectionBlock
+            key={section.id}
+            section={section}
+            ui={overview.ui}
+            className="mt-14 border-t border-border/70 pt-10"
+          />
+        ))}
+      </div>
+    </article>
+  );
+}
+
+/** One project's page: /agentx/optimizations/<slug>. */
+export function AgentXOptimizationsArticle({ slug, locale }: { slug: string; locale: Locale }) {
+  const overview = getOptimizationsOverview(locale);
+  const frameworks = getLocalizedFrameworks(locale);
+  const framework = frameworks.find((entry) => entry.slug === slug);
+  if (!framework) return null;
+  const others = frameworks.filter((entry) => entry.slug !== slug);
+
+  return (
+    <article data-testid="agentx-optimizations-article" data-framework={framework.slug}>
+      <header className="border-b border-border/70 pb-8">
+        <AgentXOptimizationsLink
+          href={`${localePrefix(locale)}/agentx/optimizations`}
+          analyticsEvent="agentx_optimizations_returned"
+          analyticsTarget="index"
+          className="text-sm text-muted-foreground transition-colors hover:text-foreground"
+        >
+          {overview.ui.backToOverview}
+        </AgentXOptimizationsLink>
+        <p className="mt-8 font-mono text-xs font-semibold tracking-eyebrow-wide text-brand uppercase">
+          {overview.eyebrow} · {overview.layerLabels[framework.layer]}
+        </p>
+        <h1 className="mt-3 text-4xl font-semibold tracking-tight text-foreground sm:text-5xl">
+          {framework.name}
+        </h1>
+        <p className="mt-5 max-w-3xl text-lg leading-8 text-muted-foreground">{framework.lead}</p>
+        <Highlights items={framework.highlights} />
+      </header>
+
+      <div className="py-10">
+        <nav aria-label={overview.ui.onThisPage} className="mb-10">
+          <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
+            {overview.ui.onThisPage}
+          </p>
+          <ul className="mt-2 grid gap-x-8 gap-y-1 sm:grid-cols-2">
+            {framework.sections.map((section) => (
+              <li key={section.id}>
+                <a
+                  href={`#${section.id}`}
+                  className="text-sm leading-6 text-muted-foreground underline decoration-border underline-offset-4 transition-colors hover:text-foreground"
+                >
+                  {section.heading}
+                </a>
+              </li>
+            ))}
+          </ul>
+        </nav>
+
+        {framework.sections.map((section, index) => (
+          <SectionBlock
+            key={section.id}
+            section={section}
+            ui={overview.ui}
+            className={index === 0 ? undefined : 'mt-14'}
+          />
+        ))}
+
+        <section aria-labelledby="other-projects" className="mt-14 border-t border-border/70 pt-10">
+          <h2 id="other-projects" className="text-base font-semibold text-foreground">
+            {overview.ui.allProjects}
+          </h2>
+          <div className="mt-4 grid gap-4 sm:grid-cols-2 lg:grid-cols-3">
+            {others.map((entry) => (
+              <FrameworkCard
+                key={entry.slug}
+                framework={entry}
+                overview={overview}
+                locale={locale}
+              />
+            ))}
+          </div>
+        </section>
+      </div>
+    </article>
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-optimizations-callout.tsx b/packages/app/src/components/datasets/agentx-optimizations-callout.tsx
new file mode 100644
index 000000000..02dcca310
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-optimizations-callout.tsx
@@ -0,0 +1,61 @@
+import { getLocalizedFrameworks, getOptimizationsOverview } from '@/lib/agentx-optimizations-zh';
+import type { Locale } from '@/lib/i18n';
+
+import { AgentXOptimizationsLink } from './agentx-optimizations-link';
+
+/**
+ * Entry point from /agentx into the optimizations pages: the headline claim,
+ * a button to the index, and one link per project so a reader who already
+ * knows which engine they care about can go straight there. Rendered inside
+ * the AgentX methodology card, so it is a section rather than its own card.
+ */
+export function AgentXOptimizationsCallout({ locale }: { locale: Locale }) {
+  const overview = getOptimizationsOverview(locale);
+  const frameworks = getLocalizedFrameworks(locale);
+  const prefix = locale === 'zh' ? '/zh' : '';
+
+  return (
+    <section
+      aria-labelledby="agentx-optimizations-title"
+      data-testid="agentx-optimizations-callout"
+      className="rounded-lg border border-primary/30 bg-primary/5"
+    >
+      <div className="px-4 py-4 sm:px-5 sm:py-5">
+        <p className="mb-2 font-mono text-2xs font-medium tracking-eyebrow text-primary uppercase">
+          {overview.eyebrow}
+        </p>
+        <h2 id="agentx-optimizations-title" className="text-lg font-semibold text-foreground">
+          {overview.title}
+        </h2>
+        <p className="mt-3 max-w-4xl text-sm leading-6 text-muted-foreground">{overview.lead}</p>
+
+        <AgentXOptimizationsLink
+          href={`${prefix}/agentx/optimizations`}
+          analyticsEvent="agentx_optimizations_opened"
+          analyticsTarget="agentx-callout"
+          data-testid="agentx-optimizations-cta"
+          className="mt-4 inline-flex min-h-11 items-center rounded-md border border-primary/40 px-4 py-2 text-sm font-semibold text-primary transition-colors hover:bg-primary/10"
+        >
+          {overview.ui.readMore} →
+        </AgentXOptimizationsLink>
+
+        <ul className="mt-4 flex flex-wrap gap-2">
+          {frameworks.map((framework) => (
+            <li key={framework.slug}>
+              <AgentXOptimizationsLink
+                href={`${prefix}/agentx/optimizations/${framework.slug}`}
+                analyticsEvent="agentx_optimizations_framework_opened"
+                analyticsTarget={framework.slug}
+                data-testid="agentx-optimizations-framework-link"
+                data-framework={framework.slug}
+                className="inline-flex min-h-11 items-center rounded-md border border-border bg-card px-3 py-1.5 text-sm text-muted-foreground transition-colors hover:border-primary/50 hover:text-primary"
+              >
+                {framework.name}
+              </AgentXOptimizationsLink>
+            </li>
+          ))}
+        </ul>
+      </div>
+    </section>
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-optimizations-link.tsx b/packages/app/src/components/datasets/agentx-optimizations-link.tsx
new file mode 100644
index 000000000..6a941821b
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-optimizations-link.tsx
@@ -0,0 +1,41 @@
+'use client';
+
+import Link from 'next/link';
+
+import { track } from '@/lib/analytics';
+
+/**
+ * Analytics-instrumented link for the AgentX optimizations pages. The pages
+ * themselves are server components, so every tracked click goes through here.
+ */
+interface AgentXOptimizationsLinkProps extends Omit<React.ComponentProps<typeof Link>, 'href'> {
+  href: string;
+  analyticsEvent:
+    | 'agentx_optimizations_opened'
+    | 'agentx_optimizations_returned'
+    | 'agentx_optimizations_framework_opened'
+    | 'agentx_optimizations_pr_opened'
+    | 'agentx_optimizations_reference_opened'
+    | 'agentx_optimizations_figure_opened';
+  analyticsTarget?: string;
+}
+
+export function AgentXOptimizationsLink({
+  href,
+  analyticsEvent,
+  analyticsTarget,
+  onClick,
+  ...props
+}: AgentXOptimizationsLinkProps) {
+  return (
+    <Link
+      {...props}
+      href={href}
+      onClick={(event) => {
+        onClick?.(event);
+        if (event.defaultPrevented) return;
+        track(analyticsEvent, analyticsTarget ? { target: analyticsTarget } : undefined);
+      }}
+    />
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-telemetry-article.tsx b/packages/app/src/components/datasets/agentx-telemetry-article.tsx
new file mode 100644
index 000000000..5fcf24e1d
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-telemetry-article.tsx
@@ -0,0 +1,203 @@
+import Image from 'next/image';
+
+import {
+  TELEMETRY_FIGURES,
+  type TelemetryGuide,
+  type TelemetryHighlight,
+  type TelemetrySection,
+} from '@/lib/agentx-telemetry';
+import { getTelemetryGuide } from '@/lib/agentx-telemetry-zh';
+import type { Locale } from '@/lib/i18n';
+
+import { AgentXTelemetryLink } from './agentx-telemetry-link';
+
+function localePrefix(locale: Locale): string {
+  return locale === 'zh' ? '/zh' : '';
+}
+
+function Highlights({ items }: { items: readonly TelemetryHighlight[] }) {
+  return (
+    <dl className="mt-8 grid grid-cols-2 gap-px overflow-hidden rounded-xl border border-border/70 bg-border/70 lg:grid-cols-4">
+      {items.map((item) => (
+        <div key={item.label} className="flex flex-col bg-card px-4 py-5">
+          <dt className="order-2 mt-1 text-sm leading-5 text-muted-foreground">{item.label}</dt>
+          <dd className="order-1 font-mono text-xl font-semibold tabular-nums text-foreground">
+            {item.value}
+          </dd>
+        </div>
+      ))}
+    </dl>
+  );
+}
+
+function SectionFigure({
+  figure,
+  ui,
+}: {
+  figure: NonNullable<TelemetrySection['figure']>;
+  ui: TelemetryGuide['ui'];
+}) {
+  const asset = TELEMETRY_FIGURES[figure.key];
+  return (
+    <AgentXTelemetryLink
+      href={asset.src}
+      analyticsEvent="agentx_telemetry_figure_opened"
+      analyticsTarget={figure.key}
+      target="_blank"
+      rel="noopener noreferrer"
+      prefetch={false}
+      className="group my-7 block rounded-xl focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2"
+    >
+      <figure
+        className="overflow-hidden rounded-xl border border-border/70 bg-card transition-colors group-hover:border-primary/50"
+        data-testid={`agentx-telemetry-figure-${figure.key}`}
+      >
+        <Image
+          src={asset.src}
+          alt={figure.alt}
+          width={asset.width}
+          height={asset.height}
+          sizes="(max-width: 768px) 100vw, 1152px"
+          quality={100}
+          className="h-auto w-full"
+        />
+        <figcaption className="border-t border-border/70 px-4 py-3 text-sm leading-6 text-muted-foreground">
+          <span className="block">{figure.caption}</span>
+          <span className="mt-1 inline-block font-medium text-foreground underline decoration-border underline-offset-4 transition-colors group-hover:text-primary group-hover:decoration-primary">
+            {ui.figureCta} ↗
+          </span>
+        </figcaption>
+      </figure>
+    </AgentXTelemetryLink>
+  );
+}
+
+function SectionBlock({
+  section,
+  ui,
+  className,
+}: {
+  section: TelemetrySection;
+  ui: TelemetryGuide['ui'];
+  className?: string;
+}) {
+  return (
+    <section
+      aria-labelledby={section.id}
+      className={className}
+      data-testid={`agentx-telemetry-section-${section.id}`}
+    >
+      <h2
+        id={section.id}
+        className="mb-4 scroll-mt-24 text-2xl font-semibold tracking-tight text-foreground"
+      >
+        {section.heading}
+      </h2>
+      <div className="space-y-4 text-base leading-7 text-muted-foreground">
+        {section.paragraphs.map((paragraph) => (
+          <p key={paragraph}>{paragraph}</p>
+        ))}
+      </div>
+      {section.bullets && section.bullets.length > 0 && (
+        <ul className="mt-5 space-y-2 border-l border-border pl-4 text-base leading-7 text-muted-foreground">
+          {section.bullets.map((bullet) => (
+            <li key={bullet}>{bullet}</li>
+          ))}
+        </ul>
+      )}
+      {section.figure && <SectionFigure figure={section.figure} ui={ui} />}
+      {section.links && section.links.length > 0 && (
+        <ul className="mt-5 flex flex-wrap gap-2">
+          {section.links.map((link) => (
+            <li key={link.href}>
+              <AgentXTelemetryLink
+                href={link.href}
+                analyticsEvent="agentx_telemetry_opened"
+                analyticsTarget={link.href}
+                target="_blank"
+                rel="noopener noreferrer"
+                prefetch={false}
+                className="inline-flex min-h-11 items-center rounded-md border border-border bg-card px-3 py-1.5 text-sm text-muted-foreground transition-colors hover:border-primary/50 hover:text-primary"
+              >
+                {link.label}
+              </AgentXTelemetryLink>
+            </li>
+          ))}
+        </ul>
+      )}
+    </section>
+  );
+}
+
+/** The /agentx/telemetry tutorial. */
+export function AgentXTelemetryArticle({ locale }: { locale: Locale }) {
+  const guide = getTelemetryGuide(locale);
+  const prefix = localePrefix(locale);
+
+  return (
+    <article data-testid="agentx-telemetry-article">
+      <header className="border-b border-border/70 pb-8">
+        <AgentXTelemetryLink
+          href={`${prefix}/agentx`}
+          analyticsEvent="agentx_telemetry_returned"
+          analyticsTarget="agentx"
+          className="text-sm text-muted-foreground transition-colors hover:text-foreground"
+        >
+          {guide.ui.backToAgentX}
+        </AgentXTelemetryLink>
+        <p className="mt-8 font-mono text-xs font-semibold tracking-eyebrow-wide text-brand uppercase">
+          {guide.eyebrow}
+        </p>
+        <h1 className="mt-3 text-4xl font-semibold tracking-tight text-foreground sm:text-5xl">
+          {guide.title}
+        </h1>
+        <p className="mt-5 max-w-3xl text-lg leading-8 text-muted-foreground">{guide.lead}</p>
+        <Highlights items={guide.highlights} />
+        <AgentXTelemetryLink
+          href={`${prefix}/inference?i_seq=agentic-traces`}
+          analyticsEvent="agentx_telemetry_results_opened"
+          analyticsTarget="telemetry-header"
+          data-testid="agentx-telemetry-results-cta"
+          className="mt-6 inline-flex min-h-11 items-center rounded-md bg-primary px-4 py-2 text-sm font-semibold text-primary-foreground transition-colors hover:bg-primary/90"
+        >
+          {guide.ui.openResults} →
+        </AgentXTelemetryLink>
+      </header>
+
+      <div className="py-10">
+        <div className="space-y-4 text-base leading-7 text-muted-foreground">
+          {guide.intro.map((paragraph) => (
+            <p key={paragraph}>{paragraph}</p>
+          ))}
+        </div>
+
+        <nav aria-label={guide.ui.onThisPage} className="mt-10">
+          <p className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase">
+            {guide.ui.onThisPage}
+          </p>
+          <ul className="mt-2 grid gap-x-8 gap-y-1 sm:grid-cols-2">
+            {guide.sections.map((section) => (
+              <li key={section.id}>
+                <a
+                  href={`#${section.id}`}
+                  className="text-sm leading-6 text-muted-foreground underline decoration-border underline-offset-4 transition-colors hover:text-foreground"
+                >
+                  {section.heading}
+                </a>
+              </li>
+            ))}
+          </ul>
+        </nav>
+
+        {guide.sections.map((section, index) => (
+          <SectionBlock
+            key={section.id}
+            section={section}
+            ui={guide.ui}
+            className={index === 0 ? 'mt-12' : 'mt-14 border-t border-border/70 pt-10'}
+          />
+        ))}
+      </div>
+    </article>
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-telemetry-callout.tsx b/packages/app/src/components/datasets/agentx-telemetry-callout.tsx
new file mode 100644
index 000000000..1566d5c65
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-telemetry-callout.tsx
@@ -0,0 +1,54 @@
+import { getTelemetryGuide } from '@/lib/agentx-telemetry-zh';
+import type { Locale } from '@/lib/i18n';
+
+import { AgentXTelemetryLink } from './agentx-telemetry-link';
+
+const STRINGS = {
+  en: {
+    blurb:
+      'Every point on an AgentX curve opens into the run behind it: eleven per-point telemetry charts, a per-request timeline, and a per-conversation flamegraph. The tutorial walks through all of them.',
+  },
+  zh: {
+    blurb:
+      '点击 AgentX 曲线上的任一数据点，都可查看该次运行的详细数据，包括 11 张单点遥测图、逐请求时间线和单会话火焰图。本教程将逐一介绍这些视图。',
+  },
+} as const;
+
+/**
+ * Entry point from /agentx into the telemetry tutorial. Rendered inside the
+ * AgentX methodology card next to the optimizations callout, so it is a
+ * section rather than its own card.
+ */
+export function AgentXTelemetryCallout({ locale }: { locale: Locale }) {
+  const guide = getTelemetryGuide(locale);
+  const t = STRINGS[locale];
+  const prefix = locale === 'zh' ? '/zh' : '';
+
+  return (
+    <section
+      aria-labelledby="agentx-telemetry-title"
+      data-testid="agentx-telemetry-callout"
+      className="rounded-lg border border-border bg-muted/20"
+    >
+      <div className="px-4 py-4 sm:px-5 sm:py-5">
+        <p className="mb-2 font-mono text-2xs font-medium tracking-eyebrow text-brand uppercase">
+          {guide.eyebrow}
+        </p>
+        <h2 id="agentx-telemetry-title" className="text-lg font-semibold text-foreground">
+          {guide.title}
+        </h2>
+        <p className="mt-3 max-w-4xl text-sm leading-6 text-muted-foreground">{t.blurb}</p>
+
+        <AgentXTelemetryLink
+          href={`${prefix}/agentx/telemetry`}
+          analyticsEvent="agentx_telemetry_opened"
+          analyticsTarget="agentx-callout"
+          data-testid="agentx-telemetry-cta"
+          className="mt-4 inline-flex min-h-11 items-center rounded-md border border-primary/40 px-4 py-2 text-sm font-semibold text-primary transition-colors hover:bg-primary/10"
+        >
+          {guide.ui.readMore} →
+        </AgentXTelemetryLink>
+      </div>
+    </section>
+  );
+}
diff --git a/packages/app/src/components/datasets/agentx-telemetry-link.tsx b/packages/app/src/components/datasets/agentx-telemetry-link.tsx
new file mode 100644
index 000000000..90afc83fc
--- /dev/null
+++ b/packages/app/src/components/datasets/agentx-telemetry-link.tsx
@@ -0,0 +1,39 @@
+'use client';
+
+import Link from 'next/link';
+
+import { track } from '@/lib/analytics';
+
+/**
+ * Analytics-instrumented link for the AgentX telemetry tutorial. The tutorial
+ * pages are server components, so every tracked click goes through here.
+ */
+interface AgentXTelemetryLinkProps extends Omit<React.ComponentProps<typeof Link>, 'href'> {
+  href: string;
+  analyticsEvent:
+    | 'agentx_telemetry_opened'
+    | 'agentx_telemetry_returned'
+    | 'agentx_telemetry_figure_opened'
+    | 'agentx_telemetry_results_opened';
+  analyticsTarget?: string;
+}
+
+export function AgentXTelemetryLink({
+  href,
+  analyticsEvent,
+  analyticsTarget,
+  onClick,
+  ...props
+}: AgentXTelemetryLinkProps) {
+  return (
+    <Link
+      {...props}
+      href={href}
+      onClick={(event) => {
+        onClick?.(event);
+        if (event.defaultPrevented) return;
+        track(analyticsEvent, analyticsTarget ? { target: analyticsTarget } : undefined);
+      }}
+    />
+  );
+}
diff --git a/packages/app/src/components/datasets/conversation-view.tsx b/packages/app/src/components/datasets/conversation-view.tsx
index 71f3098de..84a587701 100644
--- a/packages/app/src/components/datasets/conversation-view.tsx
+++ b/packages/app/src/components/datasets/conversation-view.tsx
@@ -14,8 +14,8 @@ const STRINGS = {
   en: {
     loading: 'Loading conversation…',
     notFound: 'Conversation not found.',
-    backToDataset: 'Back to dataset',
-    breadcrumbDatasets: 'Datasets',
+    backToDataset: 'Back to AgentX dataset',
+    breadcrumbDatasets: 'AgentX',
     breadcrumbConversation: 'conversation',
     mainTurns: 'Main turns',
     subagentGroups: 'Subagent groups',
@@ -30,8 +30,8 @@ const STRINGS = {
   zh: {
     loading: '正在加载对话…',
     notFound: '未找到对话。',
-    backToDataset: '返回数据集',
-    breadcrumbDatasets: '数据集',
+    backToDataset: '返回 AgentX 数据集',
+    breadcrumbDatasets: 'AgentX',
     breadcrumbConversation: '对话',
     mainTurns: '主轮次',
     subagentGroups: 'Subagent 组',
@@ -69,7 +69,7 @@ export function ConversationView({ slug, convId }: { slug: string; convId: strin
     return (
       <div className="py-12 text-center text-sm text-destructive">
         {t.notFound}{' '}
-        <Link href={`${prefix}/datasets/${slug}`} className="text-primary underline">
+        <Link href={`${prefix}/agentx/${slug}`} className="text-primary underline">
           {t.backToDataset}
         </Link>
       </div>
@@ -82,11 +82,11 @@ export function ConversationView({ slug, convId }: { slug: string; convId: strin
     <div className="flex flex-col gap-6">
       <div>
         <div className="mb-1 flex flex-wrap items-center gap-2 text-xs text-muted-foreground">
-          <Link href={`${prefix}/datasets`} className="hover:text-foreground">
+          <Link href={`${prefix}/agentx`} className="hover:text-foreground">
             {t.breadcrumbDatasets}
           </Link>
           <span>/</span>
-          <Link href={`${prefix}/datasets/${slug}`} className="hover:text-foreground">
+          <Link href={`${prefix}/agentx/${slug}`} className="hover:text-foreground">
             {slug}
           </Link>
           <span>/</span>
diff --git a/packages/app/src/components/datasets/dataset-detail.test.tsx b/packages/app/src/components/datasets/dataset-detail.test.tsx
new file mode 100644
index 000000000..d2c1bf229
--- /dev/null
+++ b/packages/app/src/components/datasets/dataset-detail.test.tsx
@@ -0,0 +1,104 @@
+// @vitest-environment jsdom
+import { act, createElement, type PropsWithChildren } from 'react';
+import { createRoot } from 'react-dom/client';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+const { conversationQueries } = vi.hoisted(() => ({
+  conversationQueries: [] as {
+    slug: string | null;
+    search?: string;
+    limit?: number;
+    offset?: number;
+    sort?: string;
+  }[],
+}));
+
+vi.mock('next/link', () => ({
+  default: ({ children, href }: PropsWithChildren<{ href: string }>) =>
+    createElement('a', { href }, children),
+}));
+vi.mock('@/components/ui/card', () => ({
+  Card: ({ children }: PropsWithChildren) => createElement('div', null, children),
+}));
+vi.mock('@/components/ui/select', () => ({
+  Select: ({ children }: PropsWithChildren) => createElement('div', null, children),
+  SelectContent: ({ children }: PropsWithChildren) => createElement('div', null, children),
+  SelectItem: ({ children }: PropsWithChildren) => createElement('div', null, children),
+  SelectTrigger: ({ children }: PropsWithChildren) => createElement('div', null, children),
+  SelectValue: () => null,
+}));
+vi.mock('@/components/datasets/distribution-card', () => ({ DistributionCard: () => null }));
+vi.mock('@/lib/analytics', () => ({ track: vi.fn() }));
+vi.mock('@/lib/use-locale', () => ({ useLocale: () => 'en' }));
+vi.mock('@/hooks/api/use-datasets', () => ({
+  useDataset: () => ({
+    data: {
+      slug: 'trace',
+      label: 'Trace dataset',
+      variant: 'test',
+      conversation_count: 100,
+      summary: {},
+      chart_data: {},
+    },
+    isLoading: false,
+    isError: false,
+  }),
+  useDatasetConversations: (args: (typeof conversationQueries)[number]) => {
+    conversationQueries.push({ ...args });
+    return { data: { total: 100, items: [] }, isFetching: false };
+  },
+}));
+
+import { DatasetDetail } from './dataset-detail';
+
+function changeInput(input: HTMLInputElement, value: string): void {
+  Object.getOwnPropertyDescriptor(HTMLInputElement.prototype, 'value')?.set?.call(input, value);
+  input.dispatchEvent(new Event('input', { bubbles: true }));
+}
+
+afterEach(() => {
+  conversationQueries.length = 0;
+  vi.useRealTimers();
+  vi.restoreAllMocks();
+});
+
+describe('DatasetDetail conversation search', () => {
+  it('keeps typing immediate and commits only the settled search with page zero', () => {
+    vi.useFakeTimers();
+    const container = document.createElement('div');
+    const root = createRoot(container);
+
+    act(() => root.render(createElement(DatasetDetail, { slug: 'trace' })));
+    act(() => vi.advanceTimersByTime(250));
+
+    const next = [...container.querySelectorAll('button')].find(
+      (button) => button.textContent === 'Next →',
+    );
+    expect(next).toBeDefined();
+    act(() => next?.click());
+    expect(conversationQueries.at(-1)).toMatchObject({ search: '', offset: 50 });
+
+    conversationQueries.length = 0;
+    const input = container.querySelector('input');
+    expect(input).not.toBeNull();
+
+    act(() => changeInput(input!, 'a'));
+    act(() => changeInput(input!, 'ab'));
+    act(() => changeInput(input!, 'abc'));
+
+    expect(input?.value).toBe('abc');
+    expect(conversationQueries.every((query) => query.search === '')).toBe(true);
+    expect(conversationQueries.every((query) => query.offset === 50)).toBe(true);
+    expect(next?.disabled).toBe(true);
+
+    act(() => vi.advanceTimersByTime(249));
+    expect(conversationQueries.at(-1)).toMatchObject({ search: '', offset: 50 });
+
+    act(() => vi.advanceTimersByTime(1));
+    expect(conversationQueries.at(-1)).toMatchObject({ search: 'abc', offset: 0 });
+    expect(next?.disabled).toBe(false);
+    expect(new Set(conversationQueries.map((query) => query.search))).toEqual(new Set(['', 'abc']));
+
+    act(() => root.unmount());
+  });
+});
diff --git a/packages/app/src/components/datasets/dataset-detail.tsx b/packages/app/src/components/datasets/dataset-detail.tsx
index eb2a12422..359917d87 100644
--- a/packages/app/src/components/datasets/dataset-detail.tsx
+++ b/packages/app/src/components/datasets/dataset-detail.tsx
@@ -1,6 +1,6 @@
 'use client';
 
-import { useState } from 'react';
+import { useEffect, useState } from 'react';
 import Link from 'next/link';
 
 import { Card } from '@/components/ui/card';
@@ -23,13 +23,14 @@ import { compact, formatPct, formatShare, perConversation } from './format';
 import { Stat } from './stat';
 
 const PAGE = 50;
+const SEARCH_DEBOUNCE_MS = 250;
 
 const STRINGS = {
   en: {
     loading: 'Loading dataset…',
     notFound: 'Dataset not found.',
-    backToDatasets: 'Back to datasets',
-    breadcrumbDatasets: '← Datasets',
+    backToDatasets: 'Back to AgentX',
+    breadcrumbDatasets: '← AgentX',
     viewOnHf: 'View on HuggingFace ↗',
     conversations: 'Conversations',
     medianReqConvo: 'Median requests / convo',
@@ -70,8 +71,8 @@ const STRINGS = {
   zh: {
     loading: '正在加载数据集…',
     notFound: '未找到数据集。',
-    backToDatasets: '返回数据集列表',
-    breadcrumbDatasets: '← 数据集',
+    backToDatasets: '返回 AgentX',
+    breadcrumbDatasets: '← AgentX',
     viewOnHf: '在 HuggingFace 查看 ↗',
     conversations: '对话数',
     medianReqConvo: '每对话中位请求数',
@@ -113,9 +114,12 @@ const STRINGS = {
 
 export function DatasetDetail({ slug }: { slug: string }) {
   const { data: dataset, isLoading, isError } = useDataset(slug);
-  const [search, setSearch] = useState('');
-  const [sort, setSort] = useState<ConversationSort>('tokens');
-  const [page, setPage] = useState(0);
+  const [searchInput, setSearchInput] = useState('');
+  const [{ search, sort, page }, setConversationQuery] = useState({
+    search: '',
+    sort: 'tokens' as ConversationSort,
+    page: 0,
+  });
   const locale = useLocale();
   const t = STRINGS[locale];
   const prefix = locale === 'zh' ? '/zh' : '';
@@ -127,22 +131,44 @@ export function DatasetDetail({ slug }: { slug: string }) {
     { value: 'id', label: t.sortId },
   ];
 
-  const { data: convs, isFetching } = useDatasetConversations({
+  useEffect(() => {
+    const timer = window.setTimeout(() => {
+      setConversationQuery((current) => {
+        if (current.search === searchInput && current.page === 0) return current;
+        return { ...current, search: searchInput, page: 0 };
+      });
+    }, SEARCH_DEBOUNCE_MS);
+    return () => window.clearTimeout(timer);
+  }, [searchInput]);
+
+  const {
+    data: convs,
+    isFetching,
+    isPlaceholderData,
+  } = useDatasetConversations({
     slug,
     search,
     sort,
     limit: PAGE,
     offset: page * PAGE,
   });
+  const paginationPending = searchInput !== search || isFetching || isPlaceholderData;
 
   if (isLoading) {
-    return <div className="py-12 text-center text-sm text-muted-foreground">{t.loading}</div>;
+    return (
+      <div
+        className="min-h-svh py-12 text-center text-sm text-muted-foreground"
+        data-testid="dataset-detail-loading"
+      >
+        {t.loading}
+      </div>
+    );
   }
   if (isError || !dataset) {
     return (
-      <div className="py-12 text-center text-sm text-destructive">
+      <div className="min-h-svh py-12 text-center text-sm text-destructive">
         {t.notFound}{' '}
-        <Link href={`${prefix}/datasets`} className="text-primary underline">
+        <Link href={`${prefix}/agentx`} className="text-primary underline">
           {t.backToDatasets}
         </Link>
       </div>
@@ -155,12 +181,12 @@ export function DatasetDetail({ slug }: { slug: string }) {
   const pageCount = Math.ceil(total / PAGE);
 
   return (
-    <div className="flex flex-col gap-6">
+    <div className="min-h-svh flex flex-col gap-6">
       {/* header */}
       <div>
         <div className="mb-1 flex items-center gap-2">
           <Link
-            href={`${prefix}/datasets`}
+            href={`${prefix}/agentx`}
             className="text-xs text-muted-foreground hover:text-foreground"
           >
             {t.breadcrumbDatasets}
@@ -281,19 +307,19 @@ export function DatasetDetail({ slug }: { slug: string }) {
           <div className="flex items-center gap-2">
             <input
               type="text"
-              value={search}
-              onChange={(e) => {
-                setSearch(e.target.value);
-                setPage(0);
-              }}
+              value={searchInput}
+              onChange={(e) => setSearchInput(e.target.value)}
               placeholder={t.searchPlaceholder}
               className="h-8 w-40 rounded-md border border-border/40 bg-background px-2 text-xs outline-none focus:border-primary"
             />
             <Select
               value={sort}
               onValueChange={(v) => {
-                setSort(v as ConversationSort);
-                setPage(0);
+                setConversationQuery((current) => ({
+                  ...current,
+                  sort: v as ConversationSort,
+                  page: 0,
+                }));
                 track('datasets_conversations_sorted', { mode: v });
               }}
             >
@@ -333,14 +359,14 @@ export function DatasetDetail({ slug }: { slug: string }) {
                   >
                     <td className="px-3 py-2">
                       <Link
-                        href={`${prefix}/datasets/${slug}/conversations/${encodeURIComponent(c.conv_id)}`}
+                        href={`${prefix}/agentx/${slug}/conversations/${encodeURIComponent(c.conv_id)}`}
                         onClick={() => track('datasets_conversation_clicked', { slug })}
                         className="font-mono text-xs text-primary hover:underline"
                       >
                         {c.conv_id.slice(0, 20)}…
                       </Link>
                       {c.models.length > 0 && (
-                        <span className="ml-2 text-[11px] text-muted-foreground">
+                        <span className="ml-2 text-2xs text-muted-foreground">
                           {t.modelSuffix(c.models.length)}
                         </span>
                       )}
@@ -370,11 +396,11 @@ export function DatasetDetail({ slug }: { slug: string }) {
           <div className="flex items-center justify-center gap-3 text-xs">
             <button
               type="button"
-              disabled={page === 0}
+              disabled={paginationPending || page === 0}
               onClick={() => {
                 const next = Math.max(0, page - 1);
                 track('datasets_conversations_page_changed', { direction: 'prev', page: next });
-                setPage(next);
+                setConversationQuery((current) => ({ ...current, page: next }));
               }}
               className="rounded-md border border-border/40 px-2 py-1 hover:bg-accent disabled:opacity-30"
             >
@@ -383,11 +409,11 @@ export function DatasetDetail({ slug }: { slug: string }) {
             <span className="text-muted-foreground">{t.pageOf(page + 1, pageCount)}</span>
             <button
               type="button"
-              disabled={page >= pageCount - 1}
+              disabled={paginationPending || page >= pageCount - 1}
               onClick={() => {
                 const next = Math.min(pageCount - 1, page + 1);
                 track('datasets_conversations_page_changed', { direction: 'next', page: next });
-                setPage(next);
+                setConversationQuery((current) => ({ ...current, page: next }));
               }}
               className="rounded-md border border-border/40 px-2 py-1 hover:bg-accent disabled:opacity-30"
             >
diff --git a/packages/app/src/components/datasets/dataset-list.tsx b/packages/app/src/components/datasets/dataset-list.tsx
index a18869854..20d67cf5d 100644
--- a/packages/app/src/components/datasets/dataset-list.tsx
+++ b/packages/app/src/components/datasets/dataset-list.tsx
@@ -26,15 +26,15 @@ const STRINGS = {
   zh: {
     loading: '正在加载数据集…',
     error: '数据集加载失败。',
-    empty: '尚未导入数据集。',
-    conversations: '对话数',
-    medianReqConvo: '每对话中位请求数',
-    meanReqConvo: '每对话平均请求数',
-    mainTurns: '主轮次',
-    subagentGroups: 'Subagent 组',
-    cachedInput: '缓存输入',
-    totalInput: '总输入',
-    totalOutput: '总输出',
+    empty: '尚无已导入的数据集。',
+    conversations: '会话数',
+    medianReqConvo: '单会话请求数中位数',
+    meanReqConvo: '单会话平均请求数',
+    mainTurns: 'main agent 轮次',
+    subagentGroups: 'subagent 组',
+    cachedInput: 'cached input 占比',
+    totalInput: 'input token 总数',
+    totalOutput: 'output token 总数',
     viewDataset: '查看数据集 →',
   },
 } as const;
@@ -46,14 +46,14 @@ function DatasetCard({ d, locale }: { d: DatasetRecord; locale: 'en' | 'zh' }) {
   const prefix = locale === 'zh' ? '/zh' : '';
   return (
     <Link
-      href={`${prefix}/datasets/${d.slug}`}
+      href={`${prefix}/agentx/${d.slug}`}
       onClick={() => track('datasets_card_clicked', { slug: d.slug })}
       className="block transition-colors hover:[&_*]:border-primary/40"
     >
       <Card className="h-full p-4 transition-colors hover:border-primary/40">
         <div className="mb-1 flex items-baseline justify-between gap-2">
           <h3 className="text-base font-semibold text-foreground">{d.label}</h3>
-          <span className="rounded-full border border-border/50 px-2 py-0.5 text-[10px] uppercase tracking-wide text-muted-foreground">
+          <span className="rounded-full border border-border/50 px-2 py-0.5 text-3xs uppercase tracking-wide text-muted-foreground">
             {d.variant}
           </span>
         </div>
diff --git a/packages/app/src/components/datasets/distribution-card.tsx b/packages/app/src/components/datasets/distribution-card.tsx
index d0cfe30f0..867b4daef 100644
--- a/packages/app/src/components/datasets/distribution-card.tsx
+++ b/packages/app/src/components/datasets/distribution-card.tsx
@@ -115,7 +115,7 @@ export function DistributionCard({
       <div className="mb-0.5 flex items-baseline justify-between gap-2">
         <span className="text-sm font-medium text-foreground">{title}</span>
         {scale === 'log' && (
-          <span className="text-[10px] uppercase tracking-wide text-muted-foreground">
+          <span className="text-3xs uppercase tracking-wide text-muted-foreground">
             {t.logScale}
           </span>
         )}
diff --git a/packages/app/src/components/datasets/trace-flamegraph.tsx b/packages/app/src/components/datasets/trace-flamegraph.tsx
index b6a764600..be67e43d0 100644
--- a/packages/app/src/components/datasets/trace-flamegraph.tsx
+++ b/packages/app/src/components/datasets/trace-flamegraph.tsx
@@ -227,7 +227,7 @@ export function TraceFlamegraph({
       </div>
 
       {braces.overflowLanes > 0 && (
-        <p className="mb-2 text-[11px] text-muted-foreground">
+        <p className="mb-2 text-2xs text-muted-foreground">
           {t.denseParallel(MAX_LANES, braces.overflowLanes)}
         </p>
       )}
@@ -337,7 +337,7 @@ export function TraceFlamegraph({
                   </div>
 
                   <div
-                    className="w-36 shrink-0 text-[11px] tabular-nums text-muted-foreground"
+                    className="w-36 shrink-0 text-2xs tabular-nums text-muted-foreground"
                     data-testid={`flamegraph-time-${row.key}`}
                   >
                     {row.timeLabel ?? '—'}
@@ -360,7 +360,7 @@ export function TraceFlamegraph({
                     </div>
                   </div>
 
-                  <div className="w-16 shrink-0 text-right text-[11px] tabular-nums text-muted-foreground">
+                  <div className="w-16 shrink-0 text-right text-2xs tabular-nums text-muted-foreground">
                     {compact(row.total)}
                   </div>
                 </div>
diff --git a/packages/app/src/components/evaluation/EvaluationContext.test.ts b/packages/app/src/components/evaluation/EvaluationContext.test.ts
new file mode 100644
index 000000000..fc237bcc0
--- /dev/null
+++ b/packages/app/src/components/evaluation/EvaluationContext.test.ts
@@ -0,0 +1,24 @@
+import { describe, expect, it } from 'vitest';
+
+import { resolveEvaluationDate } from '@/components/evaluation/EvaluationContext';
+
+describe('resolveEvaluationDate', () => {
+  const dates = ['2026-08-01', '2026-08-10', '2026-08-20'];
+
+  it('uses latest when neither evaluation nor global intent provides a date', () => {
+    expect(resolveEvaluationDate('', dates)).toBe('2026-08-20');
+  });
+
+  it('keeps an available requested date', () => {
+    expect(resolveEvaluationDate('2026-08-10', dates)).toBe('2026-08-10');
+  });
+
+  it('selects the nearest evaluation date without copying it into intent', () => {
+    expect(resolveEvaluationDate('2026-08-13', dates)).toBe('2026-08-10');
+    expect(resolveEvaluationDate('2026-08-18', dates)).toBe('2026-08-20');
+  });
+
+  it('preserves requested intent while evaluation availability is empty', () => {
+    expect(resolveEvaluationDate('2026-08-13', [])).toBe('2026-08-13');
+  });
+});
diff --git a/packages/app/src/components/evaluation/EvaluationContext.tsx b/packages/app/src/components/evaluation/EvaluationContext.tsx
index 897dd1568..ff7ac661a 100644
--- a/packages/app/src/components/evaluation/EvaluationContext.tsx
+++ b/packages/app/src/components/evaluation/EvaluationContext.tsx
@@ -14,17 +14,21 @@ import {
 import { DISPLAY_MODEL_TO_DB } from '@semianalysisai/inferencex-constants';
 import { track } from '@/lib/analytics';
 
-import { useGlobalFilters } from '@/components/GlobalFilterContext';
+import {
+  useGlobalFilterActions,
+  useGlobalFilterAvailability,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+} from '@/components/GlobalFilterContext';
 import { useUnofficialRun } from '@/components/unofficial-run-provider';
 import {
+  resolveAvailableSelection,
   useChartUIState,
   useChartToggleSet,
-  useAutoInitializeToggleSet,
   useUrlStateSync,
 } from '@/hooks/useChartContext';
 import { useEvaluations } from '@/hooks/api/use-evaluations';
 import { useUrlState } from '@/hooks/useUrlState';
-import { normalizeEvalHardwareKey } from '@/lib/chart-utils';
 import type { Model } from '@/lib/data-mappings';
 import type { EvalRow } from '@/lib/api';
 
@@ -38,42 +42,55 @@ import type { EvalChangelogEntry, EvaluationChartContextType, EvaluationChartDat
 /** @internal Exported for test provider wrapping only. */
 export const EvaluationContext = createContext<EvaluationChartContextType | undefined>(undefined);
 
+export function resolveEvaluationDate(
+  requestedDate: string,
+  availableDates: readonly string[],
+): string {
+  if (availableDates.length === 0) return requestedDate;
+  if (!requestedDate) return availableDates.at(-1)!;
+  if (availableDates.includes(requestedDate)) return requestedDate;
+
+  const target = new Date(requestedDate).getTime();
+  return availableDates.reduce((closest, date) => {
+    const closestDifference = Math.abs(new Date(closest).getTime() - target);
+    const difference = Math.abs(new Date(date).getTime() - target);
+    return difference < closestDifference ? date : closest;
+  }, availableDates[0]);
+}
+
 export function EvaluationProvider({ children }: { children: ReactNode }) {
+  const { selectedModel, effectivePrecisions } = useGlobalFilterSelection();
   const {
-    selectedModel,
     setSelectedModel,
-    selectedRunDate: globalRunDate,
-    selectedRunDateRev,
     setSelectedRunDate: setGlobalRunDate,
+    setSelectedPrecisions,
+  } = useGlobalFilterActions();
+  const { selectedRunDate: globalRunDate, selectedRunDateRev } = useGlobalFilterRun();
+  const {
     availableModels,
     availableDates: inferenceAvailableDates,
-    effectivePrecisions,
-    setSelectedPrecisions,
     availablePrecisions: globalAvailablePrecisions,
-  } = useGlobalFilters();
+    availabilityError,
+  } = useGlobalFilterAvailability();
   const { getUrlParam } = useUrlState();
-  const { data: rawRows, isLoading: loading, error: queryError } = useEvaluations();
+  const {
+    data: rawRows,
+    isLoading: loading,
+    isSuccess: evaluationsSettled,
+    error: queryError,
+  } = useEvaluations();
   const { unofficialEvalRows, localOfficialOverride } = useUnofficialRun();
 
-  const error = queryError ? queryError.message : null;
+  const error = availabilityError || (queryError ? queryError.message : null);
   const rawData: EvalRow[] = rawRows ?? [];
   const unofficialRawData: EvalRow[] = unofficialEvalRows ?? [];
 
-  const [selectedRunDate, setSelectedRunDate] = useState<string>(
-    () => getUrlParam('e_rundate') || globalRunDate || '',
-  );
+  const [dateIntent, setDateIntent] = useState(() => ({
+    date: getUrlParam('e_rundate') || '',
+    globalRevision: selectedRunDateRev,
+  }));
 
-  const handleSetSelectedRunDate = useCallback(
-    (date: string) => {
-      setSelectedRunDate(date);
-      if (inferenceAvailableDates.length === 0 || inferenceAvailableDates.includes(date)) {
-        setGlobalRunDate(date);
-      }
-    },
-    [inferenceAvailableDates, setGlobalRunDate],
-  );
-
-  const [selectedBenchmark, setSelectedBenchmark] = useState<string | undefined>(
+  const [requestedBenchmark, setRequestedBenchmark] = useState<string | undefined>(
     () => getUrlParam('e_bench') || undefined,
   );
 
@@ -92,11 +109,11 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
   } = useChartToggleSet();
 
   // Pending legend-active selection restored from `e_active` URL param.
-  // Consumed once when hwTypesWithData first populates.
+  // Consumed once when hardware availability settles.
   const [pendingActiveHardware, setPendingActiveHardware] = useState<Set<string> | null>(() => {
-    const v = getUrlParam('e_active');
-    if (!v) return null;
-    const set = new Set(v.split(',').filter(Boolean));
+    const value = getUrlParam('e_active');
+    if (!value) return null;
+    const set = new Set(value.split(',').filter(Boolean));
     return set.size > 0 ? set : null;
   });
 
@@ -119,65 +136,29 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
     return [...dates].toSorted();
   }, [rawData, selectedModel]);
 
-  const prevAvailableDatesRef = useRef<string[]>([]);
+  const selectedBenchmark =
+    requestedBenchmark && availableBenchmarks.includes(requestedBenchmark)
+      ? requestedBenchmark
+      : availableBenchmarks[0];
 
-  useEffect(() => {
-    if (availableBenchmarks.length > 0 && !selectedBenchmark) {
-      setSelectedBenchmark(availableBenchmarks[0]);
-    }
-    if (availableModels.length > 0 && !selectedModel) {
-      setSelectedModel(availableModels[0]);
-    }
-  }, [availableBenchmarks, availableModels, selectedBenchmark, setSelectedModel]);
+  const preferredRunDate =
+    selectedRunDateRev > dateIntent.globalRevision
+      ? globalRunDate
+      : dateIntent.date || globalRunDate;
+  const selectedRunDate = resolveEvaluationDate(preferredRunDate, availableDates);
 
-  useEffect(() => {
-    if (availableDates.length === 0) return;
-    const latestDate = availableDates.at(-1);
-    const prevAvailableDates = prevAvailableDatesRef.current;
-    const wasOnLatest =
-      prevAvailableDates.length > 0 && selectedRunDate === prevAvailableDates.at(-1);
-    if (!selectedRunDate || wasOnLatest || !availableDates.includes(selectedRunDate)) {
-      setSelectedRunDate(latestDate!);
-      // If no global date yet (evals loaded first), set it so inference syncs to us.
-      if (!globalRunDate) setGlobalRunDate(latestDate!);
-    }
-    prevAvailableDatesRef.current = availableDates;
-  }, [availableDates, selectedRunDate, setSelectedRunDate, globalRunDate, setGlobalRunDate]);
-
-  useEffect(() => {
-    if (!globalRunDate) return;
-    if (availableDates.length === 0) {
-      setSelectedRunDate(globalRunDate);
-      return;
-    }
-    if (availableDates.includes(globalRunDate)) {
-      setSelectedRunDate(globalRunDate);
-      return;
-    }
-    // Snap to the nearest valid date
-    const target = new Date(globalRunDate).getTime();
-    let closest = availableDates[0];
-    let minDiff = Math.abs(new Date(closest).getTime() - target);
-    for (const d of availableDates) {
-      const diff = Math.abs(new Date(d).getTime() - target);
-      if (diff < minDiff) {
-        minDiff = diff;
-        closest = d;
-      }
-    }
-    setSelectedRunDate(closest);
-  }, [globalRunDate, availableDates, selectedRunDateRev]);
-
-  const availableHardware = useMemo(() => {
-    const hwSet = new Set<string>();
-    rawData.forEach((item) => {
-      const hwKey = normalizeEvalHardwareKey(item.hardware, item.framework, item.spec_method);
-      if (hwKey !== 'unknown') hwSet.add(hwKey);
-    });
-    return [...hwSet].toSorted();
-  }, [rawData]);
-
-  useAutoInitializeToggleSet(availableHardware, enabledHardware, setEnabledHardware);
+  const handleSetSelectedRunDate = useCallback(
+    (date: string) => {
+      const updatesGlobal =
+        inferenceAvailableDates.length === 0 || inferenceAvailableDates.includes(date);
+      setDateIntent({
+        date,
+        globalRevision: selectedRunDateRev + (updatesGlobal ? 1 : 0),
+      });
+      if (updatesGlobal) setGlobalRunDate(date);
+    },
+    [inferenceAvailableDates, selectedRunDateRev, setGlobalRunDate],
+  );
 
   const availablePrecisions = useMemo(() => {
     const dbModelKeys = DISPLAY_MODEL_TO_DB[selectedModel];
@@ -256,16 +237,29 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
     [unfilteredChartData],
   );
 
+  const hardwareScopeKey = `${selectedModel}|${selectedBenchmark ?? ''}|${selectedRunDate}|${effectivePrecisions.join(',')}`;
+  const lastHardwareScopeRef = useRef('');
   useEffect(() => {
-    if (hwTypesWithData.size === 0) return;
-    if (pendingActiveHardware) {
-      const restored = new Set([...pendingActiveHardware].filter((k) => hwTypesWithData.has(k)));
-      setEnabledHardware(restored.size > 0 ? restored : hwTypesWithData);
-      setPendingActiveHardware(null);
-      return;
-    }
-    setEnabledHardware(hwTypesWithData);
-  }, [selectedModel, hwTypesWithData]);
+    const scopeChanged = lastHardwareScopeRef.current !== hardwareScopeKey;
+    const resolution = resolveAvailableSelection({
+      active: enabledHardware,
+      available: hwTypesWithData,
+      pending: pendingActiveHardware,
+      scopeChanged,
+      settled: evaluationsSettled,
+    });
+    if (!evaluationsSettled) return;
+    lastHardwareScopeRef.current = hardwareScopeKey;
+    if (resolution.selection !== enabledHardware) setEnabledHardware(resolution.selection);
+    if (resolution.consumedPending) setPendingActiveHardware(null);
+  }, [
+    enabledHardware,
+    evaluationsSettled,
+    hardwareScopeKey,
+    hwTypesWithData,
+    pendingActiveHardware,
+    setEnabledHardware,
+  ]);
 
   const selectAllHwTypes = useCallback(
     () => selectAllHwRaw(hwTypesWithData),
@@ -346,7 +340,7 @@ export function EvaluationProvider({ children }: { children: ReactNode }) {
       loading,
       error,
       selectedBenchmark,
-      setSelectedBenchmark,
+      setSelectedBenchmark: setRequestedBenchmark,
       selectedModel,
       setSelectedModel: handleSetSelectedModel,
       selectedRunDate,
diff --git a/packages/app/src/components/evaluation/ui/BarChartD3.test.ts b/packages/app/src/components/evaluation/ui/BarChartD3.test.ts
new file mode 100644
index 000000000..5b9924e55
--- /dev/null
+++ b/packages/app/src/components/evaluation/ui/BarChartD3.test.ts
@@ -0,0 +1,73 @@
+// @vitest-environment jsdom
+import * as d3 from 'd3';
+import { describe, expect, it } from 'vitest';
+
+import type { EvaluationChartData } from '../types';
+import { sizeScoreLabelBackgrounds } from './BarChartD3';
+
+function makeDatum(configLabel: string, score: number): EvaluationChartData {
+  return {
+    evalResultId: score,
+    configId: score,
+    hwKey: configLabel,
+    hardware: configLabel,
+    configLabel,
+    score,
+    model: 'model',
+    benchmark: 'benchmark',
+    specDecode: 'none',
+    date: '2026-01-01',
+    datetime: '2026-01-01T00:00:00Z',
+    precision: 'fp8',
+    framework: 'framework',
+    tp: 1,
+    ep: 1,
+    dp_attention: false,
+    conc: 1,
+    disagg: false,
+    isMultinode: false,
+    prefillTp: 1,
+    prefillEp: 1,
+    prefillDpAttention: false,
+    prefillNumWorkers: 1,
+    decodeNumWorkers: 1,
+    numPrefillGpu: 0,
+    numDecodeGpu: 1,
+  };
+}
+
+describe('sizeScoreLabelBackgrounds', () => {
+  it('finishes every text measurement before writing any label background geometry', () => {
+    const svg = d3.select(document.createElementNS('http://www.w3.org/2000/svg', 'svg'));
+    const root = svg.append('g');
+    const labelGroups = root
+      .selectAll<SVGGElement, EvaluationChartData>('.score-label-group')
+      .data([makeDatum('a', 1), makeDatum('b', 2)])
+      .join('g')
+      .attr('class', 'score-label-group');
+    labelGroups.append('text').attr('class', 'score-label');
+
+    const events: string[] = [];
+    labelGroups.each(function (_datum, index) {
+      const text = this.querySelector<SVGTextElement>('.score-label')!;
+      Object.defineProperty(text, 'getBBox', {
+        value: () => {
+          events.push(`read:${index}`);
+          return new DOMRect(index, index + 1, 10 + index, 5 + index);
+        },
+      });
+      const insertBefore = this.insertBefore.bind(this);
+      Object.defineProperty(this, 'insertBefore', {
+        value: (node: Node, child: Node | null) => {
+          events.push(`write:${index}`);
+          return insertBefore(node, child);
+        },
+      });
+    });
+
+    sizeScoreLabelBackgrounds(labelGroups);
+
+    expect(events).toEqual(['read:0', 'read:1', 'write:0', 'write:1']);
+    expect(labelGroups.select('.score-label-bg').attr('width')).toBe('20');
+  });
+});
diff --git a/packages/app/src/components/evaluation/ui/BarChartD3.tsx b/packages/app/src/components/evaluation/ui/BarChartD3.tsx
index 760e2b22e..5acec763c 100644
--- a/packages/app/src/components/evaluation/ui/BarChartD3.tsx
+++ b/packages/app/src/components/evaluation/ui/BarChartD3.tsx
@@ -9,6 +9,11 @@ import { D3Chart, type D3ChartHandle, type LayerConfig } from '@/lib/d3-chart/D3
 import { renderErrorBars } from '@/lib/d3-chart/layers/error-bars';
 import { renderPoints, updatePointsOnZoom } from '@/lib/d3-chart/layers/points';
 import { computeTooltipPosition } from '@/lib/d3-chart/layers/scatter-points';
+import {
+  attachOverlayXMarkerHandlers,
+  overlayMarkerPosition,
+  xMarkerPath,
+} from '@/lib/d3-chart/overlay-x-marker';
 import { computeLeftMargin } from '@/lib/d3-chart/dynamic-margins';
 
 import { useEvaluation } from '@/components/evaluation/EvaluationContext';
@@ -23,7 +28,10 @@ import {
 import { useLocale } from '@/lib/use-locale';
 import ChartLegend from '@/components/ui/chart-legend';
 import { Skeleton } from '@/components/ui/skeleton';
-import { useUnofficialRun } from '@/components/unofficial-run-provider';
+import {
+  useOverlayScopeReconciliation,
+  useUnofficialRun,
+} from '@/components/unofficial-run-provider';
 import { useThemeColors } from '@/hooks/useThemeColors';
 import { computeToggle } from '@/hooks/useTogglableSet';
 import { overlayRunColor, overlayRunIndex } from '@/lib/overlay-run-style';
@@ -34,9 +42,6 @@ const OVERLAY_X_HOVER_SIZE = 8;
 const OVERLAY_HIT_RADIUS = 10;
 const OVERLAY_ERROR_STROKE_WIDTH = 1.5;
 
-const getOverlayXPath = (size: number) =>
-  `M ${-size},${-size} L ${size},${size} M ${-size},${size} L ${size},${-size}`;
-
 const formatDateStr = (dateStr: string) => {
   const [year, month, day] = dateStr.split('-');
   return new Intl.DateTimeFormat('en-US', {
@@ -133,6 +138,94 @@ function formatYAxisLabels(axisGroup: d3.Selection<SVGGElement, unknown, null, u
   });
 }
 
+/**
+ * Measure every score label before writing any background geometry. Keeping the
+ * DOM-read pass separate prevents one synchronous layout per label.
+ */
+export function sizeScoreLabelBackgrounds(
+  labelGroups: d3.Selection<SVGGElement, EvaluationChartData, SVGGElement, unknown>,
+): void {
+  const measurements: { group: SVGGElement; bounds: DOMRect }[] = [];
+  labelGroups.each(function () {
+    const text = this.querySelector<SVGTextElement>('.score-label');
+    if (text) measurements.push({ group: this, bounds: text.getBBox() });
+  });
+
+  for (const { group, bounds } of measurements) {
+    d3.select(group)
+      .selectAll<SVGRectElement, EvaluationChartData>('.score-label-bg')
+      .data((datum) => [datum])
+      .join((enter) => enter.insert('rect', ':first-child').attr('class', 'score-label-bg'))
+      .attr('rx', 4)
+      .attr('ry', 4)
+      .attr('fill', 'var(--popover)')
+      .attr('stroke', 'var(--border)')
+      .attr('stroke-width', 1)
+      .attr('x', bounds.x - 5)
+      .attr('y', bounds.y - 1)
+      .attr('width', bounds.width + 10)
+      .attr('height', bounds.height + 2);
+  }
+}
+
+function renderScoreLabels(
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  data: EvaluationChartData[],
+  showLabels: boolean,
+  xScale: d3.ScaleLinear<number, number>,
+  yScale: d3.ScaleBand<string>,
+): void {
+  const labelGroups = group
+    .selectAll<SVGGElement, EvaluationChartData>('.score-label-group')
+    .data(showLabels ? data : [], (datum) => datum.configLabel)
+    .join((enter) => {
+      const labelGroup = enter.append('g').attr('class', 'score-label-group');
+      // Backgrounds are inserted only after every text node has been measured.
+      labelGroup
+        .append('text')
+        .attr('class', 'score-label')
+        .attr('text-anchor', 'start')
+        .style('fill', 'var(--foreground)')
+        .attr('font-size', '10px')
+        .attr('font-weight', '600')
+        .attr('dy', '0.35em');
+      return labelGroup;
+    })
+    .attr(
+      'transform',
+      (datum) =>
+        `translate(${xScale(datum.score) + 12},${
+          (yScale(datum.configLabel) || 0) + yScale.bandwidth() / 2
+        })`,
+    );
+
+  labelGroups.select<SVGTextElement>('.score-label').text((datum) => datum.score.toFixed(3));
+  sizeScoreLabelBackgrounds(labelGroups);
+}
+
+function updateUnofficialScoreLabels(
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  showLabels: boolean,
+): void {
+  group
+    .selectAll<SVGGElement, EvaluationChartData>('.unofficial-eval-point')
+    .each(function (datum) {
+      d3.select(this)
+        .selectAll<SVGTextElement, EvaluationChartData>('.unofficial-score-label')
+        .data(showLabels ? [datum] : [], (labelDatum) => labelDatum.configLabel)
+        .join('text')
+        .attr('class', 'unofficial-score-label')
+        .attr('x', 12)
+        .attr('text-anchor', 'start')
+        .style('fill', 'var(--foreground)')
+        .attr('font-size', '10px')
+        .attr('font-weight', '600')
+        .attr('dy', '0.35em')
+        .attr('pointer-events', 'none')
+        .text((labelDatum) => labelDatum.score.toFixed(3));
+    });
+}
+
 const EVAL_STRINGS = {
   en: {
     showLabels: 'Show Labels',
@@ -178,11 +271,9 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
     unofficialRunInfo,
     unofficialRunInfos,
     activeOverlayHwTypes,
-    setActiveOverlayHwTypes,
     allOverlayHwTypes,
-    resetOverlayHwTypes,
     localOfficialOverride,
-    setLocalOfficialOverride,
+    setUnifiedOverlaySelection,
     runIndexByUrl,
   } = useUnofficialRun();
   const chartRef = useRef<D3ChartHandle>(null);
@@ -206,6 +297,41 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
     },
     [runIndexByUrl, unofficialRunInfos, unofficialRunInfo],
   );
+  const branchForRowRef = useRef(branchForRow);
+  branchForRowRef.current = branchForRow;
+  const evaluationOverlayScope = useMemo(
+    () => new Set(unofficialChartData.map((datum) => String(datum.hwKey))),
+    [unofficialChartData],
+  );
+  const evaluationScopeRegistration = useMemo(
+    () =>
+      isUnofficialRun
+        ? {
+            scopeKey: `evaluation|${selectedModel ?? ''}|${selectedBenchmark ?? ''}|${
+              selectedRunDate ?? ''
+            }|${unofficialRunInfos.map((run) => run.url).join(',')}`,
+            officialHwTypes: hwTypesWithData,
+            overlayHwTypes: evaluationOverlayScope,
+            bestOfficialHwTypes: hwTypesWithData,
+            bestOverlayHwTypes: evaluationOverlayScope,
+            bestPerSku: false,
+            ready: !loading || hwTypesWithData.size > 0,
+          }
+        : null,
+    [
+      isUnofficialRun,
+      selectedModel,
+      selectedBenchmark,
+      selectedRunDate,
+      unofficialRunInfos,
+      hwTypesWithData,
+      evaluationOverlayScope,
+      loading,
+      activeOverlayHwTypes,
+      localOfficialOverride,
+    ],
+  );
+  useOverlayScopeReconciliation(evaluationScopeRegistration);
 
   const effectiveOfficialHardware = localOfficialOverride ?? enabledHardware;
 
@@ -228,15 +354,13 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
         if (key.startsWith('overlay:')) nextOverlay.add(key.slice(8));
         else nextOfficial.add(key);
       }
-      setLocalOfficialOverride(nextOfficial);
-      setActiveOverlayHwTypes(nextOverlay);
+      setUnifiedOverlaySelection(nextOfficial, nextOverlay);
     },
     [
       activeOverlayHwTypes,
       allUnifiedHwTypes,
       effectiveOfficialHardware,
-      setActiveOverlayHwTypes,
-      setLocalOfficialOverride,
+      setUnifiedOverlaySelection,
     ],
   );
 
@@ -446,7 +570,7 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
   );
 
   const parallelismKey = hasDisaggConfigs ? (
-    <div className="mt-2 px-1 pr-2 text-[10px] text-muted-foreground/80 leading-tight no-export">
+    <div className="mt-2 px-1 pr-2 text-3xs text-muted-foreground/80 leading-tight no-export">
       <div>
         <span className="font-mono">P(·/·/·/·)</span> prefill
         <span className="mx-1">·</span>
@@ -464,6 +588,143 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
     [unofficialChartData],
   );
 
+  const dataIdentity = useMemo(
+    () =>
+      JSON.stringify(
+        [
+          ...chartData.map((datum) => JSON.stringify(['official', datum.configLabel])),
+          ...unofficialChartData.map((datum) =>
+            JSON.stringify(['unofficial', datum.runUrl ?? '', datum.configLabel]),
+          ),
+        ].toSorted(),
+      ),
+    [chartData, unofficialChartData],
+  );
+  const metricIdentity = useMemo(
+    () =>
+      JSON.stringify({
+        official: chartData
+          .map((datum) =>
+            JSON.stringify([
+              datum.configLabel,
+              datum.score,
+              datum.errorMin ?? null,
+              datum.errorMax ?? null,
+            ]),
+          )
+          .toSorted(),
+        unofficial: unofficialChartData
+          .map((datum) =>
+            JSON.stringify([
+              datum.runUrl ?? '',
+              datum.configLabel,
+              datum.score,
+              datum.errorMin ?? null,
+              datum.errorMax ?? null,
+            ]),
+          )
+          .toSorted(),
+        xDomain,
+        yLabels,
+      }),
+    [chartData, unofficialChartData, xDomain, yLabels],
+  );
+  const officialPaletteIdentity = useMemo(
+    () =>
+      JSON.stringify(
+        chartData
+          .map((datum) =>
+            JSON.stringify([
+              datum.configLabel,
+              getCssColor(resolveColor(datum.configLabel, String(datum.hwKey))),
+            ]),
+          )
+          .toSorted(),
+      ),
+    [chartData, getCssColor, resolveColor],
+  );
+  const unofficialPaletteIdentity = useMemo(
+    () =>
+      JSON.stringify(
+        unofficialChartData
+          .map((datum) =>
+            JSON.stringify([
+              datum.runUrl ?? '',
+              overlayRunColor(overlayRunIndex(datum.runUrl ?? null, runIndexByUrl)),
+            ]),
+          )
+          .toSorted(),
+      ),
+    [runIndexByUrl, unofficialChartData],
+  );
+  const displayIdentity = `${showLabels}:${officialPaletteIdentity}:${unofficialPaletteIdentity}`;
+  const xScaleConfig = useMemo(() => ({ type: 'linear' as const, domain: xDomain }), [xDomain]);
+  const yScaleConfig = useMemo(
+    () => ({ type: 'band' as const, domain: yLabels, padding: 0.1 }),
+    [yLabels],
+  );
+  const xAxisConfig = useMemo(
+    () => ({
+      label: `${getEvalBenchmarkLabel(selectedBenchmark as EvalBenchmark)} Score`,
+      tickFormat: (datum: d3.AxisDomain) => Number(datum).toFixed(2),
+      tickCount: 5,
+    }),
+    [selectedBenchmark],
+  );
+  const yAxisConfig = useMemo(() => ({ customize: formatYAxisLabels }), []);
+  const zoomConfig = useMemo(
+    () => ({
+      enabled: true,
+      axes: 'x' as const,
+      scaleExtent: [1, 20] as [number, number],
+      resetEventName: 'evaluation_zoom_reset_evaluation-chart',
+      constrain: (transform: d3.ZoomTransform) => {
+        const k = transform.k;
+        const innerWidth =
+          (typeof window === 'undefined' ? 800 : window.innerWidth) -
+          chartMargin.left -
+          chartMargin.right;
+        const xScale = d3.scaleLinear().domain(xDomain).range([0, innerWidth]);
+        const minTx = -xScale(1) * k + innerWidth;
+        const maxTx = -xScale(0) * k;
+        const tx = minTx < maxTx ? Math.max(minTx, Math.min(maxTx, transform.x)) : transform.x;
+        return d3.zoomIdentity.translate(tx, transform.y).scale(k);
+      },
+    }),
+    [chartMargin.left, chartMargin.right, xDomain],
+  );
+  const tooltipConfig = useMemo(
+    () => ({
+      rulerType: 'crosshair' as const,
+      content: generateEvaluationTooltipContent,
+      getRulerX: (
+        datum: EvaluationChartData,
+        scale:
+          | d3.ScaleBand<string>
+          | d3.ScaleLinear<number, number>
+          | d3.ScaleLogarithmic<number, number>,
+      ) => {
+        const xScale = scale as unknown as d3.ScaleLinear<number, number>;
+        return xScale(datum.score);
+      },
+      getRulerY: (
+        datum: EvaluationChartData,
+        scale: d3.ScaleLinear<number, number> | d3.ScaleLogarithmic<number, number>,
+      ) => {
+        const bandScale = scale as unknown as d3.ScaleBand<string>;
+        return (bandScale(datum.configLabel) || 0) + bandScale.bandwidth() / 2;
+      },
+      onHoverStart: (
+        selection: d3.Selection<SVGCircleElement, EvaluationChartData, SVGGElement, unknown>,
+      ) => selection.attr('r', 8),
+      onHoverEnd: (
+        selection: d3.Selection<SVGCircleElement, EvaluationChartData, SVGGElement, unknown>,
+      ) => selection.attr('r', 6),
+      attachToLayer: 1,
+    }),
+    [],
+  );
+
   // Horizontal bar chart: yScale = band (config labels), xScale = linear (scores)
   const layers = useMemo(
     (): LayerConfig<EvaluationChartData>[] => [
@@ -534,6 +795,7 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
       {
         type: 'custom',
         key: 'mean-points',
+        displayIdentity: officialPaletteIdentity,
         render: (group, { xScale: xs, yScale: ys }) => {
           const xScale = xs as d3.ScaleLinear<number, number>;
           const yScale = ys as d3.ScaleBand<string>;
@@ -548,6 +810,13 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
             strokeWidth: 0,
           });
         },
+        onDisplayUpdate: (group) => {
+          group
+            .selectAll<SVGCircleElement, EvaluationChartData>('.point')
+            .attr('fill', (datum) =>
+              getCssColor(resolveColor(datum.configLabel, String(datum.hwKey))),
+            );
+        },
         onZoom: (group, ctx) => {
           const newXScale = ctx.newXScale as d3.ScaleLinear<number, number>;
           const yScale = ctx.yScale as d3.ScaleBand<string>;
@@ -568,10 +837,7 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
 
           const bars = group
             .selectAll<SVGGElement, EvaluationChartData>('.unofficial-error-bar')
-            .data(
-              unofficialErrorData,
-              (d) => `${d.configLabel}|${d.score}|${d.errorMin}|${d.errorMax}`,
-            )
+            .data(unofficialErrorData, (datum) => `${datum.runUrl ?? ''}|${datum.configLabel}`)
             .join((enter) => {
               const bar = enter.append('g').attr('class', 'unofficial-error-bar');
               bar
@@ -650,47 +916,29 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
       {
         type: 'custom',
         key: 'score-labels',
+        displayIdentity: showLabels ? 'visible' : 'hidden',
         render: (group, { xScale: xs, yScale: ys }) => {
-          group.selectAll('.score-label-group').remove();
-          if (!showLabels) return;
-          const xScale = xs as d3.ScaleLinear<number, number>;
-          const yScale = ys as d3.ScaleBand<string>;
-          const labelGroups = group
-            .selectAll('.score-label-group')
-            .data(chartData)
-            .join('g')
-            .attr('class', 'score-label-group')
-            .attr(
-              'transform',
-              (d) =>
-                `translate(${xScale(d.score) + 12},${(yScale(d.configLabel) || 0) + yScale.bandwidth() / 2})`,
-            );
-          labelGroups
-            .append('rect')
-            .attr('class', 'score-label-bg')
-            .attr('rx', 4)
-            .attr('ry', 4)
-            .attr('fill', 'var(--popover)')
-            .attr('stroke', 'var(--border)')
-            .attr('stroke-width', 1);
-          labelGroups
-            .append('text')
-            .attr('class', 'score-label')
-            .attr('text-anchor', 'start')
-            .style('fill', 'var(--foreground)')
-            .attr('font-size', '10px')
-            .attr('font-weight', '600')
-            .attr('dy', '0.35em')
-            .text((d) => d.score.toFixed(3));
-          labelGroups.each(function () {
-            const g = d3.select(this);
-            const bbox = (g.select('text').node() as SVGTextElement).getBBox();
-            g.select('.score-label-bg')
-              .attr('x', bbox.x - 5)
-              .attr('y', bbox.y - 1)
-              .attr('width', bbox.width + 10)
-              .attr('height', bbox.height + 2);
-          });
+          renderScoreLabels(
+            group,
+            chartData,
+            showLabels,
+            xs as d3.ScaleLinear<number, number>,
+            ys as d3.ScaleBand<string>,
+          );
+        },
+        onDisplayUpdate: (group, ctx) => {
+          const baseXScale = ctx.xScale as d3.ScaleLinear<number, number>;
+          const svgNode = ctx.layout.svg.node();
+          const currentXScale = svgNode
+            ? d3.zoomTransform(svgNode).rescaleX(baseXScale)
+            : baseXScale;
+          renderScoreLabels(
+            group,
+            chartData,
+            showLabels,
+            currentXScale,
+            ctx.yScale as d3.ScaleBand<string>,
+          );
         },
         onZoom: (group, ctx) => {
           if (!showLabels) return;
@@ -700,14 +948,17 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
             .selectAll<SVGGElement, EvaluationChartData>('.score-label-group')
             .attr(
               'transform',
-              (d) =>
-                `translate(${newXScale(d.score) + 12},${(yScale(d.configLabel) || 0) + yScale.bandwidth() / 2})`,
+              (datum) =>
+                `translate(${newXScale(datum.score) + 12},${
+                  (yScale(datum.configLabel) || 0) + yScale.bandwidth() / 2
+                })`,
             );
         },
       },
       {
         type: 'custom',
         key: 'unofficial-overlay',
+        displayIdentity: `${showLabels}:${unofficialPaletteIdentity}`,
         render: (group, { xScale: xs, yScale: ys, layout }) => {
           const xScale = xs as d3.ScaleLinear<number, number>;
           const yScale = ys as d3.ScaleBand<string>;
@@ -719,7 +970,7 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
           const tooltip = d3.select(tooltipNode);
           const overlayPoints = group
             .selectAll<SVGGElement, EvaluationChartData>('.unofficial-eval-point')
-            .data(unofficialChartData, (d) => `${d.configLabel}|${d.score}`)
+            .data(unofficialChartData, (datum) => `${datum.runUrl ?? ''}|${datum.configLabel}`)
             .join((enter) => {
               const g = enter.append('g').attr('class', 'unofficial-eval-point');
               g.append('circle')
@@ -728,7 +979,7 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
                 .attr('cursor', 'pointer');
               g.append('path')
                 .attr('class', 'unofficial-eval-x')
-                .attr('d', getOverlayXPath(OVERLAY_X_SIZE))
+                .attr('d', xMarkerPath(OVERLAY_X_SIZE))
                 .attr('fill', 'none')
                 .attr('stroke-width', 2.5)
                 .attr('stroke-linecap', 'round')
@@ -749,61 +1000,42 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
               overlayRunColor(overlayRunIndex(d.runUrl ?? null, runIndexByUrl)),
             );
 
-          overlayPoints.each(function (d) {
-            d3.select(this)
-              .selectAll<SVGTextElement, boolean>('.unofficial-score-label')
-              .data(showLabels ? [true] : [])
-              .join('text')
-              .attr('class', 'unofficial-score-label')
-              .attr('x', 12)
-              .attr('text-anchor', 'start')
-              .style('fill', 'var(--foreground)')
-              .attr('font-size', '10px')
-              .attr('font-weight', '600')
-              .attr('dy', '0.35em')
-              .attr('pointer-events', 'none')
-              .text(d.score.toFixed(3));
-          });
+          updateUnofficialScoreLabels(group, showLabels);
 
-          overlayPoints
-            .on('mouseenter', function (_event, d) {
-              if (chartRef.current?.isPinned()) return;
-              d3.select(this)
-                .select('.unofficial-eval-x')
-                .attr('d', getOverlayXPath(OVERLAY_X_HOVER_SIZE))
-                .attr('stroke-width', 3.5);
-              tooltip
-                .style('opacity', 1)
-                .style('display', 'block')
-                .style('pointer-events', 'none')
-                .html(generateEvaluationTooltipContent(d, false, branchForRow(d)));
-            })
-            .on('mousemove', (event) => {
-              if (chartRef.current?.isPinned()) return;
-              const [mx, my] = d3.pointer(event, container);
-              const pos = computeTooltipPosition(mx, my, tooltip, container);
-              tooltip.style('left', `${pos.left}px`).style('top', `${pos.top}px`);
-            })
-            .on('mouseleave', function () {
-              if (chartRef.current?.isPinned()) return;
-              d3.select(this)
-                .select('.unofficial-eval-x')
-                .attr('d', getOverlayXPath(OVERLAY_X_SIZE))
-                .attr('stroke-width', 2.5);
-              tooltip.style('opacity', 0).style('display', 'none');
-            })
-            .on('click', (event, d) => {
-              event.stopPropagation();
-              const [mx, my] = d3.pointer(event, container);
-              tooltip
-                .html(generateEvaluationTooltipContent(d, true, branchForRow(d)))
-                .style('opacity', 1)
-                .style('display', 'block')
-                .style('pointer-events', 'auto');
-              const pos = computeTooltipPosition(mx, my, tooltip, container);
-              tooltip.style('left', `${pos.left}px`).style('top', `${pos.top}px`);
-              chartRef.current?.pinTooltip(d, true);
-            });
+          attachOverlayXMarkerHandlers(overlayPoints, {
+            markerSelector: '.unofficial-eval-x',
+            normalPath: xMarkerPath(OVERLAY_X_SIZE),
+            hoverPath: xMarkerPath(OVERLAY_X_HOVER_SIZE),
+            tooltip,
+            handle: chartRef.current,
+            content: (datum, pinned) =>
+              generateEvaluationTooltipContent(datum, pinned, branchForRowRef.current(datum)),
+            position: (event) => {
+              const [mouseX, mouseY] = d3.pointer(event, container);
+              return computeTooltipPosition(mouseX, mouseY, tooltip, container);
+            },
+            rulers: {
+              show: (datum, marker) => {
+                const position = overlayMarkerPosition(marker) ?? {
+                  x: xScale(datum.score),
+                  y: (yScale(datum.configLabel) || 0) + yScale.bandwidth() / 2,
+                };
+                group.select('.ruler-group').style('display', 'block');
+                group.select('.vertical-ruler').attr('x1', position.x).attr('x2', position.x);
+                group.select('.horizontal-ruler').attr('y1', position.y).attr('y2', position.y);
+              },
+              hide: () => group.select('.ruler-group').style('display', 'none'),
+            },
+          });
+        },
+        onDisplayUpdate: (group) => {
+          group
+            .selectAll<SVGGElement, EvaluationChartData>('.unofficial-eval-point')
+            .select('.unofficial-eval-x')
+            .attr('stroke', (datum) =>
+              overlayRunColor(overlayRunIndex(datum.runUrl ?? null, runIndexByUrl)),
+            );
+          updateUnofficialScoreLabels(group, showLabels);
         },
         onZoom: (group, { newXScale, yScale: ys }) => {
           const xScale = newXScale as d3.ScaleLinear<number, number>;
@@ -822,11 +1054,12 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
       chartData,
       errorData,
       getCssColor,
+      officialPaletteIdentity,
       resolveColor,
       showLabels,
       unofficialChartData,
       unofficialErrorData,
-      branchForRow,
+      unofficialPaletteIdentity,
       runIndexByUrl,
     ],
   );
@@ -875,50 +1108,21 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
       ref={chartRef}
       chartId="evaluation-chart"
       data={chartData}
+      dataIdentity={dataIdentity}
+      metricIdentity={metricIdentity}
+      displayIdentity={displayIdentity}
       height={chartHeight}
       margin={chartMargin}
       watermark={getChartWatermark(isUnofficialRun)}
       grabCursor={false}
       caption={caption}
-      xScale={{ type: 'linear', domain: xDomain }}
-      yScale={{ type: 'band', domain: yLabels, padding: 0.1 }}
-      xAxis={{
-        label: `${getEvalBenchmarkLabel(selectedBenchmark as EvalBenchmark)} Score`,
-        tickFormat: (d) => Number(d).toFixed(2),
-        tickCount: 5,
-      }}
-      yAxis={{ customize: formatYAxisLabels }}
+      xScale={xScaleConfig}
+      yScale={yScaleConfig}
+      xAxis={xAxisConfig}
+      yAxis={yAxisConfig}
       layers={layers}
-      zoom={{
-        enabled: true,
-        axes: 'x',
-        scaleExtent: [1, 20],
-        resetEventName: 'evaluation_zoom_reset_evaluation-chart',
-        constrain: (transform) => {
-          const k = transform.k;
-          const innerWidth =
-            (typeof window === 'undefined' ? 800 : window.innerWidth) -
-            chartMargin.left -
-            chartMargin.right;
-          const xScale = d3.scaleLinear().domain(xDomain).range([0, innerWidth]);
-          const minTx = -xScale(1) * k + innerWidth;
-          const maxTx = -xScale(0) * k;
-          const tx = minTx < maxTx ? Math.max(minTx, Math.min(maxTx, transform.x)) : transform.x;
-          return d3.zoomIdentity.translate(tx, transform.y).scale(k);
-        },
-      }}
-      tooltip={{
-        rulerType: 'crosshair',
-        content: generateEvaluationTooltipContent,
-        getRulerX: (d, xs) => (xs as d3.ScaleLinear<number, number>)(d.score),
-        getRulerY: (d, ys) => {
-          const bs = ys as unknown as d3.ScaleBand<string>;
-          return (bs(d.configLabel) || 0) + bs.bandwidth() / 2;
-        },
-        onHoverStart: (sel) => sel.attr('r', 8),
-        onHoverEnd: (sel) => sel.attr('r', 6),
-        attachToLayer: 1,
-      }}
+      zoom={zoomConfig}
+      tooltip={tooltipConfig}
       legendElement={
         <ChartLegend
           variant="sidebar"
@@ -958,8 +1162,7 @@ export default function EvalBarChartD3({ caption }: { caption?: ReactNode }) {
                     label: legendT.resetFilter,
                     onClick: () => {
                       selectAllHwTypes();
-                      setLocalOfficialOverride(null);
-                      resetOverlayHwTypes();
+                      setUnifiedOverlaySelection(hwTypesWithData, allOverlayHwTypes);
                       track('evaluation_filter_reset');
                     },
                   },
diff --git a/packages/app/src/components/evaluation/ui/ChartControls.tsx b/packages/app/src/components/evaluation/ui/ChartControls.tsx
index 0ca9a115d..1fac3b96a 100644
--- a/packages/app/src/components/evaluation/ui/ChartControls.tsx
+++ b/packages/app/src/components/evaluation/ui/ChartControls.tsx
@@ -1,9 +1,8 @@
 'use client';
 
-import { useState } from 'react';
-
 import { track } from '@/lib/analytics';
 import { useLocale } from '@/lib/use-locale';
+import { useOpenDropdown } from '@/hooks/useOpenDropdown';
 import { ChevronDownIcon } from 'lucide-react';
 
 import { useEvaluation } from '@/components/evaluation/EvaluationContext';
@@ -40,14 +39,7 @@ const STRINGS = {
 
 export default function EvaluationChartControls() {
   const t = STRINGS[useLocale()];
-  const [openDropdown, setOpenDropdown] = useState<string | null>(null);
-  const handleDropdownOpenChange = (dropdownKey: string) => (isOpen: boolean) => {
-    if (isOpen) {
-      setOpenDropdown(dropdownKey);
-      return;
-    }
-    setOpenDropdown((current) => (current === dropdownKey ? null : current));
-  };
+  const { openDropdown, handleDropdownOpenChange } = useOpenDropdown<string>();
 
   const {
     selectedBenchmark,
diff --git a/packages/app/src/components/evaluation/ui/ChartDisplay.tsx b/packages/app/src/components/evaluation/ui/ChartDisplay.tsx
index 8061ea8fc..cd210208a 100644
--- a/packages/app/src/components/evaluation/ui/ChartDisplay.tsx
+++ b/packages/app/src/components/evaluation/ui/ChartDisplay.tsx
@@ -40,7 +40,7 @@ const STRINGS = {
     tableView: '表格',
     viewModeAria: '视图模式',
     heading: '准确率评估',
-    description: '基准测试结果展示不同 Chip、量化精度和推理配置下，模型质量与吞吐量之间的权衡。',
+    description: '基准测试结果展示不同芯片、量化精度和推理配置下，模型质量与吞吐量之间的权衡。',
     captionHeading: '各硬件配置的评估得分',
     sourceUnofficial: '来源：非官方',
     sourceOfficial: '来源：SemiAnalysis InferenceX™',
diff --git a/packages/app/src/components/evaluation/ui/EvalSamplesDrawer.tsx b/packages/app/src/components/evaluation/ui/EvalSamplesDrawer.tsx
index 92e4c9af4..f909eccbe 100644
--- a/packages/app/src/components/evaluation/ui/EvalSamplesDrawer.tsx
+++ b/packages/app/src/components/evaluation/ui/EvalSamplesDrawer.tsx
@@ -348,7 +348,7 @@ export default function EvalSamplesDrawer({
             </p>
           )}
           {!liveUnavailable && isUnofficial && (
-            <p className="mb-2 rounded-md border border-primary/40 bg-primary/5 px-3 py-1.5 text-[11px] text-primary">
+            <p className="mb-2 rounded-md border border-primary/40 bg-primary/5 px-3 py-1.5 text-2xs text-primary">
               {t.unofficialNotice}
             </p>
           )}
@@ -384,7 +384,7 @@ export default function EvalSamplesDrawer({
                   aria-expanded={expanded.has(s.docId)}
                 >
                   <PassFailBadge passed={s.passed} />
-                  <span className="font-mono text-[10px] text-muted-foreground tabular-nums">
+                  <span className="font-mono text-3xs text-muted-foreground tabular-nums">
                     #{s.docId}
                   </span>
                   <span className="min-w-0 flex-1 truncate text-xs">{s.prompt ?? t.noPrompt}</span>
@@ -422,7 +422,7 @@ export default function EvalSamplesDrawer({
                       <button
                         type="button"
                         onClick={() => handleShare(s.docId)}
-                        className="inline-flex items-center gap-1 rounded-md border border-border px-2 py-1 text-[11px] font-medium text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
+                        className="inline-flex items-center gap-1 rounded-md border border-border px-2 py-1 text-2xs font-medium text-muted-foreground transition-colors hover:bg-muted hover:text-foreground"
                         aria-label={t.shareSample}
                         title={t.shareSample}
                         data-testid={`eval-sample-share-${s.docId}`}
@@ -521,7 +521,7 @@ function FilterChip({ label, count, active, onClick, tone }: FilterChipProps) {
       aria-pressed={active}
     >
       {label}
-      <span className="ml-1.5 text-[10px] tabular-nums opacity-70">{count}</span>
+      <span className="ml-1.5 text-3xs tabular-nums opacity-70">{count}</span>
     </button>
   );
 }
@@ -529,14 +529,14 @@ function FilterChip({ label, count, active, onClick, tone }: FilterChipProps) {
 function PassFailBadge({ passed }: { passed: boolean | null }) {
   if (passed === null) {
     return (
-      <span className="inline-flex h-4 w-4 shrink-0 items-center justify-center rounded-full border border-muted-foreground/30 text-[10px] text-muted-foreground">
+      <span className="inline-flex h-4 w-4 shrink-0 items-center justify-center rounded-full border border-muted-foreground/30 text-3xs text-muted-foreground">
         ?
       </span>
     );
   }
   return (
     <span
-      className={`inline-flex h-4 w-4 shrink-0 items-center justify-center rounded-full border text-[10px] font-bold ${
+      className={`inline-flex h-4 w-4 shrink-0 items-center justify-center rounded-full border text-3xs font-bold ${
         passed
           ? 'border-emerald-500/40 bg-emerald-500/10 text-emerald-700 dark:text-emerald-400'
           : 'border-rose-500/40 bg-rose-500/10 text-rose-700 dark:text-rose-400'
@@ -590,16 +590,16 @@ function FewShotBlock({
               key={i}
               className="rounded border border-border/40 bg-muted/30 p-2"
             >
-              <div className="mb-1 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+              <div className="mb-1 text-3xs font-semibold uppercase tracking-wider text-muted-foreground">
                 {t.exampleQuestion(i + 1)}
               </div>
-              <pre className="mb-2 whitespace-pre-wrap font-mono text-[11px] leading-snug wrap-break-word">
+              <pre className="mb-2 whitespace-pre-wrap font-mono text-2xs leading-snug wrap-break-word">
                 {d.question}
               </pre>
-              <div className="mb-1 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+              <div className="mb-1 text-3xs font-semibold uppercase tracking-wider text-muted-foreground">
                 {t.exampleAnswer(i + 1)}
               </div>
-              <pre className="whitespace-pre-wrap font-mono text-[11px] leading-snug wrap-break-word">
+              <pre className="whitespace-pre-wrap font-mono text-2xs leading-snug wrap-break-word">
                 {d.answer}
               </pre>
             </div>
@@ -621,10 +621,10 @@ function Block({
 }) {
   return (
     <div>
-      <div className="mb-1 text-[10px] font-semibold uppercase tracking-wider text-muted-foreground">
+      <div className="mb-1 text-3xs font-semibold uppercase tracking-wider text-muted-foreground">
         {label}
       </div>
-      <pre className="max-h-72 overflow-auto whitespace-pre-wrap rounded border border-border/40 bg-muted/30 p-2 font-mono text-[11px] leading-snug wrap-break-word">
+      <pre className="max-h-72 overflow-auto whitespace-pre-wrap rounded border border-border/40 bg-muted/30 p-2 font-mono text-2xs leading-snug wrap-break-word">
         {value ?? <span className="italic text-muted-foreground">{emptyText}</span>}
       </pre>
     </div>
diff --git a/packages/app/src/components/evaluation/ui/EvaluationTable.tsx b/packages/app/src/components/evaluation/ui/EvaluationTable.tsx
index 7ac274e4e..9656aa723 100644
--- a/packages/app/src/components/evaluation/ui/EvaluationTable.tsx
+++ b/packages/app/src/components/evaluation/ui/EvaluationTable.tsx
@@ -8,8 +8,10 @@ import type { EvaluationChartData } from '@/components/evaluation/types';
 import { useUnofficialRun } from '@/components/unofficial-run-provider';
 import { type DataTableColumn, DataTable } from '@/components/ui/data-table';
 import { track } from '@/lib/analytics';
+import { notifyClientSearchChange } from '@/lib/client-navigation';
 import { overlayRunColor, overlayRunIndex } from '@/lib/overlay-run-style';
 import { useLocale } from '@/lib/use-locale';
+import { useClientSearchParams } from '@/hooks/useClientSearch';
 
 const STRINGS = {
   en: {
@@ -58,14 +60,23 @@ export default function EvaluationTable({ data }: EvaluationTableProps) {
   const sorted = useMemo(() => [...data].toSorted((a, b) => b.score - a.score), [data]);
   const hasDisaggConfigs = useMemo(() => data.some((d) => d.disagg), [data]);
   const [drawerRow, setDrawerRow] = useState<EvaluationChartData | null>(null);
-  const [sharedTarget, setSharedTarget] = useState<{
-    evalResultId: number;
-    docId?: number;
-  } | null>(null);
+  const searchParams = useClientSearchParams();
+  const sharedTarget = useMemo(() => {
+    const evalParam = searchParams.get('eval');
+    if (evalParam === null) return null;
+    const evalResultId = Number(evalParam);
+    if (!Number.isInteger(evalResultId) || evalResultId <= 0) return null;
+
+    const sampleParam = searchParams.get('sample');
+    const docId = sampleParam === null ? undefined : Number(sampleParam);
+    return {
+      evalResultId,
+      ...(docId !== undefined && Number.isInteger(docId) && docId >= 0 ? { docId } : {}),
+    };
+  }, [searchParams]);
   const trackedSharedEvalId = useRef<number | null>(null);
 
   const openDrawer = (row: EvaluationChartData) => {
-    setSharedTarget(null);
     setDrawerRow(row);
     // Notify the first-visit nudge to dismiss itself once the user has
     // discovered the affordance on their own.
@@ -79,29 +90,13 @@ export default function EvaluationTable({ data }: EvaluationTableProps) {
     });
   };
 
-  useEffect(() => {
-    const params = new URLSearchParams(window.location.search);
-    const evalParam = params.get('eval');
-    if (evalParam === null) return;
-
-    const evalResultId = Number(evalParam);
-    if (!Number.isInteger(evalResultId) || evalResultId <= 0) return;
-
-    const sampleParam = params.get('sample');
-    const docId = sampleParam === null ? undefined : Number(sampleParam);
-    setSharedTarget({
-      evalResultId,
-      ...(docId !== undefined && Number.isInteger(docId) && docId >= 0 ? { docId } : {}),
-    });
-  }, []);
-
   const closeDrawer = () => {
     setDrawerRow(null);
-    setSharedTarget(null);
     const url = new URL(window.location.href);
     url.searchParams.delete('eval');
     url.searchParams.delete('sample');
     window.history.replaceState(window.history.state, '', url);
+    notifyClientSearchChange(url.href);
   };
 
   const sharedRow = sharedTarget
@@ -234,7 +229,7 @@ export default function EvaluationTable({ data }: EvaluationTableProps) {
   return (
     <>
       {hasDisaggConfigs && (
-        <div className="mt-2 mb-2 text-[11px] text-muted-foreground/80 leading-tight">
+        <div className="mt-2 mb-2 text-2xs text-muted-foreground/80 leading-tight">
           <div>
             <span className="font-mono">P(·/·/·/·)</span> {t.prefill}
             <span className="mx-1">·</span>
diff --git a/packages/app/src/components/favorites/favorite-presets.ts b/packages/app/src/components/favorites/favorite-presets.ts
index b5ed21c3a..be92d39c2 100644
--- a/packages/app/src/components/favorites/favorite-presets.ts
+++ b/packages/app/src/components/favorites/favorite-presets.ts
@@ -149,7 +149,7 @@ export const FAVORITE_PRESETS: FavoritePreset[] = [
     titleZh: 'Kimi K3 — 首发基准测试',
     description:
       'First benchmarks of Kimi K3 across every available chip. New configurations appear here as they come online.',
-    descriptionZh: '涵盖所有可用 Chip 的 Kimi K3 首批基准测试结果。新配置上线后将在此同步更新。',
+    descriptionZh: '涵盖所有可用芯片的 Kimi K3 首批基准测试结果。新配置上线后将在此同步更新。',
     tags: ['Kimi', 'K3', 'New'],
     category: 'comparison',
     wide: true,
@@ -170,7 +170,7 @@ export const FAVORITE_PRESETS: FavoritePreset[] = [
     titleZh: 'MiniMax M3 — 首发基准测试',
     description:
       'First benchmarks of MiniMax M3 across every available chip. New configurations appear here as they come online.',
-    descriptionZh: '涵盖所有可用 Chip 的 MiniMax M3 首批基准测试结果。新配置上线后将在此同步更新。',
+    descriptionZh: '涵盖所有可用芯片的 MiniMax M3 首批基准测试结果。新配置上线后将在此同步更新。',
     tags: ['MiniMax', 'M3', 'New'],
     category: 'comparison',
     wide: true,
@@ -248,7 +248,7 @@ export const FAVORITE_PRESETS: FavoritePreset[] = [
     description:
       'Blackwell B200 vs Hopper H200 Dynamo TRTLLM throughput per chip on DeepSeek R1 (8k/1k) at FP8.',
     descriptionZh:
-      'Blackwell B200 vs Hopper H200 Dynamo TRTLLM 每 Chip 吞吐量对比，基于 DeepSeek R1 (8k/1k)，FP8 精度。',
+      'Blackwell B200 vs Hopper H200 Dynamo TRTLLM 每芯片吞吐量对比，基于 DeepSeek R1 (8k/1k)，FP8 精度。',
     tags: ['DeepSeek', 'B200', 'H200', 'Dynamo', 'FP8'],
     category: 'comparison',
     config: {
diff --git a/packages/app/src/components/feedback-modal.tsx b/packages/app/src/components/feedback-modal.tsx
index 990e28674..98e39c928 100644
--- a/packages/app/src/components/feedback-modal.tsx
+++ b/packages/app/src/components/feedback-modal.tsx
@@ -192,7 +192,7 @@ function FieldBlock({
         <label htmlFor={id} className="text-xs font-medium">
           {label}
         </label>
-        <span className="text-[10px] tabular-nums text-muted-foreground">
+        <span className="text-3xs tabular-nums text-muted-foreground">
           {value.length}/{MAX_LEN}
         </span>
       </div>
diff --git a/packages/app/src/components/feedback-viewer/FeedbackViewer.tsx b/packages/app/src/components/feedback-viewer/FeedbackViewer.tsx
index e6501a135..b877a8470 100644
--- a/packages/app/src/components/feedback-viewer/FeedbackViewer.tsx
+++ b/packages/app/src/components/feedback-viewer/FeedbackViewer.tsx
@@ -329,7 +329,7 @@ function FeedbackRow({ row }: { row: DecryptedRow }) {
         )}
         {row.wantToSee && <FieldRow label={t.wouldLikeToSee} value={row.wantToSee} tone="want" />}
         {row.userAgent && (
-          <p className="text-[10px] text-muted-foreground/70 font-mono">{row.userAgent}</p>
+          <p className="text-3xs text-muted-foreground/70 font-mono">{row.userAgent}</p>
         )}
       </div>
     </Card>
@@ -353,7 +353,7 @@ function FieldRow({
         : 'border-l-2 border-sky-500/40';
   return (
     <div className={`pl-3 ${toneClass}`}>
-      <div className="text-[10px] uppercase tracking-wide text-muted-foreground">{label}</div>
+      <div className="text-3xs uppercase tracking-wide text-muted-foreground">{label}</div>
       <p className="text-sm whitespace-pre-wrap break-words">{value}</p>
     </div>
   );
diff --git a/packages/app/src/components/footer/footer.tsx b/packages/app/src/components/footer/footer.tsx
index a4d36179a..5364b98a0 100644
--- a/packages/app/src/components/footer/footer.tsx
+++ b/packages/app/src/components/footer/footer.tsx
@@ -12,7 +12,7 @@ import { StarButton } from './footer-star-cta';
 const STRINGS = {
   en: {
     description:
-      'Continuous open-source inference benchmarking. Real-world, reproducible, auditable performance data trusted by trillion dollar AI infrastructure operators like OpenAI, Meta, Oracle, Microsoft, etc.',
+      'Continuous open-source agentic inference benchmarking. Real-world, reproducible, auditable performance data trusted by trillion dollar AI infrastructure operators like OpenAI, Meta, Oracle, Microsoft, etc.',
     semianalysis: 'SemiAnalysis',
     mainSite: 'Main Site',
     newsletter: 'Newsletter',
@@ -23,48 +23,54 @@ const STRINGS = {
     cookiePolicy: 'Cookie Policy',
     contribute: 'Contribute',
     benchmarks: 'Benchmarks',
-    frontend: 'Frontend',
+    agentxHarness: 'AgentX Harness',
+    visualization: 'Visualization',
     more: 'More',
     supporters: 'Supporters',
-    datasets: 'Datasets',
+    agentx: 'AgentX',
+    telemetry: 'Telemetry',
     articles: 'Articles',
     apiReference: 'API Reference',
     gpuReliability: 'Chip Reliability',
     perfPerDollar: 'Performance per Dollar',
+    modelArchitectures: 'Model Architectures',
     glossary: 'AI Inference Glossary',
-    languageLink: '中文版',
-    languageHref: '/zh',
-    languageHrefLang: 'zh-CN',
+    chipSpecs: 'Chip Specs & Pricing',
+    rankings: 'GPU Rankings',
+    runPages: 'Model on GPU Results',
     cta: 'If this data helps your work, consider starring us on GitHub or sharing with your network.',
     rights: 'All rights reserved.',
   },
   zh: {
     description:
-      '持续的开源推理基准测试。真实、可复现、可审计的性能数据，获得 OpenAI、Meta、Oracle、Microsoft 等万亿美元级 AI 基础设施运营方的信赖。',
+      'InferenceX 持续开展开源的 agentic 推理基准测试，发布来自真实环境、可复现、可审计的性能数据，并获得 OpenAI、Meta、Oracle、Microsoft 等万亿美元级 AI 基础设施运营方的信赖。',
     semianalysis: 'SemiAnalysis',
-    mainSite: '官方网站',
+    mainSite: 'SemiAnalysis 官网',
     newsletter: '订阅通讯',
-    about: '关于我们',
+    about: '关于 SemiAnalysis',
     legal: '法律信息',
-    landAcknowledgement: '土地致谢',
+    landAcknowledgement: '原住民传统领地声明',
     privacyPolicy: '隐私政策',
     cookiePolicy: 'Cookie 政策',
     contribute: '参与贡献',
     benchmarks: '基准测试仓库',
-    frontend: '前端仓库',
+    agentxHarness: 'AgentX 测试框架',
+    visualization: '可视化工具',
     more: '更多',
-    supporters: '支持者',
-    datasets: '数据集',
-    articles: '文章',
-    gpuReliability: 'Chip 可靠性',
-    apiReference: 'API 参考文档',
+    supporters: '业界评价',
+    agentx: 'AgentX',
+    telemetry: '遥测数据',
+    articles: '技术文章',
+    gpuReliability: '芯片可靠性',
+    apiReference: 'API 文档',
     perfPerDollar: '每美元性能',
+    modelArchitectures: '模型架构',
     glossary: 'AI 推理术语表',
-    languageLink: 'English',
-    languageHref: '/',
-    languageHrefLang: 'en',
-    cta: '如果这些数据对您的工作有帮助，欢迎在 GitHub 上为我们加星或分享给您的同事。',
-    rights: '保留所有权利。',
+    chipSpecs: '芯片规格与价格',
+    rankings: 'GPU 排行榜',
+    runPages: '模型在 GPU 上的实测结果',
+    cta: '如果这些数据对您的工作有帮助，欢迎在 GitHub 上点个 Star，或分享给同事。',
+    rights: '版权所有。',
   },
 } as const;
 
@@ -106,7 +112,10 @@ export const Footer = ({ starCount }: { starCount?: number | null }) => {
           </div>
 
           {/* Center — Links */}
-          <div data-testid="footer-links" className="grid grid-cols-3 gap-x-6 gap-y-8">
+          <div
+            data-testid="footer-links"
+            className="grid grid-cols-2 xl:grid-cols-[minmax(0,1fr)_minmax(0,1.35fr)_minmax(0,1fr)_minmax(0,2fr)] gap-x-6 gap-y-8 break-words hyphens-auto min-w-0"
+          >
             <div data-testid="footer-links-semianalysis" className="flex flex-col gap-2.5">
               <span className="text-sm font-medium text-foreground">{t.semianalysis}</span>
               <a
@@ -177,78 +186,124 @@ export const Footer = ({ starCount }: { starCount?: number | null }) => {
                 {t.benchmarks}
               </a>
               <a
-                data-testid="footer-link-frontend"
+                data-testid="footer-link-agentx-harness"
+                href="https://github.com/SemiAnalysisAI/agentx-harness"
+                target="_blank"
+                rel="noopener noreferrer"
+                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+              >
+                {t.agentxHarness}
+              </a>
+              <a
+                data-testid="footer-link-visualization"
                 href="https://github.com/SemiAnalysisAI/InferenceX-app"
                 target="_blank"
                 rel="noopener noreferrer"
                 className="text-sm text-muted-foreground hover:text-foreground transition-colors"
               >
-                {t.frontend}
+                {t.visualization}
               </a>
             </div>
-            <div data-testid="footer-links-more" className="flex flex-col gap-2.5">
+            {/* "More" holds many links, so it gets a double-width column with
+                the links flowing in two sub-columns. This keeps every group on
+                one balanced row instead of wrapping below the others. */}
+            <div
+              data-testid="footer-links-more"
+              className="col-span-2 xl:col-span-1 flex flex-col gap-2.5"
+            >
               <span className="text-sm font-medium text-foreground">{t.more}</span>
-              <Link
-                data-testid="footer-link-supporters"
-                href={`${prefix}/quotes`}
-                onClick={() => track('footer_supporters_clicked')}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.supporters}
-              </Link>
-              <Link
-                data-testid="footer-link-datasets"
-                href={`${prefix}/datasets`}
-                onClick={() => track('footer_datasets_clicked')}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.datasets}
-              </Link>
-              <Link
-                data-testid="footer-link-articles"
-                href={`${prefix}/blog`}
-                onClick={() => track('footer_articles_clicked')}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.articles}
-              </Link>
-              <Link
-                data-testid="footer-link-api"
-                href={`${prefix}/api`}
-                onClick={() => track('footer_api_clicked')}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.apiReference}
-              </Link>
-              <Link
-                data-testid="footer-link-reliability"
-                href={`${prefix}/reliability`}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.gpuReliability}
-              </Link>
-              <Link
-                data-testid="footer-link-compare-per-dollar"
-                href={`${prefix}/compare-per-dollar`}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.perfPerDollar}
-              </Link>
-              <Link
-                data-testid="footer-link-glossary"
-                href={`${prefix}/glossary`}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.glossary}
-              </Link>
-              <Link
-                data-testid="footer-link-zh"
-                href={t.languageHref}
-                hrefLang={t.languageHrefLang}
-                className="text-sm text-muted-foreground hover:text-foreground transition-colors"
-              >
-                {t.languageLink}
-              </Link>
+              <div className="grid grid-cols-2 gap-x-6 gap-y-2.5">
+                <Link
+                  data-testid="footer-link-supporters"
+                  href={`${prefix}/quotes`}
+                  onClick={() => track('footer_supporters_clicked')}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.supporters}
+                </Link>
+                <Link
+                  data-testid="footer-link-agentx"
+                  href={`${prefix}/agentx`}
+                  onClick={() => track('footer_agentx_clicked')}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.agentx}
+                </Link>
+                <Link
+                  data-testid="footer-link-telemetry"
+                  href={`${prefix}/inference/agentic`}
+                  onClick={() => track('footer_telemetry_clicked')}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.telemetry}
+                </Link>
+                <Link
+                  data-testid="footer-link-articles"
+                  href={`${prefix}/blog`}
+                  onClick={() => track('footer_articles_clicked')}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.articles}
+                </Link>
+                <Link
+                  data-testid="footer-link-api"
+                  href={`${prefix}/api`}
+                  onClick={() => track('footer_api_clicked')}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.apiReference}
+                </Link>
+                <Link
+                  data-testid="footer-link-reliability"
+                  href={`${prefix}/reliability`}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.gpuReliability}
+                </Link>
+                <Link
+                  data-testid="footer-link-compare-per-dollar"
+                  href={`${prefix}/compare-per-dollar`}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.perfPerDollar}
+                </Link>
+                <Link
+                  data-testid="footer-link-model-architectures"
+                  // English-only route (not zh-mirrored), so no locale prefix.
+                  href="/model"
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.modelArchitectures}
+                </Link>
+                <Link
+                  data-testid="footer-link-glossary"
+                  href={`${prefix}/glossary`}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.glossary}
+                </Link>
+                <Link
+                  data-testid="footer-link-chips"
+                  href={`${prefix}/chips`}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.chipSpecs}
+                </Link>
+                <Link
+                  data-testid="footer-link-rankings"
+                  href={`${prefix}/rankings`}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.rankings}
+                </Link>
+                <Link
+                  data-testid="footer-link-run"
+                  href={`${prefix}/run`}
+                  className="text-sm text-muted-foreground hover:text-foreground transition-colors"
+                >
+                  {t.runPages}
+                </Link>
+              </div>
             </div>
           </div>
 
diff --git a/packages/app/src/components/glossary/glossary-browser.tsx b/packages/app/src/components/glossary/glossary-browser.tsx
index 496790169..72f6a2624 100644
--- a/packages/app/src/components/glossary/glossary-browser.tsx
+++ b/packages/app/src/components/glossary/glossary-browser.tsx
@@ -265,7 +265,7 @@ export function GlossaryBrowser({
                             </span>
                           )}
                         </div>
-                        <p className="mt-1 text-[0.68rem] font-semibold tracking-[0.14em] text-muted-foreground uppercase">
+                        <p className="mt-1 text-2xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
                           {categoryLabels?.[entry.category] ?? entry.category}
                         </p>
                       </div>
@@ -287,7 +287,7 @@ export function GlossaryBrowser({
         </div>
       ) : (
         <div className="border-t border-border/50 px-6 py-20 text-center">
-          <p className="font-mono text-xs tracking-[0.18em] text-brand uppercase">
+          <p className="font-mono text-xs tracking-eyebrow text-brand uppercase">
             {labels.noMatch}
           </p>
           <h2 className="mt-3 text-2xl font-semibold">{labels.noResultsTitle}</h2>
diff --git a/packages/app/src/components/gpu-power/GpuPowerDisplay.tsx b/packages/app/src/components/gpu-power/GpuPowerDisplay.tsx
index ee2919392..e6206c386 100644
--- a/packages/app/src/components/gpu-power/GpuPowerDisplay.tsx
+++ b/packages/app/src/components/gpu-power/GpuPowerDisplay.tsx
@@ -1,10 +1,11 @@
 'use client';
 
+import { useQuery } from '@tanstack/react-query';
 import { track } from '@/lib/analytics';
 import * as d3 from 'd3';
 import { BarChart3, Check, Link as LinkIcon, Lock, Loader2, ScatterChart } from 'lucide-react';
 import { useRouter } from 'next/navigation';
-import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import { useCallback, useMemo, useState } from 'react';
 
 import { Button } from '@/components/ui/button';
 import { Card } from '@/components/ui/card';
@@ -22,17 +23,17 @@ import {
   SelectValue,
 } from '@/components/ui/select';
 
+import { localePath } from '@/lib/i18n';
 import { relockFeatureGate } from '@/lib/use-feature-gate';
 import { useLocale } from '@/lib/use-locale';
+import { useClientSearchParams } from '@/hooks/useClientSearch';
 
 import GpuCorrelationChart from './GpuCorrelationChart';
 import GpuMetricsChart from './GpuPowerChart';
 import GpuStatsTable from './GpuStatsTable';
 import {
   type GpuMetricKey,
-  type GpuMetricsArtifact,
   type GpuPowerApiResponse,
-  type GpuPowerRunInfo,
   ALL_METRIC_OPTIONS,
   getAvailableMetrics,
 } from './types';
@@ -70,7 +71,7 @@ const STRINGS = {
   zh: {
     heading: 'PowerX',
     descPre: '输入 GitHub Actions 运行 ID，可视化',
-    descPost: '产物中 Chip 指标的时间变化趋势。',
+    descPost: '产物中芯片指标的时间变化趋势。',
     relockButton: '重新锁定功能入口',
     runIdLabel: '运行 ID',
     runIdPlaceholder: '例如 22806827144',
@@ -91,7 +92,7 @@ const STRINGS = {
     metricCorrelation: '指标相关性',
     resetFilter: '重置筛选',
     downsample: '降采样',
-    perGpuStats: '每 Chip 统计信息',
+    perGpuStats: '每芯片统计信息',
     rows: '行',
   },
 } as const;
@@ -112,147 +113,167 @@ const GPU_METRICS_VIEW_OPTIONS: SegmentedToggleOption<GpuMetricsView>[] = [
     title: 'Correlation scatter',
   },
 ];
+async function fetchGpuPowerRun(runId: string, signal: AbortSignal): Promise<GpuPowerApiResponse> {
+  const response = await fetch(`/api/gpu-metrics?runId=${encodeURIComponent(runId)}`, {
+    cache: 'no-store',
+    signal,
+  });
+  const result = (await response.json()) as GpuPowerApiResponse | { error: string };
+  if (!response.ok) {
+    throw new Error('error' in result ? result.error : 'Failed to fetch chip metrics');
+  }
+  return result as GpuPowerApiResponse;
+}
 
 export default function GpuMetricsDisplay() {
   const router = useRouter();
-  const t = STRINGS[useLocale()];
-  const [runIdInput, setRunIdInput] = useState('22806827144');
-  const [loading, setLoading] = useState(false);
-  const [error, setError] = useState<string | null>(null);
-  const [artifacts, setArtifacts] = useState<GpuMetricsArtifact[]>([]);
-  const [selectedArtifact, setSelectedArtifact] = useState<string>('');
-  const [selectedMetric, setSelectedMetric] = useState<GpuMetricKey>('power');
-  const [runInfo, setRunInfo] = useState<GpuPowerRunInfo | null>(null);
-  const [visibleGpus, setVisibleGpus] = useState<Set<number>>(new Set());
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const searchParams = useClientSearchParams();
+  const searchKey = searchParams.toString();
+  const urlRunId = searchParams.get('gm_runId')?.trim() || null;
+  const [runIdDraft, setRunIdDraft] = useState<{ searchKey: string; value: string } | null>(null);
+  const runIdInput =
+    runIdDraft?.searchKey === searchKey ? runIdDraft.value : (urlRunId ?? '22806827144');
+  const [submittedRun, setSubmittedRun] = useState<{
+    searchKey: string;
+    runId: string;
+  } | null>(null);
+  const requestedRunId = submittedRun?.searchKey === searchKey ? submittedRun.runId : urlRunId;
+  const query = useQuery({
+    queryKey: ['gpu-power-run', requestedRunId] as const,
+    queryFn: ({ signal }) => fetchGpuPowerRun(requestedRunId!, signal),
+    enabled: Boolean(requestedRunId),
+    staleTime: 0,
+    gcTime: 0,
+    retry: false,
+    refetchOnMount: 'always',
+  });
+  const artifacts = query.data?.artifacts ?? [];
+  const runInfo = query.data?.runInfo ?? null;
+  const loading = Boolean(requestedRunId) && query.isFetching;
+  const error = query.error instanceof Error ? query.error.message : null;
+
+  const [selection, setSelection] = useState<{
+    searchKey: string;
+    runId: string;
+    artifact?: string;
+    metric?: GpuMetricKey;
+  } | null>(null);
+  const selectionApplies = selection?.searchKey === searchKey && selection.runId === requestedRunId;
+  const selectedArtifactCandidate = selectionApplies
+    ? selection.artifact
+    : (searchParams.get('gm_artifact') ?? undefined);
+  const selectedArtifact =
+    selectedArtifactCandidate &&
+    artifacts.some((artifact) => artifact.name === selectedArtifactCandidate)
+      ? selectedArtifactCandidate
+      : (artifacts[0]?.name ?? '');
+  const currentData = useMemo(
+    () => artifacts.find((artifact) => artifact.name === selectedArtifact)?.data ?? [],
+    [artifacts, selectedArtifact],
+  );
+  const availableMetrics = useMemo(() => getAvailableMetrics(currentData), [currentData]);
+  const urlMetric = searchParams.get('gm_metric');
+  const selectedMetricCandidate = selectionApplies ? selection.metric : urlMetric;
+  const selectedMetric =
+    selectedMetricCandidate &&
+    ALL_METRIC_OPTIONS.some((metric) => metric.key === selectedMetricCandidate) &&
+    availableMetrics.some((metric) => metric.key === selectedMetricCandidate)
+      ? (selectedMetricCandidate as GpuMetricKey)
+      : 'power';
+
+  const [gpuSelection, setGpuSelection] = useState<{
+    scopeKey: string;
+    values: Set<number>;
+  } | null>(null);
+  const allGpuIndices = useMemo(
+    () => [...new Set(currentData.map((datum) => datum.index))].toSorted((a, b) => a - b),
+    [currentData],
+  );
+  const gpuScopeKey = `${requestedRunId ?? ''}|${selectedArtifact}`;
+  const visibleGpus = useMemo(
+    () =>
+      gpuSelection?.scopeKey === gpuScopeKey
+        ? new Set(gpuSelection.values)
+        : new Set(allGpuIndices),
+    [gpuSelection, gpuScopeKey, allGpuIndices],
+  );
+
   const [copied, setCopied] = useState(false);
   const [isLegendExpanded, setIsLegendExpanded] = useState(true);
   const [downsample, setDownsample] = useState(true);
-  // View toggle + correlation
   const [chartView, setChartView] = useState<GpuMetricsView>('chart');
   const [corrXMetric, setCorrXMetric] = useState<GpuMetricKey>('power');
   const [corrYMetric, setCorrYMetric] = useState<GpuMetricKey>('temperature');
-  // URL state
-  const pendingUrlState = useRef<{ artifact?: string; metric?: string } | null>(null);
-
-  const loadRun = useCallback(
-    async (runId: string, urlDefaults?: { artifact?: string; metric?: string }) => {
-      track('gpu_metrics_load_run', { runId });
-      setLoading(true);
-      setError(null);
-      if (urlDefaults) {
-        pendingUrlState.current = urlDefaults;
-      }
-      try {
-        const response = await fetch(`/api/gpu-metrics?runId=${encodeURIComponent(runId)}`);
-        const result: GpuPowerApiResponse | { error: string } = await response.json();
-        if (!response.ok) {
-          throw new Error('error' in result ? result.error : 'Failed to fetch chip metrics');
-        }
-        const apiResult = result as GpuPowerApiResponse;
-        setArtifacts(apiResult.artifacts);
-        setRunInfo(apiResult.runInfo);
-
-        const pending = pendingUrlState.current;
-        const targetArtifact =
-          pending?.artifact && apiResult.artifacts.some((a) => a.name === pending.artifact)
-            ? pending.artifact
-            : (apiResult.artifacts[0]?.name ?? '');
-        setSelectedArtifact(targetArtifact);
-
-        if (pending?.metric && ALL_METRIC_OPTIONS.some((m) => m.key === pending.metric)) {
-          setSelectedMetric(pending.metric as GpuMetricKey);
-        }
-        pendingUrlState.current = null;
-
-        const targetData = apiResult.artifacts.find((a) => a.name === targetArtifact)?.data ?? [];
-        const gpuIndices = new Set(targetData.map((d) => d.index));
-        setVisibleGpus(gpuIndices);
-      } catch (caughtError) {
-        setError(caughtError instanceof Error ? caughtError.message : 'Unknown error');
-        setArtifacts([]);
-        setRunInfo(null);
-        setSelectedArtifact('');
-        pendingUrlState.current = null;
-      } finally {
-        setLoading(false);
-      }
-    },
-    [],
-  );
 
   const handleLoad = useCallback(() => {
-    const trimmed = runIdInput.trim();
-    if (!trimmed) return;
-    loadRun(trimmed);
-  }, [runIdInput, loadRun]);
-
-  useEffect(() => {
-    const params = new URLSearchParams(window.location.search);
-    const urlRunId = params.get('gm_runId');
-    if (!urlRunId) return;
-    setRunIdInput(urlRunId);
-    loadRun(urlRunId, {
-      artifact: params.get('gm_artifact') ?? undefined,
-      metric: params.get('gm_metric') ?? undefined,
-    });
-  }, [loadRun]);
+    const runId = runIdInput.trim();
+    if (!runId) return;
+    track('gpu_metrics_load_run', { runId });
+    if (runId === requestedRunId) {
+      void query.refetch();
+      return;
+    }
+    setSubmittedRun({ searchKey, runId });
+  }, [runIdInput, requestedRunId, query, searchKey]);
 
   const handleArtifactChange = useCallback(
     (name: string) => {
       track('gpu_metrics_artifact_selected', { artifact: name });
-      setSelectedArtifact(name);
-      const artifact = artifacts.find((a) => a.name === name);
-      if (artifact) {
-        setVisibleGpus(new Set(artifact.data.map((d) => d.index)));
-        // Reset metric if current selection isn't available in new artifact
-        const available = getAvailableMetrics(artifact.data);
-        setSelectedMetric((prev) => (available.some((m) => m.key === prev) ? prev : 'power'));
-      }
+      const artifact = artifacts.find((candidate) => candidate.name === name);
+      const metrics = getAvailableMetrics(artifact?.data ?? []);
+      setSelection({
+        searchKey,
+        runId: requestedRunId ?? '',
+        artifact: name,
+        metric: metrics.some((metric) => metric.key === selectedMetric) ? selectedMetric : 'power',
+      });
     },
-    [artifacts],
+    [artifacts, requestedRunId, searchKey, selectedMetric],
   );
 
-  const handleMetricChange = useCallback((value: string) => {
-    track('gpu_metrics_metric_changed', { metric: value });
-    setSelectedMetric(value as GpuMetricKey);
-  }, []);
-
-  const currentData = useMemo(
-    () => artifacts.find((a) => a.name === selectedArtifact)?.data ?? [],
-    [artifacts, selectedArtifact],
+  const handleMetricChange = useCallback(
+    (value: string) => {
+      track('gpu_metrics_metric_changed', { metric: value });
+      setSelection({
+        searchKey,
+        runId: requestedRunId ?? '',
+        artifact: selectedArtifact,
+        metric: value as GpuMetricKey,
+      });
+    },
+    [requestedRunId, searchKey, selectedArtifact],
   );
 
-  const availableMetrics = useMemo(() => getAvailableMetrics(currentData), [currentData]);
-
-  const toggleGpu = useCallback((gpuIndex: number) => {
-    track('gpu_metrics_gpu_toggled', { gpuIndex });
-    setVisibleGpus((prev) => {
-      const next = new Set(prev);
-      if (next.has(gpuIndex)) {
-        next.delete(gpuIndex);
-      } else {
-        next.add(gpuIndex);
-      }
-      return next;
-    });
-  }, []);
+  const toggleGpu = useCallback(
+    (gpuIndex: number) => {
+      track('gpu_metrics_gpu_toggled', { gpuIndex });
+      const next = new Set(visibleGpus);
+      if (next.has(gpuIndex)) next.delete(gpuIndex);
+      else next.add(gpuIndex);
+      setGpuSelection({ scopeKey: gpuScopeKey, values: next });
+    },
+    [gpuScopeKey, visibleGpus],
+  );
 
-  const removeGpu = useCallback((hw: string) => {
-    setVisibleGpus((prev) => {
-      const next = new Set(prev);
+  const removeGpu = useCallback(
+    (hw: string) => {
+      const next = new Set(visibleGpus);
       next.delete(Number(hw));
-      return next;
-    });
-  }, []);
+      setGpuSelection({ scopeKey: gpuScopeKey, values: next });
+    },
+    [gpuScopeKey, visibleGpus],
+  );
 
   const handleShare = useCallback(async () => {
     const params = new URLSearchParams();
-    params.set('gm_runId', runIdInput.trim());
+    params.set('gm_runId', requestedRunId ?? runIdInput.trim());
     if (selectedArtifact) params.set('gm_artifact', selectedArtifact);
     if (selectedMetric !== 'power') params.set('gm_metric', selectedMetric);
     const url = `${window.location.origin}${window.location.pathname}?${params.toString()}#gpu-metrics`;
     track('gpu_metrics_share_link_copied', {
-      runId: runIdInput.trim(),
+      runId: requestedRunId ?? runIdInput.trim(),
       artifact: selectedArtifact,
       metric: selectedMetric,
     });
@@ -269,21 +290,15 @@ export default function GpuMetricsDisplay() {
     setCopied(true);
     setTimeout(() => setCopied(false), 2000);
     window.dispatchEvent(new CustomEvent('inferencex:action'));
-  }, [runIdInput, selectedArtifact, selectedMetric]);
-
-  const metricConfig = ALL_METRIC_OPTIONS.find((m) => m.key === selectedMetric)!;
-
-  const allGpuIndices = useMemo(
-    () => [...new Set(currentData.map((d) => d.index))].toSorted((a, b) => a - b),
-    [currentData],
-  );
+  }, [requestedRunId, runIdInput, selectedArtifact, selectedMetric]);
 
+  const metricConfig = ALL_METRIC_OPTIONS.find((metric) => metric.key === selectedMetric)!;
   const allGpusSelected =
-    allGpuIndices.length > 0 && allGpuIndices.every((i) => visibleGpus.has(i));
+    allGpuIndices.length > 0 && allGpuIndices.every((index) => visibleGpus.has(index));
   const selectAllGpus = useCallback(() => {
-    setVisibleGpus(new Set(allGpuIndices));
+    setGpuSelection({ scopeKey: gpuScopeKey, values: new Set(allGpuIndices) });
     track('gpu_metrics_gpu_reset_filter');
-  }, [allGpuIndices]);
+  }, [allGpuIndices, gpuScopeKey]);
 
   const handleChartViewChange = useCallback((value: GpuMetricsView) => {
     setChartView(value);
@@ -311,7 +326,7 @@ export default function GpuMetricsDisplay() {
                 onClick={() => {
                   relockFeatureGate();
                   track('powerx_relocked');
-                  router.push('/inference');
+                  router.push(localePath('/inference', locale));
                 }}
                 title="Re-lock feature gate"
               >
@@ -329,7 +344,7 @@ export default function GpuMetricsDisplay() {
                 data-testid="gpu-metrics-run-input"
                 placeholder={t.runIdPlaceholder}
                 value={runIdInput}
-                onChange={(e) => setRunIdInput(e.target.value)}
+                onChange={(event) => setRunIdDraft({ searchKey, value: event.target.value })}
                 onKeyDown={(e) => {
                   if (e.key === 'Enter') handleLoad();
                 }}
diff --git a/packages/app/src/components/gpu-specs/gpu-specs-content.tsx b/packages/app/src/components/gpu-specs/gpu-specs-content.tsx
index c6b0b9267..4b6b29acd 100644
--- a/packages/app/src/components/gpu-specs/gpu-specs-content.tsx
+++ b/packages/app/src/components/gpu-specs/gpu-specs-content.tsx
@@ -57,7 +57,7 @@ function VendorBadge({ vendor }: { vendor: GpuSpec['vendor'] }) {
   const isNvidia = vendor === 'nvidia';
   return (
     <span
-      className={`inline-flex items-center rounded-md px-1.5 py-0.5 text-[10px] font-medium ${
+      className={`inline-flex items-center rounded-md px-1.5 py-0.5 text-3xs font-medium ${
         isNvidia ? 'bg-[#76b900]/15 text-[#76b900]' : 'bg-[#ed1c24]/15 text-[#ed1c24]'
       }`}
     >
@@ -99,22 +99,22 @@ const STRINGS = {
       'Intra-node scale-up interconnect topology for each Chip SKU, showing Chip → NVSwitch or direct Chip-to-Chip connectivity.',
   },
   zh: {
-    heading: 'Chip 规格',
-    description: 'InferenceX™ 基准测试中使用的 Chip 硬件规格，包括计算性能、显存带宽和互联详情。',
+    heading: '芯片规格',
+    description: 'InferenceX™ 基准测试中使用的芯片硬件规格，包括计算性能、显存带宽和互联详情。',
     viewTable: '表格',
     viewChart: '图表',
     viewRadar: '雷达图',
-    colGpu: 'Chip',
+    colGpu: '芯片',
     colMemory: '显存',
     colMemBw: '显存带宽',
     colScaleUp: '纵向扩展',
     colScaleUpBw: '纵向扩展带宽',
-    colWorldSize: '域内 Chip 数',
+    colWorldSize: '域内芯片数',
     colScaleUpDomainMem: '纵向扩展域显存',
     colScaleUpDomainMemBw: '纵向扩展域显存带宽',
     colScaleUpTopology: '纵向扩展拓扑',
     colScaleUpSwitch: '纵向扩展交换机',
-    colScaleOutBwPerGpu: '每 Chip 横向扩展带宽',
+    colScaleOutBwPerGpu: '每芯片横向扩展带宽',
     colScaleOutTech: '横向扩展技术',
     colScaleOutSwitch: '横向扩展交换机',
     colScaleOutTopology: '横向扩展拓扑',
@@ -122,9 +122,9 @@ const STRINGS = {
     footnote1: '密集 Tensor Core 峰值 TFLOP/s（不含稀疏加速）。',
     footnote2: 'InferenceX™ 机柜级测试不使用横向扩展。',
     scaleOutHeading: '横向扩展拓扑图',
-    scaleOutDescription: '每台服务器的横向扩展网络拓扑，展示 Chip → NIC → Leaf 交换机的连接方式。',
+    scaleOutDescription: '每台服务器的横向扩展网络拓扑，展示芯片 → NIC → Leaf 交换机的连接方式。',
     scaleUpHeading: '纵向扩展拓扑图',
-    scaleUpDescription: '节点内纵向扩展互联拓扑，展示 Chip → NVSwitch 或 Chip 直连方式。',
+    scaleUpDescription: '节点内纵向扩展互联拓扑，展示芯片 → NVSwitch 或芯片直连方式。',
   },
 } as const;
 
diff --git a/packages/app/src/components/gpu-specs/scale-up-topology-diagram.tsx b/packages/app/src/components/gpu-specs/scale-up-topology-diagram.tsx
index d7bae1e30..6b60a97c1 100644
--- a/packages/app/src/components/gpu-specs/scale-up-topology-diagram.tsx
+++ b/packages/app/src/components/gpu-specs/scale-up-topology-diagram.tsx
@@ -113,7 +113,7 @@ export const ScaleUpTopologyDiagram = forwardRef<
         aria-label={`Expand ${spec.name} scale-up topology diagram`}
       >
         <ScaleUpTopologyD3 spec={spec} config={compactConfig} compact />
-        <p className="text-[10px] text-muted-foreground mt-1 text-center">Click to expand</p>
+        <p className="text-3xs text-muted-foreground mt-1 text-center">Click to expand</p>
       </button>
 
       <Dialog open={open} onOpenChange={setOpen}>
diff --git a/packages/app/src/components/gpu-specs/topology-diagram.tsx b/packages/app/src/components/gpu-specs/topology-diagram.tsx
index ec4442eb8..6df7fb5e5 100644
--- a/packages/app/src/components/gpu-specs/topology-diagram.tsx
+++ b/packages/app/src/components/gpu-specs/topology-diagram.tsx
@@ -111,7 +111,7 @@ export const TopologyDiagram = forwardRef<
         aria-label={`Expand ${spec.name} topology diagram`}
       >
         <TopologyD3 spec={spec} config={compactConfig} compact />
-        <p className="text-[10px] text-muted-foreground mt-1 text-center">Click to expand</p>
+        <p className="text-3xs text-muted-foreground mt-1 text-center">Click to expand</p>
       </button>
 
       <Dialog open={open} onOpenChange={setOpen}>
diff --git a/packages/app/src/components/header/header.tsx b/packages/app/src/components/header/header.tsx
index a8c58186e..4b8608df6 100644
--- a/packages/app/src/components/header/header.tsx
+++ b/packages/app/src/components/header/header.tsx
@@ -7,30 +7,42 @@ import { useCallback, useEffect, useRef, useState } from 'react';
 import { track } from '@/lib/analytics';
 
 import { ModeToggle } from '@/components/ui/mode-toggle';
+import { NewBadge } from '@/components/ui/new-badge';
 import { MinecraftToggles } from '@/components/minecraft/minecraft-toggles';
-import { CLIENT_SEARCH_CHANGE_EVENT, navigateInApp } from '@/lib/client-navigation';
+import { navigateInApp } from '@/lib/client-navigation';
+import { DASHBOARD_ROUTES } from '@/lib/dashboard-routes';
+import { useClientPathname } from '@/hooks/useClientPathname';
+import { useClientSearch } from '@/hooks/useClientSearch';
 import { hasZhSibling, isZhPathname, switchLocalePath, ZH_PREFIX, zhPath } from '@/lib/i18n';
-import { NAV_LABELS_ZH } from '@/lib/tab-meta-zh';
+import { NAV_LABELS_ZH, type HeaderNavHref } from '@/lib/tab-meta-zh';
 import { cn } from '@/lib/utils';
 
 import { GitHubStars } from './GithubStars';
 
-/** Dashboard tab paths that should highlight the "Dashboard" nav link. */
-const DASHBOARD_TABS = [
-  '/inference',
-  '/evaluation',
-  '/historical',
-  '/calculator',
-  '/reliability',
-  '/gpu-specs',
-  '/gpu-metrics',
-  '/collectivex',
-  '/submissions',
-  '/current-inferencex-image',
-];
+const DASHBOARD_TABS = DASHBOARD_ROUTES.map((route) => route.path);
 
-const NAV_LINKS = [
+interface NavLink {
+  href: HeaderNavHref;
+  label: string;
+  testId: string;
+  event: string;
+  badge?: {
+    en: string;
+    zh: string;
+  };
+}
+
+const NAV_LINKS: readonly NavLink[] = [
   { href: '/', label: 'Home', testId: 'nav-link-home', event: 'header_home_clicked' },
+  // AgentX sits directly after Home: it is the flagship benchmark, and the
+  // surface every entry below it ultimately explains.
+  {
+    href: '/agentx',
+    label: 'AgentX',
+    testId: 'nav-link-agentx',
+    event: 'header_agentx_clicked',
+    badge: { en: 'NEW', zh: '新' },
+  },
   {
     href: '/overview',
     label: 'Overview',
@@ -49,6 +61,12 @@ const NAV_LINKS = [
     testId: 'nav-link-compare',
     event: 'header_compare_clicked',
   },
+  {
+    href: '/blog',
+    label: 'Articles',
+    testId: 'nav-link-articles',
+    event: 'header_articles_clicked',
+  },
   { href: '/about', label: 'About', testId: 'nav-link-about', event: 'header_about_clicked' },
 ] as const;
 
@@ -61,7 +79,11 @@ function isActive(pathname: string, href: string): boolean {
       : pathname.slice(ZH_PREFIX.length)
     : pathname;
   if (href === '/') return enPathname === '/';
-  if (href === '/inference') return DASHBOARD_TABS.some((tab) => enPathname.startsWith(tab));
+  // Dashboard owns every tab path, including the telemetry catalog beneath
+  // `/inference`, which now lives in the footer rather than the primary nav.
+  if (href === '/inference') {
+    return DASHBOARD_TABS.some((tab) => enPathname.startsWith(tab));
+  }
   // Exact match or a child path under `<href>/...`. The bare `startsWith` would
   // light up `/compare` when the user is on `/compare-per-dollar/...` since the
   // latter starts with the literal string `/compare`.
@@ -81,47 +103,33 @@ function isCurrentPage(pathname: string, displayHref: string): boolean {
 
 /** EN ↔ 中文 switcher; maps the current page to its sibling in the other language. */
 function LanguageToggle({
-  pathname,
+  pathname: routerPathname,
   router,
 }: {
   pathname: string;
   router: ReturnType<typeof useRouter>;
 }) {
+  // The toggle maps the CURRENT address to its sibling, so it must see the
+  // live pathname — per-model dashboard routes rewrite the URL with
+  // replaceClientPathname on model switches, which usePathname ignores.
+  const pathname = useClientPathname(routerPathname);
   const isZh = isZhPathname(pathname);
   const target = switchLocalePath(pathname);
-  // The root layout is reused for search-only App Router transitions, so keep
-  // this persistent link synchronized with both browser history and the app's
-  // explicit soft-navigation signal.
-  const [search, setSearch] = useState('');
-  useEffect(() => {
-    const sync = (event: Event) => {
-      setSearch(
-        event instanceof CustomEvent && typeof event.detail === 'string'
-          ? event.detail
-          : window.location.search,
-      );
-    };
-    sync(new Event('initial'));
-    window.addEventListener('popstate', sync);
-    window.addEventListener(CLIENT_SEARCH_CHANGE_EVENT, sync);
-    return () => {
-      window.removeEventListener('popstate', sync);
-      window.removeEventListener(CLIENT_SEARCH_CHANGE_EVENT, sync);
-    };
-  }, [pathname]);
+  const search = useClientSearch();
+  const isOverview = isActive(pathname, '/overview');
   return (
     <Link
       href={target + search}
       // Only /overview rewrites this href per interaction, which would
       // re-prefetch its force-dynamic sibling on every selector commit.
       // Everywhere else the href is stable, so let Next prefetch it.
-      prefetch={isActive(pathname, '/overview') ? false : undefined}
+      prefetch={isOverview ? false : undefined}
       data-testid="language-toggle"
       hrefLang={isZh ? 'en' : 'zh-CN'}
       className="inline-flex items-center min-h-11 px-2 rounded-md text-sm font-medium text-muted-foreground hover:text-foreground hover:bg-muted transition-colors whitespace-nowrap"
       onClick={(event) => {
         track('header_language_toggled', { to: isZh ? 'en' : 'zh' });
-        navigateInApp(event, router, target + search);
+        if (!isOverview) navigateInApp(event, router, target + search);
       }}
     >
       {isZh ? 'EN' : '中文'}
@@ -142,9 +150,14 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
     ? NAV_LINKS.map((link) => ({
         ...link,
         label: NAV_LABELS_ZH[link.href] ?? link.label,
+        badgeLabel: link.badge?.zh,
         displayHref: hasZhSibling(link.href) ? zhPath(link.href) : link.href,
       }))
-    : NAV_LINKS.map((link) => ({ ...link, displayHref: link.href }));
+    : NAV_LINKS.map((link) => ({
+        ...link,
+        badgeLabel: link.badge?.en,
+        displayHref: link.href,
+      }));
 
   // Close menu on route change
   useEffect(() => {
@@ -202,15 +215,15 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
           </Link>
 
           {/* Desktop nav */}
-          <nav className="hidden lg:flex items-center gap-1">
-            {navLinks.map(({ href, displayHref, label, testId, event }) => (
+          <nav className="hidden items-center gap-1 xl:flex">
+            {navLinks.map(({ href, displayHref, label, badgeLabel, testId, event }) => (
               <Link
                 key={href}
                 data-testid={testId}
                 href={displayHref}
                 prefetch={isActive(pathname, href) ? false : undefined}
                 className={cn(
-                  'px-3 py-1.5 rounded-md text-sm font-medium transition-colors',
+                  'inline-flex items-center gap-1.5 px-3 py-1.5 rounded-md text-sm font-medium transition-colors',
                   isActive(pathname, href)
                     ? 'text-brand bg-brand/10'
                     : 'text-muted-foreground hover:text-foreground hover:bg-muted',
@@ -228,7 +241,12 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
                   }
                 }}
               >
-                {label}
+                <span>{label}</span>
+                {badgeLabel && (
+                  <NewBadge data-nav-badge="agentx" data-new-badge="agentx-nav">
+                    {badgeLabel}
+                  </NewBadge>
+                )}
               </Link>
             ))}
           </nav>
@@ -247,7 +265,7 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
             <ModeToggle />
 
             {/* Mobile hamburger */}
-            <div ref={menuRef} className="relative lg:hidden">
+            <div ref={menuRef} className="relative xl:hidden">
               <button
                 type="button"
                 data-testid="mobile-menu-toggle"
@@ -276,7 +294,7 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
                   data-testid="mobile-menu"
                   className="absolute right-0 top-full mt-2 z-50 flex flex-col rounded-lg border border-border bg-background p-1.5 shadow-lg min-w-40"
                 >
-                  {navLinks.map(({ href, displayHref, label, event }) => (
+                  {navLinks.map(({ href, displayHref, label, badgeLabel, event }) => (
                     <Link
                       key={href}
                       href={displayHref}
@@ -298,7 +316,16 @@ export const Header = ({ starCount }: { starCount?: number | null }) => {
                         }
                       }}
                     >
-                      {label}
+                      <span>{label}</span>
+                      {badgeLabel && (
+                        <NewBadge
+                          data-nav-badge="agentx"
+                          data-new-badge="agentx-nav"
+                          className="ml-1.5"
+                        >
+                          {badgeLabel}
+                        </NewBadge>
+                      )}
                     </Link>
                   ))}
                   <span className="flex items-center gap-2 px-3 sm:hidden">
diff --git a/packages/app/src/components/inference/InferenceContext.test.ts b/packages/app/src/components/inference/InferenceContext.test.ts
new file mode 100644
index 000000000..5795a0f93
--- /dev/null
+++ b/packages/app/src/components/inference/InferenceContext.test.ts
@@ -0,0 +1,147 @@
+// @vitest-environment jsdom
+import { act, createElement, memo, useMemo, useState } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import {
+  InferenceContextsProvider,
+  resolveE2eXAxisMetric,
+  resolveEffectiveXAxisMode,
+  useInferenceActions,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
+import { Sequence } from '@/lib/data-mappings';
+import type {
+  InferenceActionsContextType,
+  InferenceDataContextType,
+  InferenceDisplayContextType,
+  InferenceFiltersContextType,
+} from '@/components/inference/types';
+
+describe('inference requested and effective axis selectors', () => {
+  it('preserves URL mode while sequence availability is unresolved', () => {
+    expect(
+      resolveEffectiveXAxisMode('e2e-normalized-interactivity', Sequence.EightK_OneK, false),
+    ).toBe('e2e-normalized-interactivity');
+  });
+
+  it('falls back from an agentic-only mode for a resolved fixed sequence', () => {
+    expect(
+      resolveEffectiveXAxisMode('e2e-normalized-interactivity', Sequence.EightK_OneK, true),
+    ).toBe('interactivity');
+  });
+
+  it('restores the requested mode when the effective sequence supports it', () => {
+    expect(
+      resolveEffectiveXAxisMode('e2e-normalized-interactivity', Sequence.AgenticTraces, true),
+    ).toBe('e2e-normalized-interactivity');
+  });
+
+  it('derives TTFT metric from sequence kind and percentile', () => {
+    expect(resolveE2eXAxisMetric('p90_ttft', 'ttft', Sequence.AgenticTraces, 'p75')).toBe(
+      'p75_ttft',
+    );
+    expect(resolveE2eXAxisMetric('p90_ttft', 'ttft', Sequence.EightK_OneK, 'p75')).toBe(
+      'median_ttft',
+    );
+  });
+
+  it('uses natural E2E x-axis and preserves inactive requested metric', () => {
+    expect(resolveE2eXAxisMetric('p90_ttft', 'e2e', Sequence.AgenticTraces, 'p75')).toBeNull();
+    expect(resolveE2eXAxisMetric('p90_ttft', 'interactivity', Sequence.AgenticTraces, 'p75')).toBe(
+      'p90_ttft',
+    );
+  });
+});
+
+describe('inference domain providers', () => {
+  let container: HTMLDivElement;
+  let root: Root;
+
+  const data = {} as InferenceDataContextType;
+  const filters = {} as InferenceFiltersContextType;
+  const display = { highContrast: false } as InferenceDisplayContextType;
+  const actions = { setHighContrast: vi.fn() } as unknown as InferenceActionsContextType;
+
+  beforeEach(() => {
+    container = document.createElement('div');
+    document.body.append(container);
+    root = createRoot(container);
+  });
+
+  afterEach(() => {
+    act(() => root.unmount());
+    container.remove();
+  });
+
+  it('does not rerender filter-only consumers for a display toggle', () => {
+    const filterRender = vi.fn();
+    let toggleHighContrast!: () => void;
+
+    const FilterConsumer = memo(() => {
+      useInferenceFilters();
+      filterRender();
+      return null;
+    });
+
+    function DisplayToggleHarness() {
+      const [highContrast, setHighContrast] = useState(false);
+      toggleHighContrast = () => setHighContrast((current) => !current);
+      const nextDisplay = useMemo(() => ({ ...display, highContrast }), [highContrast]);
+      return createElement(InferenceContextsProvider, {
+        data,
+        filters,
+        display: nextDisplay,
+        actions,
+        children: createElement(FilterConsumer),
+      });
+    }
+
+    act(() => root.render(createElement(DisplayToggleHarness)));
+    expect(filterRender).toHaveBeenCalledTimes(1);
+
+    act(() => toggleHighContrast());
+    expect(filterRender).toHaveBeenCalledTimes(1);
+  });
+
+  it('keeps the action context and setter identities stable when data arrives', () => {
+    const seenContexts: InferenceActionsContextType[] = [];
+    const seenSetters: InferenceActionsContextType['setHighContrast'][] = [];
+    const invokedRevisions: number[] = [];
+    let publishData!: () => void;
+
+    function ActionConsumer({ revision }: { revision: number }) {
+      const currentActions = useInferenceActions();
+      seenContexts.push(currentActions);
+      seenSetters.push(currentActions.setHighContrast);
+      return createElement('output', null, revision);
+    }
+
+    function DataArrivalHarness() {
+      const [revision, setRevision] = useState(0);
+      publishData = () => setRevision((current) => current + 1);
+      const nextData = useMemo(() => ({ ...data, loading: revision === 0 }), [revision]);
+      const nextActions = {
+        ...actions,
+        setHighContrast: () => invokedRevisions.push(revision),
+      };
+      return createElement(InferenceContextsProvider, {
+        data: nextData,
+        filters,
+        display,
+        actions: nextActions,
+        children: createElement(ActionConsumer, { revision }),
+      });
+    }
+
+    act(() => root.render(createElement(DataArrivalHarness)));
+    act(() => publishData());
+
+    expect(seenContexts).toHaveLength(2);
+    expect(seenContexts[1]).toBe(seenContexts[0]);
+    expect(seenSetters[1]).toBe(seenSetters[0]);
+
+    seenSetters[0](true);
+    expect(invokedRevisions).toEqual([1]);
+  });
+});
diff --git a/packages/app/src/components/inference/InferenceContext.tsx b/packages/app/src/components/inference/InferenceContext.tsx
index 991c583dc..29476d2fa 100644
--- a/packages/app/src/components/inference/InferenceContext.tsx
+++ b/packages/app/src/components/inference/InferenceContext.tsx
@@ -13,6 +13,7 @@ import {
 } from 'react';
 
 import { DISPLAY_MODEL_TO_DB, rowToSequence } from '@semianalysisai/inferencex-constants';
+import type { BenchmarkRow } from '@/lib/api';
 import { track } from '@/lib/analytics';
 import {
   FAVORITE_PRESETS,
@@ -20,9 +21,23 @@ import {
   matchesPresetHwFilter,
 } from '@/components/favorites/favorite-presets';
 
-import { useGlobalFilters } from '@/components/GlobalFilterContext';
+import {
+  useGlobalFilterActions,
+  useGlobalFilterAvailability,
+  useGlobalFilterRun,
+  useGlobalFilterSelection,
+  useGlobalFilterWorkflow,
+} from '@/components/GlobalFilterContext';
 import { useUnofficialRun } from '@/components/unofficial-run-provider';
-import type { InferenceChartContextType, InferenceData } from '@/components/inference/types';
+import type {
+  InferenceActionsContextType,
+  InferenceData,
+  InferenceDataContextType,
+  InferenceDisplayContextType,
+  InferenceFiltersContextType,
+  TokenRevenuePriceSource,
+} from '@/components/inference/types';
+import { resolveMetricConfigKey } from '@/components/inference/metric-registry';
 import { Button } from '@/components/ui/button';
 import {
   Dialog,
@@ -39,10 +54,18 @@ import {
   useUrlStateSync,
 } from '@/hooks/useChartContext';
 import { useUrlState } from '@/hooks/useUrlState';
+import { useOpenRouterPricing } from '@/hooks/api/use-openrouter-pricing';
+import { DEFAULT_Y_AXIS_METRIC } from '@/lib/url-state';
 import { computeToggle } from '@/hooks/useTogglableSet';
 import { buildAvailabilityHwKey } from '@/lib/chart-utils';
 import { getHardwareConfig, getModelSortIndex, isKnownGpu } from '@/lib/constants';
-import { MODEL_PREFIX_MAPPING, sequenceKind } from '@/lib/data-mappings';
+import {
+  getOpenRouterModelId,
+  MODEL_PREFIX_MAPPING,
+  Sequence,
+  sequenceKind,
+} from '@/lib/data-mappings';
+import { NORMALIZED_TOKEN_REVENUE_PRICING } from './token-revenue';
 import {
   EngineComparisonConflictToast,
   type EngineComparisonConflictDetail,
@@ -55,7 +78,6 @@ import {
   type ExclusionConflictPolicy,
 } from '@/lib/exclusion';
 import { filterRunsByModel, getDisplayLabel } from '@/lib/utils';
-
 import {
   isAgenticOnlyXAxisMode,
   useChartData,
@@ -78,15 +100,95 @@ import {
   type SpecMode,
 } from './utils/quickFilters';
 
-/** @internal Exported for test provider wrapping only. */
-export const InferenceContext = createContext<InferenceChartContextType | undefined>(undefined);
+const InferenceDataContext = createContext<InferenceDataContextType | undefined>(undefined);
+const InferenceFiltersContext = createContext<InferenceFiltersContextType | undefined>(undefined);
+const InferenceDisplayContext = createContext<InferenceDisplayContextType | undefined>(undefined);
+const InferenceActionsContext = createContext<InferenceActionsContextType | undefined>(undefined);
+
+function useStableInferenceActions(
+  actions: InferenceActionsContextType,
+): InferenceActionsContextType {
+  const actionsRef = useRef(actions);
+  actionsRef.current = actions;
+
+  return useMemo(() => {
+    const stableActions = {} as InferenceActionsContextType;
+    const writableActions = stableActions as unknown as Record<
+      keyof InferenceActionsContextType,
+      (...args: never[]) => unknown
+    >;
+    for (const key of Object.keys(actions) as (keyof InferenceActionsContextType)[]) {
+      writableActions[key] = (...args: never[]) => {
+        const currentAction = actionsRef.current[key] as (...currentArgs: never[]) => unknown;
+        return currentAction(...args);
+      };
+    }
+    return stableActions;
+  }, []);
+}
+
+/** @internal Shared by the route provider and focused context-isolation tests. */
+export function InferenceContextsProvider({
+  children,
+  data,
+  filters,
+  display,
+  actions,
+}: {
+  children: ReactNode;
+  data: InferenceDataContextType;
+  filters: InferenceFiltersContextType;
+  display: InferenceDisplayContextType;
+  actions: InferenceActionsContextType;
+}) {
+  const stableActions = useStableInferenceActions(actions);
+  return (
+    <InferenceActionsContext.Provider value={stableActions}>
+      <InferenceDataContext.Provider value={data}>
+        <InferenceFiltersContext.Provider value={filters}>
+          <InferenceDisplayContext.Provider value={display}>
+            {children}
+          </InferenceDisplayContext.Provider>
+        </InferenceFiltersContext.Provider>
+      </InferenceDataContext.Provider>
+    </InferenceActionsContext.Provider>
+  );
+}
+
+export function resolveEffectiveXAxisMode(
+  requestedMode: XAxisMode,
+  sequence: Parameters<typeof sequenceKind>[0],
+  sequenceResolved: boolean,
+): XAxisMode {
+  if (!sequenceResolved) return requestedMode;
+  return sequenceKind(sequence) === 'fixed-seq' && isAgenticOnlyXAxisMode(requestedMode)
+    ? 'interactivity'
+    : requestedMode;
+}
+
+export function resolveE2eXAxisMetric(
+  requestedMetric: string | null,
+  mode: XAxisMode,
+  sequence: Parameters<typeof sequenceKind>[0],
+  percentile: string,
+): string | null {
+  if (mode === 'ttft') {
+    return sequenceKind(sequence) === 'agentic' ? `${percentile}_ttft` : 'median_ttft';
+  }
+  if (mode === 'e2e') return null;
+  return requestedMetric;
+}
 
 export function InferenceProvider({
   children,
   activeTab,
   initialActiveHwTypes,
   compareGpuPair,
+  benchmarkQueryScope,
+  initialBenchmarkModel,
+  initialBenchmarkRows,
   initialYAxisMetric,
+  autoSelectAllGpus = false,
 }: {
   children: ReactNode;
   activeTab: string;
@@ -101,43 +203,54 @@ export function InferenceProvider({
    * registry GPU base keys so other hardware never appears on the legend or plots.
    */
   compareGpuPair?: readonly [string, string];
+  /** Isolates pair-filtered compare hydration from canonical dashboard queries. */
+  benchmarkQueryScope?: string;
+  initialBenchmarkModel?: string;
+  initialBenchmarkRows?: BenchmarkRow[];
   /**
    * Initial y-axis metric key when the URL has no `?i_metric=` param. Used by
    * `/compare-per-dollar/[slug]` to default the chart to
    * `y_costh` (Cost per Million Total Tokens — Owning Hyperscaler) instead of
-   * the dashboard's default `y_tpPerGpu`. URL param still wins so existing
+   * the dashboard's default `y_tokensPerDollarN`. URL param still wins so existing
    * shared links are unaffected.
    */
   initialYAxisMetric?: string;
+  /**
+   * When true, all chip configs with data for the current model/scenario/
+   * precision selection are selected once availability settles and no
+   * `?i_gpus=` URL param is present. Used by the `/model/[slug]` pages, which
+   * embed the chart without the selector controls — an empty default there
+   * would leave the embed permanently blank. One-shot: after the initial
+   * auto-selection (or when the URL provided chips), user deselections stick.
+   */
+  autoSelectAllGpus?: boolean;
 }) {
   const isActive =
     activeTab === 'inference' || activeTab === 'historical' || activeTab === 'compare';
 
+  const { selectedModel, effectiveSequence, sequenceResolved, effectivePrecisions } =
+    useGlobalFilterSelection();
   const {
-    selectedModel,
     setSelectedModel,
-    effectiveSequence,
-    sequenceResolved,
     setSelectedSequence,
-    effectivePrecisions,
     setSelectedPrecisions,
-    selectedRunDate,
     setSelectedRunDate,
-    selectedRunId,
     setSelectedRunId,
+  } = useGlobalFilterActions();
+  const { selectedRunDate, selectedRunId, effectiveRunDate } = useGlobalFilterRun();
+  const {
     availableModels,
     availableSequences,
     availablePrecisions,
     availableDates,
-    effectiveRunDate,
     availabilityRows,
-    workflowInfo,
-    availableRuns,
-    workflowError,
-  } = useGlobalFilters();
+    availabilitySettled,
+    availabilityError,
+  } = useGlobalFilterAvailability();
+  const { availableRuns, workflowError } = useGlobalFilterWorkflow();
   const { isUnofficialRun } = useUnofficialRun();
 
-  const { getUrlParam, setUrlParam, setUrlParams } = useUrlState();
+  const { getUrlParam, setUrlParams } = useUrlState();
 
   const [overviewHistoryPair, setOverviewHistoryPair] = useState(() => {
     const currentConfigKey = getUrlParam('i_overview_current');
@@ -221,10 +334,15 @@ export function InferenceProvider({
     };
   }, [selectedGpuState, sequenceResolved, exclusion, exclusionPolicy, defaultExclusionGroup]);
   const selectedGPUs = selectedGpuResolution?.selection ?? selectedGpuState;
+  const notifiedGpuResolutionRef = useRef('');
   useEffect(() => {
-    if (!selectedGpuResolution) return;
-    setSelectedGpuState(selectedGpuResolution.selection);
-    setUrlParam('i_gpus', selectedGpuResolution.selection.join(','));
+    if (!selectedGpuResolution) {
+      notifiedGpuResolutionRef.current = '';
+      return;
+    }
+    const resolutionKey = `${selectedGpuResolution.selection.join(',')}|${selectedGpuResolution.dropped.join(',')}|${selectedGpuResolution.partial.join(',')}`;
+    if (resolutionKey === notifiedGpuResolutionRef.current) return;
+    notifiedGpuResolutionRef.current = resolutionKey;
     if (selectedGpuResolution.dropped.length > 0 || selectedGpuResolution.partial.length > 0) {
       setEngineConflict({
         kind: 'resolved',
@@ -233,14 +351,34 @@ export function InferenceProvider({
         partial: selectedGpuResolution.partial,
       });
     }
-  }, [selectedGpuResolution, setUrlParam]);
-  const [selectedYAxisMetric, setSelectedYAxisMetric] = useState<string>(
-    () => getUrlParam('i_metric') || initialYAxisMetric || 'y_tpPerGpu',
+  }, [selectedGpuResolution]);
+  const [selectedYAxisMetric, setSelectedYAxisMetric] = useState<string>(() =>
+    resolveMetricConfigKey(getUrlParam('i_metric'), initialYAxisMetric ?? DEFAULT_Y_AXIS_METRIC),
+  );
+  const [tokenRevenuePriceSource, setTokenRevenuePriceSource] = useState<TokenRevenuePriceSource>(
+    () => (getUrlParam('i_revenue') === 'openrouter' ? 'openrouter' : 'normalized'),
+  );
+  const openRouterModelId = getOpenRouterModelId(selectedModel);
+  const openRouterPricingQuery = useOpenRouterPricing(
+    openRouterModelId,
+    selectedYAxisMetric === 'y_tokenRevenuePerGpuHour' && tokenRevenuePriceSource === 'openrouter',
   );
+  const tokenRevenuePricing =
+    tokenRevenuePriceSource === 'normalized'
+      ? NORMALIZED_TOKEN_REVENUE_PRICING
+      : (openRouterPricingQuery.data ?? null);
+  const openRouterPricingError =
+    tokenRevenuePriceSource === 'openrouter'
+      ? openRouterModelId
+        ? openRouterPricingQuery.error instanceof Error
+          ? openRouterPricingQuery.error.message
+          : null
+        : 'No OpenRouter model mapping is configured.'
+      : null;
   const [selectedXAxisMetric, setSelectedXAxisMetric] = useState<string | null>(
     () => getUrlParam('i_xmetric') || 'p90_ttft',
   );
-  const [selectedE2eXAxisMetric, setSelectedE2eXAxisMetric] = useState<string | null>(
+  const [requestedE2eXAxisMetric] = useState<string | null>(
     () => getUrlParam('i_e2e_xmetric') || 'p90_ttft',
   );
   // Selected chart variant. Initialize from URL only — SSR cannot read URL, so
@@ -252,31 +390,37 @@ export function InferenceProvider({
   // value (if any) in a post-mount effect — keeps server + client first render
   // identical and avoids "didn't match" hydration warnings when the URL holds
   // a non-default mode.
-  const [selectedXAxisMode, setSelectedXAxisMode] = useState<XAxisMode>('interactivity');
+  const [requestedXAxisMode, setRequestedXAxisMode] = useState<XAxisMode>('interactivity');
   const xAxisModeFromUrlRef = useRef(false);
   useEffect(() => {
     if (xAxisModeFromUrlRef.current) return;
-    const v = getUrlParam('i_xmode');
-    if (v && (X_AXIS_MODES as readonly string[]).includes(v)) {
+    const value = getUrlParam('i_xmode');
+    if (value && (X_AXIS_MODES as readonly string[]).includes(value)) {
       xAxisModeFromUrlRef.current = true;
-      setSelectedXAxisMode(v as XAxisMode);
+      setRequestedXAxisMode(value as XAxisMode);
     }
     // eslint-disable-next-line react-hooks/exhaustive-deps
   }, []);
-  // Wrap the setter so a button click also aligns selectedE2eXAxisMetric — the
-  // existing useChartData pipeline keys off that flag for the e2e chart's x-axis.
+  const selectedXAxisMode = resolveEffectiveXAxisMode(
+    requestedXAxisMode,
+    effectiveSequence,
+    sequenceResolved,
+  );
   const handleSetXAxisMode = useCallback((mode: XAxisMode) => {
     xAxisModeFromUrlRef.current = true;
-    setSelectedXAxisMode(mode);
-    // The e2e chart's x-axis metric is reconciled in a separate effect below,
-    // because it depends on sequence kind (fixed-seq has no p90_* metrics) and
-    // the agentic percentile, both of which can change independently.
+    setRequestedXAxisMode(mode);
   }, []);
   // Latency percentile applied to the chart x-axis for agentic scenarios.
   // Values: 'p90' | 'p99'. Non-agentic charts ignore.
   const [selectedPercentile, setSelectedPercentile] = useState<string>(
     () => getUrlParam('i_pctl') || 'p90',
   );
+  const selectedE2eXAxisMetric = resolveE2eXAxisMetric(
+    requestedE2eXAxisMetric,
+    selectedXAxisMode,
+    effectiveSequence,
+    selectedPercentile,
+  );
   const [scaleType, setScaleType] = useState<'auto' | 'linear' | 'log'>(
     () => (getUrlParam('i_scale') as 'auto' | 'linear' | 'log') || 'auto',
   );
@@ -319,10 +463,19 @@ export function InferenceProvider({
     }),
     [quickFilterVendors, quickFilterFrameworks, quickFilterDeployment, quickFilterSpec],
   );
-  // The Historical Trends tab hides the quick-filter pills (hideGpuComparison), so
-  // don't silently narrow its chart with selections carried in via share links or
-  // the inference tab — there would be no pill to clear them.
-  const dataQuickFilters = activeTab === 'historical' ? EMPTY_QUICK_FILTERS : quickFilters;
+  // Historical Trends hides Quick Filters, so never apply invisible selections there.
+  // Agentic charts expose vendor, framework, and deployment filters, but speculative
+  // decoding is intentionally not an agentic option. Ignore stale `i_spec` state from
+  // an older shared URL without discarding it when the reader returns to fixed-seq.
+  const dataQuickFilters = useMemo(
+    () =>
+      activeTab === 'historical'
+        ? EMPTY_QUICK_FILTERS
+        : effectiveSequence === Sequence.AgenticTraces
+          ? { ...quickFilters, spec: [] }
+          : quickFilters,
+    [activeTab, effectiveSequence, quickFilters],
+  );
   const { highContrast, setHighContrast, isLegendExpanded, setIsLegendExpanded } = useChartUIState({
     urlPrefix: 'i_',
   });
@@ -345,14 +498,13 @@ export function InferenceProvider({
   const [showPointLabels, setShowPointLabels] = useState(initialLabelState.showPointLabels);
   const [logScale, setLogScale] = useState(() => getUrlParam('i_log') === '1');
   const [useAdvancedLabels, setUseAdvancedLabels] = useState(initialLabelState.useAdvancedLabels);
+  const [showConcurrencyLabels, setShowConcurrencyLabels] = useState(
+    () => getUrlParam('i_conclabel') === '1',
+  );
   const [showGradientLabels, setShowGradientLabels] = useState(
     () => getUrlParam('i_gradlabel') === '1',
   );
   const [showLineLabels, setShowLineLabels] = useState(initialLabelState.showLineLabels);
-  const [showSpeedOverlay, setShowSpeedOverlay] = useState(() => getUrlParam('i_speed') === '1');
-  const [showMinecraftOverlay, setShowMinecraftOverlay] = useState(
-    () => getUrlParam('i_mc') === '1',
-  );
   const [userCosts, setUserCosts] = useState<Record<string, number | undefined> | null>(null);
   const [userPowers, setUserPowers] = useState<Record<string, number | undefined> | null>(null);
 
@@ -392,12 +544,11 @@ export function InferenceProvider({
   );
 
   const filteredAvailableRuns = useMemo(
-    () => filterRunsByModel(availableRuns, modelPrefixes, [...effectivePrecisions]),
+    () => filterRunsByModel(availableRuns, modelPrefixes, [...effectivePrecisions]) ?? {},
     [availableRuns, modelPrefixes, effectivePrecisions],
   );
 
   const effectiveSelectedRunId = useMemo(() => {
-    if (!filteredAvailableRuns) return selectedRunId;
     const filteredRunIds = Object.keys(filteredAvailableRuns);
     if (filteredRunIds.length === 0 || filteredRunIds.includes(selectedRunId)) return selectedRunId;
     return filteredRunIds.reduce((max, id) => (id > max ? id : max), filteredRunIds[0]);
@@ -406,7 +557,7 @@ export function InferenceProvider({
   // The latest run for this model on the selected date. GitHub run ids increase
   // monotonically with time, so the lexicographically-greatest id is the newest run.
   const latestRunIdForModel = useMemo(() => {
-    const ids = filteredAvailableRuns ? Object.keys(filteredAvailableRuns) : [];
+    const ids = Object.keys(filteredAvailableRuns);
     return ids.length > 0 ? ids.reduce((max, id) => (id > max ? id : max), ids[0]) : '';
   }, [filteredAvailableRuns]);
 
@@ -491,6 +642,8 @@ export function InferenceProvider({
     selectedDateRange,
     userCosts,
     userPowers,
+    tokenRevenuePricing,
+    tokenRevenuePriceSource,
     effectiveRunDate,
     // Gate benchmark fetching on sequenceResolved: before availability loads we
     // don't yet know the model's real sequence, and the selection (e.g. an
@@ -507,6 +660,8 @@ export function InferenceProvider({
     asOfRunId,
     dataQuickFilters,
     overviewHistoryPair,
+    benchmarkQueryScope,
+    selectedModel === initialBenchmarkModel ? initialBenchmarkRows : undefined,
   );
 
   // For GPU comparison date picker — use shared availability data from global filters
@@ -571,6 +726,49 @@ export function InferenceProvider({
       }));
   }, [availabilityRows, dbModelKeys, effectiveSequence, effectivePrecisions, selectedModel]);
 
+  // One-shot auto-selection of every available chip config (see the
+  // `autoSelectAllGpus` prop doc). The selection is pre-resolved through the
+  // exclusion groups so the embed doesn't open with a cross-engine conflict
+  // toast — setting an already-resolved list keeps `selectedGpuResolution`
+  // null, which is the no-notification path.
+  const autoSelectedGpusRef = useRef(false);
+  useEffect(() => {
+    if (!autoSelectAllGpus || autoSelectedGpusRef.current) return;
+    if (!gpuUrlHydrated || !availabilitySettled || !sequenceResolved) return;
+    if (getUrlParam('i_gpus') || selectedGpuState.length > 0) {
+      autoSelectedGpusRef.current = true;
+      return;
+    }
+    if (availableGPUs.length === 0) return;
+    autoSelectedGpusRef.current = true;
+    const all = new Set(availableGPUs.map((gpu) => gpu.value));
+    // `exclusion` is null when the model/scenario has no comparability guard —
+    // in that case every available chip config is selectable together.
+    const selection = exclusion
+      ? [
+          ...resolveExclusionGroups(
+            all,
+            new Set(),
+            exclusion,
+            exclusionPolicy,
+            defaultExclusionGroup,
+          ).result,
+        ]
+      : [...all];
+    setSelectedGpuState(selection);
+  }, [
+    autoSelectAllGpus,
+    gpuUrlHydrated,
+    availabilitySettled,
+    sequenceResolved,
+    exclusion,
+    exclusionPolicy,
+    defaultExclusionGroup,
+    availableGPUs,
+    selectedGpuState,
+    getUrlParam,
+  ]);
+
   useEffect(() => {
     if (!sequenceResolved) return;
     const labelState = resolveLabelState(labelScenarioKind, {
@@ -584,62 +782,6 @@ export function InferenceProvider({
     setShowLineLabels(labelState.showLineLabels);
   }, [labelScenarioKind, sequenceResolved, getUrlParam]);
 
-  // Reconcile the x-axis mode with the scenario kind:
-  //  - On mount with no `i_xmode` URL param: snap to the kind's natural default
-  //    (E2E Normalized Interactivity for agentic — the "north star" e2e-interactivity view —
-  //    and interactivity for fixed-sequence scenarios). The state was initialized
-  //    to a SSR-stable constant so server and client render the same DOM; this
-  //    effect fixes it up after hydration.
-  //  - When the user later switches sequence kinds: snap to the new kind's
-  //    natural default (the prior selection was for a different kind, so it
-  //    doesn't carry over).
-  const lastSeqKindRef = useRef<ReturnType<typeof sequenceKind> | null>(null);
-  useEffect(() => {
-    // Wait for availability to resolve the sequence. Before it does,
-    // `effectiveSequence` is a fixed-seq placeholder; recording that kind here
-    // would make the later switch to agentic look like a user-driven kind
-    // change and clobber a URL-restored `i_xmode` with the kind's default.
-    if (!sequenceResolved) return;
-    const kind = sequenceKind(effectiveSequence);
-    const isInitialMount = lastSeqKindRef.current === null;
-    const isAgenticOnlyMode = isAgenticOnlyXAxisMode(selectedXAxisMode);
-    if (!isInitialMount && lastSeqKindRef.current === kind) {
-      if (kind === 'fixed-seq' && isAgenticOnlyMode) {
-        handleSetXAxisMode('interactivity');
-      }
-      return;
-    }
-    lastSeqKindRef.current = kind;
-    if (
-      isInitialMount &&
-      xAxisModeFromUrlRef.current &&
-      !(kind === 'fixed-seq' && isAgenticOnlyMode)
-    ) {
-      // URL-restored agentic-only mode on a fixed-seq sequence makes no sense
-      // — fall through to the default snap below.
-      return;
-    }
-    handleSetXAxisMode(kind === 'agentic' ? 'e2e-normalized-interactivity' : 'interactivity');
-  }, [sequenceResolved, effectiveSequence, selectedXAxisMode, handleSetXAxisMode]);
-
-  // Reconcile selectedE2eXAxisMetric whenever the mode, sequence kind, or
-  // agentic percentile changes. For fixed-seq the JSONB only carries
-  // median_* / p99_* (no p90_*), so the TTFT button there has to point at
-  // median_ttft — otherwise the chart goes blank. For agentic, we point at
-  // the user's chosen percentile so the dropdown actually drives the axis.
-  useEffect(() => {
-    const isAgentic = sequenceKind(effectiveSequence) === 'agentic';
-    if (selectedXAxisMode === 'ttft') {
-      setSelectedE2eXAxisMetric(isAgentic ? `${selectedPercentile}_ttft` : 'median_ttft');
-    } else if (selectedXAxisMode === 'e2e') {
-      // null = use the chart-config natural x (median_e2el), which useChartData
-      // rewrites to <pctl>_e2el for agentic via withPercentile().
-      setSelectedE2eXAxisMetric(null);
-    }
-    // 'interactivity' mode renders the interactivity chart, which keys off
-    // selectedXAxisMetric (not the e2e one), so nothing to do here.
-  }, [selectedXAxisMode, effectiveSequence, selectedPercentile]);
-
   // Ref guard: when true, filter changes don't clear the active preset.
   // FavoritePresetsDropdown sets this while applying a preset so its own
   // programmatic setter calls don't accidentally deactivate it.
@@ -762,8 +904,16 @@ export function InferenceProvider({
     [setSelectedDateRange, clearScopedSelectionOnChange],
   );
 
-  const loading = chartDataLoading;
-  const error = workflowError || chartDataError;
+  // A failed availability lookup leaves `sequenceResolved` false forever, so
+  // the gated benchmark query never produces rows and `chartDataLoading` would
+  // pin the chart on its first-load skeleton. Availability errors are terminal:
+  // drop the loading flag so ChartDisplay surfaces the error instead.
+  const openRouterPricingLoading =
+    selectedYAxisMetric === 'y_tokenRevenuePerGpuHour' &&
+    tokenRevenuePriceSource === 'openrouter' &&
+    openRouterPricingQuery.isLoading;
+  const loading = availabilityError ? false : chartDataLoading || openRouterPricingLoading;
+  const error = availabilityError || workflowError || chartDataError;
 
   // ── Toggle sets ───────────────────────────────────────────────────────────
 
@@ -946,8 +1096,12 @@ export function InferenceProvider({
   }, [graphs, hwTypesWithData, selectedXAxisMode, selectedYAxisMetric]);
 
   const setBestPerSkuAndApply = useCallback(
-    (enabled: boolean) => {
+    (enabled: boolean, options?: { applySelection?: boolean }) => {
       setBestPerSku(enabled);
+      // Overlay-mode legend edits own a temporary unified selection. They can
+      // disable the automatic mode without replacing the context selection
+      // that should be restored when the overlay is dismissed.
+      if (options?.applySelection === false) return;
       const target = enabled ? bestHwTypes : selectableHwTypes;
       setActiveHwTypes(resolveHwSelection(target).result);
       setActivePresetId(null);
@@ -1050,7 +1204,7 @@ export function InferenceProvider({
   const precisionsKey = effectivePrecisions.join(',');
   const hwResetKey = `${selectedModel}|${effectiveSequence}|${precisionsKey}|${
     isUnofficialRun ? 'preview' : 'official'
-  }|${selectedYAxisMetric}|${selectedXAxisMode}`;
+  }`;
   const lastHwResetKeyRef = useRef('');
 
   // Restore legend-active selection from URL on first availability of
@@ -1104,9 +1258,14 @@ export function InferenceProvider({
     if (pendingHwFilterRef.current) return;
     if (pendingActiveHwTypes) return;
     if (selectableHwTypes.size === 0) return;
-    if (lastHwResetKeyRef.current === hwResetKey) return;
-    lastHwResetKeyRef.current = hwResetKey;
+    const scopeChanged = lastHwResetKeyRef.current !== hwResetKey;
     const presetFilter = presetHwFilterRef.current;
+    // Metric changes must preserve manual legend subsets, but automatic
+    // selections still need to follow the newly selected axes. In particular,
+    // Best per SKU is metric-aware and would otherwise keep the previous
+    // metric's winners while its toggle remained enabled.
+    if (!scopeChanged && !bestPerSku && !presetFilter) return;
+    lastHwResetKeyRef.current = hwResetKey;
     if (presetFilter) {
       const filtered = new Set(
         [...selectableHwTypes].filter((k) => matchesPresetHwFilter(k, presetFilter, selectedModel)),
@@ -1148,13 +1307,14 @@ export function InferenceProvider({
     resolveHwSelection,
   ]);
 
-  // Remove selected GPUs that no longer have data for current filters
+  // Remove selected GPUs only after successful availability has settled. An
+  // empty successful scope is different from the transient empty loading state.
   useEffect(() => {
-    if (selectedGPUs.length === 0 || availableGPUs.length === 0) return;
-    const validKeys = new Set(availableGPUs.map((g) => g.value));
-    const valid = selectedGPUs.filter((g) => validKeys.has(g));
+    if (!availabilitySettled || !sequenceResolved || selectedGPUs.length === 0) return;
+    const validKeys = new Set(availableGPUs.map((gpu) => gpu.value));
+    const valid = selectedGPUs.filter((gpu) => validKeys.has(gpu));
     if (valid.length !== selectedGPUs.length) setSelectedGpuState(valid);
-  }, [availableGPUs]);
+  }, [availabilitySettled, sequenceResolved, availableGPUs, selectedGPUs]);
 
   useEffect(() => {
     if (!gpuUrlHydrated) return;
@@ -1184,7 +1344,9 @@ export function InferenceProvider({
   }, [allDateIds, setActiveDates]);
 
   useEffect(() => {
-    if (selectedYAxisMetric !== 'y_costUser') setUserCosts((prev) => (prev === null ? prev : null));
+    if (selectedYAxisMetric !== 'y_costUser' && selectedYAxisMetric !== 'y_tokensPerDollarUser') {
+      setUserCosts((prev) => (prev === null ? prev : null));
+    }
     if (selectedYAxisMetric !== 'y_powerUser')
       setUserPowers((prev) => (prev === null ? prev : null));
   }, [selectedModel, effectiveSequence, effectivePrecisions, selectedYAxisMetric]);
@@ -1285,6 +1447,8 @@ export function InferenceProvider({
   useUrlStateSync(
     {
       i_metric: selectedYAxisMetric,
+      i_revenue:
+        selectedYAxisMetric === 'y_tokenRevenuePerGpuHour' ? tokenRevenuePriceSource : 'normalized',
       i_pctl: selectedPercentile,
       i_gpus: selectedGPUs.join(','),
       i_dates: selectedDates.join(','),
@@ -1301,10 +1465,9 @@ export function InferenceProvider({
       i_scale: scaleType,
       i_legend: isLegendExpanded ? '' : '0',
       i_advlabel: serializedLabelState.i_advlabel,
+      i_conclabel: showConcurrencyLabels ? '1' : '',
       i_gradlabel: showGradientLabels ? '1' : '',
       i_linelabel: serializedLabelState.i_linelabel,
-      i_speed: showSpeedOverlay ? '1' : '',
-      i_mc: showMinecraftOverlay ? '1' : '',
       i_active: iActiveStr,
       i_vendor: quickFilterVendors.join(','),
       i_fw: quickFilterFrameworks.join(','),
@@ -1313,6 +1476,7 @@ export function InferenceProvider({
     },
     [
       selectedYAxisMetric,
+      tokenRevenuePriceSource,
       selectedXAxisMetric,
       selectedE2eXAxisMetric,
       selectedXAxisMode,
@@ -1327,10 +1491,9 @@ export function InferenceProvider({
       logScale,
       isLegendExpanded,
       useAdvancedLabels,
+      showConcurrencyLabels,
       showGradientLabels,
       showLineLabels,
-      showSpeedOverlay,
-      showMinecraftOverlay,
       iActiveStr,
       quickFilterVendors,
       quickFilterFrameworks,
@@ -1440,11 +1603,11 @@ export function InferenceProvider({
   // filteredAvailableRuns / effectiveSelectedRunId are computed above the data
   // fetch (so the chart can query "as of" the selected run).
   //
-  // NOTE: We intentionally do NOT sync effectiveSelectedRunId back to
-  // GlobalFilterContext (setSelectedRunId). That would cause a full tree
-  // re-render on every precision change because filteredAvailableRuns
-  // depends on effectivePrecisions. Instead, InferenceContext exposes
-  // effectiveSelectedRunId directly (line ~499).
+  // NOTE: We intentionally do not sync effectiveSelectedRunId back through the
+  // global run action. filteredAvailableRuns depends on effectivePrecisions, so
+  // doing so would replace the user's global run intent with an inference-only
+  // fallback whenever the precision scope changes. The inference filter domain
+  // exposes effectiveSelectedRunId directly instead.
 
   const handleDateRangeDialogOk = () => {
     clearOverviewHistoryPair();
@@ -1453,165 +1616,184 @@ export function InferenceProvider({
     setShowDateRangeDialog(false);
   };
 
-  // ── Context value ─────────────────────────────────────────────────────────
+  // ── Domain context values ─────────────────────────────────────────────────
 
-  const value = useMemo(
+  const dataValue = useMemo<InferenceDataContextType>(
     () => ({
-      activeHwTypes,
       hwTypesWithData,
-      toggleHwType,
-      removeHwType,
-      selectAllHwTypes,
-      bestPerSku,
-      setBestPerSku: setBestPerSkuAndApply,
-      resolveComparisonSelection: resolveHwSelection,
-      toggleComparisonSelection,
       hardwareConfig,
       graphs,
-      selectedModel,
-      setSelectedModel: setSelectedModelAndClear,
-      selectedSequence: effectiveSequence,
-      setSelectedSequence: setSelectedSequenceAndClear,
-      selectedPrecisions: effectivePrecisions,
-      setSelectedPrecisions: setSelectedPrecisionsAndClear,
-      isLegendExpanded,
-      setIsLegendExpanded,
-      hideNonOptimal,
-      setHideNonOptimal,
-      showPointLabels,
-      setShowPointLabels,
-      highContrast,
-      setHighContrast,
-      logScale,
-      setLogScale,
-      selectedXAxisMetric,
-      setSelectedXAxisMetric,
-      selectedE2eXAxisMetric,
-      setSelectedE2eXAxisMetric,
-      selectedXAxisMode,
-      setSelectedXAxisMode: handleSetXAxisMode,
-      scaleType,
-      setScaleType,
-      quickFilters,
-      availableQuickFilters,
-      setQuickFilterVendors,
-      setQuickFilterFrameworks,
-      setQuickFilterDeployment,
-      setQuickFilterSpec,
       loading,
       error,
-      workflowInfo,
-      selectedYAxisMetric,
-      setSelectedYAxisMetric: setSelectedYAxisMetricAndClear,
-      selectedPercentile,
-      setSelectedPercentile,
-      selectedGPUs,
-      setSelectedGPUs: setSelectedGPUsAndClear,
+      availableQuickFilters,
       availableGPUs,
-      selectedDates,
-      setSelectedDates: setSelectedDatesAndClear,
-      setSelectedDatesFromRunExpansion: setSelectedDates,
-      selectedDateRange,
-      setSelectedDateRange: setSelectedDateRangeAndClear,
-      activeDates,
-      setActiveDates,
-      toggleActiveDate,
-      removeActiveDate,
-      selectAllActiveDates,
-      selectedRunDate,
-      setSelectedRunDate,
-      userCosts,
-      setUserCosts,
       availableDates,
       dateRangeAvailableDates,
       isCheckingAvailableDates,
       availableRuns: filteredAvailableRuns,
-      selectedRunId: effectiveSelectedRunId,
-      setSelectedRunId,
       availablePrecisions,
       availableSequences,
       availableModels,
+    }),
+    [
+      hwTypesWithData,
+      hardwareConfig,
+      graphs,
+      loading,
+      error,
+      availableQuickFilters,
+      availableGPUs,
+      availableDates,
+      dateRangeAvailableDates,
+      isCheckingAvailableDates,
+      filteredAvailableRuns,
+      availablePrecisions,
+      availableSequences,
+      availableModels,
+    ],
+  );
+
+  const filtersValue = useMemo<InferenceFiltersContextType>(
+    () => ({
+      activeHwTypes,
+      bestPerSku,
+      selectedModel,
+      selectedSequence: effectiveSequence,
+      selectedPrecisions: effectivePrecisions,
+      quickFilters,
+      selectedGPUs,
+      selectedDates,
+      selectedDateRange,
+      activeDates,
+      userCosts,
+      selectedRunDate,
+      selectedRunId: effectiveSelectedRunId,
       userPowers,
-      setUserPowers,
-      useAdvancedLabels,
-      setUseAdvancedLabels,
-      showGradientLabels,
-      setShowGradientLabels,
-      showLineLabels,
-      setShowLineLabels,
-      showSpeedOverlay,
-      setShowSpeedOverlay,
-      showMinecraftOverlay,
-      setShowMinecraftOverlay,
-      setHwFilter: setPendingHwFilter,
       activePresetId,
-      setActivePresetId,
       presetGuardRef,
       compareGpuPair: compareGpuPair ?? null,
     }),
     [
       activeHwTypes,
-      hwTypesWithData,
-      toggleHwType,
-      removeHwType,
-      selectAllHwTypes,
       bestPerSku,
-      setBestPerSkuAndApply,
-      resolveHwSelection,
-      toggleComparisonSelection,
-
-      hardwareConfig,
-      graphs,
-      loading,
-      error,
-      workflowInfo,
       selectedModel,
       effectiveSequence,
       effectivePrecisions,
-      selectedYAxisMetric,
-      selectedXAxisMetric,
-      selectedE2eXAxisMetric,
-      selectedXAxisMode,
-      scaleType,
       quickFilters,
-      availableQuickFilters,
       selectedGPUs,
       selectedDates,
       selectedDateRange,
       activeDates,
-      toggleActiveDate,
-      removeActiveDate,
-      selectAllActiveDates,
+      userCosts,
       selectedRunDate,
-      availableDates,
-      dateRangeAvailableDates,
-      isCheckingAvailableDates,
-      availableGPUs,
-      filteredAvailableRuns,
       effectiveSelectedRunId,
-      availablePrecisions,
-      availableSequences,
-      availableModels,
+      userPowers,
+      activePresetId,
+      compareGpuPair,
+    ],
+  );
+
+  const displayValue = useMemo<InferenceDisplayContextType>(
+    () => ({
+      selectedYAxisMetric,
+      tokenRevenuePriceSource,
+      tokenRevenuePricing,
+      openRouterModelId,
+      openRouterPricingLoading,
+      openRouterPricingError,
+      selectedPercentile,
+      selectedXAxisMetric,
+      selectedE2eXAxisMetric,
+      selectedXAxisMode,
+      scaleType,
+      isLegendExpanded,
       hideNonOptimal,
       showPointLabels,
       highContrast,
       logScale,
+      useAdvancedLabels,
+      showConcurrencyLabels,
+      showGradientLabels,
+      showLineLabels,
+    }),
+    [
+      selectedYAxisMetric,
+      tokenRevenuePriceSource,
+      tokenRevenuePricing,
+      openRouterModelId,
+      openRouterPricingLoading,
+      openRouterPricingError,
+      selectedPercentile,
+      selectedXAxisMetric,
+      selectedE2eXAxisMetric,
+      selectedXAxisMode,
+      scaleType,
       isLegendExpanded,
+      hideNonOptimal,
+      showPointLabels,
+      highContrast,
+      logScale,
       useAdvancedLabels,
+      showConcurrencyLabels,
       showGradientLabels,
       showLineLabels,
-      showSpeedOverlay,
-      showMinecraftOverlay,
-      userCosts,
-      userPowers,
-      activePresetId,
-      compareGpuPair,
     ],
   );
 
+  const actionsValue: InferenceActionsContextType = {
+    toggleActiveDate,
+    removeActiveDate,
+    selectAllActiveDates,
+    toggleHwType,
+    removeHwType,
+    selectAllHwTypes,
+    setBestPerSku: setBestPerSkuAndApply,
+    resolveComparisonSelection: resolveHwSelection,
+    toggleComparisonSelection,
+    setSelectedModel: setSelectedModelAndClear,
+    setSelectedSequence: setSelectedSequenceAndClear,
+    setSelectedPrecisions: setSelectedPrecisionsAndClear,
+    setSelectedYAxisMetric: setSelectedYAxisMetricAndClear,
+    setTokenRevenuePriceSource,
+    setSelectedPercentile,
+    setSelectedXAxisMetric,
+    setSelectedXAxisMode: handleSetXAxisMode,
+    setScaleType,
+    setQuickFilterVendors,
+    setQuickFilterFrameworks,
+    setQuickFilterDeployment,
+    setQuickFilterSpec,
+    setIsLegendExpanded,
+    setHideNonOptimal,
+    setShowPointLabels,
+    setHighContrast,
+    setLogScale,
+    setUseAdvancedLabels,
+    setShowConcurrencyLabels,
+    setShowGradientLabels,
+    setShowLineLabels,
+    setSelectedGPUs: setSelectedGPUsAndClear,
+    setSelectedDates: setSelectedDatesAndClear,
+    setSelectedDatesFromRunExpansion: setSelectedDates,
+    setSelectedDateRange: setSelectedDateRangeAndClear,
+    setUserCosts,
+    setSelectedRunDate,
+    setSelectedRunId,
+    setUserPowers,
+    setHwFilter: setPendingHwFilter,
+    setActivePresetId,
+  };
+
   return (
-    <InferenceContext.Provider value={value}>
-      {children}
+    <>
+      <InferenceContextsProvider
+        data={dataValue}
+        filters={filtersValue}
+        display={displayValue}
+        actions={actionsValue}
+      >
+        {children}
+      </InferenceContextsProvider>
       <EngineComparisonConflictToast
         detail={isUnofficialRun ? null : engineConflict}
         onDismiss={dismissEngineConflict}
@@ -1630,14 +1812,29 @@ export function InferenceProvider({
           </DialogFooter>
         </DialogContent>
       </Dialog>
-    </InferenceContext.Provider>
+    </>
   );
 }
 
-export function useInference() {
-  const context = useContext(InferenceContext);
+function useRequiredInferenceContext<T>(context: T | undefined, hookName: string): T {
   if (context === undefined) {
-    throw new Error('useInference must be used within an InferenceProvider');
+    throw new Error(`${hookName} must be used within an InferenceProvider`);
   }
   return context;
 }
+
+export function useInferenceData() {
+  return useRequiredInferenceContext(useContext(InferenceDataContext), 'useInferenceData');
+}
+
+export function useInferenceFilters() {
+  return useRequiredInferenceContext(useContext(InferenceFiltersContext), 'useInferenceFilters');
+}
+
+export function useInferenceDisplay() {
+  return useRequiredInferenceContext(useContext(InferenceDisplayContext), 'useInferenceDisplay');
+}
+
+export function useInferenceActions() {
+  return useRequiredInferenceContext(useContext(InferenceActionsContext), 'useInferenceActions');
+}
diff --git a/packages/app/src/components/inference/agentic-catalog/agentic-catalog-hero.tsx b/packages/app/src/components/inference/agentic-catalog/agentic-catalog-hero.tsx
new file mode 100644
index 000000000..a4ae49a26
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-catalog/agentic-catalog-hero.tsx
@@ -0,0 +1,72 @@
+import { Card } from '@/components/ui/card';
+import { Eyebrow } from '@/components/ui/eyebrow';
+import { getTelemetryGuide } from '@/lib/agentx-telemetry-zh';
+import type { Locale } from '@/lib/i18n';
+
+import { AgenticCatalogLink } from './agentic-catalog-link';
+
+/**
+ * Short hero for `/inference/agentic`, drawn from the same
+ * `AGENTX_TELEMETRY_GUIDE` that renders the full tutorial at
+ * `/agentx/telemetry`. Reusing the guide's eyebrow, title, lead, and
+ * highlights keeps the catalog's framing in lockstep with the tutorial —
+ * including the highlight counts, which `agentx-telemetry.test.ts` pins to
+ * what the detail page actually ships.
+ */
+
+const STRINGS = {
+  en: {
+    catalogLead:
+      'Every card below opens the telemetry behind one AgentX configuration. Pick a model and a serving stack, and the detail page shows that run charted request by request.',
+    readTutorial: 'Read the telemetry tutorial',
+  },
+  zh: {
+    catalogLead:
+      '下方每张卡片都会打开对应 AgentX 配置背后的遥测数据。选择一个模型与推理服务栈，详情页会逐请求呈现该次运行的图表。',
+    readTutorial: '阅读遥测数据教程',
+  },
+} as const;
+
+export function AgenticCatalogHero({ locale }: { locale: Locale }) {
+  const guide = getTelemetryGuide(locale);
+  const t = STRINGS[locale];
+  const prefix = locale === 'zh' ? '/zh' : '';
+
+  return (
+    <section data-testid="agentic-catalog-hero">
+      <Card>
+        <Eyebrow as="p">{guide.eyebrow}</Eyebrow>
+        <h1 className="mt-2 max-w-3xl text-2xl font-bold tracking-tight text-foreground lg:text-3xl">
+          {guide.title}
+        </h1>
+        <p className="mt-3 max-w-4xl text-base leading-7 text-muted-foreground lg:text-lg">
+          {guide.lead}
+        </p>
+        <p className="mt-3 max-w-4xl text-sm leading-6 text-muted-foreground">{t.catalogLead}</p>
+
+        <dl className="mt-6 grid grid-cols-2 gap-px overflow-hidden rounded-xl border border-border/70 bg-border/70 lg:grid-cols-4">
+          {guide.highlights.map((item) => (
+            <div key={item.label} className="flex flex-col bg-card px-4 py-4">
+              <dt className="order-2 mt-1 text-sm leading-5 text-muted-foreground">{item.label}</dt>
+              <dd className="order-1 font-mono text-xl font-semibold tabular-nums text-foreground">
+                {item.value}
+              </dd>
+            </div>
+          ))}
+        </dl>
+
+        <div className="mt-6">
+          <AgenticCatalogLink
+            data-testid="agentic-catalog-tutorial-link"
+            href={`${prefix}/agentx/telemetry`}
+            analyticsEvent="agentic_catalog_tutorial_clicked"
+            className="inline-flex min-h-11 items-center gap-2 rounded-md bg-brand px-5 py-2.5 text-sm font-semibold text-primary-foreground shadow-sm transition-colors hover:bg-brand/90"
+          >
+            {t.readTutorial}
+            <span aria-hidden="true">→</span>
+          </AgenticCatalogLink>
+        </div>
+      </Card>
+    </section>
+  );
+}
diff --git a/packages/app/src/components/inference/agentic-catalog/agentic-catalog-link.tsx b/packages/app/src/components/inference/agentic-catalog/agentic-catalog-link.tsx
new file mode 100644
index 000000000..075861c45
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-catalog/agentic-catalog-link.tsx
@@ -0,0 +1,88 @@
+'use client';
+
+import { ArrowRight } from 'lucide-react';
+import Link from 'next/link';
+
+import { track } from '@/lib/analytics';
+
+/**
+ * Analytics-instrumented links for the `/inference/agentic` catalog. The page
+ * and its cards are server components, so every tracked click routes here —
+ * the same arrangement `compare-index-tracked-link.tsx` uses for `/compare`.
+ */
+interface AgenticCatalogLinkProps extends Omit<React.ComponentProps<typeof Link>, 'href'> {
+  href: string;
+  analyticsEvent: 'agentic_catalog_tutorial_clicked' | 'agentic_catalog_point_clicked';
+  analyticsTarget?: string;
+}
+
+export function AgenticCatalogLink({
+  href,
+  analyticsEvent,
+  analyticsTarget,
+  onClick,
+  ...props
+}: AgenticCatalogLinkProps) {
+  return (
+    <Link
+      {...props}
+      href={href}
+      onClick={(event) => {
+        onClick?.(event);
+        if (event.defaultPrevented) return;
+        track(analyticsEvent, analyticsTarget ? { target: analyticsTarget } : undefined);
+      }}
+    />
+  );
+}
+
+export interface AgenticCatalogCardStrings {
+  /** "{n} points · conc {min}–{max}" summary line. */
+  points: (count: number) => string;
+  concurrency: (min: number, max: number) => string;
+}
+
+export function AgenticCatalogCardLink({
+  href,
+  label,
+  archLine,
+  precisionLabel,
+  detailLine,
+  target,
+}: {
+  href: string;
+  label: string;
+  archLine: string;
+  precisionLabel: string;
+  detailLine: string;
+  target: string;
+}) {
+  return (
+    <AgenticCatalogLink
+      href={href}
+      analyticsEvent="agentic_catalog_point_clicked"
+      analyticsTarget={target}
+      data-testid={`agentic-catalog-card-${target}`}
+      className="group relative flex flex-col rounded-xl border border-border bg-background/20 p-5 backdrop-blur-[2px] transition-all duration-200 hover:scale-[1.01] hover:border-brand/50 hover:shadow-lg hover:shadow-brand/5"
+    >
+      <div className="absolute inset-y-3 left-0 w-0.5 rounded-full bg-brand/60 transition-all duration-200 group-hover:inset-y-2 group-hover:bg-brand" />
+      <div className="flex items-start justify-between gap-2">
+        <div className="min-w-0 flex-1">
+          <div className="flex flex-wrap items-center gap-2">
+            <h4 className="text-sm font-semibold leading-tight transition-colors duration-200 group-hover:text-brand">
+              {label}
+            </h4>
+            <span className="inline-flex min-h-5 items-center rounded-full border border-brand/30 bg-brand/10 px-2 font-mono text-3xs font-semibold leading-none text-brand">
+              {precisionLabel}
+            </span>
+          </div>
+          <p className="mt-1.5 text-xs leading-relaxed text-muted-foreground">{archLine}</p>
+          <p className="mt-1 font-mono text-2xs leading-relaxed text-muted-foreground">
+            {detailLine}
+          </p>
+        </div>
+        <ArrowRight className="size-4 shrink-0 text-muted-foreground transition-all duration-200 group-hover:translate-x-0.5 group-hover:text-brand" />
+      </div>
+    </AgenticCatalogLink>
+  );
+}
diff --git a/packages/app/src/components/inference/agentic-catalog/agentic-catalog-list.tsx b/packages/app/src/components/inference/agentic-catalog/agentic-catalog-list.tsx
new file mode 100644
index 000000000..a6e247eb3
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-catalog/agentic-catalog-list.tsx
@@ -0,0 +1,101 @@
+import { Card } from '@/components/ui/card';
+import type { AgenticCatalogModelGroup } from '@/lib/agentic-catalog';
+import type { Locale } from '@/lib/i18n';
+
+import { AgenticCatalogCardLink } from './agentic-catalog-link';
+
+/**
+ * Model-grouped catalog of the AgentX runs that have stored telemetry, laid
+ * out the way `/compare` lays out its comparison catalog: a summary card, then
+ * one `<section>` per model holding a grid of link cards.
+ *
+ * One card per (hardware, framework, precision) config rather than per point.
+ * There are hundreds of trace-backed points, and the detail page carries a
+ * sibling navigator, so a card opens that config's representative point and
+ * the reader moves between concurrencies from there.
+ */
+
+const STRINGS = {
+  en: {
+    eyebrow: 'Telemetry catalog',
+    heading: 'AgentX runs with stored telemetry',
+    summary: (configs: number, points: number, models: number) =>
+      `${configs.toLocaleString()} serving configurations across ${models} models, covering ${points.toLocaleString()} benchmark points with per-request telemetry. Each card opens that configuration's headline point; the detail page's navigator moves between its concurrencies.`,
+    empty: 'No AgentX runs currently have stored telemetry.',
+    configs: (n: number) => `${n} configuration${n === 1 ? '' : 's'}`,
+    points: (n: number) => `${n.toLocaleString()} point${n === 1 ? '' : 's'}`,
+    detail: (points: number, minConc: number, maxConc: number, date: string) =>
+      `${points} points · conc ${minConc}–${maxConc} · ${date}`,
+  },
+  zh: {
+    eyebrow: '遥测数据目录',
+    heading: '已存储遥测数据的 AgentX 运行',
+    summary: (configs: number, points: number, models: number) =>
+      `覆盖 ${models} 个模型的 ${configs.toLocaleString()} 种推理服务配置，共 ${points.toLocaleString()} 个带有逐请求遥测数据的基准测试数据点。每张卡片打开该配置的代表性数据点，详情页的导航器可在其各并发档位之间切换。`,
+    empty: '当前没有 AgentX 运行存储了遥测数据。',
+    configs: (n: number) => `${n} 种配置`,
+    points: (n: number) => `${n.toLocaleString()} 个数据点`,
+    detail: (points: number, minConc: number, maxConc: number, date: string) =>
+      `${points} 个数据点 · 并发 ${minConc}–${maxConc} · ${date}`,
+  },
+} as const;
+
+export function AgenticCatalogList({
+  groups,
+  locale,
+}: {
+  groups: AgenticCatalogModelGroup[];
+  locale: Locale;
+}) {
+  const t = STRINGS[locale];
+  const prefix = locale === 'zh' ? '/zh' : '';
+  const configCount = groups.reduce((sum, group) => sum + group.cards.length, 0);
+  const pointCount = groups.reduce((sum, group) => sum + group.totalPoints, 0);
+
+  return (
+    <>
+      <section id="telemetry-catalog" data-testid="agentic-catalog-summary">
+        <Card>
+          <p className="font-mono text-xs font-semibold tracking-eyebrow text-muted-foreground uppercase">
+            {t.eyebrow}
+          </p>
+          <h2 className="mt-2 text-2xl font-bold tracking-tight lg:text-3xl">{t.heading}</h2>
+          <p className="mt-3 max-w-4xl text-base text-muted-foreground lg:text-lg">
+            {groups.length === 0 ? t.empty : t.summary(configCount, pointCount, groups.length)}
+          </p>
+        </Card>
+      </section>
+
+      {groups.map((group) => (
+        <section key={group.key} id={group.key} data-testid={`agentic-catalog-model-${group.key}`}>
+          <Card className="flex flex-col gap-4">
+            <div>
+              <h3 className="text-xl font-bold tracking-tight lg:text-2xl">{group.label}</h3>
+              <p className="mt-1 text-sm text-muted-foreground">
+                {t.configs(group.cards.length)} · {t.points(group.totalPoints)}
+              </p>
+            </div>
+            <div className="grid grid-cols-1 gap-3 sm:grid-cols-2 lg:grid-cols-3">
+              {group.cards.map((card) => (
+                <AgenticCatalogCardLink
+                  key={card.id}
+                  href={`${prefix}/inference/agentic/${card.id}`}
+                  label={card.label}
+                  archLine={`${card.vendor} · ${card.arch}`}
+                  precisionLabel={card.precisionLabel}
+                  detailLine={t.detail(card.points, card.minConc, card.maxConc, card.latestDate)}
+                  // Precision is part of the key: a model/SKU/engine can ship
+                  // both an FP4 and an FP8 config, and the two are separate
+                  // cards that must not collide on the same test id.
+                  target={`${group.key}-${card.hardwareKey}-${card.frameworkLabel}-${card.precisionLabel}`
+                    .toLowerCase()
+                    .replaceAll(/[^a-z0-9]+/gu, '-')}
+                />
+              ))}
+            </div>
+          </Card>
+        </section>
+      ))}
+    </>
+  );
+}
diff --git a/packages/app/src/components/inference/agentic-point/agentic-point-detail-copy.test.ts b/packages/app/src/components/inference/agentic-point/agentic-point-detail-copy.test.ts
new file mode 100644
index 000000000..43e4fb423
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/agentic-point-detail-copy.test.ts
@@ -0,0 +1,13 @@
+import { describe, expect, it } from 'vitest';
+
+import { AGENTIC_POINT_DETAIL_STRINGS } from './agentic-point-detail';
+
+describe('agentic point detail copy', () => {
+  it('composes the Chinese warmup note without repeating warmup', () => {
+    const copy = AGENTIC_POINT_DETAIL_STRINGS.zh;
+    const note = `${copy.warmupNotePrefix}${copy.warmupWord}${copy.warmupNoteBody}`;
+
+    expect(note).toMatch(/^当前显示 warmup 阶段/u);
+    expect(note).not.toContain('warmup warmup');
+  });
+});
diff --git a/packages/app/src/components/inference/agentic-point/agentic-point-detail.tsx b/packages/app/src/components/inference/agentic-point/agentic-point-detail.tsx
index 16741a84c..75c7f861e 100644
--- a/packages/app/src/components/inference/agentic-point/agentic-point-detail.tsx
+++ b/packages/app/src/components/inference/agentic-point/agentic-point-detail.tsx
@@ -1,32 +1,39 @@
 'use client';
 
 import Link from 'next/link';
-import { usePathname, useRouter, useSearchParams } from 'next/navigation';
-import { useCallback, useMemo, useState } from 'react';
+import dynamic from 'next/dynamic';
+import { usePathname, useRouter } from 'next/navigation';
+import { useEffect, useMemo, useState } from 'react';
 import { ArrowLeft } from 'lucide-react';
 
 import { useAgenticAggregates } from '@/hooks/api/use-agentic-aggregates';
+import { useRequestChartData } from '@/hooks/api/use-request-chart-data';
 import { useRequestTimeline } from '@/hooks/api/use-request-timeline';
-import { useTraceServerMetrics } from '@/hooks/api/use-trace-server-metrics';
+import {
+  useTraceServerMetrics,
+  useTraceServerMetricSource,
+} from '@/hooks/api/use-trace-server-metrics';
 import { useBenchmarkSiblings } from '@/hooks/api/use-benchmark-siblings';
+import { NudgeEngine } from '@/components/nudge-engine';
 import { SegmentedToggle, type SegmentedToggleOption } from '@/components/ui/segmented-toggle';
 import { track } from '@/lib/analytics';
 import { useLocale } from '@/lib/use-locale';
 import { isZhPathname, ZH_PREFIX } from '@/lib/i18n';
+import { withChartState } from '@/lib/url-state';
 
 import { AggregatesGrid } from './aggregates-grid';
 import { MetricSourceToolbar } from './metric-source-toolbar';
 import {
   phaseBoundarySec,
+  sliceRequestChartDataByPhase,
   sliceServerSeriesByPhase,
-  sliceTimelineByPhase,
   timelineHasWarmup,
   type ServerSeriesLike,
   type StagePhase,
 } from './phase-slice';
 import { PointSummary } from './point-summary';
 import { RequestMetricOverTime, SequenceMetricCard } from './request-metric-cards';
-import { RequestTimelineView } from './request-timeline';
+import { ServerLogViewer } from './server-log-viewer';
 import {
   CumulativeUniqueInputTokensCard,
   InflightUniqueTokensCard,
@@ -39,8 +46,9 @@ import {
 } from './server-metric-cards';
 import { SiblingNav } from './sibling-nav';
 import type { ThroughputSeriesKey } from './time-series-math';
+import { useDetailView, type DetailView } from './use-detail-view';
 
-const STRINGS = {
+export const AGENTIC_POINT_DETAIL_STRINGS = {
   en: {
     back: 'Back',
     inferenceChart: 'Inference chart',
@@ -53,12 +61,15 @@ const STRINGS = {
     perPoint: 'Per-point',
     requestTimeline: 'Request timeline',
     aggregatesAcrossConfigs: 'Aggregates across configs',
+    logs: 'Logs',
+    detailView: 'Detail view',
     warmupWord: 'warmup',
     warmupNotePrefix: 'Showing the ',
     warmupNoteBody:
       ' phase — a cache-warming pass whose outputs are capped at 1 token. Warmup OSL ≈ 1, and interactivity/decode are blank (single-token outputs have no inter-token latency).',
     warmupNoServerData:
       ' Warmup server-side metrics aren’t available for this point, so the server charts below are empty — the request-level charts above still reflect warmup.',
+    metricSourceError: 'The selected server-metrics source could not be loaded.',
   },
   zh: {
     back: '返回',
@@ -72,57 +83,51 @@ const STRINGS = {
     perPoint: '单点',
     requestTimeline: '请求时间线',
     aggregatesAcrossConfigs: '跨配置聚合',
+    logs: '日志',
+    detailView: '详情视图',
     warmupWord: 'warmup',
     warmupNotePrefix: '当前显示 ',
     warmupNoteBody:
-      ' 预热阶段——该阶段用于缓存预热，输出被限制为 1 个 token。预热阶段 OSL ≈ 1，交互性/解码指标为空（单 token 输出没有 token 间延迟）。',
+      ' 阶段——该阶段用于建立 cache 状态，输出被限制为 1 个 token。warmup 阶段 OSL ≈ 1，交互性/解码指标为空（单 token 输出没有 token 间延迟）。',
     warmupNoServerData:
-      ' 该数据点没有预热阶段的服务器端指标，因此下方服务器图表为空——上方请求级图表仍反映预热阶段数据。',
+      ' 该数据点没有 warmup 阶段的服务器端指标，因此下方服务器图表为空——上方请求级图表仍反映 warmup 阶段数据。',
+    metricSourceError: '无法加载所选服务器指标来源。',
   },
 } as const;
 
+const RequestTimelineView = dynamic(() =>
+  import('./request-timeline').then((module_) => module_.RequestTimelineView),
+);
+
 interface Props {
   id: number;
 }
 
-type DetailView = 'point' | 'timeline' | 'aggregates';
-
-const isDetailView = (value: string | null): value is DetailView =>
-  value === 'point' || value === 'timeline' || value === 'aggregates';
-
-/** URL-persisted detail view (`?view=`; per-point is the unadorned default). */
-function useDetailView(): [DetailView, (nextView: DetailView) => void] {
-  const router = useRouter();
-  const pathname = usePathname();
-  const searchParams = useSearchParams();
-  const requestedView = searchParams.get('view');
-  const view: DetailView = isDetailView(requestedView) ? requestedView : 'point';
-  const setView = useCallback(
-    (nextView: DetailView) => {
-      const nextParams = new URLSearchParams(searchParams.toString());
-      if (nextView === 'point') nextParams.delete('view');
-      else nextParams.set('view', nextView);
-      const query = nextParams.toString();
-      router.replace(query ? `${pathname}?${query}` : pathname, { scroll: false });
-      track('inference_agentic_detail_view_changed', { view: nextView });
-    },
-    [pathname, router, searchParams],
-  );
-  return [view, setView];
-}
-
 export function AgenticPointDetail({ id }: Props) {
   const router = useRouter();
   const pathname = usePathname();
   const locale = useLocale();
-  const t = STRINGS[locale];
+  const t = AGENTIC_POINT_DETAIL_STRINGS[locale];
   const isZh = isZhPathname(pathname);
-  const inferenceHref = isZh ? `${ZH_PREFIX}/inference` : '/inference';
+  const inferenceBaseHref = isZh ? `${ZH_PREFIX}/inference` : '/inference';
+  // Carry the chart state the reader arrived with back to the chart. The link
+  // used to be a bare path, so it could only ever land on the default model
+  // with the default legend, no matter what the reader was looking at.
+  //
+  // Resolved after mount, not during render: `withChartState` reads the
+  // in-memory param store (seeded from this page's own URL at load), which
+  // does not exist on the server — computing it during render would make the
+  // server and client markup disagree.
+  const [inferenceHref, setInferenceHref] = useState(inferenceBaseHref);
+  useEffect(() => {
+    setInferenceHref(withChartState(inferenceBaseHref));
+  }, [inferenceBaseHref]);
   const viewOptions: SegmentedToggleOption<DetailView>[] = useMemo(
     () => [
       { value: 'point', label: t.perPoint, testId: 'detail-view-point' },
       { value: 'timeline', label: t.requestTimeline, testId: 'detail-view-timeline' },
       { value: 'aggregates', label: t.aggregatesAcrossConfigs, testId: 'detail-view-aggregates' },
+      { value: 'logs', label: t.logs, testId: 'detail-view-logs' },
     ],
     [t],
   );
@@ -143,33 +148,42 @@ export function AgenticPointDetail({ id }: Props) {
   // shows how the metric varies across the SKU.
   const siblingIds = siblingsData?.siblings.map((s) => s.id) ?? [];
   const aggregatesQuery = useAgenticAggregates(siblingIds, view === 'aggregates');
-  // Per-request timeline used by the timeline view AND every per-point
-  // request-derived chart (ISL/OSL, latency-over-time, in-flight), so fetch
-  // whenever we're on either view.
-  const timelineQuery = useRequestTimeline(id, view === 'timeline' || view === 'point');
-  const timeline = timelineQuery.data;
+  // The default charts use a compact nine-field request projection. The much
+  // larger source-rich Gantt payload is fetched only after opening Timeline.
+  const requestChartQuery = useRequestChartData(id, view === 'point');
+  const requestChartData = requestChartQuery.data;
+  const timelineQuery = useRequestTimeline(id, view === 'timeline');
 
   // Warmup vs profiling stage. Only meaningful when the point actually has a
   // warmup phase (older runs are profiling-only) — when absent the toggle is
   // hidden and everything falls back to the full (profiling) run.
   const [phase, setPhase] = useState<StagePhase>('profiling');
-  const hasWarmup = useMemo(() => timelineHasWarmup(timeline), [timeline]);
+  const hasWarmup = useMemo(() => timelineHasWarmup(requestChartData), [requestChartData]);
   const effectivePhase: StagePhase = hasWarmup ? phase : 'profiling';
 
   // Server-metric boundary on the chart's own t-axis (rebased through absolute
   // ns — see phase-slice header for the origin-gap invariant). Request charts
   // get a phase-scoped timeline (filtered + rebased) so they share a 0-based
   // axis with the server charts for the selected phase.
-  const boundarySec = useMemo(() => phaseBoundarySec(metrics, timeline), [metrics, timeline]);
-  const phaseTimeline = useMemo(
-    () => (timeline ? sliceTimelineByPhase(timeline, effectivePhase) : null),
-    [timeline, effectivePhase],
+  const boundarySec = useMemo(
+    () => phaseBoundarySec(metrics, requestChartData),
+    [metrics, requestChartData],
+  );
+  const phaseRequestData = useMemo(
+    () =>
+      requestChartData ? sliceRequestChartDataByPhase(requestChartData, effectivePhase) : null,
+    [requestChartData, effectivePhase],
   );
 
   const metricSources = metrics?.metricSources ?? [];
   const selectedMetricSource = metricSources.find(({ source }) => source.id === metricSourceId);
+  const metricSourceQuery = useTraceServerMetricSource(
+    id,
+    metricSourceId,
+    view === 'point' && metricSourceId !== 'all',
+  );
   const baseServerSeries: ServerSeriesLike | undefined = useMemo(() => {
-    const src = metrics?.metricSources?.find((m) => m.source.id === metricSourceId);
+    const src = metricSourceQuery.data;
     if (src) {
       return {
         kvCacheUsage: src.kvCacheUsage,
@@ -183,8 +197,8 @@ export function AgenticPointDetail({ id }: Props) {
         kvCacheUsageByEngine: src.kvCacheUsageByEngine,
       };
     }
-    return metrics ?? undefined;
-  }, [metrics, metricSourceId]);
+    return metricSourceId === 'all' ? (metrics ?? undefined) : undefined;
+  }, [metrics, metricSourceId, metricSourceQuery.data]);
   // Phase-sliced server series (+ matching durationS) consumed by every server
   // chart. Null only when there are no server metrics at all.
   const sliced = useMemo(
@@ -212,16 +226,27 @@ export function AgenticPointDetail({ id }: Props) {
 
   return (
     <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-4 py-6">
+      {/* Points readers at /agentx/telemetry, which explains every chart below.
+          Non-centered, so the card never covers the charts it describes. */}
+      <NudgeEngine scope="agentic-detail" />
+
       <div className="flex items-center gap-2">
         <button
           type="button"
-          onClick={() => router.back()}
+          onClick={() => {
+            track('inference_agentic_detail_back_clicked', { id });
+            router.back();
+          }}
           className="inline-flex items-center gap-1 text-sm text-muted-foreground hover:text-foreground"
         >
           <ArrowLeft className="size-4" /> {t.back}
         </button>
         <span className="text-sm text-muted-foreground">·</span>
-        <Link href={inferenceHref} className="text-sm text-muted-foreground hover:text-foreground">
+        <Link
+          href={inferenceHref}
+          onClick={() => track('inference_agentic_detail_chart_link_clicked', { id })}
+          className="text-sm text-muted-foreground hover:text-foreground"
+        >
           {t.inferenceChart}
         </Link>
       </div>
@@ -238,25 +263,26 @@ export function AgenticPointDetail({ id }: Props) {
         <div className="text-sm text-muted-foreground">{t.loadingPoint}</div>
       ) : null}
 
-      {metricsQuery.isError && (
+      {view !== 'logs' && metricsQuery.isError && (
         <div className="rounded-lg border border-destructive/40 bg-destructive/10 p-4 text-sm text-destructive">
           Failed to load trace data for benchmark point #{id}.
         </div>
       )}
-      {metricsQuery.data === null && !metricsQuery.isLoading && (
+      {view !== 'logs' && metricsQuery.data === null && !metricsQuery.isLoading && (
         <div className="rounded-lg border border-border/40 bg-card/40 p-4 text-sm text-muted-foreground">
           No stored trace_replay blob for benchmark point #{id}. This point predates the aiperf
           time-series capture, or its source artifacts have expired on GitHub.
         </div>
       )}
 
-      <div className="flex items-center justify-between gap-3">
+      <div className="flex min-w-0 items-center justify-between gap-3">
         <SegmentedToggle
           value={view}
           options={viewOptions}
           onValueChange={setView}
-          ariaLabel="Detail view"
+          ariaLabel={t.detailView}
           testId="detail-view-toggle"
+          className="max-w-full overflow-x-auto"
           buttonClassName="px-3 py-1.5 text-sm"
         />
         {view === 'aggregates' && (
@@ -284,7 +310,9 @@ export function AgenticPointDetail({ id }: Props) {
         />
       )}
 
-      {view === 'aggregates' ? (
+      {view === 'logs' ? (
+        <ServerLogViewer id={id} enabled />
+      ) : view === 'aggregates' ? (
         <AggregatesGrid
           siblings={siblingsData?.siblings ?? []}
           aggregates={aggregatesQuery.data}
@@ -320,30 +348,35 @@ export function AgenticPointDetail({ id }: Props) {
               {!slicedHasServerData && t.warmupNoServerData}
             </p>
           )}
+          {metricSourceQuery.isError && (
+            <p className="rounded-md border-l-2 border-destructive/60 bg-destructive/10 px-3 py-2 text-xs text-destructive">
+              {t.metricSourceError}
+            </p>
+          )}
           <div className="grid grid-cols-1 lg:grid-cols-2 gap-4">
             <SequenceMetricCard
               metric="isl"
-              timeline={phaseTimeline}
-              timelineLoading={timelineQuery.isLoading}
+              timeline={phaseRequestData}
+              timelineLoading={requestChartQuery.isLoading}
             />
             <SequenceMetricCard
               metric="osl"
-              timeline={phaseTimeline}
-              timelineLoading={timelineQuery.isLoading}
+              timeline={phaseRequestData}
+              timelineLoading={requestChartQuery.isLoading}
             />
 
             <RequestMetricOverTime
               title={t.interactivityOverTime}
               metric="interactivity"
-              timeline={phaseTimeline}
-              isLoading={timelineQuery.isLoading}
+              timeline={phaseRequestData}
+              isLoading={requestChartQuery.isLoading}
             />
 
             <RequestMetricOverTime
               title={t.ttftOverTime}
               metric="ttft"
-              timeline={phaseTimeline}
-              isLoading={timelineQuery.isLoading}
+              timeline={phaseRequestData}
+              isLoading={requestChartQuery.isLoading}
               latencySelector
             />
 
@@ -351,8 +384,8 @@ export function AgenticPointDetail({ id }: Props) {
 
             <RequestActivityCard
               sliced={sliced}
-              phaseTimeline={phaseTimeline}
-              timelineLoading={timelineQuery.isLoading}
+              phaseTimeline={phaseRequestData}
+              timelineLoading={requestChartQuery.isLoading}
               view={requestActivityView}
               onViewChange={setRequestActivityView}
             />
@@ -371,8 +404,8 @@ export function AgenticPointDetail({ id }: Props) {
             <CumulativeUniqueInputTokensCard sliced={sliced} />
 
             <InflightUniqueTokensCard
-              phaseTimeline={phaseTimeline}
-              timelineLoading={timelineQuery.isLoading}
+              phaseTimeline={phaseRequestData}
+              timelineLoading={requestChartQuery.isLoading}
               kvCachePoolTokens={metrics?.kvCachePoolTokens ?? null}
             />
           </div>
diff --git a/packages/app/src/components/inference/agentic-point/aggregate-chart.tsx b/packages/app/src/components/inference/agentic-point/aggregate-chart.tsx
index d4526d24d..c9d90d254 100644
--- a/packages/app/src/components/inference/agentic-point/aggregate-chart.tsx
+++ b/packages/app/src/components/inference/agentic-point/aggregate-chart.tsx
@@ -5,7 +5,7 @@ import { useMemo } from 'react';
 import { ChartHover, type HoverItem } from './chart-hover';
 import { ChartEmpty, PERCENTILE_COLORS } from './chart-shared';
 
-export type PercentileKey = 'mean' | 'p50' | 'p75' | 'p90' | 'p99';
+export type PercentileKey = 'mean' | 'p50' | 'p75' | 'p90' | 'p95' | 'p99';
 
 interface PercentileLine {
   key: PercentileKey;
@@ -19,6 +19,7 @@ const PERCENTILE_LINES: PercentileLine[] = [
   { key: 'p50', label: 'P50', color: PERCENTILE_COLORS.p50 },
   { key: 'p75', label: 'P75', color: PERCENTILE_COLORS.p75 },
   { key: 'p90', label: 'P90', color: PERCENTILE_COLORS.p90 },
+  { key: 'p95', label: 'P95', color: PERCENTILE_COLORS.p95 },
   { key: 'p99', label: 'P99', color: PERCENTILE_COLORS.p99 },
 ];
 
@@ -37,7 +38,7 @@ export interface AggregatePoint {
 
 /**
  * Multi-line chart: one x-position per sibling config, one line per
- * percentile (mean/p50/p75/p90/p99). Designed for the "Aggregates across
+ * percentile (mean/p50/p75/p90/p95/p99). Designed for the "Aggregates across
  * configs" view on the agentic detail page.
  */
 export function AggregateChart({
diff --git a/packages/app/src/components/inference/agentic-point/aggregates-grid.tsx b/packages/app/src/components/inference/agentic-point/aggregates-grid.tsx
index 09252940e..34330ab51 100644
--- a/packages/app/src/components/inference/agentic-point/aggregates-grid.tsx
+++ b/packages/app/src/components/inference/agentic-point/aggregates-grid.tsx
@@ -19,6 +19,7 @@ function toAggPoint(
     values.p50 = pct.p50;
     values.p75 = pct.p75;
     values.p90 = pct.p90;
+    values.p95 = pct.p95;
     values.p99 = pct.p99;
   }
   return { id: sibling.id, label: sibling.label, values };
diff --git a/packages/app/src/components/inference/agentic-point/distribution.tsx b/packages/app/src/components/inference/agentic-point/distribution.tsx
index 6573d60ca..f11d9f20b 100644
--- a/packages/app/src/components/inference/agentic-point/distribution.tsx
+++ b/packages/app/src/components/inference/agentic-point/distribution.tsx
@@ -2,8 +2,11 @@
 
 import { useMemo } from 'react';
 
+import { useLocale } from '@/lib/use-locale';
+
 import { ChartHover, type HoverItem } from './chart-hover';
 import { CHART_PAD, ChartEmpty, PERCENTILE_COLORS, fmtCount } from './chart-shared';
+import { logHistogram, logTicks, positiveValues } from './lognormal';
 import { quantile } from './time-series-math';
 
 const PAD = CHART_PAD;
@@ -15,9 +18,42 @@ const GUIDES = [
   { label: 'p95', q: 0.95, color: PERCENTILE_COLORS.p95 },
 ] as const;
 
+const STRINGS = {
+  en: {
+    requests: 'requests',
+    range: 'range',
+    logScale: 'log scale',
+    excluded: (n: string, unit: string) =>
+      `${n} requests with 0 ${unit} excluded from the log axis`,
+    bin: 'Bin',
+    count: 'Count',
+    cumulative: 'Cumulative',
+    valueAxis: (unit: string) => `value (${unit}, log scale)`,
+    countAxis: 'count',
+  },
+  zh: {
+    requests: '个请求',
+    range: '范围',
+    logScale: '对数刻度',
+    excluded: (n: string, unit: string) => `${n} 个请求为 0 ${unit}，已从对数轴中排除`,
+    bin: '区间',
+    count: '数量',
+    cumulative: '累计',
+    valueAxis: (unit: string) => `数值（${unit}，对数刻度）`,
+    countAxis: '数量',
+  },
+} as const;
+
 /**
- * Bar histogram with vertical p50/p75/p90/p95 guide lines. Designed for the
- * detail-page card — fills its container width via `viewBox` + 100% width.
+ * Bar histogram of per-request sequence lengths with vertical p50/p75/p90/p95
+ * guide lines. Designed for the detail-page card — fills its container width
+ * via `viewBox` + 100% width.
+ *
+ * Bins are uniform in ln(value), not in value: ISL/OSL span orders of magnitude,
+ * so linear bins pile nearly every request into the leftmost few and stretch a
+ * near-empty tail across the rest of the axis. On a log axis the same roughly
+ * lognormal data reads as an ordinary bell instead.
+ *
  * Hover shows the bin range + count + cumulative percentile.
  */
 export function Distribution({
@@ -31,33 +67,42 @@ export function Distribution({
   width?: number;
   height?: number;
 }) {
+  const t = STRINGS[useLocale()];
   const W = width;
   const H = height;
 
   const computed = useMemo(() => {
-    if (values.length === 0) return null;
-    const sorted = [...values].toSorted((a, b) => a - b);
-    const min = sorted[0]!;
-    const max = sorted.at(-1)!;
-    const range = Math.max(1e-9, max - min);
-    const innerW = W - PAD.left - PAD.right;
-    const innerH = H - PAD.top - PAD.bottom;
-    const nBins = Math.min(50, Math.max(15, Math.ceil(Math.sqrt(values.length))));
-    const counts: number[] = Array.from({ length: nBins }, () => 0);
-    for (const v of values) {
-      const i = Math.min(nBins - 1, Math.floor(((v - min) / range) * nBins));
-      counts[i]!++;
-    }
-    return { sorted, min, max, range, innerW, innerH, nBins, counts };
+    const positive = positiveValues(values);
+    if (positive.length === 0) return null;
+    const nBins = Math.min(50, Math.max(15, Math.ceil(Math.sqrt(positive.length))));
+    const histogram = logHistogram(positive, nBins);
+    if (!histogram) return null;
+    return {
+      sorted: positive.toSorted((a, b) => a - b),
+      histogram,
+      // Zero-token requests are real but unplottable on a log axis; they are
+      // reported under the chart rather than silently folded into bin one.
+      excluded: values.length - positive.length,
+      innerW: W - PAD.left - PAD.right,
+      innerH: H - PAD.top - PAD.bottom,
+    };
   }, [values, W, H]);
 
   if (!computed) {
-    return <ChartEmpty />;
+    return <ChartEmpty height={H} />;
   }
-  const { sorted, min, max, range, innerW, innerH, nBins, counts } = computed;
-  const maxCount = Math.max(...counts, 1);
-  const xScale = (v: number) => PAD.left + ((v - min) / range) * innerW;
-  const yScale = (c: number) => PAD.top + (1 - c / maxCount) * innerH;
+  const { sorted, histogram, excluded, innerW, innerH } = computed;
+  const { counts, edges, lnMin, lnMax } = histogram;
+  const min = edges[0]!;
+  const max = edges.at(-1)!;
+  const nBins = counts.length;
+
+  const yMax = Math.max(...counts, 1);
+
+  const lnSpan = lnMax - lnMin;
+  const xScale = (v: number) =>
+    PAD.left + ((Math.log(Math.max(v, Number.MIN_VALUE)) - lnMin) / lnSpan) * innerW;
+  const yScale = (c: number) => PAD.top + (1 - Math.min(c, yMax) / yMax) * innerH;
   const barW = innerW / nBins;
 
   const fmt = fmtCount;
@@ -65,30 +110,30 @@ export function Distribution({
   // Hover: report the bin range under cursor, its count, and what percentile
   // the bin's midpoint represents in the empirical distribution.
   const resolve = (fraction: number) => {
-    const v = min + fraction * range;
-    const binIdx = Math.min(nBins - 1, Math.floor(((v - min) / range) * nBins));
-    const binLo = min + (binIdx * range) / nBins;
-    const binHi = min + ((binIdx + 1) * range) / nBins;
+    const binIdx = Math.min(nBins - 1, Math.max(0, Math.floor(fraction * nBins)));
+    const binLo = edges[binIdx]!;
+    const binHi = edges[binIdx + 1]!;
     const count = counts[binIdx] ?? 0;
     // Cumulative % at the bin's right edge.
     let cumCount = 0;
     for (let i = 0; i <= binIdx; i++) cumCount += counts[i] ?? 0;
-    const cumPct = (cumCount / values.length) * 100;
+    const cumPct = (cumCount / sorted.length) * 100;
     const items: HoverItem[] = [
-      { color: 'currentColor', label: 'Bin', value: `${fmt(binLo)}–${fmt(binHi)} ${unit}` },
-      { color: 'currentColor', label: 'Count', value: count.toLocaleString() },
-      { color: 'currentColor', label: 'Cumulative', value: `${cumPct.toFixed(1)}%` },
+      { color: 'currentColor', label: t.bin, value: `${fmt(binLo)}–${fmt(binHi)} ${unit}` },
+      { color: 'currentColor', label: t.count, value: count.toLocaleString() },
+      { color: 'currentColor', label: t.cumulative, value: `${cumPct.toFixed(1)}%` },
     ];
     return { items };
   };
 
-  const xTickVals = [min, min + range / 3, min + (2 * range) / 3, max];
-  const yTickVals = Array.from({ length: 5 }, (_, i) => (maxCount * i) / 4);
+  const xTickVals = logTicks(min, max);
+  const yTickVals = Array.from({ length: 5 }, (_, i) => (yMax * i) / 4);
 
   return (
     <div className="w-full">
       <div className="mb-2 text-xs text-muted-foreground">
-        {values.length.toLocaleString()} requests · range {fmt(min)}–{fmt(max)} {unit}
+        {sorted.length.toLocaleString()} {t.requests} · {t.range} {fmt(min)}–{fmt(max)} {unit} ·{' '}
+        {t.logScale}
       </div>
       <ChartHover pad={PAD} width={W} height={H} resolve={resolve}>
         {/* y-axis gridlines + labels */}
@@ -120,7 +165,7 @@ export function Distribution({
 
         {/* Bars */}
         {counts.map((c, i) => {
-          const h = (c / maxCount) * innerH;
+          const h = (Math.min(c, yMax) / yMax) * innerH;
           const x = PAD.left + i * barW;
           const y = PAD.top + (innerH - h);
           return (
@@ -188,7 +233,7 @@ export function Distribution({
           opacity={0.55}
           textAnchor="middle"
         >
-          value ({unit})
+          {t.valueAxis(unit)}
         </text>
         <text
           x={10}
@@ -199,7 +244,7 @@ export function Distribution({
           textAnchor="middle"
           transform={`rotate(-90 10 ${H / 2})`}
         >
-          count
+          {t.countAxis}
         </text>
 
         {/* Percentile legend chips */}
@@ -207,7 +252,6 @@ export function Distribution({
           const chipY = H - 8;
           const chipW = innerW / GUIDES.length;
           return GUIDES.map(({ label: ql, q, color }, i) => {
-            const v = quantile(sorted, q);
             const x = PAD.left + i * chipW;
             return (
               <g key={ql}>
@@ -221,13 +265,18 @@ export function Distribution({
                   strokeDasharray="5 3"
                 />
                 <text x={x + 18} y={chipY} fontSize={11} fill="currentColor" opacity={0.9}>
-                  {ql} {fmt(v)}
+                  {ql} {fmt(quantile(sorted, q))}
                 </text>
               </g>
             );
           });
         })()}
       </ChartHover>
+      {excluded > 0 && (
+        <div className="mt-1 text-xs text-muted-foreground">
+          {t.excluded(excluded.toLocaleString(), unit)}
+        </div>
+      )}
     </div>
   );
 }
diff --git a/packages/app/src/components/inference/agentic-point/log-text.test.ts b/packages/app/src/components/inference/agentic-point/log-text.test.ts
new file mode 100644
index 000000000..56d782186
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/log-text.test.ts
@@ -0,0 +1,97 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  buildLogLines,
+  isNearLogBottom,
+  logLineSeverity,
+  readableLogText,
+  utf16IndexAtCodePointOffset,
+} from './log-text';
+
+describe('readableLogText', () => {
+  it('removes ANSI color and OSC title sequences', () => {
+    expect(readableLogText('\u001B[1;33mwarning\u001B[0m\n\u001B]0;worker-1\u0007ready')).toBe(
+      'warning\nready',
+    );
+  });
+
+  it('normalizes carriage-return line endings without changing tabs', () => {
+    expect(readableLogText('one\r\ntwo\rthree\tvalue')).toBe('one\ntwo\nthree\tvalue');
+  });
+});
+
+describe('utf16IndexAtCodePointOffset', () => {
+  it('maps database character offsets across non-BMP characters', () => {
+    const value = 'before 🚀 match';
+    expect(utf16IndexAtCodePointOffset(value, 9)).toBe(10);
+    expect(value.slice(utf16IndexAtCodePointOffset(value, 9)!)).toBe('match');
+  });
+
+  it('accepts the end boundary and rejects invalid or out-of-range offsets', () => {
+    expect(utf16IndexAtCodePointOffset('a🚀', 2)).toBe(3);
+    expect(utf16IndexAtCodePointOffset('a🚀', 3)).toBeNull();
+    expect(utf16IndexAtCodePointOffset('text', -1)).toBeNull();
+  });
+});
+
+describe('isNearLogBottom', () => {
+  it('does not auto-load an unscrollable log with one very long line', () => {
+    expect(isNearLogBottom({ scrollTop: 0, clientHeight: 448, scrollHeight: 448 })).toBe(false);
+  });
+
+  it('loads only after a vertically overflowing viewport is scrolled near the bottom', () => {
+    expect(isNearLogBottom({ scrollTop: 100, clientHeight: 448, scrollHeight: 1_000 })).toBe(false);
+    expect(isNearLogBottom({ scrollTop: 500, clientHeight: 448, scrollHeight: 1_000 })).toBe(true);
+  });
+});
+
+describe('logLineSeverity', () => {
+  it('classifies conventional uppercase level tokens', () => {
+    expect(logLineSeverity('ERROR 2026-06-12 engine iteration timed out')).toBe('error');
+    expect(logLineSeverity('Traceback (most recent call last)')).toBe('error');
+    expect(logLineSeverity('WARNING [scheduler.py:1120] preempted')).toBe('warn');
+    expect(logLineSeverity('DEBUG [loggers.py:12] step=1')).toBe('debug');
+    expect(logLineSeverity('INFO [api_server.py:1421] ready')).toBe('info');
+  });
+
+  it('ignores level words that only appear far into a long line', () => {
+    expect(logLineSeverity(`INFO ${'x'.repeat(400)} ERROR`)).toBe('info');
+  });
+});
+
+describe('buildLogLines', () => {
+  it('splits plain text into lines without a trailing empty line', () => {
+    const lines = buildLogLines([{ text: 'INFO one\nWARNING two\n', highlighted: false }]);
+    expect(lines).toHaveLength(2);
+    expect(lines[0]).toEqual({
+      text: 'INFO one',
+      pieces: [{ text: 'INFO one', highlighted: false }],
+      severity: 'info',
+    });
+    expect(lines[1]?.severity).toBe('warn');
+  });
+
+  it('keeps a mid-line highlight inside the line it starts on', () => {
+    const lines = buildLogLines([
+      { text: 'INFO before ', highlighted: false },
+      { text: 'router ready', highlighted: true },
+      { text: ' after\nnext line', highlighted: false },
+    ]);
+    expect(lines).toHaveLength(2);
+    // `text` rejoins the pieces, so the memoized plain-row path and the
+    // per-piece highlight path always render the same characters.
+    expect(lines[0]?.text).toBe('INFO before router ready after');
+    expect(lines[0]?.pieces).toEqual([
+      { text: 'INFO before ', highlighted: false },
+      { text: 'router ready', highlighted: true },
+      { text: ' after', highlighted: false },
+    ]);
+    expect(lines[1]?.pieces).toEqual([{ text: 'next line', highlighted: false }]);
+  });
+
+  it('preserves blank lines and an empty highlight so the jump target renders', () => {
+    expect(buildLogLines([{ text: 'a\n\nb', highlighted: false }])).toHaveLength(3);
+    const empty = buildLogLines([{ text: '', highlighted: true }]);
+    expect(empty[0]?.pieces).toEqual([{ text: '', highlighted: true }]);
+  });
+});
diff --git a/packages/app/src/components/inference/agentic-point/log-text.ts b/packages/app/src/components/inference/agentic-point/log-text.ts
new file mode 100644
index 000000000..c8734dd0e
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/log-text.ts
@@ -0,0 +1,105 @@
+// oxlint-disable-next-line no-control-regex -- terminal logs intentionally contain ANSI controls
+const ANSI_OSC_PATTERN = /\u001B\][^\u0007]*(?:\u0007|\u001B\\)/gu;
+// oxlint-disable-next-line no-control-regex -- terminal logs intentionally contain ANSI controls
+const ANSI_CSI_PATTERN = /\u001B\[[0-?]*[ -/]*[@-~]/gu;
+
+/** Remove terminal control sequences while preserving the original log text. */
+export function readableLogText(value: string): string {
+  return value
+    .replace(ANSI_OSC_PATTERN, '')
+    .replace(ANSI_CSI_PATTERN, '')
+    .replaceAll('\r\n', '\n')
+    .replaceAll('\r', '\n');
+}
+
+/** Convert a PostgreSQL character offset to a JavaScript UTF-16 string index. */
+export function utf16IndexAtCodePointOffset(value: string, offset: number): number | null {
+  if (!Number.isInteger(offset) || offset < 0) return null;
+  let codePointOffset = 0;
+  let utf16Index = 0;
+  while (utf16Index < value.length && codePointOffset < offset) {
+    const codePoint = value.codePointAt(utf16Index);
+    utf16Index += codePoint !== undefined && codePoint > 65_535 ? 2 : 1;
+    codePointOffset++;
+  }
+  return codePointOffset === offset ? utf16Index : null;
+}
+
+/** Only user-driven vertical scrolling should request the next log chunk. */
+export function isNearLogBottom({
+  scrollTop,
+  clientHeight,
+  scrollHeight,
+}: {
+  scrollTop: number;
+  clientHeight: number;
+  scrollHeight: number;
+}): boolean {
+  return scrollTop > 0 && scrollTop + clientHeight >= scrollHeight - 160;
+}
+
+/** Conventional log levels we tint in the viewer gutter and line background. */
+export type LogSeverity = 'error' | 'warn' | 'debug' | 'info';
+
+/** One run of characters inside a rendered line; `highlighted` marks a search hit. */
+export interface LogLinePiece {
+  text: string;
+  highlighted: boolean;
+}
+
+export interface LogLine {
+  /** The whole line as plain text — the concatenation of `pieces`. */
+  text: string;
+  pieces: LogLinePiece[];
+  severity: LogSeverity;
+}
+
+// Only the head of a line is scanned so a long INFO line that happens to quote
+// the word ERROR further along is not mis-tinted (and so severity stays O(1)).
+const SEVERITY_SCAN_LENGTH = 200;
+const ERROR_PATTERN = /\b(?:ERROR|FATAL|CRITICAL|PANIC)\b|^Traceback \(most recent call last\)/;
+const WARN_PATTERN = /\b(?:WARNING|WARN)\b/;
+const DEBUG_PATTERN = /\b(?:DEBUG|TRACE)\b/;
+
+/** Classify a log line by its conventional uppercase level token. */
+export function logLineSeverity(line: string): LogSeverity {
+  const head = line.length > SEVERITY_SCAN_LENGTH ? line.slice(0, SEVERITY_SCAN_LENGTH) : line;
+  if (ERROR_PATTERN.test(head)) return 'error';
+  if (WARN_PATTERN.test(head)) return 'warn';
+  if (DEBUG_PATTERN.test(head)) return 'debug';
+  return 'info';
+}
+
+/**
+ * Split ordered text runs into rendered lines, preserving which runs are search
+ * highlights. Runs are split rather than the joined string so a highlight that
+ * starts mid-line stays a single `<mark>` inside that one line.
+ */
+export function buildLogLines(segments: readonly LogLinePiece[]): LogLine[] {
+  const lines: LogLine[] = [];
+  let pieces: LogLinePiece[] = [];
+  let plainText = '';
+
+  const endLine = () => {
+    lines.push({ text: plainText, pieces, severity: logLineSeverity(plainText) });
+    pieces = [];
+    plainText = '';
+  };
+
+  for (const segment of segments) {
+    const parts = segment.text.split('\n');
+    for (const [index, part] of parts.entries()) {
+      if (index > 0) endLine();
+      // Empty highlighted runs are kept so the jump target always has a `<mark>`.
+      if (part !== '' || segment.highlighted) {
+        pieces.push({ text: part, highlighted: segment.highlighted });
+        plainText += part;
+      }
+    }
+  }
+  endLine();
+
+  // A trailing newline ends the last line rather than starting an empty one.
+  if (lines.length > 1 && lines.at(-1)?.pieces.length === 0) lines.pop();
+  return lines;
+}
diff --git a/packages/app/src/components/inference/agentic-point/lognormal.test.ts b/packages/app/src/components/inference/agentic-point/lognormal.test.ts
new file mode 100644
index 000000000..d80355cfa
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/lognormal.test.ts
@@ -0,0 +1,87 @@
+import { describe, expect, it } from 'vitest';
+
+import { logHistogram, logTicks, positiveValues } from './lognormal';
+
+describe('positiveValues', () => {
+  it('drops zero, negative, and non-finite samples that cannot be logged', () => {
+    expect(positiveValues([5, 0, -3, Number.NaN, Number.POSITIVE_INFINITY, 2])).toEqual([5, 2]);
+  });
+});
+
+describe('logHistogram', () => {
+  it('bins with equal width in log space and counts every positive sample', () => {
+    const histogram = logHistogram([1, 10, 100, 1000], 3);
+    expect(histogram).not.toBeNull();
+    expect(histogram!.counts.reduce((a, b) => a + b, 0)).toBe(4);
+    expect(histogram!.edges).toHaveLength(4);
+    // Equal ratios between edges is what "equal width in log space" means.
+    const ratios = histogram!.edges.slice(1).map((edge, i) => edge / histogram!.edges[i]!);
+    for (const ratio of ratios) expect(ratio).toBeCloseTo(10, 6);
+  });
+
+  it('places the maximum sample in the last bin rather than past the end', () => {
+    const histogram = logHistogram([1, 2, 4, 8, 16], 4);
+    expect(histogram!.counts.at(-1)).toBeGreaterThanOrEqual(1);
+    expect(histogram!.counts.reduce((a, b) => a + b, 0)).toBe(5);
+  });
+
+  it('widens a single-valued range so the spike still has a bin', () => {
+    const histogram = logHistogram([256, 256, 256], 8);
+    expect(histogram).not.toBeNull();
+    expect(histogram!.lnMax).toBeGreaterThan(histogram!.lnMin);
+    expect(histogram!.counts.reduce((a, b) => a + b, 0)).toBe(3);
+  });
+
+  it('returns null when no sample can be logged', () => {
+    expect(logHistogram([], 10)).toBeNull();
+    expect(logHistogram([0, -4], 10)).toBeNull();
+  });
+
+  it('handles request counts above the JavaScript variadic argument limit', () => {
+    const values = Array.from({ length: 200_000 }, (_, index) => (index % 4096) + 1);
+
+    const histogram = logHistogram(values, 32);
+
+    expect(histogram).not.toBeNull();
+    expect(histogram!.counts.reduce((sum, count) => sum + count, 0)).toBe(values.length);
+    expect(histogram!.edges[0]).toBeCloseTo(1, 10);
+    expect(histogram!.edges.at(-1)).toBeCloseTo(4096, 6);
+  });
+});
+
+describe('logTicks', () => {
+  it('anchors the endpoints and puts decades in between', () => {
+    const ticks = logTicks(90, 120_000);
+    expect(ticks[0]).toBe(90);
+    expect(ticks.at(-1)).toBe(120_000);
+    expect(ticks).toContain(1000);
+    expect(ticks).toContain(10_000);
+    expect(ticks.toSorted((a, b) => a - b)).toEqual(ticks);
+  });
+
+  it('subdivides a narrow range so the axis is not left bare', () => {
+    const ticks = logTicks(100, 400);
+    expect(ticks).toContain(200);
+    expect(ticks.length).toBeGreaterThan(2);
+  });
+
+  it('drops a decade that would print on top of an endpoint label', () => {
+    // 95 sits just below 100; both labels would collide at the axis origin.
+    const ticks = logTicks(95, 100_000);
+    expect(ticks[0]).toBe(95);
+    expect(ticks).not.toContain(100);
+    expect(ticks).toContain(1000);
+  });
+
+  it('always ends at the data maximum, even when the last decade crowds it', () => {
+    // 10,200 is a hair above 10,000: the decade must yield, not the endpoint.
+    const ticks = logTicks(10, 10_200);
+    expect(ticks.at(-1)).toBe(10_200);
+    expect(ticks).not.toContain(10_000);
+  });
+
+  it('returns nothing for a range a log axis cannot draw', () => {
+    expect(logTicks(0, 100)).toEqual([]);
+    expect(logTicks(100, 100)).toEqual([]);
+  });
+});
diff --git a/packages/app/src/components/inference/agentic-point/lognormal.ts b/packages/app/src/components/inference/agentic-point/lognormal.ts
new file mode 100644
index 000000000..3a3f2a5be
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/lognormal.ts
@@ -0,0 +1,109 @@
+/**
+ * Log-spaced binning for the ISL/OSL distribution charts.
+ *
+ * Per-request sequence lengths span orders of magnitude (a few hundred tokens
+ * to tens of thousands), so a linear histogram collapses almost every request
+ * into the leftmost bins and leaves a long, unreadable tail. Binning uniformly
+ * in ln(x) instead spreads the mass out, and the roughly lognormal shape then
+ * reads as an ordinary bell.
+ *
+ * Every function here takes and returns plain numbers so the chart component
+ * stays presentational and this math is unit-testable on its own.
+ */
+
+/** Uniform-in-ln bins over a positive range. */
+export interface LogHistogram {
+  /** `bins + 1` bin edges in the original units, ascending. */
+  edges: number[];
+  /** Count per bin; `edges.length - 1` entries. */
+  counts: number[];
+  /** Bin width in ln space — constant, and what the fitted curve scales by. */
+  lnStep: number;
+  /** ln of the first and last edge. */
+  lnMin: number;
+  lnMax: number;
+}
+
+/**
+ * Only strictly positive samples can be logged. A request that produced zero
+ * tokens is real data, so callers report how many were set aside rather than
+ * silently folding them into the first bin.
+ */
+export function positiveValues(values: readonly number[]): number[] {
+  return values.filter((v) => Number.isFinite(v) && v > 0);
+}
+
+/**
+ * Histogram with bins of equal width in ln(x). `bins` is clamped to at least 1.
+ * When every sample is the same value the range is widened by a factor of two
+ * either side so the single spike still has somewhere to sit.
+ */
+export function logHistogram(values: readonly number[], bins: number): LogHistogram | null {
+  const positive = positiveValues(values);
+  if (positive.length === 0) return null;
+
+  const nBins = Math.max(1, Math.floor(bins));
+  // Do not spread this array into Math.min/Math.max. Large agentic runs have
+  // hundreds of thousands of requests, which exceeds JavaScript engines'
+  // variadic argument limit and crashes the entire point page.
+  let min = positive[0]!;
+  let max = min;
+  for (let i = 1; i < positive.length; i++) {
+    const value = positive[i]!;
+    if (value < min) min = value;
+    if (value > max) max = value;
+  }
+  let lnMin = Math.log(min);
+  let lnMax = Math.log(max);
+  if (!(lnMax > lnMin)) {
+    lnMin -= Math.LN2;
+    lnMax += Math.LN2;
+  }
+
+  const lnStep = (lnMax - lnMin) / nBins;
+  const counts: number[] = Array.from({ length: nBins }, () => 0);
+  for (const value of positive) {
+    const index = Math.min(nBins - 1, Math.floor((Math.log(value) - lnMin) / lnStep));
+    counts[index]!++;
+  }
+
+  const edges = Array.from({ length: nBins + 1 }, (_, i) => Math.exp(lnMin + i * lnStep));
+  return { edges, counts, lnStep, lnMin, lnMax };
+}
+
+/**
+ * Tick values for a log axis: every power of ten in range, plus its 2x and 5x
+ * steps when the range is narrow enough that decades alone would leave the axis
+ * nearly bare. Always includes the endpoints so the axis is anchored.
+ */
+export function logTicks(min: number, max: number): number[] {
+  if (!(min > 0) || !(max > min)) return [];
+  const decades = Math.log10(max) - Math.log10(min);
+  const multipliers = decades > 3 ? [1] : decades > 1.5 ? [1, 5] : [1, 2, 5];
+  const candidates: number[] = [min];
+  for (let exp = Math.floor(Math.log10(min)); exp <= Math.ceil(Math.log10(max)); exp++) {
+    for (const multiplier of multipliers) {
+      const tick = multiplier * 10 ** exp;
+      if (tick > min && tick < max) candidates.push(tick);
+    }
+  }
+  candidates.push(max);
+
+  // A data endpoint can land just short of a decade (min 95, first decade 100),
+  // which would print two labels on top of each other. Keep a tick only once it
+  // clears a fraction of the axis, and let the max endpoint evict its neighbour
+  // rather than be dropped itself — the axis has to end at the data.
+  const minGap = decades * 0.06;
+  const kept: number[] = [];
+  for (const tick of candidates) {
+    const previous = kept.at(-1);
+    if (previous === undefined || Math.log10(tick) - Math.log10(previous) >= minGap) {
+      kept.push(tick);
+    }
+  }
+  if (kept.at(-1) !== max) {
+    if (kept.length > 1) kept.pop();
+    kept.push(max);
+  }
+  return kept;
+}
diff --git a/packages/app/src/components/inference/agentic-point/metric-source-toolbar.tsx b/packages/app/src/components/inference/agentic-point/metric-source-toolbar.tsx
index e56ddeeee..b4094adfa 100644
--- a/packages/app/src/components/inference/agentic-point/metric-source-toolbar.tsx
+++ b/packages/app/src/components/inference/agentic-point/metric-source-toolbar.tsx
@@ -1,6 +1,6 @@
 'use client';
 
-import type { MetricSource, MetricSourceSeries } from '@/hooks/api/use-trace-server-metrics';
+import type { MetricSource, MetricSourceDescriptor } from '@/hooks/api/use-trace-server-metrics';
 import { SegmentedToggle, type SegmentedToggleOption } from '@/components/ui/segmented-toggle';
 import {
   Select,
@@ -58,8 +58,8 @@ export function MetricSourceToolbar({
   hasWarmup: boolean;
   phase: StagePhase;
   onPhaseChange: (phase: StagePhase) => void;
-  metricSources: MetricSourceSeries[];
-  selectedSource: MetricSourceSeries | undefined;
+  metricSources: MetricSourceDescriptor[];
+  selectedSource: MetricSourceDescriptor | undefined;
   onSourceChange: (id: string) => void;
   /** Adapter reported in analytics when the selected source lookup misses. */
   fallbackAdapter: string | undefined;
diff --git a/packages/app/src/components/inference/agentic-point/phase-slice.test.ts b/packages/app/src/components/inference/agentic-point/phase-slice.test.ts
index ef6cdaabb..1db42acac 100644
--- a/packages/app/src/components/inference/agentic-point/phase-slice.test.ts
+++ b/packages/app/src/components/inference/agentic-point/phase-slice.test.ts
@@ -1,11 +1,13 @@
 import { describe, expect, it } from 'vitest';
 
 import type { RequestRecord, RequestTimeline } from '@/hooks/api/use-request-timeline';
+import type { RequestChartData } from '@/hooks/api/use-request-chart-data';
 import {
   phaseBoundaryNs,
   phaseBoundarySec,
   requestsForPhase,
   sliceServerSeriesByPhase,
+  sliceRequestChartDataByPhase,
   sliceTimelineByPhase,
   timelineHasWarmup,
   type ServerSeriesLike,
@@ -210,3 +212,45 @@ describe('sliceTimelineByPhase', () => {
     expect(out.startNs).toBe(1_100);
   });
 });
+
+describe('sliceRequestChartDataByPhase', () => {
+  it('filters and rebases the compact request payload without Gantt-only fields', () => {
+    const data: RequestChartData = {
+      version: 1,
+      timelineVersion: 5,
+      startNs: 1_000,
+      endNs: 4_000,
+      durationS: 3,
+      requests: [
+        {
+          cid: 'warmup',
+          phase: 'warmup',
+          start: 0,
+          end: 50,
+          ttftMs: 1,
+          tpotMs: 2,
+          isl: 3,
+          osl: 4,
+          cancelled: false,
+        },
+        {
+          cid: 'profiling',
+          phase: 'profiling',
+          start: 100,
+          end: 300,
+          ttftMs: 5,
+          tpotMs: 6,
+          isl: 7,
+          osl: 8,
+          cancelled: false,
+        },
+      ],
+    };
+
+    const result = sliceRequestChartDataByPhase(data, 'profiling');
+    expect(result.requests).toEqual([
+      expect.objectContaining({ cid: 'profiling', start: 0, end: 200 }),
+    ]);
+    expect(result.startNs).toBe(1_100);
+  });
+});
diff --git a/packages/app/src/components/inference/agentic-point/phase-slice.ts b/packages/app/src/components/inference/agentic-point/phase-slice.ts
index e6e17719e..2983b38e8 100644
--- a/packages/app/src/components/inference/agentic-point/phase-slice.ts
+++ b/packages/app/src/components/inference/agentic-point/phase-slice.ts
@@ -15,6 +15,7 @@
  * off by the origin gap. This rebasing lives in `phaseBoundarySec` only.
  */
 
+import type { RequestChartData, RequestChartRecord } from '@/hooks/api/use-request-chart-data';
 import type { RequestRecord, RequestTimeline } from '@/hooks/api/use-request-timeline';
 import type {
   QueueDepthPoint,
@@ -42,7 +43,9 @@ export interface ServerSeriesLike {
 }
 
 /** True when the timeline contains at least one non-profiling (warmup) request. */
-export function timelineHasWarmup(timeline: RequestTimeline | null | undefined): boolean {
+type RequestPhaseData = Pick<RequestChartData, 'startNs' | 'durationS' | 'requests'>;
+
+export function timelineHasWarmup(timeline: RequestPhaseData | null | undefined): boolean {
   return Boolean(timeline?.requests.some((r) => r.phase !== 'profiling'));
 }
 
@@ -52,7 +55,7 @@ export function timelineHasWarmup(timeline: RequestTimeline | null | undefined):
  * Returns null unless BOTH a warmup and a profiling request exist (nothing to
  * split otherwise).
  */
-export function phaseBoundaryNs(timeline: RequestTimeline | null | undefined): number | null {
+export function phaseBoundaryNs(timeline: RequestPhaseData | null | undefined): number | null {
   if (!timeline) return null;
   let hasWarmup = false;
   let minProfilingStart: number | null = null;
@@ -77,7 +80,7 @@ export function phaseBoundaryNs(timeline: RequestTimeline | null | undefined): n
  */
 export function phaseBoundarySec(
   serverMetrics: Pick<TraceServerMetrics, 'startNs'> | null | undefined,
-  timeline: RequestTimeline | null | undefined,
+  timeline: RequestPhaseData | null | undefined,
 ): number | null {
   if (!serverMetrics) return null;
   const boundaryNs = phaseBoundaryNs(timeline);
@@ -186,3 +189,24 @@ export function sliceTimelineByPhase(
     phase === 'warmup' ? boundaryOff / 1e9 : Math.max(1, timeline.durationS - boundaryOff / 1e9);
   return { ...timeline, startNs: timeline.startNs + shift, requests, durationS };
 }
+
+/** Compact-data counterpart to `sliceTimelineByPhase` for summary charts. */
+export function sliceRequestChartDataByPhase(
+  data: RequestChartData,
+  phase: StagePhase,
+): RequestChartData {
+  const boundaryNs = phaseBoundaryNs(data);
+  if (boundaryNs === null) return data;
+  const boundaryOff = boundaryNs - data.startNs;
+  const inPhase = (request: RequestChartRecord) =>
+    phase === 'warmup' ? request.start < boundaryOff : request.start >= boundaryOff;
+  const shift = phase === 'profiling' ? boundaryOff : 0;
+  const requests = data.requests.filter(inPhase).map((request) => ({
+    ...request,
+    start: request.start - shift,
+    end: request.end - shift,
+  }));
+  const durationS =
+    phase === 'warmup' ? boundaryOff / 1e9 : Math.max(1, data.durationS - boundaryOff / 1e9);
+  return { ...data, startNs: data.startNs + shift, requests, durationS };
+}
diff --git a/packages/app/src/components/inference/agentic-point/point-summary.test.ts b/packages/app/src/components/inference/agentic-point/point-summary.test.ts
index 1e74ceeaf..7fef6f8d0 100644
--- a/packages/app/src/components/inference/agentic-point/point-summary.test.ts
+++ b/packages/app/src/components/inference/agentic-point/point-summary.test.ts
@@ -107,8 +107,8 @@ describe('PointSummary', () => {
     );
     localeState.pathname = '/inference/agentic/206885';
 
-    expect(html).toContain('卸载类型');
-    expect(html).toContain('KV 卸载引擎');
+    expect(html).toContain('offload 类型');
+    expect(html).toContain('KV offload 引擎');
     expect(html).toContain('HiCache');
     expect(html).toContain('路由器');
     expect(html).toContain('SGLang Router 0.3.2');
diff --git a/packages/app/src/components/inference/agentic-point/point-summary.tsx b/packages/app/src/components/inference/agentic-point/point-summary.tsx
index 0ba20c875..541c34104 100644
--- a/packages/app/src/components/inference/agentic-point/point-summary.tsx
+++ b/packages/app/src/components/inference/agentic-point/point-summary.tsx
@@ -33,12 +33,12 @@ const STRINGS = {
     disagg: '解耦',
     multiNodeAggregate: '多节点聚合',
     githubRun: 'GitHub Actions 运行 →',
-    offloadType: '卸载类型',
-    offloadBackend: 'KV 卸载引擎',
+    offloadType: 'offload 类型',
+    offloadBackend: 'KV offload 引擎',
     transferEngine: 'KV 传输引擎',
     router: '路由器',
     concurrency: '并发数',
-    gpuCacheHit: 'Chip Cache 命中率',
+    gpuCacheHit: '芯片 Cache 命中率',
     cpuCacheHit: 'CPU Cache 命中率',
     enabledLegacy: '已启用（旧版数据）',
     disabledLegacy: '已禁用（旧版数据）',
diff --git a/packages/app/src/components/inference/agentic-point/request-metric-cards.tsx b/packages/app/src/components/inference/agentic-point/request-metric-cards.tsx
index 8ca85ac94..233ae5f96 100644
--- a/packages/app/src/components/inference/agentic-point/request-metric-cards.tsx
+++ b/packages/app/src/components/inference/agentic-point/request-metric-cards.tsx
@@ -1,8 +1,8 @@
 'use client';
 
-import { useState } from 'react';
+import { useMemo, useState } from 'react';
 
-import type { RequestTimeline } from '@/hooks/api/use-request-timeline';
+import type { RequestChartData } from '@/hooks/api/use-request-chart-data';
 import { SegmentedToggle, type SegmentedToggleOption } from '@/components/ui/segmented-toggle';
 import { track } from '@/lib/analytics';
 
@@ -47,16 +47,18 @@ export function RequestMetricOverTime({
 }: {
   title: string;
   metric: RequestMetric;
-  timeline: RequestTimeline | null | undefined;
+  timeline: RequestChartData | null | undefined;
   isLoading: boolean;
   latencySelector?: boolean;
 }) {
   const [percentile, setPercentile] = useState<RequestPercentile>('p90');
   const [latencyMetric, setLatencyMetric] = useState<'ttft' | 'e2e'>('ttft');
   const selectedMetric = latencySelector ? latencyMetric : metric;
-  const result = timeline
-    ? rollingRequestMetric(timeline.requests, selectedMetric, percentile, 50)
-    : null;
+  const result = useMemo(
+    () =>
+      timeline ? rollingRequestMetric(timeline.requests, selectedMetric, percentile, 50) : null,
+    [timeline, selectedMetric, percentile],
+  );
   const metricLabel =
     selectedMetric === 'ttft' ? 'TTFT' : selectedMetric === 'e2e' ? 'E2E latency' : 'Interactivity';
   const color =
@@ -151,7 +153,7 @@ export function SequenceMetricCard({
 }: {
   metric: 'isl' | 'osl';
   /** Phase-scoped timeline — distribution values + in-flight are both derived from it. */
-  timeline: RequestTimeline | null | undefined;
+  timeline: RequestChartData | null | undefined;
   timelineLoading: boolean;
 }) {
   const [view, setView] = useState<SequenceMetricView>('distribution');
@@ -160,11 +162,19 @@ export function SequenceMetricCard({
   const testPrefix = `${metric}-metric`;
   // Per-request ISL/OSL for the selected phase (request_timeline carries both,
   // so the distribution honours the warmup/profiling toggle for free).
-  const values = timeline
-    ? timeline.requests
-        .map((r) => r[metric])
-        .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))
-    : undefined;
+  const values = useMemo(
+    () =>
+      timeline
+        ? timeline.requests
+            .map((r) => r[metric])
+            .filter((v): v is number => typeof v === 'number' && Number.isFinite(v))
+        : undefined,
+    [timeline, metric],
+  );
+  const inflightData = useMemo(
+    () => (timeline ? averageSequenceLengthInFlight(timeline.requests, metric) : null),
+    [timeline, metric],
+  );
   return (
     <ExpandableChart
       title={view === 'distribution' ? `${fullName} distribution` : `Average ${acronym} in flight`}
@@ -193,7 +203,7 @@ export function SequenceMetricCard({
           return timelineLoading ? <ChartSkeleton /> : <ChartEmpty />;
         }
         if (!timeline) return timelineLoading ? <ChartSkeleton /> : <ChartEmpty />;
-        const raw = averageSequenceLengthInFlight(timeline.requests, metric);
+        const raw = inflightData ?? [];
         return (
           <div>
             {metric === 'osl' && (
diff --git a/packages/app/src/components/inference/agentic-point/request-timeline.test.ts b/packages/app/src/components/inference/agentic-point/request-timeline.test.ts
index cf43f5aea..cffeb5f95 100644
--- a/packages/app/src/components/inference/agentic-point/request-timeline.test.ts
+++ b/packages/app/src/components/inference/agentic-point/request-timeline.test.ts
@@ -167,16 +167,77 @@ describe('subagent timeline hierarchy', () => {
 
   it('deep-links a main-agent aux request to the parent conversation without sa', () => {
     expect(conversationHref('slug', { ...request(0, 10), cid: 'abc123::aux:red:002', ti: 3 })).toBe(
-      '/datasets/slug/conversations/abc123?turn=3',
+      '/agentx/slug/conversations/abc123?turn=3',
     );
   });
 });
 
+describe('reused conversation replay lanes', () => {
+  it('renders overlapping replays of one source conversation on separate rows', () => {
+    const records = [
+      { ...request(0, 100), cid: 'reused-conversation', ti: 26, ri: 0 },
+      { ...request(110, 180), cid: 'reused-conversation', ti: 27, ri: 0 },
+      { ...request(50, 140), cid: 'reused-conversation', ti: 40, ri: 1 },
+      {
+        ...request(70, 120),
+        cid: 'reused-conversation::sa:subagent_001_abcd',
+        ti: 40,
+        ri: 1,
+      },
+    ];
+
+    const rows = buildRequestTimelineRows(records, 'conversation', new Set());
+    const parents = rows.filter((row) => row.kind === 'parent');
+    expect(parents.map((row) => row.label)).toEqual([
+      'reused-conversation · replay 1',
+      'reused-conversation · replay 2',
+    ]);
+    expect(parents.map((row) => row.requests.map((record) => record.ti))).toEqual([[26, 27], [40]]);
+    expect(rows.find((row) => row.kind === 'subagent')?.key).toContain('replay-lane:1');
+  });
+
+  it('localizes replay labels in the Chinese timeline', () => {
+    const rows = buildRequestTimelineRows(
+      [{ ...request(0, 10), cid: 'conversation', ri: 2 }],
+      'conversation',
+      new Set(),
+      undefined,
+      'zh',
+    );
+    expect(rows[0]?.label).toBe('conversation · 重放 3');
+  });
+
+  it('keeps a replay lane stable across warmup and profiling filters', () => {
+    const records = [
+      { ...request(0, 10), cid: 'conversation', ri: 0, phase: 'warmup' },
+      { ...request(100, 110), cid: 'conversation', ri: 0, phase: 'profiling' },
+      { ...request(50, 60), cid: 'conversation', ri: 1, phase: 'profiling' },
+    ];
+    const index = computeStableRowIndex(records, 'conversation');
+    const warmup = buildRequestTimelineRows(
+      records.filter((record) => record.phase === 'warmup'),
+      'conversation',
+      new Set(),
+      index,
+    );
+    const profiling = buildRequestTimelineRows(
+      records.filter((record) => record.phase === 'profiling'),
+      'conversation',
+      new Set(),
+      index,
+    );
+
+    expect(warmup[0]?.label).toBe('conversation · replay 1');
+    expect(profiling[0]?.label).toBe('conversation · replay 1');
+    expect(warmup[0]?.color).toBe(profiling[0]?.color);
+  });
+});
+
 describe('conversationHref', () => {
   it('builds a turn-carrying dataset link for a main-conversation request', () => {
     expect(
       conversationHref('cc-traces-weka-062126', { ...request(0, 10), cid: 'abc123', ti: 4 }),
-    ).toBe('/datasets/cc-traces-weka-062126/conversations/abc123?turn=4');
+    ).toBe('/agentx/cc-traces-weka-062126/conversations/abc123?turn=4');
   });
 
   it('carries the subagent id and strips the ::sa suffix from the conv id', () => {
@@ -186,7 +247,7 @@ describe('conversationHref', () => {
         cid: 'abc123::sa:subagent_001_bf1c5c16:s2',
         ti: 7,
       }),
-    ).toBe('/datasets/slug/conversations/abc123?turn=7&sa=subagent_001_bf1c5c16');
+    ).toBe('/agentx/slug/conversations/abc123?turn=7&sa=subagent_001_bf1c5c16');
   });
 
   it('uses raw source provenance for flattened-agent dataset links', () => {
@@ -199,7 +260,7 @@ describe('conversationHref', () => {
         srcOuter: 204,
         srcKind: 'weka_flat',
       }),
-    ).toBe('/datasets/slug/conversations/02bc0afb13f7a2d9efa86c28511261d85c0e?turn=3&raw=204');
+    ).toBe('/agentx/slug/conversations/02bc0afb13f7a2d9efa86c28511261d85c0e?turn=3&raw=204');
   });
 
   it('uses raw nested source provenance for subagent child links', () => {
@@ -214,7 +275,7 @@ describe('conversationHref', () => {
         srcKind: 'weka_subagent',
       }),
     ).toBe(
-      '/datasets/slug/conversations/117ebe75819d050f308a0a81647893abd02d?turn=16&raw=39&inner=16',
+      '/agentx/slug/conversations/117ebe75819d050f308a0a81647893abd02d?turn=16&raw=39&inner=16',
     );
   });
 });
diff --git a/packages/app/src/components/inference/agentic-point/request-timeline.tsx b/packages/app/src/components/inference/agentic-point/request-timeline.tsx
index 04ee24be0..f7599bf3c 100644
--- a/packages/app/src/components/inference/agentic-point/request-timeline.tsx
+++ b/packages/app/src/components/inference/agentic-point/request-timeline.tsx
@@ -18,8 +18,10 @@ import {
   PADDING_RIGHT,
   ROW_GAP,
   ROW_HEIGHT,
+  ROW_SPAN,
   TIMELINE_BODY_MAX_HEIGHT,
   timelineSvgHeight,
+  visibleTimelineRowRange,
 } from './timeline-layout';
 import {
   buildRequestTimelineRows,
@@ -47,6 +49,7 @@ import {
 export {
   buildRequestTimelineRows,
   computeStableRowIndex,
+  conversationReplayKey,
   conversationHref,
   datasetConvId,
   requestIdleStats,
@@ -86,7 +89,7 @@ const PHASE_OPTIONS: Record<'en' | 'zh', SegmentedToggleOption<PhaseFilter>[]> =
   ],
   zh: [
     { value: 'profiling', label: '性能剖析', testId: 'timeline-phase-profiling' },
-    { value: 'warmup', label: '预热', testId: 'timeline-phase-warmup' },
+    { value: 'warmup', label: 'warmup', testId: 'timeline-phase-warmup' },
   ],
 };
 
@@ -108,6 +111,8 @@ export function RequestTimelineView({
   const [rowMode, setRowMode] = useState<RowMode>('conversation');
   const [phaseFilter, setPhaseFilter] = useState<PhaseFilter>('profiling');
   const [tooltip, setTooltip] = useState<TooltipData | null>(null);
+  const [scrollTop, setScrollTop] = useState(0);
+  const [viewportHeight, setViewportHeight] = useState(TIMELINE_BODY_MAX_HEIGHT);
 
   // The scroll container (vertical row scroll + horizontal chart scroll) and a
   // ref mirror of the live view state, so click-through can snapshot the exact
@@ -185,7 +190,7 @@ export function RequestTimelineView({
     () => sliceTimelineByPhase(data, hasWarmup ? phaseFilter : 'profiling').requests,
     [data, phaseFilter, hasWarmup],
   );
-  // Stable order/color per conversation (or worker), computed over the FULL
+  // Stable order/color per replay lane (or worker), computed over the FULL
   // request set — NOT the phase-filtered subset — so a row keeps its position
   // and color when the user toggles between warmup and profiling.
   const stableRowIndex = useMemo(
@@ -193,8 +198,8 @@ export function RequestTimelineView({
     [data.requests, rowMode],
   );
   const rows = useMemo(
-    () => buildRequestTimelineRows(filtered, rowMode, expandedSubagents, stableRowIndex),
-    [filtered, rowMode, expandedSubagents, stableRowIndex],
+    () => buildRequestTimelineRows(filtered, rowMode, expandedSubagents, stableRowIndex, locale),
+    [filtered, rowMode, expandedSubagents, stableRowIndex, locale],
   );
   const idleStats = useMemo(() => requestIdleStats(filtered), [filtered]);
 
@@ -264,6 +269,21 @@ export function RequestTimelineView({
   }, [pointId]);
 
   const svgHeight = timelineSvgHeight(rows.length);
+  const visibleRange = useMemo(
+    () => visibleTimelineRowRange(rows.length, scrollTop, viewportHeight),
+    [rows.length, scrollTop, viewportHeight],
+  );
+  const visibleRows = rows.slice(visibleRange.start, visibleRange.end);
+
+  useLayoutEffect(() => {
+    const element = scrollRef.current;
+    if (!element) return;
+    const updateHeight = () => setViewportHeight(element.clientHeight);
+    updateHeight();
+    const observer = new ResizeObserver(updateHeight);
+    observer.observe(element);
+    return () => observer.disconnect();
+  }, [rows.length]);
 
   // Native (non-passive) wheel handler: React's synthetic onWheel is attached
   // passively, so preventDefault there is silently ignored and shift+scroll
@@ -456,25 +476,28 @@ export function RequestTimelineView({
             window's bottom edge (rather than the bottom of the tall content). */}
         <div
           ref={scrollRef}
+          data-testid="request-timeline-scroll"
           className="overflow-auto"
           style={{ maxHeight: TIMELINE_BODY_MAX_HEIGHT }}
+          onScroll={(event) => setScrollTop(event.currentTarget.scrollTop)}
         >
           <div className="flex w-max">
             {/* Label column — pinned left (sticky) so it stays put during
                 horizontal scroll, while scrolling vertically with the rows. */}
             <div
               className="sticky left-0 z-10 flex-shrink-0 border-r border-border/60 bg-card"
-              style={{ width: LABEL_WIDTH }}
+              style={{ width: LABEL_WIDTH, height: svgHeight }}
             >
               <div
-                className="border-b border-border/60 flex items-end px-2 pb-1"
+                className="sticky top-0 z-20 border-b border-border/60 bg-card flex items-end px-2 pb-1"
                 style={{ height: HEADER_HEIGHT }}
               >
-                <span className="text-[9px] font-mono font-bold uppercase tracking-[0.15em] text-muted-foreground">
+                <span className="text-[9px] font-mono font-bold uppercase tracking-eyebrow text-muted-foreground">
                   {rowMode === 'conversation' ? 'Conversation' : 'Worker'}
                 </span>
               </div>
-              {rows.map((row) => {
+              {visibleRows.map((row, visibleIndex) => {
+                const rowIndex = visibleRange.start + visibleIndex;
                 const isSubagentRow = row.kind === 'subagent';
                 const isChildRow = row.kind === 'stream' || row.kind === 'aux';
                 const isExpandable = isSubagentRow && (row.streamCount ?? 1) > 1;
@@ -485,6 +508,10 @@ export function RequestTimelineView({
                     data-timeline-row-kind={row.kind}
                     className="flex items-center gap-1 overflow-hidden pr-2"
                     style={{
+                      position: 'absolute',
+                      top: HEADER_HEIGHT + rowIndex * ROW_SPAN,
+                      left: 0,
+                      right: 0,
                       height: ROW_HEIGHT + ROW_GAP,
                       paddingLeft: 4 + row.depth * 10,
                     }}
@@ -497,7 +524,7 @@ export function RequestTimelineView({
                         aria-label={isExpanded ? 'Collapse streams' : 'Expand streams'}
                         title={isExpanded ? 'Collapse streams' : 'Expand streams'}
                       >
-                        <span className="text-[10px] leading-none">{isExpanded ? '▾' : '▸'}</span>
+                        <span className="text-3xs leading-none">{isExpanded ? '▾' : '▸'}</span>
                       </button>
                     ) : (
                       <span className="size-3.5 shrink-0" />
@@ -510,7 +537,7 @@ export function RequestTimelineView({
                       }}
                     />
                     <span
-                      className="text-[10px] font-mono truncate"
+                      className="text-3xs font-mono truncate"
                       style={{
                         color: row.color,
                         opacity: isChildRow ? 0.7 : isSubagentRow ? 0.85 : 1,
@@ -535,12 +562,13 @@ export function RequestTimelineView({
             {/* Chart column — horizontal scrolling is handled by the window
                 container above so its scrollbar stays pinned to the window's
                 bottom edge; double-click anywhere resets the zoom. */}
-            <div className="flex-shrink-0">
+            <div className="flex-shrink-0" style={{ height: svgHeight }}>
               <svg
                 ref={zoomSvgRef}
+                data-testid="request-timeline-svg"
                 width={CHART_WIDTH}
-                height={svgHeight}
-                className="block"
+                height={viewportHeight}
+                className="sticky top-0 block bg-card"
                 style={{ cursor: isZoomed ? 'grab' : 'crosshair' }}
                 onMouseDown={handleMouseDown}
                 onMouseMove={handleMouseMove}
@@ -549,7 +577,10 @@ export function RequestTimelineView({
                 onDoubleClick={resetZoom}
               >
                 <TimelineBars
-                  rows={rows}
+                  rows={visibleRows}
+                  firstRowIndex={visibleRange.start}
+                  scrollTop={scrollTop}
+                  viewportHeight={viewportHeight}
                   expandedSubagents={expandedSubagents}
                   dataStart={dataStart}
                   vStart={vStart}
@@ -568,7 +599,7 @@ export function RequestTimelineView({
                     x1={cursor.xPx}
                     x2={cursor.xPx}
                     y1={0}
-                    y2={svgHeight}
+                    y2={viewportHeight}
                     stroke="currentColor"
                     strokeWidth={1}
                     opacity={0.45}
@@ -582,7 +613,7 @@ export function RequestTimelineView({
       </div>
 
       {/* Footer — interaction hint only. */}
-      <div className="flex items-center px-1 text-[11px] text-muted-foreground">
+      <div className="flex items-center px-1 text-2xs text-muted-foreground">
         <span className="ml-auto opacity-70">
           shift+scroll to zoom · drag to pan · double-click to reset
         </span>
diff --git a/packages/app/src/components/inference/agentic-point/server-log-viewer.tsx b/packages/app/src/components/inference/agentic-point/server-log-viewer.tsx
new file mode 100644
index 000000000..4e121bde6
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/server-log-viewer.tsx
@@ -0,0 +1,659 @@
+'use client';
+
+import {
+  ArrowDownToLine,
+  Check,
+  ChevronDown,
+  Copy,
+  Download,
+  FileWarning,
+  LoaderCircle,
+  Search,
+  Terminal,
+} from 'lucide-react';
+import { Fragment, memo, useCallback, useEffect, useMemo, useRef, useState } from 'react';
+
+import {
+  Select,
+  SelectContent,
+  SelectItem,
+  SelectTrigger,
+  SelectValue,
+} from '@/components/ui/select';
+import { Button } from '@/components/ui/button';
+import { Input } from '@/components/ui/input';
+import { useServerLogFiles } from '@/hooks/api/use-server-log-files';
+import { useServerLogSearch } from '@/hooks/api/use-server-log-search';
+import { SERVER_LOG_CHUNK_SIZE, useServerLog } from '@/hooks/api/use-server-log';
+import { track } from '@/lib/analytics';
+import { cn } from '@/lib/utils';
+import { useLocale } from '@/lib/use-locale';
+
+import {
+  buildLogLines,
+  isNearLogBottom,
+  readableLogText,
+  utf16IndexAtCodePointOffset,
+  type LogSeverity,
+} from './log-text';
+
+const STRINGS = {
+  en: {
+    title: 'Log files',
+    description: 'Raw .log and .out files captured for this benchmark point.',
+    fileLabel: 'Log file',
+    fileOne: 'file',
+    fileMany: 'files',
+    searchLabel: 'Search all log files',
+    searchPlaceholder: 'Search full log contents…',
+    searchHint: 'Searches every stored file, including content not loaded below.',
+    searching: 'Searching…',
+    searchError: 'The logs could not be searched. Try again in a moment.',
+    noMatches: 'No matches found.',
+    match: 'match',
+    matches: 'matches',
+    firstMatches: 'Showing the first',
+    character: 'character',
+    goToMatch: 'Go to in logs',
+    loading: 'Loading log files…',
+    error: 'The log files could not be loaded. Try again in a moment.',
+    missing: 'No log files are stored for this benchmark point.',
+    copy: 'Copy loaded logs',
+    copied: 'Copied',
+    download: 'Download selected log',
+    loadMore: 'Load next 64 KiB',
+    loadingMore: 'Loading…',
+    loadMoreError: 'The next chunk could not be loaded. The loaded text is still available.',
+    loadedCharacters: 'characters loaded',
+    endOfLog: 'End of stored log',
+  },
+  zh: {
+    title: '日志文件',
+    description: '该基准测试数据点采集的原始 .log 和 .out 文件。',
+    fileLabel: '日志文件',
+    fileOne: '个文件',
+    fileMany: '个文件',
+    searchLabel: '搜索所有日志文件',
+    searchPlaceholder: '搜索完整日志内容……',
+    searchHint: '搜索所有已存储文件，包括下方尚未加载的内容。',
+    searching: '正在搜索……',
+    searchError: '无法搜索日志，请稍后重试。',
+    noMatches: '未找到匹配内容。',
+    match: '处匹配',
+    matches: '处匹配',
+    firstMatches: '显示前',
+    character: '字符',
+    goToMatch: '在日志中定位',
+    loading: '正在加载日志文件……',
+    error: '无法加载日志文件，请稍后重试。',
+    missing: '该基准测试数据点没有已存储的日志文件。',
+    copy: '复制已加载日志',
+    copied: '已复制',
+    download: '下载当前日志',
+    loadMore: '继续加载 64 KiB',
+    loadingMore: '正在加载……',
+    loadMoreError: '无法加载下一段内容，已加载的文本仍可查看。',
+    loadedCharacters: '个字符已加载',
+    endOfLog: '已到达日志末尾',
+  },
+} as const;
+
+interface Props {
+  id: number;
+  enabled: boolean;
+  analyticsContext?: 'agentic' | 'fixed-sequence';
+}
+
+interface LogJumpTarget {
+  fileName: string;
+  offset: number;
+  match: string;
+  requestId: number;
+}
+
+interface LogSelection {
+  pointId: number;
+  fileName: string;
+  initialOffset: number;
+  jumpTarget: LogJumpTarget | null;
+}
+
+const SEARCH_JUMP_CONTEXT_SIZE = 16 * 1024;
+
+/** Console surface. Kept dark in every theme — these are raw terminal artifacts. */
+const LOG_SURFACE = 'bg-[#0b0e13]';
+
+/** Per-severity line tint. Error and warning rows also carry a left accent bar. */
+const LINE_TONE: Record<LogSeverity, string> = {
+  error: 'border-l-rose-400/80 bg-rose-500/8 text-rose-200',
+  warn: 'border-l-amber-300/70 bg-amber-300/6 text-amber-100',
+  debug: 'text-zinc-500',
+  info: 'text-zinc-300',
+};
+
+/** Gutter cell. Empty on purpose — the number comes from a CSS counter, so it
+    stays out of `textContent` and out of anything the reader copies. */
+const GUTTER_CLASS = cn(
+  'log-gutter sticky left-0 z-10 border-r border-white/8 px-3 text-right text-zinc-600 tabular-nums select-none',
+  LOG_SURFACE,
+);
+const LINE_CLASS = 'border-l-2 border-transparent pr-4 pl-2.5';
+
+/**
+ * One plain log line, memoized on primitives. Loading a chunk rebuilds the whole
+ * line array, so without this every already-rendered line would reconcile again;
+ * a megabyte-scale log is ~10k lines and that dominates the load-more click.
+ */
+const LogRow = memo(({ text, severity }: { text: string; severity: LogSeverity }) => (
+  <>
+    <span aria-hidden="true" className={GUTTER_CLASS} />
+    <span className={cn(LINE_CLASS, LINE_TONE[severity])}>{text}</span>
+  </>
+));
+LogRow.displayName = 'LogRow';
+
+/** Split a path so the directory reads as context and the filename as the value. */
+function FilePath({ path, className }: { path: string; className?: string }) {
+  const separator = path.lastIndexOf('/');
+  if (separator === -1) return <span className={className}>{path}</span>;
+  return (
+    <span className={className}>
+      <span className="text-muted-foreground">{path.slice(0, separator + 1)}</span>
+      {path.slice(separator + 1)}
+    </span>
+  );
+}
+
+/** Full-width bordered notice used by the loading, error, and empty states. */
+function LogViewerNotice({
+  icon,
+  tone = 'muted',
+  children,
+}: {
+  icon: React.ReactNode;
+  tone?: 'muted' | 'destructive';
+  children: React.ReactNode;
+}) {
+  return (
+    <div
+      className={cn(
+        'flex items-center gap-3 rounded-lg border px-4 py-6 text-sm',
+        tone === 'destructive'
+          ? 'border-destructive/40 bg-destructive/10 text-destructive'
+          : 'border-border/40 bg-card/40 text-muted-foreground',
+      )}
+    >
+      <span
+        className={cn(
+          'flex size-9 shrink-0 items-center justify-center rounded-md border',
+          tone === 'destructive'
+            ? 'border-destructive/30 bg-destructive/10'
+            : 'border-border/50 bg-background',
+        )}
+      >
+        {icon}
+      </span>
+      <span className="min-w-0">{children}</span>
+    </div>
+  );
+}
+
+export function ServerLogViewer({ id, enabled, analyticsContext = 'agentic' }: Props) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const filesQuery = useServerLogFiles(id, enabled);
+  const files = filesQuery.data ?? [];
+  const [selection, setSelection] = useState<LogSelection | null>(null);
+  const [searchInput, setSearchInput] = useState('');
+  const [searchTerm, setSearchTerm] = useState('');
+  const activeSelection = selection?.pointId === id ? selection : null;
+  const requestedFile = activeSelection?.fileName ?? null;
+  const initialOffset = activeSelection?.initialOffset ?? 0;
+  const jumpTarget = activeSelection?.jumpTarget ?? null;
+  const selectedFile =
+    requestedFile && files.includes(requestedFile) ? requestedFile : (files[0] ?? null);
+
+  const query = useServerLog(id, selectedFile, enabled && selectedFile !== null, initialOffset);
+  const searchQuery = useServerLogSearch(id, searchTerm, enabled);
+  const [copied, setCopied] = useState(false);
+  const rawLog = useMemo(
+    () => query.data?.pages.map((page) => page?.serverLog ?? '').join('') ?? '',
+    [query.data],
+  );
+  const log = useMemo(() => readableLogText(rawLog), [rawLog]);
+  const formatter = useMemo(
+    () => new Intl.NumberFormat(locale === 'zh' ? 'zh-CN' : 'en-US'),
+    [locale],
+  );
+  const logViewportRef = useRef<HTMLPreElement>(null);
+  const jumpHighlightRef = useRef<HTMLElement>(null);
+  const jumpRequestIdRef = useRef(0);
+  const scrolledJumpRequestIdRef = useRef(0);
+  const suppressScrollLoadUntilRef = useRef(0);
+  const nextOffset = query.data?.pages.at(-1)?.nextOffset ?? 0;
+  const loadedOffset = query.data?.pages[0]?.offset ?? initialOffset;
+  const analyticsPrefix =
+    analyticsContext === 'agentic' ? 'inference_agentic' : 'inference_fixed_seq';
+
+  const highlightedLog = useMemo(() => {
+    if (!jumpTarget || jumpTarget.fileName !== selectedFile) return null;
+    const relativeOffset = jumpTarget.offset - loadedOffset;
+    if (relativeOffset < 0) return null;
+    const matchCodePointLength = [...jumpTarget.match].length;
+    const matchStart = utf16IndexAtCodePointOffset(rawLog, relativeOffset);
+    const matchEnd = utf16IndexAtCodePointOffset(rawLog, relativeOffset + matchCodePointLength);
+    if (matchStart === null || matchEnd === null) return null;
+    return {
+      before: readableLogText(rawLog.slice(0, matchStart)),
+      match: readableLogText(rawLog.slice(matchStart, matchEnd)),
+      after: readableLogText(rawLog.slice(matchEnd)),
+    };
+  }, [jumpTarget, loadedOffset, rawLog, selectedFile]);
+
+  // Rendered lines carry the search highlight as a piece so a mid-line match
+  // stays one `<mark>` on the line it starts on.
+  const logLines = useMemo(
+    () =>
+      buildLogLines(
+        highlightedLog
+          ? [
+              { text: highlightedLog.before, highlighted: false },
+              { text: highlightedLog.match, highlighted: true },
+              { text: highlightedLog.after, highlighted: false },
+            ]
+          : [{ text: log, highlighted: false }],
+      ),
+    [highlightedLog, log],
+  );
+  // The jump target is the first highlighted run: it carries the scroll ref and
+  // the test id, while any continuation runs on later lines render unmarked.
+  const jumpPiece = useMemo(() => {
+    for (const [lineIndex, line] of logLines.entries()) {
+      const pieceIndex = line.pieces.findIndex((piece) => piece.highlighted);
+      if (pieceIndex !== -1) return { lineIndex, pieceIndex };
+    }
+    return null;
+  }, [logLines]);
+
+  const downloadUrl = useMemo(() => {
+    if (!selectedFile) return '#';
+    return `/api/v1/server-log?${new URLSearchParams({
+      id: String(id),
+      file: selectedFile,
+      download: '1',
+    })}`;
+  }, [id, selectedFile]);
+
+  useEffect(() => {
+    const timeout = window.setTimeout(() => setSearchTerm(searchInput.trim()), 300);
+    return () => window.clearTimeout(timeout);
+  }, [searchInput]);
+
+  useEffect(() => {
+    if (!searchTerm) return;
+    track(`${analyticsPrefix}_logs_searched`, { id, queryLength: searchTerm.length });
+  }, [analyticsPrefix, id, searchTerm]);
+
+  useEffect(() => {
+    if (
+      !highlightedLog ||
+      !jumpTarget ||
+      !jumpHighlightRef.current ||
+      scrolledJumpRequestIdRef.current === jumpTarget.requestId
+    ) {
+      return;
+    }
+    scrolledJumpRequestIdRef.current = jumpTarget.requestId;
+    suppressScrollLoadUntilRef.current = Date.now() + 500;
+    jumpHighlightRef.current.scrollIntoView({ block: 'center' });
+    jumpHighlightRef.current.focus({ preventScroll: true });
+  }, [highlightedLog, jumpTarget]);
+
+  const copyLoadedLog = async () => {
+    await navigator.clipboard.writeText(log);
+    setCopied(true);
+    track(`${analyticsPrefix}_logs_copied`, {
+      id,
+      fileName: selectedFile,
+      loadedCharacters: log.length,
+    });
+    window.setTimeout(() => setCopied(false), 1500);
+  };
+
+  const selectFile = (fileName: string) => {
+    suppressScrollLoadUntilRef.current = 0;
+    setSelection({ pointId: id, fileName, initialOffset: 0, jumpTarget: null });
+    setCopied(false);
+    track(`${analyticsPrefix}_log_file_selected`, { id, fileName });
+  };
+
+  const goToSearchMatch = (result: Omit<LogJumpTarget, 'requestId'>) => {
+    const startOffset = Math.max(0, result.offset - SEARCH_JUMP_CONTEXT_SIZE);
+    jumpRequestIdRef.current += 1;
+    setSelection({
+      pointId: id,
+      fileName: result.fileName,
+      initialOffset: startOffset,
+      jumpTarget: { ...result, requestId: jumpRequestIdRef.current },
+    });
+    setCopied(false);
+    track(`${analyticsPrefix}_log_search_match_opened`, {
+      id,
+      fileName: result.fileName,
+      offset: result.offset,
+      queryLength: searchTerm.length,
+    });
+  };
+
+  const loadMore = useCallback(
+    (trigger: 'button' | 'scroll') => {
+      track(`${analyticsPrefix}_log_chunk_loaded`, {
+        id,
+        fileName: selectedFile,
+        offset: nextOffset,
+        chunkSize: SERVER_LOG_CHUNK_SIZE,
+        trigger,
+      });
+      void query.fetchNextPage();
+    },
+    [analyticsPrefix, id, nextOffset, query.fetchNextPage, selectedFile],
+  );
+
+  const handleLogScroll = useCallback(
+    (event: React.UIEvent<HTMLPreElement>) => {
+      const viewport = event.currentTarget;
+      if (
+        Date.now() < suppressScrollLoadUntilRef.current ||
+        !isNearLogBottom(viewport) ||
+        !query.hasNextPage ||
+        query.isFetchingNextPage ||
+        query.isFetchNextPageError
+      ) {
+        return;
+      }
+      loadMore('scroll');
+    },
+    [loadMore, query.hasNextPage, query.isFetchNextPageError, query.isFetchingNextPage],
+  );
+
+  const hasLoadedLog = query.data?.pages[0] !== undefined && query.data.pages[0] !== null;
+
+  if (filesQuery.isLoading || (!hasLoadedLog && selectedFile !== null && query.isLoading)) {
+    return (
+      <LogViewerNotice icon={<LoaderCircle className="size-4 animate-spin" aria-hidden="true" />}>
+        {t.loading}
+      </LogViewerNotice>
+    );
+  }
+  if (filesQuery.isError || (query.isError && !hasLoadedLog)) {
+    return (
+      <LogViewerNotice
+        tone="destructive"
+        icon={<FileWarning className="size-4" aria-hidden="true" />}
+      >
+        {t.error}
+      </LogViewerNotice>
+    );
+  }
+  if (files.length === 0 || selectedFile === null || query.data?.pages[0] === null) {
+    return (
+      <LogViewerNotice icon={<Terminal className="size-4" aria-hidden="true" />}>
+        {t.missing}
+      </LogViewerNotice>
+    );
+  }
+
+  return (
+    <section
+      className="overflow-hidden rounded-lg border border-border/60 bg-card/40 shadow-sm"
+      data-testid="agentic-server-log-viewer"
+      data-log-context={analyticsContext}
+    >
+      <header className="flex min-w-0 items-start gap-3 border-b border-border/60 bg-muted/20 px-4 py-3">
+        <span className="mt-0.5 flex size-8 shrink-0 items-center justify-center rounded-md border border-border/60 bg-background text-muted-foreground">
+          <Terminal className="size-4" aria-hidden="true" />
+        </span>
+        <div className="min-w-0 flex-1">
+          <h2 className="text-sm font-semibold text-foreground">{t.title}</h2>
+          <p className="mt-0.5 text-xs text-muted-foreground">{t.description}</p>
+        </div>
+        <span className="shrink-0 rounded-full border border-border/60 bg-background px-2.5 py-0.5 font-mono text-2xs tabular-nums text-muted-foreground">
+          {formatter.format(files.length)} {files.length === 1 ? t.fileOne : t.fileMany}
+        </span>
+      </header>
+
+      <div className="grid min-w-0 gap-2 border-b border-border/60 bg-background/30 px-4 py-3">
+        <div className="flex min-w-0 flex-wrap items-center gap-2">
+          <label className="sr-only" htmlFor="agentic-log-file">
+            {t.fileLabel}
+          </label>
+          <Select value={selectedFile} onValueChange={selectFile}>
+            <SelectTrigger
+              id="agentic-log-file"
+              size="sm"
+              className="w-full min-w-0 shrink overflow-hidden font-mono text-xs sm:w-72 lg:w-96 *:data-[slot=select-value]:min-w-0 *:data-[slot=select-value]:truncate"
+              title={selectedFile}
+            >
+              <SelectValue />
+            </SelectTrigger>
+            <SelectContent className="max-w-[calc(100vw-2rem)]">
+              {files.map((fileName) => (
+                <SelectItem
+                  key={fileName}
+                  value={fileName}
+                  textValue={fileName}
+                  className="max-w-full font-mono text-xs whitespace-normal break-all"
+                  title={fileName}
+                >
+                  <FilePath path={fileName} />
+                </SelectItem>
+              ))}
+            </SelectContent>
+          </Select>
+
+          <label className="sr-only" htmlFor="agentic-log-search">
+            {t.searchLabel}
+          </label>
+          <div className="relative min-w-48 flex-1">
+            <Search
+              className="pointer-events-none absolute top-1/2 left-3 size-3.5 -translate-y-1/2 text-muted-foreground"
+              aria-hidden="true"
+            />
+            <Input
+              id="agentic-log-search"
+              type="search"
+              value={searchInput}
+              onChange={(event) => setSearchInput(event.target.value)}
+              placeholder={t.searchPlaceholder}
+              className="h-8 pr-9 pl-8.5 font-mono text-xs md:text-xs"
+              data-testid="server-log-search"
+            />
+            {searchQuery.isFetching ? (
+              <LoaderCircle
+                className="absolute top-1/2 right-3 size-3.5 -translate-y-1/2 animate-spin text-muted-foreground"
+                aria-label={t.searching}
+              />
+            ) : null}
+          </div>
+
+          {/* Wraps rather than clipping: the pair does not fit beside the search field
+              on a phone, so the group breaks onto its own line(s). */}
+          <div className="flex min-w-0 flex-wrap items-center gap-2">
+            <Button asChild variant="outline" size="sm" className="text-xs">
+              <a
+                href={downloadUrl}
+                download
+                onClick={() =>
+                  track(`${analyticsPrefix}_log_downloaded`, { id, fileName: selectedFile })
+                }
+                data-testid="download-selected-server-log"
+              >
+                <Download className="size-3.5" />
+                {t.download}
+              </a>
+            </Button>
+            <Button
+              type="button"
+              onClick={() => void copyLoadedLog()}
+              disabled={log.length === 0}
+              variant="outline"
+              size="sm"
+              className="text-xs"
+              data-testid="copy-loaded-server-log"
+            >
+              {copied ? (
+                <Check className="size-3.5 text-emerald-500" />
+              ) : (
+                <Copy className="size-3.5" />
+              )}
+              {copied ? t.copied : t.copy}
+            </Button>
+          </div>
+        </div>
+
+        <p className="text-2xs text-muted-foreground">{t.searchHint}</p>
+
+        {searchTerm && searchQuery.isError ? (
+          <p className="text-xs text-destructive" role="alert">
+            {t.searchError}
+          </p>
+        ) : null}
+        {searchTerm && searchQuery.data ? (
+          <div className="grid gap-1.5 pt-1" data-testid="server-log-search-results">
+            <p
+              className="font-mono text-3xs font-semibold tracking-eyebrow text-muted-foreground uppercase"
+              aria-live="polite"
+            >
+              {searchQuery.data.truncated ? `${t.firstMatches} ` : ''}
+              {formatter.format(searchQuery.data.matches.length)}{' '}
+              {searchQuery.data.matches.length === 1 ? t.match : t.matches}
+            </p>
+            {searchQuery.data.matches.length === 0 ? (
+              <p className="rounded-md border border-border/60 bg-muted/20 px-3 py-2 text-xs text-muted-foreground">
+                {t.noMatches}
+              </p>
+            ) : (
+              <div
+                className={cn(
+                  'max-h-72 divide-y divide-white/5 overflow-auto rounded-md border border-border/60',
+                  LOG_SURFACE,
+                )}
+              >
+                {searchQuery.data.matches.map((result) => (
+                  <article key={`${result.fileName}:${result.offset}`} className="grid gap-1.5 p-3">
+                    <div className="flex min-w-0 items-center gap-3 font-mono text-2xs">
+                      <span className="min-w-0 truncate text-zinc-300" title={result.fileName}>
+                        <FilePath path={result.fileName} className="[&>span]:text-zinc-500" />
+                      </span>
+                      <span className="ml-auto shrink-0 text-zinc-500 tabular-nums">
+                        {t.character} {formatter.format(result.offset + 1)}
+                      </span>
+                      <Button
+                        type="button"
+                        variant="ghost"
+                        size="sm"
+                        className="h-6 shrink-0 gap-1 px-2 text-2xs text-amber-200/90 hover:bg-amber-300/10 hover:text-amber-100"
+                        onClick={() => goToSearchMatch(result)}
+                        data-testid="go-to-server-log-match"
+                      >
+                        <ArrowDownToLine className="size-3" />
+                        {t.goToMatch}
+                      </Button>
+                    </div>
+                    <pre className="max-h-20 overflow-hidden border-l-2 border-white/10 pl-2.5 font-mono text-xs leading-5 break-all whitespace-pre-wrap text-zinc-400">
+                      {readableLogText(result.before)}
+                      <mark className="rounded-xs bg-amber-300 px-0.5 text-zinc-950">
+                        {readableLogText(result.match)}
+                      </mark>
+                      {readableLogText(result.after)}
+                    </pre>
+                  </article>
+                ))}
+              </div>
+            )}
+          </div>
+        ) : null}
+      </div>
+
+      <pre
+        ref={logViewportRef}
+        className={cn(
+          'max-h-[70vh] min-h-[28rem] overflow-auto font-mono text-xs leading-5 selection:bg-sky-400/30',
+          LOG_SURFACE,
+        )}
+        data-testid="server-log-content"
+        onScroll={handleLogScroll}
+        tabIndex={0}
+      >
+        <code className="log-lines grid w-max min-w-full grid-cols-[auto_1fr] py-3">
+          {logLines.map((line, lineIndex) =>
+            // Only the few lines a search match spans need per-piece markup; the
+            // rest go through the memoized row so appends stay cheap.
+            line.pieces.some((piece) => piece.highlighted) ? (
+              <Fragment key={lineIndex}>
+                <span aria-hidden="true" className={GUTTER_CLASS} />
+                <span className={cn(LINE_CLASS, LINE_TONE[line.severity])}>
+                  {line.pieces.map((piece, pieceIndex) => {
+                    if (!piece.highlighted)
+                      return <Fragment key={pieceIndex}>{piece.text}</Fragment>;
+                    const isJumpTarget =
+                      jumpPiece?.lineIndex === lineIndex && jumpPiece.pieceIndex === pieceIndex;
+                    return (
+                      <mark
+                        key={pieceIndex}
+                        ref={isJumpTarget ? jumpHighlightRef : undefined}
+                        className="rounded-xs bg-amber-300 px-0.5 text-zinc-950 ring-2 ring-amber-300/30 ring-offset-2 ring-offset-[#0b0e13] outline-none"
+                        data-testid={isJumpTarget ? 'server-log-jump-highlight' : undefined}
+                        tabIndex={-1}
+                      >
+                        {piece.text}
+                      </mark>
+                    );
+                  })}
+                </span>
+              </Fragment>
+            ) : (
+              <LogRow key={lineIndex} text={line.text} severity={line.severity} />
+            ),
+          )}
+        </code>
+      </pre>
+
+      <footer className="flex flex-col gap-2 border-t border-border/60 bg-muted/20 px-4 py-2.5 text-xs text-muted-foreground sm:flex-row sm:items-center sm:justify-between">
+        <span className="font-mono text-2xs tabular-nums">
+          {formatter.format(log.length)} {t.loadedCharacters}
+        </span>
+        <div className="flex flex-col items-start gap-2 sm:flex-row sm:items-center sm:justify-end">
+          {query.isFetchNextPageError ? (
+            <span className="text-destructive" role="alert">
+              {t.loadMoreError}
+            </span>
+          ) : null}
+          {query.hasNextPage ? (
+            <Button
+              type="button"
+              onClick={() => loadMore('button')}
+              disabled={query.isFetchingNextPage}
+              variant="outline"
+              size="sm"
+              className="h-7 text-xs"
+              data-testid="load-more-server-log"
+            >
+              {query.isFetchingNextPage ? (
+                <LoaderCircle className="size-3.5 animate-spin" />
+              ) : (
+                <ChevronDown className="size-3.5" />
+              )}
+              {query.isFetchingNextPage ? t.loadingMore : t.loadMore}
+            </Button>
+          ) : (
+            <span className="inline-flex items-center gap-1.5 font-mono text-2xs">
+              <Check className="size-3.5 text-emerald-500" aria-hidden="true" />
+              {t.endOfLog}
+            </span>
+          )}
+        </div>
+      </footer>
+    </section>
+  );
+}
diff --git a/packages/app/src/components/inference/agentic-point/server-metric-cards.tsx b/packages/app/src/components/inference/agentic-point/server-metric-cards.tsx
index 3cc023de4..545298f41 100644
--- a/packages/app/src/components/inference/agentic-point/server-metric-cards.tsx
+++ b/packages/app/src/components/inference/agentic-point/server-metric-cards.tsx
@@ -1,7 +1,9 @@
 'use client';
 
-import type { RequestTimeline } from '@/hooks/api/use-request-timeline';
-import type { MetricSourceSeries, QueueDepthPoint } from '@/hooks/api/use-trace-server-metrics';
+import { useMemo } from 'react';
+
+import type { RequestChartData } from '@/hooks/api/use-request-chart-data';
+import type { MetricSourceDescriptor, QueueDepthPoint } from '@/hooks/api/use-trace-server-metrics';
 import { SegmentedToggle, type SegmentedToggleOption } from '@/components/ui/segmented-toggle';
 import { track } from '@/lib/analytics';
 
@@ -18,6 +20,7 @@ import {
   buildThroughputChartSeries,
   inflightUniqueTokens,
   rollingAverage,
+  rollingRatioFromComponents,
   timeRollingAverage,
   toggleThroughputSeries,
   type ThroughputSeriesKey,
@@ -62,6 +65,15 @@ const DP_RANK_PALETTE = [
   '#eab308',
 ];
 
+/**
+ * Bare DP ranks read as "DP 3"; disaggregated runs get role- or worker-
+ * qualified labels from the ETL ("decode", "prefill 0", "0 (a01a)") that
+ * already name themselves, so those pass through untouched.
+ */
+function engineSeriesName(engineLabel: string): string {
+  return /^\d+$/u.test(engineLabel) ? `DP ${engineLabel}` : engineLabel;
+}
+
 export function KvCacheUtilizationCard({ sliced }: { sliced: SlicedServerSeries }) {
   return (
     <ExpandableChart
@@ -77,21 +89,29 @@ export function KvCacheUtilizationCard({ sliced }: { sliced: SlicedServerSeries
         // than one, draw one line per rank in distinct colors so
         // load skew is visible at a glance; cluster-average sits on
         // top in white so it stands out.
-        const perEngine = serverSeries.kvCacheUsageByEngine ?? [];
-        const hasPerEngine = perEngine.length > 1;
+        const allEngines = serverSeries.kvCacheUsageByEngine ?? [];
+        // Decide off the point's own engine count, not the phase-sliced one:
+        // this also drives the average line's name, color and stroke, so
+        // keying it to the slice would make the chart change identity when
+        // you switch between the Warmup and Profiling tabs.
+        const hasPerEngine = allEngines.length > 1;
+        // Colors come from the unsliced position so a rank keeps its color
+        // across phases; engines with no points in this phase are dropped
+        // afterwards so they don't take a legend slot with no line.
+        const perEngine = allEngines
+          .map((e, i) => ({
+            name: engineSeriesName(e.engineLabel),
+            data: rollingAverage(e.points, 50),
+            color: DP_RANK_PALETTE[i % DP_RANK_PALETTE.length]!,
+            // Thin + translucent so the Avg line on top reads as
+            // the headline number, not just one more series.
+            strokeWidth: 1,
+            strokeOpacity: 0.5,
+          }))
+          .filter((s) => s.data.length > 0);
         // Render order matters: per-engine first → average drawn on top.
         const series = [
-          ...(hasPerEngine
-            ? perEngine.map((e, i) => ({
-                name: `DP ${e.engineLabel}`,
-                data: rollingAverage(e.points, 50),
-                color: DP_RANK_PALETTE[i % DP_RANK_PALETTE.length]!,
-                // Thin + translucent so the Avg line on top reads as
-                // the headline number, not just one more series.
-                strokeWidth: 1,
-                strokeOpacity: 0.5,
-              }))
-            : []),
+          ...(hasPerEngine ? perEngine : []),
           {
             name: hasHost
               ? 'Chip HBM (avg n=50)'
@@ -143,11 +163,15 @@ export function RequestActivityCard({
   onViewChange,
 }: {
   sliced: SlicedServerSeries;
-  phaseTimeline: RequestTimeline | null;
+  phaseTimeline: RequestChartData | null;
   timelineLoading: boolean;
   view: RequestActivityView;
   onViewChange: (view: RequestActivityView) => void;
 }) {
+  const completedRequests = useMemo(
+    () => (phaseTimeline ? cumulativeCompletedRequests(phaseTimeline.requests) : null),
+    [phaseTimeline],
+  );
   return (
     <ExpandableChart
       title={view === 'queue' ? 'Request queue depth' : 'Cumulative completed requests'}
@@ -176,7 +200,7 @@ export function RequestActivityCard({
               series={[
                 {
                   name: 'Completed requests',
-                  data: cumulativeCompletedRequests(phaseTimeline.requests),
+                  data: completedRequests ?? [],
                   color: '#3b82f6',
                   strokeWidth: 2.5,
                 },
@@ -240,20 +264,32 @@ export function RequestActivityCard({
 }
 
 export function PrefixCacheHitRateCard({ sliced }: { sliced: SlicedServerSeries }) {
+  const hitRateData = useMemo(() => {
+    if (!sliced) return [];
+    const serverSeries = sliced.series;
+    const weighted = rollingRatioFromComponents(
+      serverSeries.prefixCacheHitRate,
+      serverSeries.prefixCacheHitsTps,
+      serverSeries.prefillTps,
+      50,
+    );
+    // Older stored rows may not have the component rate series. Preserve
+    // their existing chart rather than turning it into an empty state.
+    return weighted.length > 0 ? weighted : rollingAverage(serverSeries.prefixCacheHitRate, 50);
+  }, [sliced]);
+
   return (
     <ExpandableChart
       title="Prefix cache hit rate per interval"
       render={(expanded) => {
         const size = expanded ? CHART_SIZES.expanded : CHART_SIZES.inline;
         if (!sliced) return <ChartSkeleton />;
-        const serverSeries = sliced.series;
         return (
           <TimeSeriesChart
             series={[
               {
                 name: 'Chip (HBM, avg n=50)',
-                data: rollingAverage(serverSeries.prefixCacheHitRate, 50),
-                rawData: serverSeries.prefixCacheHitRate,
+                data: hitRateData,
                 color: '#a855f7',
                 strokeWidth: 2,
               },
@@ -277,7 +313,7 @@ export function ThroughputCard({
   onSelectedChange,
 }: {
   sliced: SlicedServerSeries;
-  selectedSource: MetricSourceSeries | undefined;
+  selectedSource: MetricSourceDescriptor | undefined;
   selected: ReadonlySet<ThroughputSeriesKey>;
   onSelectedChange: (next: ReadonlySet<ThroughputSeriesKey>) => void;
 }) {
@@ -418,11 +454,20 @@ export function InflightUniqueTokensCard({
   timelineLoading,
   kvCachePoolTokens,
 }: {
-  phaseTimeline: RequestTimeline | null;
+  phaseTimeline: RequestChartData | null;
   timelineLoading: boolean;
   /** KV-cache pool size in tokens (vLLM only) — drawn as a constant ceiling. */
   kvCachePoolTokens: number | null;
 }) {
+  const inflightSeries = useMemo(() => {
+    if (!phaseTimeline) return null;
+    const raw = inflightUniqueTokens(phaseTimeline.requests);
+    return {
+      raw,
+      smoothed: timeRollingAverage(raw, 30),
+      cumulative: cumulativeTimeAverage(raw),
+    };
+  }, [phaseTimeline]);
   return (
     <ExpandableChart
       title="Unique input tokens in flight"
@@ -439,8 +484,8 @@ export function InflightUniqueTokensCard({
         // independent (cross-conv prefix sharing adds <1pp in
         // practice). Smooth with a 30s time-weighted rolling average
         // so brief turn-handoff dips don't dominate the chart.
-        const raw = inflightUniqueTokens(phaseTimeline.requests);
-        const smoothed = timeRollingAverage(raw, 30);
+        const raw = inflightSeries?.raw ?? [];
+        const smoothed = inflightSeries?.smoothed ?? [];
         // KV-cache pool size (vLLM only) drawn as a constant ceiling so
         // you can see how close the working set gets to eviction
         // pressure. Phase-independent — it's a static config value.
@@ -457,7 +502,7 @@ export function InflightUniqueTokensCard({
               },
               {
                 name: 'Cumulative average',
-                data: cumulativeTimeAverage(raw),
+                data: inflightSeries?.cumulative ?? [],
                 color: '#ef4444',
                 strokeWidth: 3,
               },
diff --git a/packages/app/src/components/inference/agentic-point/sibling-nav.test.ts b/packages/app/src/components/inference/agentic-point/sibling-nav.test.ts
index 2c1b562df..c69dc5b35 100644
--- a/packages/app/src/components/inference/agentic-point/sibling-nav.test.ts
+++ b/packages/app/src/components/inference/agentic-point/sibling-nav.test.ts
@@ -12,11 +12,15 @@ function sibling(overrides: Partial<BenchmarkSibling> = {}): BenchmarkSibling {
     decode_tp: 0,
     decode_ep: 0,
     decode_pp: 1,
+    decode_dcp_size: null,
+    decode_pcp_size: null,
     decode_dp_attention: false,
     decode_num_workers: 0,
     prefill_tp: 8,
     prefill_ep: 1,
     prefill_pp: 2,
+    prefill_dcp_size: null,
+    prefill_pcp_size: null,
     prefill_dp_attention: false,
     prefill_num_workers: 1,
     num_prefill_gpu: 16,
@@ -40,6 +44,19 @@ describe('chipLabel', () => {
     expect(chipLabel(sibling({ decode_tp: 8, decode_ep: 1 }))).toBe('TP8PP2 • c=2');
   });
 
+  it('uses the shared point-label format for aggregate DCP and PCP', () => {
+    expect(
+      chipLabel(
+        sibling({
+          prefill_dcp_size: 8,
+          decode_dcp_size: 8,
+          prefill_pcp_size: 1,
+          decode_pcp_size: 1,
+        }),
+      ),
+    ).toBe('TP8PP2/DCP8 • c=2');
+  });
+
   it('keeps prefill and decode roles separate only for disaggregated serving', () => {
     expect(
       chipLabel(
@@ -48,9 +65,13 @@ describe('chipLabel', () => {
           decode_tp: 8,
           decode_ep: 1,
           decode_pp: 1,
+          prefill_dcp_size: 2,
+          prefill_pcp_size: 4,
+          decode_dcp_size: 8,
+          decode_pcp_size: 1,
           decode_num_workers: 1,
         }),
       ),
-    ).toBe('1xTP8PP2+1xTP8 • c=2');
+    ).toBe('1xTP8PP2/DCP2/PCP4+1xTP8/DCP8 • c=2');
   });
 });
diff --git a/packages/app/src/components/inference/agentic-point/sibling-nav.tsx b/packages/app/src/components/inference/agentic-point/sibling-nav.tsx
index 4dca8fceb..e417790aa 100644
--- a/packages/app/src/components/inference/agentic-point/sibling-nav.tsx
+++ b/packages/app/src/components/inference/agentic-point/sibling-nav.tsx
@@ -5,7 +5,10 @@ import { usePathname, useRouter } from 'next/navigation';
 import { ChevronLeft, ChevronRight } from 'lucide-react';
 
 import type { BenchmarkSibling, BenchmarkSku } from '@/hooks/api/use-benchmark-siblings';
-import { parallelismLabel } from '@/components/inference/utils/parallelism-label';
+import {
+  meaningfulParallelismSize,
+  parallelismLabel,
+} from '@/components/inference/utils/parallelism-label';
 import {
   Select,
   SelectContent,
@@ -71,6 +74,12 @@ export function chipLabel(s: BenchmarkSibling): string {
     !s.disagg && (s.prefill_pp !== null || s.decode_pp !== null)
       ? Math.max(s.prefill_pp ?? 1, s.decode_pp ?? 1)
       : undefined;
+  const aggregateDcp = s.disagg
+    ? undefined
+    : meaningfulParallelismSize(s.prefill_dcp_size, s.decode_dcp_size);
+  const aggregatePcp = s.disagg
+    ? undefined
+    : meaningfulParallelismSize(s.prefill_pcp_size, s.decode_pcp_size);
   // Same parallelism labeler the chart points use (TP/EP/TEP/DEP/DPA…).
   const parallel = parallelismLabel({
     tp: usePrefill ? s.prefill_tp : s.decode_tp,
@@ -80,17 +89,31 @@ export function chipLabel(s: BenchmarkSibling): string {
         ? (s.prefill_pp ?? undefined)
         : (s.decode_pp ?? undefined)
       : aggregatePp,
+    dcp: s.disagg
+      ? usePrefill
+        ? (s.prefill_dcp_size ?? undefined)
+        : (s.decode_dcp_size ?? undefined)
+      : aggregateDcp,
+    pcp: s.disagg
+      ? usePrefill
+        ? (s.prefill_pcp_size ?? undefined)
+        : (s.decode_pcp_size ?? undefined)
+      : aggregatePcp,
     dpAttention: usePrefill ? s.prefill_dp_attention : s.decode_dp_attention,
     disagg: s.disagg,
     isMultinode: s.is_multinode,
     prefillTp: s.prefill_tp,
     prefillEp: s.prefill_ep,
     prefillPp: s.prefill_pp ?? undefined,
+    prefillDcp: s.prefill_dcp_size ?? undefined,
+    prefillPcp: s.prefill_pcp_size ?? undefined,
     prefillDpAttention: s.prefill_dp_attention,
     prefillNumWorkers: s.prefill_num_workers,
     decodeTp: s.decode_tp,
     decodeEp: s.decode_ep,
     decodePp: s.decode_pp ?? undefined,
+    decodeDcp: s.decode_dcp_size ?? undefined,
+    decodePcp: s.decode_pcp_size ?? undefined,
     decodeDpAttention: s.decode_dp_attention,
     decodeNumWorkers: s.decode_num_workers,
   });
@@ -111,12 +134,16 @@ const SORT_OPTIONS: { value: SortMode; label: string }[] = [
 // Group key for the "parallelism" sort: ep first (so TP/EP1 sorts ahead of
 // EP/TEP/DEP groups), then tp, then dp-attention, then disagg — every config
 // of one parallelism lands together, ordered by concurrency within.
-const parallelRank = (s: BenchmarkSibling): [number, number, number, number, number] => {
+const parallelRank = (
+  s: BenchmarkSibling,
+): [number, number, number, number, number, number, number] => {
   const usePrefill = !s.disagg && s.decode_tp <= 0 && s.prefill_tp > 0;
   return [
     usePrefill ? s.prefill_ep : s.decode_ep,
     usePrefill ? s.prefill_tp : s.decode_tp,
     (usePrefill ? s.prefill_pp : s.decode_pp) ?? 1,
+    (usePrefill ? s.prefill_dcp_size : s.decode_dcp_size) ?? 1,
+    (usePrefill ? s.prefill_pcp_size : s.decode_pcp_size) ?? 1,
     (usePrefill ? s.prefill_dp_attention : s.decode_dp_attention) ? 1 : 0,
     s.disagg ? 1 : 0,
   ];
diff --git a/packages/app/src/components/inference/agentic-point/time-series-math.test.ts b/packages/app/src/components/inference/agentic-point/time-series-math.test.ts
index 73eb5cf27..0cf72ed1d 100644
--- a/packages/app/src/components/inference/agentic-point/time-series-math.test.ts
+++ b/packages/app/src/components/inference/agentic-point/time-series-math.test.ts
@@ -14,6 +14,8 @@ import {
   interpAt,
   maxTimeSeriesValue,
   rollingAverage,
+  rollingRatioFromComponents,
+  rollingRatioOfSums,
   rollingRequestMetric,
   timeRollingAverage,
   toggleThroughputSeries,
@@ -135,6 +137,15 @@ describe('rollingRequestMetric', () => {
       { t: 2, value: 0.2 },
     ]);
   });
+
+  it('computes exact cumulative percentiles for a large reverse-ordered latency set', () => {
+    const requests = Array.from({ length: 25_000 }, (_, index) =>
+      request(index + 1, 25_000 - index, 10),
+    );
+    const result = rollingRequestMetric(requests, 'ttft', 'p90');
+    expect(result.cumulative).toHaveLength(requests.length);
+    expect(result.cumulative.at(-1)?.value).toBeCloseTo(22.5001, 4);
+  });
 });
 
 describe('timeRollingAverage', () => {
@@ -177,6 +188,16 @@ describe('timeRollingAverage', () => {
     const data = [{ t: 0, value: 1 }];
     expect(timeRollingAverage(data, 0)).toBe(data);
   });
+
+  it('handles a six-figure event series without rescanning each prefix', () => {
+    const data = Array.from({ length: 100_000 }, (_, index) => ({
+      t: index / 10,
+      value: 7,
+    }));
+    const result = timeRollingAverage(data, 30);
+    expect(result).toHaveLength(data.length);
+    expect(result.at(-1)?.value).toBeCloseTo(7, 9);
+  });
 });
 
 describe('rollingAverage', () => {
@@ -191,6 +212,127 @@ describe('rollingAverage', () => {
   });
 });
 
+describe('rollingRatioOfSums', () => {
+  it('weights cache-hit ratios by token volume instead of averaging interval ratios', () => {
+    const result = rollingRatioOfSums(
+      [
+        { t: 0, value: 0 },
+        { t: 1, value: 100 },
+        { t: 2, value: 0 },
+      ],
+      [
+        { t: 0, value: 100 },
+        { t: 1, value: 0 },
+        { t: 2, value: 100 },
+      ],
+      3,
+    );
+
+    // The middle interval has a zero denominator and would yield an infinite
+    // pointwise ratio. Across the shared window it is 100 / 200 = 50%.
+    expect(result[1]).toEqual({ t: 1, value: 0.5 });
+    expect(result.every((point) => point.value <= 1)).toBe(true);
+  });
+
+  it('treats a missing numerator tick as zero and bounds residual timing skew', () => {
+    const result = rollingRatioOfSums(
+      [
+        { t: 1, value: 50 },
+        { t: 3, value: 110 },
+      ],
+      [
+        { t: 0, value: 100 },
+        { t: 1, value: 100 },
+        { t: 2, value: 100 },
+        { t: 3, value: 100 },
+      ],
+      1,
+    );
+
+    expect(result).toEqual([
+      { t: 0, value: 0 },
+      { t: 1, value: 0.5 },
+      { t: 2, value: 0 },
+      { t: 3, value: 1 },
+    ]);
+  });
+
+  it('handles six-figure series in linear time', () => {
+    const denominator = Array.from({ length: 100_000 }, (_, t) => ({ t, value: 100 }));
+    const numerator = denominator.map(({ t }) => ({ t, value: 80 }));
+    const result = rollingRatioOfSums(numerator, denominator, 50);
+
+    expect(result).toHaveLength(denominator.length);
+    expect(result.at(-1)?.value).toBeCloseTo(0.8, 12);
+  });
+});
+
+describe('rollingRatioFromComponents', () => {
+  it('preserves the stored query denominator when hits exceed prompt throughput', () => {
+    const result = rollingRatioFromComponents(
+      [
+        { t: 0, value: 0.5 },
+        { t: 1, value: 0.5 },
+      ],
+      [
+        { t: 0, value: 200 },
+        { t: 1, value: 200 },
+      ],
+      [
+        { t: 0, value: 100 },
+        { t: 1, value: 100 },
+      ],
+      2,
+    );
+
+    // The original query rate is 200 / 0.5 = 400. Using prompt throughput
+    // as the denominator would incorrectly clamp both intervals to 100%.
+    expect(result.map((point) => point.value)).toEqual([0.5, 0.5]);
+  });
+
+  it('uses prompt throughput only to weight a zero-hit interval', () => {
+    const result = rollingRatioFromComponents(
+      [
+        { t: 0, value: 0 },
+        { t: 1, value: 0.5 },
+      ],
+      [
+        { t: 0, value: 0 },
+        { t: 1, value: 50 },
+      ],
+      [
+        { t: 0, value: 100 },
+        { t: 1, value: 100 },
+      ],
+      2,
+    );
+
+    expect(result).toEqual([
+      { t: 0, value: 0.25 },
+      { t: 1, value: 0.5 },
+    ]);
+  });
+
+  it('retains hit-only ticks omitted from the stored ratio timeline', () => {
+    const result = rollingRatioFromComponents(
+      [{ t: 1, value: 0 }],
+      [
+        { t: 0, value: 100 },
+        { t: 1, value: 0 },
+      ],
+      [{ t: 1, value: 100 }],
+      3,
+    );
+
+    // t=0 has hits but no queries, while the matching query delta lands at
+    // t=1. The centered window must retain both halves: 100 / 100 = 100%.
+    expect(result).toEqual([
+      { t: 0, value: 1 },
+      { t: 1, value: 1 },
+    ]);
+  });
+});
+
 describe('cumulativeAverage', () => {
   it('hides the startup interval without removing it from later averages', () => {
     const result = cumulativeAverage(
diff --git a/packages/app/src/components/inference/agentic-point/time-series-math.ts b/packages/app/src/components/inference/agentic-point/time-series-math.ts
index d0a4587a6..f5b055972 100644
--- a/packages/app/src/components/inference/agentic-point/time-series-math.ts
+++ b/packages/app/src/components/inference/agentic-point/time-series-math.ts
@@ -5,7 +5,7 @@
  * here is unit-testable in isolation (see time-series-math.test.ts).
  */
 
-import type { RequestRecord } from '@/hooks/api/use-request-timeline';
+import type { RequestChartRecord } from '@/hooks/api/use-request-chart-data';
 import type { TimeSeriesPoint } from '@/hooks/api/use-trace-server-metrics';
 
 /** One drawable line in a TimeSeriesChart. */
@@ -96,7 +96,7 @@ export function interpAt(data: TimeSeriesPoint[], t: number): number | null {
  * P90 interactivity = 1 / P90 TPOT (a conservative tail-latency view).
  */
 export function rollingRequestMetric(
-  requests: readonly RequestRecord[],
+  requests: readonly RequestChartRecord[],
   metric: RequestMetric,
   percentile: RequestPercentile,
   windowSize = 50,
@@ -132,17 +132,44 @@ export function rollingRequestMetric(
     const latencyMs = quantile(sorted, q);
     return { t, value: metric === 'interactivity' ? 1000 / latencyMs : latencyMs / 1000 };
   });
-  const prefixLatencies: number[] = [];
-  const cumulative = samples.map(({ t, latencyMs }) => {
-    let lo = 0;
-    let hi = prefixLatencies.length;
-    while (lo < hi) {
-      const mid = (lo + hi) >> 1;
-      if (prefixLatencies[mid]! <= latencyMs) lo = mid + 1;
-      else hi = mid;
+
+  // Exact expanding percentiles without inserting into a growing sorted array
+  // (which is O(n²) from repeated `splice` shifts on six-figure request sets).
+  // Coordinate-compress all observed values, then track prefix counts in a
+  // Fenwick tree so insertion and kth-value lookup are both O(log n).
+  const latencyValues = [...new Set(samples.map(({ latencyMs }) => latencyMs))].toSorted(
+    (a, b) => a - b,
+  );
+  const latencyIndex = new Map(latencyValues.map((value, index) => [value, index + 1]));
+  const counts = new Uint32Array(latencyValues.length + 1);
+  const addLatency = (index: number) => {
+    for (let cursor = index; cursor < counts.length; cursor += cursor & -cursor) {
+      counts[cursor]! += 1;
+    }
+  };
+  const kthLatency = (zeroBasedRank: number): number => {
+    let index = 0;
+    let target = zeroBasedRank + 1;
+    let step = 1;
+    while (step * 2 < counts.length) step *= 2;
+    for (; step > 0; step >>= 1) {
+      const next = index + step;
+      if (next < counts.length && counts[next]! < target) {
+        index = next;
+        target -= counts[next]!;
+      }
     }
-    prefixLatencies.splice(lo, 0, latencyMs);
-    const cumulativeLatencyMs = quantile(prefixLatencies, q);
+    return latencyValues[index]!;
+  };
+  const cumulative = samples.map(({ t, latencyMs }, index) => {
+    addLatency(latencyIndex.get(latencyMs)!);
+    const count = index + 1;
+    const position = (count - 1) * q;
+    const lowRank = Math.floor(position);
+    const highRank = Math.ceil(position);
+    const low = kthLatency(lowRank);
+    const high = kthLatency(highRank);
+    const cumulativeLatencyMs = low + (high - low) * (position - lowRank);
     return {
       t,
       value: metric === 'interactivity' ? 1000 / cumulativeLatencyMs : cumulativeLatencyMs / 1000,
@@ -162,23 +189,27 @@ export function rollingRequestMetric(
  */
 export function timeRollingAverage(data: TimeSeriesPoint[], windowS: number): TimeSeriesPoint[] {
   if (data.length === 0 || windowS <= 0) return data;
+  // Prefix integral of the input step function. `areaAt[i]` is the integral
+  // from t=0 through data[i].t, with the first value extended back to zero.
+  // A moving left cursor then evaluates every trailing window in O(n) total.
+  const areaAt = new Float64Array(data.length);
+  areaAt[0] = data[0]!.value * data[0]!.t;
+  for (let index = 1; index < data.length; index += 1) {
+    const previous = data[index - 1]!;
+    const current = data[index]!;
+    areaAt[index] = areaAt[index - 1]! + previous.value * (current.t - previous.t);
+  }
   const out: TimeSeriesPoint[] = Array.from({ length: data.length });
+  let left = 0;
   for (let i = 0; i < data.length; i++) {
     const tEnd = data[i]!.t;
     const tStart = Math.max(0, tEnd - windowS);
-    // Find the first sample j whose t is >= tStart; the step value at
-    // tStart is data[j-1].value if j > 0, else data[0].value.
-    let j = 0;
-    while (j < data.length && data[j]!.t < tStart) j++;
-    let prevT = tStart;
-    let prevV = j > 0 ? data[j - 1]!.value : data[0]!.value;
-    let area = 0;
-    for (; j <= i; j++) {
-      const curT = data[j]!.t;
-      area += prevV * (curT - prevT);
-      prevT = curT;
-      prevV = data[j]!.value;
-    }
+    while (left + 1 <= i && data[left + 1]!.t <= tStart) left += 1;
+    const areaAtStart =
+      tStart <= data[0]!.t
+        ? data[0]!.value * tStart
+        : areaAt[left]! + data[left]!.value * (tStart - data[left]!.t);
+    const area = areaAt[i]! - areaAtStart;
     const dur = tEnd - tStart;
     out[i] = { t: tEnd, value: dur > 0 ? area / dur : data[i]!.value };
   }
@@ -204,6 +235,158 @@ export function rollingAverage(data: TimeSeriesPoint[], windowSize: number): Tim
   return out;
 }
 
+/**
+ * Centered, volume-weighted ratio over matching rate series.
+ *
+ * Counter families that describe one logical fraction can publish their
+ * deltas in adjacent scrape buckets. Averaging the pointwise ratios gives a
+ * tiny-denominator bucket the same weight as a million-token bucket and can
+ * therefore produce impossible percentages. Sum the rates over the window
+ * first, then divide. The semantic upper bound is applied only after that
+ * aggregation as a guard against residual counter timing skew.
+ *
+ * The denominator owns the timeline. Missing numerator samples contribute
+ * zero, which is the correct interpretation for a no-hit interval.
+ */
+export function rollingRatioOfSums(
+  numerator: TimeSeriesPoint[],
+  denominator: TimeSeriesPoint[],
+  windowSize: number,
+  upperBound = 1,
+): TimeSeriesPoint[] {
+  if (denominator.length === 0 || windowSize <= 0) return [];
+
+  const numeratorByT = new Map<number, number>();
+  for (const point of numerator) {
+    if (!Number.isFinite(point.t) || !Number.isFinite(point.value)) continue;
+    numeratorByT.set(point.t, (numeratorByT.get(point.t) ?? 0) + Math.max(0, point.value));
+  }
+
+  // The ETL emits sorted canonical-grid series. Coalesce defensively so a
+  // duplicate denominator timestamp cannot receive the numerator twice.
+  const rows: { t: number; numerator: number; denominator: number }[] = [];
+  for (const point of denominator) {
+    if (!Number.isFinite(point.t) || !Number.isFinite(point.value)) continue;
+    const previous = rows.at(-1);
+    if (previous?.t === point.t) {
+      previous.denominator += Math.max(0, point.value);
+      continue;
+    }
+    rows.push({
+      t: point.t,
+      numerator: numeratorByT.get(point.t) ?? 0,
+      denominator: Math.max(0, point.value),
+    });
+  }
+  if (rows.length === 0) return [];
+
+  const numeratorPrefix = new Float64Array(rows.length + 1);
+  const denominatorPrefix = new Float64Array(rows.length + 1);
+  for (let i = 0; i < rows.length; i++) {
+    numeratorPrefix[i + 1] = numeratorPrefix[i]! + rows[i]!.numerator;
+    denominatorPrefix[i + 1] = denominatorPrefix[i]! + rows[i]!.denominator;
+  }
+
+  const leftSpan = Math.floor((windowSize - 1) / 2);
+  const rightSpan = windowSize - leftSpan - 1;
+  const boundedUpper = Number.isFinite(upperBound) ? Math.max(0, upperBound) : Infinity;
+  const out: TimeSeriesPoint[] = [];
+  for (let i = 0; i < rows.length; i++) {
+    const start = Math.max(0, i - leftSpan);
+    const end = Math.min(rows.length, i + rightSpan + 1);
+    const denominatorSum = denominatorPrefix[end]! - denominatorPrefix[start]!;
+    if (denominatorSum <= 0) continue;
+    const numeratorSum = numeratorPrefix[end]! - numeratorPrefix[start]!;
+    out.push({
+      t: rows[i]!.t,
+      value: Math.min(boundedUpper, Math.max(0, numeratorSum / denominatorSum)),
+    });
+  }
+  return out;
+}
+
+/**
+ * Volume-weight a stored ratio without substituting a different denominator.
+ *
+ * `ratio = numerator / denominator`, so a positive interval lets us recover
+ * the original denominator exactly as `numerator / ratio`. This matters for
+ * vLLM, where prefix-cache queries are not interchangeable with prompt-token
+ * throughput. A zero-hit interval cannot be inverted; only there do we use
+ * the supplied denominator proxy to retain its zero-valued weight.
+ */
+export function rollingRatioFromComponents(
+  ratio: TimeSeriesPoint[],
+  numerator: TimeSeriesPoint[],
+  zeroRatioDenominator: TimeSeriesPoint[],
+  windowSize: number,
+  upperBound = 1,
+): TimeSeriesPoint[] {
+  const ratioByT = new Map<number, number>();
+  for (const point of ratio) {
+    if (!Number.isFinite(point.t) || !Number.isFinite(point.value) || point.value < 0) continue;
+    ratioByT.set(point.t, point.value);
+  }
+  const numeratorByT = new Map<number, number>();
+  for (const point of numerator) {
+    if (!Number.isFinite(point.t) || !Number.isFinite(point.value)) continue;
+    numeratorByT.set(point.t, (numeratorByT.get(point.t) ?? 0) + Math.max(0, point.value));
+  }
+  const fallbackByT = new Map(
+    zeroRatioDenominator.map((point) => [point.t, Math.max(0, point.value)]),
+  );
+  const recoveredNumerator: TimeSeriesPoint[] = [];
+  const recoveredDenominator: TimeSeriesPoint[] = [];
+
+  // The stored ratio omits `queries=0` ticks, but a lagging hit counter can
+  // still publish a positive numerator there. Walk the sorted union so those
+  // hits participate in the same centered window as their adjacent queries.
+  const timeline: number[] = [];
+  let ratioIndex = 0;
+  let numeratorIndex = 0;
+  while (ratioIndex < ratio.length || numeratorIndex < numerator.length) {
+    while (ratioIndex < ratio.length && !Number.isFinite(ratio[ratioIndex]!.t)) ratioIndex++;
+    while (numeratorIndex < numerator.length && !Number.isFinite(numerator[numeratorIndex]!.t)) {
+      numeratorIndex++;
+    }
+    const ratioT = ratio[ratioIndex]?.t ?? Infinity;
+    const numeratorT = numerator[numeratorIndex]?.t ?? Infinity;
+    const t = Math.min(ratioT, numeratorT);
+    if (!Number.isFinite(t)) break;
+    if (timeline.at(-1) !== t) timeline.push(t);
+    while (ratio[ratioIndex]?.t === t) ratioIndex++;
+    while (numerator[numeratorIndex]?.t === t) numeratorIndex++;
+  }
+
+  for (const t of timeline) {
+    const pointRatio = ratioByT.get(t);
+    const observedNumerator = numeratorByT.get(t) ?? 0;
+    // No stored ratio means the ETL saw no positive query denominator. Keep
+    // a zero denominator entry so a hit-only tick is not silently dropped.
+    if (pointRatio === undefined) {
+      if (observedNumerator > 0) {
+        recoveredNumerator.push({ t, value: observedNumerator });
+        recoveredDenominator.push({ t, value: 0 });
+      }
+      continue;
+    }
+
+    const fallbackDenominator = fallbackByT.get(t) ?? 0;
+    const denominator =
+      pointRatio > 0 && observedNumerator > 0
+        ? observedNumerator / pointRatio
+        : fallbackDenominator;
+    if (!Number.isFinite(denominator) || denominator <= 0) continue;
+
+    recoveredDenominator.push({ t, value: denominator });
+    recoveredNumerator.push({
+      t,
+      value: observedNumerator > 0 ? observedNumerator : pointRatio * denominator,
+    });
+  }
+
+  return rollingRatioOfSums(recoveredNumerator, recoveredDenominator, windowSize, upperBound);
+}
+
 /**
  * Expanding-window cumulative mean from index 0..i.
  *
@@ -262,7 +445,9 @@ export function cumulativeTimeAverage(data: TimeSeriesPoint[]): TimeSeriesPoint[
  * Cumulative count of successfully completed (non-cancelled) requests by end
  * time. Phase is the caller's concern — pass a phase-scoped timeline.
  */
-export function cumulativeCompletedRequests(requests: readonly RequestRecord[]): TimeSeriesPoint[] {
+export function cumulativeCompletedRequests(
+  requests: readonly RequestChartRecord[],
+): TimeSeriesPoint[] {
   const completionTimes = requests
     .filter((request) => !request.cancelled)
     .map((request) => request.end / 1e9)
@@ -277,7 +462,7 @@ export function cumulativeCompletedRequests(requests: readonly RequestRecord[]):
  * OSL uses the request's final observed length across its whole lifetime.
  */
 export function averageSequenceLengthInFlight(
-  requests: readonly RequestRecord[],
+  requests: readonly RequestChartRecord[],
   metric: 'isl' | 'osl',
 ): TimeSeriesPoint[] {
   const events = new Map<number, { tokenDelta: number; countDelta: number }>();
diff --git a/packages/app/src/components/inference/agentic-point/timeline-bars.tsx b/packages/app/src/components/inference/agentic-point/timeline-bars.tsx
index a5444cb2f..eb65a5436 100644
--- a/packages/app/src/components/inference/agentic-point/timeline-bars.tsx
+++ b/packages/app/src/components/inference/agentic-point/timeline-bars.tsx
@@ -10,7 +10,7 @@ import {
   PADDING_RIGHT,
   ROW_GAP,
   ROW_HEIGHT,
-  timelineSvgHeight,
+  ROW_SPAN,
 } from './timeline-layout';
 import { formatTickLabel } from './timeline-format';
 import { conversationHref, type RequestTimelineRow } from './timeline-rows';
@@ -30,6 +30,10 @@ const NICE_TICK_MS = [
 
 export interface TimelineBarsProps {
   rows: RequestTimelineRow[];
+  /** Absolute index of rows[0] within the full virtualized row list. */
+  firstRowIndex: number;
+  scrollTop: number;
+  viewportHeight: number;
   expandedSubagents: ReadonlySet<string>;
   /** Absolute ns timestamp of the visible data's origin (min credit). */
   dataStart: number;
@@ -52,6 +56,9 @@ export interface TimelineBarsProps {
 export const TimelineBars = memo(
   ({
     rows,
+    firstRowIndex,
+    scrollTop,
+    viewportHeight,
     expandedSubagents,
     dataStart,
     vStart,
@@ -61,7 +68,6 @@ export const TimelineBars = memo(
     onBarLeave,
     onBarClick,
   }: TimelineBarsProps) => {
-    const svgHeight = timelineSvgHeight(rows.length);
     const visibleDur = Math.max(vEnd - vStart, 1);
     const scale = (CHART_WIDTH - PADDING_RIGHT) / visibleDur;
     // Local coords: convert ns offset from dataStart to x px.
@@ -79,6 +85,16 @@ export const TimelineBars = memo(
 
     return (
       <>
+        <defs>
+          <clipPath id="timeline-body-clip">
+            <rect
+              x={0}
+              y={HEADER_HEIGHT}
+              width={CHART_WIDTH}
+              height={Math.max(0, viewportHeight - HEADER_HEIGHT)}
+            />
+          </clipPath>
+        </defs>
         {/* Header / time-axis baseline */}
         <line
           x1={0}
@@ -100,7 +116,7 @@ export const TimelineBars = memo(
                 x1={x}
                 y1={HEADER_HEIGHT}
                 x2={x}
-                y2={svgHeight}
+                y2={viewportHeight}
                 stroke="currentColor"
                 opacity={0.08}
                 strokeDasharray="2 4"
@@ -119,133 +135,136 @@ export const TimelineBars = memo(
           );
         })}
 
-        {/* Row separators */}
-        {rows.map((row, idx) => (
-          <line
-            key={`sep-${row.key}`}
-            x1={0}
-            y1={HEADER_HEIGHT + idx * (ROW_HEIGHT + ROW_GAP)}
-            x2={CHART_WIDTH}
-            y2={HEADER_HEIGHT + idx * (ROW_HEIGHT + ROW_GAP)}
-            stroke="currentColor"
-            opacity={0.04}
-          />
-        ))}
+        <g clipPath="url(#timeline-body-clip)">
+          {/* Row separators */}
+          {rows.map((row, idx) => (
+            <line
+              key={`sep-${row.key}`}
+              x1={0}
+              y1={HEADER_HEIGHT + (firstRowIndex + idx) * ROW_SPAN - scrollTop}
+              x2={CHART_WIDTH}
+              y2={HEADER_HEIGHT + (firstRowIndex + idx) * ROW_SPAN - scrollTop}
+              stroke="currentColor"
+              opacity={0.04}
+            />
+          ))}
 
-        {/* Request bars */}
-        {rows.map((row, rowIdx) => {
-          const yTop = HEADER_HEIGHT + rowIdx * (ROW_HEIGHT + ROW_GAP) + 2;
-          const barH = ROW_HEIGHT - 4;
-          // For multi-stream subagent containers, suppress the union
-          // bars when expanded — the child stream rows draw them
-          // individually instead, so we'd double-draw otherwise.
-          if (
-            row.kind === 'subagent' &&
-            (row.streamCount ?? 1) > 1 &&
-            expandedSubagents.has(row.key)
-          ) {
-            return null;
-          }
-          return row.requests.map((req) => {
-            const xCredit = xOf(req.credit);
-            const xStart = xOf(req.start);
-            const xEnd = xOf(req.end);
-            // Cull bars entirely outside the visible window so big
-            // benchmarks don't render thousands of zero-width rects.
-            if (xEnd < -2 || xCredit > CHART_WIDTH + 2) return null;
-            const runW = Math.max(xEnd - xStart, 1);
-            const queueW = Math.max(xStart - xCredit, 0);
-            const phaseColor = PHASE_COLORS[req.phase] ?? PHASE_COLORS.unknown!;
-            const barKey = `${req.cid}-${req.ti}-${req.start}`;
-            const barChildren = (
-              <>
-                {/* Queue lead-in (faint) — only drawn when noticeable. */}
-                {queueW >= 1 && (
+          {/* Request bars */}
+          {rows.map((row, rowIdx) => {
+            const yTop =
+              HEADER_HEIGHT + (firstRowIndex + rowIdx) * (ROW_HEIGHT + ROW_GAP) - scrollTop + 2;
+            const barH = ROW_HEIGHT - 4;
+            // For multi-stream subagent containers, suppress the union
+            // bars when expanded — the child stream rows draw them
+            // individually instead, so we'd double-draw otherwise.
+            if (
+              row.kind === 'subagent' &&
+              (row.streamCount ?? 1) > 1 &&
+              expandedSubagents.has(row.key)
+            ) {
+              return null;
+            }
+            return row.requests.map((req) => {
+              const xCredit = xOf(req.credit);
+              const xStart = xOf(req.start);
+              const xEnd = xOf(req.end);
+              // Cull bars entirely outside the visible window so big
+              // benchmarks don't render thousands of zero-width rects.
+              if (xEnd < -2 || xCredit > CHART_WIDTH + 2) return null;
+              const runW = Math.max(xEnd - xStart, 1);
+              const queueW = Math.max(xStart - xCredit, 0);
+              const phaseColor = PHASE_COLORS[req.phase] ?? PHASE_COLORS.unknown!;
+              const barKey = `${req.cid}-${req.ti}-${req.start}`;
+              const barChildren = (
+                <>
+                  {/* Queue lead-in (faint) — only drawn when noticeable. */}
+                  {queueW >= 1 && (
+                    <rect
+                      x={xCredit}
+                      y={yTop + barH / 2 - 1}
+                      width={queueW}
+                      height={2}
+                      fill={row.color}
+                      opacity={0.35}
+                    />
+                  )}
+                  {/* Main bar — opacity stepped down with depth so
+                parent > subagent > stream reads visually. */}
                   <rect
-                    x={xCredit}
-                    y={yTop + barH / 2 - 1}
-                    width={queueW}
-                    height={2}
+                    x={xStart}
+                    y={yTop}
+                    width={runW}
+                    height={barH}
+                    rx={2}
                     fill={row.color}
-                    opacity={0.35}
+                    opacity={
+                      req.cancelled
+                        ? 0.35
+                        : row.kind === 'stream' || row.kind === 'aux'
+                          ? 0.5
+                          : row.kind === 'subagent'
+                            ? 0.6
+                            : 0.85
+                    }
                   />
-                )}
-                {/* Main bar — opacity stepped down with depth so
-                parent > subagent > stream reads visually. */}
-                <rect
-                  x={xStart}
-                  y={yTop}
-                  width={runW}
-                  height={barH}
-                  rx={2}
-                  fill={row.color}
-                  opacity={
-                    req.cancelled
-                      ? 0.35
-                      : row.kind === 'stream' || row.kind === 'aux'
-                        ? 0.5
-                        : row.kind === 'subagent'
-                          ? 0.6
-                          : 0.85
-                  }
-                />
-                {/* Phase strip at bottom */}
-                <rect
-                  x={xStart}
-                  y={yTop + barH - 2}
-                  width={runW}
-                  height={2}
-                  rx={1}
-                  fill={phaseColor}
-                  opacity={0.85}
-                />
-                {/* Cancelled X overlay */}
-                {req.cancelled && runW > 6 && (
-                  <line
-                    x1={xStart + 1}
-                    y1={yTop + 1}
-                    x2={xStart + runW - 1}
-                    y2={yTop + barH - 1}
-                    stroke="currentColor"
-                    strokeWidth={0.7}
-                    opacity={0.6}
+                  {/* Phase strip at bottom */}
+                  <rect
+                    x={xStart}
+                    y={yTop + barH - 2}
+                    width={runW}
+                    height={2}
+                    rx={1}
+                    fill={phaseColor}
+                    opacity={0.85}
                   />
-                )}
-              </>
-            );
-            // No source dataset → not linkable; plain group.
-            if (!datasetSlug) {
+                  {/* Cancelled X overlay */}
+                  {req.cancelled && runW > 6 && (
+                    <line
+                      x1={xStart + 1}
+                      y1={yTop + 1}
+                      x2={xStart + runW - 1}
+                      y2={yTop + barH - 1}
+                      stroke="currentColor"
+                      strokeWidth={0.7}
+                      opacity={0.6}
+                    />
+                  )}
+                </>
+              );
+              // No source dataset → not linkable; plain group.
+              if (!datasetSlug) {
+                return (
+                  <g
+                    key={barKey}
+                    onMouseMove={(e) => onBarHover(e, row, req)}
+                    onMouseLeave={onBarLeave}
+                  >
+                    {barChildren}
+                  </g>
+                );
+              }
+              // Linkable: render a real SVG anchor with the conversation
+              // href so the browser's native "open in new tab" works
+              // (right-click menu, ⌘/Ctrl-click, middle-click). Plain
+              // left-click stays an in-app navigation; modified or
+              // non-primary clicks fall through to the browser. Suppress
+              // the native link drag so it doesn't fight the pan gesture.
               return (
-                <g
+                <a
                   key={barKey}
+                  href={conversationHref(datasetSlug, req)}
                   onMouseMove={(e) => onBarHover(e, row, req)}
                   onMouseLeave={onBarLeave}
+                  onClick={(e) => onBarClick(e, req)}
+                  onDragStart={(e) => e.preventDefault()}
+                  style={{ cursor: 'pointer' }}
                 >
                   {barChildren}
-                </g>
+                </a>
               );
-            }
-            // Linkable: render a real SVG anchor with the conversation
-            // href so the browser's native "open in new tab" works
-            // (right-click menu, ⌘/Ctrl-click, middle-click). Plain
-            // left-click stays an in-app navigation; modified or
-            // non-primary clicks fall through to the browser. Suppress
-            // the native link drag so it doesn't fight the pan gesture.
-            return (
-              <a
-                key={barKey}
-                href={conversationHref(datasetSlug, req)}
-                onMouseMove={(e) => onBarHover(e, row, req)}
-                onMouseLeave={onBarLeave}
-                onClick={(e) => onBarClick(e, req)}
-                onDragStart={(e) => e.preventDefault()}
-                style={{ cursor: 'pointer' }}
-              >
-                {barChildren}
-              </a>
-            );
-          });
-        })}
+            });
+          })}
+        </g>
       </>
     );
   },
diff --git a/packages/app/src/components/inference/agentic-point/timeline-layout.test.ts b/packages/app/src/components/inference/agentic-point/timeline-layout.test.ts
new file mode 100644
index 000000000..61bdf1ce4
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/timeline-layout.test.ts
@@ -0,0 +1,16 @@
+import { describe, expect, it } from 'vitest';
+
+import { HEADER_HEIGHT, ROW_SPAN, visibleTimelineRowRange } from './timeline-layout';
+
+describe('visibleTimelineRowRange', () => {
+  it('bounds a huge timeline to viewport rows plus overscan', () => {
+    const range = visibleTimelineRowRange(100_000, HEADER_HEIGHT + 50_000 * ROW_SPAN, 480);
+    expect(range.start).toBe(49_996);
+    expect(range.end - range.start).toBeLessThanOrEqual(28);
+  });
+
+  it('clamps the first and final viewport', () => {
+    expect(visibleTimelineRowRange(10, 0, 100)).toEqual({ start: 0, end: 8 });
+    expect(visibleTimelineRowRange(10, 10_000, 480)).toEqual({ start: 10, end: 10 });
+  });
+});
diff --git a/packages/app/src/components/inference/agentic-point/timeline-layout.ts b/packages/app/src/components/inference/agentic-point/timeline-layout.ts
index 7043e4871..c5eed1d35 100644
--- a/packages/app/src/components/inference/agentic-point/timeline-layout.ts
+++ b/packages/app/src/components/inference/agentic-point/timeline-layout.ts
@@ -14,8 +14,24 @@ export const ROW_GAP = 3;
 export const HEADER_HEIGHT = 24;
 export const PADDING_RIGHT = 12;
 export const CHART_WIDTH = 920;
+export const ROW_SPAN = ROW_HEIGHT + ROW_GAP;
 
 /** Chart height for a given row count (header + rows + bottom padding). */
 export function timelineSvgHeight(rowCount: number): number {
-  return HEADER_HEIGHT + rowCount * (ROW_HEIGHT + ROW_GAP) + 6;
+  return HEADER_HEIGHT + rowCount * ROW_SPAN + 6;
+}
+
+/** Rows intersecting the scroll viewport, with a small paint overscan. */
+export function visibleTimelineRowRange(
+  rowCount: number,
+  scrollTop: number,
+  viewportHeight: number,
+  overscan = 4,
+): { start: number; end: number } {
+  const firstVisible = Math.floor(Math.max(0, scrollTop - HEADER_HEIGHT) / ROW_SPAN);
+  const lastVisible = Math.ceil(Math.max(0, scrollTop + viewportHeight - HEADER_HEIGHT) / ROW_SPAN);
+  return {
+    start: Math.min(rowCount, Math.max(0, firstVisible - overscan)),
+    end: Math.min(rowCount, lastVisible + overscan),
+  };
 }
diff --git a/packages/app/src/components/inference/agentic-point/timeline-rows.ts b/packages/app/src/components/inference/agentic-point/timeline-rows.ts
index 12f6e99e7..feeef7ed0 100644
--- a/packages/app/src/components/inference/agentic-point/timeline-rows.ts
+++ b/packages/app/src/components/inference/agentic-point/timeline-rows.ts
@@ -5,13 +5,14 @@
  */
 
 import type { RequestRecord } from '@/hooks/api/use-request-timeline';
+import type { Locale } from '@/lib/i18n';
 
 export type RowMode = 'conversation' | 'worker';
 
 /**
  * The dataset conversation id for a request: the cid with any subagent/forked
  * suffix (`::sa:…`, `::fa:…`) stripped. This is exactly the `conv_id` stored in
- * dataset_conversations, so it deep-links into /datasets/<slug>/conversations/.
+ * dataset_conversations, so it deep-links into /agentx/<slug>/conversations/.
  */
 export function datasetConvId(cid: string): string {
   const i = cid.indexOf('::');
@@ -51,7 +52,7 @@ export function conversationHref(datasetSlug: string, req: RequestRecord): strin
   }
   const sa = subagentIdOf(req.cid);
   if (sa && !params.has('inner')) params.set('sa', sa);
-  return `/datasets/${datasetSlug}/conversations/${encodeURIComponent(convId)}?${params.toString()}`;
+  return `/agentx/${datasetSlug}/conversations/${encodeURIComponent(convId)}?${params.toString()}`;
 }
 
 /** Human label for where a request came from (raw trace index or replay turn). */
@@ -187,6 +188,17 @@ export function splitTimelineCid(cid: string): {
   return { parent, subagentBase: raw, stream: null, aux: null };
 }
 
+/**
+ * Top-level grouping identity for conversation mode. AIPerf can replay the
+ * same source conversation in several trajectory lanes at once, so the source
+ * id alone is not an execution lane. Legacy timelines have no `ri` and retain
+ * the historical conversation-only grouping.
+ */
+export function conversationReplayKey(request: RequestRecord): string {
+  const parent = splitTimelineCid(request.cid).parent;
+  return request.ri === undefined ? parent : `${parent}::replay-lane:${request.ri}`;
+}
+
 /**
  * Stable order/color index for the top-level row groups (conversations in
  * conversation mode, workers in worker mode), keyed by group id and computed
@@ -214,7 +226,7 @@ export function computeStableRowIndex(
   const inProfiling = new Set<string>();
   const inWarmup = new Set<string>();
   for (const r of requests) {
-    const key = mode === 'conversation' ? splitTimelineCid(r.cid).parent : r.wid;
+    const key = mode === 'conversation' ? conversationReplayKey(r) : r.wid;
     const cur = firstStart.get(key);
     if (cur === undefined || r.start < cur) firstStart.set(key, r.start);
     if (r.phase === 'profiling') inProfiling.add(key);
@@ -256,6 +268,7 @@ export function buildRequestTimelineRows(
   mode: RowMode,
   expandedSubagents: ReadonlySet<string>,
   stableRowIndex?: ReadonlyMap<string, number>,
+  locale: Locale = 'en',
 ): RequestTimelineRow[] {
   const index = stableRowIndex ?? computeStableRowIndex(requests, mode);
   const colorFor = (key: string) =>
@@ -298,7 +311,9 @@ export function buildRequestTimelineRows(
     aux: Map<string, RequestRecord[]>;
   }
   interface Tree {
+    groupKey: string;
     parentCid: string;
+    replayIndex?: number;
     parentReqs: RequestRecord[];
     // Aux lanes hanging directly off the main agent (`<cid>::aux:…`).
     parentAux: Map<string, RequestRecord[]>;
@@ -309,16 +324,19 @@ export function buildRequestTimelineRows(
   const trees = new Map<string, Tree>();
   for (const r of requests) {
     const { parent, subagentBase, stream, aux } = splitTimelineCid(r.cid);
-    let tree = trees.get(parent);
+    const groupKey = conversationReplayKey(r);
+    let tree = trees.get(groupKey);
     if (!tree) {
       tree = {
+        groupKey,
         parentCid: parent,
+        replayIndex: r.ri,
         parentReqs: [],
         parentAux: new Map(),
         subagents: new Map(),
         firstStart: Number.POSITIVE_INFINITY,
       };
-      trees.set(parent, tree);
+      trees.set(groupKey, tree);
     }
     if (subagentBase === null && aux !== null) {
       const list = tree.parentAux.get(aux);
@@ -346,17 +364,21 @@ export function buildRequestTimelineRows(
   }
 
   const sortedTrees = [...trees.values()].toSorted(
-    (a, b) => orderOf(a.parentCid) - orderOf(b.parentCid) || a.firstStart - b.firstStart,
+    (a, b) => orderOf(a.groupKey) - orderOf(b.groupKey) || a.firstStart - b.firstStart,
   );
   const rows: RequestTimelineRow[] = [];
   for (const tree of sortedTrees) {
-    const color = colorFor(tree.parentCid);
+    const color = colorFor(tree.groupKey);
     // Parent row (use a placeholder key if the parent itself wasn't replayed).
     tree.parentReqs.sort((a, b) => a.start - b.start);
-    const parentRowKey = tree.parentReqs.length > 0 ? tree.parentCid : `__parent_${tree.parentCid}`;
+    const parentRowKey = tree.parentReqs.length > 0 ? tree.groupKey : `__parent_${tree.groupKey}`;
+    const replayLabel =
+      tree.replayIndex === undefined
+        ? tree.parentCid
+        : `${tree.parentCid} · ${locale === 'zh' ? '重放' : 'replay'} ${tree.replayIndex + 1}`;
     rows.push({
       key: parentRowKey,
-      label: tree.parentCid,
+      label: replayLabel,
       color,
       requests: tree.parentReqs,
       depth: 0,
@@ -372,7 +394,7 @@ export function buildRequestTimelineRows(
     for (const [auxId, reqs] of parentAuxEntries) {
       reqs.sort((a, b) => a.start - b.start);
       rows.push({
-        key: `${tree.parentCid}::aux:${auxId}`,
+        key: `${tree.groupKey}::aux:${auxId}`,
         label: `aux ${auxId} · parallel`,
         color,
         requests: reqs,
@@ -397,7 +419,7 @@ export function buildRequestTimelineRows(
       return aStart - bStart;
     });
     for (const [saBase, lanes] of subagentEntries) {
-      const subagentKey = `${tree.parentCid}::sa:${saBase}`;
+      const subagentKey = `${tree.groupKey}::sa:${saBase}`;
       // Union of primary stream requests for collapsed-view bars. Aux lanes
       // stay separate so their overlap remains visible as parallel work.
       const allReqs: RequestRecord[] = [];
diff --git a/packages/app/src/components/inference/agentic-point/timeline-tooltips.tsx b/packages/app/src/components/inference/agentic-point/timeline-tooltips.tsx
index 7aa63efc9..dd04b05b6 100644
--- a/packages/app/src/components/inference/agentic-point/timeline-tooltips.tsx
+++ b/packages/app/src/components/inference/agentic-point/timeline-tooltips.tsx
@@ -20,7 +20,7 @@ export function TimelineTooltip({ data, linkable }: { data: TooltipData; linkabl
   const queueMs = (req.start - req.credit) / 1e6;
   return (
     <div
-      className="fixed z-50 pointer-events-none rounded-md border border-border bg-card p-2.5 shadow-lg text-[11px]"
+      className="fixed z-50 pointer-events-none rounded-md border border-border bg-card p-2.5 shadow-lg text-2xs"
       style={{ left: data.x + 12, top: data.y - 10, maxWidth: 280 }}
     >
       <div className="flex items-center gap-2 font-medium text-foreground">
@@ -71,11 +71,11 @@ export function TimelineTooltip({ data, linkable }: { data: TooltipData; linkabl
         <span>Worker</span>
         <span className="text-foreground text-right truncate">{shortenWid(req.wid)}</span>
       </div>
-      <div className="mt-1.5 pt-1 border-t border-border/40 text-[10px] text-muted-foreground">
+      <div className="mt-1.5 pt-1 border-t border-border/40 text-3xs text-muted-foreground">
         Started at {formatTickLabel(req.start)}
       </div>
       {linkable && (
-        <div className="mt-1 text-[10px] font-medium text-primary">
+        <div className="mt-1 text-3xs font-medium text-primary">
           Click to view this conversation in the dataset →
         </div>
       )}
@@ -115,7 +115,7 @@ export function CursorPopover({
       : cursor.clientX - 220;
   return (
     <div
-      className="fixed z-40 pointer-events-none rounded-md border border-border bg-card/95 backdrop-blur p-2 shadow-lg text-[11px] font-mono"
+      className="fixed z-40 pointer-events-none rounded-md border border-border bg-card/95 backdrop-blur p-2 shadow-lg text-2xs font-mono"
       style={{ left, top: cursor.clientY - 60, minWidth: 180 }}
     >
       <div className="flex justify-between gap-3 text-foreground">
@@ -127,9 +127,9 @@ export function CursorPopover({
       <div className="mt-1 pt-1 border-t border-border/40 grid grid-cols-2 gap-x-3 gap-y-0.5 text-muted-foreground">
         <span>In flight</span>
         <span className="text-foreground text-right tabular-nums">{inflight}</span>
-        <span className="pl-3 text-[10px]">running</span>
+        <span className="pl-3 text-3xs">running</span>
         <span className="text-foreground text-right tabular-nums">{running}</span>
-        <span className="pl-3 text-[10px]">waiting</span>
+        <span className="pl-3 text-3xs">waiting</span>
         <span className="text-foreground text-right tabular-nums">{waiting}</span>
         <span>Completed</span>
         <span className="text-foreground text-right tabular-nums">{completed}</span>
diff --git a/packages/app/src/components/inference/agentic-point/use-detail-view.test.tsx b/packages/app/src/components/inference/agentic-point/use-detail-view.test.tsx
new file mode 100644
index 000000000..4fe54b399
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/use-detail-view.test.tsx
@@ -0,0 +1,122 @@
+// @vitest-environment jsdom
+import { act, createElement } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { track } from '@/lib/analytics';
+
+import { useDetailView } from './use-detail-view';
+
+vi.mock('@/lib/analytics', () => ({ track: vi.fn() }));
+
+let container: HTMLDivElement;
+let root: Root;
+
+function DetailViewProbe() {
+  const [view, setView] = useDetailView();
+  return createElement(
+    'div',
+    null,
+    createElement('output', { 'data-testid': 'view' }, view),
+    createElement('button', { onClick: () => setView('point') }, 'point'),
+    createElement('button', { onClick: () => setView('timeline') }, 'timeline'),
+    createElement('button', { onClick: () => setView('aggregates') }, 'aggregates'),
+  );
+}
+
+function nativeReplace(href: string, state: unknown = window.history.state): void {
+  History.prototype.replaceState.call(window.history, state, '', href);
+}
+
+function renderProbe(): void {
+  act(() => root.render(createElement(DetailViewProbe)));
+}
+
+function click(label: string): void {
+  const button = [...container.querySelectorAll('button')].find(
+    (candidate) => candidate.textContent === label,
+  );
+  if (!button) throw new Error(`Missing ${label} button`);
+  act(() => button.dispatchEvent(new MouseEvent('click', { bubbles: true })));
+}
+
+function renderedView(): string | null {
+  return container.querySelector('[data-testid="view"]')?.textContent ?? null;
+}
+
+beforeEach(() => {
+  Reflect.deleteProperty(window.history, 'replaceState');
+  nativeReplace('/inference/agentic/42');
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  Reflect.deleteProperty(window.history, 'replaceState');
+  nativeReplace('/');
+  vi.clearAllMocks();
+});
+
+describe('useDetailView', () => {
+  it('replaces client-owned view state without Next navigation and preserves the rest of the URL', () => {
+    nativeReplace(
+      '/inference/agentic/42?unofficialruns=11&view=timeline&sample=abc#server-metrics',
+      { route: 'agentic-point' },
+    );
+    const nextReplace = vi.fn();
+    Object.defineProperty(window.history, 'replaceState', {
+      configurable: true,
+      value: nextReplace,
+    });
+    const historyLength = window.history.length;
+
+    renderProbe();
+    expect(renderedView()).toBe('timeline');
+
+    click('aggregates');
+    expect(renderedView()).toBe('aggregates');
+    expect(window.location.pathname).toBe('/inference/agentic/42');
+    expect(window.location.hash).toBe('#server-metrics');
+    expect(new URLSearchParams(window.location.search).get('view')).toBe('aggregates');
+    expect(new URLSearchParams(window.location.search).get('unofficialruns')).toBe('11');
+    expect(new URLSearchParams(window.location.search).get('sample')).toBe('abc');
+    expect(window.history.state).toEqual({ route: 'agentic-point' });
+    expect(window.history.length).toBe(historyLength);
+    expect(nextReplace).not.toHaveBeenCalled();
+
+    click('point');
+    expect(renderedView()).toBe('point');
+    expect(new URLSearchParams(window.location.search).has('view')).toBe(false);
+    expect(new URLSearchParams(window.location.search).get('unofficialruns')).toBe('11');
+    expect(new URLSearchParams(window.location.search).get('sample')).toBe('abc');
+    expect(window.location.hash).toBe('#server-metrics');
+    expect(nextReplace).not.toHaveBeenCalled();
+    expect(track).toHaveBeenNthCalledWith(1, 'inference_agentic_detail_view_changed', {
+      view: 'aggregates',
+    });
+    expect(track).toHaveBeenNthCalledWith(2, 'inference_agentic_detail_view_changed', {
+      view: 'point',
+    });
+  });
+
+  it('updates the mounted selection when Back or forward changes the URL', () => {
+    nativeReplace('/inference/agentic/42?unofficialruns=11');
+    renderProbe();
+    expect(renderedView()).toBe('point');
+
+    act(() => {
+      nativeReplace('/inference/agentic/42?unofficialruns=11&view=timeline');
+      window.dispatchEvent(new PopStateEvent('popstate'));
+    });
+    expect(renderedView()).toBe('timeline');
+
+    act(() => {
+      nativeReplace('/inference/agentic/42?unofficialruns=11&view=aggregates');
+      window.dispatchEvent(new PopStateEvent('popstate'));
+    });
+    expect(renderedView()).toBe('aggregates');
+  });
+});
diff --git a/packages/app/src/components/inference/agentic-point/use-detail-view.ts b/packages/app/src/components/inference/agentic-point/use-detail-view.ts
new file mode 100644
index 000000000..1a97a85cc
--- /dev/null
+++ b/packages/app/src/components/inference/agentic-point/use-detail-view.ts
@@ -0,0 +1,30 @@
+'use client';
+
+import { useCallback } from 'react';
+
+import { useClientSearchParams } from '@/hooks/useClientSearch';
+import { track } from '@/lib/analytics';
+import { replaceClientSearch } from '@/lib/client-navigation';
+
+export type DetailView = 'point' | 'timeline' | 'aggregates' | 'logs';
+
+const isDetailView = (value: string | null): value is DetailView =>
+  value === 'point' || value === 'timeline' || value === 'aggregates' || value === 'logs';
+
+/** URL-persisted detail view (`?view=`; per-point is the unadorned default). */
+export function useDetailView(): [DetailView, (nextView: DetailView) => void] {
+  const searchParams = useClientSearchParams();
+  const requestedView = searchParams.get('view');
+  const view: DetailView = isDetailView(requestedView) ? requestedView : 'point';
+  const setView = useCallback((nextView: DetailView) => {
+    // Read at write time so URL state owned by other mounted controls cannot be
+    // lost if it changed since this component's last render.
+    const nextParams = new URLSearchParams(window.location.search);
+    if (nextView === 'point') nextParams.delete('view');
+    else nextParams.set('view', nextView);
+    replaceClientSearch(nextParams);
+    track('inference_agentic_detail_view_changed', { view: nextView });
+  }, []);
+
+  return [view, setView];
+}
diff --git a/packages/app/src/components/inference/axis-metric-explanations.test.ts b/packages/app/src/components/inference/axis-metric-explanations.test.ts
new file mode 100644
index 000000000..a2178fe50
--- /dev/null
+++ b/packages/app/src/components/inference/axis-metric-explanations.test.ts
@@ -0,0 +1,150 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  METRIC_EXPLANATIONS,
+  metricRowLabel,
+  resolveXAxisKind,
+  X_AXIS_EXPLANATIONS,
+  X_AXIS_KINDS,
+  xAxisPercentileFromLabel,
+} from './axis-metric-explanations';
+import { METRIC_REGISTRY, type MetricKey } from './metric-registry';
+
+const CJK = /[\u4E00-\u9FFF]/u;
+const metricKeys = Object.keys(METRIC_REGISTRY) as MetricKey[];
+
+describe('METRIC_EXPLANATIONS completeness', () => {
+  it('covers every METRIC_REGISTRY key and nothing else', () => {
+    expect(Object.keys(METRIC_EXPLANATIONS).sort()).toEqual([...metricKeys].sort());
+  });
+
+  it.each(metricKeys)('%s has non-empty bilingual description and formula', (key) => {
+    const entry = METRIC_EXPLANATIONS[key];
+    expect(entry.description.en.length).toBeGreaterThan(20);
+    expect(entry.description.zh.length).toBeGreaterThan(10);
+    expect(entry.formula.en.length).toBeGreaterThan(5);
+    expect(entry.formula.zh.length).toBeGreaterThan(5);
+    // Chinese strings must actually be Chinese, not copied English.
+    expect(entry.description.zh).toMatch(CJK);
+    expect(entry.formula.zh).toMatch(CJK);
+    expect(entry.description.en).not.toMatch(CJK);
+    // Formulas are structural: they must show a computation, not a number soup.
+    expect(entry.formula.en).toMatch(/[÷×=]/u);
+  });
+
+  it('per-MW metrics describe the per-MW normalization', () => {
+    for (const key of ['tpPerMw', 'inputTputPerMw', 'outputTputPerMw', 'powerUser'] as const) {
+      expect(METRIC_EXPLANATIONS[key].formula.en).toContain('all-in utility MW');
+    }
+  });
+
+  it('cost metrics show the $/Mtok formula', () => {
+    for (const key of [
+      'costh',
+      'costn',
+      'costr',
+      'costhOutput',
+      'costnOutput',
+      'costrOutput',
+      'costhi',
+      'costni',
+      'costri',
+      'costUser',
+    ] as const) {
+      expect(METRIC_EXPLANATIONS[key].formula.en).toContain('$/Mtok =');
+    }
+  });
+
+  it('states both available sale-price sources behind token revenue', () => {
+    const explanation = METRIC_EXPLANATIONS.tokenRevenuePerGpuHour;
+    expect(explanation.description.en).toContain('$1 per million');
+    expect(explanation.description.en).toContain('OpenRouter');
+    expect(explanation.description.zh).toContain('每百万 1 美元');
+    expect(explanation.description.zh).toContain('OpenRouter');
+    expect(explanation.formula.en).toContain('$/GPU/hr =');
+  });
+});
+
+describe('X_AXIS_EXPLANATIONS', () => {
+  it('covers every x-axis kind', () => {
+    expect(Object.keys(X_AXIS_EXPLANATIONS).sort()).toEqual([...X_AXIS_KINDS].sort());
+  });
+
+  it.each(X_AXIS_KINDS)('%s has bilingual name and description', (kind) => {
+    const entry = X_AXIS_EXPLANATIONS[kind];
+    expect(entry.name.en(null).length).toBeGreaterThan(3);
+    expect(entry.name.zh(null)).toMatch(CJK);
+    expect(entry.description.en.length).toBeGreaterThan(20);
+    expect(entry.description.zh).toMatch(CJK);
+  });
+
+  it('renders percentile prefixes in row names', () => {
+    expect(X_AXIS_EXPLANATIONS.ttft.name.en('P90')).toBe('P90 Time To First Token (s)');
+    expect(X_AXIS_EXPLANATIONS.ttft.name.en(null)).toBe('Time To First Token (s)');
+    expect(X_AXIS_EXPLANATIONS.ttft.name.zh('Median')).toContain('中位');
+    expect(X_AXIS_EXPLANATIONS.e2eNormalizedInteractivity.name.zh('P90')).toContain('P90 ');
+  });
+});
+
+describe('resolveXAxisKind', () => {
+  const base = {
+    xAxisField: 'intvty',
+    isDerivedNormalizedInteractivity: false,
+  };
+
+  it('interactivity chart plots interactivity by default', () => {
+    expect(resolveXAxisKind('interactivity', base)).toBe('interactivity');
+  });
+
+  it('interactivity chart plots TTFT when the resolved field is a TTFT column', () => {
+    expect(resolveXAxisKind('interactivity', { ...base, xAxisField: 'p90_ttft' })).toBe('ttft');
+    expect(resolveXAxisKind('interactivity', { ...base, xAxisField: 'median_ttft' })).toBe('ttft');
+  });
+
+  it('interactivity chart keeps interactivity for input metrics without a TTFT override', () => {
+    // Input metric with no `*_x` config override: `resolveXAxisField` falls
+    // back to the chart's natural x, so the footer must say interactivity.
+    expect(resolveXAxisKind('interactivity', { ...base, xAxisField: 'p90_intvty' })).toBe(
+      'interactivity',
+    );
+  });
+
+  it('e2e chart plots end-to-end latency by default', () => {
+    expect(resolveXAxisKind('e2e', { ...base, xAxisField: 'e2el' })).toBe('e2eLatency');
+  });
+
+  it('e2e chart plots TTFT under the ttft x-axis mode', () => {
+    expect(resolveXAxisKind('e2e', { ...base, xAxisField: 'p99_ttft' })).toBe('ttft');
+  });
+
+  it('e2e chart plots normalized interactivity under the derived agentic mode', () => {
+    expect(
+      resolveXAxisKind('e2e', {
+        xAxisField: 'p90_ttft',
+        isDerivedNormalizedInteractivity: true,
+      }),
+    ).toBe('e2eNormalizedInteractivity');
+  });
+});
+
+describe('xAxisPercentileFromLabel', () => {
+  it('extracts percentile words', () => {
+    expect(xAxisPercentileFromLabel('P90 Time To First Token (s)')).toBe('P90');
+    expect(xAxisPercentileFromLabel('Median Time To First Token (s)')).toBe('Median');
+    expect(xAxisPercentileFromLabel('P75 E2E Normalized Interactivity (tok/s/user)')).toBe('P75');
+    expect(xAxisPercentileFromLabel('P99.9 End-to-end Latency (s)')).toBe('P99.9');
+  });
+
+  it('returns null when there is no percentile prefix', () => {
+    expect(xAxisPercentileFromLabel('Interactivity (tok/s/user)')).toBeNull();
+    expect(xAxisPercentileFromLabel('End-to-end Latency (s)')).toBeNull();
+  });
+});
+
+describe('metricRowLabel', () => {
+  it('resolves locale-aware titles from the registry', () => {
+    expect(metricRowLabel('tpPerGpu', 'en')).toBe('Token Throughput per Chip');
+    expect(metricRowLabel('tpPerGpu', 'zh')).toBe('每芯片 token 吞吐量');
+    expect(metricRowLabel('tokenRevenuePerGpuHour', 'en')).toBe('Token Revenue per GPU Hour');
+  });
+});
diff --git a/packages/app/src/components/inference/axis-metric-explanations.ts b/packages/app/src/components/inference/axis-metric-explanations.ts
new file mode 100644
index 000000000..042f5d060
--- /dev/null
+++ b/packages/app/src/components/inference/axis-metric-explanations.ts
@@ -0,0 +1,521 @@
+/**
+ * @file axis-metric-explanations.ts
+ * @description Bilingual plain-English explanations (and, for y-axis metrics,
+ * structural formulas) backing the expandable axis-metric footer rendered
+ * below each inference chart. Y-axis entries are keyed by `METRIC_REGISTRY`
+ * keys; x-axis entries are keyed by the resolved x-axis kind. A unit test
+ * asserts completeness against the registry, so adding a metric without an
+ * explanation fails CI.
+ *
+ * Formulas are structural descriptions of the derived-field math in
+ * `buildDerivedChartFields` (src/lib/chart-utils.ts) and the glossary
+ * (src/lib/glossary.ts) — they describe how a value is computed, they do not
+ * restate assumed constants.
+ */
+import { METRIC_REGISTRY, type MetricKey } from './metric-registry';
+
+export interface LocalizedText {
+  en: string;
+  zh: string;
+}
+
+export interface MetricExplanation {
+  /** 1–3 sentence plain-language explanation of the metric. */
+  description: LocalizedText;
+  /** Structural formula, rendered in monospace in the footer. */
+  formula: LocalizedText;
+}
+
+/** Cost-basis flavor shared by the $/¥/cost metric families. */
+type CostBasis = 'h' | 'n' | 'r';
+type TokenType = 'total' | 'output' | 'input';
+
+const COST_BASIS_EN: Record<CostBasis, string> = {
+  h: 'all-in hourly ownership cost of a hyperscaler operator',
+  n: 'all-in hourly ownership cost of a Neocloud Giant operator',
+  r: 'all-in hourly cost of a 3-year rental contract',
+};
+
+const COST_BASIS_ZH: Record<CostBasis, string> = {
+  h: '超大规模云厂商自有硬件的每小时全包成本',
+  n: 'Neocloud Giant 自有硬件的每小时全包成本',
+  r: '3 年期租赁合同的每小时全包成本',
+};
+
+const TOKEN_TYPE_EN: Record<TokenType, string> = {
+  total: 'total tokens (input + output)',
+  output: 'output tokens',
+  input: 'input tokens',
+};
+
+const TOKEN_TYPE_ZH: Record<TokenType, string> = {
+  total: '总 token（输入 + 输出）',
+  output: '输出 token',
+  input: '输入 token',
+};
+
+const TOKEN_RATE_EN: Record<TokenType, string> = {
+  total: 'total tok/s/chip',
+  output: 'output tok/s/chip',
+  input: 'input tok/s/chip',
+};
+
+const TOKEN_RATE_ZH: Record<TokenType, string> = {
+  total: '总 tok/s/chip',
+  output: '输出 tok/s/chip',
+  input: '输入 tok/s/chip',
+};
+
+function throughputPerChip(tokenType: TokenType): MetricExplanation {
+  return {
+    description: {
+      en:
+        `Rate of ${TOKEN_TYPE_EN[tokenType]} the serving replica handles each second, divided ` +
+        'across every chip serving it. Normalizing per chip keeps different parallelism setups ' +
+        'and replica sizes comparable.',
+      zh:
+        `服务副本每秒处理的${TOKEN_TYPE_ZH[tokenType]}速率，平均到服务该副本的每一块芯片上。` +
+        '按芯片归一化后，不同并行配置和副本规模之间可以直接比较。',
+    },
+    formula: {
+      en: `tok/s/chip = ${TOKEN_TYPE_EN[tokenType]} per second ÷ number of chips serving the replica`,
+      zh: `tok/s/chip = 每秒${TOKEN_TYPE_ZH[tokenType]}数 ÷ 服务该副本的芯片数`,
+    },
+  };
+}
+
+function throughputPerMw(tokenType: TokenType): MetricExplanation {
+  return {
+    description: {
+      en:
+        `Rate of ${TOKEN_TYPE_EN[tokenType]} produced per megawatt of all-in utility power ` +
+        'provisioned for the hardware — the full data-center power budget attributed to the ' +
+        'chips, not just their own draw. It answers how much output a fixed power budget buys.',
+      zh:
+        `按硬件全电源配置（all-in utility）功率归一化的${TOKEN_TYPE_ZH[tokenType]}速率——` +
+        '即分摊到芯片上的整个数据中心电力预算，而不仅是芯片自身功耗。' +
+        '它回答的是固定电力预算能换来多少产出。',
+    },
+    formula: {
+      en: `tok/s/MW = ${TOKEN_RATE_EN[tokenType]} ÷ all-in utility MW provisioned per chip`,
+      zh: `tok/s/MW = ${TOKEN_RATE_ZH[tokenType]} ÷ 每芯片全电源配置功率（MW）`,
+    },
+  };
+}
+
+function tokenRevenuePerGpuHour(): MetricExplanation {
+  return {
+    description: {
+      en:
+        'Gross token revenue a GPU could earn per hour at this operating point. The normalized ' +
+        'source prices every input and output token at $1 per million; the OpenRouter source uses ' +
+        "the selected model's current public input and output prices. This turns the " +
+        'throughput/interactivity tradeoff into a business-facing SLA curve.',
+      zh:
+        '表示该运行点下每块 GPU 每小时可获得的 token 毛收入。标准化模式将输入和输出 token 均按每百万 1 美元计价；' +
+        'OpenRouter 模式采用所选模型当前公开的输入和输出价格。该指标把吞吐量与交互性的权衡转换为面向业务的 SLA 曲线。',
+    },
+    formula: {
+      en: '$/GPU/hr = total tok/s/GPU × (input share × input $/M + output share × output $/M) × 3,600 ÷ 1,000,000',
+      zh: '$/GPU/hr = 总 tok/s/GPU ×（输入占比 × 输入 $/百万 + 输出占比 × 输出 $/百万）× 3,600 ÷ 1,000,000',
+    },
+  };
+}
+
+function costPerMillion(basis: CostBasis, tokenType: TokenType): MetricExplanation {
+  return {
+    description: {
+      en:
+        `Estimated infrastructure cost of producing one million ${TOKEN_TYPE_EN[tokenType]} at ` +
+        `this operating point, priced with the ${COST_BASIS_EN[basis]}. Lower is cheaper.`,
+      zh:
+        `在该运行点生产一百万${TOKEN_TYPE_ZH[tokenType]}的基础设施成本估算，` +
+        `按${COST_BASIS_ZH[basis]}计价。数值越低越便宜。`,
+    },
+    formula: {
+      en: `$/Mtok = all-in cost per chip-hour ($) × 1,000,000 ÷ (3,600 × ${TOKEN_RATE_EN[tokenType]})`,
+      zh: `$/Mtok = 每芯片小时全包成本（$）× 1,000,000 ÷（3,600 × ${TOKEN_RATE_ZH[tokenType]}）`,
+    },
+  };
+}
+
+function tokensPerDollar(basis: CostBasis, tokenType: TokenType): MetricExplanation {
+  return {
+    description: {
+      en:
+        `How many ${TOKEN_TYPE_EN[tokenType]} one US dollar of infrastructure spend buys, ` +
+        `priced with the ${COST_BASIS_EN[basis]}. It is the reciprocal of cost per token, so ` +
+        'higher means cheaper.',
+      zh:
+        `1 美元基础设施开支能换来多少${TOKEN_TYPE_ZH[tokenType]}，` +
+        `按${COST_BASIS_ZH[basis]}计价。它是单 token 成本的倒数，数值越高越便宜。`,
+    },
+    formula: {
+      en: `tok/$ = (${TOKEN_RATE_EN[tokenType]} × 3,600) ÷ all-in cost per chip-hour ($)`,
+      zh: `tok/$ =（${TOKEN_RATE_ZH[tokenType]} × 3,600）÷ 每芯片小时全包成本（$）`,
+    },
+  };
+}
+
+function tokensPerRmb(basis: CostBasis, tokenType: TokenType): MetricExplanation {
+  return {
+    description: {
+      en:
+        `How many ${TOKEN_TYPE_EN[tokenType]} one Chinese yuan of infrastructure spend buys, ` +
+        `priced with the ${COST_BASIS_EN[basis]} converted at a fixed USD→CNY exchange rate. ` +
+        'Higher means cheaper.',
+      zh:
+        `1 元人民币基础设施开支能换来多少${TOKEN_TYPE_ZH[tokenType]}，` +
+        `按${COST_BASIS_ZH[basis]}以固定 USD→CNY 汇率折算计价。数值越高越便宜。`,
+    },
+    formula: {
+      en:
+        `tok/¥ = (${TOKEN_RATE_EN[tokenType]} × 3,600) ÷ ` +
+        '(all-in cost per chip-hour ($) × USD→CNY exchange rate)',
+      zh:
+        `tok/¥ =（${TOKEN_RATE_ZH[tokenType]} × 3,600）÷` +
+        '（每芯片小时全包成本（$）× USD→CNY 汇率）',
+    },
+  };
+}
+
+function provisionedJoules(tokenType: TokenType): MetricExplanation {
+  return {
+    description: {
+      en:
+        `Electrical energy attributed to each of the ${TOKEN_TYPE_EN[tokenType]}, assuming the ` +
+        'full all-in utility power provisioned for the hardware is drawn. It is a provisioned ' +
+        'upper bound, not a telemetry measurement.',
+      zh:
+        `平均到每个${TOKEN_TYPE_ZH[tokenType]}上的电能，假设硬件按全电源配置功率满额用电。` +
+        '这是按配置功率计算的上限值，不是遥测实测值。',
+    },
+    formula: {
+      en: `J/tok = all-in utility power per chip (W) ÷ ${TOKEN_RATE_EN[tokenType]}`,
+      zh: `J/tok = 每芯片全电源配置功率（W）÷ ${TOKEN_RATE_ZH[tokenType]}`,
+    },
+  };
+}
+
+type MeasuredPhase = 'run' | 'prefill' | 'decode';
+
+const MEASURED_PHASE_EN: Record<MeasuredPhase, string> = {
+  run: 'the whole benchmark run',
+  prefill: 'the prefill phase (prompt processing)',
+  decode: 'the decode phase (token generation)',
+};
+
+const MEASURED_PHASE_ZH: Record<MeasuredPhase, string> = {
+  run: '整个基准测试运行期间',
+  prefill: 'prefill 阶段（处理提示词）',
+  decode: 'decode 阶段（生成 token）',
+};
+
+function measuredPower(phase: MeasuredPhase): MetricExplanation {
+  return {
+    description: {
+      en:
+        `Average per-chip accelerator power actually drawn during ${MEASURED_PHASE_EN[phase]}, ` +
+        'read from runner telemetry. Unlike the all-in provisioned metrics, this reflects real ' +
+        'measured draw, not the provisioned budget.',
+      zh:
+        `${MEASURED_PHASE_ZH[phase]}每块加速器芯片的实际平均功耗，来自运行器遥测数据。` +
+        '与全电源配置类指标不同，它反映的是真实实测功耗，而不是按配置计算的预算值。',
+    },
+    formula: {
+      en: `W = mean of sampled per-chip accelerator power draw over ${MEASURED_PHASE_EN[phase]}`,
+      zh: `W = ${MEASURED_PHASE_ZH[phase]}每芯片加速器功耗采样值的平均`,
+    },
+  };
+}
+
+function measuredJoulesPerToken(tokenType: TokenType): MetricExplanation {
+  return {
+    description: {
+      en:
+        `Measured accelerator energy consumed per ${
+          tokenType === 'total' ? 'token (including prompt tokens)' : `${tokenType} token`
+        }, from runner power telemetry integrated over the run. Lower means the system converts ` +
+        'electricity into tokens more efficiently.',
+      zh:
+        `每个${TOKEN_TYPE_ZH[tokenType]}消耗的加速器实测能耗，由运行器功耗遥测在整个运行期间积分得到。` +
+        '数值越低，说明系统把电能转化为 token 的效率越高。',
+    },
+    formula: {
+      en: `J/tok = measured accelerator energy over the run ÷ ${TOKEN_TYPE_EN[tokenType]} processed`,
+      zh: `J/tok = 运行期间加速器实测能耗 ÷ 处理的${TOKEN_TYPE_ZH[tokenType]}数`,
+    },
+  };
+}
+
+/**
+ * Every `METRIC_REGISTRY` key gets a bilingual explanation and a structural
+ * formula. Grounded in `buildDerivedChartFields` (src/lib/chart-utils.ts) and
+ * the glossary entries for throughput, cost per million tokens, tokens per
+ * dollar, tokens per megawatt, and energy per token.
+ */
+export const METRIC_EXPLANATIONS: Record<MetricKey, MetricExplanation> = {
+  tpPerGpu: throughputPerChip('total'),
+  inputTputPerGpu: throughputPerChip('input'),
+  outputTputPerGpu: throughputPerChip('output'),
+  tokenRevenuePerGpuHour: tokenRevenuePerGpuHour(),
+  tpPerMw: throughputPerMw('total'),
+  inputTputPerMw: throughputPerMw('input'),
+  outputTputPerMw: throughputPerMw('output'),
+  costh: costPerMillion('h', 'total'),
+  costn: costPerMillion('n', 'total'),
+  costr: costPerMillion('r', 'total'),
+  costhOutput: costPerMillion('h', 'output'),
+  costnOutput: costPerMillion('n', 'output'),
+  costrOutput: costPerMillion('r', 'output'),
+  costhi: costPerMillion('h', 'input'),
+  costni: costPerMillion('n', 'input'),
+  costri: costPerMillion('r', 'input'),
+  tokensPerDollarH: tokensPerDollar('h', 'total'),
+  tokensPerDollarN: tokensPerDollar('n', 'total'),
+  tokensPerDollarR: tokensPerDollar('r', 'total'),
+  outputTokensPerDollarH: tokensPerDollar('h', 'output'),
+  outputTokensPerDollarN: tokensPerDollar('n', 'output'),
+  outputTokensPerDollarR: tokensPerDollar('r', 'output'),
+  inputTokensPerDollarH: tokensPerDollar('h', 'input'),
+  inputTokensPerDollarN: tokensPerDollar('n', 'input'),
+  inputTokensPerDollarR: tokensPerDollar('r', 'input'),
+  tokensPerRmbH: tokensPerRmb('h', 'total'),
+  tokensPerRmbN: tokensPerRmb('n', 'total'),
+  tokensPerRmbR: tokensPerRmb('r', 'total'),
+  outputTokensPerRmbH: tokensPerRmb('h', 'output'),
+  outputTokensPerRmbN: tokensPerRmb('n', 'output'),
+  outputTokensPerRmbR: tokensPerRmb('r', 'output'),
+  inputTokensPerRmbH: tokensPerRmb('h', 'input'),
+  inputTokensPerRmbN: tokensPerRmb('n', 'input'),
+  inputTokensPerRmbR: tokensPerRmb('r', 'input'),
+  costUser: {
+    description: {
+      en:
+        'Estimated infrastructure cost of producing one million total tokens, priced with the ' +
+        'hourly cost you entered in the custom-values panel instead of a modeled cost basis.',
+      zh:
+        '生产一百万总 token 的基础设施成本估算，按你在自定义值面板中输入的每小时成本计价，' +
+        '而不是按内置成本模型计价。',
+    },
+    formula: {
+      en: '$/Mtok = user-entered cost per chip-hour ($) × 1,000,000 ÷ (3,600 × total tok/s/chip)',
+      zh: '$/Mtok = 自定义每芯片小时成本（$）× 1,000,000 ÷（3,600 × 总 tok/s/chip）',
+    },
+  },
+  tokensPerDollarUser: {
+    description: {
+      en:
+        'How many total tokens one US dollar buys, priced with the hourly cost you entered in ' +
+        'the custom-values panel. Higher means cheaper under your own cost assumptions.',
+      zh:
+        '1 美元能换来多少总 token，按你在自定义值面板中输入的每小时成本计价。' +
+        '在你自己的成本假设下，数值越高越便宜。',
+    },
+    formula: {
+      en: 'tok/$ = (total tok/s/chip × 3,600) ÷ user-entered cost per chip-hour ($)',
+      zh: 'tok/$ =（总 tok/s/chip × 3,600）÷ 自定义每芯片小时成本（$）',
+    },
+  },
+  powerUser: {
+    description: {
+      en:
+        'Total token throughput normalized by the all-in utility power you entered in the ' +
+        'custom-values panel, instead of the modeled per-chip power budget.',
+      zh: '按你在自定义值面板中输入的全电源配置功率归一化的总 token 吞吐量，而不是按内置的每芯片电力预算。',
+    },
+    formula: {
+      en: 'tok/s/MW = total tok/s/chip ÷ user-entered all-in utility MW per chip',
+      zh: 'tok/s/MW = 总 tok/s/chip ÷ 自定义每芯片全电源配置功率（MW）',
+    },
+  },
+  jTotal: provisionedJoules('total'),
+  jOutput: provisionedJoules('output'),
+  jInput: provisionedJoules('input'),
+  measuredAvgPower: measuredPower('run'),
+  measuredPrefillAvgPower: measuredPower('prefill'),
+  measuredDecodeAvgPower: measuredPower('decode'),
+  measuredJPerOutputToken: measuredJoulesPerToken('output'),
+  measuredJPerInputToken: measuredJoulesPerToken('input'),
+  measuredJPerTotalToken: measuredJoulesPerToken('total'),
+  measuredJPerSuccessfulQuery: {
+    description: {
+      en:
+        'Measured accelerator energy consumed per successfully completed request, from runner ' +
+        'power telemetry. It charges the energy of the whole run only to requests that finished ' +
+        'successfully.',
+      zh:
+        '每个成功完成的请求消耗的加速器实测能耗，来自运行器功耗遥测。' +
+        '整个运行的能耗只计入成功完成的请求。',
+    },
+    formula: {
+      en: 'J/query = measured accelerator energy over the run ÷ successfully completed requests',
+      zh: 'J/query = 运行期间加速器实测能耗 ÷ 成功完成的请求数',
+    },
+  },
+  measuredWhPerSuccessfulQuery: {
+    description: {
+      en:
+        'The same measured energy per successful request expressed in watt-hours, a more ' +
+        'familiar household unit (1 Wh = 3,600 J).',
+      zh: '与每次成功请求实测能耗相同的量，换算成更直观的瓦时单位（1 Wh = 3,600 J）。',
+    },
+    formula: {
+      en: 'Wh/query = measured J per successful query ÷ 3,600',
+      zh: 'Wh/query = 每次成功请求实测能耗（J）÷ 3,600',
+    },
+  },
+  measuredPowerPercentTdp: {
+    description: {
+      en:
+        'Measured average per-chip power as a share of the accelerator’s rated TDP. Values well ' +
+        'below 100% suggest the workload leaves thermal or power headroom on the table.',
+      zh:
+        '每芯片实测平均功耗占加速器额定 TDP 的百分比。' +
+        '明显低于 100% 说明该工作负载没有用满散热或供电余量。',
+    },
+    formula: {
+      en: '% TDP = measured average power per chip (W) ÷ rated TDP (W) × 100',
+      zh: '% TDP = 每芯片实测平均功耗（W）÷ 额定 TDP（W）× 100',
+    },
+  },
+};
+
+/**
+ * The four logical x-axis metrics an inference chart can plot, independent of
+ * the percentile prefix. Mirrors the branch logic in `resolveXAxisField` plus
+ * the derived agentic x-axis mode handled in `ChartDisplay`.
+ */
+export type XAxisKind = 'interactivity' | 'e2eLatency' | 'ttft' | 'e2eNormalizedInteractivity';
+
+export const X_AXIS_KINDS: readonly XAxisKind[] = [
+  'interactivity',
+  'e2eLatency',
+  'ttft',
+  'e2eNormalizedInteractivity',
+];
+
+export interface XAxisExplanation {
+  /** Display name for the footer row; `pctl` is e.g. 'P90' | 'Median' | null. */
+  name: {
+    en: (pctl: string | null) => string;
+    zh: (pctl: string | null) => string;
+  };
+  description: LocalizedText;
+}
+
+const zhPctl = (pctl: string | null): string =>
+  pctl === null ? '' : pctl === 'Median' ? '中位' : `${pctl} `;
+
+const enPctl = (pctl: string | null): string => (pctl === null ? '' : `${pctl} `);
+
+export const X_AXIS_EXPLANATIONS: Record<XAxisKind, XAxisExplanation> = {
+  interactivity: {
+    name: {
+      en: (pctl) => `${enPctl(pctl)}Interactivity (tok/s/user)`,
+      zh: (pctl) => `${zhPctl(pctl)}交互性（tok/s/user）`,
+    },
+    description: {
+      en:
+        'Interactivity is the rate at which a single user receives generated tokens while the ' +
+        'model streams its answer — how quickly new words appear on screen. Higher values feel ' +
+        'snappier; operators trade it against batch throughput.',
+      zh:
+        '交互性（interactivity）指模型流式输出回答时，单个用户接收生成 token 的速率——' +
+        '即新内容出现在屏幕上的快慢。数值越高体验越流畅；运营方需要在交互性与批量吞吐量之间权衡。',
+    },
+  },
+  e2eLatency: {
+    name: {
+      en: (pctl) => `${enPctl(pctl)}End-to-end Latency (s)`,
+      zh: (pctl) => `${zhPctl(pctl)}端到端延迟（s）`,
+    },
+    description: {
+      en:
+        'End-to-end latency is the total wall-clock time from submitting a request until the ' +
+        'complete response finishes — the wait before the first token plus the entire ' +
+        'generation. Lower is better.',
+      zh:
+        '端到端延迟指从提交请求到完整响应结束的总耗时——包括收到首个 token 之前的等待和之后的全部生成过程。' +
+        '数值越低越好。',
+    },
+  },
+  ttft: {
+    name: {
+      en: (pctl) => `${enPctl(pctl)}Time To First Token (s)`,
+      zh: (pctl) => `${zhPctl(pctl)}首 token 延迟（TTFT，s）`,
+    },
+    description: {
+      en:
+        'Time to first token (TTFT) is the delay from submitting a request until the first ' +
+        'generated token arrives — the “thinking…” pause before the answer starts. Percentile ' +
+        'prefixes (Median, P90, P99) report that delay across all requests in the run.',
+      zh:
+        '首 token 延迟（TTFT，Time To First Token）指从提交请求到收到第一个生成 token 的等待时间——' +
+        '也就是回答开始前的“思考”停顿。百分位前缀（Median、P90、P99）表示该延迟在整个运行的所有请求上的分布。',
+    },
+  },
+  e2eNormalizedInteractivity: {
+    name: {
+      en: (pctl) => `${enPctl(pctl)}E2E Normalized Interactivity (tok/s/user)`,
+      zh: (pctl) => `${zhPctl(pctl)}端到端归一化交互性（tok/s/user）`,
+    },
+    description: {
+      en:
+        'E2E normalized interactivity is the effective per-user token rate over a complete ' +
+        'request: output tokens divided by end-to-end latency. Unlike plain interactivity it ' +
+        'also counts the wait before the first token, so slow prefill lowers the score.',
+      zh:
+        '端到端归一化交互性是完整请求内的等效单用户 token 速率：输出 token 数除以端到端延迟。' +
+        '与普通交互性不同，它把首个 token 之前的等待也计算在内，prefill 越慢得分越低。',
+    },
+  },
+};
+
+/**
+ * Resolve which logical x-axis metric a chart is currently plotting.
+ * Classifies off the x-axis data field that `resolveXAxisField` actually
+ * resolved for the chart's current state, so the footer can never disagree
+ * with the drawn axis — `resolveXAxisField` is the single source of truth for
+ * the field, including the input-metric fallback where a metric without a
+ * `*_x` config override keeps plotting the chart's natural x. Trace-derived
+ * agentic x-axis modes bypass that resolver entirely, so they arrive as an
+ * explicit flag. Applies to both the official pipeline and `?unofficialrun=`
+ * overlays — the overlay path shares the chart's resolved x-axis, so one
+ * footer row describes both.
+ */
+export function resolveXAxisKind(
+  chartType: 'interactivity' | 'e2e',
+  opts: {
+    /** Resolved x-axis data field from `resolveXAxisField` (e.g. `p90_ttft`). */
+    xAxisField: string;
+    /** Agentic trace-derived normalized-interactivity x-axis mode. */
+    isDerivedNormalizedInteractivity: boolean;
+  },
+): XAxisKind {
+  if (opts.isDerivedNormalizedInteractivity) return 'e2eNormalizedInteractivity';
+  if (opts.xAxisField.endsWith('ttft')) return 'ttft';
+  return chartType === 'e2e' ? 'e2eLatency' : 'interactivity';
+}
+
+/**
+ * Extract the percentile word from a resolved x-axis label (e.g.
+ * "P90 Time To First Token (s)" → "P90"). The chart pipelines always render
+ * the percentile prefix in this English form, including on /zh pages.
+ */
+export function xAxisPercentileFromLabel(xAxisLabel: string): string | null {
+  const match = /^(?<pctl>Median|Mean|P\d+(?:\.\d+)?)\s/iu.exec(xAxisLabel);
+  if (!match?.groups?.pctl) return null;
+  const pctl = match.groups.pctl;
+  return pctl.toLowerCase() === 'median'
+    ? 'Median'
+    : pctl.toLowerCase() === 'mean'
+      ? 'Mean'
+      : pctl.toUpperCase();
+}
+
+/** Locale-aware y-axis row label straight from the metric registry. */
+export function metricRowLabel(metricKey: MetricKey, locale: 'en' | 'zh'): string {
+  const metric = METRIC_REGISTRY[metricKey];
+  return locale === 'zh' ? metric.titleZh : metric.title;
+}
diff --git a/packages/app/src/components/inference/hooks/useChartData.test.ts b/packages/app/src/components/inference/hooks/useChartData.test.ts
index 94c9075d3..2313456d0 100644
--- a/packages/app/src/components/inference/hooks/useChartData.test.ts
+++ b/packages/app/src/components/inference/hooks/useChartData.test.ts
@@ -1,6 +1,6 @@
 import { describe, it, expect } from 'vitest';
 
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
+import chartDefinitions from '@/components/inference/metric-registry';
 import { dedupeAgenticHistoryRuns } from '@/lib/benchmark-run-selection';
 
 import type { InferenceData } from '@/components/inference/types';
@@ -15,6 +15,7 @@ import {
   filterByGPU,
   derivedModeRoofline,
   flipRooflineDirection,
+  supportsPointTokenMetric,
 } from './useChartData';
 
 interface DedupeInput {
@@ -29,6 +30,9 @@ interface DedupeInput {
   date: string;
   workflow_run_id?: number;
   run_started_at?: string | null;
+  curve_date?: string;
+  curve_workflow_run_id?: number;
+  curve_run_started_at?: string | null;
 }
 
 const drow = (over: Partial<DedupeInput> = {}): DedupeInput => ({
@@ -93,9 +97,24 @@ describe('dedupeRowsToLatestPerConfig', () => {
 
   it('dedupes mixed agentic spec methods as one curve', () => {
     const rows = [
-      drow({ id: 1, benchmark_type: 'agentic_traces', spec_method: 'none', date: '2026-06-01' }),
-      drow({ id: 2, benchmark_type: 'agentic_traces', spec_method: 'mtp', date: '2026-06-03' }),
-      drow({ id: 3, benchmark_type: 'agentic_traces', spec_method: 'eagle', date: '2026-06-03' }),
+      drow({
+        id: 1,
+        benchmark_type: 'agentic_traces',
+        spec_method: 'none',
+        date: '2026-06-01',
+      }),
+      drow({
+        id: 2,
+        benchmark_type: 'agentic_traces',
+        spec_method: 'mtp',
+        date: '2026-06-03',
+      }),
+      drow({
+        id: 3,
+        benchmark_type: 'agentic_traces',
+        spec_method: 'eagle',
+        date: '2026-06-03',
+      }),
     ];
 
     expect(dedupeRowsToLatestPerConfig(rows).map((r) => r.id)).toEqual([2, 3]);
@@ -103,8 +122,18 @@ describe('dedupeRowsToLatestPerConfig', () => {
 
   it('continues deduping fixed-sequence spec methods independently', () => {
     const rows = [
-      drow({ id: 1, benchmark_type: 'single_turn', spec_method: 'none', date: '2026-06-01' }),
-      drow({ id: 2, benchmark_type: 'single_turn', spec_method: 'mtp', date: '2026-06-03' }),
+      drow({
+        id: 1,
+        benchmark_type: 'single_turn',
+        spec_method: 'none',
+        date: '2026-06-01',
+      }),
+      drow({
+        id: 2,
+        benchmark_type: 'single_turn',
+        spec_method: 'mtp',
+        date: '2026-06-03',
+      }),
     ];
 
     expect(dedupeRowsToLatestPerConfig(rows).map((r) => r.id)).toEqual([1, 2]);
@@ -169,6 +198,52 @@ describe('dedupeRowsToLatestPerConfig', () => {
       'eagle',
     ]);
   });
+
+  it('keeps cross-day points carried into one append-only snapshot', () => {
+    const rows = [
+      drow({
+        id: 1,
+        date: '2026-06-01',
+        workflow_run_id: 10,
+        curve_date: '2026-06-03',
+        curve_workflow_run_id: 12,
+        curve_run_started_at: '2026-06-03T12:00:00Z',
+      }),
+      drow({
+        id: 2,
+        date: '2026-06-03',
+        workflow_run_id: 12,
+        curve_date: '2026-06-03',
+        curve_workflow_run_id: 12,
+        curve_run_started_at: '2026-06-03T12:00:00Z',
+      }),
+    ];
+
+    expect(dedupeRowsToLatestPerConfig(rows).map((r) => r.id)).toEqual([1, 2]);
+  });
+
+  it('keeps same-day agentic points carried from an earlier producer run', () => {
+    const rows = [
+      drow({
+        id: 1,
+        benchmark_type: 'agentic_traces',
+        workflow_run_id: 20,
+        run_started_at: '2026-06-03T10:00:00Z',
+        curve_workflow_run_id: 21,
+        curve_run_started_at: '2026-06-03T12:00:00Z',
+      }),
+      drow({
+        id: 2,
+        benchmark_type: 'agentic_traces',
+        workflow_run_id: 21,
+        run_started_at: '2026-06-03T12:00:00Z',
+        curve_workflow_run_id: 21,
+        curve_run_started_at: '2026-06-03T12:00:00Z',
+      }),
+    ];
+
+    expect(dedupeRowsToLatestPerConfig(rows).map((r) => r.id)).toEqual([1, 2]);
+  });
 });
 
 describe('dedupeAgenticHistoryRuns', () => {
@@ -320,9 +395,19 @@ describe('filterByGPU', () => {
 describe('filterOverviewHistoryRows', () => {
   it('keeps only the serving envelope encoded by the Overview history link', () => {
     const rows = [
-      drow({ id: 1, hardware: 'mi355x', framework: 'sglang', precision: 'fp8' }),
+      drow({
+        id: 1,
+        hardware: 'mi355x',
+        framework: 'sglang',
+        precision: 'fp8',
+      }),
       drow({ id: 2, hardware: 'mi355x', framework: 'vllm', precision: 'fp4' }),
-      drow({ id: 3, hardware: 'mi355x', framework: 'sglang', precision: 'fp4' }),
+      drow({
+        id: 3,
+        hardware: 'mi355x',
+        framework: 'sglang',
+        precision: 'fp4',
+      }),
     ];
     const key = JSON.stringify(['qwen3.5', 'mi355x', 'vllm', 'none', 'fp4', false, false, 'off']);
 
@@ -409,8 +494,7 @@ describe('derived higher-is-better x-axis rooflines', () => {
   // The E2E Normalized Interactivity mode renders on the e2e chart definition (lower-x-is-better)
   // but its x-axis is higher-is-better, like interactivity. ChartDisplay
   // therefore mirrors each configured e2e corner horizontally rather than
-  // hardcoding one corner — hardcoding `upper_left` inverted the frontier for
-  // cost and joules metrics, whose good direction is a LOWER corner.
+  // hardcoding one corner.
   it('mirroring the e2e corner reproduces the interactivity corner for every y-metric', () => {
     const defs = chartDefinitions as Record<string, unknown>[];
     const e2e = defs.find((d) => d.chartType === 'e2e')!;
@@ -428,12 +512,13 @@ describe('derived higher-is-better x-axis rooflines', () => {
     }
   });
 
-  it('covers the cost metrics Bugbot flagged, not just throughput', () => {
+  it('keeps purchasing power on an upper corner and cost/joules on lower corners', () => {
     const defs = chartDefinitions as Record<string, unknown>[];
     const e2e = defs.find((d) => d.chartType === 'e2e')!;
-    // Throughput wants an upper corner, cost a lower one — a single hardcoded
-    // corner cannot serve both.
     expect(derivedModeRoofline(e2e.y_tpPerGpu_roofline as 'upper_right', true)).toBe('upper_left');
+    expect(derivedModeRoofline(e2e.y_tokensPerDollarH_roofline as 'upper_right', true)).toBe(
+      'upper_left',
+    );
     expect(derivedModeRoofline(e2e.y_costh_roofline as 'lower_left', true)).toBe('lower_right');
     expect(derivedModeRoofline(e2e.y_jTotal_roofline as 'lower_left', true)).toBe('lower_right');
   });
@@ -442,6 +527,58 @@ describe('derived higher-is-better x-axis rooflines', () => {
     expect(derivedModeRoofline('upper_right', false)).toBe('upper_right');
     expect(derivedModeRoofline(undefined, true)).toBeUndefined();
   });
+
+  it('defines every purchasing-power metric as a separate upper-is-better option', () => {
+    const defs = chartDefinitions as Record<string, unknown>[];
+    const interactivity = defs.find((d) => d.chartType === 'interactivity')!;
+    const e2e = defs.find((d) => d.chartType === 'e2e')!;
+    const metrics = [
+      'y_tokensPerDollarH',
+      'y_tokensPerDollarN',
+      'y_tokensPerDollarR',
+      'y_outputTokensPerDollarH',
+      'y_outputTokensPerDollarN',
+      'y_outputTokensPerDollarR',
+      'y_inputTokensPerDollarH',
+      'y_inputTokensPerDollarN',
+      'y_inputTokensPerDollarR',
+      'y_tokensPerDollarUser',
+    ];
+
+    for (const metric of metrics) {
+      expect(interactivity[`${metric}_title`]).toContain('per $1');
+      expect(interactivity[`${metric}_titleZh`]).toContain('每 1 美元');
+      expect(interactivity[`${metric}_roofline`]).toBe('upper_left');
+      expect(e2e[`${metric}_roofline`]).toBe('upper_right');
+    }
+    expect(interactivity.y_cost_limit).toBe(5);
+    expect(e2e.y_cost_limit).toBe(5);
+  });
+
+  it('preserves every existing cost-per-million metric as a separate lower-is-better option', () => {
+    const defs = chartDefinitions as Record<string, unknown>[];
+    const interactivity = defs.find((d) => d.chartType === 'interactivity')!;
+    const e2e = defs.find((d) => d.chartType === 'e2e')!;
+    const metrics = [
+      'y_costh',
+      'y_costn',
+      'y_costr',
+      'y_costhOutput',
+      'y_costnOutput',
+      'y_costrOutput',
+      'y_costhi',
+      'y_costni',
+      'y_costri',
+      'y_costUser',
+    ];
+
+    for (const metric of metrics) {
+      expect(interactivity[`${metric}_title`]).toContain('Cost per Million');
+      expect(interactivity[`${metric}_titleZh`]).toContain('每百万');
+      expect(interactivity[`${metric}_roofline`]).toBe('lower_right');
+      expect(e2e[`${metric}_roofline`]).toBe('lower_left');
+    }
+  });
 });
 
 const scopePoint = (hwKey: string, extra: Record<string, unknown> = {}): InferenceData =>
@@ -468,7 +605,9 @@ describe('applyScopeFilters', () => {
   // the user's legend selection with whatever set it is handed and never
   // re-widens, so a universe that shrinks when a Measured Energy axis is
   // picked deletes the telemetry-less configs for good.
-  const withTelemetry = scopePoint('b200_sglang', { measuredAvgPower: { y: 900, roof: false } });
+  const withTelemetry = scopePoint('b200_sglang', {
+    measuredAvgPower: { y: 900, roof: false },
+  });
   const withoutTelemetry = scopePoint('h200_vllm');
   const points = [withTelemetry, withoutTelemetry];
 
@@ -497,3 +636,25 @@ describe('applyScopeFilters', () => {
     expect(scoped.map((p) => p.hwKey)).toEqual(['h200_vllm']);
   });
 });
+
+describe('supportsPointTokenMetric', () => {
+  it('uses the source snapshot date when a curve is stamped with the selected display date', () => {
+    const julyRubin = scopePoint('vr200_rubin-july', {
+      date: '2026-08-22',
+      actualDate: '2026-07-01',
+    });
+
+    expect(supportsPointTokenMetric(julyRubin, 'output')).toBe(true);
+    expect(supportsPointTokenMetric(julyRubin, 'total')).toBe(false);
+    expect(supportsPointTokenMetric(julyRubin, 'input')).toBe(false);
+  });
+
+  it('does not restrict future Vera Rubin snapshots without capability metadata', () => {
+    const futureRubin = scopePoint('vr200_rubin-july', {
+      date: '2026-10-01',
+      actualDate: '2026-09-01',
+    });
+
+    expect(supportsPointTokenMetric(futureRubin, 'total')).toBe(true);
+  });
+});
diff --git a/packages/app/src/components/inference/hooks/useChartData.ts b/packages/app/src/components/inference/hooks/useChartData.ts
index 1f76ebbb1..8d0022152 100644
--- a/packages/app/src/components/inference/hooks/useChartData.ts
+++ b/packages/app/src/components/inference/hooks/useChartData.ts
@@ -3,21 +3,27 @@ import { useMemo, useRef } from 'react';
 import { useQueries } from '@tanstack/react-query';
 import { rowToSequence } from '@semianalysisai/inferencex-constants';
 
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
+import chartDefinitions, {
+  tokenMetricTypeForConfigKey,
+} from '@/components/inference/metric-registry';
 import type {
   AggDataEntry,
   ChartDefinition,
   HardwareConfig,
   InferenceData,
   RenderableGraph,
+  TokenRevenuePricing,
+  TokenRevenuePriceSource,
   YAxisMetricKey,
 } from '@/components/inference/types';
+import { applyTokenRevenuePricing } from '@/components/inference/token-revenue';
 import { partitionChartDataByLimits } from '@/components/inference/utils';
 import {
   parseComparisonEntry,
   resolveComparisonEntries,
 } from '@/components/inference/utils/comparisonEntry';
 import { useBenchmarks, benchmarkQueryOptions } from '@/hooks/api/use-benchmarks';
+import type { BenchmarkRow } from '@/lib/api';
 import {
   GPU_ALIAS_TO_CANONICAL,
   getModelSortIndex,
@@ -25,12 +31,15 @@ import {
 } from '@/lib/constants';
 import { mergeRunScopedRows, transformBenchmarkRows } from '@/lib/benchmark-transform';
 import {
+  benchmarkCurveDate,
   dedupeAgenticHistoryRuns,
   dedupeRowsToLatestPerConfig as dedupeLatestBenchmarkSeries,
 } from '@/lib/benchmark-run-selection';
 import { Sequence, type Model } from '@/lib/data-mappings';
 import { calculateCostsForGpus, calculatePowerForGpus } from '@/lib/utils';
+import { remapInferencePoint } from '@/lib/chart-utils';
 import { overviewServingSeriesKey, type OverviewServingSeriesRow } from '@/lib/overview-data';
+import { supportsChartTokenMetric, type TokenMetricType } from '@/lib/supplemental-benchmarks';
 import { resolveXAxisField } from '@/components/inference/utils/resolveXAxisField';
 import {
   applyQuickFilters,
@@ -40,9 +49,8 @@ import {
 } from '@/components/inference/utils/quickFilters';
 
 /**
- * Chart x-axis variant selected by the mode buttons above the plot. This is
- * the single definition — InferenceContext (URL/state) and ChartDisplay
- * (buttons) import it from here.
+ * Chart x-axis variant selected by the mode buttons above the plot. The
+ * inference provider and ChartDisplay import this single definition.
  */
 export type XAxisMode = 'ttft' | 'e2e' | 'interactivity' | 'e2e-normalized-interactivity';
 
@@ -124,9 +132,9 @@ export function flipRooflineDirection(dir: RooflineDirection): RooflineDirection
  * Roofline corner for a trace-derived x-axis mode. Derived modes render on the
  * e2e chart definition, whose corners assume lower-x-is-better; when the
  * derived metric is higher-is-better (E2E Normalized Interactivity) the corner mirrors
- * horizontally. This keeps the y-metric's own good direction — throughput
- * lands on an upper corner, cost and joules on a lower one — where hardcoding
- * a single corner inverted the frontier for the cost metrics.
+ * horizontally. This keeps the y-metric's own good direction — throughput and
+ * tokens-per-dollar purchasing power land on an upper corner, while cost and
+ * joules land on a lower one.
  */
 export function derivedModeRoofline(
   configuredE2eCorner: RooflineDirection | undefined,
@@ -199,6 +207,14 @@ export function applyScopeFilters(
   return scoped;
 }
 
+/** Apply snapshot-scoped metric support using the source date, not its display date. */
+export function supportsPointTokenMetric(
+  point: Pick<InferenceData, 'hwKey' | 'date' | 'actualDate'>,
+  tokenType: TokenMetricType,
+): boolean {
+  return supportsChartTokenMetric(String(point.hwKey), point.actualDate ?? point.date, tokenType);
+}
+
 export function useChartData(
   selectedModel: Model,
   selectedSequence: Sequence,
@@ -211,6 +227,8 @@ export function useChartData(
   selectedDateRange: { startDate: string; endDate: string },
   userCosts: Record<string, number | undefined> | null,
   userPowers: Record<string, number | undefined> | null,
+  tokenRevenuePricing: TokenRevenuePricing | null,
+  tokenRevenuePriceSource: TokenRevenuePriceSource,
   selectedRunDate?: string,
   enabled = true,
   latestAvailableDate?: string,
@@ -242,6 +260,8 @@ export function useChartData(
     currentConfigKey: string;
     baselineConfigKey: string;
   },
+  benchmarkQueryScope?: string,
+  initialBenchmarkRows?: BenchmarkRow[],
 ) {
   // When the selected date is the latest available, use '' (empty string) to match
   // the initial no-date query key, reusing the eagerly-fetched benchmarks from the
@@ -267,7 +287,16 @@ export function useChartData(
     data: baseRows,
     isLoading: baseLoading,
     error: baseError,
-  } = useBenchmarks(selectedModel, queryDate, enabled, asOfRunId);
+  } = useBenchmarks(
+    selectedModel,
+    queryDate,
+    enabled,
+    asOfRunId,
+    undefined,
+    undefined,
+    !asOfRunId && queryDate === '' ? initialBenchmarkRows : undefined,
+    benchmarkQueryScope,
+  );
   const {
     data: runRows,
     isLoading: runLoading,
@@ -341,9 +370,11 @@ export function useChartData(
     // an offload=on sweep can't hide a differently-dated offload=off series.
     const deduped = dedupeRowsToLatestPerConfig(seqFiltered);
 
-    const mainRows = deduped.map((r) =>
-      selectedRunDate ? { ...r, date: selectedRunDate, actualDate: r.date } : r,
-    );
+    const mainRows = deduped.map((r) => ({
+      ...r,
+      date: selectedRunDate ?? benchmarkCurveDate(r),
+      actualDate: r.date,
+    }));
     if (comparisonDates.length === 0) return mainRows;
     const extraRows = comparisonQueries.flatMap((q, i) => {
       const filtered = filterOverviewHistoryRows(
@@ -352,7 +383,11 @@ export function useChartData(
       );
       const selected =
         selectedSequence === Sequence.AgenticTraces ? dedupeAgenticHistoryRuns(filtered) : filtered;
-      return selected.map((r) => ({ ...r, date: comparisonDates[i], actualDate: r.date }));
+      return selected.map((r) => ({
+        ...r,
+        date: comparisonDates[i],
+        actualDate: r.date,
+      }));
     });
     return [...mainRows, ...extraRows];
   }, [
@@ -368,7 +403,10 @@ export function useChartData(
   // Transform filtered rows into chart data
   const { chartData, hardwareConfig: rawHardwareConfig } = useMemo(() => {
     if (rows.length === 0)
-      return { chartData: [] as InferenceData[][], hardwareConfig: {} as HardwareConfig };
+      return {
+        chartData: [] as InferenceData[][],
+        hardwareConfig: {} as HardwareConfig,
+      };
     return transformBenchmarkRows(rows, selectedPercentile);
   }, [rows, selectedPercentile]);
 
@@ -480,9 +518,6 @@ export function useChartData(
         const xAxisFlipped =
           xAxisField !== naturalX && !(chartDef.chartType === 'e2e' && isTtftOverride);
 
-        const yLabelKey = `${selectedYAxisMetric}_label` as keyof ChartDefinition;
-        const dynamicYLabel = chartDef[yLabelKey];
-
         const rooflineOverrides: Partial<ChartDefinition> = {};
         if (xAxisFlipped) {
           for (const key of Object.keys(chartDef) as (keyof ChartDefinition)[]) {
@@ -495,10 +530,36 @@ export function useChartData(
           }
         }
 
+        const revenueLabels: Partial<ChartDefinition> =
+          selectedYAxisMetric === 'y_tokenRevenuePerGpuHour'
+            ? tokenRevenuePriceSource === 'openrouter'
+              ? {
+                  y_tokenRevenuePerGpuHour_label:
+                    'Token Revenue per GPU Hour at OpenRouter Pricing ($/GPU/hr)',
+                  y_tokenRevenuePerGpuHour_labelZh:
+                    '按 OpenRouter 价格计算的每 GPU 小时 token 收入（$/GPU/hr）',
+                  y_tokenRevenuePerGpuHour_title:
+                    'Token Revenue per GPU Hour at OpenRouter Pricing',
+                  y_tokenRevenuePerGpuHour_titleZh:
+                    '按 OpenRouter 价格计算的每 GPU 小时 token 收入',
+                }
+              : {
+                  y_tokenRevenuePerGpuHour_label:
+                    'Token Revenue per GPU Hour at $1/M tok ($/GPU/hr)',
+                  y_tokenRevenuePerGpuHour_labelZh:
+                    '按 $1/百万 token 计价的每 GPU 小时 token 收入（$/GPU/hr）',
+                  y_tokenRevenuePerGpuHour_title: 'Token Revenue per GPU Hour at $1/M tok',
+                  y_tokenRevenuePerGpuHour_titleZh: '按 $1/百万 token 计价的每 GPU 小时 token 收入',
+                }
+            : {};
+        const yLabelKey = `${selectedYAxisMetric}_label` as keyof ChartDefinition;
+        const dynamicYLabel = { ...chartDef, ...revenueLabels }[yLabelKey];
+
         return {
           chartDefinition: {
             ...chartDef,
             ...rooflineOverrides,
+            ...revenueLabels,
             heading: chartHeading,
             x_label: xAxisLabel,
             y_label: dynamicYLabel === null ? undefined : String(dynamicYLabel),
@@ -513,20 +574,34 @@ export function useChartData(
       selectedE2eXAxisMetric,
       selectedPercentile,
       selectedSequence,
+      tokenRevenuePriceSource,
     ],
   );
 
   // Build renderable graphs (data processing + stable chart definitions)
   const graphs: RenderableGraph[] = useMemo(() => {
     if (chartData.length === 0) return [];
+    if (
+      selectedYAxisMetric === 'y_tokenRevenuePerGpuHour' &&
+      tokenRevenuePriceSource === 'openrouter' &&
+      !tokenRevenuePricing
+    ) {
+      return [];
+    }
 
     let dataSource: InferenceData[][] = chartData;
-    if (selectedYAxisMetric === 'y_costUser' && userCosts) {
+    if (
+      (selectedYAxisMetric === 'y_costUser' || selectedYAxisMetric === 'y_tokensPerDollarUser') &&
+      userCosts
+    ) {
       dataSource = chartData.map((d) => calculateCostsForGpus(d, userCosts));
     }
     if (selectedYAxisMetric === 'y_powerUser' && userPowers) {
       dataSource = chartData.map((d) => calculatePowerForGpus(d, userPowers));
     }
+    if (selectedYAxisMetric === 'y_tokenRevenuePerGpuHour') {
+      dataSource = chartData.map((d) => applyTokenRevenuePricing(d, tokenRevenuePricing));
+    }
 
     const result = stableChartDefinitions.map(
       ({ chartDefinition, metricKey, xAxisField }, index) => {
@@ -543,27 +618,14 @@ export function useChartData(
         // Filter to points that have the selected metric, then remap x/y.
         // Intentional cost/TTFT outliers are partitioned only after this step
         // so ScatterGraph can retain them for dashed boundary continuations.
-        const hasMetric = filteredData.some((d) => metricKey in d);
+        const tokenType = tokenMetricTypeForConfigKey(selectedYAxisMetric);
+        const metricData = filteredData.filter(
+          (d) => metricKey in d && supportsPointTokenMetric(d, tokenType),
+        );
+        const hasMetric = metricData.length > 0;
         const isTtftX = typeof xAxisField === 'string' && xAxisField.endsWith('_ttft');
         const mappedData = hasMetric
-          ? filteredData
-              .filter((d) => metricKey in d)
-              .map((d: InferenceData) => {
-                const yValue = (d[metricKey] as { y: number })?.y ?? d.y;
-                const roof = (d[metricKey] as { roof: boolean })?.roof ?? false;
-                // xAxisField is `keyof AggDataEntry`; InferenceData embeds those
-                // fields via `Partial<Omit<AggDataEntry, ...>>`, so a typed
-                // accessor catches a future field rename (silent fallthrough to
-                // d.x would otherwise mask the regression).
-                const xCandidate = (d as Partial<AggDataEntry>)[xAxisField];
-                const xValue = typeof xCandidate === 'number' ? xCandidate : d.x;
-                return {
-                  ...d,
-                  x: xValue,
-                  y: yValue,
-                  roof,
-                };
-              })
+          ? metricData.map((d) => remapInferencePoint(d, metricKey, xAxisField))
           : [];
 
         const isAgentic = selectedSequence === Sequence.AgenticTraces;
@@ -593,6 +655,8 @@ export function useChartData(
     selectedGPUs,
     userCosts,
     userPowers,
+    tokenRevenuePricing,
+    tokenRevenuePriceSource,
     stableChartDefinitions,
     compareGpuPair,
     selectedPercentile,
@@ -611,5 +675,12 @@ export function useChartData(
     [chartData, selectedGPUs, quickFilters, compareGpuPair],
   );
 
-  return { graphs, selectionPoints, loading, error, hardwareConfig, availableQuickFilters };
+  return {
+    graphs,
+    selectionPoints,
+    loading,
+    error,
+    hardwareConfig,
+    availableQuickFilters,
+  };
 }
diff --git a/packages/app/src/components/inference/hooks/useInterpolatedTrendData.test.ts b/packages/app/src/components/inference/hooks/useInterpolatedTrendData.test.ts
index a0e17a1dc..7ee3fa4c4 100644
--- a/packages/app/src/components/inference/hooks/useInterpolatedTrendData.test.ts
+++ b/packages/app/src/components/inference/hooks/useInterpolatedTrendData.test.ts
@@ -2,7 +2,15 @@ import { describe, it, expect } from 'vitest';
 
 import type { InferenceData } from '@/components/inference/types';
 
-import { interpolateMetricAtInteractivity } from './useInterpolatedTrendData';
+import type { BenchmarkRow } from '@/lib/api';
+
+import {
+  interpolateMetricAtInteractivity,
+  rowToLightweightPoint,
+  rowSupportsTrendMetric,
+  trendMetricDependencies,
+} from './useInterpolatedTrendData';
+import { SUPPLEMENTAL_BENCHMARK_ROWS } from '@/lib/supplemental-benchmarks';
 
 // ─── Factory ───
 
@@ -27,6 +35,97 @@ function makePoint(overrides: Partial<InferenceData> = {}): InferenceData {
   } as InferenceData;
 }
 
+function makeBenchmarkRow(overrides: Partial<BenchmarkRow> = {}): BenchmarkRow {
+  return {
+    id: 1,
+    hardware: 'h200',
+    framework: 'trt',
+    model: 'dsr1',
+    precision: 'fp8',
+    spec_method: 'none',
+    disagg: false,
+    is_multinode: false,
+    prefill_tp: 8,
+    prefill_ep: 1,
+    prefill_dp_attention: false,
+    prefill_num_workers: 0,
+    decode_tp: 8,
+    decode_ep: 1,
+    decode_dp_attention: false,
+    decode_num_workers: 0,
+    num_prefill_gpu: 8,
+    num_decode_gpu: 8,
+    benchmark_type: 'single_turn',
+    offload_mode: 'off',
+    isl: 1024,
+    osl: 1024,
+    conc: 64,
+    image: 'test',
+    metrics: {
+      tput_per_gpu: 400,
+      median_intvty: 20,
+    },
+    date: '2026-03-01',
+    run_url: null,
+    ...overrides,
+  };
+}
+
+describe('rowToLightweightPoint', () => {
+  it('preserves legacy output-throughput trends when the explicit field is absent', () => {
+    const point = rowToLightweightPoint(makeBenchmarkRow(), [
+      'outputTputPerGpu',
+      'costhOutput',
+      'outputTokensPerDollarH',
+    ]);
+
+    expect(point?.outputTputPerGpu?.y).toBe(400);
+    expect(point?.costhOutput?.y).toBeGreaterThan(0);
+    expect(point?.outputTokensPerDollarH?.y).toBeGreaterThan(0);
+  });
+
+  it('applies OpenRouter input and output prices to historical points', () => {
+    const point = rowToLightweightPoint(
+      makeBenchmarkRow({
+        metrics: {
+          tput_per_gpu: 400,
+          input_tput_per_gpu: 320,
+          output_tput_per_gpu: 80,
+          median_intvty: 20,
+        },
+      }),
+      ['tokenRevenuePerGpuHour'],
+      {
+        source: 'openrouter',
+        inputPerMillion: 1.122,
+        outputPerMillion: 3.366,
+        openRouterModelId: 'deepseek/deepseek-v4-pro-0813',
+      },
+    );
+
+    expect(point?.tokenRevenuePerGpuHour?.y).toBeCloseTo(2.261952, 10);
+  });
+});
+
+describe('rowSupportsTrendMetric', () => {
+  it('limits the July Vera Rubin snapshot to output-token trend metrics', () => {
+    const julyRubin = SUPPLEMENTAL_BENCHMARK_ROWS.find((row) => row.hardware === 'vr200')!;
+
+    expect(rowSupportsTrendMetric(julyRubin, 'y_outputTputPerGpu')).toBe(true);
+    expect(rowSupportsTrendMetric(julyRubin, 'y_tpPerGpu')).toBe(false);
+    expect(rowSupportsTrendMetric(julyRubin, 'y_inputTputPerGpu')).toBe(false);
+  });
+
+  it('leaves a future Vera Rubin snapshot unrestricted without capability metadata', () => {
+    const futureRubin = {
+      ...SUPPLEMENTAL_BENCHMARK_ROWS.find((row) => row.hardware === 'vr200')!,
+      date: '2026-09-01',
+    };
+
+    expect(rowSupportsTrendMetric(futureRubin, 'y_tpPerGpu')).toBe(true);
+  });
+});
+
 // ─── Tests ───
 
 describe('interpolateMetricAtInteractivity', () => {
@@ -85,16 +184,84 @@ describe('interpolateMetricAtInteractivity', () => {
     expect(result!).toBeGreaterThanOrEqual(0);
   });
 
-  it('works with cost metrics (costh)', () => {
+  it('interpolates tokens per dollar as a throughput-proportional metric', () => {
     const points = [
-      makePoint({ x: 20, tpPerGpu: { y: 800, roof: false }, costh: { y: 1, roof: false } }),
-      makePoint({ x: 40, tpPerGpu: { y: 600, roof: false }, costh: { y: 1.5, roof: false } }),
-      makePoint({ x: 60, tpPerGpu: { y: 400, roof: false }, costh: { y: 2, roof: false } }),
+      makePoint({
+        x: 20,
+        tpPerGpu: { y: 800, roof: false },
+        tokensPerDollarH: { y: 800_000, roof: false },
+      }),
+      makePoint({
+        x: 40,
+        tpPerGpu: { y: 600, roof: false },
+        tokensPerDollarH: { y: 600_000, roof: false },
+      }),
+      makePoint({
+        x: 60,
+        tpPerGpu: { y: 400, roof: false },
+        tokensPerDollarH: { y: 400_000, roof: false },
+      }),
     ];
-    const result = interpolateMetricAtInteractivity(points, 30, 'costh');
-    expect(result).not.toBeNull();
-    expect(result!).toBeGreaterThan(0.9);
-    expect(result!).toBeLessThan(1.6);
+    const result = interpolateMetricAtInteractivity(points, 30, 'tokensPerDollarH');
+    expect(result).toBeCloseTo(700_000, 0);
+  });
+
+  it('uses the matching output-throughput Pareto knots for output tokens per dollar', () => {
+    const points = [
+      makePoint({
+        x: 20,
+        tpPerGpu: { y: 1000, roof: false },
+        outputTputPerGpu: { y: 100, roof: false },
+        outputTokensPerDollarH: { y: 100_000, roof: false },
+      }),
+      makePoint({
+        x: 40,
+        tpPerGpu: { y: 800, roof: false },
+        outputTputPerGpu: { y: 700, roof: false },
+        outputTokensPerDollarH: { y: 700_000, roof: false },
+      }),
+      makePoint({
+        x: 60,
+        tpPerGpu: { y: 900, roof: false },
+        outputTputPerGpu: { y: 500, roof: false },
+        outputTokensPerDollarH: { y: 500_000, roof: false },
+      }),
+    ];
+
+    // The total-throughput frontier drops x=40, while the output-throughput
+    // frontier keeps x=40 and drops x=20. The midpoint must follow the latter.
+    expect(interpolateMetricAtInteractivity(points, 50, 'outputTokensPerDollarH')).toBeCloseTo(
+      581_250,
+      0,
+    );
+  });
+
+  it('uses the matching input-throughput Pareto knots for input tokens per dollar', () => {
+    const points = [
+      makePoint({
+        x: 20,
+        tpPerGpu: { y: 1000, roof: false },
+        inputTputPerGpu: { y: 100, roof: false },
+        inputTokensPerDollarH: { y: 200_000, roof: false },
+      }),
+      makePoint({
+        x: 40,
+        tpPerGpu: { y: 800, roof: false },
+        inputTputPerGpu: { y: 700, roof: false },
+        inputTokensPerDollarH: { y: 1_400_000, roof: false },
+      }),
+      makePoint({
+        x: 60,
+        tpPerGpu: { y: 900, roof: false },
+        inputTputPerGpu: { y: 500, roof: false },
+        inputTokensPerDollarH: { y: 1_000_000, roof: false },
+      }),
+    ];
+
+    expect(interpolateMetricAtInteractivity(points, 50, 'inputTokensPerDollarH')).toBeCloseTo(
+      1_162_500,
+      0,
+    );
   });
 
   it('filters dominated points via Pareto front', () => {
@@ -156,6 +323,30 @@ describe('interpolateMetricAtInteractivity', () => {
     expect(result!).toBeLessThan(3.5);
   });
 
+  it.each([
+    'measuredJPerSuccessfulQuery',
+    'measuredWhPerSuccessfulQuery',
+    'measuredPowerPercentTdp',
+  ] as const)('interpolates the derived measured metric %s', (metricKey) => {
+    const points = [
+      makePoint({
+        x: 20,
+        tpPerGpu: { y: 800, roof: false },
+        [metricKey]: { y: 80, roof: false },
+      }),
+      makePoint({
+        x: 60,
+        tpPerGpu: { y: 400, roof: false },
+        [metricKey]: { y: 40, roof: false },
+      }),
+    ];
+
+    const result = interpolateMetricAtInteractivity(points, 40, metricKey);
+    expect(result).not.toBeNull();
+    expect(result!).toBeGreaterThan(40);
+    expect(result!).toBeLessThan(80);
+  });
+
   it('returns null when metric field is missing from data points', () => {
     const points = [
       makePoint({ x: 20, tpPerGpu: { y: 800, roof: false } }),
@@ -199,6 +390,25 @@ describe('interpolateMetricAtInteractivity', () => {
     expect(result!).toBeLessThan(700);
   });
 
+  it('keeps token revenue proportional to the interpolated total throughput', () => {
+    const points = [
+      makePoint({
+        x: 20,
+        tpPerGpu: { y: 800, roof: false },
+        tokenRevenuePerGpuHour: { y: 2.88, roof: false },
+      }),
+      makePoint({
+        x: 60,
+        tpPerGpu: { y: 400, roof: false },
+        tokenRevenuePerGpuHour: { y: 1.44, roof: false },
+      }),
+    ];
+
+    const throughput = interpolateMetricAtInteractivity(points, 40, 'tpPerGpu');
+    const revenue = interpolateMetricAtInteractivity(points, 40, 'tokenRevenuePerGpuHour');
+    expect(revenue).toBeCloseTo(throughput! * 0.0036, 10);
+  });
+
   it('returns exact boundary value at the lowest frontier x', () => {
     const points = [
       makePoint({ x: 10, tpPerGpu: { y: 1000, roof: false } }),
@@ -221,3 +431,17 @@ describe('interpolateMetricAtInteractivity', () => {
     expect(result!).toBeCloseTo(200, 0);
   });
 });
+
+describe('trendMetricDependencies', () => {
+  it('selects only the metric and matching throughput needed by reciprocal formulas', () => {
+    expect(trendMetricDependencies('costhOutput')).toEqual([
+      'tpPerGpu',
+      'costhOutput',
+      'outputTputPerGpu',
+    ]);
+  });
+
+  it('does not route custom user metrics through benchmark-derived history fields', () => {
+    expect(trendMetricDependencies('costUser')).toEqual(['tpPerGpu']);
+  });
+});
diff --git a/packages/app/src/components/inference/hooks/useInterpolatedTrendData.ts b/packages/app/src/components/inference/hooks/useInterpolatedTrendData.ts
index bfef739d3..47c06962d 100644
--- a/packages/app/src/components/inference/hooks/useInterpolatedTrendData.ts
+++ b/packages/app/src/components/inference/hooks/useInterpolatedTrendData.ts
@@ -2,7 +2,21 @@ import { useEffect, useMemo, useRef, useState } from 'react';
 
 import { sequenceToIslOsl } from '@semianalysisai/inferencex-constants';
 
-import type { InferenceData, TrendDataPoint, YAxisMetricKey } from '@/components/inference/types';
+import type {
+  InferenceData,
+  TokenRevenuePricing,
+  TrendDataPoint,
+  YAxisMetricKey,
+} from '@/components/inference/types';
+import {
+  applyTokenRevenuePricing,
+  NORMALIZED_TOKEN_REVENUE_PRICING,
+} from '@/components/inference/token-revenue';
+import {
+  isBenchmarkMetricKey,
+  resolveMetricConfigKey,
+  tokenMetricTypeForConfigKey,
+} from '@/components/inference/metric-registry';
 import {
   hermiteInterpolate,
   monotoneSlopes,
@@ -11,89 +25,61 @@ import {
   reciprocalMetricAt,
 } from '@/components/calculator/useThroughputData';
 import { useBenchmarkHistory } from '@/hooks/api/use-benchmark-history';
-import { getHardwareKey } from '@/lib/chart-utils';
-import { getGpuSpecs, isKnownGpu } from '@/lib/constants';
+import { buildDerivedChartFields, getHardwareKey, type DerivedMetricKey } from '@/lib/chart-utils';
+import { isKnownGpu } from '@/lib/constants';
 import { rowToAggDataEntry } from '@/lib/benchmark-transform';
 import type { BenchmarkRow } from '@/lib/api';
-import { dedupeAgenticHistoryRuns } from '@/lib/benchmark-run-selection';
+import { benchmarkCurveDate, dedupeAgenticHistoryRuns } from '@/lib/benchmark-run-selection';
 import { Sequence, type Model } from '@/lib/data-mappings';
+import { supportsTokenMetric } from '@/lib/supplemental-benchmarks';
 
-// Trend points never sit on a roofline — they're synthetic per-(date, config)
-// aggregates, not the per-load Pareto-frontier points the chart marks. Hardcode
-// roof:false so the field shape lines up with InferenceData without a cast.
-const wrapMetric = (n: number): { y: number; roof: boolean } => ({ y: n, roof: false });
+/** Snapshot-scoped token metric support for raw historical rows. */
+export function rowSupportsTrendMetric(row: BenchmarkRow, selectedYAxisMetric: string): boolean {
+  return supportsTokenMetric(row, tokenMetricTypeForConfigKey(selectedYAxisMetric));
+}
 
 /**
  * Build a lightweight InferenceData-compatible point from a raw BenchmarkRow.
- * Skips the expensive transformBenchmarkRows pipeline (rooflines, cost derivations)
- * since the trend interpolation only needs x (interactivity), tpPerGpu, and metric values.
+ * This deliberately skips full chart transformation and derives only the
+ * selected trend metric plus its frontier/interpolation dependencies.
  */
-function rowToLightweightPoint(row: BenchmarkRow): InferenceData | null {
+export function rowToLightweightPoint(
+  row: BenchmarkRow,
+  requestedMetrics: readonly DerivedMetricKey[],
+  tokenRevenuePricing: TokenRevenuePricing | null = NORMALIZED_TOKEN_REVENUE_PRICING,
+): InferenceData | null {
   const entry = rowToAggDataEntry(row);
   const hwKey = getHardwareKey(entry);
+  // Historical rows predating explicit output-throughput telemetry used total
+  // throughput as output throughput. Preserve that production fallback for
+  // output cost, purchasing-power, and energy trend metrics.
+  const derivedEntry =
+    row.metrics.output_tput_per_gpu === null || row.metrics.output_tput_per_gpu === undefined
+      ? { ...entry, output_tput_per_gpu: entry.tput_per_gpu }
+      : entry;
   if (!isKnownGpu(hwKey)) return null;
 
-  const m = row.metrics;
-  const tput = m.tput_per_gpu ?? 0;
-  const outputTput = m.output_tput_per_gpu ?? tput;
-  const inputTput = m.input_tput_per_gpu ?? 0;
-  const specs = getGpuSpecs(hwKey);
-  const power = specs.power;
-
-  const tokPerHr = (tput * 3600) / 1_000_000;
-  const outTokPerHr = (outputTput * 3600) / 1_000_000;
-  const inTokPerHr = (inputTput * 3600) / 1_000_000;
-
-  // Build metric objects matching InferenceData shape. Measured-power keys are
-  // only set when the runner-side aggregate_power.py emitted them — leaving the
-  // field undefined lets extractMetric return null and the trend show a real
-  // gap instead of a flat-zero line.
-  const point: InferenceData = {
-    x: m.median_intvty ?? 0,
-    y: tput,
+  const point = {
+    x: row.metrics.median_intvty ?? 0,
+    y: row.metrics.tput_per_gpu ?? 0,
     hwKey,
     precision: row.precision,
     tp: row.decode_tp,
     conc: row.conc,
-    date: row.date,
-    tpPerGpu: wrapMetric(tput),
-    outputTputPerGpu: wrapMetric(outputTput),
-    inputTputPerGpu: wrapMetric(inputTput),
-    tpPerMw: wrapMetric(power > 0 ? (tput * 1000) / power : 0),
-    // Cost per million tokens (total / output / input)
-    costh: wrapMetric(tokPerHr ? specs.costh / tokPerHr : 0),
-    costn: wrapMetric(tokPerHr ? specs.costn / tokPerHr : 0),
-    costr: wrapMetric(tokPerHr ? specs.costr / tokPerHr : 0),
-    costhOutput: wrapMetric(outTokPerHr ? specs.costh / outTokPerHr : 0),
-    costnOutput: wrapMetric(outTokPerHr ? specs.costn / outTokPerHr : 0),
-    costrOutput: wrapMetric(outTokPerHr ? specs.costr / outTokPerHr : 0),
-    costhi: wrapMetric(inTokPerHr ? specs.costh / inTokPerHr : 0),
-    costni: wrapMetric(inTokPerHr ? specs.costn / inTokPerHr : 0),
-    costri: wrapMetric(inTokPerHr ? specs.costr / inTokPerHr : 0),
-    // Energy: J/token = W / tok/s
-    jTotal: wrapMetric(power > 0 && tput ? (power * 1000) / tput : 0),
-    ...(outputTput ? { jOutput: wrapMetric(power > 0 ? (power * 1000) / outputTput : 0) } : {}),
-    ...(inputTput ? { jInput: wrapMetric(power > 0 ? (power * 1000) / inputTput : 0) } : {}),
-    ...(typeof entry.avg_power_w === 'number'
-      ? { measuredAvgPower: { y: entry.avg_power_w, roof: false } }
-      : {}),
-    ...(typeof entry.joules_per_output_token === 'number'
-      ? { measuredJPerOutputToken: { y: entry.joules_per_output_token, roof: false } }
-      : {}),
-    ...(typeof entry.joules_per_total_token === 'number'
-      ? { measuredJPerTotalToken: { y: entry.joules_per_total_token, roof: false } }
-      : {}),
-    ...(typeof entry.prefill_avg_power_w === 'number'
-      ? { measuredPrefillAvgPower: { y: entry.prefill_avg_power_w, roof: false } }
-      : {}),
-    ...(typeof entry.decode_avg_power_w === 'number'
-      ? { measuredDecodeAvgPower: { y: entry.decode_avg_power_w, roof: false } }
-      : {}),
-    ...(typeof entry.joules_per_input_token === 'number'
-      ? { measuredJPerInputToken: { y: entry.joules_per_input_token, roof: false } }
-      : {}),
-  };
-  return point;
+    date: benchmarkCurveDate(row),
+    tput_per_gpu: entry.tput_per_gpu,
+    input_tput_per_gpu: entry.input_tput_per_gpu,
+    output_tput_per_gpu: entry.output_tput_per_gpu,
+    isl: entry.isl,
+    osl: entry.osl,
+    total_prompt_tokens: entry.total_prompt_tokens,
+    total_generation_tokens: entry.total_generation_tokens,
+    ...buildDerivedChartFields(derivedEntry, hwKey, requestedMetrics),
+  } as InferenceData;
+
+  return requestedMetrics.includes('tokenRevenuePerGpuHour')
+    ? applyTokenRevenuePricing([point], tokenRevenuePricing)[0]!
+    : point;
 }
 
 /**
@@ -124,9 +110,71 @@ const RECIPROCAL_OF_THROUGHPUT: Partial<Record<YAxisMetricKey, YAxisMetricKey>>
   jInput: 'inputTputPerGpu',
 };
 
+/**
+ * Business metrics mapped to the throughput they scale. When the multiplier is
+ * constant, interpolation preserves that identity. OpenRouter revenue can use
+ * a point-specific input/output mix; in that case multiplier recovery fails
+ * safely and the metric itself is splined on the total-throughput frontier.
+ */
+const PROPORTIONAL_TO_THROUGHPUT: Partial<Record<YAxisMetricKey, YAxisMetricKey>> = {
+  tokenRevenuePerGpuHour: 'tpPerGpu',
+  tokensPerDollarH: 'tpPerGpu',
+  tokensPerDollarN: 'tpPerGpu',
+  tokensPerDollarR: 'tpPerGpu',
+  outputTokensPerDollarH: 'outputTputPerGpu',
+  outputTokensPerDollarN: 'outputTputPerGpu',
+  outputTokensPerDollarR: 'outputTputPerGpu',
+  inputTokensPerDollarH: 'inputTputPerGpu',
+  inputTokensPerDollarN: 'inputTputPerGpu',
+  inputTokensPerDollarR: 'inputTputPerGpu',
+  tokensPerRmbH: 'tpPerGpu',
+  tokensPerRmbN: 'tpPerGpu',
+  tokensPerRmbR: 'tpPerGpu',
+  outputTokensPerRmbH: 'outputTputPerGpu',
+  outputTokensPerRmbN: 'outputTputPerGpu',
+  outputTokensPerRmbR: 'outputTputPerGpu',
+  inputTokensPerRmbH: 'inputTputPerGpu',
+  inputTokensPerRmbN: 'inputTputPerGpu',
+  inputTokensPerRmbR: 'inputTputPerGpu',
+};
+
+export function trendMetricDependencies(metricKey: YAxisMetricKey): DerivedMetricKey[] {
+  const dependencies = new Set<DerivedMetricKey>(['tpPerGpu']);
+  if (!isBenchmarkMetricKey(metricKey)) return [...dependencies];
+  dependencies.add(metricKey);
+  const throughputKey =
+    RECIPROCAL_OF_THROUGHPUT[metricKey] ?? PROPORTIONAL_TO_THROUGHPUT[metricKey];
+  if (throughputKey && isBenchmarkMetricKey(throughputKey)) dependencies.add(throughputKey);
+  return [...dependencies];
+}
+
+function recoverProportionalMultiplier(
+  values: readonly number[],
+  throughputs: readonly number[],
+): number | null {
+  const RELATIVE_TOLERANCE = 1e-3;
+  let multiplier: number | null = null;
+
+  for (let i = 0; i < values.length; i += 1) {
+    const value = values[i];
+    const throughput = throughputs[i];
+    if (value === undefined || throughput === undefined) continue;
+    if (!(value > 0) || !(throughput > 0)) continue;
+
+    const candidate = value / throughput;
+    if (multiplier === null) {
+      multiplier = candidate;
+    } else if (Math.abs(candidate - multiplier) > Math.abs(multiplier) * RELATIVE_TOLERANCE) {
+      return null;
+    }
+  }
+
+  return multiplier;
+}
+
 /**
  * Interpolate a selected metric at a target interactivity for a set of InferenceData points
- * from a single GPU. Uses Pareto front (throughput-based frontier) + monotone cubic Hermite spline.
+ * from a single GPU. Uses a throughput-based Pareto front + monotone cubic Hermite spline.
  *
  * Exported for unit testing.
  */
@@ -137,11 +185,19 @@ export function interpolateMetricAtInteractivity(
 ): number | null {
   if (points.length === 0) return null;
 
-  // Build Pareto front on interactivity(x) vs throughput(y)
+  // Proportional business metrics use the corresponding total/output/input
+  // throughput frontier so their knots exactly match the serving envelope.
+  const proportionalThroughputKey = PROPORTIONAL_TO_THROUGHPUT[metricKey];
+  const frontierThroughputKey = proportionalThroughputKey ?? 'tpPerGpu';
+  for (const point of points) {
+    if (extractMetric(point, frontierThroughputKey) === null) return null;
+  }
+
+  // Build Pareto front on interactivity(x) vs the applicable throughput(y).
   const frontier = paretoFrontUpperLeft<InferenceData>(
     points,
     (p) => p.x,
-    (p) => p.tpPerGpu.y,
+    (p) => extractMetric(p, frontierThroughputKey)!,
   );
   if (frontier.length === 0) return null;
 
@@ -172,9 +228,22 @@ export function interpolateMetricAtInteractivity(
     metricYs.push(v);
   }
 
+  // When a business metric is a fixed multiple of throughput, spline the
+  // matching throughput and apply that multiplier so the derived curve cannot
+  // drift from its throughput/interactivity Pareto curve. OpenRouter revenue
+  // falls through to a direct metric spline when the token mix varies by point.
+  if (proportionalThroughputKey) {
+    const tputYs = sorted.map((p) => extractMetric(p, proportionalThroughputKey)!);
+    const multiplier = recoverProportionalMultiplier(metricYs, tputYs);
+    if (multiplier !== null) {
+      const tputSlopes = monotoneSlopes(xs, tputYs);
+      const tput = hermiteInterpolate(xs, tputYs, tputSlopes, targetInteractivity);
+      return Math.max(0, tput) * multiplier;
+    }
+  }
+
   // Cost and energy per token are `constant / throughput`. Spline that
-  // throughput and re-derive rather than splining the metric, so the value
-  // agrees with the per-point figures on the inference chart.
+  // throughput and re-derive rather than splining the metric, preserving the identity.
   const throughputKey = RECIPROCAL_OF_THROUGHPUT[metricKey];
   if (throughputKey) {
     const tputYs: number[] = [];
@@ -216,6 +285,7 @@ interface UseInterpolatedTrendDataParams {
   selectedYAxisMetric: string;
   targetInteractivity: number;
   availableDates: string[];
+  tokenRevenuePricing?: TokenRevenuePricing | null;
   enabled: boolean;
 }
 
@@ -239,6 +309,7 @@ export function useInterpolatedTrendData({
   selectedPrecisions,
   selectedYAxisMetric,
   targetInteractivity,
+  tokenRevenuePricing = NORMALIZED_TOKEN_REVENUE_PRICING,
   enabled,
 }: UseInterpolatedTrendDataParams): UseInterpolatedTrendDataResult {
   const seqIslOsl = useMemo(() => sequenceToIslOsl(selectedSequence), [selectedSequence]);
@@ -247,8 +318,10 @@ export function useInterpolatedTrendData({
     enabled ? selectedModel : '',
     seqIslOsl?.isl ?? 0,
     seqIslOsl?.osl ?? 0,
-    selectedSequence === Sequence.AgenticTraces ? 'agentic_traces' : undefined,
+    selectedSequence === Sequence.AgenticTraces ? { benchmarkType: 'agentic_traces' } : undefined,
   );
+  const trendMetricKey = resolveMetricConfigKey(selectedYAxisMetric).slice(2) as YAxisMetricKey;
+  const requestedMetrics = useMemo(() => trendMetricDependencies(trendMetricKey), [trendMetricKey]);
 
   // Build lightweight InferenceData points grouped by date and hwKey.
   // Skips the full transformBenchmarkRows pipeline (~100x faster for ~100 dates).
@@ -259,14 +332,16 @@ export function useInterpolatedTrendData({
 
     for (const row of dedupeAgenticHistoryRuns(allRows)) {
       if (!selectedPrecisions.includes(row.precision)) continue;
+      if (!rowSupportsTrendMetric(row, selectedYAxisMetric)) continue;
 
-      const point = rowToLightweightPoint(row);
+      const point = rowToLightweightPoint(row, requestedMetrics, tokenRevenuePricing);
       if (!point) continue;
 
-      let dateMap = result.get(row.date);
+      const curveDate = benchmarkCurveDate(row);
+      let dateMap = result.get(curveDate);
       if (!dateMap) {
         dateMap = new Map();
-        result.set(row.date, dateMap);
+        result.set(curveDate, dateMap);
       }
 
       const hwKey = point.hwKey as string;
@@ -281,12 +356,12 @@ export function useInterpolatedTrendData({
     }
 
     return result;
-  }, [allRows, selectedPrecisions]);
+  }, [allRows, selectedPrecisions, requestedMetrics, selectedYAxisMetric, tokenRevenuePricing]);
 
   // Interpolation memo — instant when slider moves or metric changes
   const { trendLines, hwKeysWithData } = useMemo(() => {
     const resultMap = new Map<string, Map<string, TrendDataPoint>>();
-    const metricKey = selectedYAxisMetric.replace('y_', '') as YAxisMetricKey;
+    const metricKey = trendMetricKey;
 
     for (const [date, byGroupKey] of dateGroupedData) {
       for (const [groupKey, points] of byGroupKey) {
@@ -296,7 +371,6 @@ export function useInterpolatedTrendData({
           metricKey,
         );
         if (interpolated === null) continue;
-
         if (!resultMap.has(groupKey)) resultMap.set(groupKey, new Map());
         resultMap.get(groupKey)!.set(date, {
           date,
@@ -319,7 +393,12 @@ export function useInterpolatedTrendData({
         // Extend line to today if the last point is before today
         const last = points.at(-1)!;
         if (last.date < today) {
-          points.push({ date: today, value: last.value, x: last.x, synthetic: true });
+          points.push({
+            date: today,
+            value: last.value,
+            x: last.x,
+            synthetic: true,
+          });
         }
         lines.set(groupKey, points);
         // Return base hwKey for legend filtering
@@ -331,7 +410,7 @@ export function useInterpolatedTrendData({
     }
 
     return { trendLines: lines, hwKeysWithData: keysWithData };
-  }, [dateGroupedData, targetInteractivity, selectedYAxisMetric]);
+  }, [dateGroupedData, targetInteractivity, trendMetricKey]);
 
   // Artificial progress that ramps up while the API call is in flight
   const [progress, setProgress] = useState(0);
@@ -353,7 +432,12 @@ export function useInterpolatedTrendData({
   }, [isLoading]);
 
   if (!enabled) {
-    return { trendLines: new Map(), hwKeysWithData: [], loading: false, progress: 0 };
+    return {
+      trendLines: new Map(),
+      hwKeysWithData: [],
+      loading: false,
+      progress: 0,
+    };
   }
 
   return { trendLines, hwKeysWithData, loading: isLoading, progress };
diff --git a/packages/app/src/components/inference/inference-chart-config.json b/packages/app/src/components/inference/inference-chart-config.json
deleted file mode 100644
index 392909349..000000000
--- a/packages/app/src/components/inference/inference-chart-config.json
+++ /dev/null
@@ -1,337 +0,0 @@
-[
-  {
-    "chartType": "interactivity",
-    "heading": "vs. Interactivity",
-    "x": "median_intvty",
-    "x_label": "Interactivity (tok/s/user)",
-    "y": "tput_per_gpu",
-    "y_tpPerGpu": "tpPerGpu.y",
-    "y_tpPerGpu_label": "Token Throughput per Chip (tok/s/chip)",
-    "y_tpPerGpu_labelZh": "每 Chip token 吞吐量（tok/s/chip）",
-    "y_tpPerGpu_title": "Token Throughput per Chip",
-    "y_tpPerGpu_titleZh": "每 Chip token 吞吐量",
-    "y_tpPerGpu_roofline": "upper_left",
-    "y_inputTputPerGpu": "inputTputPerGpu.y",
-    "y_inputTputPerGpu_label": "Input Token Throughput per Chip (tok/s/chip)",
-    "y_inputTputPerGpu_labelZh": "每 Chip 输入 token 吞吐量（tok/s/chip）",
-    "y_inputTputPerGpu_title": "Input Token Throughput per Chip",
-    "y_inputTputPerGpu_titleZh": "每 Chip 输入 token 吞吐量",
-    "y_inputTputPerGpu_roofline": "upper_left",
-    "y_inputTputPerGpu_x": "p90_ttft",
-    "y_inputTputPerGpu_x_label": "P90 Time To First Token (s)",
-    "y_inputTputPerGpu_heading": "vs. P90 Time To First Token",
-    "y_outputTputPerGpu": "outputTputPerGpu.y",
-    "y_outputTputPerGpu_label": "Output Token Throughput per Chip (tok/s/chip)",
-    "y_outputTputPerGpu_labelZh": "每 Chip 输出 token 吞吐量（tok/s/chip）",
-    "y_outputTputPerGpu_title": "Output Token Throughput per Chip",
-    "y_outputTputPerGpu_titleZh": "每 Chip 输出 token 吞吐量",
-    "y_outputTputPerGpu_roofline": "upper_left",
-    "y_tpPerMw": "tpPerMw.y",
-    "y_tpPerMw_label": "Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_tpPerMw_labelZh": "每全电源配置兆瓦 token 吞吐量（tok/s/MW）",
-    "y_tpPerMw_title": "Token Throughput per All in Utility MW",
-    "y_tpPerMw_titleZh": "每全电源配置兆瓦 token 吞吐量",
-    "y_tpPerMw_roofline": "upper_left",
-    "y_inputTputPerMw": "inputTputPerMw.y",
-    "y_inputTputPerMw_label": "Input Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_inputTputPerMw_labelZh": "每全电源配置兆瓦输入 token 吞吐量（tok/s/MW）",
-    "y_inputTputPerMw_title": "Input Token Throughput per All in Utility MW",
-    "y_inputTputPerMw_titleZh": "每全电源配置兆瓦输入 token 吞吐量",
-    "y_inputTputPerMw_roofline": "upper_left",
-    "y_outputTputPerMw": "outputTputPerMw.y",
-    "y_outputTputPerMw_label": "Output Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_outputTputPerMw_labelZh": "每全电源配置兆瓦输出 token 吞吐量（tok/s/MW）",
-    "y_outputTputPerMw_title": "Output Token Throughput per All in Utility MW",
-    "y_outputTputPerMw_titleZh": "每全电源配置兆瓦输出 token 吞吐量",
-    "y_outputTputPerMw_roofline": "upper_left",
-    "y_costh": "costh.y",
-    "y_costh_label": "Cost per Million Total Tokens ($)",
-    "y_costh_labelZh": "每百万总 token 成本（$）",
-    "y_costh_title": "Cost per Million Total Tokens (Owning - Hyperscaler)",
-    "y_costh_titleZh": "每百万总 token 成本（自有 - 超大规模）",
-    "y_costh_roofline": "lower_right",
-    "y_costn": "costn.y",
-    "y_costn_label": "Cost per Million Total Tokens ($)",
-    "y_costn_labelZh": "每百万总 token 成本（$）",
-    "y_costn_title": "Cost per Million Total Tokens (Owning - Neocloud Giant)",
-    "y_costn_titleZh": "每百万总 token 成本（自有 - Neocloud Giant）",
-    "y_costn_roofline": "lower_right",
-    "y_costr": "costr.y",
-    "y_costr_label": "Cost per Million Total Tokens ($)",
-    "y_costr_labelZh": "每百万总 token 成本（$）",
-    "y_costr_title": "Cost per Million Total Tokens (3 Year Rental)",
-    "y_costr_titleZh": "每百万总 token 成本（3 年租赁）",
-    "y_costr_roofline": "lower_right",
-    "y_costhOutput": "costhOutput.y",
-    "y_costhOutput_label": "Cost per Million Output Tokens ($)",
-    "y_costhOutput_labelZh": "每百万输出 token 成本（$）",
-    "y_costhOutput_title": "Cost per Million Output Tokens (Owning - Hyperscaler)",
-    "y_costhOutput_titleZh": "每百万输出 token 成本（自有 - 超大规模）",
-    "y_costhOutput_roofline": "lower_right",
-    "y_costnOutput": "costnOutput.y",
-    "y_costnOutput_label": "Cost per Million Output Tokens ($)",
-    "y_costnOutput_labelZh": "每百万输出 token 成本（$）",
-    "y_costnOutput_title": "Cost per Million Output Tokens (Owning - Neocloud Giant)",
-    "y_costnOutput_titleZh": "每百万输出 token 成本（自有 - Neocloud Giant）",
-    "y_costnOutput_roofline": "lower_right",
-    "y_costrOutput": "costrOutput.y",
-    "y_costrOutput_label": "Cost per Million Output Tokens ($)",
-    "y_costrOutput_labelZh": "每百万输出 token 成本（$）",
-    "y_costrOutput_title": "Cost per Million Output Tokens (3 Year Rental)",
-    "y_costrOutput_titleZh": "每百万输出 token 成本（3 年租赁）",
-    "y_costrOutput_roofline": "lower_right",
-    "y_costhi": "costhi.y",
-    "y_costhi_label": "Cost per Million Input Tokens ($)",
-    "y_costhi_labelZh": "每百万输入 token 成本（$）",
-    "y_costhi_title": "Cost per Million Input Tokens (Owning - Hyperscaler)",
-    "y_costhi_titleZh": "每百万输入 token 成本（自有 - 超大规模）",
-    "y_costhi_roofline": "lower_right",
-    "y_costni": "costni.y",
-    "y_costni_label": "Cost per Million Input Tokens ($)",
-    "y_costni_labelZh": "每百万输入 token 成本（$）",
-    "y_costni_title": "Cost per Million Input Tokens (Owning - Neocloud Giant)",
-    "y_costni_titleZh": "每百万输入 token 成本（自有 - Neocloud Giant）",
-    "y_costni_roofline": "lower_right",
-    "y_costri": "costri.y",
-    "y_costri_label": "Cost per Million Input Tokens ($)",
-    "y_costri_labelZh": "每百万输入 token 成本（$）",
-    "y_costri_title": "Cost per Million Input Tokens (3 Year Rental)",
-    "y_costri_titleZh": "每百万输入 token 成本（3 年租赁）",
-    "y_costri_roofline": "lower_right",
-    "y_costUser": "costUser.y",
-    "y_costUser_label": "Cost per Million Total Tokens ($)",
-    "y_costUser_labelZh": "每百万总 token 成本（$）",
-    "y_costUser_title": "Cost per Million Total Tokens (Custom User Values)",
-    "y_costUser_titleZh": "每百万总 token 成本（自定义值）",
-    "y_costUser_roofline": "lower_right",
-    "y_powerUser": "powerUser.y",
-    "y_powerUser_label": "Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_powerUser_labelZh": "每全电源配置兆瓦 token 吞吐量（tok/s/MW）",
-    "y_powerUser_title": "Token Throughput per All in Utility MW (Custom User Values)",
-    "y_powerUser_titleZh": "每全电源配置兆瓦 token 吞吐量（自定义值）",
-    "y_powerUser_roofline": "upper_left",
-    "y_jTotal": "jTotal.y",
-    "y_jTotal_label": "All-in Provisioned J per Total Token (J/tok)",
-    "y_jTotal_labelZh": "每总 token 全电源配置能耗（J/tok）",
-    "y_jTotal_title": "All-in Provisioned Joules per Total Token",
-    "y_jTotal_titleZh": "每总 token 全电源配置焦耳能耗",
-    "y_jTotal_roofline": "lower_right",
-    "y_jOutput": "jOutput.y",
-    "y_jOutput_label": "All-in Provisioned J per Output Token (J/tok)",
-    "y_jOutput_labelZh": "每输出 token 全电源配置能耗（J/tok）",
-    "y_jOutput_title": "All-in Provisioned Joules per Output Token",
-    "y_jOutput_titleZh": "每输出 token 全电源配置焦耳能耗",
-    "y_jOutput_roofline": "lower_right",
-    "y_jInput": "jInput.y",
-    "y_jInput_label": "All-in Provisioned J per Input Token (J/tok)",
-    "y_jInput_labelZh": "每输入 token 全电源配置能耗（J/tok）",
-    "y_jInput_title": "All-in Provisioned Joules per Input Token",
-    "y_jInput_titleZh": "每输入 token 全电源配置焦耳能耗",
-    "y_jInput_roofline": "lower_right",
-    "y_measuredAvgPower": "measuredAvgPower.y",
-    "y_measuredAvgPower_label": "Measured Avg Power per Chip (W)",
-    "y_measuredAvgPower_labelZh": "每 Chip 实测平均功耗（W）",
-    "y_measuredAvgPower_title": "Measured Average Power per Chip",
-    "y_measuredAvgPower_titleZh": "每 Chip 实测平均功耗",
-    "y_measuredAvgPower_roofline": "lower_right",
-    "y_measuredPrefillAvgPower": "measuredPrefillAvgPower.y",
-    "y_measuredPrefillAvgPower_label": "Measured Prefill Power per Chip (W)",
-    "y_measuredPrefillAvgPower_labelZh": "每 Chip 实测 Prefill 功耗（W）",
-    "y_measuredPrefillAvgPower_title": "Measured Prefill Power per Chip",
-    "y_measuredPrefillAvgPower_titleZh": "每 Chip 实测 Prefill 功耗",
-    "y_measuredPrefillAvgPower_roofline": "lower_right",
-    "y_measuredDecodeAvgPower": "measuredDecodeAvgPower.y",
-    "y_measuredDecodeAvgPower_label": "Measured Decode Power per Chip (W)",
-    "y_measuredDecodeAvgPower_labelZh": "每 Chip 实测 Decode 功耗（W）",
-    "y_measuredDecodeAvgPower_title": "Measured Decode Power per Chip",
-    "y_measuredDecodeAvgPower_titleZh": "每 Chip 实测 Decode 功耗",
-    "y_measuredDecodeAvgPower_roofline": "lower_right",
-    "y_measuredJPerOutputToken": "measuredJPerOutputToken.y",
-    "y_measuredJPerOutputToken_label": "Measured J per Output Token (J/tok)",
-    "y_measuredJPerOutputToken_labelZh": "每输出 token 实测能耗（J/tok）",
-    "y_measuredJPerOutputToken_title": "Measured Joules per Output Token",
-    "y_measuredJPerOutputToken_titleZh": "每输出 token 实测焦耳能耗",
-    "y_measuredJPerOutputToken_roofline": "lower_right",
-    "y_measuredJPerInputToken": "measuredJPerInputToken.y",
-    "y_measuredJPerInputToken_label": "Measured J per Input Token (J/tok)",
-    "y_measuredJPerInputToken_labelZh": "每输入 token 实测能耗（J/tok）",
-    "y_measuredJPerInputToken_title": "Measured Joules per Input Token",
-    "y_measuredJPerInputToken_titleZh": "每输入 token 实测焦耳能耗",
-    "y_measuredJPerInputToken_roofline": "lower_right",
-    "y_measuredJPerTotalToken": "measuredJPerTotalToken.y",
-    "y_measuredJPerTotalToken_label": "Measured J per Token (J/tok)",
-    "y_measuredJPerTotalToken_labelZh": "每 token 实测能耗（J/tok）",
-    "y_measuredJPerTotalToken_title": "Measured Joules per Token (incl. prompt)",
-    "y_measuredJPerTotalToken_titleZh": "每 token 实测焦耳能耗（含提示词）",
-    "y_measuredJPerTotalToken_roofline": "lower_right",
-    "y_cost_limit": 5,
-    "y_latency_limit": 60
-  },
-  {
-    "chartType": "e2e",
-    "heading": "vs. End-to-end Latency",
-    "x": "median_e2el",
-    "x_label": "End-to-end Latency (s)",
-    "y": "tput_per_gpu",
-    "y_tpPerGpu": "tpPerGpu.y",
-    "y_tpPerGpu_label": "Token Throughput per Chip (tok/s/chip)",
-    "y_tpPerGpu_labelZh": "每 Chip token 吞吐量（tok/s/chip）",
-    "y_tpPerGpu_title": "Token Throughput per Chip",
-    "y_tpPerGpu_titleZh": "每 Chip token 吞吐量",
-    "y_tpPerGpu_roofline": "upper_right",
-    "y_inputTputPerGpu": "inputTputPerGpu.y",
-    "y_inputTputPerGpu_label": "Input Token Throughput per Chip (tok/s/chip)",
-    "y_inputTputPerGpu_labelZh": "每 Chip 输入 token 吞吐量（tok/s/chip）",
-    "y_inputTputPerGpu_title": "Input Token Throughput per Chip",
-    "y_inputTputPerGpu_titleZh": "每 Chip 输入 token 吞吐量",
-    "y_inputTputPerGpu_roofline": "upper_right",
-    "y_inputTputPerGpu_x": "p90_ttft",
-    "y_inputTputPerGpu_x_label": "P90 Time To First Token (s)",
-    "y_outputTputPerGpu": "outputTputPerGpu.y",
-    "y_outputTputPerGpu_label": "Output Token Throughput per Chip (tok/s/chip)",
-    "y_outputTputPerGpu_labelZh": "每 Chip 输出 token 吞吐量（tok/s/chip）",
-    "y_outputTputPerGpu_title": "Output Token Throughput per Chip",
-    "y_outputTputPerGpu_titleZh": "每 Chip 输出 token 吞吐量",
-    "y_outputTputPerGpu_roofline": "upper_right",
-    "y_tpPerMw": "tpPerMw.y",
-    "y_tpPerMw_label": "Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_tpPerMw_labelZh": "每全电源配置兆瓦 token 吞吐量（tok/s/MW）",
-    "y_tpPerMw_title": "Token Throughput per All in Utility MW",
-    "y_tpPerMw_titleZh": "每全电源配置兆瓦 token 吞吐量",
-    "y_tpPerMw_roofline": "upper_right",
-    "y_inputTputPerMw": "inputTputPerMw.y",
-    "y_inputTputPerMw_label": "Input Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_inputTputPerMw_labelZh": "每全电源配置兆瓦输入 token 吞吐量（tok/s/MW）",
-    "y_inputTputPerMw_title": "Input Token Throughput per All in Utility MW",
-    "y_inputTputPerMw_titleZh": "每全电源配置兆瓦输入 token 吞吐量",
-    "y_inputTputPerMw_roofline": "upper_right",
-    "y_outputTputPerMw": "outputTputPerMw.y",
-    "y_outputTputPerMw_label": "Output Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_outputTputPerMw_labelZh": "每全电源配置兆瓦输出 token 吞吐量（tok/s/MW）",
-    "y_outputTputPerMw_title": "Output Token Throughput per All in Utility MW",
-    "y_outputTputPerMw_titleZh": "每全电源配置兆瓦输出 token 吞吐量",
-    "y_outputTputPerMw_roofline": "upper_right",
-    "y_costh": "costh.y",
-    "y_costh_label": "Cost per Million Total Tokens ($)",
-    "y_costh_labelZh": "每百万总 token 成本（$）",
-    "y_costh_title": "Cost per Million Total Tokens (Owning - Hyperscaler)",
-    "y_costh_titleZh": "每百万总 token 成本（自有 - 超大规模）",
-    "y_costh_roofline": "lower_left",
-    "y_costn": "costn.y",
-    "y_costn_label": "Cost per Million Total Tokens ($)",
-    "y_costn_labelZh": "每百万总 token 成本（$）",
-    "y_costn_title": "Cost per Million Total Tokens (Owning - Neocloud Giant)",
-    "y_costn_titleZh": "每百万总 token 成本（自有 - Neocloud Giant）",
-    "y_costn_roofline": "lower_left",
-    "y_costr": "costr.y",
-    "y_costr_label": "Cost per Million Total Tokens ($)",
-    "y_costr_labelZh": "每百万总 token 成本（$）",
-    "y_costr_title": "Cost per Million Total Tokens (3 Year Rental)",
-    "y_costr_titleZh": "每百万总 token 成本（3 年租赁）",
-    "y_costr_roofline": "lower_left",
-    "y_costhOutput": "costhOutput.y",
-    "y_costhOutput_label": "Cost per Million Output Tokens ($)",
-    "y_costhOutput_labelZh": "每百万输出 token 成本（$）",
-    "y_costhOutput_title": "Cost per Million Output Tokens (Owning - Hyperscaler)",
-    "y_costhOutput_titleZh": "每百万输出 token 成本（自有 - 超大规模）",
-    "y_costhOutput_roofline": "lower_left",
-    "y_costnOutput": "costnOutput.y",
-    "y_costnOutput_label": "Cost per Million Output Tokens ($)",
-    "y_costnOutput_labelZh": "每百万输出 token 成本（$）",
-    "y_costnOutput_title": "Cost per Million Output Tokens (Owning - Neocloud Giant)",
-    "y_costnOutput_titleZh": "每百万输出 token 成本（自有 - Neocloud Giant）",
-    "y_costnOutput_roofline": "lower_left",
-    "y_costrOutput": "costrOutput.y",
-    "y_costrOutput_label": "Cost per Million Output Tokens ($)",
-    "y_costrOutput_labelZh": "每百万输出 token 成本（$）",
-    "y_costrOutput_title": "Cost per Million Output Tokens (3 Year Rental)",
-    "y_costrOutput_titleZh": "每百万输出 token 成本（3 年租赁）",
-    "y_costrOutput_roofline": "lower_left",
-    "y_costhi": "costhi.y",
-    "y_costhi_label": "Cost per Million Input Tokens ($)",
-    "y_costhi_labelZh": "每百万输入 token 成本（$）",
-    "y_costhi_title": "Cost per Million Input Tokens (Owning - Hyperscaler)",
-    "y_costhi_titleZh": "每百万输入 token 成本（自有 - 超大规模）",
-    "y_costhi_roofline": "lower_left",
-    "y_costni": "costni.y",
-    "y_costni_label": "Cost per Million Input Tokens ($)",
-    "y_costni_labelZh": "每百万输入 token 成本（$）",
-    "y_costni_title": "Cost per Million Input Tokens (Owning - Neocloud Giant)",
-    "y_costni_titleZh": "每百万输入 token 成本（自有 - Neocloud Giant）",
-    "y_costni_roofline": "lower_left",
-    "y_costri": "costri.y",
-    "y_costri_label": "Cost per Million Input Tokens ($)",
-    "y_costri_labelZh": "每百万输入 token 成本（$）",
-    "y_costri_title": "Cost per Million Input Tokens (3 Year Rental)",
-    "y_costri_titleZh": "每百万输入 token 成本（3 年租赁）",
-    "y_costri_roofline": "lower_left",
-    "y_costUser": "costUser.y",
-    "y_costUser_label": "Cost per Million Total Tokens ($)",
-    "y_costUser_labelZh": "每百万总 token 成本（$）",
-    "y_costUser_title": "Cost per Million Total Tokens (Custom User Values)",
-    "y_costUser_titleZh": "每百万总 token 成本（自定义值）",
-    "y_costUser_roofline": "lower_left",
-    "y_powerUser": "powerUser.y",
-    "y_powerUser_label": "Token Throughput per All in Utility MW (tok/s/MW)",
-    "y_powerUser_labelZh": "每全电源配置兆瓦 token 吞吐量（tok/s/MW）",
-    "y_powerUser_title": "Token Throughput per All in Utility MW (Custom User Values)",
-    "y_powerUser_titleZh": "每全电源配置兆瓦 token 吞吐量（自定义值）",
-    "y_powerUser_roofline": "upper_right",
-    "y_jTotal": "jTotal.y",
-    "y_jTotal_label": "All-in Provisioned J per Total Token (J/tok)",
-    "y_jTotal_labelZh": "每总 token 全电源配置能耗（J/tok）",
-    "y_jTotal_title": "All-in Provisioned Joules per Total Token",
-    "y_jTotal_titleZh": "每总 token 全电源配置焦耳能耗",
-    "y_jTotal_roofline": "lower_left",
-    "y_jOutput": "jOutput.y",
-    "y_jOutput_label": "All-in Provisioned J per Output Token (J/tok)",
-    "y_jOutput_labelZh": "每输出 token 全电源配置能耗（J/tok）",
-    "y_jOutput_title": "All-in Provisioned Joules per Output Token",
-    "y_jOutput_titleZh": "每输出 token 全电源配置焦耳能耗",
-    "y_jOutput_roofline": "lower_left",
-    "y_jInput": "jInput.y",
-    "y_jInput_label": "All-in Provisioned J per Input Token (J/tok)",
-    "y_jInput_labelZh": "每输入 token 全电源配置能耗（J/tok）",
-    "y_jInput_title": "All-in Provisioned Joules per Input Token",
-    "y_jInput_titleZh": "每输入 token 全电源配置焦耳能耗",
-    "y_jInput_roofline": "lower_left",
-    "y_measuredAvgPower": "measuredAvgPower.y",
-    "y_measuredAvgPower_label": "Measured Avg Power per Chip (W)",
-    "y_measuredAvgPower_labelZh": "每 Chip 实测平均功耗（W）",
-    "y_measuredAvgPower_title": "Measured Average Power per Chip",
-    "y_measuredAvgPower_titleZh": "每 Chip 实测平均功耗",
-    "y_measuredAvgPower_roofline": "lower_left",
-    "y_measuredPrefillAvgPower": "measuredPrefillAvgPower.y",
-    "y_measuredPrefillAvgPower_label": "Measured Prefill Power per Chip (W)",
-    "y_measuredPrefillAvgPower_labelZh": "每 Chip 实测 Prefill 功耗（W）",
-    "y_measuredPrefillAvgPower_title": "Measured Prefill Power per Chip",
-    "y_measuredPrefillAvgPower_titleZh": "每 Chip 实测 Prefill 功耗",
-    "y_measuredPrefillAvgPower_roofline": "lower_left",
-    "y_measuredDecodeAvgPower": "measuredDecodeAvgPower.y",
-    "y_measuredDecodeAvgPower_label": "Measured Decode Power per Chip (W)",
-    "y_measuredDecodeAvgPower_labelZh": "每 Chip 实测 Decode 功耗（W）",
-    "y_measuredDecodeAvgPower_title": "Measured Decode Power per Chip",
-    "y_measuredDecodeAvgPower_titleZh": "每 Chip 实测 Decode 功耗",
-    "y_measuredDecodeAvgPower_roofline": "lower_left",
-    "y_measuredJPerOutputToken": "measuredJPerOutputToken.y",
-    "y_measuredJPerOutputToken_label": "Measured J per Output Token (J/tok)",
-    "y_measuredJPerOutputToken_labelZh": "每输出 token 实测能耗（J/tok）",
-    "y_measuredJPerOutputToken_title": "Measured Joules per Output Token",
-    "y_measuredJPerOutputToken_titleZh": "每输出 token 实测焦耳能耗",
-    "y_measuredJPerOutputToken_roofline": "lower_left",
-    "y_measuredJPerInputToken": "measuredJPerInputToken.y",
-    "y_measuredJPerInputToken_label": "Measured J per Input Token (J/tok)",
-    "y_measuredJPerInputToken_labelZh": "每输入 token 实测能耗（J/tok）",
-    "y_measuredJPerInputToken_title": "Measured Joules per Input Token",
-    "y_measuredJPerInputToken_titleZh": "每输入 token 实测焦耳能耗",
-    "y_measuredJPerInputToken_roofline": "lower_left",
-    "y_measuredJPerTotalToken": "measuredJPerTotalToken.y",
-    "y_measuredJPerTotalToken_label": "Measured J per Token (J/tok)",
-    "y_measuredJPerTotalToken_labelZh": "每 token 实测能耗（J/tok）",
-    "y_measuredJPerTotalToken_title": "Measured Joules per Token (incl. prompt)",
-    "y_measuredJPerTotalToken_titleZh": "每 token 实测焦耳能耗（含提示词）",
-    "y_measuredJPerTotalToken_roofline": "lower_left",
-    "y_cost_limit": 5,
-    "y_latency_limit": 60
-  }
-]
diff --git a/packages/app/src/components/inference/log-viewer/benchmark-log-detail.tsx b/packages/app/src/components/inference/log-viewer/benchmark-log-detail.tsx
new file mode 100644
index 000000000..bd9261fb1
--- /dev/null
+++ b/packages/app/src/components/inference/log-viewer/benchmark-log-detail.tsx
@@ -0,0 +1,68 @@
+'use client';
+
+import { ArrowLeft } from 'lucide-react';
+import Link from 'next/link';
+import { usePathname, useRouter } from 'next/navigation';
+
+import { ServerLogViewer } from '@/components/inference/agentic-point/server-log-viewer';
+import { track } from '@/lib/analytics';
+import { isZhPathname, ZH_PREFIX } from '@/lib/i18n';
+import { useLocale } from '@/lib/use-locale';
+
+const STRINGS = {
+  en: {
+    back: 'Back',
+    inferenceChart: 'Inference chart',
+    eyebrow: 'Fixed-sequence benchmark',
+    title: 'Benchmark logs',
+  },
+  zh: {
+    back: '返回',
+    inferenceChart: '推理图表',
+    eyebrow: '固定序列长度基准测试',
+    title: '基准测试日志',
+  },
+} as const;
+
+export function BenchmarkLogDetail({ id }: { id: number }) {
+  const router = useRouter();
+  const pathname = usePathname();
+  const t = STRINGS[useLocale()];
+  const inferenceHref = isZhPathname(pathname) ? `${ZH_PREFIX}/inference` : '/inference';
+
+  return (
+    <main className="container mx-auto flex flex-col gap-4 px-4 py-6 lg:px-8">
+      <nav className="flex items-center gap-2" aria-label={t.title}>
+        <button
+          type="button"
+          onClick={() => {
+            track('inference_fixed_seq_logs_back_clicked', { id });
+            router.back();
+          }}
+          className="inline-flex items-center gap-1 text-sm text-muted-foreground hover:text-foreground"
+        >
+          <ArrowLeft className="size-4" aria-hidden="true" /> {t.back}
+        </button>
+        <span className="text-sm text-muted-foreground" aria-hidden="true">
+          ·
+        </span>
+        <Link
+          href={inferenceHref}
+          className="text-sm text-muted-foreground hover:text-foreground"
+          onClick={() => track('inference_fixed_seq_logs_chart_clicked', { id })}
+        >
+          {t.inferenceChart}
+        </Link>
+      </nav>
+
+      <header>
+        <p className="font-mono text-xs uppercase tracking-wider text-muted-foreground">
+          {t.eyebrow} · #{id}
+        </p>
+        <h1 className="mt-1 text-xl font-semibold text-foreground">{t.title}</h1>
+      </header>
+
+      <ServerLogViewer id={id} enabled analyticsContext="fixed-sequence" />
+    </main>
+  );
+}
diff --git a/packages/app/src/components/inference/log-viewer/fixed-sequence-log-dialog.tsx b/packages/app/src/components/inference/log-viewer/fixed-sequence-log-dialog.tsx
new file mode 100644
index 000000000..e0a77e074
--- /dev/null
+++ b/packages/app/src/components/inference/log-viewer/fixed-sequence-log-dialog.tsx
@@ -0,0 +1,48 @@
+'use client';
+
+import {
+  Dialog,
+  DialogContent,
+  DialogDescription,
+  DialogHeader,
+  DialogTitle,
+} from '@/components/ui/dialog';
+import { ServerLogViewer } from '@/components/inference/agentic-point/server-log-viewer';
+import { useLocale } from '@/lib/use-locale';
+
+const STRINGS = {
+  en: {
+    title: 'Benchmark logs',
+    description: 'Stored runtime logs for this fixed-sequence benchmark point.',
+  },
+  zh: {
+    title: '基准测试日志',
+    description: '该固定序列长度基准测试数据点的已存储运行时日志。',
+  },
+} as const;
+
+interface Props {
+  pointId: number | null;
+  onOpenChange: (open: boolean) => void;
+}
+
+export function FixedSequenceLogDialog({ pointId, onOpenChange }: Props) {
+  const t = STRINGS[useLocale()];
+
+  return (
+    <Dialog open={pointId !== null} onOpenChange={onOpenChange}>
+      <DialogContent
+        className="flex max-h-[94vh] w-[min(96vw,80rem)] max-w-[min(96vw,80rem)] flex-col gap-3 overflow-y-auto p-3 sm:p-5"
+        data-testid="fixed-sequence-log-dialog"
+      >
+        <DialogHeader className="sr-only">
+          <DialogTitle>{t.title}</DialogTitle>
+          <DialogDescription>{t.description}</DialogDescription>
+        </DialogHeader>
+        {pointId === null ? null : (
+          <ServerLogViewer id={pointId} enabled analyticsContext="fixed-sequence" />
+        )}
+      </DialogContent>
+    </Dialog>
+  );
+}
diff --git a/packages/app/src/components/inference/measured-power-direction.test.ts b/packages/app/src/components/inference/measured-power-direction.test.ts
index ae307e431..a8762e110 100644
--- a/packages/app/src/components/inference/measured-power-direction.test.ts
+++ b/packages/app/src/components/inference/measured-power-direction.test.ts
@@ -1,6 +1,6 @@
 import { describe, it, expect } from 'vitest';
 
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
+import chartDefinitions from '@/components/inference/metric-registry';
 import type { ChartDefinition, InferenceData } from '@/components/inference/types';
 import { sortRowsByYMetric } from '@/components/inference/ui/inference-table-sort';
 import {
@@ -30,6 +30,11 @@ const MEASURED_POWER_METRICS = [
   'y_measuredDecodeAvgPower',
 ] as const;
 
+const QUERY_ENERGY_METRICS = [
+  'y_measuredJPerSuccessfulQuery',
+  'y_measuredWhPerSuccessfulQuery',
+] as const;
+
 const defs = chartDefinitions as unknown as ChartDefinition[];
 const interactivityDef = defs.find((d) => d.chartType === 'interactivity')!;
 const e2eDef = defs.find((d) => d.chartType === 'e2e')!;
@@ -133,4 +138,27 @@ describe('measured-power Pareto direction', () => {
       }
     }
   });
+
+  it.each(QUERY_ENERGY_METRICS)('%s is bilingual and lower-is-better', (metric) => {
+    expect(declaredDirection(interactivityDef, metric)).toBe('lower_right');
+    expect(declaredDirection(e2eDef, metric)).toBe('lower_left');
+    for (const chartDef of [interactivityDef, e2eDef]) {
+      expect(chartDef[metric]).toMatch(/\.y$/u);
+      expect(chartDef[`${metric}_label`]).toBeTruthy();
+      expect(chartDef[`${metric}_labelZh`]).toBeTruthy();
+    }
+  });
+
+  it('leaves %TDP without a Pareto direction on either block', () => {
+    // %TDP is a utilization gauge, not an efficiency frontier: a config running
+    // hotter is not "worse" along an axis the roofline can order, so declaring a
+    // corner would draw a frontier with no meaning. The axis still ships as a
+    // plottable, bilingual metric — it just never anchors a roofline.
+    for (const chartDef of [interactivityDef, e2eDef]) {
+      expect(chartDef.y_measuredPowerPercentTdp).toMatch(/\.y$/u);
+      expect(chartDef['y_measuredPowerPercentTdp_label']).toBeTruthy();
+      expect(chartDef['y_measuredPowerPercentTdp_labelZh']).toBeTruthy();
+      expect(declaredDirection(chartDef, 'y_measuredPowerPercentTdp')).toBeUndefined();
+    }
+  });
 });
diff --git a/packages/app/src/components/inference/metric-registry.test.ts b/packages/app/src/components/inference/metric-registry.test.ts
new file mode 100644
index 000000000..0fd7f0171
--- /dev/null
+++ b/packages/app/src/components/inference/metric-registry.test.ts
@@ -0,0 +1,81 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  chartDefinitions,
+  DEFAULT_METRIC_CONFIG_KEY,
+  isBenchmarkMetricKey,
+  METRIC_CONFIG_KEYS,
+  METRIC_CONTROL_GROUPS,
+  METRIC_REGISTRY,
+  resolveMetricConfigKey,
+  tokenMetricTypeForConfigKey,
+} from './metric-registry';
+
+describe('metric registry', () => {
+  it('derives chart-specific roofline corners from one polarity owner', () => {
+    const [interactivity, e2e] = chartDefinitions;
+    expect(interactivity.chartType).toBe('interactivity');
+    expect(e2e.chartType).toBe('e2e');
+    expect(interactivity.y_tpPerGpu_roofline).toBe('upper_left');
+    expect(e2e.y_tpPerGpu_roofline).toBe('upper_right');
+    expect(interactivity.y_tokenRevenuePerGpuHour_roofline).toBe('upper_left');
+    expect(e2e.y_tokenRevenuePerGpuHour_roofline).toBe('upper_right');
+    expect(interactivity.y_costh_roofline).toBe('lower_right');
+    expect(e2e.y_costh_roofline).toBe('lower_left');
+    expect(interactivity.y_measuredPowerPercentTdp_roofline).toBeUndefined();
+    expect(e2e.y_measuredPowerPercentTdp_roofline).toBeUndefined();
+  });
+
+  it('preserves metric-specific x overrides and bilingual labels', () => {
+    const interactivity = chartDefinitions[0];
+
+    expect(interactivity.y_inputTputPerGpu_x).toBe('p90_ttft');
+    expect(interactivity.y_inputTputPerGpu_heading).toBe('vs. P90 Time To First Token');
+    expect(interactivity.y_inputTputPerGpu_labelZh).toBe(METRIC_REGISTRY.inputTputPerGpu.labelZh);
+  });
+
+  it('lists every canonical metric exactly once across controls', () => {
+    const controlMetrics = METRIC_CONTROL_GROUPS.flatMap((group) => group.metrics);
+
+    expect(new Set(controlMetrics).size).toBe(controlMetrics.length);
+    expect(controlMetrics.toSorted()).toEqual(METRIC_CONFIG_KEYS.toSorted());
+  });
+});
+
+describe('metric compatibility', () => {
+  it('maps the legacy base y field to canonical throughput', () => {
+    expect(resolveMetricConfigKey('y')).toBe('y_tpPerGpu');
+  });
+
+  it('keeps the legacy fallback independent from the dashboard default', () => {
+    expect(resolveMetricConfigKey(undefined, 'y')).toBe('y_tpPerGpu');
+    expect(DEFAULT_METRIC_CONFIG_KEY).toBe('y_tokensPerDollarN');
+  });
+
+  it('falls back safely for unknown persisted metric values', () => {
+    expect(resolveMetricConfigKey('y_removedMetric')).toBe(DEFAULT_METRIC_CONFIG_KEY);
+    expect(resolveMetricConfigKey('arbitrary')).toBe(DEFAULT_METRIC_CONFIG_KEY);
+    expect(isBenchmarkMetricKey('removedMetric')).toBe(false);
+  });
+
+  it('preserves valid benchmark, derived, and custom metric identities', () => {
+    expect(resolveMetricConfigKey('y_tpPerGpu')).toBe('y_tpPerGpu');
+    expect(resolveMetricConfigKey('y_measuredJPerSuccessfulQuery')).toBe(
+      'y_measuredJPerSuccessfulQuery',
+    );
+    expect(resolveMetricConfigKey('y_costUser')).toBe('y_costUser');
+    expect(isBenchmarkMetricKey('tpPerGpu')).toBe(true);
+    expect(isBenchmarkMetricKey('tokenRevenuePerGpuHour')).toBe(true);
+    expect(isBenchmarkMetricKey('measuredJPerSuccessfulQuery')).toBe(true);
+    expect(isBenchmarkMetricKey('costUser')).toBe(false);
+  });
+
+  it('classifies output, input, and total token metric options', () => {
+    expect(tokenMetricTypeForConfigKey('y_outputTputPerGpu')).toBe('output');
+    expect(tokenMetricTypeForConfigKey('y_jOutput')).toBe('output');
+    expect(tokenMetricTypeForConfigKey('y_costhi')).toBe('input');
+    expect(tokenMetricTypeForConfigKey('y_tpPerGpu')).toBe('total');
+    expect(tokenMetricTypeForConfigKey('y_tokenRevenuePerGpuHour')).toBe('total');
+    expect(tokenMetricTypeForConfigKey('y_measuredAvgPower')).toBe('total');
+  });
+});
diff --git a/packages/app/src/components/inference/metric-registry.ts b/packages/app/src/components/inference/metric-registry.ts
new file mode 100644
index 000000000..88bb1b16d
--- /dev/null
+++ b/packages/app/src/components/inference/metric-registry.ts
@@ -0,0 +1,624 @@
+import type { ChartDefinition } from './types';
+import type { TokenMetricType } from '@/lib/supplemental-benchmarks';
+
+export type RooflineDirection = 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
+
+export interface MetricDefinition {
+  field: `${string}.y`;
+  label: string;
+  labelZh: string;
+  title: string;
+  titleZh: string;
+  polarity?: 'higher' | 'lower';
+  x?: string;
+  source?: 'custom';
+  xLabel?: string;
+  heading?: string;
+}
+
+export const METRIC_REGISTRY = {
+  tpPerGpu: {
+    field: 'tpPerGpu.y',
+    label: 'Token Throughput per Chip (tok/s/chip)',
+    labelZh: '每芯片 token 吞吐量（tok/s/chip）',
+    title: 'Token Throughput per Chip',
+    titleZh: '每芯片 token 吞吐量',
+    polarity: 'higher',
+  },
+  inputTputPerGpu: {
+    field: 'inputTputPerGpu.y',
+    label: 'Input Token Throughput per Chip (tok/s/chip)',
+    labelZh: '每芯片输入 token 吞吐量（tok/s/chip）',
+    title: 'Input Token Throughput per Chip',
+    titleZh: '每芯片输入 token 吞吐量',
+    polarity: 'higher',
+    x: 'p90_ttft',
+    xLabel: 'P90 Time To First Token (s)',
+    heading: 'vs. P90 Time To First Token',
+  },
+  outputTputPerGpu: {
+    field: 'outputTputPerGpu.y',
+    label: 'Output Token Throughput per Chip (tok/s/chip)',
+    labelZh: '每芯片输出 token 吞吐量（tok/s/chip）',
+    title: 'Output Token Throughput per Chip',
+    titleZh: '每芯片输出 token 吞吐量',
+    polarity: 'higher',
+  },
+  tokenRevenuePerGpuHour: {
+    field: 'tokenRevenuePerGpuHour.y',
+    label: 'Token Revenue per GPU Hour ($/GPU/hr)',
+    labelZh: '每 GPU 小时 token 收入（$/GPU/hr）',
+    title: 'Token Revenue per GPU Hour',
+    titleZh: '每 GPU 小时 token 收入',
+    polarity: 'higher',
+  },
+  tpPerMw: {
+    field: 'tpPerMw.y',
+    label: 'Token Throughput per All in Utility MW (tok/s/MW)',
+    labelZh: '每全电源配置兆瓦 token 吞吐量（tok/s/MW）',
+    title: 'Token Throughput per All in Utility MW',
+    titleZh: '每全电源配置兆瓦 token 吞吐量',
+    polarity: 'higher',
+  },
+  inputTputPerMw: {
+    field: 'inputTputPerMw.y',
+    label: 'Input Token Throughput per All in Utility MW (tok/s/MW)',
+    labelZh: '每全电源配置兆瓦输入 token 吞吐量（tok/s/MW）',
+    title: 'Input Token Throughput per All in Utility MW',
+    titleZh: '每全电源配置兆瓦输入 token 吞吐量',
+    polarity: 'higher',
+  },
+  outputTputPerMw: {
+    field: 'outputTputPerMw.y',
+    label: 'Output Token Throughput per All in Utility MW (tok/s/MW)',
+    labelZh: '每全电源配置兆瓦输出 token 吞吐量（tok/s/MW）',
+    title: 'Output Token Throughput per All in Utility MW',
+    titleZh: '每全电源配置兆瓦输出 token 吞吐量',
+    polarity: 'higher',
+  },
+  costh: {
+    field: 'costh.y',
+    label: 'Cost per Million Total Tokens ($)',
+    labelZh: '每百万总 token 成本（$）',
+    title: 'Cost per Million Total Tokens (Owning - Hyperscaler)',
+    titleZh: '每百万总 token 成本（自有 - 超大规模）',
+    polarity: 'lower',
+  },
+  costn: {
+    field: 'costn.y',
+    label: 'Cost per Million Total Tokens ($)',
+    labelZh: '每百万总 token 成本（$）',
+    title: 'Cost per Million Total Tokens (Owning - Neocloud Giant)',
+    titleZh: '每百万总 token 成本（自有 - Neocloud Giant）',
+    polarity: 'lower',
+  },
+  costr: {
+    field: 'costr.y',
+    label: 'Cost per Million Total Tokens ($)',
+    labelZh: '每百万总 token 成本（$）',
+    title: 'Cost per Million Total Tokens (3 Year Rental)',
+    titleZh: '每百万总 token 成本（3 年租赁）',
+    polarity: 'lower',
+  },
+  costhOutput: {
+    field: 'costhOutput.y',
+    label: 'Cost per Million Output Tokens ($)',
+    labelZh: '每百万输出 token 成本（$）',
+    title: 'Cost per Million Output Tokens (Owning - Hyperscaler)',
+    titleZh: '每百万输出 token 成本（自有 - 超大规模）',
+    polarity: 'lower',
+  },
+  costnOutput: {
+    field: 'costnOutput.y',
+    label: 'Cost per Million Output Tokens ($)',
+    labelZh: '每百万输出 token 成本（$）',
+    title: 'Cost per Million Output Tokens (Owning - Neocloud Giant)',
+    titleZh: '每百万输出 token 成本（自有 - Neocloud Giant）',
+    polarity: 'lower',
+  },
+  costrOutput: {
+    field: 'costrOutput.y',
+    label: 'Cost per Million Output Tokens ($)',
+    labelZh: '每百万输出 token 成本（$）',
+    title: 'Cost per Million Output Tokens (3 Year Rental)',
+    titleZh: '每百万输出 token 成本（3 年租赁）',
+    polarity: 'lower',
+  },
+  costhi: {
+    field: 'costhi.y',
+    label: 'Cost per Million Input Tokens ($)',
+    labelZh: '每百万输入 token 成本（$）',
+    title: 'Cost per Million Input Tokens (Owning - Hyperscaler)',
+    titleZh: '每百万输入 token 成本（自有 - 超大规模）',
+    polarity: 'lower',
+  },
+  costni: {
+    field: 'costni.y',
+    label: 'Cost per Million Input Tokens ($)',
+    labelZh: '每百万输入 token 成本（$）',
+    title: 'Cost per Million Input Tokens (Owning - Neocloud Giant)',
+    titleZh: '每百万输入 token 成本（自有 - Neocloud Giant）',
+    polarity: 'lower',
+  },
+  costri: {
+    field: 'costri.y',
+    label: 'Cost per Million Input Tokens ($)',
+    labelZh: '每百万输入 token 成本（$）',
+    title: 'Cost per Million Input Tokens (3 Year Rental)',
+    titleZh: '每百万输入 token 成本（3 年租赁）',
+    polarity: 'lower',
+  },
+  tokensPerDollarH: {
+    field: 'tokensPerDollarH.y',
+    label: 'Total Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的总 token 数（tok/$）',
+    title: 'Total Tokens per $1 USD (Owning - Hyperscaler)',
+    titleZh: '每 1 美元可购买的总 token 数（自有 - 超大规模）',
+    polarity: 'higher',
+  },
+  tokensPerDollarN: {
+    field: 'tokensPerDollarN.y',
+    label: 'Total Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的总 token 数（tok/$）',
+    title: 'Total Tokens per $1 USD (Owning - Neocloud Giant)',
+    titleZh: '每 1 美元可购买的总 token 数（自有 - Neocloud Giant）',
+    polarity: 'higher',
+  },
+  tokensPerDollarR: {
+    field: 'tokensPerDollarR.y',
+    label: 'Total Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的总 token 数（tok/$）',
+    title: 'Total Tokens per $1 USD (3 Year Rental)',
+    titleZh: '每 1 美元可购买的总 token 数（3 年租赁）',
+    polarity: 'higher',
+  },
+  outputTokensPerDollarH: {
+    field: 'outputTokensPerDollarH.y',
+    label: 'Output Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的输出 token 数（tok/$）',
+    title: 'Output Tokens per $1 USD (Owning - Hyperscaler)',
+    titleZh: '每 1 美元可购买的输出 token 数（自有 - 超大规模）',
+    polarity: 'higher',
+  },
+  outputTokensPerDollarN: {
+    field: 'outputTokensPerDollarN.y',
+    label: 'Output Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的输出 token 数（tok/$）',
+    title: 'Output Tokens per $1 USD (Owning - Neocloud Giant)',
+    titleZh: '每 1 美元可购买的输出 token 数（自有 - Neocloud Giant）',
+    polarity: 'higher',
+  },
+  outputTokensPerDollarR: {
+    field: 'outputTokensPerDollarR.y',
+    label: 'Output Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的输出 token 数（tok/$）',
+    title: 'Output Tokens per $1 USD (3 Year Rental)',
+    titleZh: '每 1 美元可购买的输出 token 数（3 年租赁）',
+    polarity: 'higher',
+  },
+  inputTokensPerDollarH: {
+    field: 'inputTokensPerDollarH.y',
+    label: 'Input Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的输入 token 数（tok/$）',
+    title: 'Input Tokens per $1 USD (Owning - Hyperscaler)',
+    titleZh: '每 1 美元可购买的输入 token 数（自有 - 超大规模）',
+    polarity: 'higher',
+  },
+  inputTokensPerDollarN: {
+    field: 'inputTokensPerDollarN.y',
+    label: 'Input Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的输入 token 数（tok/$）',
+    title: 'Input Tokens per $1 USD (Owning - Neocloud Giant)',
+    titleZh: '每 1 美元可购买的输入 token 数（自有 - Neocloud Giant）',
+    polarity: 'higher',
+  },
+  inputTokensPerDollarR: {
+    field: 'inputTokensPerDollarR.y',
+    label: 'Input Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的输入 token 数（tok/$）',
+    title: 'Input Tokens per $1 USD (3 Year Rental)',
+    titleZh: '每 1 美元可购买的输入 token 数（3 年租赁）',
+    polarity: 'higher',
+  },
+  tokensPerRmbH: {
+    field: 'tokensPerRmbH.y',
+    label: 'Total Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的总 token 数（tok/¥）',
+    title: 'Total Tokens per ¥1 RMB (Owning - Hyperscaler)',
+    titleZh: '每 1 元人民币可购买的总 token 数（自有 - 超大规模）',
+    polarity: 'higher',
+  },
+  tokensPerRmbN: {
+    field: 'tokensPerRmbN.y',
+    label: 'Total Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的总 token 数（tok/¥）',
+    title: 'Total Tokens per ¥1 RMB (Owning - Neocloud Giant)',
+    titleZh: '每 1 元人民币可购买的总 token 数（自有 - Neocloud Giant）',
+    polarity: 'higher',
+  },
+  tokensPerRmbR: {
+    field: 'tokensPerRmbR.y',
+    label: 'Total Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的总 token 数（tok/¥）',
+    title: 'Total Tokens per ¥1 RMB (3 Year Rental)',
+    titleZh: '每 1 元人民币可购买的总 token 数（3 年租赁）',
+    polarity: 'higher',
+  },
+  outputTokensPerRmbH: {
+    field: 'outputTokensPerRmbH.y',
+    label: 'Output Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的输出 token 数（tok/¥）',
+    title: 'Output Tokens per ¥1 RMB (Owning - Hyperscaler)',
+    titleZh: '每 1 元人民币可购买的输出 token 数（自有 - 超大规模）',
+    polarity: 'higher',
+  },
+  outputTokensPerRmbN: {
+    field: 'outputTokensPerRmbN.y',
+    label: 'Output Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的输出 token 数（tok/¥）',
+    title: 'Output Tokens per ¥1 RMB (Owning - Neocloud Giant)',
+    titleZh: '每 1 元人民币可购买的输出 token 数（自有 - Neocloud Giant）',
+    polarity: 'higher',
+  },
+  outputTokensPerRmbR: {
+    field: 'outputTokensPerRmbR.y',
+    label: 'Output Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的输出 token 数（tok/¥）',
+    title: 'Output Tokens per ¥1 RMB (3 Year Rental)',
+    titleZh: '每 1 元人民币可购买的输出 token 数（3 年租赁）',
+    polarity: 'higher',
+  },
+  inputTokensPerRmbH: {
+    field: 'inputTokensPerRmbH.y',
+    label: 'Input Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的输入 token 数（tok/¥）',
+    title: 'Input Tokens per ¥1 RMB (Owning - Hyperscaler)',
+    titleZh: '每 1 元人民币可购买的输入 token 数（自有 - 超大规模）',
+    polarity: 'higher',
+  },
+  inputTokensPerRmbN: {
+    field: 'inputTokensPerRmbN.y',
+    label: 'Input Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的输入 token 数（tok/¥）',
+    title: 'Input Tokens per ¥1 RMB (Owning - Neocloud Giant)',
+    titleZh: '每 1 元人民币可购买的输入 token 数（自有 - Neocloud Giant）',
+    polarity: 'higher',
+  },
+  inputTokensPerRmbR: {
+    field: 'inputTokensPerRmbR.y',
+    label: 'Input Tokens per ¥1 RMB (tok/¥)',
+    labelZh: '每 1 元人民币可购买的输入 token 数（tok/¥）',
+    title: 'Input Tokens per ¥1 RMB (3 Year Rental)',
+    titleZh: '每 1 元人民币可购买的输入 token 数（3 年租赁）',
+    polarity: 'higher',
+  },
+  costUser: {
+    field: 'costUser.y',
+    label: 'Cost per Million Total Tokens ($)',
+    labelZh: '每百万总 token 成本（$）',
+    title: 'Cost per Million Total Tokens (Custom User Values)',
+    titleZh: '每百万总 token 成本（自定义值）',
+    polarity: 'lower',
+    source: 'custom',
+  },
+  tokensPerDollarUser: {
+    field: 'tokensPerDollarUser.y',
+    label: 'Total Tokens per $1 USD (tok/$)',
+    labelZh: '每 1 美元可购买的总 token 数（tok/$）',
+    title: 'Total Tokens per $1 USD (Custom User Values)',
+    titleZh: '每 1 美元可购买的总 token 数（自定义值）',
+    polarity: 'higher',
+    source: 'custom',
+  },
+  powerUser: {
+    field: 'powerUser.y',
+    label: 'Token Throughput per All in Utility MW (tok/s/MW)',
+    labelZh: '每全电源配置兆瓦 token 吞吐量（tok/s/MW）',
+    title: 'Token Throughput per All in Utility MW (Custom User Values)',
+    titleZh: '每全电源配置兆瓦 token 吞吐量（自定义值）',
+    polarity: 'higher',
+    source: 'custom',
+  },
+  jTotal: {
+    field: 'jTotal.y',
+    label: 'All-in Provisioned J per Total Token (J/tok)',
+    labelZh: '每总 token 全电源配置能耗（J/tok）',
+    title: 'All-in Provisioned Joules per Total Token',
+    titleZh: '每总 token 全电源配置焦耳能耗',
+    polarity: 'lower',
+  },
+  jOutput: {
+    field: 'jOutput.y',
+    label: 'All-in Provisioned J per Output Token (J/tok)',
+    labelZh: '每输出 token 全电源配置能耗（J/tok）',
+    title: 'All-in Provisioned Joules per Output Token',
+    titleZh: '每输出 token 全电源配置焦耳能耗',
+    polarity: 'lower',
+  },
+  jInput: {
+    field: 'jInput.y',
+    label: 'All-in Provisioned J per Input Token (J/tok)',
+    labelZh: '每输入 token 全电源配置能耗（J/tok）',
+    title: 'All-in Provisioned Joules per Input Token',
+    titleZh: '每输入 token 全电源配置焦耳能耗',
+    polarity: 'lower',
+  },
+  measuredAvgPower: {
+    field: 'measuredAvgPower.y',
+    label: 'Measured Avg Power per Chip (W)',
+    labelZh: '每芯片实测平均功耗（W）',
+    title: 'Measured Average Power per Chip',
+    titleZh: '每芯片实测平均功耗',
+    polarity: 'lower',
+  },
+  measuredPrefillAvgPower: {
+    field: 'measuredPrefillAvgPower.y',
+    label: 'Measured Prefill Power per Chip (W)',
+    labelZh: '每芯片实测 Prefill 功耗（W）',
+    title: 'Measured Prefill Power per Chip',
+    titleZh: '每芯片实测 Prefill 功耗',
+    polarity: 'lower',
+  },
+  measuredDecodeAvgPower: {
+    field: 'measuredDecodeAvgPower.y',
+    label: 'Measured Decode Power per Chip (W)',
+    labelZh: '每芯片实测 Decode 功耗（W）',
+    title: 'Measured Decode Power per Chip',
+    titleZh: '每芯片实测 Decode 功耗',
+    polarity: 'lower',
+  },
+  measuredJPerOutputToken: {
+    field: 'measuredJPerOutputToken.y',
+    label: 'Measured J per Output Token (J/tok)',
+    labelZh: '每输出 token 实测能耗（J/tok）',
+    title: 'Measured Joules per Output Token',
+    titleZh: '每输出 token 实测焦耳能耗',
+    polarity: 'lower',
+  },
+  measuredJPerInputToken: {
+    field: 'measuredJPerInputToken.y',
+    label: 'Measured J per Input Token (J/tok)',
+    labelZh: '每输入 token 实测能耗（J/tok）',
+    title: 'Measured Joules per Input Token',
+    titleZh: '每输入 token 实测焦耳能耗',
+    polarity: 'lower',
+  },
+  measuredJPerTotalToken: {
+    field: 'measuredJPerTotalToken.y',
+    label: 'Measured J per Token (J/tok)',
+    labelZh: '每 token 实测能耗（J/tok）',
+    title: 'Measured Joules per Token (incl. prompt)',
+    titleZh: '每 token 实测焦耳能耗（含提示词）',
+    polarity: 'lower',
+  },
+  measuredJPerSuccessfulQuery: {
+    field: 'measuredJPerSuccessfulQuery.y',
+    label: 'Measured J per Successful Query (J/query)',
+    labelZh: '每次成功请求实测能耗（J/query）',
+    title: 'Measured Joules per Successful Query',
+    titleZh: '每次成功请求实测焦耳能耗',
+    polarity: 'lower',
+  },
+  measuredWhPerSuccessfulQuery: {
+    field: 'measuredWhPerSuccessfulQuery.y',
+    label: 'Measured Wh per Successful Query (Wh/query)',
+    labelZh: '每次成功请求实测能耗（Wh/query）',
+    title: 'Measured Watt-hours per Successful Query',
+    titleZh: '每次成功请求实测瓦时能耗',
+    polarity: 'lower',
+  },
+  measuredPowerPercentTdp: {
+    field: 'measuredPowerPercentTdp.y',
+    label: 'Measured Average Power (% TDP)',
+    labelZh: '实测平均功耗（TDP 占比）',
+    title: 'Measured Average Power as Percent of TDP',
+    titleZh: '实测平均功耗占 TDP 百分比',
+  },
+} as const satisfies Record<string, MetricDefinition>;
+
+export type MetricKey = keyof typeof METRIC_REGISTRY;
+export type MetricConfigKey = `y_${MetricKey}`;
+export type CustomMetricKey = {
+  [Key in MetricKey]: (typeof METRIC_REGISTRY)[Key] extends { source: 'custom' } ? Key : never;
+}[MetricKey];
+export type BenchmarkMetricKey = Exclude<MetricKey, CustomMetricKey>;
+export type BenchmarkMetricConfigKey = `y_${BenchmarkMetricKey}`;
+
+export const DEFAULT_METRIC_CONFIG_KEY = 'y_tokensPerDollarN' satisfies MetricConfigKey;
+
+export function isMetricKey(metricKey: string): metricKey is MetricKey {
+  return Object.hasOwn(METRIC_REGISTRY, metricKey);
+}
+
+export function isBenchmarkMetricKey(metricKey: string): metricKey is BenchmarkMetricKey {
+  return isMetricKey(metricKey) && !('source' in METRIC_REGISTRY[metricKey]);
+}
+
+/** Token basis represented by a y-axis option. Non-output metrics deliberately
+ * resolve to total/input so output-only snapshots cannot leak into them. */
+export function tokenMetricTypeForConfigKey(metric: string): TokenMetricType {
+  const normalized = metric.toLowerCase();
+  if (normalized.includes('output')) return 'output';
+  if (normalized.includes('input') || /cost[hnr]i$/u.test(normalized)) return 'input';
+  return 'total';
+}
+
+/**
+ * Resolve persisted metric state to a canonical config key.
+ *
+ * `y` was the original chart field for total throughput and remains a
+ * read-only share-link alias. Unknown persisted values fall back rather than
+ * escaping the registry and reaching chart/data lookups unchecked.
+ */
+export function resolveMetricConfigKey(
+  metricConfigKey: string | null | undefined,
+  fallback?: string,
+): MetricConfigKey {
+  if (metricConfigKey === 'y') return 'y_tpPerGpu';
+  if (metricConfigKey?.startsWith('y_')) {
+    const metricKey = metricConfigKey.slice(2);
+    if (isMetricKey(metricKey)) return metricConfigKey as MetricConfigKey;
+  }
+
+  if (fallback === 'y') return 'y_tpPerGpu';
+  if (fallback?.startsWith('y_')) {
+    const fallbackMetricKey = fallback.slice(2);
+    if (isMetricKey(fallbackMetricKey)) return fallback as MetricConfigKey;
+  }
+
+  return DEFAULT_METRIC_CONFIG_KEY;
+}
+
+export const METRIC_CONFIG_KEYS = Object.keys(METRIC_REGISTRY).map(
+  (key) => `y_${key}` as MetricConfigKey,
+);
+export const BENCHMARK_METRIC_CONFIG_KEYS = METRIC_CONFIG_KEYS.filter(
+  (configKey): configKey is BenchmarkMetricConfigKey => isBenchmarkMetricKey(configKey.slice(2)),
+);
+
+export interface MetricControlGroup {
+  label: string;
+  labelZh: string;
+  metrics: readonly MetricConfigKey[];
+  gated?: boolean;
+}
+
+export const METRIC_CONTROL_GROUPS: readonly MetricControlGroup[] = [
+  {
+    label: 'Throughput',
+    labelZh: '吞吐量',
+    metrics: [
+      'y_tpPerGpu',
+      'y_inputTputPerGpu',
+      'y_outputTputPerGpu',
+      'y_tpPerMw',
+      'y_inputTputPerMw',
+      'y_outputTputPerMw',
+    ],
+  },
+  {
+    label: 'Token Revenue per GPU Hour',
+    labelZh: '每 GPU 小时 token 收入',
+    metrics: ['y_tokenRevenuePerGpuHour'],
+  },
+  {
+    label: 'Total Tokens per $1 USD',
+    labelZh: '每 1 美元可购买的总 token 数',
+    metrics: ['y_tokensPerDollarH', 'y_tokensPerDollarN', 'y_tokensPerDollarR'],
+  },
+  {
+    label: 'Total Tokens per ¥1 CNY',
+    labelZh: '每 1 元人民币可购买的总 token 数',
+    metrics: ['y_tokensPerRmbH', 'y_tokensPerRmbN', 'y_tokensPerRmbR'],
+  },
+  {
+    label: 'Output Tokens per $1 USD',
+    labelZh: '每 1 美元可购买的输出 token 数',
+    metrics: ['y_outputTokensPerDollarH', 'y_outputTokensPerDollarN', 'y_outputTokensPerDollarR'],
+  },
+  {
+    label: 'Output Tokens per ¥1 CNY',
+    labelZh: '每 1 元人民币可购买的输出 token 数',
+    metrics: ['y_outputTokensPerRmbH', 'y_outputTokensPerRmbN', 'y_outputTokensPerRmbR'],
+  },
+  {
+    label: 'Input Tokens per $1 USD',
+    labelZh: '每 1 美元可购买的输入 token 数',
+    metrics: ['y_inputTokensPerDollarH', 'y_inputTokensPerDollarN', 'y_inputTokensPerDollarR'],
+  },
+  {
+    label: 'Input Tokens per ¥1 CNY',
+    labelZh: '每 1 元人民币可购买的输入 token 数',
+    metrics: ['y_inputTokensPerRmbH', 'y_inputTokensPerRmbN', 'y_inputTokensPerRmbR'],
+  },
+  {
+    label: 'Cost per Million Total Tokens',
+    labelZh: '每百万总 token 成本',
+    metrics: ['y_costh', 'y_costn', 'y_costr'],
+  },
+  {
+    label: 'Cost per Million Output Tokens',
+    labelZh: '每百万输出 token 成本',
+    metrics: ['y_costhOutput', 'y_costnOutput', 'y_costrOutput'],
+  },
+  {
+    label: 'Cost per Million Input Tokens',
+    labelZh: '每百万输入 token 成本',
+    metrics: ['y_costhi', 'y_costni', 'y_costri'],
+  },
+  {
+    label: 'All-in Provisioned Energy per Token',
+    labelZh: '每 token 全电源配置能耗',
+    metrics: ['y_jTotal', 'y_jOutput', 'y_jInput'],
+  },
+  {
+    label: 'Measured Energy',
+    labelZh: '实测能耗',
+    metrics: [
+      'y_measuredPrefillAvgPower',
+      'y_measuredDecodeAvgPower',
+      'y_measuredAvgPower',
+      'y_measuredJPerInputToken',
+      'y_measuredJPerOutputToken',
+      'y_measuredJPerTotalToken',
+      'y_measuredJPerSuccessfulQuery',
+      'y_measuredWhPerSuccessfulQuery',
+      'y_measuredPowerPercentTdp',
+    ],
+  },
+  {
+    label: 'Custom User Values',
+    labelZh: '自定义值',
+    metrics: ['y_tokensPerDollarUser', 'y_costUser', 'y_powerUser'],
+  },
+];
+
+function rooflineDirection(
+  chartType: 'interactivity' | 'e2e',
+  polarity: 'higher' | 'lower',
+): RooflineDirection {
+  if (chartType === 'interactivity') return polarity === 'higher' ? 'upper_left' : 'lower_right';
+  return polarity === 'higher' ? 'upper_right' : 'lower_left';
+}
+
+function buildChartDefinition(chartType: 'interactivity' | 'e2e'): ChartDefinition {
+  const definition: ChartDefinition = {
+    chartType,
+    heading: chartType === 'interactivity' ? 'vs. Interactivity' : 'vs. End-to-end Latency',
+    x: chartType === 'interactivity' ? 'median_intvty' : 'median_e2el',
+    x_label:
+      chartType === 'interactivity' ? 'Interactivity (tok/s/user)' : 'End-to-end Latency (s)',
+    y: 'tput_per_gpu',
+    y_cost_limit: 5,
+    y_latency_limit: 60,
+  };
+
+  for (const [key, metric] of Object.entries(METRIC_REGISTRY) as [
+    MetricKey,
+    (typeof METRIC_REGISTRY)[MetricKey],
+  ][]) {
+    const configKey = `y_${key}`;
+    definition[configKey] = metric.field;
+    definition[`${configKey}_label`] = metric.label;
+    definition[`${configKey}_labelZh`] = metric.labelZh;
+    definition[`${configKey}_title`] = metric.title;
+    definition[`${configKey}_titleZh`] = metric.titleZh;
+    if ('polarity' in metric && metric.polarity) {
+      definition[`${configKey}_roofline`] = rooflineDirection(chartType, metric.polarity);
+    }
+    if ('x' in metric) definition[`${configKey}_x`] = metric.x;
+    if ('xLabel' in metric) definition[`${configKey}_x_label`] = metric.xLabel;
+    if ('heading' in metric && chartType === 'interactivity') {
+      definition[`${configKey}_heading`] = metric.heading;
+    }
+  }
+
+  return definition;
+}
+
+export const chartDefinitions = [
+  buildChartDefinition('interactivity'),
+  buildChartDefinition('e2e'),
+] satisfies ChartDefinition[];
+
+export default chartDefinitions;
diff --git a/packages/app/src/components/inference/replay/ReplayPanel.tsx b/packages/app/src/components/inference/replay/ReplayPanel.tsx
index 3d36bca0f..8638bffdf 100644
--- a/packages/app/src/components/inference/replay/ReplayPanel.tsx
+++ b/packages/app/src/components/inference/replay/ReplayPanel.tsx
@@ -2,11 +2,19 @@
 
 import { Pause, Play, RotateCcw, Video } from 'lucide-react';
 import { flushSync } from 'react-dom';
-import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
+import {
+  useCallback,
+  useEffect,
+  useMemo,
+  useReducer,
+  useRef,
+  useState,
+  type SetStateAction,
+} from 'react';
 
 import { sequenceToIslOsl } from '@semianalysisai/inferencex-constants';
 
-import { useInference } from '@/components/inference/InferenceContext';
+import { useInferenceDisplay, useInferenceFilters } from '@/components/inference/InferenceContext';
 import ScatterGraph from '@/components/inference/ui/ScatterGraph';
 import type { ChartDefinition } from '@/components/inference/types';
 import { Button } from '@/components/ui/button';
@@ -24,7 +32,12 @@ import { track } from '@/lib/analytics';
 import { Sequence } from '@/lib/data-mappings';
 import { cn } from '@/lib/utils';
 
-import { buildReplayTimeline, computeFullRunDomain } from './buildReplayTimeline';
+import {
+  buildReplayTimeline,
+  computeFullRunDomain,
+  type ReplayTimeline,
+  type StepDomain,
+} from './buildReplayTimeline';
 import type { Mp4ExportError, Mp4ExportStage } from './exportMp4';
 import { buildFrameData, dateAtFraction, shouldCommitFraction, spanMs } from './replayFrameData';
 import { useReducedMotion } from './useReducedMotion';
@@ -51,6 +64,36 @@ interface ReplayPanelProps {
 const SPEED_OPTIONS: readonly number[] = [0.25, 0.5, 0.75, 1, 1.25, 1.5, 1.75, 2];
 const REPLAY_BODY_MIN_HEIGHT = 480;
 
+interface ReplayPlaybackState {
+  timeline: ReplayTimeline | null;
+  fraction: number;
+  playing: boolean;
+}
+
+type ReplayPlaybackAction =
+  | { type: 'resetTimeline'; timeline: ReplayTimeline | null }
+  | { type: 'setFraction'; value: SetStateAction<number> }
+  | { type: 'setPlaying'; value: SetStateAction<boolean> };
+
+function replayPlaybackReducer(
+  state: ReplayPlaybackState,
+  action: ReplayPlaybackAction,
+): ReplayPlaybackState {
+  if (action.type === 'resetTimeline') {
+    return { timeline: action.timeline, fraction: 0, playing: false };
+  }
+  if (action.type === 'setFraction') {
+    return {
+      ...state,
+      fraction: typeof action.value === 'function' ? action.value(state.fraction) : action.value,
+    };
+  }
+  return {
+    ...state,
+    playing: typeof action.value === 'function' ? action.value(state.playing) : action.value,
+  };
+}
+
 /**
  * Replay panel that drives the actual `<ScatterGraph>` with interpolated frame
  * data per tick. React re-renders every frame; ScatterGraph's `transitionDuration`
@@ -65,38 +108,32 @@ export default function ReplayPanel({
   yLabel,
   xLabel,
 }: ReplayPanelProps) {
-  const inference = useInference();
-  const { selectedModel, selectedSequence, activeHwTypes } = inference;
+  const { selectedModel, selectedSequence, selectedPrecisions, activeHwTypes } =
+    useInferenceFilters();
+  const { selectedE2eXAxisMetric, selectedXAxisMetric, selectedYAxisMetric } =
+    useInferenceDisplay();
 
   const { isl = 0, osl = 0 } = sequenceToIslOsl(selectedSequence) ?? {};
   const history = useBenchmarkHistory(
     selectedModel,
     isl,
     osl,
-    selectedSequence === Sequence.AgenticTraces ? 'agentic_traces' : undefined,
+    selectedSequence === Sequence.AgenticTraces ? { benchmarkType: 'agentic_traces' } : undefined,
   );
 
   const effectiveX =
-    chartDefinition.chartType === 'e2e'
-      ? inference.selectedE2eXAxisMetric
-      : inference.selectedXAxisMetric;
+    chartDefinition.chartType === 'e2e' ? selectedE2eXAxisMetric : selectedXAxisMetric;
 
   const timeline = useMemo(() => {
     if (!history.data) return null;
     return buildReplayTimeline(
       history.data,
       chartDefinition,
-      inference.selectedYAxisMetric,
+      selectedYAxisMetric,
       effectiveX ?? null,
-      inference.selectedPrecisions,
+      selectedPrecisions,
     );
-  }, [
-    history.data,
-    chartDefinition,
-    inference.selectedYAxisMetric,
-    effectiveX,
-    inference.selectedPrecisions,
-  ]);
+  }, [history.data, chartDefinition, selectedYAxisMetric, effectiveX, selectedPrecisions]);
 
   // Fixed axes for the whole run: take the extent across every step (not just
   // the current frame) for the active hardware, so the axes stay put and the
@@ -107,6 +144,37 @@ export default function ReplayPanel({
     [timeline, activeHwTypes],
   );
 
+  return (
+    <ReplayPanelContent
+      parentChartId={parentChartId}
+      chartDefinition={chartDefinition}
+      yLabel={yLabel}
+      xLabel={xLabel}
+      timeline={timeline}
+      fixedExtent={fixedExtent}
+      selectedModel={selectedModel}
+      isLoading={history.isLoading}
+    />
+  );
+}
+
+interface ReplayPanelContentProps extends ReplayPanelProps {
+  timeline: ReplayTimeline | null;
+  fixedExtent: StepDomain | null;
+  selectedModel: string;
+  isLoading: boolean;
+}
+
+function ReplayPanelContent({
+  parentChartId,
+  chartDefinition,
+  yLabel,
+  xLabel,
+  timeline,
+  fixedExtent,
+  selectedModel,
+  isLoading,
+}: ReplayPanelContentProps) {
   // Track the SVG's position inside our relative wrapper so the date overlay
   // can anchor its bottom-right to the chart plot's top-right (the wrapper
   // also contains the legend, so we can't anchor to the wrapper edge).
@@ -175,8 +243,24 @@ export default function ReplayPanel({
 
   const panelRef = useRef<HTMLDivElement>(null);
 
-  const [fraction, setFraction] = useState(0);
-  const [playing, setPlaying] = useState(false);
+  const [playback, dispatchPlayback] = useReducer(replayPlaybackReducer, {
+    timeline,
+    fraction: 0,
+    playing: false,
+  });
+  let { fraction, playing } = playback;
+  const timelineChanged = playback.timeline !== timeline;
+  if (timelineChanged) {
+    fraction = 0;
+    playing = false;
+    dispatchPlayback({ type: 'resetTimeline', timeline });
+  }
+  const setFraction = useCallback((value: SetStateAction<number>) => {
+    dispatchPlayback({ type: 'setFraction', value });
+  }, []);
+  const setPlaying = useCallback((value: SetStateAction<boolean>) => {
+    dispatchPlayback({ type: 'setPlaying', value });
+  }, []);
   const [speed, setSpeed] = useState(1);
   // Fixed axes (default) freeze the coordinate space to the whole run so the
   // frontier visibly expands over time; turning this off lets the axes refit to
@@ -212,7 +296,8 @@ export default function ReplayPanel({
   // so the predicate compares like-with-like — comparing against the
   // React-committed value lags by a frame and would no-op a backward scrub
   // that crosses a quantum boundary.
-  const fractionRef = useRef(0);
+  const fractionRef = useRef(fraction);
+  if (timelineChanged) fractionRef.current = 0;
   const commitFraction = useCallback((next: number, opts?: { force?: boolean }) => {
     const clamped = next < 0 ? 0 : Math.min(1, next);
     const prev = fractionRef.current;
@@ -274,12 +359,6 @@ export default function ReplayPanel({
     };
   }, [playing, timeline, prefersReducedMotion]);
 
-  useEffect(() => {
-    fractionRef.current = 0;
-    setFraction(0);
-    setPlaying(false);
-  }, [timeline]);
-
   const frameData = useMemo(
     () => (timeline ? buildFrameData(timeline, fraction) : []),
     [timeline, fraction],
@@ -464,7 +543,7 @@ export default function ReplayPanel({
     }
   }, [chartDefinition.chartType, parentChartId, selectedModel, timeline, hasWebCodecs]);
 
-  if (history.isLoading || !timeline) {
+  if (isLoading || !timeline) {
     return (
       <div
         className="p-4 sm:p-6 flex flex-col"
diff --git a/packages/app/src/components/inference/replay/__tests__/buildReplayTimeline.test.ts b/packages/app/src/components/inference/replay/__tests__/buildReplayTimeline.test.ts
index 6dd9aac94..ff9b11d00 100644
--- a/packages/app/src/components/inference/replay/__tests__/buildReplayTimeline.test.ts
+++ b/packages/app/src/components/inference/replay/__tests__/buildReplayTimeline.test.ts
@@ -216,6 +216,31 @@ describe('buildReplayTimeline', () => {
     expect(t.configs.length).toBeGreaterThanOrEqual(2);
   });
 
+  it('keeps same-coordinate recipe variants as separate replay timelines', () => {
+    const rows = [
+      baseRow({
+        recipe_fingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+        metrics: { tput_per_gpu: 400, median_intvty: 20 },
+      }),
+      baseRow({
+        recipe_fingerprint: 'bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb',
+        metrics: { tput_per_gpu: 500, median_intvty: 25 },
+      }),
+    ];
+
+    const timeline = buildReplayTimeline(rows, interactivityChartDef, 'y_tpPerGpu', null, ['fp4']);
+
+    expect(timeline.dates).toEqual(['2025-01-01']);
+    expect(timeline.configs).toHaveLength(2);
+    expect(timeline.configs.map(({ configId }) => configId).toSorted()).toEqual([
+      'h100_trt|fp4|0|32|0|0|0|recipe-aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+      'h100_trt|fp4|0|32|0|0|0|recipe-bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb',
+    ]);
+    expect(timeline.configs.map(({ stepValues }) => stepValues[0].y).toSorted()).toEqual([
+      400, 500,
+    ]);
+  });
+
   it('keeps overlapping agentic MTP and standard-decoding replay points distinct', () => {
     const rows = [
       baseRow({
diff --git a/packages/app/src/components/inference/replay/buildReplayTimeline.ts b/packages/app/src/components/inference/replay/buildReplayTimeline.ts
index 717b12840..e397bf550 100644
--- a/packages/app/src/components/inference/replay/buildReplayTimeline.ts
+++ b/packages/app/src/components/inference/replay/buildReplayTimeline.ts
@@ -86,6 +86,7 @@ export function computeFullRunDomain(
 const buildReplayPointConfigId = (point: InferenceData): string => {
   let key = `${point.hwKey}|${point.precision}|${point.tp}|${point.conc}|${point.decode_ep ?? 0}|${point.prefill_tp ?? 0}|${point.prefill_ep ?? 0}`;
   if (point.disagg) key += `|disagg|${point.num_prefill_gpu ?? 0}|${point.num_decode_gpu ?? 0}`;
+  if (point.recipe_fingerprint) key += `|recipe-${point.recipe_fingerprint}`;
   // Preserve the pre-existing replay identity for fixed-sequence points. Agentic
   // curves need only the decode-method suffix because their hwKey now merges it.
   return key + agenticSpecDecodingKeySuffix(point);
diff --git a/packages/app/src/components/inference/replay/useReducedMotion.ts b/packages/app/src/components/inference/replay/useReducedMotion.ts
index adb12439a..9de80fbaa 100644
--- a/packages/app/src/components/inference/replay/useReducedMotion.ts
+++ b/packages/app/src/components/inference/replay/useReducedMotion.ts
@@ -1,16 +1,28 @@
-import { useEffect, useState } from 'react';
+import { useSyncExternalStore } from 'react';
 
 const QUERY = '(prefers-reduced-motion: reduce)';
 
+const NOOP = () => {};
+
+function subscribe(onStoreChange: () => void): () => void {
+  if (typeof window === 'undefined' || typeof window.matchMedia !== 'function') return NOOP;
+  const mediaQuery = window.matchMedia(QUERY);
+  mediaQuery.addEventListener('change', onStoreChange);
+  return () => mediaQuery.removeEventListener('change', onStoreChange);
+}
+
+function getSnapshot(): boolean {
+  return (
+    typeof window !== 'undefined' &&
+    typeof window.matchMedia === 'function' &&
+    window.matchMedia(QUERY).matches
+  );
+}
+
+function getServerSnapshot(): boolean {
+  return false;
+}
+
 export function useReducedMotion(): boolean {
-  const [reduced, setReduced] = useState(false);
-  useEffect(() => {
-    if (typeof window === 'undefined' || typeof window.matchMedia !== 'function') return;
-    const mq = window.matchMedia(QUERY);
-    setReduced(mq.matches);
-    const onChange = (e: MediaQueryListEvent) => setReduced(e.matches);
-    mq.addEventListener('change', onChange);
-    return () => mq.removeEventListener('change', onChange);
-  }, []);
-  return reduced;
+  return useSyncExternalStore(subscribe, getSnapshot, getServerSnapshot);
 }
diff --git a/packages/app/src/components/inference/token-revenue.test.ts b/packages/app/src/components/inference/token-revenue.test.ts
new file mode 100644
index 000000000..019ef11d2
--- /dev/null
+++ b/packages/app/src/components/inference/token-revenue.test.ts
@@ -0,0 +1,144 @@
+import { describe, expect, it } from 'vitest';
+
+import type { InferenceData, TokenRevenuePricing } from './types';
+import {
+  applyTokenRevenuePricing,
+  formatTokenPrice,
+  inputTokenShareForRevenue,
+  NORMALIZED_TOKEN_REVENUE_PRICING,
+  tokenRevenuePerGpuHour,
+} from './token-revenue';
+
+function point(overrides: Partial<InferenceData> = {}): InferenceData {
+  return {
+    x: 20,
+    y: 2_000,
+    hwKey: 'b200',
+    precision: 'fp8',
+    tp: 8,
+    conc: 16,
+    date: '2026-08-27',
+    tput_per_gpu: 2_000,
+    input_tput_per_gpu: 1_600,
+    output_tput_per_gpu: 400,
+    tpPerGpu: { y: 2_000, roof: false },
+    tpPerMw: { y: 1_000, roof: false },
+    costh: { y: 1, roof: false },
+    costn: { y: 1, roof: false },
+    costr: { y: 1, roof: false },
+    costhi: { y: 1, roof: false },
+    costni: { y: 1, roof: false },
+    costri: { y: 1, roof: false },
+    ...overrides,
+  } as InferenceData;
+}
+
+const openRouterPricing: TokenRevenuePricing = {
+  source: 'openrouter',
+  inputPerMillion: 2,
+  outputPerMillion: 10,
+  openRouterModelId: 'example/model',
+};
+
+describe('token revenue', () => {
+  it('formats subtitle prices without unnecessary trailing zeroes', () => {
+    expect(formatTokenPrice(1)).toBe('1');
+    expect(formatTokenPrice(1.122)).toBe('1.122');
+    expect(formatTokenPrice(0.000001)).toBe('0.000001');
+  });
+
+  it('keeps the normalized $1/M axis equal to million total tokens per GPU hour', () => {
+    expect(tokenRevenuePerGpuHour(point(), NORMALIZED_TOKEN_REVENUE_PRICING)).toBe(7.2);
+  });
+
+  it('prices compatible aggregate input and output rates separately', () => {
+    const aggregate = point({
+      tput_per_gpu: 1_000,
+      input_tput_per_gpu: 800,
+      output_tput_per_gpu: 200,
+    });
+    expect(inputTokenShareForRevenue(aggregate)).toBe(0.8);
+    expect(tokenRevenuePerGpuHour(aggregate, openRouterPricing)).toBeCloseTo(12.96, 10);
+  });
+
+  it('uses the structural sequence mix for disaggregated per-role rates', () => {
+    const disaggregated = point({
+      disagg: true,
+      tput_per_gpu: 1_000,
+      input_tput_per_gpu: 1_600,
+      output_tput_per_gpu: 200,
+      isl: 8_192,
+      osl: 1_024,
+    });
+    expect(inputTokenShareForRevenue(disaggregated)).toBeCloseTo(8 / 9, 10);
+    expect(tokenRevenuePerGpuHour(disaggregated, openRouterPricing)).toBeCloseTo(10.4, 10);
+  });
+
+  it('uses measured prompt and generation totals for agentic disaggregated rows', () => {
+    const agentic = point({
+      disagg: true,
+      tput_per_gpu: 1_000,
+      input_tput_per_gpu: 3_000,
+      output_tput_per_gpu: 100,
+      isl: null,
+      osl: null,
+      total_prompt_tokens: 1_300,
+      total_generation_tokens: 10,
+    });
+    expect(inputTokenShareForRevenue(agentic)).toBeCloseTo(130 / 131, 10);
+    expect(tokenRevenuePerGpuHour(agentic, openRouterPricing)).toBeCloseTo(
+      (1_000 * 3_600 * ((130 / 131) * 2 + (1 / 131) * 10)) / 1_000_000,
+      10,
+    );
+  });
+
+  it('rejects incomplete fixed-sequence token mixes', () => {
+    const partialSequence = point({
+      tput_per_gpu: 1_000,
+      input_tput_per_gpu: 1_600,
+      output_tput_per_gpu: 200,
+      isl: 8_192,
+      osl: null,
+      total_prompt_tokens: undefined,
+      total_generation_tokens: undefined,
+    });
+
+    expect(inputTokenShareForRevenue(partialSequence)).toBeNull();
+    expect(tokenRevenuePerGpuHour(partialSequence, openRouterPricing)).toBeNull();
+  });
+
+  it('rejects incomplete agentic token mixes', () => {
+    const partialAgentic = point({
+      tput_per_gpu: 1_000,
+      input_tput_per_gpu: 3_000,
+      output_tput_per_gpu: 100,
+      isl: null,
+      osl: null,
+      total_prompt_tokens: 1_300,
+      total_generation_tokens: undefined,
+    });
+
+    expect(inputTokenShareForRevenue(partialAgentic)).toBeNull();
+    expect(tokenRevenuePerGpuHour(partialAgentic, openRouterPricing)).toBeNull();
+  });
+
+  it('removes the normalized placeholder when OpenRouter pricing is unavailable', () => {
+    const original = point({ tokenRevenuePerGpuHour: { y: 7.2, roof: false } });
+    const [cleared] = applyTokenRevenuePricing([original], null);
+    expect(cleared).not.toHaveProperty('tokenRevenuePerGpuHour');
+    expect(original.tokenRevenuePerGpuHour).toEqual({ y: 7.2, roof: false });
+  });
+
+  it('does not invent OpenRouter revenue when the token mix is unavailable', () => {
+    const unknownMix = point({
+      input_tput_per_gpu: 0,
+      output_tput_per_gpu: 0,
+      isl: null,
+      osl: null,
+      total_prompt_tokens: undefined,
+      total_generation_tokens: undefined,
+    });
+
+    expect(tokenRevenuePerGpuHour(unknownMix, openRouterPricing)).toBeNull();
+  });
+});
diff --git a/packages/app/src/components/inference/token-revenue.ts b/packages/app/src/components/inference/token-revenue.ts
new file mode 100644
index 000000000..01cd4c969
--- /dev/null
+++ b/packages/app/src/components/inference/token-revenue.ts
@@ -0,0 +1,92 @@
+import type { InferenceData, TokenRevenuePricing } from './types';
+
+const SECONDS_PER_HOUR = 3_600;
+const TOKENS_PER_MILLION = 1_000_000;
+const RATE_SUM_RELATIVE_TOLERANCE = 0.01;
+
+export const NORMALIZED_TOKEN_REVENUE_PRICING: TokenRevenuePricing = {
+  source: 'normalized',
+  inputPerMillion: 1,
+  outputPerMillion: 1,
+};
+
+/** Format a $/M tok sale price consistently across revenue UI surfaces. */
+export function formatTokenPrice(value: number): string {
+  return value.toLocaleString('en-US', { maximumFractionDigits: 6 });
+}
+
+/**
+ * Fraction of a point's per-GPU token throughput that is input.
+ *
+ * Aggregate rows report compatible input/output/total per-GPU rates, so their
+ * measured split wins. Disaggregated rows report input per prefill GPU and
+ * output per decode GPU while total throughput uses every GPU; for those rows
+ * the fixed ISL:OSL shape or measured agentic prompt:generation totals supply
+ * the like-for-like split.
+ */
+export function inputTokenShareForRevenue(point: InferenceData): number | null {
+  const total = point.tput_per_gpu ?? 0;
+  const input = point.input_tput_per_gpu ?? 0;
+  const output = point.output_tput_per_gpu ?? 0;
+  const sum = input + output;
+
+  if (total > 0 && sum > 0 && Math.abs(sum / total - 1) <= RATE_SUM_RELATIVE_TOLERANCE) {
+    return input / sum;
+  }
+
+  const { isl, osl } = point;
+  if (typeof isl === 'number' && typeof osl === 'number' && isl + osl > 0) {
+    return isl / (isl + osl);
+  }
+
+  const prompt = point.total_prompt_tokens;
+  const generation = point.total_generation_tokens;
+  if (typeof prompt === 'number' && typeof generation === 'number' && prompt + generation > 0) {
+    return prompt / (prompt + generation);
+  }
+
+  return null;
+}
+
+/** Gross token revenue at one benchmark operating point, in $/GPU/hr. */
+export function tokenRevenuePerGpuHour(
+  point: InferenceData,
+  pricing: TokenRevenuePricing,
+): number | null {
+  const total = point.tput_per_gpu ?? 0;
+  if (!(total > 0)) return 0;
+
+  // Equal input/output prices collapse exactly to total throughput and do not
+  // require token-mix telemetry. This is the normalized $1/M default.
+  if (pricing.inputPerMillion === pricing.outputPerMillion) {
+    return (total * SECONDS_PER_HOUR * pricing.inputPerMillion) / TOKENS_PER_MILLION;
+  }
+
+  const inputShare = inputTokenShareForRevenue(point);
+  if (inputShare === null) return null;
+  const blendedPriceAtPoint =
+    inputShare * pricing.inputPerMillion + (1 - inputShare) * pricing.outputPerMillion;
+  return (total * SECONDS_PER_HOUR * blendedPriceAtPoint) / TOKENS_PER_MILLION;
+}
+
+/**
+ * Recompute the revenue field without mutating transformed benchmark points.
+ * A null price removes the normalized placeholder while OpenRouter is loading
+ * or unavailable, so the chart never presents $1/M revenue as live pricing.
+ */
+export function applyTokenRevenuePricing(
+  points: InferenceData[],
+  pricing: TokenRevenuePricing | null,
+): InferenceData[] {
+  return points.map((point) => {
+    const next = { ...point };
+    delete next.tokenRevenuePerGpuHour;
+    if (!pricing) return next;
+
+    const revenue = tokenRevenuePerGpuHour(point, pricing);
+    if (revenue !== null) {
+      next.tokenRevenuePerGpuHour = { y: revenue, roof: false };
+    }
+    return next;
+  });
+}
diff --git a/packages/app/src/components/inference/types.ts b/packages/app/src/components/inference/types.ts
index 71cb9d86f..0c99174d8 100644
--- a/packages/app/src/components/inference/types.ts
+++ b/packages/app/src/components/inference/types.ts
@@ -1,7 +1,11 @@
 import type React from 'react';
+import type { WorkerPower } from '@semianalysisai/inferencex-db/queries/benchmarks';
 
 import type { HardwareEntry } from '@/lib/constants';
 import type { Model, Sequence } from '@/lib/data-mappings';
+import type { MetricKey } from './metric-registry';
+
+export type { WorkerPower };
 
 /**
  * Role of a single worker process in a multinode / disaggregated deployment.
@@ -17,36 +21,6 @@ import type { Model, Sequence } from '@/lib/data-mappings';
  */
 export type WorkerRole = 'prefill' | 'decode' | 'agg' | 'frontend';
 
-/**
- * Per-worker measured power entry emitted by the runner's aggregate_power.py
- * for multinode and disaggregated runs. The chart layer can use these to
- * surface a stacked breakdown of where energy is spent across worker types.
- *
- * `hosts` lists the node hostnames whose perfmon CSVs were rolled up into
- * this worker entry (a single-node worker has one host; a multinode decode
- * worker spanning 4 nodes has four). Optional because pre-multinode versions
- * of aggregate_power.py didn't emit it.
- *
- * `avg_temp_c`, `peak_temp_c`, `avg_util_pct`, `avg_mem_used_mb` mirror the
- * cluster-wide telemetry scalars and are only present when the perfmon CSVs
- * include the corresponding sample columns. Each is optional so callers can
- * distinguish "field absent from this run" from "field present and equal to 0".
- */
-export interface WorkerPower {
-  // `string` rather than `WorkerRole` so the type lines up with what we get
-  // from the JSONB column without an unsafe cast at every boundary. Chart
-  // code can still narrow on the literal values it understands.
-  role: string;
-  worker_idx: number;
-  hosts?: string[];
-  num_gpus: number;
-  avg_power_w: number;
-  avg_temp_c?: number;
-  peak_temp_c?: number;
-  avg_util_pct?: number;
-  avg_mem_used_mb?: number;
-}
-
 /**
  * Represents an aggregated data entry, typically from a raw data source.
  * This interface contains various performance metrics.
@@ -84,6 +58,8 @@ export interface AggDataEntry {
   rawMetricKeys?: string[];
   /** Stable per-point id from benchmark_results — for trace_replay lookups. */
   id?: number;
+  /** Stable identity for recipe variants that share topology and concurrency. */
+  recipe_fingerprint?: string;
   hw: string;
   mtp?: string;
   hwKey: string;
@@ -137,7 +113,10 @@ export interface AggDataEntry {
   'p99.9_e2el': number;
   // Measured GPU telemetry (emitted by runner's aggregate_power.py).
   // Optional because historical runs predate the fields.
+  power_valid?: number;
+  power_metric_schema_version?: number;
   avg_power_w?: number;
+  joules_per_successful_query?: number;
   joules_per_output_token?: number;
   joules_per_total_token?: number;
   // Multinode / disagg-only measured power. The aggregate_power.py runner
@@ -145,13 +124,13 @@ export interface AggDataEntry {
   // and decode workers (single-node disagg or multinode disagg). Single-node
   // aggregated configs leave these undefined.
   // - prefill_avg_power_w / decode_avg_power_w: mean per-GPU draw (W) within each role
-  // - joules_per_input_token: prefill_energy / total_input_tokens (prefill GPUs only)
-  // The disagg decode-only J/output is carried by joules_per_output_token above
-  // (the runner overrides it to decode_energy / total_output_tokens on disagg) —
-  // there is no separate _decode field.
+  // Unprefixed joules fields are whole-deployment metrics in schema version 2.
+  // Explicit prefill/decode keys retain role-local energy breakdowns.
   prefill_avg_power_w?: number;
   decode_avg_power_w?: number;
   joules_per_input_token?: number;
+  prefill_joules_per_input_token?: number;
+  decode_joules_per_output_token?: number;
   // Cluster-wide GPU telemetry beyond power (temperature, utilization, memory).
   // Emitted by aggregate_power.py when the perfmon CSVs include the matching
   // sample columns. Optional because older runs (and runs without the relevant
@@ -189,6 +168,14 @@ export interface AggDataEntry {
   decode_ep?: number;
   /** Decode-side pipeline parallelism — see {@link AggDataEntry.pp}. */
   decode_pp?: number;
+  /** Prefill worker's decode-context parallel width, when emitted by the runtime. */
+  prefill_dcp_size?: number;
+  /** Decode worker's decode-context parallel width, when emitted by the runtime. */
+  decode_dcp_size?: number;
+  /** Prefill worker's prefill-context parallel width, when emitted by the runtime. */
+  prefill_pcp_size?: number;
+  /** Decode worker's prefill-context parallel width, when emitted by the runtime. */
+  decode_pcp_size?: number;
   decode_dp_attention?: boolean | string;
   decode_num_workers?: number;
   image?: string;
@@ -256,6 +243,7 @@ export interface InferenceData extends Partial<Omit<AggDataEntry, AggDataConflic
   // Chart-specific derived fields
   x: number;
   y: number;
+  roof?: boolean;
   hidden?: boolean;
   /**
    * Whether this point sits on the canonical
@@ -286,9 +274,12 @@ export interface InferenceData extends Partial<Omit<AggDataEntry, AggDataConflic
   tpPerGpu: { y: number; roof: boolean };
   outputTputPerGpu?: { y: number; roof: boolean };
   inputTputPerGpu?: { y: number; roof: boolean };
+  /** Gross token revenue using the selected normalized or OpenRouter prices. */
+  tokenRevenuePerGpuHour?: { y: number; roof: boolean };
   tpPerMw: { y: number; roof: boolean };
   inputTputPerMw?: { y: number; roof: boolean };
   outputTputPerMw?: { y: number; roof: boolean };
+  // Cost per million tokens.
   costh: { y: number; roof: boolean };
   costn: { y: number; roof: boolean };
   costr: { y: number; roof: boolean };
@@ -299,6 +290,27 @@ export interface InferenceData extends Partial<Omit<AggDataEntry, AggDataConflic
   costni: { y: number; roof: boolean };
   costri: { y: number; roof: boolean };
   costUser?: { y: number; roof: boolean };
+  // Tokens purchasable per $1.
+  tokensPerDollarH?: { y: number; roof: boolean };
+  tokensPerDollarN?: { y: number; roof: boolean };
+  tokensPerDollarR?: { y: number; roof: boolean };
+  outputTokensPerDollarH?: { y: number; roof: boolean };
+  outputTokensPerDollarN?: { y: number; roof: boolean };
+  outputTokensPerDollarR?: { y: number; roof: boolean };
+  inputTokensPerDollarH?: { y: number; roof: boolean };
+  inputTokensPerDollarN?: { y: number; roof: boolean };
+  inputTokensPerDollarR?: { y: number; roof: boolean };
+  // Tokens purchasable per ¥1 — the $ metrics converted at USD_TO_CNY.
+  tokensPerRmbH?: { y: number; roof: boolean };
+  tokensPerRmbN?: { y: number; roof: boolean };
+  tokensPerRmbR?: { y: number; roof: boolean };
+  outputTokensPerRmbH?: { y: number; roof: boolean };
+  outputTokensPerRmbN?: { y: number; roof: boolean };
+  outputTokensPerRmbR?: { y: number; roof: boolean };
+  inputTokensPerRmbH?: { y: number; roof: boolean };
+  inputTokensPerRmbN?: { y: number; roof: boolean };
+  inputTokensPerRmbR?: { y: number; roof: boolean };
+  tokensPerDollarUser?: { y: number; roof: boolean };
   powerUser?: { y: number; roof: boolean };
 
   // All-in provisioned Joules per token
@@ -315,6 +327,9 @@ export interface InferenceData extends Partial<Omit<AggDataEntry, AggDataConflic
   measuredJPerOutputToken?: { y: number; roof: boolean };
   measuredJPerTotalToken?: { y: number; roof: boolean };
   measuredJPerInputToken?: { y: number; roof: boolean };
+  measuredJPerSuccessfulQuery?: { y: number; roof: boolean };
+  measuredWhPerSuccessfulQuery?: { y: number; roof: boolean };
+  measuredPowerPercentTdp?: { y: number; roof: boolean };
 }
 
 /** Why a chart-ready point was intentionally excluded from the visible plot. */
@@ -332,33 +347,19 @@ export interface ClippedInferenceData {
 /**
  * Keys of InferenceData that have the roofline metric structure ({y, roof}).
  */
-export type YAxisMetricKey =
-  | 'tpPerGpu'
-  | 'outputTputPerGpu'
-  | 'inputTputPerGpu'
-  | 'tpPerMw'
-  | 'inputTputPerMw'
-  | 'outputTputPerMw'
-  | 'costh'
-  | 'costn'
-  | 'costr'
-  | 'costhOutput'
-  | 'costnOutput'
-  | 'costrOutput'
-  | 'costhi'
-  | 'costni'
-  | 'costri'
-  | 'costUser'
-  | 'powerUser'
-  | 'jTotal'
-  | 'jOutput'
-  | 'jInput'
-  | 'measuredAvgPower'
-  | 'measuredPrefillAvgPower'
-  | 'measuredDecodeAvgPower'
-  | 'measuredJPerOutputToken'
-  | 'measuredJPerTotalToken'
-  | 'measuredJPerInputToken';
+export type YAxisMetricKey = MetricKey;
+
+export type TokenRevenuePriceSource = 'normalized' | 'openrouter';
+
+export interface TokenRevenuePricing {
+  source: TokenRevenuePriceSource;
+  /** Published or assumed input-token sale price, $/M tok. */
+  inputPerMillion: number;
+  /** Published or assumed output-token sale price, $/M tok. */
+  outputPerMillion: number;
+  /** Exact OpenRouter catalog id when `source` is `openrouter`. */
+  openRouterModelId?: string;
+}
 
 /**
  * Defines the configuration and labels for a specific chart.
@@ -381,113 +382,7 @@ export interface ChartDefinition {
   x_label: string;
   y: keyof AggDataEntry;
   y_label?: string;
-  y_tpPerGpu?: string;
-  y_tpPerGpu_label?: string;
-  y_tpPerGpu_title?: string;
-  y_tpPerGpu_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_outputTputPerGpu?: string;
-  y_outputTputPerGpu_label?: string;
-  y_outputTputPerGpu_title?: string;
-  y_outputTputPerGpu_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_inputTputPerGpu?: string;
-  y_inputTputPerGpu_label?: string;
-  y_inputTputPerGpu_title?: string;
-  y_inputTputPerGpu_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_inputTputPerGpu_x?: string;
-  y_inputTputPerGpu_x_label?: string;
-  y_inputTputPerGpu_heading?: string;
-  y_tpPerMw?: string;
-  y_tpPerMw_label?: string;
-  y_tpPerMw_title?: string;
-  y_tpPerMw_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_inputTputPerMw?: string;
-  y_inputTputPerMw_label?: string;
-  y_inputTputPerMw_title?: string;
-  y_inputTputPerMw_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_outputTputPerMw?: string;
-  y_outputTputPerMw_label?: string;
-  y_outputTputPerMw_title?: string;
-  y_outputTputPerMw_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_costh?: string;
-  y_costh_label?: string;
-  y_costh_title?: string;
-  y_costh_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_costn?: string;
-  y_costn_label?: string;
-  y_costn_title?: string;
-  y_costn_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_costr?: string;
-  y_costr_label?: string;
-  y_costr_title?: string;
-  y_costr_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  // Cost per million output tokens
-  y_costhOutput?: string;
-  y_costhOutput_label?: string;
-  y_costhOutput_title?: string;
-  y_costhOutput_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_costnOutput?: string;
-  y_costnOutput_label?: string;
-  y_costnOutput_title?: string;
-  y_costnOutput_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_costrOutput?: string;
-  y_costrOutput_label?: string;
-  y_costrOutput_title?: string;
-  y_costrOutput_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  // Cost per million input tokens
-  y_costhi?: string;
-  y_costhi_label?: string;
-  y_costhi_title?: string;
-  y_costhi_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_costni?: string;
-  y_costni_label?: string;
-  y_costni_title?: string;
-  y_costni_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_costri?: string;
-  y_costri_label?: string;
-  y_costri_title?: string;
-  y_costri_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  // All-in provisioned Joules per token
-  y_jTotal?: string;
-  y_jTotal_label?: string;
-  y_jTotal_title?: string;
-  y_jTotal_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_jOutput?: string;
-  y_jOutput_label?: string;
-  y_jOutput_title?: string;
-  y_jOutput_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_jInput?: string;
-  y_jInput_label?: string;
-  y_jInput_title?: string;
-  y_jInput_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  // Measured power / energy from runner GPU telemetry
-  y_measuredAvgPower?: string;
-  y_measuredAvgPower_label?: string;
-  y_measuredAvgPower_title?: string;
-  // Not explicitly set in the config — ScatterGraph falls back to lower_right
-  // (matches "lower power at the same interactivity is more efficient").
-  // The field stays in the type for parity with the other y_* metrics and
-  // so a future config can override the default.
-  y_measuredAvgPower_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_measuredPrefillAvgPower?: string;
-  y_measuredPrefillAvgPower_label?: string;
-  y_measuredPrefillAvgPower_title?: string;
-  y_measuredPrefillAvgPower_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_measuredDecodeAvgPower?: string;
-  y_measuredDecodeAvgPower_label?: string;
-  y_measuredDecodeAvgPower_title?: string;
-  y_measuredDecodeAvgPower_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_measuredJPerOutputToken?: string;
-  y_measuredJPerOutputToken_label?: string;
-  y_measuredJPerOutputToken_title?: string;
-  y_measuredJPerOutputToken_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_measuredJPerInputToken?: string;
-  y_measuredJPerInputToken_label?: string;
-  y_measuredJPerInputToken_title?: string;
-  y_measuredJPerInputToken_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
-  y_measuredJPerTotalToken?: string;
-  y_measuredJPerTotalToken_label?: string;
-  y_measuredJPerTotalToken_title?: string;
-  y_measuredJPerTotalToken_roofline?: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right';
+
   y_cost_limit?: number;
   y_latency_limit?: number;
 }
@@ -566,8 +461,8 @@ export interface ScatterGraphProps {
   overlayData?: OverlayData;
   /**
    * D3 transition duration in ms used when data or scales change. Defaults to
-   * the regular interactive value (750). The replay panel passes 0 so frames
-   * snap to interpolated positions instead of fighting a 750ms tween.
+   * the regular 300ms interactive value. The replay panel passes 0 so frames
+   * snap to interpolated positions instead of fighting an in-flight tween.
    */
   transitionDuration?: number;
   /**
@@ -629,52 +524,6 @@ export interface LegendItemProps {
   onClick: (key: string) => void;
 }
 
-/**
- * Props for the WorkflowInfoDisplay component.
- * @interface WorkflowInfoDisplayProps
- * @property {string} [runId] - The ID of the workflow run.
- * @property {string} [runUrl] - The URL to the workflow run details.
- * @property {string} [runDate] - The date of the workflow run.
- * @property {string} [runTimezone] - The timezone of the workflow run date.
- */
-export interface WorkflowInfoDisplayProps {
-  runId?: string;
-  runUrl?: string;
-  runDate?: string;
-  runTimezone?: string;
-}
-
-/**
- * Represents the configuration of models, sequences, and precisions.
- * @interface ModelConfig
- * @property {object} [modelName: string] - An object where keys are model names.
- * @property {object} [modelName: string].[sequence: string] - An object where keys are sequence names.
- * @property {string[]} [modelName: string].[sequence: string] - An array of available precisions for the given model and sequence.
- */
-export type ModelConfig = Record<string, Record<string, string[]>>;
-
-/**
- * Represents information about a workflow run.
- * @interface WorkflowInfo
- * @property {string} runInfoBySequence - Object mapping sequence types to their run information.
- * @property {string} run_date - The date when the workflow was run.
- * @property {ModelConfig} modelConfig - Configuration details for models, sequences, and precisions.
- */
-export interface WorkflowInfo {
-  runInfoBySequence: Record<
-    string,
-    {
-      runId: string;
-      runDate: string;
-      runUrl: string;
-      changelog?: ChangelogMetadata;
-    }
-  >;
-  run_date: string;
-  modelConfig: ModelConfig;
-  gpus: HardwareConfig;
-}
-
 /**
  * Represents information about a single workflow run by sequence.
  * @interface RunInfo
@@ -715,142 +564,121 @@ export interface QuickFilters {
  */
 export type AvailableQuickFilters = QuickFilters;
 
-/**
- * Defines the shape of the context object provided by `InferenceChartContext`.
- * @interface InferenceChartContextType
- * @property {Set<string>} activeHwTypes - A set of currently active hardware types for filtering.
- * @property {Set<string>} hwTypesWithData - A set of all hardware types present in the current dataset.
- * @property {(hw: string) => void} toggleHwType - Function to toggle the active state of a hardware type.
- * @property {HardwareConfig} hardwareConfig - The hardware configuration map.
- * @property {RenderableGraph[]} graphs - An array of graphs ready for rendering.
- * @property {string} selectedModel - The currently selected model.
- * @property {(model: string) => void} setSelectedModel - Function to set the selected model.
- * @property {string} selectedSequence - The currently selected sequence.
- * @property {(sequence: string) => void} setSelectedSequence - Function to set the selected sequence.
- * @property {string} selectedPrecision - The currently selected precision.
- * @property {(precision: string) => void} setSelectedPrecision - Function to set the selected precision.
- * @property {boolean} loading - Indicates if data is currently being loaded.
- * @property {string | null} error - Any error message encountered during data loading, or null if no error.
- * @property {WorkflowInfo | null} workflowInfo - Information about the workflow run, or null if not yet loaded.
- */
-export interface InferenceChartContextType {
+/** Fetched and derived benchmark data shared by inference consumers. */
+export interface InferenceDataContextType {
+  hwTypesWithData: Set<string>;
+  hardwareConfig: HardwareConfig;
+  graphs: RenderableGraph[];
+  loading: boolean;
+  error: string | null;
+  availableQuickFilters: AvailableQuickFilters;
+  availableGPUs: { value: string; label: string }[];
+  availableDates: string[];
+  dateRangeAvailableDates: string[];
+  isCheckingAvailableDates: boolean;
+  availableRuns: Record<string, RunInfo>;
+  availablePrecisions: string[];
+  availableSequences: Sequence[];
+  availableModels: string[];
+}
+
+/** Inference-only workflow, filter, and date selection state. */
+export interface InferenceFiltersContextType {
   activeHwTypes: Set<string>;
+  bestPerSku: boolean;
+  selectedModel: Model;
+  selectedSequence: Sequence;
+  selectedPrecisions: string[];
+  quickFilters: QuickFilters;
+  selectedGPUs: string[];
+  selectedDates: string[];
+  selectedDateRange: { startDate: string; endDate: string };
+  activeDates: Set<string>;
+  userCosts: Record<string, number | undefined> | null;
+  selectedRunDate: string;
+  selectedRunId: string;
+  userPowers: Record<string, number | undefined> | null;
+  activePresetId: string | null;
+  presetGuardRef: React.RefObject<boolean>;
+  compareGpuPair: readonly [string, string] | null;
+}
+
+/** Axis choices and visual presentation state. */
+export interface InferenceDisplayContextType {
+  selectedYAxisMetric: string;
+  tokenRevenuePriceSource: TokenRevenuePriceSource;
+  tokenRevenuePricing: TokenRevenuePricing | null;
+  openRouterModelId: string | null;
+  openRouterPricingLoading: boolean;
+  openRouterPricingError: string | null;
+  selectedPercentile: string;
+  selectedXAxisMetric: string | null;
+  selectedE2eXAxisMetric: string | null;
+  selectedXAxisMode: 'ttft' | 'e2e' | 'interactivity' | 'e2e-normalized-interactivity';
+  scaleType: 'auto' | 'linear' | 'log';
+  isLegendExpanded: boolean;
+  hideNonOptimal: boolean;
+  showPointLabels: boolean;
+  highContrast: boolean;
+  logScale: boolean;
+  useAdvancedLabels: boolean;
+  showConcurrencyLabels: boolean;
+  showGradientLabels: boolean;
+  showLineLabels: boolean;
+}
+
+/** Stable commands that mutate inference state. */
+export interface InferenceActionsContextType {
   toggleActiveDate: (date: string) => void;
   removeActiveDate: (date: string) => void;
   selectAllActiveDates: () => void;
-  activeDates: Set<string>;
-  hwTypesWithData: Set<string>;
   toggleHwType: (hw: string) => void;
   removeHwType: (hw: string) => void;
   selectAllHwTypes: () => void;
-  /** Whether clean dashboard loads automatically keep the best configuration per physical SKU. */
-  bestPerSku: boolean;
-  setBestPerSku: (enabled: boolean) => void;
-  /** Resolve automatic official + `overlay:` hardware selections under the active scope rule. */
+  setBestPerSku: (enabled: boolean, options?: { applySelection?: boolean }) => void;
   resolveComparisonSelection: (
     proposed: Set<string>,
     prev?: Set<string>,
   ) => { result: Set<string>; keptGroup: string | null; droppedGroups: string[] };
-  /** Apply one official/overlay toggle; returns null when a cross-engine add is blocked. */
   toggleComparisonSelection: (
     prev: Set<string>,
     item: string,
     allItems: Set<string>,
   ) => Set<string> | null;
-  hardwareConfig: HardwareConfig;
-  graphs: RenderableGraph[];
-  selectedModel: Model;
   setSelectedModel: (model: Model) => void;
-  selectedSequence: Sequence;
   setSelectedSequence: (sequence: Sequence) => void;
-  selectedPrecisions: string[];
   setSelectedPrecisions: (precisions: string[]) => void;
-  loading: boolean;
-  error: string | null;
-  workflowInfo: any;
-  selectedYAxisMetric: string;
   setSelectedYAxisMetric: (metric: string) => void;
-  /** Latency percentile for the x-axis under agentic scenarios (median/p90/p99/p99.9). */
-  selectedPercentile: string;
-  setSelectedPercentile: (p: string) => void;
-  selectedXAxisMetric: string | null;
+  setTokenRevenuePriceSource: (source: TokenRevenuePriceSource) => void;
+  setSelectedPercentile: (percentile: string) => void;
   setSelectedXAxisMetric: (metric: string | null) => void;
-  selectedE2eXAxisMetric: string | null;
-  setSelectedE2eXAxisMetric: (metric: string | null) => void;
-  /**
-   * Which chart variant the user wants to see — the inference card shows one chart
-   * at a time, picked by the big buttons above the chart.
-   * - 'ttft'          → e2e chartType with x-axis forced to p90_ttft
-   * - 'e2e'           → e2e chartType with the chart-config default x-axis (median_e2el / p90_e2el)
-   * - 'interactivity' → interactivity chartType (x = median_intvty / p90_intvty)
-   * - 'e2e-normalized-interactivity'      → agentic-only; x = slow-tail per-request OSL / E2E latency
-   *                     in tok/s/user (live-computed from trace blobs)
-   */
-  selectedXAxisMode: 'ttft' | 'e2e' | 'interactivity' | 'e2e-normalized-interactivity';
   setSelectedXAxisMode: (
     mode: 'ttft' | 'e2e' | 'interactivity' | 'e2e-normalized-interactivity',
   ) => void;
-  scaleType: 'auto' | 'linear' | 'log';
   setScaleType: (type: 'auto' | 'linear' | 'log') => void;
-  /** Coarse vendor / framework / deployment / mtp-stp filters applied to the chart point set. */
-  quickFilters: QuickFilters;
-  /** Quick-filter values that have data for the current model (drives pill enable/disable). */
-  availableQuickFilters: AvailableQuickFilters;
   setQuickFilterVendors: (vendors: string[]) => void;
   setQuickFilterFrameworks: (frameworks: string[]) => void;
   setQuickFilterDeployment: (modes: DeploymentMode[]) => void;
   setQuickFilterSpec: (modes: SpecMode[]) => void;
-  setIsLegendExpanded: (metric: boolean) => void;
-  isLegendExpanded: boolean;
-  hideNonOptimal: boolean;
+  setIsLegendExpanded: (expanded: boolean) => void;
   setHideNonOptimal: (hide: boolean) => void;
-  showPointLabels: boolean;
   setShowPointLabels: (show: boolean) => void;
-  highContrast: boolean;
   setHighContrast: (highContrast: boolean) => void;
-  logScale: boolean;
   setLogScale: (logScale: boolean) => void;
-  useAdvancedLabels: boolean;
   setUseAdvancedLabels: (useAdvancedLabels: boolean) => void;
-  showGradientLabels: boolean;
+  setShowConcurrencyLabels: (showConcurrencyLabels: boolean) => void;
   setShowGradientLabels: (showGradientLabels: boolean) => void;
-  showLineLabels: boolean;
   setShowLineLabels: (showLineLabels: boolean) => void;
-  showSpeedOverlay: boolean;
-  setShowSpeedOverlay: (showSpeedOverlay: boolean) => void;
-  showMinecraftOverlay: boolean;
-  setShowMinecraftOverlay: (showMinecraftOverlay: boolean) => void;
-  selectedGPUs: string[];
   setSelectedGPUs: (gpus: string[]) => void;
-  availableGPUs: { value: string; label: string }[];
-  selectedDates: string[];
-  /** Accepts a value or a state-updater fn (for safe rapid successive adds). */
   setSelectedDates: (dates: string[] | ((prev: string[]) => string[])) => void;
-  /** Internal date-to-run normalization; preserves an Overview exact-pair scope. */
   setSelectedDatesFromRunExpansion: (dates: string[] | ((prev: string[]) => string[])) => void;
-  selectedDateRange: { startDate: string; endDate: string };
   setSelectedDateRange: (dateRange: { startDate: string; endDate: string }) => void;
-  userCosts: Record<string, number | undefined> | null;
   setUserCosts: (userCosts: Record<string, number | undefined> | null) => void;
-  selectedRunDate: string;
   setSelectedRunDate: (date: string) => void;
-  availableDates: string[];
-  dateRangeAvailableDates: string[];
-  isCheckingAvailableDates: boolean;
-  availableRuns: Record<string, RunInfo> | null;
-  selectedRunId: string;
   setSelectedRunId: (runId: string) => void;
-  availablePrecisions: string[];
-  availableSequences: Sequence[];
-  availableModels: string[];
-  userPowers: Record<string, number | undefined> | null;
   setUserPowers: (userPowers: Record<string, number | undefined> | null) => void;
   setHwFilter: (filter: string[] | null) => void;
-  activePresetId: string | null;
   setActivePresetId: (id: string | null) => void;
-  presetGuardRef: React.RefObject<boolean>;
-  /** Compare pages only: slug GPU pair used to filter benchmark series. */
-  compareGpuPair: readonly [string, string] | null;
 }
 export interface CalculateUserCostsRequest {
   model: string;
@@ -905,5 +733,6 @@ export interface ChangelogMetadata {
     pr_link: string | null;
     head_ref?: string;
     evals_only?: boolean;
+    append_only?: boolean;
   }[];
 }
diff --git a/packages/app/src/components/inference/ui/AgenticOptimizationNote.tsx b/packages/app/src/components/inference/ui/AgenticOptimizationNote.tsx
index 020d0c38b..7c158fcdc 100644
--- a/packages/app/src/components/inference/ui/AgenticOptimizationNote.tsx
+++ b/packages/app/src/components/inference/ui/AgenticOptimizationNote.tsx
@@ -22,14 +22,14 @@ const STRINGS = {
   },
 } as const;
 
-/** Agentic-only legend note; point tooltips carry the exact optimization method. */
+/** Agentic-only info-footer note; point tooltips carry the exact optimization method. */
 export function AgenticOptimizationNote() {
   const t = STRINGS[useLocale()];
 
   return (
     <div
       data-testid="agentic-optimization-note"
-      className="mt-2 flex w-full items-center gap-1 px-1 pr-2 text-xs italic text-blue-600 dark:text-blue-400"
+      className="flex w-full items-center gap-1 px-1 pr-2 text-xs italic text-blue-600 dark:text-blue-400"
     >
       <span>*{t.label}</span>
       <TooltipProvider delayDuration={100}>
diff --git a/packages/app/src/components/inference/ui/AxisMetricFooter.test.tsx b/packages/app/src/components/inference/ui/AxisMetricFooter.test.tsx
new file mode 100644
index 000000000..712d59976
--- /dev/null
+++ b/packages/app/src/components/inference/ui/AxisMetricFooter.test.tsx
@@ -0,0 +1,160 @@
+// @vitest-environment jsdom
+
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const mocks = vi.hoisted(() => ({
+  track: vi.fn(),
+  locale: { value: 'en' as 'en' | 'zh' },
+}));
+
+vi.mock('@/lib/analytics', () => ({ track: mocks.track }));
+vi.mock('@/lib/use-locale', () => ({ useLocale: () => mocks.locale.value }));
+
+import AxisMetricFooter from './AxisMetricFooter';
+
+declare global {
+  var IS_REACT_ACT_ENVIRONMENT: boolean;
+}
+
+globalThis.IS_REACT_ACT_ENVIRONMENT = true;
+
+let container: HTMLDivElement;
+let root: Root;
+
+beforeEach(() => {
+  mocks.track.mockClear();
+  mocks.locale.value = 'en';
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+});
+
+function render(props?: Partial<React.ComponentProps<typeof AxisMetricFooter>>) {
+  act(() => {
+    root.render(
+      <AxisMetricFooter
+        chartId="chart-0"
+        metricKey="tokensPerDollarH"
+        xAxisKind="interactivity"
+        xAxisLabel="Interactivity (tok/s/user)"
+        {...props}
+      />,
+    );
+  });
+}
+
+function click(el: Element) {
+  act(() => {
+    el.dispatchEvent(new MouseEvent('click', { bubbles: true }));
+  });
+}
+
+describe('AxisMetricFooter', () => {
+  it('renders notices in a dedicated section only when provided', () => {
+    render();
+    expect(container.querySelector('[data-testid="axis-metric-notices-chart-0"]')).toBeNull();
+    render({ notices: <span data-testid="footer-notice">KV offload ON</span> });
+    const notices = container.querySelector('[data-testid="axis-metric-notices-chart-0"]');
+    expect(notices).not.toBeNull();
+    expect(notices!.querySelector('[data-testid="footer-notice"]')!.textContent).toBe(
+      'KV offload ON',
+    );
+  });
+
+  it('renders one collapsed row per axis', () => {
+    render();
+    const xRow = container.querySelector('[data-testid="axis-metric-row-x-chart-0"]');
+    const yRow = container.querySelector('[data-testid="axis-metric-row-y-chart-0"]');
+    expect(xRow).not.toBeNull();
+    expect(yRow).not.toBeNull();
+    expect(xRow!.getAttribute('aria-expanded')).toBe('false');
+    expect(yRow!.getAttribute('aria-expanded')).toBe('false');
+    expect(xRow!.textContent).toContain('X-axis: Interactivity (tok/s/user)');
+    expect(yRow!.textContent).toContain('Y-axis: Total Tokens per $1 USD (Owning - Hyperscaler)');
+    // Bodies exist for aria-controls but stay hidden until toggled.
+    const xBody = container.querySelector<HTMLElement>(
+      '[data-testid="axis-metric-body-x-chart-0"]',
+    );
+    expect(xBody!.hidden).toBe(true);
+    expect(xRow!.getAttribute('aria-controls')).toBe(xBody!.id);
+  });
+
+  it('expands the x-axis row with an explanation and no formula', () => {
+    render();
+    const xRow = container.querySelector('[data-testid="axis-metric-row-x-chart-0"]')!;
+    click(xRow);
+    expect(xRow.getAttribute('aria-expanded')).toBe('true');
+    const xBody = container.querySelector<HTMLElement>(
+      '[data-testid="axis-metric-body-x-chart-0"]',
+    )!;
+    expect(xBody.hidden).toBe(false);
+    expect(xBody.textContent).toContain('rate at which a single user receives generated tokens');
+    // The x-axis row has no formula; the y-axis formula lives in the (still hidden) y body.
+    expect(xBody.querySelector('code')).toBeNull();
+    const yBody = container.querySelector<HTMLElement>(
+      '[data-testid="axis-metric-body-y-chart-0"]',
+    )!;
+    expect(yBody.hidden).toBe(true);
+  });
+
+  it('expands the y-axis row with explanation plus formula, then collapses', () => {
+    render();
+    const yRow = container.querySelector('[data-testid="axis-metric-row-y-chart-0"]')!;
+    click(yRow);
+    const yBody = container.querySelector<HTMLElement>(
+      '[data-testid="axis-metric-body-y-chart-0"]',
+    )!;
+    expect(yBody.hidden).toBe(false);
+    const formula = container.querySelector('[data-testid="axis-metric-formula-chart-0"]')!;
+    expect(formula.textContent).toBe(
+      'tok/$ = (total tok/s/chip × 3,600) ÷ all-in cost per chip-hour ($)',
+    );
+    click(yRow);
+    expect(yRow.getAttribute('aria-expanded')).toBe('false');
+    expect(yBody.hidden).toBe(true);
+  });
+
+  it('tracks toggle analytics with chart, axis, metric, and expanded state', () => {
+    render();
+    const yRow = container.querySelector('[data-testid="axis-metric-row-y-chart-0"]')!;
+    click(yRow);
+    expect(mocks.track).toHaveBeenCalledWith('axis_metric_footer_toggled', {
+      chart: 'chart-0',
+      axis: 'y',
+      metric: 'tokensPerDollarH',
+      expanded: true,
+    });
+    click(yRow);
+    expect(mocks.track).toHaveBeenLastCalledWith('axis_metric_footer_toggled', {
+      chart: 'chart-0',
+      axis: 'y',
+      metric: 'tokensPerDollarH',
+      expanded: false,
+    });
+  });
+
+  it('shows the percentile prefix for TTFT x-axes', () => {
+    render({ xAxisKind: 'ttft', xAxisLabel: 'P90 Time To First Token (s)' });
+    const xRow = container.querySelector('[data-testid="axis-metric-row-x-chart-0"]')!;
+    expect(xRow.textContent).toContain('X-axis: P90 Time To First Token (s)');
+  });
+
+  it('renders Chinese strings on /zh pages', () => {
+    mocks.locale.value = 'zh';
+    render({ metricKey: 'tpPerMw', xAxisKind: 'e2eLatency', xAxisLabel: 'End-to-end Latency (s)' });
+    const xRow = container.querySelector('[data-testid="axis-metric-row-x-chart-0"]')!;
+    const yRow = container.querySelector('[data-testid="axis-metric-row-y-chart-0"]')!;
+    expect(xRow.textContent).toContain('X 轴：端到端延迟（s）');
+    expect(yRow.textContent).toContain('Y 轴：每全电源配置兆瓦 token 吞吐量');
+    click(yRow);
+    const formula = container.querySelector('[data-testid="axis-metric-formula-chart-0"]')!;
+    expect(formula.textContent).toContain('每芯片全电源配置功率（MW）');
+  });
+});
diff --git a/packages/app/src/components/inference/ui/AxisMetricFooter.tsx b/packages/app/src/components/inference/ui/AxisMetricFooter.tsx
new file mode 100644
index 000000000..3c2e6610c
--- /dev/null
+++ b/packages/app/src/components/inference/ui/AxisMetricFooter.tsx
@@ -0,0 +1,169 @@
+'use client';
+
+import { Info, Plus } from 'lucide-react';
+import { useState, type ReactNode } from 'react';
+
+import {
+  METRIC_EXPLANATIONS,
+  metricRowLabel,
+  X_AXIS_EXPLANATIONS,
+  xAxisPercentileFromLabel,
+  type XAxisKind,
+} from '@/components/inference/axis-metric-explanations';
+import type { MetricKey } from '@/components/inference/metric-registry';
+import { track } from '@/lib/analytics';
+import { useLocale } from '@/lib/use-locale';
+import { cn } from '@/lib/utils';
+
+const STRINGS = {
+  en: {
+    xAxisPrefix: 'X-axis: ',
+    yAxisPrefix: 'Y-axis: ',
+    formula: 'Formula',
+  },
+  zh: {
+    xAxisPrefix: 'X 轴：',
+    yAxisPrefix: 'Y 轴：',
+    formula: '计算公式',
+  },
+} as const;
+
+interface AxisMetricFooterProps {
+  /** Chart instance id (e.g. "chart-0") for testids and analytics context. */
+  chartId: string;
+  /** Selected y-axis metric registry key (without the `y_` prefix). */
+  metricKey: MetricKey;
+  /** Resolved logical x-axis metric for the chart's current state. */
+  xAxisKind: XAxisKind;
+  /**
+   * The chart's resolved English x-axis label (percentile-adjusted), used
+   * only to extract the percentile prefix for the row name.
+   */
+  xAxisLabel: string;
+  /**
+   * Chart-level notices (KV-offload halo key, agentic optimization note,
+   * ATOM engine footnote) rendered as the info section's final block.
+   */
+  notices?: ReactNode;
+}
+
+interface FooterRow {
+  axis: 'x' | 'y';
+  /** Registry key for analytics (metric key or x-axis kind). */
+  key: string;
+  label: string;
+  description: string;
+  formula?: string;
+}
+
+/**
+ * Expandable per-axis metric explainer rendered below each inference chart.
+ * One row per axis: the x-axis row explains the plotted latency/interactivity
+ * metric, the y-axis row explains the selected metric and shows its
+ * structural formula. Explanations describe the metric itself, so they apply
+ * equally to official runs and `?unofficialrun=` overlay points, which share
+ * the chart's resolved axes. Marked `no-export` so PNG exports (which capture
+ * the chart element) never include it.
+ */
+export default function AxisMetricFooter({
+  chartId,
+  metricKey,
+  xAxisKind,
+  xAxisLabel,
+  notices,
+}: AxisMetricFooterProps) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const [expanded, setExpanded] = useState<{ x: boolean; y: boolean }>({ x: false, y: false });
+
+  const xExplanation = X_AXIS_EXPLANATIONS[xAxisKind];
+  const pctl = xAxisPercentileFromLabel(xAxisLabel);
+  const yExplanation = METRIC_EXPLANATIONS[metricKey];
+
+  const rows: FooterRow[] = [
+    {
+      axis: 'x',
+      key: xAxisKind,
+      label: `${t.xAxisPrefix}${xExplanation.name[locale](pctl)}`,
+      description: xExplanation.description[locale],
+    },
+    {
+      axis: 'y',
+      key: metricKey,
+      label: `${t.yAxisPrefix}${metricRowLabel(metricKey, locale)}`,
+      description: yExplanation.description[locale],
+      formula: yExplanation.formula[locale],
+    },
+  ];
+
+  const toggleRow = (row: FooterRow) => {
+    setExpanded((prev) => {
+      const next = !prev[row.axis];
+      track('axis_metric_footer_toggled', {
+        chart: chartId,
+        axis: row.axis,
+        metric: row.key,
+        expanded: next,
+      });
+      return { ...prev, [row.axis]: next };
+    });
+  };
+
+  return (
+    <div className="no-export mt-4" data-testid={`axis-metric-footer-${chartId}`}>
+      {rows.map((row) => {
+        const isOpen = expanded[row.axis];
+        const bodyId = `${chartId}-axis-metric-${row.axis}`;
+        return (
+          <div key={row.axis} className="border-t border-border/60">
+            <button
+              type="button"
+              className="flex w-full cursor-pointer items-center gap-2 py-2.5 text-left"
+              aria-expanded={isOpen}
+              aria-controls={bodyId}
+              data-testid={`axis-metric-row-${row.axis}-${chartId}`}
+              onClick={() => toggleRow(row)}
+            >
+              <Info aria-hidden="true" className="size-3.5 shrink-0 text-muted-foreground" />
+              <span className="flex-1 text-sm text-muted-foreground">{row.label}</span>
+              <Plus
+                aria-hidden="true"
+                className={cn(
+                  'size-4 shrink-0 text-muted-foreground transition-transform duration-200',
+                  isOpen && 'rotate-45',
+                )}
+              />
+            </button>
+            <div
+              id={bodyId}
+              hidden={!isOpen}
+              className="pb-3 pl-[1.375rem]"
+              data-testid={`axis-metric-body-${row.axis}-${chartId}`}
+            >
+              <p className="text-sm text-muted-foreground">{row.description}</p>
+              {row.formula && (
+                <p className="mt-2">
+                  <span className="sr-only">{t.formula}: </span>
+                  <code
+                    className="rounded bg-muted px-1.5 py-0.5 font-mono text-xs text-muted-foreground"
+                    data-testid={`axis-metric-formula-${chartId}`}
+                  >
+                    {row.formula}
+                  </code>
+                </p>
+              )}
+            </div>
+          </div>
+        );
+      })}
+      {notices && (
+        <div
+          className="space-y-2 border-t border-border/60 py-2.5"
+          data-testid={`axis-metric-notices-${chartId}`}
+        >
+          {notices}
+        </div>
+      )}
+    </div>
+  );
+}
diff --git a/packages/app/src/components/inference/ui/ChartControls.tsx b/packages/app/src/components/inference/ui/ChartControls.tsx
index 8882128c2..4c0ef0604 100644
--- a/packages/app/src/components/inference/ui/ChartControls.tsx
+++ b/packages/app/src/components/inference/ui/ChartControls.tsx
@@ -3,10 +3,17 @@
 import { useEffect, useMemo, useState } from 'react';
 
 import { track } from '@/lib/analytics';
+import { replaceRouterPathname } from '@/lib/client-navigation';
+import { AGENTX_NEW_MODEL_DISPLAY_NAMES } from '@/lib/compare-agentx';
+import { inferenceModelRouteForSelection } from '@/lib/inference-model-slug';
 import { useFeatureGate } from '@/lib/use-feature-gate';
-import { cn } from '@/lib/utils';
 
-import { useInference } from '@/components/inference/InferenceContext';
+import {
+  useInferenceActions,
+  useInferenceData,
+  useInferenceDisplay,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
 import {
   ModelSelector,
   ScenarioSelector,
@@ -25,11 +32,10 @@ import {
 } from '@/components/ui/select';
 import { SearchableSelect } from '@/components/ui/searchable-select';
 import { TooltipProvider } from '@/components/ui/tooltip';
-import { Button } from '@/components/ui/button';
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
-import type { ChartDefinition, DeploymentMode, SpecMode } from '@/components/inference/types';
-import { resolveComparisonEntries } from '@/components/inference/utils/comparisonEntry';
-import { FRAMEWORK_FAMILIES } from '@/components/inference/utils/quickFilters';
+import { METRIC_CONTROL_GROUPS, METRIC_REGISTRY } from '@/components/inference/metric-registry';
+import { formatTokenPrice } from '@/components/inference/token-revenue';
+import { useOpenDropdown } from '@/hooks/useOpenDropdown';
+import { ModelArchitectureInfoLink } from './ModelArchitectureInfoLink';
 import { Sequence, type Model, type Percentile } from '@/lib/data-mappings';
 import { useLocale } from '@/lib/use-locale';
 
@@ -37,7 +43,7 @@ const STRINGS = {
   en: {
     yAxisMetric: 'Y-Axis Metric',
     yAxisMetricTooltip:
-      "The performance metric displayed on the chart's Y-axis. Options include throughput (tokens/sec), cost per million tokens, and custom user-defined values.",
+      "The performance metric displayed on the chart's Y-axis. Options include throughput, token revenue per GPU hour, cost per million tokens, tokens per $1 USD or ¥1 CNY, and custom user-defined values.",
     xAxisMetric: 'X-Axis Metric',
     xAxisMetricTooltip:
       "The latency metric displayed on the chart's X-axis: P90 Time To First Token.",
@@ -55,22 +61,21 @@ const STRINGS = {
     comparisonDateRangeTooltip:
       'Select the start and end dates for the historical comparison. The chart will show performance data for the selected chip configs across this time range.',
     dateRangePlaceholder: 'Select date range',
-    quickFilters: 'Quick Filters',
-    quickFiltersTooltip:
-      'Narrow the chart by chip vendor, serving framework, deployment mode (single-node, multi-node aggregate, or disaggregated), and speculative decoding. Selecting none in a group shows all.',
-    filterVendor: 'Vendor',
-    filterFramework: 'Framework',
-    filterDeployment: 'Deployment',
-    singleNode: 'Single-node',
-    multiNode: 'Multi-node',
-    disaggregated: 'Disaggregated',
-    filterSpecDecoding: 'Spec Decoding',
-    noData: 'No data for the current selection',
+    revenuePriceSource: 'Revenue Price Source',
+    revenuePriceSourceTooltip:
+      'Choose the token sale prices used for revenue. Normalized prices input and output at $1/M tok. OpenRouter reads the selected model’s current public input and output prices.',
+    normalizedPrice: 'Normalized ($1/M input + output)',
+    openRouterPrice: 'OpenRouter current pricing',
+    openRouterLoading: 'Loading OpenRouter pricing…',
+    openRouterUnavailable: 'OpenRouter pricing is unavailable for this model.',
+    openRouterSummary: (input: string, output: string) =>
+      `Input $${input}/M tok · Output $${output}/M tok`,
+    viewOpenRouter: 'View OpenRouter pricing',
   },
   zh: {
     yAxisMetric: 'Y 轴指标',
     yAxisMetricTooltip:
-      '图表 Y 轴显示的性能指标。包括吞吐量（token/秒）、每百万 token 成本以及自定义用户值。',
+      '图表 Y 轴显示的性能指标，包括吞吐量、每 GPU 小时 token 收入、每百万 token 成本、每 1 美元可购买的 token 数以及自定义值。',
     xAxisMetric: 'X 轴指标',
     xAxisMetricTooltip: '图表 X 轴显示的延迟指标：P90 Time To First Token。',
     xAxisScale: 'X 轴刻度',
@@ -79,128 +84,36 @@ const STRINGS = {
     scaleAuto: '自动',
     scaleLinear: '线性',
     scaleLog: '对数',
-    gpuConfig: 'Chip 配置',
+    gpuConfig: '芯片配置',
     gpuConfigTooltip:
-      '最多选择 4 个 Chip 配置以对比其历史性能趋势。可用于追踪软件更新对特定硬件的影响。',
-    gpuConfigPlaceholder: '选择 Chip 配置进行对比',
+      '最多选择 4 个芯片配置以对比其历史性能趋势。可用于追踪软件更新对特定硬件的影响。',
+    gpuConfigPlaceholder: '选择芯片配置进行对比',
     comparisonDateRange: '对比日期范围',
     comparisonDateRangeTooltip:
-      '选择历史对比的起止日期。图表将展示所选 Chip 配置在此时间范围内的性能数据。',
+      '选择历史对比的起止日期。图表将展示所选芯片配置在此时间范围内的性能数据。',
     dateRangePlaceholder: '选择日期范围',
-    quickFilters: '快捷筛选',
-    quickFiltersTooltip:
-      '按 Chip 厂商、推理框架、部署模式（单节点、多节点聚合或分离式）和投机解码筛选图表。某组不选则显示全部。',
-    filterVendor: '厂商',
-    filterFramework: '框架',
-    filterDeployment: '部署模式',
-    singleNode: '单节点',
-    multiNode: '多节点聚合',
-    disaggregated: '分离式',
-    filterSpecDecoding: '投机解码',
-    noData: '当前选择无可用数据',
+    revenuePriceSource: '收入计价来源',
+    revenuePriceSourceTooltip:
+      '选择计算 token 收入所用的售价。标准化模式将输入和输出 token 均按 $1/百万计价；OpenRouter 模式读取所选模型当前公开的输入和输出价格。',
+    normalizedPrice: '标准化（输入和输出均为 $1/百万）',
+    openRouterPrice: 'OpenRouter 当前价格',
+    openRouterLoading: '正在加载 OpenRouter 价格…',
+    openRouterUnavailable: 'OpenRouter 暂无该模型的价格。',
+    openRouterSummary: (input: string, output: string) =>
+      `输入 $${input}/百万 token · 输出 $${output}/百万 token`,
+    viewOpenRouter: '查看 OpenRouter 定价',
   },
 } as const;
 
-/**
- * Y-axis metric options from static chart config JSON — available immediately, no API wait.
- *
- * Groups marked `gated: true` are hidden unless the konami-code feature gate is unlocked
- * (see useFeatureGate). Use this for surfaces that are wired but whose underlying data
- * pipeline is in the rollout phase (e.g. measured-power telemetry waiting on a runner-
- * side aggregation PR to start populating the DB).
- */
-const METRIC_GROUPS: {
-  label: string;
-  labelZh: string;
-  metrics: string[];
-  gated?: boolean;
-}[] = [
-  {
-    label: 'Throughput',
-    labelZh: '吞吐量',
-    metrics: [
-      'y_tpPerGpu',
-      'y_inputTputPerGpu',
-      'y_outputTputPerGpu',
-      'y_tpPerMw',
-      'y_inputTputPerMw',
-      'y_outputTputPerMw',
-    ],
-  },
-  {
-    label: 'Cost per Million Total Tokens',
-    labelZh: '每百万总 token 成本',
-    metrics: ['y_costh', 'y_costn', 'y_costr'],
-  },
-  {
-    label: 'Cost per Million Output Tokens',
-    labelZh: '每百万输出 token 成本',
-    metrics: ['y_costhOutput', 'y_costnOutput', 'y_costrOutput'],
-  },
-  {
-    label: 'Cost per Million Input Tokens',
-    labelZh: '每百万输入 token 成本',
-    metrics: ['y_costhi', 'y_costni', 'y_costri'],
-  },
-  {
-    label: 'All-in Provisioned Energy per Token',
-    labelZh: '每 token 全电源配置能耗',
-    metrics: ['y_jTotal', 'y_jOutput', 'y_jInput'],
-  },
-  {
-    label: 'Measured Energy',
-    labelZh: '实测能耗',
-    metrics: [
-      'y_measuredPrefillAvgPower',
-      'y_measuredDecodeAvgPower',
-      'y_measuredAvgPower',
-      'y_measuredJPerInputToken',
-      'y_measuredJPerOutputToken',
-      'y_measuredJPerTotalToken',
-    ],
-  },
-  { label: 'Custom User Values', labelZh: '自定义值', metrics: ['y_costUser', 'y_powerUser'] },
-];
+const METRIC_GROUPS = METRIC_CONTROL_GROUPS;
 
-/** Map from metric key → human-readable title (e.g. "Token Throughput per GPU") */
-const METRIC_TITLE_MAP = (() => {
-  const chartDef = (chartDefinitions as ChartDefinition[])[0];
-  const map = new Map<string, string>();
-  for (const key of Object.keys(chartDef)) {
-    if (key.startsWith('y_') && key.endsWith('_title')) {
-      map.set(key.replace('_title', ''), chartDef[key as keyof ChartDefinition] as string);
-    }
-  }
-  return map;
-})();
-
-const METRIC_TITLE_ZH_MAP = (() => {
-  const chartDef = (chartDefinitions as ChartDefinition[])[0];
-  const map = new Map<string, string>();
-  for (const key of Object.keys(chartDef)) {
-    if (key.startsWith('y_') && key.endsWith('_titleZh')) {
-      const metricKey = key.replace('_titleZh', '');
-      map.set(metricKey, chartDef[key] as string);
-    }
-  }
-  return map;
-})();
+const METRIC_TITLE_MAP = new Map(
+  Object.entries(METRIC_REGISTRY).map(([key, metric]) => [`y_${key}`, metric.title]),
+);
 
-/** Quick-filter pill groups: vendor, deployment mode, spec-decoding method. */
-const QUICK_FILTER_VENDORS: { value: string; label: string }[] = [
-  { value: 'NVIDIA', label: 'NVIDIA' },
-  { value: 'AMD', label: 'AMD' },
-];
-const QUICK_FILTER_DEPLOYMENT: DeploymentMode[] = ['single-node', 'multi-node', 'disagg'];
-const QUICK_FILTER_SPEC: { value: SpecMode; label: string }[] = [
-  { value: 'mtp', label: 'MTP' },
-  { value: 'stp', label: 'STP' },
-];
-
-/** Toggle a value in/out of a quick-filter selection array. */
-function toggleValue<T extends string>(current: T[], value: T): T[] {
-  return current.includes(value) ? current.filter((v) => v !== value) : [...current, value];
-}
+const METRIC_TITLE_ZH_MAP = new Map(
+  Object.entries(METRIC_REGISTRY).map(([key, metric]) => [`y_${key}`, metric.titleZh]),
+);
 
 interface ChartControlsProps {
   /** Hide GPU Config selector and related date pickers (used by Historical Trends tab) */
@@ -217,52 +130,45 @@ export default function ChartControls({ hideGpuComparison = false }: ChartContro
   const [mounted, setMounted] = useState(false);
   useEffect(() => setMounted(true), []);
 
-  const [openDropdown, setOpenDropdown] = useState<string | null>(null);
-  const handleDropdownOpenChange = (dropdownKey: string) => (open: boolean) => {
-    if (open) {
-      setOpenDropdown(dropdownKey);
-      return;
-    }
-    setOpenDropdown((current) => (current === dropdownKey ? null : current));
-  };
+  const { openDropdown, handleDropdownOpenChange } = useOpenDropdown<string>();
 
+  const { selectedModel, selectedSequence, selectedPrecisions, selectedGPUs, selectedDateRange } =
+    useInferenceFilters();
   const {
-    selectedModel,
-    setSelectedModel,
-    selectedSequence,
-    setSelectedSequence,
-    selectedPrecisions,
-    setSelectedPrecisions,
-    selectedYAxisMetric,
-    setSelectedYAxisMetric,
-    selectedPercentile,
-    setSelectedPercentile,
     graphs,
-    selectedGPUs,
-    setSelectedGPUs,
     availableGPUs,
-    selectedDates,
-    selectedDateRange,
-    setSelectedDateRange,
     dateRangeAvailableDates,
     isCheckingAvailableDates,
     availablePrecisions,
     availableSequences,
     availableModels,
+  } = useInferenceData();
+  const {
+    selectedYAxisMetric,
+    tokenRevenuePriceSource,
+    tokenRevenuePricing,
+    openRouterModelId,
+    openRouterPricingLoading,
+    openRouterPricingError,
+    selectedPercentile,
     selectedXAxisMetric,
-    setSelectedXAxisMetric,
     scaleType,
+  } = useInferenceDisplay();
+  const {
+    setSelectedModel,
+    setSelectedSequence,
+    setSelectedPrecisions,
+    setSelectedYAxisMetric,
+    setTokenRevenuePriceSource,
+    setSelectedPercentile,
+    setSelectedGPUs,
+    setSelectedDateRange,
+    setSelectedXAxisMetric,
     setScaleType,
-    quickFilters,
-    availableQuickFilters,
-    setQuickFilterVendors,
-    setQuickFilterFrameworks,
-    setQuickFilterDeployment,
-    setQuickFilterSpec,
-  } = useInference();
+  } = useInferenceActions();
 
-  // Y-axis metric options — built from static chart config JSON (no API dependency).
-  // Hidden groups (Measured Energy) appear only after the ↑↑↓↓ feature gate unlocks.
+  // Y-axis options come from the canonical registry and need no API data.
+  // Gated groups appear only after the feature gate unlocks.
   const featureGateUnlocked = useFeatureGate();
   const visibleGroups = useMemo(
     () => METRIC_GROUPS.filter((g) => !g.gated || featureGateUnlocked),
@@ -308,6 +214,15 @@ export default function ChartControls({ hideGpuComparison = false }: ChartContro
 
   const handleModelChange = (value: Model) => {
     setSelectedModel(value);
+    // A deliberate pick moves the URL onto the model's indexable subroute in
+    // place — no reload, RSC refetch, or scroll reset. Kept out of the model
+    // state effects on purpose: programmatic changes (back-nav restore,
+    // config load, auto-switch) must not rewrite the URL, and event handlers
+    // are immune to Strict Mode's double-invoked effects. `g_model` is
+    // dropped because the path now carries the model — a lingering share
+    // param would override it on the next snapshot read.
+    const target = inferenceModelRouteForSelection(window.location.pathname, value);
+    if (target !== null) replaceRouterPathname(target, ['g_model']);
     track('inference_model_selected', {
       model: value,
     });
@@ -363,28 +278,6 @@ export default function ChartControls({ hideGpuComparison = false }: ChartContro
     });
   };
 
-  const handleQuickFilterToggle = (
-    category: 'vendor' | 'framework' | 'deployment' | 'spec',
-    value: string,
-  ) => {
-    const wasActive =
-      category === 'vendor'
-        ? quickFilters.vendors.includes(value)
-        : category === 'framework'
-          ? quickFilters.frameworks.includes(value)
-          : category === 'deployment'
-            ? quickFilters.deployment.includes(value as DeploymentMode)
-            : quickFilters.spec.includes(value as SpecMode);
-    if (category === 'vendor') setQuickFilterVendors(toggleValue(quickFilters.vendors, value));
-    else if (category === 'framework')
-      setQuickFilterFrameworks(toggleValue(quickFilters.frameworks, value));
-    else if (category === 'deployment')
-      setQuickFilterDeployment(toggleValue(quickFilters.deployment, value as DeploymentMode));
-    else setQuickFilterSpec(toggleValue(quickFilters.spec, value as SpecMode));
-    // `active` is the state *after* this toggle.
-    track('inference_quick_filter_toggled', { category, value, active: !wasActive });
-  };
-
   const isInputMetric = (() => {
     const chartDef = graphs[0]?.chartDefinition;
     if (!chartDef) return false;
@@ -401,70 +294,6 @@ export default function ChartControls({ hideGpuComparison = false }: ChartContro
     });
   };
 
-  // Quick-filter pill groups. Each option carries an `available` flag (has data
-  // for the current model); the render disables unavailable options unless they
-  // are currently selected, so a selection can always be toggled back off. The
-  // Framework group is data-driven — only families present (or selected) are
-  // offered, so it's omitted entirely when none resolve (e.g. while data loads).
-  const fwSelected = quickFilters.frameworks;
-  const frameworkOptions = FRAMEWORK_FAMILIES.filter(
-    (f) => availableQuickFilters.frameworks.includes(f.key) || fwSelected.includes(f.key),
-  ).map((f) => ({
-    value: f.key,
-    label: f.label,
-    available: availableQuickFilters.frameworks.includes(f.key),
-  }));
-  const quickFilterGroups: {
-    key: 'vendor' | 'framework' | 'deployment' | 'spec';
-    label: string;
-    options: readonly { value: string; label: string; available: boolean }[];
-    selected: readonly string[];
-  }[] = [
-    {
-      key: 'vendor',
-      label: t.filterVendor,
-      options: QUICK_FILTER_VENDORS.map((o) => ({
-        ...o,
-        available: availableQuickFilters.vendors.includes(o.value),
-      })),
-      selected: quickFilters.vendors,
-    },
-    ...(frameworkOptions.length > 0
-      ? [
-          {
-            key: 'framework' as const,
-            label: t.filterFramework,
-            options: frameworkOptions,
-            selected: quickFilters.frameworks,
-          },
-        ]
-      : []),
-    {
-      key: 'deployment',
-      label: t.filterDeployment,
-      options: QUICK_FILTER_DEPLOYMENT.map((value) => ({
-        value,
-        label:
-          value === 'single-node'
-            ? t.singleNode
-            : value === 'multi-node'
-              ? t.multiNode
-              : t.disaggregated,
-        available: availableQuickFilters.deployment.includes(value),
-      })),
-      selected: quickFilters.deployment,
-    },
-    {
-      key: 'spec',
-      label: t.filterSpecDecoding,
-      options: QUICK_FILTER_SPEC.map((o) => ({
-        ...o,
-        available: availableQuickFilters.spec.includes(o.value),
-      })),
-      selected: quickFilters.spec,
-    },
-  ];
-
   return (
     <TooltipProvider delayDuration={0}>
       <div className="flex flex-col gap-4">
@@ -476,6 +305,8 @@ export default function ChartControls({ hideGpuComparison = false }: ChartContro
             onOpenChange={handleDropdownOpenChange('model')}
             availableModels={availableModels}
             data-testid="model-selector"
+            trailing={<ModelArchitectureInfoLink model={selectedModel} locale={locale} />}
+            newModels={AGENTX_NEW_MODEL_DISPLAY_NAMES}
           />
           <ScenarioSelector
             value={selectedSequence}
@@ -521,11 +352,70 @@ export default function ChartControls({ hideGpuComparison = false }: ChartContro
                 options: g.options,
               }))}
               searchPlaceholder={locale === 'zh' ? '搜索…' : undefined}
+              searchAriaLabel={locale === 'zh' ? '搜索指标选项' : undefined}
               noResultsLabel={locale === 'zh' ? '无结果' : undefined}
               clearSearchLabel={locale === 'zh' ? '清除搜索' : undefined}
             />
           </div>
 
+          {mounted && selectedYAxisMetric === 'y_tokenRevenuePerGpuHour' && (
+            <div className="flex flex-col space-y-1.5 lg:col-span-2">
+              <LabelWithTooltip
+                htmlFor="token-revenue-price-source"
+                label={t.revenuePriceSource}
+                tooltip={t.revenuePriceSourceTooltip}
+              />
+              <Select
+                value={tokenRevenuePriceSource}
+                onValueChange={(value) => {
+                  setTokenRevenuePriceSource(value as 'normalized' | 'openrouter');
+                  track('inference_token_revenue_price_source_selected', { source: value });
+                }}
+              >
+                <SelectTrigger
+                  id="token-revenue-price-source"
+                  data-testid="token-revenue-price-source"
+                  className="w-full"
+                >
+                  <SelectValue />
+                </SelectTrigger>
+                <SelectContent portalled={false}>
+                  <SelectItem value="normalized">{t.normalizedPrice}</SelectItem>
+                  <SelectItem value="openrouter">{t.openRouterPrice}</SelectItem>
+                </SelectContent>
+              </Select>
+              {tokenRevenuePriceSource === 'openrouter' && (
+                <p data-testid="openrouter-price-summary" className="text-xs text-muted-foreground">
+                  {openRouterPricingLoading
+                    ? t.openRouterLoading
+                    : openRouterPricingError || !tokenRevenuePricing
+                      ? t.openRouterUnavailable
+                      : t.openRouterSummary(
+                          formatTokenPrice(tokenRevenuePricing.inputPerMillion),
+                          formatTokenPrice(tokenRevenuePricing.outputPerMillion),
+                        )}{' '}
+                  {openRouterModelId && (
+                    <a
+                      href={`https://openrouter.ai/${openRouterModelId}`}
+                      target="_blank"
+                      rel="noopener noreferrer"
+                      className="underline underline-offset-2"
+                      data-testid="openrouter-pricing-link"
+                      onClick={() => {
+                        track('inference_openrouter_pricing_opened', {
+                          model: selectedModel,
+                          openRouterModelId,
+                        });
+                      }}
+                    >
+                      {t.viewOpenRouter}
+                    </a>
+                  )}
+                </p>
+              )}
+            </div>
+          )}
+
           {graphs.some((g) => g.chartDefinition?.chartType === 'interactivity') &&
             isInputMetric &&
             selectedSequence !== Sequence.AgenticTraces && (
@@ -616,64 +506,10 @@ export default function ChartControls({ hideGpuComparison = false }: ChartContro
                 placeholder={t.dateRangePlaceholder}
                 availableDates={dateRangeAvailableDates}
                 isCheckingAvailableDates={isCheckingAvailableDates}
-                className={
-                  // Note (wenyao): a pinned run (`date~rID`) can only reach the chart through
-                  // selectedDates, never through a range, so demanding a range here raises a
-                  // false alarm on comparisons that are already complete.
-                  resolveComparisonEntries(selectedDates, selectedDateRange).length === 0
-                    ? 'border-red-500 ring-4 ring-red-500/40 animate-pulse'
-                    : ''
-                }
               />
             </div>
           )}
         </div>
-
-        {!hideGpuComparison && (
-          <div className="flex flex-col space-y-1.5" data-testid="quick-filters">
-            <LabelWithTooltip
-              htmlFor="quick-filters"
-              label={t.quickFilters}
-              tooltip={t.quickFiltersTooltip}
-            />
-            <div className="flex flex-wrap items-center gap-x-6 gap-y-2">
-              {quickFilterGroups.map((group) => (
-                <div key={group.key} className="flex items-center gap-1.5">
-                  <span className="text-xs font-medium text-muted-foreground">{group.label}:</span>
-                  <div className="flex flex-wrap gap-1">
-                    {group.options.map((option) => {
-                      const active = (group.selected as readonly string[]).includes(option.value);
-                      // Disable options with no data, but keep a selected one
-                      // clickable so it can always be toggled back off.
-                      const disabled = !option.available && !active;
-                      return (
-                        <Button
-                          key={option.value}
-                          type="button"
-                          size="sm"
-                          variant={active ? 'default' : 'outline'}
-                          aria-pressed={active}
-                          disabled={disabled}
-                          title={disabled ? t.noData : undefined}
-                          // Active pills use the brand color (blue in light, amber in dark)
-                          // rather than the amber primary fill.
-                          className={cn(
-                            'h-7 rounded-full px-3 text-xs',
-                            active && 'bg-brand hover:bg-brand/90',
-                          )}
-                          data-testid={`quick-filter-${group.key}-${option.value}`}
-                          onClick={() => handleQuickFilterToggle(group.key, option.value)}
-                        >
-                          {option.label}
-                        </Button>
-                      );
-                    })}
-                  </div>
-                </div>
-              ))}
-            </div>
-          </div>
-        )}
       </div>
     </TooltipProvider>
   );
diff --git a/packages/app/src/components/inference/ui/ChartDisplay.tsx b/packages/app/src/components/inference/ui/ChartDisplay.tsx
index dfc4c9a18..50ad304b6 100644
--- a/packages/app/src/components/inference/ui/ChartDisplay.tsx
+++ b/packages/app/src/components/inference/ui/ChartDisplay.tsx
@@ -1,12 +1,22 @@
 'use client';
 import { DISPLAY_MODEL_TO_DB } from '@semianalysisai/inferencex-constants';
 import { track } from '@/lib/analytics';
-import dynamic from 'next/dynamic';
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
-import { BarChart3, Check, ChevronDown, Table2 } from 'lucide-react';
+import { BarChart3, Table2 } from 'lucide-react';
 
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
-import { useInference } from '@/components/inference/InferenceContext';
+import chartDefinitions, {
+  tokenMetricTypeForConfigKey,
+  type MetricKey,
+} from '@/components/inference/metric-registry';
+import { resolveXAxisKind } from '@/components/inference/axis-metric-explanations';
+import { resolveXAxisField } from '@/components/inference/utils/resolveXAxisField';
+import { applyTokenRevenuePricing, formatTokenPrice } from '@/components/inference/token-revenue';
+import {
+  useInferenceActions,
+  useInferenceData,
+  useInferenceDisplay,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
 import type {
   ChartDefinition,
   HardwareConfig,
@@ -23,24 +33,28 @@ import {
 } from '@/components/inference/utils/comparisonEntry';
 import { dataRunsForDate } from '@/components/inference/utils/runEnumeration';
 import { matchesQuickFilters } from '@/components/inference/utils/quickFilters';
-import { canonicalNormalizedFrontierIds } from '@/components/inference/utils/canonicalFrontier';
 import { bestSeriesPerSku } from '@/components/inference/utils/best-series-per-sku';
 import InferenceTable from '@/components/inference/ui/InferenceTable';
 import ScatterGraph from '@/components/inference/ui/ScatterGraph';
 import { Card } from '@/components/ui/card';
 import { ChartButtons } from '@/components/ui/chart-buttons';
+import { ShareButton } from '@/components/ui/share-button';
 import { type SegmentedToggleOption, SegmentedToggle } from '@/components/ui/segmented-toggle';
-import { Popover, PopoverContent, PopoverTrigger } from '@/components/ui/popover';
 import { Tabs, TabsList, TabsTrigger } from '@/components/ui/tabs';
-import { ChartShareActions, MetricAssumptionNotes } from '@/components/ui/chart-display-helpers';
+import { MetricAssumptionNotes } from '@/components/ui/chart-display-helpers';
 import { UnofficialDomainNotice } from '@/components/ui/unofficial-domain-notice';
+import { ModelLogo } from '@/components/ui/model-logo';
 import { metricLabel, metricTitle } from '@/lib/chart-utils';
 import { exportToCsv } from '@/lib/csv-export';
 import { inferenceChartToCsv } from '@/lib/csv-export-helpers';
 import { knownIssueCsvNote, matchKnownConfigIssues } from '@/lib/known-issues';
-import { cn, getDisplayLabel } from '@/lib/utils';
+import { getDisplayLabel, getFrameworkLabel } from '@/lib/utils';
+import { supportsChartTokenMetric } from '@/lib/supplemental-benchmarks';
 import { Skeleton } from '@/components/ui/skeleton';
-import { useUnofficialRun } from '@/components/unofficial-run-provider';
+import {
+  useOverlayScopeReconciliation,
+  useUnofficialRun,
+} from '@/components/unofficial-run-provider';
 import {
   type Model,
   type Precision,
@@ -61,11 +75,16 @@ import {
   useDerivedAgenticMetrics,
   type DerivedAgenticMetric,
 } from '@/hooks/api/use-derived-agentic-metrics';
+import { useResidentSequenceLengths } from '@/hooks/api/use-resident-sequence-lengths';
 import { getHardwareConfig, hardwareKeyMatchesAnyBase } from '@/lib/constants';
 import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
-import type { Locale } from '@/lib/i18n';
 import { useLocale } from '@/lib/use-locale';
 
+import { ATOM_FOOTNOTE_MARKER, AtomEngineFootnote } from '@/components/ui/atom-engine-footnote';
+import { AgenticOptimizationNote } from '@/components/inference/ui/AgenticOptimizationNote';
+import { OffloadHaloLegendKey } from '@/components/inference/ui/OffloadHaloLegendKey';
+
+import AxisMetricFooter from './AxisMetricFooter';
 import ChartControls from './ChartControls';
 import ComparisonChangelog from './ComparisonChangelog';
 import CustomCosts from './CustomCosts';
@@ -73,11 +92,8 @@ import CustomPowers from './CustomPowers';
 import GPUGraph from './GPUGraph';
 import ReplayLauncher, { type ReplayLauncherHandle } from '../replay/ReplayLauncher';
 
-const ModelArchitectureDiagram = dynamic(() => import('./ModelArchitectureDiagram'), {
-  ssr: false,
-  loading: () => <Skeleton className="h-40 w-full" />,
-});
 import WorkflowInfoDisplay from './WorkflowInfoDisplay';
+import { NormalizedInteractivityHelpLink } from './NormalizedInteractivityHelpLink';
 
 type InferenceViewMode = 'chart' | 'table';
 
@@ -85,42 +101,44 @@ const STRINGS = {
   en: {
     inferencePerformance: 'Inference Performance',
     inferencePerformanceDesc:
-      'Inference performance metrics across different models, hardware configurations, and serving parameters.',
+      'Agentic inference metrics from the AgentX scenario and fixed-sequence inference metrics across models, hardware configurations, and serving parameters.',
     chart: 'Chart',
     table: 'Table',
     sourceUnofficial: 'Source: UNOFFICIAL',
     sourceOfficial: 'Source: SemiAnalysis InferenceX™',
+    revenuePrices: (input: string, output: string) =>
+      `Input $${input}/M tok · Output $${output}/M tok`,
     updated: 'Updated:',
     e2eNormIntvtyDisclaimer:
       'E2E Normalized Interactivity requires persisted per-request traces, so unofficial-run overlays are unavailable for this experimental view.',
-    selectDateRange: 'Select a date range or add a run to view chip comparison',
+    completedSequenceLengths: (count: string) =>
+      `Completed requests across all resident points (n=${count})`,
     viewMode: 'View mode',
     vsTtft: (word: string) => `vs. ${word} Time To First Token`,
     vsE2eLatency: (pctl?: string) =>
       pctl ? `vs. ${pctl} End-to-end Latency` : 'vs. End-to-end Latency',
-    advancedXAxis: 'Advanced',
-    advancedXAxisWith: (label: string) => `Advanced: ${label}`,
   },
   zh: {
     inferencePerformance: '推理性能',
-    inferencePerformanceDesc: '不同模型、硬件配置和服务参数下的推理性能指标。',
+    inferencePerformanceDesc:
+      '不同模型、硬件配置和服务参数下，来自 AgentX 场景的智能体推理指标与固定序列推理指标。',
     chart: '图表',
     table: '表格',
     sourceUnofficial: '来源：非官方',
     sourceOfficial: '来源：SemiAnalysis InferenceX™',
+    revenuePrices: (input: string, output: string) =>
+      `输入 $${input}/百万 token · 输出 $${output}/百万 token`,
     updated: '更新时间：',
     e2eNormIntvtyDisclaimer:
       '端到端归一化交互性需要持久化的逐请求 trace 数据，因此该实验性视图不支持非官方运行覆盖。',
-    selectDateRange: '请选择日期范围或添加运行以查看 Chip 对比',
+    completedSequenceLengths: (count: string) => `当前所有数据点的已完成请求（n=${count}）`,
     viewMode: '视图模式',
     vsTtft: (word: string) => `vs. ${word === 'Median' ? '中位' : word} 首 token 延迟（TTFT）`,
     vsE2eLatency: (pctl?: string) => (pctl ? `vs. ${pctl} 端到端延迟` : 'vs. 端到端延迟'),
-    advancedXAxis: '高级',
-    advancedXAxisWith: (label: string) => `高级：${label}`,
   },
 } as const;
 
-// Translate the "vs. …" chart-heading suffix from inference-chart-config.json
+// Translate the "vs. …" chart-heading suffix from the metric registry
 // into Chinese. useChartData rewrites the heading with the selected percentile
 // for agentic sequences (e.g. "vs. P90 Interactivity"), so this matches the
 // pattern instead of a fixed string; unknown headings pass through unchanged.
@@ -152,103 +170,6 @@ const X_AXIS_MODE_BUTTONS: { value: XAxisMode; label: string; labelZh: string }[
   { value: 'ttft', label: 'TTFT', labelZh: 'TTFT' },
 ];
 
-/**
- * X-axis modes tucked behind the "Advanced" menu on agentic charts.
- *
- * AgentX headlines E2E Normalized Interactivity, so the three per-request
- * latency views are secondary there and would otherwise crowd the strip. They
- * stay flat top-level tabs on every other scenario: E2E Normalized
- * Interactivity is agentic-only, so collapsing them elsewhere would leave the
- * strip with nothing in it.
- */
-const ADVANCED_X_AXIS_MODES: readonly XAxisMode[] = ['interactivity', 'e2e', 'ttft'];
-
-const isAdvancedXAxisMode = (mode: XAxisMode): boolean => ADVANCED_X_AXIS_MODES.includes(mode);
-
-/**
- * "Advanced" x-axis picker for agentic charts. Styled to sit in the tab strip
- * beside the real tabs, but it is a menu button rather than a `TabsTrigger`:
- * Radix would otherwise treat it as a fifth tab stop and steal arrow-key
- * navigation from the modes inside it. The active mode's label is shown on the
- * trigger so the strip still says which metric the x-axis is plotting.
- */
-function AdvancedXAxisMenu({
-  selected,
-  onSelect,
-  locale,
-  labels,
-}: {
-  selected: XAxisMode;
-  onSelect: (mode: XAxisMode) => void;
-  locale: Locale;
-  labels: { advanced: string; advancedWith: (label: string) => string };
-}) {
-  const [open, setOpen] = useState(false);
-  const active = isAdvancedXAxisMode(selected);
-  const options = X_AXIS_MODE_BUTTONS.filter(({ value }) => isAdvancedXAxisMode(value));
-  const activeLabel = options.find(({ value }) => value === selected);
-  const activeText = activeLabel && (locale === 'zh' ? activeLabel.labelZh : activeLabel.label);
-
-  return (
-    <Popover open={open} onOpenChange={setOpen}>
-      <PopoverTrigger
-        data-testid="x-axis-mode-advanced"
-        data-state={active ? 'active' : 'inactive'}
-        // No aria-label: the accessible name comes from the visible text, so
-        // screen readers announce "Advanced: TTFT" rather than a bare
-        // "Advanced" that hides which metric the x-axis is plotting.
-        className={cn(
-          'relative inline-flex items-center justify-center gap-1.5',
-          'border-b-2 border-transparent px-4 py-2',
-          'text-sm font-semibold whitespace-nowrap',
-          'text-muted-foreground hover:border-muted-foreground/30',
-          'data-[state=active]:text-secondary dark:data-[state=active]:text-primary',
-          'data-[state=active]:border-secondary dark:data-[state=active]:border-primary',
-          'transition-colors duration-200',
-          'focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring',
-          'min-w-[130px] sm:min-w-[140px] flex-1 sm:flex-initial cursor-pointer',
-        )}
-      >
-        {active && activeText ? labels.advancedWith(activeText) : labels.advanced}
-        <ChevronDown
-          className={cn('size-4 transition-transform', open && 'rotate-180')}
-          aria-hidden
-        />
-      </PopoverTrigger>
-      <PopoverContent align="center" className="w-52 p-1" data-testid="x-axis-mode-advanced-menu">
-        <ul className="flex flex-col">
-          {options.map(({ value, label, labelZh }) => {
-            const isActive = value === selected;
-            return (
-              <li key={value}>
-                <button
-                  type="button"
-                  data-testid={`x-axis-mode-${value}`}
-                  aria-current={isActive}
-                  onClick={() => {
-                    setOpen(false);
-                    onSelect(value);
-                  }}
-                  className={cn(
-                    'flex w-full items-center justify-between rounded-sm px-2 py-1.5 text-sm',
-                    'cursor-pointer transition-colors',
-                    isActive
-                      ? 'bg-accent text-secondary dark:text-primary font-medium'
-                      : 'text-muted-foreground hover:bg-accent hover:text-foreground',
-                  )}
-                >
-                  {locale === 'zh' ? labelZh : label}
-                  {isActive && <Check className="size-4" aria-hidden />}
-                </button>
-              </li>
-            );
-          })}
-        </ul>
-      </PopoverContent>
-    </Popover>
-  );
-}
-
 /** Presentation and data plumbing for trace-derived agentic x-axis modes. */
 interface DerivedXModeSpec {
   xLabel: (percentileLabel: string) => string;
@@ -286,41 +207,55 @@ const VIEW_MODE_OPTIONS: SegmentedToggleOption<InferenceViewMode>[] = [
   },
 ];
 
+export function formatTokenLength(value: number): string {
+  const rounded = Math.round(value);
+  if (rounded < 1_000) return String(rounded);
+  if (rounded < 10_000) return `${(rounded / 1_000).toFixed(1).replace(/\.0$/u, '')}k`;
+  if (rounded < 1_000_000) return `${Math.round(rounded / 1_000)}k`;
+  return `${(rounded / 1_000_000).toFixed(2).replace(/\.0+$/u, '')}m`;
+}
+
 /**
  * Renders the inference chart cards, captions, and overlay controls for the current filtered
  * benchmark data.
+ *
+ * `embedded` renders the chart without the header section (title, description,
+ * share actions, and selector controls) — used by the `/model/[slug]` pages,
+ * which seed the model/scenario/metric via providers instead of user-facing
+ * selectors. The run-date changelog strip and the charts themselves remain.
  */
-export default function ChartDisplay() {
+export default function ChartDisplay({ embedded = false }: { embedded?: boolean } = {}) {
   const locale = useLocale();
   const t = STRINGS[locale];
+  const { graphs, loading, error, dateRangeAvailableDates } = useInferenceData();
   const {
-    graphs,
-    loading,
-    error,
-    workflowInfo,
-    selectedYAxisMetric,
-    selectedXAxisMetric,
-    selectedE2eXAxisMetric,
     selectedGPUs,
     selectedPrecisions,
     selectedDates,
-    setSelectedDates,
-    setSelectedDatesFromRunExpansion,
     selectedDateRange,
-    dateRangeAvailableDates,
     selectedModel,
     selectedSequence,
     selectedRunDate,
-    setIsLegendExpanded,
     activeHwTypes,
     bestPerSku,
     activeDates,
-    selectedPercentile,
     compareGpuPair,
+    quickFilters,
+  } = useInferenceFilters();
+  const {
+    selectedYAxisMetric,
+    selectedXAxisMetric,
+    selectedE2eXAxisMetric,
+    selectedPercentile,
     selectedXAxisMode,
+    tokenRevenuePricing,
+  } = useInferenceDisplay();
+  const {
+    setSelectedDates,
+    setSelectedDatesFromRunExpansion,
+    setIsLegendExpanded,
     setSelectedXAxisMode,
-    quickFilters,
-  } = useInference();
+  } = useInferenceActions();
   const selectedBenchmarkType: 'single_turn' | 'agentic_traces' =
     selectedSequence === Sequence.AgenticTraces ? 'agentic_traces' : 'single_turn';
   const workflowInfoBenchmarkType =
@@ -431,9 +366,7 @@ export default function ChartDisplay() {
     getOverlayData,
     isUnofficialRun,
     activeOverlayHwTypes,
-    setActiveOverlayHwTypes,
     localOfficialOverride,
-    setLocalOfficialOverride,
   } = useUnofficialRun();
 
   // Compute overlay data for each chart type — must match useChartData processing
@@ -470,17 +403,22 @@ export default function ChartDisplay() {
 
       const effectiveXMetric = chartType === 'e2e' ? selectedE2eXAxisMetric : selectedXAxisMetric;
       const isAgentic = sequenceKind(selectedSequence) === 'agentic';
+      const tokenType = tokenMetricTypeForConfigKey(selectedYAxisMetric);
+      const pricedData =
+        selectedYAxisMetric === 'y_tokenRevenuePerGpuHour'
+          ? applyTokenRevenuePricing(rawData.data, tokenRevenuePricing)
+          : rawData.data;
+      const capableData = pricedData.filter((point) =>
+        supportsChartTokenMetric(String(point.hwKey), point.date, tokenType),
+      );
       const processed = processOverlayChartDataWithClipping(
-        rawData.data,
+        capableData,
         chartType,
         selectedYAxisMetric,
         effectiveXMetric,
         {
           isAgentic,
           selectedPercentile,
-          // Unofficial rows lack persisted request traces, so they cannot be
-          // admitted to the normalized north-star frontier on any agentic axis.
-          restrictToNormalizedFrontier: isAgentic,
         },
       );
 
@@ -531,6 +469,7 @@ export default function ChartDisplay() {
     selectedE2eXAxisMetric,
     selectedPercentile,
     selectedXAxisMode,
+    tokenRevenuePricing,
     compareGpuPair,
   ]);
 
@@ -601,55 +540,49 @@ export default function ChartDisplay() {
   ]);
   const overlayRowsScopeKey = `${selectedModel}|${selectedSequence}|${selectedPrecisions.join(
     ',',
-  )}|${unofficialRunInfos.map((run) => run.url).join(',')}`;
-  const [appliedOverlayRowsScopeKey, setAppliedOverlayRowsScopeKey] = useState(overlayRowsScopeKey);
-  const overlayRowsScopeChanged =
-    isUnofficialRun && appliedOverlayRowsScopeKey !== overlayRowsScopeKey;
-  const selectedOfficialHwTypes = overlayRowsScopeChanged
-    ? officialScope
-    : isUnofficialRun
-      ? (localOfficialOverride ?? activeHwTypes)
-      : activeHwTypes;
-  // Preview tables follow the same policy as ScatterGraph: preserve every
-  // active engine family instead of applying the production comparison guard.
-  const scopedActiveOverlayHwTypes = useMemo(() => {
-    const activeScopedOverlayKeys = new Set(
-      [...activeOverlayHwTypes].filter((key) => overlayScope.has(key)),
-    );
-    return overlayRowsScopeChanged ? scopedBestSelections.overlay : activeScopedOverlayKeys;
-  }, [activeOverlayHwTypes, overlayScope, overlayRowsScopeChanged, scopedBestSelections.overlay]);
-  useEffect(() => {
-    const merged = new Set(activeOverlayHwTypes);
-    overlayScope.forEach((key) => merged.delete(key));
-    scopedActiveOverlayHwTypes.forEach((key) => merged.add(key));
-    let selectionChanged = merged.size !== activeOverlayHwTypes.size;
-    if (!selectionChanged) {
-      for (const key of merged) {
-        if (!activeOverlayHwTypes.has(key)) {
-          selectionChanged = true;
-          break;
-        }
-      }
-    }
-    if (selectionChanged) setActiveOverlayHwTypes(merged);
-    // A scope change can render once before its official graphs arrive. Do not
-    // persist that transient empty set as an intentional legend selection.
-    if (overlayRowsScopeChanged && (!loading || officialScope.size > 0)) {
-      setLocalOfficialOverride(scopedBestSelections.official);
-      setAppliedOverlayRowsScopeKey(overlayRowsScopeKey);
-    }
-  }, [
-    overlayRowsScopeChanged,
-    overlayRowsScopeKey,
-    activeOverlayHwTypes,
-    loading,
-    officialScope,
-    scopedBestSelections.official,
-    overlayScope,
-    scopedActiveOverlayHwTypes,
-    setActiveOverlayHwTypes,
-    setLocalOfficialOverride,
-  ]);
+  )}|${unofficialRunInfos.map((run) => run.url).join(',')}|official:${[...officialScope]
+    .toSorted()
+    .join(',')}|overlay:${[...overlayScope].toSorted().join(',')}`;
+  const selectedOfficialHwTypes = isUnofficialRun
+    ? (localOfficialOverride ?? activeHwTypes)
+    : activeHwTypes;
+  const scopedActiveOverlayHwTypes = useMemo(
+    () => new Set([...activeOverlayHwTypes].filter((key) => overlayScope.has(key))),
+    [activeOverlayHwTypes, overlayScope],
+  );
+  // Caption spec badges (TCO $/chip/hr, Power/Chip) should only quote chips
+  // that can actually appear on the plot: the active official legend selection
+  // plus any active unofficial-overlay selection.
+  const captionHwKeys = useMemo(
+    () => new Set([...selectedOfficialHwTypes, ...scopedActiveOverlayHwTypes]),
+    [selectedOfficialHwTypes, scopedActiveOverlayHwTypes],
+  );
+  const overlayScopeRegistration = useMemo(
+    () =>
+      isUnofficialRun
+        ? {
+            scopeKey: overlayRowsScopeKey,
+            officialHwTypes: officialScope,
+            overlayHwTypes: overlayScope,
+            bestOfficialHwTypes: scopedBestSelections.official,
+            bestOverlayHwTypes: scopedBestSelections.overlay,
+            bestPerSku,
+            ready: !loading || officialScope.size > 0,
+          }
+        : null,
+    [
+      isUnofficialRun,
+      overlayRowsScopeKey,
+      officialScope,
+      overlayScope,
+      scopedBestSelections,
+      bestPerSku,
+      loading,
+      activeOverlayHwTypes,
+      localOfficialOverride,
+    ],
+  );
+  useOverlayScopeReconciliation(overlayScopeRegistration);
 
   const visibleComparisonRows = useCallback(
     (officialRows: InferenceData[], overlay: OverlayData | null | undefined) => {
@@ -717,11 +650,38 @@ export default function ChartDisplay() {
   }, [effectiveGraphs, selectedXAxisMode]);
 
   const isAgenticSequence = sequenceKind(selectedSequence) === 'agentic';
+  const residentPointIds = useMemo(() => {
+    if (!isAgenticSequence) return [] as number[];
+    const ids = new Set<number>();
+    for (const graph of visibleGraphs) {
+      const points = [...graph.data, ...(graph.clippedData ?? []).map((entry) => entry.point)];
+      for (const point of points) {
+        if (
+          selectedPrecisions.includes(point.precision) &&
+          point.benchmark_type === 'agentic_traces' &&
+          isPersistedBenchmarkId(point.id)
+        ) {
+          ids.add(point.id);
+        }
+      }
+    }
+    return [...ids];
+  }, [isAgenticSequence, selectedPrecisions, visibleGraphs]);
+  // Unofficial-run artifacts are transformed in memory and do not have
+  // persisted aggregate_stats sketches. Suppress the subtitle in overlay mode
+  // rather than presenting official-only values as if they covered the overlay.
+  const residentSequenceLengthsQuery = useResidentSequenceLengths(
+    residentPointIds,
+    isAgenticSequence && !isUnofficialRun,
+  );
+  const residentSequenceLengths =
+    residentSequenceLengthsQuery.data?.coveredPoints ===
+    residentSequenceLengthsQuery.data?.requestedPoints
+      ? residentSequenceLengthsQuery.data
+      : null;
   const useDerivedXAxis = isAgenticSequence && isAgenticOnlyXAxisMode(selectedXAxisMode);
   const derivedTargetIds = useMemo(() => {
-    // Every agentic x-axis is classified by the normalized north-star
-    // frontier, so all modes need the persisted trace-derived metric.
-    if (!isAgenticSequence) return [] as number[];
+    if (!useDerivedXAxis) return [] as number[];
     const ids = new Set<number>();
     for (const graph of visibleGraphs) {
       const points = [...graph.data, ...(graph.clippedData ?? []).map((entry) => entry.point)];
@@ -732,11 +692,11 @@ export default function ChartDisplay() {
       }
     }
     return [...ids];
-  }, [isAgenticSequence, visibleGraphs]);
+  }, [useDerivedXAxis, visibleGraphs]);
   const derivedQuery = useDerivedAgenticMetrics(derivedTargetIds, isAgenticSequence);
   const derivedMetrics = derivedQuery.data;
-  const isCanonicalFrontierLoading =
-    isAgenticSequence &&
+  const isDerivedXAxisLoading =
+    useDerivedXAxis &&
     derivedTargetIds.length > 0 &&
     (derivedQuery.isPending || derivedQuery.isFetching) &&
     !derivedMetrics;
@@ -746,40 +706,25 @@ export default function ChartDisplay() {
     if (!isAgenticSequence) return visibleGraphs;
     if (!derivedMetrics) {
       // Legacy AgentX axes can still render transient/non-persisted rows, which
-      // have no ids to request. Persisted rows remain gated on their derived
-      // metrics so every displayed frontier can enforce canonical eligibility.
+      // have no ids to request.
       if (!derivedSpec && derivedTargetIds.length === 0) return visibleGraphs;
       return visibleGraphs.map((graph) => ({ ...graph, data: [], clippedData: [] }));
     }
     return visibleGraphs.map((graph) => {
       const rooflineKey = `${selectedYAxisMetric}_roofline` as keyof typeof graph.chartDefinition;
-      // The normalized axis is higher-is-better. Compute its true Pareto
-      // direction once, regardless of which x-axis is currently displayed.
       const configuredCorner = graph.chartDefinition[rooflineKey] as RooflineDirection | undefined;
-      const canonicalCorner =
+      const derivedCorner =
         graph.chartDefinition.chartType === 'e2e'
           ? derivedModeRoofline(configuredCorner, true)
           : configuredCorner;
-      const allPoints = [...graph.data, ...(graph.clippedData ?? []).map((entry) => entry.point)];
-      const canonicalIds = canonicalNormalizedFrontierIds(
-        allPoints,
-        derivedMetrics,
-        selectedPercentile,
-        canonicalCorner,
-      );
 
       const preparePoint = (point: InferenceData): InferenceData | null => {
         const pointId = isPersistedBenchmarkId(point.id) ? point.id : null;
-        const stamped = {
-          ...point,
-          isOnNormalizedInteractivityFrontier:
-            canonicalIds === null ? undefined : pointId !== null && canonicalIds.has(pointId),
-        };
-        if (!derivedSpec) return stamped;
+        if (!derivedSpec) return point;
         if (pointId === null) return null;
         const raw = derivedSpec.value(derivedMetrics[pointId], selectedPercentile);
         if (raw === null || raw === undefined || !Number.isFinite(raw)) return null;
-        return { ...stamped, x: derivedSpec.toX(raw) };
+        return { ...point, x: derivedSpec.toX(raw) };
       };
 
       const data = graph.data
@@ -800,7 +745,7 @@ export default function ChartDisplay() {
         ...graph.chartDefinition,
         x_label: xLabelFn(selectedPercentile.toUpperCase()),
         y_latency_limit: undefined,
-        ...(canonicalCorner ? { [rooflineKey]: canonicalCorner } : {}),
+        ...(derivedCorner ? { [rooflineKey]: derivedCorner } : {}),
       };
       return { ...graph, chartDefinition, data, clippedData };
     });
@@ -816,7 +761,7 @@ export default function ChartDisplay() {
   ]);
 
   const displayGraphs =
-    isFirstLoad || isCanonicalFrontierLoading
+    isFirstLoad || isDerivedXAxisLoading
       ? [
           <Card key="skeleton-0">
             <Skeleton className="h-7 w-2/4 mb-1" />
@@ -831,6 +776,62 @@ export default function ChartDisplay() {
               selectedDateRange.startDate && selectedDateRange.endDate && selectedGPUs.length > 0,
             );
             const replayAvailable = getViewMode(graphIndex) === 'chart' && !isTimelineMode;
+            // Which logical metric the x-axis plots right now. Classify off
+            // the field `resolveXAxisField` resolves for this chart's current
+            // state — the same resolver both chart pipelines plot with — so
+            // the footer always matches the drawn axis (e.g. an input metric
+            // without a `*_x` override keeps the natural interactivity x).
+            // Trace-derived agentic modes bypass that resolver, hence the flag.
+            const footerXAxisKind = resolveXAxisKind(graph.chartDefinition.chartType, {
+              xAxisField: resolveXAxisField(
+                graph.chartDefinition,
+                selectedYAxisMetric,
+                graph.chartDefinition.chartType === 'e2e'
+                  ? selectedE2eXAxisMetric
+                  : selectedXAxisMetric,
+                { isAgentic: isAgenticSequence, percentile: selectedPercentile },
+              ).xAxisField,
+              isDerivedNormalizedInteractivity: Boolean(derivedSpec),
+            });
+            // Notices for the axis-metric info footer: the KV-offload halo
+            // key, the agentic optimization note, and the ATOM engine
+            // footnote. Detected from the same data the chart plots —
+            // official points plus any loaded unofficial-run overlay for
+            // this chart type — so they moved out of the legend without
+            // changing when they appear.
+            // GPU/date comparison renders GPUGraph, which plots official
+            // points only — skip the unofficial overlay there so the footer
+            // can't advertise a halo or ATOM series that isn't on the chart.
+            const isGpuComparison =
+              selectedGPUs.length > 0 &&
+              ((selectedDateRange.startDate && selectedDateRange.endDate) ||
+                selectedDates.length > 0);
+            const footerOverlay = isGpuComparison
+              ? undefined
+              : selectUnofficialOverlayForMode(
+                  selectedXAxisMode,
+                  graph.chartDefinition.chartType,
+                  overlayDataByChartType,
+                );
+            const footerPoints = [
+              ...graph.data,
+              ...(footerOverlay?.data ?? []),
+              ...(footerOverlay?.clippedData ?? []).map((entry) => entry.point),
+            ];
+            const hasOffloadHalo = footerPoints.some((point) => point.offload_mode === 'on');
+            const hasAtomSeries = footerPoints.some(
+              (point) =>
+                point.framework !== undefined &&
+                getFrameworkLabel(point.framework).includes(ATOM_FOOTNOTE_MARKER),
+            );
+            const footerNotices =
+              hasOffloadHalo || isAgenticSequence || hasAtomSeries ? (
+                <>
+                  {hasOffloadHalo && <OffloadHaloLegendKey />}
+                  {isAgenticSequence && <AgenticOptimizationNote />}
+                  {hasAtomSeries && <AtomEngineFootnote />}
+                </>
+              ) : undefined;
             return (
               <section key={graphIndex} className="pt-8 md:pt-0">
                 <figure data-testid="chart-figure" className="relative rounded-lg">
@@ -844,13 +845,16 @@ export default function ChartDisplay() {
                           : 'interactivity'
                     }
                     leadingControls={
-                      <SegmentedToggle
-                        value={getViewMode(graphIndex)}
-                        options={viewModeOptions}
-                        onValueChange={(v) => handleViewModeChange(graphIndex, v)}
-                        ariaLabel={t.viewMode}
-                        testId={`inference-view-toggle-${graphIndex}`}
-                      />
+                      <>
+                        <SegmentedToggle
+                          value={getViewMode(graphIndex)}
+                          options={viewModeOptions}
+                          onValueChange={(v) => handleViewModeChange(graphIndex, v)}
+                          ariaLabel={t.viewMode}
+                          testId={`inference-view-toggle-${graphIndex}`}
+                        />
+                        {!embedded && <ShareButton className="h-7" />}
+                      </>
                     }
                     hideImageExport={getViewMode(graphIndex) === 'table'}
                     setIsLegendExpanded={setIsLegendExpanded}
@@ -904,7 +908,7 @@ export default function ChartDisplay() {
                       );
                     }}
                   />
-                  <Card>
+                  <Card data-coach-mark-root="">
                     {(() => {
                       const chartCaption = (
                         <>
@@ -964,12 +968,26 @@ export default function ChartDisplay() {
                             })()}
                           </h2>
                           <p className="text-sm text-muted-foreground mb-2">
+                            <ModelLogo model={graph.model as Model} className="mr-1.5" />
                             {getModelLabel(graph.model as Model)} •{' '}
                             {selectedPrecisions
                               .map((prec) => getPrecisionLabel(prec as Precision))
                               .join(', ')}{' '}
                             • {getSequenceLabel(graph.sequence as Sequence)} •{' '}
                             {isUnofficialRun ? t.sourceUnofficial : t.sourceOfficial}
+                            {selectedYAxisMetric === 'y_tokenRevenuePerGpuHour' &&
+                              tokenRevenuePricing && (
+                                <>
+                                  {' '}
+                                  •{' '}
+                                  <span data-testid="token-revenue-subtitle-prices">
+                                    {t.revenuePrices(
+                                      formatTokenPrice(tokenRevenuePricing.inputPerMillion),
+                                      formatTokenPrice(tokenRevenuePricing.outputPerMillion),
+                                    )}
+                                  </span>
+                                </>
+                              )}
                             {selectedRunDate && (
                               <>
                                 {' '}
@@ -986,7 +1004,32 @@ export default function ChartDisplay() {
                               </>
                             )}
                           </p>
-                          <MetricAssumptionNotes selectedYAxisMetric={selectedYAxisMetric} />
+                          {residentSequenceLengths && (
+                            <p
+                              className="mb-2 text-xs text-muted-foreground"
+                              data-testid="resident-sequence-lengths"
+                            >
+                              {t.completedSequenceLengths(
+                                residentSequenceLengths.isl.n.toLocaleString(
+                                  locale === 'zh' ? 'zh-CN' : 'en-US',
+                                ),
+                              )}{' '}
+                              · ISL p50 {formatTokenLength(residentSequenceLengths.isl.p50)} · p75{' '}
+                              {formatTokenLength(residentSequenceLengths.isl.p75)} · p90{' '}
+                              {formatTokenLength(residentSequenceLengths.isl.p90)} · p95{' '}
+                              {formatTokenLength(residentSequenceLengths.isl.p95)} · p99{' '}
+                              {formatTokenLength(residentSequenceLengths.isl.p99)} | OSL p50{' '}
+                              {formatTokenLength(residentSequenceLengths.osl.p50)} · p75{' '}
+                              {formatTokenLength(residentSequenceLengths.osl.p75)} · p90{' '}
+                              {formatTokenLength(residentSequenceLengths.osl.p90)} · p95{' '}
+                              {formatTokenLength(residentSequenceLengths.osl.p95)} · p99{' '}
+                              {formatTokenLength(residentSequenceLengths.osl.p99)}
+                            </p>
+                          )}
+                          <MetricAssumptionNotes
+                            selectedYAxisMetric={selectedYAxisMetric}
+                            activeHwKeys={captionHwKeys}
+                          />
                           {isUnofficialRun &&
                             selectedXAxisMode === 'e2e-normalized-interactivity' && (
                               <p className="mb-2 text-xs text-muted-foreground">
@@ -1036,9 +1079,7 @@ export default function ChartDisplay() {
                         );
                       }
 
-                      return selectedGPUs.length > 0 &&
-                        ((selectedDateRange.startDate && selectedDateRange.endDate) ||
-                          selectedDates.length > 0) ? (
+                      return isGpuComparison ? (
                         <GPUGraph
                           chartId={`chart-${graphIndex}`}
                           modelLabel={graph.model}
@@ -1068,18 +1109,16 @@ export default function ChartDisplay() {
                               ) ?? undefined
                             }
                           />
-                          {selectedGPUs.length > 0 &&
-                            (!selectedDateRange.startDate || !selectedDateRange.endDate) &&
-                            selectedDates.length === 0 && (
-                              <div className="absolute inset-0 flex items-center justify-center bg-background/60 backdrop-blur-[2px] rounded-lg z-10">
-                                <p className="text-sm font-medium text-muted-foreground bg-background/90 border border-border rounded-md px-4 py-2 shadow-sm">
-                                  {t.selectDateRange}
-                                </p>
-                              </div>
-                            )}
                         </div>
                       );
                     })()}
+                    <AxisMetricFooter
+                      chartId={`chart-${graphIndex}`}
+                      metricKey={selectedYAxisMetric.replace(/^y_/u, '') as MetricKey}
+                      xAxisKind={footerXAxisKind}
+                      xAxisLabel={graph.chartDefinition.x_label}
+                      notices={footerNotices}
+                    />
                     {replayAvailable && (
                       <ReplayLauncher
                         ref={(handle) => {
@@ -1102,16 +1141,26 @@ export default function ChartDisplay() {
       <section className="relative z-20">
         <Card>
           <div className="flex flex-col gap-4">
-            <div className="flex items-start justify-between">
-              <div>
-                <h2 className="text-lg font-semibold mb-2">{t.inferencePerformance}</h2>
-                <p className="text-muted-foreground text-sm mb-4">{t.inferencePerformanceDesc}</p>
-              </div>
-              <ChartShareActions />
-            </div>
-            <ChartControls />
-            <ModelArchitectureDiagram model={selectedModel} />
-            {selectedGPUs.length === 0 && <WorkflowInfoDisplay workflowInfo={workflowInfo} />}
+            {!embedded && (
+              <>
+                <div className="flex items-start justify-between">
+                  <div>
+                    <h2 className="text-lg font-semibold mb-2">{t.inferencePerformance}</h2>
+                    <p className="text-muted-foreground text-sm mb-4">
+                      {t.inferencePerformanceDesc}
+                    </p>
+                  </div>
+                  {/* The chart-row ShareButton lives in ChartButtons, which is desktop-only
+                      (`hidden md:flex`); keep a header Share on mobile so small screens
+                      don't lose the share entry point. */}
+                  <div className="md:hidden">
+                    <ShareButton />
+                  </div>
+                </div>
+                <ChartControls />
+              </>
+            )}
+            {selectedGPUs.length === 0 && <WorkflowInfoDisplay />}
             {selectedGPUs.length > 0 && (
               <ComparisonChangelog
                 changelogs={changelogs}
@@ -1119,6 +1168,7 @@ export default function ChartDisplay() {
                 selectedPrecisions={selectedPrecisions}
                 modelDbKeys={modelDbKeys}
                 selectedSequence={selectedSequence}
+                defaultExpanded={!embedded}
                 loading={changelogsLoading}
                 totalDatesQueried={totalDatesQueried}
                 selectedDates={selectedDates}
@@ -1141,7 +1191,8 @@ export default function ChartDisplay() {
         </Card>
       </section>
 
-      {selectedYAxisMetric === 'y_costUser' && (
+      {(selectedYAxisMetric === 'y_costUser' ||
+        selectedYAxisMetric === 'y_tokensPerDollarUser') && (
         <section>
           <CustomCosts loading={loading} />
         </section>
@@ -1151,16 +1202,7 @@ export default function ChartDisplay() {
           <CustomPowers loading={loading} />
         </section>
       )}
-      {/*
-        Manual activation: with Radix's default automatic mode, merely focusing
-        a trigger fires onValueChange. On agentic the strip renders a single tab
-        while the selected mode may live in the Advanced menu, so tabbing to
-        that lone trigger would silently snap the x-axis back to E2E Normalized
-        Interactivity. Manual activation also suits a control whose every change
-        redraws the chart — arrow keys move focus, Enter/Space commits.
-      */}
       <Tabs
-        activationMode="manual"
         value={selectedXAxisMode}
         onValueChange={(value) => {
           setSelectedXAxisMode(value as XAxisMode);
@@ -1173,32 +1215,41 @@ export default function ChartDisplay() {
           className="flex-wrap justify-center gap-x-1 gap-y-1.5 sm:gap-x-1.5"
         >
           {X_AXIS_MODE_BUTTONS.filter(({ value }) => {
-            // Before mount, render the flat strip so SSR and first client render match.
+            // Before mount, render all buttons so SSR and first client render match.
             if (!mounted) return true;
-            if (isAgenticOnlyXAxisMode(value)) return isAgenticSequence;
-            // On agentic these three move into the Advanced menu below.
-            return !(isAgenticSequence && isAdvancedXAxisMode(value));
-          }).map(({ value, label, labelZh }) => (
-            <TabsTrigger
-              key={value}
-              value={value}
-              data-testid={`x-axis-mode-${value}`}
-              className="min-w-[130px] sm:min-w-[140px] flex-1 sm:flex-initial justify-center"
-            >
-              {locale === 'zh' ? labelZh : label}
-            </TabsTrigger>
-          ))}
-          {mounted && isAgenticSequence && (
-            <AdvancedXAxisMenu
-              selected={selectedXAxisMode}
-              onSelect={(mode) => {
-                setSelectedXAxisMode(mode);
-                track('latency_x_axis_mode_selected', { mode });
-              }}
-              locale={locale}
-              labels={{ advanced: t.advancedXAxis, advancedWith: t.advancedXAxisWith }}
-            />
-          )}
+            return !isAgenticOnlyXAxisMode(value) || isAgenticSequence;
+          }).map(({ value, label, labelZh }) => {
+            const modeLabel = locale === 'zh' ? labelZh : label;
+            if (value !== 'e2e-normalized-interactivity') {
+              return (
+                <TabsTrigger
+                  key={value}
+                  value={value}
+                  data-testid={`x-axis-mode-${value}`}
+                  className="min-w-[130px] sm:min-w-[140px] flex-1 sm:flex-initial justify-center"
+                >
+                  {modeLabel}
+                </TabsTrigger>
+              );
+            }
+
+            return (
+              <span
+                key={value}
+                role="presentation"
+                className="relative flex flex-1 sm:flex-initial"
+              >
+                <TabsTrigger
+                  value={value}
+                  data-testid={`x-axis-mode-${value}`}
+                  className="min-w-[130px] flex-1 justify-center pr-9 sm:min-w-[140px] sm:flex-initial"
+                >
+                  {modeLabel}
+                </TabsTrigger>
+                <NormalizedInteractivityHelpLink locale={locale} />
+              </span>
+            );
+          })}
         </TabsList>
       </Tabs>
       <div className="flex flex-col gap-4">{displayGraphs}</div>
diff --git a/packages/app/src/components/inference/ui/ChartTooltip.tsx b/packages/app/src/components/inference/ui/ChartTooltip.tsx
deleted file mode 100644
index dd9e27ec7..000000000
--- a/packages/app/src/components/inference/ui/ChartTooltip.tsx
+++ /dev/null
@@ -1,66 +0,0 @@
-'use client';
-
-import { useInference } from '@/components/inference/InferenceContext';
-
-interface TooltipContentProps<TValue, _TName> {
-  active?: boolean;
-  payload?: {
-    payload?: {
-      hwKey?: string | number;
-      tp?: number;
-      conc?: number;
-      x?: TValue;
-      y?: TValue;
-      [key: string]: unknown;
-    };
-    [key: string]: unknown;
-  }[];
-  [key: string]: unknown;
-}
-
-export default function ChartTooltip({ active, payload }: TooltipContentProps<number, string>) {
-  const { hardwareConfig } = useInference();
-  const pointPayload = payload?.at(-1)?.payload;
-  if (active && pointPayload) {
-    return (
-      <div className="bg-accent p-2 border rounded-sm">
-        <p>{`Chip: ${hardwareConfig[pointPayload.hwKey as keyof typeof hardwareConfig].gpu}`}</p>
-        <p>{`Total Chips: ${pointPayload.tp}`}</p>
-        <p>{`Deployment: ${
-          pointPayload.disagg
-            ? 'Disaggregated'
-            : pointPayload.is_multinode
-              ? 'Multi-node aggregate'
-              : 'Single-node aggregate'
-        }`}</p>
-        {(pointPayload.ep !== null && pointPayload.ep !== undefined) ||
-        (pointPayload.prefill_ep !== null && pointPayload.prefill_ep !== undefined) ? (
-          pointPayload.is_multinode && pointPayload.disagg ? (
-            <>
-              <p>{`Prefill: ${pointPayload.num_prefill_gpu ?? '?'} Chips, TP: ${pointPayload.prefill_tp ?? pointPayload.tp}, EP: ${pointPayload.prefill_ep ?? pointPayload.ep ?? 0}, DPA: ${(pointPayload.prefill_dp_attention ?? pointPayload.dp_attention) ? 'True' : 'False'}, Workers: ${pointPayload.prefill_num_workers ?? 1}`}</p>
-              <p>{`Decode: ${pointPayload.num_decode_gpu ?? '?'} Chips, TP: ${pointPayload.decode_tp ?? pointPayload.tp}, EP: ${pointPayload.decode_ep ?? pointPayload.ep ?? 0}, DPA: ${(pointPayload.decode_dp_attention ?? pointPayload.dp_attention) ? 'True' : 'False'}, Workers: ${pointPayload.decode_num_workers ?? 1}`}</p>
-            </>
-          ) : (
-            <>
-              <p>{`Tensor Parallelism: ${pointPayload.decode_tp ?? pointPayload.tp}`}</p>
-              {typeof pointPayload.pp === 'number' && pointPayload.pp > 1 && (
-                <p>{`Pipeline Parallelism: ${pointPayload.pp}`}</p>
-              )}
-              {pointPayload.ep !== null && pointPayload.ep !== undefined && (
-                <p>{`Expert Parallelism: ${pointPayload.ep}`}</p>
-              )}
-              <p>{`DP Attention: ${pointPayload.dp_attention ? 'True' : 'False'}`}</p>
-            </>
-          )
-        ) : (
-          <p>{`Parallelism Strategy: ${pointPayload.tp} Chip${(pointPayload.tp ?? 0) > 1 ? 's' : ''}`}</p>
-        )}
-        <p>{`Concurrent Requests: ${pointPayload.conc} Users`}</p>
-        <p>{`X: ${pointPayload.x}`}</p>
-        <p>{`Y: ${pointPayload.y}`}</p>
-      </div>
-    );
-  }
-
-  return null;
-}
diff --git a/packages/app/src/components/inference/ui/ComparisonChangelog.tsx b/packages/app/src/components/inference/ui/ComparisonChangelog.tsx
index f87a8084c..2075f8ced 100644
--- a/packages/app/src/components/inference/ui/ComparisonChangelog.tsx
+++ b/packages/app/src/components/inference/ui/ComparisonChangelog.tsx
@@ -81,6 +81,8 @@ interface ComparisonChangelogProps {
   onAddDate: (date: string) => void;
   onRemoveDate: (date: string) => void;
   onAddAllDates: (dates: string[]) => void;
+  /** Initial expanded state of the changelog body (collapsed in embedded contexts). */
+  defaultExpanded?: boolean;
   /** Earliest date the selected GPU config has benchmark data */
   firstAvailableDate?: string;
 }
@@ -99,8 +101,9 @@ export default function ComparisonChangelog({
   onRemoveDate,
   onAddAllDates,
   firstAvailableDate,
+  defaultExpanded = true,
 }: ComparisonChangelogProps) {
-  const [isExpanded, setIsExpanded] = useState(true);
+  const [isExpanded, setIsExpanded] = useState(defaultExpanded);
   const benchmarkType =
     selectedSequence === Sequence.AgenticTraces ? 'agentic_traces' : 'single_turn';
   const changelogBenchmarkType =
diff --git a/packages/app/src/components/inference/ui/CustomGpuValuePanel.tsx b/packages/app/src/components/inference/ui/CustomGpuValuePanel.tsx
index 3e934922a..4337173a7 100644
--- a/packages/app/src/components/inference/ui/CustomGpuValuePanel.tsx
+++ b/packages/app/src/components/inference/ui/CustomGpuValuePanel.tsx
@@ -2,7 +2,11 @@
 
 import React, { memo, useCallback, useEffect, useRef, useState } from 'react';
 
-import { useInference } from '@/components/inference/InferenceContext';
+import {
+  useInferenceActions,
+  useInferenceDisplay,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
 import {
   buildAppliedCustomGpuValues,
   buildDefaultCustomGpuValues,
@@ -20,14 +24,34 @@ import {
 } from '@/components/ui/input-group';
 import { Skeleton } from '@/components/ui/skeleton';
 import { track } from '@/lib/analytics';
+import { useLocale } from '@/lib/use-locale';
 import { HW_REGISTRY, type HwEntry } from '@semianalysisai/inferencex-constants';
 type GpuValuePanelKind = 'costs' | 'powers';
 
+const STRINGS = {
+  en: {
+    reset: 'Reset',
+    resetToDefaults: 'Reset to defaults',
+    calculate: 'Calculate',
+    invalidNumber: 'Must be a valid number',
+    nonNegative: 'Must be a non-negative number',
+  },
+  zh: {
+    reset: '重置',
+    resetToDefaults: '恢复默认值',
+    calculate: '计算',
+    invalidNumber: '请输入有效数字',
+    nonNegative: '请输入非负数',
+  },
+} as const;
+
 const PANEL_CONFIG: Record<
   GpuValuePanelKind,
   {
     title: string;
+    titleZh: string;
     description: string;
+    descriptionZh: string;
     sectionTestId: string;
     calculateTestId: string;
     inputIdPrefix: string;
@@ -38,8 +62,11 @@ const PANEL_CONFIG: Record<
 > = {
   costs: {
     title: 'Custom Chip Costs',
+    titleZh: '自定义芯片成本',
     description:
-      'Enter your own TCO (Total Cost of Ownership) values for each chip in $/chip/hr. These values will be used to calculate custom cost metrics.',
+      'Enter your own TCO (Total Cost of Ownership) values for each chip in $/chip/hr. These values determine the selected token cost metric.',
+    descriptionZh:
+      '输入每个芯片的自定义 TCO（总拥有成本），单位为 $/chip/hr。这些值用于计算所选 token 成本指标。',
     sectionTestId: 'custom-costs-section',
     calculateTestId: 'custom-costs-calculate',
     inputIdPrefix: 'cost-input',
@@ -49,8 +76,11 @@ const PANEL_CONFIG: Record<
   },
   powers: {
     title: 'Custom Chip Powers',
+    titleZh: '自定义芯片功耗',
     description:
       'Enter your own Token Throughput per All in Utility MW (tok/s/MW) values for each chip. These values will be used to calculate custom power metrics.',
+    descriptionZh:
+      '输入每个芯片的自定义全电源配置兆瓦 token 吞吐量（tok/s/MW）。这些值用于计算自定义功耗指标。',
     sectionTestId: 'custom-powers-section',
     calculateTestId: 'custom-powers-calculate',
     // Preserve legacy input IDs so existing Cypress selectors keep passing.
@@ -124,16 +154,15 @@ function renderSkeleton(title: string, description: string) {
 
 const CustomGpuValuePanel = memo(
   ({ loading, kind }: { loading: boolean; kind: GpuValuePanelKind }) => {
-    const {
-      selectedYAxisMetric,
-      selectedPrecisions,
-      selectedModel,
-      selectedSequence,
-      setUserCosts,
-      setUserPowers,
-    } = useInference();
+    const locale = useLocale();
+    const t = STRINGS[locale];
+    const { selectedYAxisMetric } = useInferenceDisplay();
+    const { selectedPrecisions, selectedModel, selectedSequence } = useInferenceFilters();
+    const { setUserCosts, setUserPowers } = useInferenceActions();
 
     const config = PANEL_CONFIG[kind];
+    const title = locale === 'zh' ? config.titleZh : config.title;
+    const description = locale === 'zh' ? config.descriptionZh : config.description;
     const applyValues = kind === 'costs' ? setUserCosts : setUserPowers;
 
     const [inputErrors, setInputErrors] = useState<Record<string, string>>({});
@@ -186,18 +215,27 @@ const CustomGpuValuePanel = memo(
       }
     }, [defaultValues, selectedModel, selectedPrecisions, selectedSequence, selectedYAxisMetric]);
 
-    const handleInputChange = useCallback((gpuKey: string, value: string) => {
-      const validationError = validateCustomGpuValueInput(value);
+    const handleInputChange = useCallback(
+      (gpuKey: string, value: string) => {
+        const validationError = validateCustomGpuValueInput(value);
+        const localizedError =
+          locale === 'zh' && validationError === STRINGS.en.invalidNumber
+            ? t.invalidNumber
+            : locale === 'zh' && validationError === STRINGS.en.nonNegative
+              ? t.nonNegative
+              : validationError;
 
-      setInputErrors((prev) => ({
-        ...prev,
-        [gpuKey]: validationError,
-      }));
-      setLastCalculatedValues((prev) => ({
-        ...prev,
-        [gpuKey]: value,
-      }));
-    }, []);
+        setInputErrors((prev) => ({
+          ...prev,
+          [gpuKey]: localizedError,
+        }));
+        setLastCalculatedValues((prev) => ({
+          ...prev,
+          [gpuKey]: value,
+        }));
+      },
+      [locale, t.invalidNumber, t.nonNegative],
+    );
 
     const handleReset = useCallback(() => {
       track(config.resetEvent, {
@@ -226,13 +264,13 @@ const CustomGpuValuePanel = memo(
     }, [applyValues, config, inputErrors, lastCalculatedValues, selectedYAxisMetric, stableGpus]);
 
     if (loading || stableGpus.length === 0) {
-      return renderSkeleton(config.title, config.description);
+      return renderSkeleton(title, description);
     }
 
     return (
       <Card data-testid={config.sectionTestId}>
-        <h2 className="text-lg font-semibold mb-2">{config.title}</h2>
-        <p className="text-muted-foreground text-sm mb-4">{config.description}</p>
+        <h2 className="text-lg font-semibold mb-2">{title}</h2>
+        <p className="text-muted-foreground text-sm mb-4">{description}</p>
 
         <div className="flex flex-col gap-4">
           <div className="flex-1 grid grid-cols-2 md:grid-cols-3 gap-4">
@@ -254,13 +292,13 @@ const CustomGpuValuePanel = memo(
             <Button
               onClick={handleReset}
               variant="ghost"
-              aria-label="Reset to defaults"
-              title="Reset to defaults"
+              aria-label={t.resetToDefaults}
+              title={t.resetToDefaults}
             >
-              Reset
+              {t.reset}
             </Button>
             <Button data-testid={config.calculateTestId} onClick={handleRecalculate}>
-              Calculate
+              {t.calculate}
             </Button>
           </div>
         </div>
diff --git a/packages/app/src/components/inference/ui/EnlargedHitAreaDot.tsx b/packages/app/src/components/inference/ui/EnlargedHitAreaDot.tsx
deleted file mode 100644
index 14c2bfd6a..000000000
--- a/packages/app/src/components/inference/ui/EnlargedHitAreaDot.tsx
+++ /dev/null
@@ -1,41 +0,0 @@
-import React from 'react';
-
-interface DotProps {
-  cx?: number;
-  cy?: number;
-  fill?: string;
-  fillOpacity?: number;
-  stroke?: string;
-  strokeWidth?: number;
-  r?: number;
-  [key: string]: unknown;
-}
-
-const EnlargedHitAreaDot = (props: DotProps) => {
-  const { cx, cy, fill, fillOpacity, stroke, strokeWidth, r = 3 } = props;
-
-  // Recharts can pass NaN or undefined coordinates during transitions; this prevents errors.
-  // Always return a <g> element to satisfy Recharts' type expectation.
-  return (
-    <g>
-      {typeof cx === 'number' && typeof cy === 'number' && (
-        <>
-          {/* The larger, transparent circle for an increased hit area */}
-          <circle cx={cx} cy={cy} r={10} fill="transparent" />
-          {/* The original, visible dot */}
-          <circle
-            cx={cx}
-            cy={cy}
-            r={r}
-            fill={fill}
-            fillOpacity={fillOpacity}
-            stroke={stroke}
-            strokeWidth={strokeWidth}
-          />
-        </>
-      )}
-    </g>
-  );
-};
-
-export default EnlargedHitAreaDot;
diff --git a/packages/app/src/components/inference/ui/GPUGraph.tsx b/packages/app/src/components/inference/ui/GPUGraph.tsx
index 1171ab66b..1ff7c8306 100644
--- a/packages/app/src/components/inference/ui/GPUGraph.tsx
+++ b/packages/app/src/components/inference/ui/GPUGraph.tsx
@@ -1,19 +1,31 @@
 'use client';
 
 import { track } from '@/lib/analytics';
+import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
+import { useEphemeralUrlState } from '@/hooks/useUrlState';
+import { rememberChartStateInUrl } from '@/lib/url-state';
 import * as d3 from 'd3';
-import React, { useCallback, useEffect, useMemo, useRef } from 'react';
+import dynamic from 'next/dynamic';
+import React, { useCallback, useEffect, useMemo, useRef, useState } from 'react';
 import { useTheme } from 'next-themes';
 
-import { useInference } from '@/components/inference/InferenceContext';
+import {
+  useInferenceActions,
+  useInferenceData,
+  useInferenceDisplay,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
 import ChartLegend from '@/components/ui/chart-legend';
-import { getHardwareConfig, getModelSortIndex } from '@/lib/constants';
+import { Button } from '@/components/ui/button';
+import { OFFICIAL_PREVIEW_SERIES } from '@/components/official-preview-notice';
+import { getHardwareConfig, getModelSortIndex, hardwareKeyMatchesAnyBase } from '@/lib/constants';
 import { getChartWatermark, Sequence } from '@/lib/data-mappings';
-import { generateGpuDateColors } from '@/lib/dynamic-colors';
+import { generateGpuDateColors, generateHighContrastGpuDateColors } from '@/lib/dynamic-colors';
 import { useLocale } from '@/lib/use-locale';
 import { formatNumber, getDisplayLabel, updateRepoUrl } from '@/lib/utils';
 import { useThemeColors } from '@/hooks/useThemeColors';
 import { useTraceAvailability } from '@/hooks/api/use-trace-availability';
+import { useLogAvailability } from '@/hooks/api/use-log-availability';
 import { D3Chart } from '@/lib/d3-chart/D3Chart';
 import type {
   CustomLayerConfig,
@@ -46,23 +58,30 @@ import {
   comparisonEntrySortValue,
   resolveComparisonEntries,
 } from '@/components/inference/utils/comparisonEntry';
-import {
-  generateGPUGraphTooltipContent,
-  getPointLabel,
-} from '@/components/inference/utils/tooltipUtils';
+import { generateGPUGraphTooltipContent } from '@/components/inference/utils/tooltipUtils';
+import { pointLabelText } from '@/components/inference/ui/point-label';
+import { scatterPointConfigId } from '@/components/inference/utils/point-identity';
 import {
   type KnownIssueAnnotation,
-  measureLegendRightInset,
-  renderKnownIssueAnnotations,
+  createKnownIssueLayer,
 } from '@/components/inference/utils/knownIssueAnnotations';
 import { matchKnownConfigIssues, pointMatchesIssue } from '@/lib/known-issues';
+import { renderOffloadHalo } from '@/components/inference/utils/offload-halo';
 import {
-  OFFLOAD_HALO_DASHARRAY,
-  OFFLOAD_HALO_RADIUS,
-  OFFLOAD_HALO_STROKE_WIDTH,
-  OffloadHaloLegendKey,
-} from '@/components/inference/ui/OffloadHaloLegendKey';
-import { AgenticOptimizationNote } from '@/components/inference/ui/AgenticOptimizationNote';
+  parallelismLabelBoxes,
+  placeLineLabels,
+  placeEndpointLineLabels,
+  renderLineLabels,
+  updateRenderedLineLabels,
+  type LineLabelSeries,
+} from '@/components/inference/ui/line-label-layer';
+import { QuickFiltersDialog } from '@/components/inference/ui/QuickFiltersDialog';
+
+const FixedSequenceLogDialog = dynamic(() =>
+  import('@/components/inference/log-viewer/fixed-sequence-log-dialog').then(
+    (module) => module.FixedSequenceLogDialog,
+  ),
+);
 
 const CHART_MARGIN = { top: 24, right: 10, bottom: 60, left: 60 };
 
@@ -84,8 +103,10 @@ const GPU_STRINGS = {
     optimalOnly: 'Optimal Only',
     labels: 'Labels',
     parallelismLabels: 'Parallelism Labels',
+    concurrencyLabels: '# Concurrent Sessions',
     lineLabels: 'Line Labels',
     resetFilter: 'Reset filter',
+    quickFilters: (count: number) => (count > 0 ? `Quick Filters (${count})` : 'Quick Filters'),
   },
   zh: {
     logScale: '对数缩放',
@@ -93,8 +114,10 @@ const GPU_STRINGS = {
     optimalOnly: '仅最优',
     labels: '标签',
     parallelismLabels: '并行配置标签',
+    concurrencyLabels: '并发会话数',
     lineLabels: '曲线标签',
     resetFilter: '重置筛选',
+    quickFilters: (count: number) => (count > 0 ? `快捷筛选（${count}）` : '快捷筛选'),
   },
 } as const;
 
@@ -109,39 +132,67 @@ const GPUGraph = React.memo(
     caption,
     runNumbering: providedRunNumbering,
   }: ScatterGraphProps) => {
+    const { hardwareConfig } = useInferenceData();
     const {
-      hardwareConfig,
       selectedPrecisions,
-      selectedYAxisMetric,
       selectedGPUs,
       selectedDateRange,
       selectedDates,
       selectedSequence,
+      quickFilters,
+      activeDates,
+    } = useInferenceFilters();
+    const {
+      selectedYAxisMetric,
+      hideNonOptimal,
+      showPointLabels,
+      logScale,
+      isLegendExpanded,
+      useAdvancedLabels,
+      showConcurrencyLabels,
+      highContrast,
+      showLineLabels,
+    } = useInferenceDisplay();
+    const {
       setSelectedDates,
       toggleActiveDate,
       removeActiveDate,
-      activeDates,
-      hideNonOptimal,
       setHideNonOptimal,
-      showPointLabels,
       setShowPointLabels,
-      logScale,
       setLogScale,
-      isLegendExpanded,
       setIsLegendExpanded,
-      useAdvancedLabels,
       setUseAdvancedLabels,
-      highContrast,
+      setShowConcurrencyLabels,
       setHighContrast,
       selectAllActiveDates,
-      showLineLabels,
       setShowLineLabels,
-    } = useInference();
+      setQuickFilterVendors,
+      setQuickFilterFrameworks,
+      setQuickFilterDeployment,
+      setQuickFilterSpec,
+    } = useInferenceActions();
     const locale = useLocale();
     const legendT = GPU_STRINGS[locale];
+    const ephemeralUrlState = useEphemeralUrlState();
     const { resolvedTheme } = useTheme();
     const chartRef = useRef<D3ChartHandle>(null);
-    const hasOffloadHalo = useMemo(() => data.some((point) => point.offload_mode === 'on'), [data]);
+    const [quickFiltersOpen, setQuickFiltersOpen] = useState(false);
+    const quickFilterCount =
+      quickFilters.vendors.length +
+      quickFilters.frameworks.length +
+      quickFilters.deployment.length +
+      (selectedSequence === Sequence.AgenticTraces ? 0 : quickFilters.spec.length);
+    const clearQuickFilters = useCallback(() => {
+      setQuickFilterVendors([]);
+      setQuickFilterFrameworks([]);
+      setQuickFilterDeployment([]);
+      setQuickFilterSpec([]);
+    }, [
+      setQuickFilterVendors,
+      setQuickFilterFrameworks,
+      setQuickFilterDeployment,
+      setQuickFilterSpec,
+    ]);
 
     // Shared date+GPU pairs. `dates` holds comparison-series entries (plain dates
     // and/or specific-run entries); a same-day range endpoint is dropped when that
@@ -193,9 +244,13 @@ const GPUGraph = React.memo(
       return ids;
     }, [gpuDatePairs]);
 
+    // High contrast keys off the GPU (not `date_gpu`) so each hardware config
+    // gets exactly one hue; the dates within a config are separated by the
+    // lightness ramp built below rather than by unrelated hues.
     const { resolveColor, getCssColor } = useThemeColors({
       highContrast,
       identifiers: graphIdentifiers,
+      hcKeys: gpuDatePairs.sortedGPUs,
     });
 
     // Dynamic GPU×date color map
@@ -206,25 +261,48 @@ const GPUGraph = React.memo(
       return generateGpuDateColors(sortedGPUs, dates.length, theme);
     }, [gpuDatePairs, resolvedTheme]);
 
+    // High-contrast GPU×date color map: one iwanthue hue per GPU, ramped across
+    // the compared dates so a config's runs stay recognisably the same color
+    // while still reading oldest → newest.
+    const hcGpuDateColorMap = useMemo(() => {
+      const { dates, sortedGPUs } = gpuDatePairs;
+      if (!highContrast || sortedGPUs.length === 0 || dates.length === 0) return {};
+      const theme = resolvedTheme === 'dark' || resolvedTheme === 'minecraft' ? 'dark' : 'light';
+      const baseColors: Record<string, string> = {};
+      for (const gpu of sortedGPUs) baseColors[gpu] = getCssColor(resolveColor(gpu));
+      return generateHighContrastGpuDateColors(baseColors, dates.length, theme);
+    }, [gpuDatePairs, highContrast, resolvedTheme, resolveColor, getCssColor]);
+
     const allGraphs = useMemo(() => {
       const { dates, sortedGPUs } = gpuDatePairs;
       const result: { date: string; color: string; hwKey: string; id: string }[] = [];
       sortedGPUs.forEach((gpu) => {
         dates.forEach((date, dateIndex) => {
           const id = `${date}_${gpu}`;
-          const dynamicColor = gpuDateColorMap[`${dateIndex}_${gpu}`];
+          const compositeKey = `${dateIndex}_${gpu}`;
+          const dynamicColor = gpuDateColorMap[compositeKey];
           result.push({
             date,
             hwKey: gpu,
             id,
             color: highContrast
-              ? getCssColor(resolveColor(id))
+              ? hcGpuDateColorMap[compositeKey] || getCssColor(resolveColor(gpu))
               : dynamicColor || 'var(--foreground)',
           });
         });
       });
       return result;
-    }, [gpuDatePairs, gpuDateColorMap, highContrast, resolveColor, getCssColor]);
+    }, [gpuDatePairs, gpuDateColorMap, hcGpuDateColorMap, highContrast, resolveColor, getCssColor]);
+
+    const paletteIdentity = useMemo(
+      () =>
+        [
+          resolvedTheme ?? 'system',
+          highContrast ? 'high-contrast' : 'standard',
+          ...allGraphs.map(({ id, color }) => `${id}:${color}`),
+        ].join('|'),
+      [resolvedTheme, highContrast, allGraphs],
+    );
 
     const groupedData = useMemo(
       () =>
@@ -291,78 +369,94 @@ const GPUGraph = React.memo(
     const agenticIds = useMemo(
       () =>
         filteredData.flatMap((point) =>
-          point.benchmark_type === 'agentic_traces' && typeof point.id === 'number'
+          point.benchmark_type === 'agentic_traces' && isPersistedBenchmarkId(point.id)
             ? [point.id]
             : [],
         ),
       [filteredData],
     );
     const { data: traceAvailability } = useTraceAvailability(agenticIds);
+    const traceAvailabilityRef = useRef(traceAvailability);
+    traceAvailabilityRef.current = traceAvailability;
+
+    // Log availability applies to every persisted official point in the
+    // comparison, including fixed-sequence runs.
+    const persistedPointIds = useMemo(
+      () => filteredData.flatMap((point) => (isPersistedBenchmarkId(point.id) ? [point.id] : [])),
+      [filteredData],
+    );
+    const { data: logAvailability } = useLogAvailability(persistedPointIds);
+    const logAvailabilityRef = useRef(logAvailability);
+    logAvailabilityRef.current = logAvailability;
+    const [fixedLogPointId, setFixedLogPointId] = useState<number | null>(null);
 
     // Warning annotations for visible series with known upstream issues —
     // same treatment the scatter view gets, applied to the date-comparison view.
     // Lines here are colored per (gpu, date) pair, so take the first active
     // pair's color as the series swatch.
-    const knownIssueAnnotations = useMemo(
-      (): KnownIssueAnnotation[] =>
-        matchKnownConfigIssues(modelLabel, filteredData).map((issue) => {
-          const cfg = getHardwareConfig(issue.hwKey, modelLabel);
-          const colorEntry = allGraphs.find(
-            (entry) => entry.hwKey === issue.hwKey && activeDates.has(entry.id),
-          );
-          return {
-            issue,
-            label: cfg ? getDisplayLabel(cfg) : issue.hwKey,
-            color: getCssColor(colorEntry?.color ?? resolveColor(issue.hwKey)),
-            points: filteredData
-              .filter((p) => pointMatchesIssue(issue, p))
-              .map((p) => ({ x: p.x, y: p.y })),
-          };
-        }),
-      [modelLabel, filteredData, allGraphs, activeDates, resolveColor, getCssColor],
-    );
+    const knownIssueAnnotations = useMemo((): KnownIssueAnnotation[] => {
+      const annotations: KnownIssueAnnotation[] = matchKnownConfigIssues(
+        modelLabel,
+        filteredData,
+      ).map((issue) => {
+        const cfg = getHardwareConfig(issue.hwKey, modelLabel);
+        const colorEntry = allGraphs.find(
+          (entry) => entry.hwKey === issue.hwKey && activeDates.has(entry.id),
+        );
+        return {
+          issue,
+          label: cfg ? getDisplayLabel(cfg) : issue.hwKey,
+          color: getCssColor(colorEntry?.color ?? resolveColor(issue.hwKey)),
+          points: filteredData
+            .filter((p) => pointMatchesIssue(issue, p))
+            .map((p) => ({ x: p.x, y: p.y })),
+        };
+      });
+      for (const previewConfig of OFFICIAL_PREVIEW_SERIES) {
+        const previewPoints = filteredData.filter((point) =>
+          hardwareKeyMatchesAnyBase(String(point.hwKey), previewConfig.baseGpuKeys),
+        );
+        if (previewPoints.length === 0) continue;
+
+        const hwKey = String(previewPoints[0]!.hwKey);
+        const colorEntry = allGraphs.find(
+          (entry) => entry.hwKey === hwKey && activeDates.has(entry.id),
+        );
+        const previewCopy = previewConfig.strings[locale];
+        annotations.push({
+          preview: {
+            id: previewConfig.id,
+            summary: previewCopy.title,
+            detail: previewCopy.chartDetail,
+          },
+          label: getDisplayLabel(getHardwareConfig(hwKey, modelLabel)),
+          color: getCssColor(colorEntry?.color ?? resolveColor(hwKey)),
+          points: previewPoints.map((point) => ({ x: point.x, y: point.y })),
+        });
+      }
+      return annotations;
+    }, [modelLabel, filteredData, allGraphs, activeDates, resolveColor, getCssColor, locale]);
 
-    const drawKnownIssues = (
-      ctx: RenderContext,
-      xScale: ContinuousScale,
-      yScale: ContinuousScale,
-    ) => {
-      renderKnownIssueAnnotations(ctx.layout.g, ctx.layout.defs, {
-        chartId,
-        width: ctx.width,
-        height: ctx.height,
-        xScale,
-        yScale,
-        annotations: knownIssueAnnotations,
-        // Only measure the legend overlap when there are boxes to place —
-        // this runs on every zoom frame, and the measurement forces layout.
-        rightInset:
-          knownIssueAnnotations.length === 0
-            ? 0
-            : measureLegendRightInset(
-                chartId,
-                ctx.layout.svg.node(),
-                ctx.layout.margin.left,
-                ctx.width,
-              ),
-        background: getCssColor('--background'),
-        foreground: getCssColor('--foreground'),
-        mutedForeground: getCssColor('--muted-foreground'),
-        onLinkClick: (a) =>
-          track('inference_known_issue_clicked', {
-            hwKey: a.issue.hwKey,
-            issue: a.issue.issueRef,
+    const knownIssueLayer = useMemo(
+      () =>
+        createKnownIssueLayer(
+          () => ({
+            chartId,
+            annotations: knownIssueAnnotations,
+            background: getCssColor('--background'),
+            foreground: getCssColor('--foreground'),
+            mutedForeground: getCssColor('--muted-foreground'),
+            onLinkClick: (annotation) =>
+              annotation.issue &&
+              track('inference_known_issue_clicked', {
+                hwKey: annotation.issue.hwKey,
+                issue: annotation.issue.issueRef,
+              }),
           }),
-      });
-    };
-    const knownIssueLayer: CustomLayerConfig = {
-      type: 'custom',
-      key: 'known-issues',
-      render: (_zoomGroup, ctx) =>
-        drawKnownIssues(ctx, ctx.xScale as ContinuousScale, ctx.yScale as ContinuousScale),
-      onZoom: (_zoomGroup, ctx) =>
-        drawKnownIssues(ctx, ctx.newXScale as ContinuousScale, ctx.newYScale as ContinuousScale),
-    };
+          paletteIdentity,
+        ),
+      [chartId, knownIssueAnnotations, getCssColor, paletteIdentity],
+    );
 
     // Compute scale domains
     const xExtent = useMemo(() => {
@@ -386,6 +480,42 @@ const GPUGraph = React.memo(
       return [yMin, yExtent[1] * 1.05] as [number, number];
     }, [filteredData, logScale]);
 
+    const dataIdentity = useMemo(
+      () =>
+        filteredData
+          .map((point) => `${point.date}:${scatterPointConfigId(point)}`)
+          .toSorted()
+          .join('|'),
+      [filteredData],
+    );
+    // Tooltip-only trace availability is deliberately excluded from chart
+    // identity; the long-lived D3 content callback reads its latest value via
+    // traceAvailabilityRef.
+    const metricIdentity = useMemo(
+      () =>
+        [
+          useAdvancedLabels ? 'advanced-labels' : 'basic-labels',
+          showConcurrencyLabels ? 'conc-labels' : 'no-conc-labels',
+          selectedYAxisMetric,
+          `linear:${xExtent.join(',')}`,
+          `${logScale ? 'log' : 'linear'}:${yDomain.join(',')}`,
+          ...filteredData.map(
+            (point) => `${point.date}:${scatterPointConfigId(point)}:${point.x}:${point.y}`,
+          ),
+        ]
+          .toSorted()
+          .join('|'),
+      [
+        selectedYAxisMetric,
+        useAdvancedLabels,
+        showConcurrencyLabels,
+        xExtent,
+        logScale,
+        yDomain,
+        filteredData,
+      ],
+    );
+
     // Color resolver for points/rooflines
     const getColor = useMemo(
       () => (d: InferenceData) => {
@@ -419,285 +549,94 @@ const GPUGraph = React.memo(
     // same combo dedupe down to the longest roofline so the label rides the
     // line that has the most placement options. Labels track the active filter
     // (`activeDates`) so removing a series via the legend hides its label too.
-    const lineLabelLayer: CustomLayerConfig = useMemo(
-      () => ({
+    const lineLabelLayer: CustomLayerConfig = useMemo(() => {
+      const buildSeries = (): LineLabelSeries<InferenceData>[] => {
+        const bestByGraph = new Map<
+          string,
+          { key: string; graphId: string; points: InferenceData[] }
+        >();
+        for (const [key, points] of Object.entries(rooflines)) {
+          if (points.length < 2 || !isRooflineVisible(key)) continue;
+          const graphId = key.slice(0, key.lastIndexOf('_'));
+          const previous = bestByGraph.get(graphId);
+          if (!previous || points.length > previous.points.length) {
+            bestByGraph.set(graphId, { key, graphId, points });
+          }
+        }
+        return [...bestByGraph.values()].map(({ key, graphId, points }) => ({
+          key,
+          seriesId: graphId,
+          label: labelTextFor(points, runNumbering),
+          color: getRooflineColor(key),
+          points,
+        }));
+      };
+      const placeLabels = (
+        xScale: ContinuousScale,
+        yScale: ContinuousScale,
+        zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+      ) => {
+        if (!showLineLabels) return [];
+        const series = buildSeries();
+        return chartDefinition.chartType === 'interactivity'
+          ? placeLineLabels(series, xScale, yScale, {
+              collisionWidth: 160,
+              obstacles: parallelismLabelBoxes(zoomGroup.node()),
+            })
+          : placeEndpointLineLabels(series, xScale, yScale);
+      };
+
+      return {
         type: 'custom',
         key: 'line-labels',
+        displayIdentity: `${showLineLabels ? 'visible' : 'hidden'}:${paletteIdentity}`,
         render: (zoomGroup, ctx) => {
-          // Always run the data-join so toggling the switch off cleans the DOM.
-          interface LineLabel {
-            key: string;
-            graphId: string;
-            label: string;
-            color: string;
-            x: number;
-            y: number;
-            visible: boolean;
-          }
-
-          if (!showLineLabels) {
-            zoomGroup.selectAll('.line-label').remove();
-            return;
-          }
-
-          const xScale = ctx.xScale as ContinuousScale;
-          const yScale = ctx.yScale as ContinuousScale;
-          const isInteractivity = chartDefinition.chartType === 'interactivity';
-          const LABEL_H = 18;
-          const LABEL_W = 160;
-
-          // Pick longest roofline per (date, hwKey) so we get one label per series.
-          const bestByGraph = new Map<string, { key: string; pts: InferenceData[] }>();
-          for (const [key, pts] of Object.entries(rooflines)) {
-            if (pts.length < 2) continue;
-            const graphId = key.slice(0, key.lastIndexOf('_'));
-            if (!isRooflineVisible(key)) continue;
-            const prev = bestByGraph.get(graphId);
-            if (!prev || pts.length > prev.pts.length) bestByGraph.set(graphId, { key, pts });
-          }
-
-          const lineLabels: LineLabel[] = [];
-
-          if (isInteractivity) {
-            // Greedy placement: try start → midpoint → 2/3-along → endpoint.
-            const placed: { x: number; y: number }[] = [];
-            const collides = (cx: number, cy: number) =>
-              placed.some((p) => Math.abs(p.y - cy) < LABEL_H && Math.abs(p.x - cx) < LABEL_W);
-
-            const sorted = [...bestByGraph.entries()].toSorted(
-              ([, a], [, b]) => yScale(a.pts[0].y) - yScale(b.pts[0].y),
-            );
-            for (const [graphId, { key, pts }] of sorted) {
-              const candidates = [
-                pts[Math.min(1, pts.length - 1)],
-                pts[Math.floor(pts.length / 2)],
-                pts[Math.max(0, Math.floor((pts.length * 2) / 3))],
-                pts.at(-1)!,
-              ];
-              const labelText = labelTextFor(pts, runNumbering);
-              let placedLabel = false;
-              for (const pt of candidates) {
-                const px = xScale(pt.x);
-                const py = yScale(pt.y);
-                if (!collides(px, py)) {
-                  lineLabels.push({
-                    key,
-                    graphId,
-                    label: labelText,
-                    color: getRooflineColor(key),
-                    x: px,
-                    y: py,
-                    visible: true,
-                  });
-                  placed.push({ x: px, y: py });
-                  placedLabel = true;
-                  break;
-                }
-              }
-              if (!placedLabel) {
-                const pt = pts[0];
-                lineLabels.push({
-                  key,
-                  graphId,
-                  label: labelText,
-                  color: getRooflineColor(key),
-                  x: xScale(pt.x),
-                  y: yScale(pt.y),
-                  visible: false,
-                });
-              }
-            }
-          } else {
-            // TTFT / E2EL: endpoint labels with vertical nudge to avoid overlap.
-            for (const [graphId, { key, pts }] of bestByGraph.entries()) {
-              const pt = pts.at(-1)!;
-              lineLabels.push({
-                key,
-                graphId,
-                label: labelTextFor(pts, runNumbering),
-                color: getRooflineColor(key),
-                x: xScale(pt.x),
-                y: yScale(pt.y),
-                visible: true,
-              });
-            }
-            if (lineLabels.length > 1) {
-              const yRange = yScale.range();
-              const top = Math.min(yRange[0], yRange[1]) + LABEL_H;
-              const bottom = Math.max(yRange[0], yRange[1]) - LABEL_H;
-              lineLabels.sort((a, b) => a.y - b.y);
-              for (let pass = 0; pass < 5; pass++) {
-                for (let i = 1; i < lineLabels.length; i++) {
-                  const overlap = lineLabels[i - 1].y + LABEL_H - lineLabels[i].y;
-                  if (overlap > 0) {
-                    const half = overlap / 2;
-                    lineLabels[i - 1].y -= half;
-                    lineLabels[i].y += half;
-                  }
-                }
-                for (const l of lineLabels) {
-                  l.y = Math.max(top, Math.min(bottom, l.y));
-                }
-              }
-            }
-          }
-
-          const llSel = zoomGroup
-            .selectAll<SVGGElement, LineLabel>('.line-label')
-            .data(lineLabels, (d) => d.key)
-            .join(
-              (enter) => {
-                const g = enter
-                  .append('g')
-                  .attr('class', 'line-label')
-                  .style('pointer-events', 'none');
-                g.append('rect').attr('class', 'll-bg').attr('rx', 4).attr('ry', 4);
-                g.append('text')
-                  .attr('class', 'll-text')
-                  .attr('text-anchor', 'start')
-                  .attr('dominant-baseline', 'central')
-                  .attr('fill', 'white')
-                  .attr('font-size', '10px')
-                  .attr('font-weight', '600');
-                return g;
-              },
-              (update) => update,
-              (exit) => exit.remove(),
-            )
-            .attr('data-line-key', (d) => d.key)
-            .attr('data-graph-id', (d) => d.graphId)
-            .attr('transform', (d) => `translate(${d.x + 8},${d.y - 14})`)
-            .style('opacity', (d) => (d.visible ? 0.95 : 0));
-
-          // Size each label's background to its text in two passes — write all
-          // texts, then measure all bboxes — so the batch forces one layout
-          // instead of one per label (mirrors ScatterGraph's label loops).
-          llSel.each(function (d) {
-            d3.select(this).select<SVGTextElement>('.ll-text').text(d.label);
-          });
-          const llMeasured: { node: SVGGElement; d: LineLabel; bbox: DOMRect }[] = [];
-          llSel.each(function (d) {
-            const text = this.querySelector<SVGTextElement>('.ll-text');
-            if (text) llMeasured.push({ node: this, d, bbox: text.getBBox() });
-          });
-          for (const { node, d, bbox } of llMeasured) {
-            const px = 5;
-            const py = 3;
-            d3.select(node)
-              .select('.ll-bg')
-              .attr('x', bbox.x - px)
-              .attr('y', bbox.y - py)
-              .attr('width', bbox.width + px * 2)
-              .attr('height', bbox.height + py * 2)
-              .attr('fill', d.color);
-          }
+          renderLineLabels(
+            zoomGroup,
+            placeLabels(ctx.xScale as ContinuousScale, ctx.yScale as ContinuousScale, zoomGroup),
+            {
+              seriesAttribute: 'data-graph-id',
+              opacity: showLineLabels ? 0.95 : 0,
+            },
+          );
+        },
+        onDisplayUpdate: (zoomGroup, ctx) => {
+          const transform = d3.zoomTransform(ctx.layout.svg.node()!);
+          renderLineLabels(
+            zoomGroup,
+            placeLabels(
+              transform.rescaleX(ctx.xScale as ContinuousScale),
+              transform.rescaleY(ctx.yScale as ContinuousScale),
+              zoomGroup,
+            ),
+            {
+              seriesAttribute: 'data-graph-id',
+              opacity: showLineLabels ? 0.95 : 0,
+            },
+          );
+          zoomGroup.selectAll('.line-label').raise();
         },
         onZoom: (zoomGroup, ctx) => {
-          if (!showLineLabels) return;
-          const newXScale = ctx.newXScale as ContinuousScale;
-          const newYScale = ctx.newYScale as ContinuousScale;
-
-          // Mirror the placement algorithm with the zoomed scales so labels
-          // stay anchored to their rooflines without jumping on zoom.
-          const isInteractivity = chartDefinition.chartType === 'interactivity';
-          const LABEL_H = 18;
-          const LABEL_W = 160;
-
-          const bestByGraph = new Map<string, { key: string; pts: InferenceData[] }>();
-          for (const [key, pts] of Object.entries(rooflines)) {
-            if (pts.length < 2 || !isRooflineVisible(key)) continue;
-            const graphId = key.slice(0, key.lastIndexOf('_'));
-            const prev = bestByGraph.get(graphId);
-            if (!prev || pts.length > prev.pts.length) bestByGraph.set(graphId, { key, pts });
-          }
-
-          const zoomResults = new Map<string, { x: number; y: number; vis: boolean }>();
-          if (isInteractivity) {
-            const placed: { x: number; y: number }[] = [];
-            const collides = (cx: number, cy: number) =>
-              placed.some((p) => Math.abs(p.y - cy) < LABEL_H && Math.abs(p.x - cx) < LABEL_W);
-            const sorted = [...bestByGraph.entries()].toSorted(
-              ([, a], [, b]) => newYScale(a.pts[0].y) - newYScale(b.pts[0].y),
-            );
-            for (const [, { key, pts }] of sorted) {
-              const candidates = [
-                pts[Math.min(1, pts.length - 1)],
-                pts[Math.floor(pts.length / 2)],
-                pts[Math.max(0, Math.floor((pts.length * 2) / 3))],
-                pts.at(-1)!,
-              ];
-              let found = false;
-              for (const pt of candidates) {
-                const px = newXScale(pt.x);
-                const py = newYScale(pt.y);
-                if (!collides(px, py)) {
-                  zoomResults.set(key, { x: px, y: py, vis: true });
-                  placed.push({ x: px, y: py });
-                  found = true;
-                  break;
-                }
-              }
-              if (!found) {
-                zoomResults.set(key, {
-                  x: newXScale(pts[0].x),
-                  y: newYScale(pts[0].y),
-                  vis: false,
-                });
-              }
-            }
-          } else {
-            interface ZL {
-              key: string;
-              x: number;
-              y: number;
-            }
-            const zls: ZL[] = [];
-            for (const [, { key, pts }] of bestByGraph.entries()) {
-              const pt = pts.at(-1)!;
-              zls.push({ key, x: newXScale(pt.x), y: newYScale(pt.y) });
-            }
-            if (zls.length > 1) {
-              const yRange = newYScale.range();
-              const top = Math.min(yRange[0], yRange[1]) + LABEL_H;
-              const bottom = Math.max(yRange[0], yRange[1]) - LABEL_H;
-              zls.sort((a, b) => a.y - b.y);
-              for (let pass = 0; pass < 5; pass++) {
-                for (let i = 1; i < zls.length; i++) {
-                  const overlap = zls[i - 1].y + LABEL_H - zls[i].y;
-                  if (overlap > 0) {
-                    const half = overlap / 2;
-                    zls[i - 1].y -= half;
-                    zls[i].y += half;
-                  }
-                }
-                for (const z of zls) z.y = Math.max(top, Math.min(bottom, z.y));
-              }
-            }
-            for (const z of zls) zoomResults.set(z.key, { x: z.x, y: z.y, vis: true });
-          }
-
-          zoomGroup.selectAll<SVGGElement, unknown>('.line-label').each(function () {
-            const el = d3.select(this);
-            const k = el.attr('data-line-key');
-            const zl = zoomResults.get(k);
-            if (zl) {
-              el.attr('transform', `translate(${zl.x + 8},${zl.y - 14})`);
-              el.style('opacity', zl.vis ? 0.95 : 0);
-            } else {
-              el.style('opacity', 0);
-            }
-          });
+          updateRenderedLineLabels(
+            zoomGroup,
+            placeLabels(
+              ctx.newXScale as ContinuousScale,
+              ctx.newYScale as ContinuousScale,
+              zoomGroup,
+            ),
+            { opacity: showLineLabels ? 0.95 : 0 },
+          );
         },
-      }),
-      [
-        showLineLabels,
-        rooflines,
-        isRooflineVisible,
-        getRooflineColor,
-        chartDefinition.chartType,
-        runNumbering,
-      ],
-    );
+      };
+    }, [
+      showLineLabels,
+      rooflines,
+      isRooflineVisible,
+      getRooflineColor,
+      chartDefinition.chartType,
+      runNumbering,
+      paletteIdentity,
+    ]);
 
     // Dismiss tooltip when pinned point's combo is hidden
     useEffect(() => {
@@ -760,8 +699,22 @@ const GPUGraph = React.memo(
               <p className="text-xs">
                 Please change the model, sequence, precision, date range or chip selection.
               </p>
+              <Button
+                type="button"
+                size="sm"
+                variant="outline"
+                className="mt-4"
+                data-testid="gpu-empty-quick-filters"
+                onClick={() => {
+                  setQuickFiltersOpen(true);
+                  track('inference_quick_filters_dialog_opened', { source: 'timeline_empty' });
+                }}
+              >
+                {legendT.quickFilters(quickFilterCount)}
+              </Button>
             </div>
           </div>
+          <QuickFiltersDialog open={quickFiltersOpen} onOpenChange={setQuickFiltersOpen} />
         </div>
       );
     }
@@ -770,6 +723,9 @@ const GPUGraph = React.memo(
       <D3Chart<InferenceData>
         ref={chartRef}
         chartId={chartId}
+        dataIdentity={dataIdentity}
+        metricIdentity={metricIdentity}
+        displayIdentity={`${showPointLabels}:${paletteIdentity}:${selectedPrecisions.join(',')}`}
         data={filteredData}
         margin={CHART_MARGIN}
         watermark={getChartWatermark()}
@@ -805,17 +761,18 @@ const GPUGraph = React.memo(
             config: {
               getColor,
               hideLabels: !showPointLabels,
-              // Match ScatterGraph: append the concurrency (C=) to the
-              // parallelism/tp label so compare-mode points are annotated the
-              // same way as the single-run scatter chart.
-              getLabelText: (d) =>
-                useAdvancedLabels ? `${getPointLabel(d)}\nC=${d.conc}` : `${d.tp}\nC=${d.conc}`,
+              // Match ScatterGraph: concurrency (C=) is appended only when the
+              // advanced "# Concurrent Sessions" toggle is on, so compare-mode
+              // points are annotated the same way as the single-run scatter
+              // chart.
+              getLabelText: (d) => pointLabelText(d, useAdvancedLabels, showConcurrencyLabels),
               foreground: 'var(--foreground)',
               dataAttrs: {
                 series: (d) => `${d.date}_${d.hwKey}`,
               },
               selectedPrecisions,
             },
+            keyFn: (point) => `${point.date}:${scatterPointConfigId(point)}`,
           },
           lineLabelLayer,
           knownIssueLayer,
@@ -848,7 +805,12 @@ const GPUGraph = React.memo(
               selectedYAxisMetric,
               hardwareConfig,
               runUrl: d.run_url ? updateRepoUrl(d.run_url) : undefined,
-              hasTrace: typeof d.id === 'number' ? traceAvailability?.[d.id] === true : false,
+              hasTrace: isPersistedBenchmarkId(d.id)
+                ? traceAvailabilityRef.current?.[d.id] === true
+                : false,
+              hasLog: isPersistedBenchmarkId(d.id)
+                ? logAvailabilityRef.current?.[d.id] === true
+                : false,
               locale,
             }),
           getRulerX: (d, xScale) => (xScale as d3.ScaleLinear<number, number>)(d.x),
@@ -873,15 +835,39 @@ const GPUGraph = React.memo(
             const tooltipEl = chartRef.current?.getTooltipElement();
             if (!tooltipEl) return;
             const viewBtn = tooltipEl.querySelector('[data-action="view-charts"]');
-            if (!viewBtn || typeof d.id !== 'number') return;
-            viewBtn.addEventListener('click', (event) => {
-              event.stopPropagation();
-              track('gpu_timeseries_view_charts_opened', {
-                id: d.id,
-                hwKey: String(d.hwKey),
-                conc: d.conc,
+            if (viewBtn && isPersistedBenchmarkId(d.id)) {
+              viewBtn.addEventListener('click', (event) => {
+                event.stopPropagation();
+                // Full-document navigation: stamp the chart state onto THIS
+                // history entry first, or Back returns to a bare /inference
+                // that rebuilds from defaults. Skipped in ephemeral scopes
+                // (/model embeds): the store holds the primary dashboard's
+                // state there, not this chart's.
+                if (!ephemeralUrlState) rememberChartStateInUrl();
+                track('gpu_timeseries_view_charts_opened', {
+                  id: d.id,
+                  hwKey: String(d.hwKey),
+                  conc: d.conc,
+                });
               });
-            });
+            }
+            const logsBtn = tooltipEl.querySelector('[data-action="view-logs"]');
+            if (logsBtn && typeof d.id === 'number') {
+              logsBtn.addEventListener('click', (event) => {
+                event.stopPropagation();
+                if (d.benchmark_type !== 'agentic_traces') {
+                  event.preventDefault();
+                  setFixedLogPointId(d.id!);
+                  chartRef.current?.dismissTooltip();
+                }
+                track('gpu_timeseries_view_logs_opened', {
+                  id: d.id,
+                  hwKey: String(d.hwKey),
+                  conc: d.conc,
+                  benchmarkType: d.benchmark_type ?? 'single_turn',
+                });
+              });
+            }
             // Pinning updates D3Chart's React state. GPU comparison rebuilds
             // several inline layer configs on that render, whose cleanup can
             // briefly hide the otherwise-pinned portal tooltip. Restore its
@@ -899,7 +885,10 @@ const GPUGraph = React.memo(
         onRender={(ctx: RenderContext) => {
           // Apply log tick format on initial render (needs the built scale)
           if (logScale) {
-            const yScale = ctx.yScale as d3.ScaleLogarithmic<number, number>;
+            const yScale = (ctx.renderedYScale ?? ctx.yScale) as d3.ScaleLogarithmic<
+              number,
+              number
+            >;
             ctx.layout.yAxisGroup.call(
               d3.axisLeft(yScale).ticks(10).tickFormat(logTickFormat(yScale)) as any,
             );
@@ -913,26 +902,12 @@ const GPUGraph = React.memo(
             .selectAll('.dot-group, .roofline-path')
             .style('transition', 'opacity 150ms ease');
 
-          // Offload halo: dashed ring on every point that used KV offload
-          // (mirrors ScatterGraph so compare mode shows the same CPU-offload
-          // indicator). The ring is a child of the dot-group, so it travels
-          // with the point on zoom/pan without a separate onZoom pass.
+          // The halo stays inside the point group, so normal zoom transforms
+          // carry it without a separate update pass.
           ctx.layout.zoomGroup
             .selectAll<SVGGElement, InferenceData>('.dot-group')
-            .each(function (d) {
-              const showHalo = d.offload_mode === 'on';
-              d3.select(this)
-                .selectAll<SVGCircleElement, boolean>('.offload-halo')
-                .data(showHalo ? [true] : [])
-                .join('circle')
-                .attr('class', 'offload-halo')
-                .attr('r', OFFLOAD_HALO_RADIUS)
-                .attr('fill', 'none')
-                .attr('stroke', 'var(--foreground)')
-                .attr('stroke-width', OFFLOAD_HALO_STROKE_WIDTH)
-                .attr('stroke-dasharray', OFFLOAD_HALO_DASHARRAY)
-                .attr('opacity', 0.9)
-                .attr('pointer-events', 'none');
+            .each(function (point) {
+              renderOffloadHalo(d3.select(this), point, 'var(--foreground)');
             });
         }}
         legendElement={
@@ -966,6 +941,7 @@ const GPUGraph = React.memo(
               {
                 id: 'gpu-log-scale',
                 label: legendT.logScale,
+                advanced: true,
                 checked: logScale,
                 onCheckedChange: (c) => {
                   setLogScale(c);
@@ -975,6 +951,7 @@ const GPUGraph = React.memo(
               {
                 id: 'gpu-high-contrast',
                 label: legendT.highContrast,
+                advanced: true,
                 checked: highContrast,
                 onCheckedChange: (c) => {
                   setHighContrast(c);
@@ -993,6 +970,7 @@ const GPUGraph = React.memo(
               {
                 id: 'gpu-point-labels',
                 label: legendT.labels,
+                advanced: true,
                 checked: showPointLabels,
                 onCheckedChange: (c) => {
                   setShowPointLabels(c);
@@ -1002,6 +980,7 @@ const GPUGraph = React.memo(
               {
                 id: 'gpu-parallelism-labels',
                 label: legendT.parallelismLabels,
+                advanced: true,
                 checked: useAdvancedLabels,
                 onCheckedChange: (c) => {
                   setUseAdvancedLabels(c);
@@ -1014,12 +993,26 @@ const GPUGraph = React.memo(
               {
                 id: 'gpu-line-labels',
                 label: legendT.lineLabels,
+                advanced: true,
                 checked: showLineLabels,
                 onCheckedChange: (c) => {
                   setShowLineLabels(c);
                   track('interactivity_line_labels_toggled', { enabled: c });
                 },
               },
+              {
+                id: 'gpu-concurrency-labels',
+                label: legendT.concurrencyLabels,
+                advanced: true,
+                checked: showConcurrencyLabels,
+                onCheckedChange: (c) => {
+                  setShowConcurrencyLabels(c);
+                  track('interactivity_concurrency_labels_toggled', { enabled: c });
+                  // Concurrency is a point-label annotation; turning it on is
+                  // pointless if labels are hidden, so auto-enable Labels.
+                  if (c && !showPointLabels) setShowPointLabels(true);
+                },
+              },
             ]}
             actions={[
               {
@@ -1027,18 +1020,33 @@ const GPUGraph = React.memo(
                 label: legendT.resetFilter,
                 onClick: () => {
                   selectAllActiveDates();
+                  clearQuickFilters();
                   track('gpu_timeseries_reset_filter');
                 },
               },
+              {
+                id: 'gpu-quick-filters',
+                label: legendT.quickFilters(quickFilterCount),
+                onClick: () => {
+                  setQuickFiltersOpen(true);
+                  track('inference_quick_filters_dialog_opened', { source: 'timeline_legend' });
+                },
+              },
             ]}
             precisionIndicators={selectedPrecisions}
+            hideAtomFootnote
             keyIndicators={
-              hasOffloadHalo || selectedSequence === Sequence.AgenticTraces ? (
-                <>
-                  {hasOffloadHalo && <OffloadHaloLegendKey />}
-                  {selectedSequence === Sequence.AgenticTraces && <AgenticOptimizationNote />}
-                </>
-              ) : undefined
+              <>
+                {fixedLogPointId === null ? null : (
+                  <FixedSequenceLogDialog
+                    pointId={fixedLogPointId}
+                    onOpenChange={(open) => {
+                      if (!open) setFixedLogPointId(null);
+                    }}
+                  />
+                )}
+                <QuickFiltersDialog open={quickFiltersOpen} onOpenChange={setQuickFiltersOpen} />
+              </>
             }
           />
         }
diff --git a/packages/app/src/components/inference/ui/InferenceTable.test.ts b/packages/app/src/components/inference/ui/InferenceTable.test.ts
index d5356b72e..6f699cb48 100644
--- a/packages/app/src/components/inference/ui/InferenceTable.test.ts
+++ b/packages/app/src/components/inference/ui/InferenceTable.test.ts
@@ -1,6 +1,7 @@
 import { describe, it, expect } from 'vitest';
 
 import type { ChartDefinition, InferenceData } from '@/components/inference/types';
+import { formatInferenceTableNumber } from '@/components/inference/ui/InferenceTable';
 
 // Test the pure logic used by InferenceTable — sorting and value resolution
 import { getNestedYValue } from '@/lib/chart-utils';
@@ -18,6 +19,9 @@ const CHART_DEF = {
   y_costh: 'costh.y',
   y_costh_label: 'Cost per Million Total Tokens ($)',
   y_costh_roofline: 'lower_right',
+  y_tokensPerDollarH: 'tokensPerDollarH.y',
+  y_tokensPerDollarH_label: 'Total Tokens per $1 (tok/$)',
+  y_tokensPerDollarH_roofline: 'upper_left',
 } as unknown as ChartDefinition;
 
 function makePoint(overrides: Partial<InferenceData>): InferenceData {
@@ -37,6 +41,7 @@ function makePoint(overrides: Partial<InferenceData>): InferenceData {
     costhi: { y: 0.2, roof: false },
     costni: { y: 0.15, roof: false },
     costri: { y: 0.1, roof: false },
+    tokensPerDollarH: { y: 2_000_000, roof: false },
     ...overrides,
   } as InferenceData;
 }
@@ -59,21 +64,21 @@ describe('InferenceTable sorting logic', () => {
     expect(getNestedYValue(sorted[2], yPath)).toBe(100);
   });
 
-  it('sorts by Y value ascending for lower_right roofline (cost)', () => {
+  it('sorts tokens-per-dollar purchasing power descending', () => {
     const points = [
-      makePoint({ costh: { y: 0.8, roof: false } }),
-      makePoint({ costh: { y: 0.2, roof: true } }),
-      makePoint({ costh: { y: 0.5, roof: false } }),
+      makePoint({ tokensPerDollarH: { y: 800_000, roof: false } }),
+      makePoint({ tokensPerDollarH: { y: 200_000, roof: false } }),
+      makePoint({ tokensPerDollarH: { y: 1_500_000, roof: true } }),
     ];
 
-    const yPath = CHART_DEF.y_costh as string;
+    const yPath = CHART_DEF.y_tokensPerDollarH as string;
     const sorted = [...points].toSorted(
-      (a, b) => getNestedYValue(a, yPath) - getNestedYValue(b, yPath),
+      (a, b) => getNestedYValue(b, yPath) - getNestedYValue(a, yPath),
     );
 
-    expect(getNestedYValue(sorted[0], yPath)).toBe(0.2);
-    expect(getNestedYValue(sorted[1], yPath)).toBe(0.5);
-    expect(getNestedYValue(sorted[2], yPath)).toBe(0.8);
+    expect(getNestedYValue(sorted[0], yPath)).toBe(1_500_000);
+    expect(getNestedYValue(sorted[1], yPath)).toBe(800_000);
+    expect(getNestedYValue(sorted[2], yPath)).toBe(200_000);
   });
 });
 
@@ -83,13 +88,35 @@ describe('getNestedYValue', () => {
     expect(getNestedYValue(point, 'tpPerGpu.y')).toBe(42);
   });
 
-  it('resolves nested cost metric path (costh.y)', () => {
+  it('resolves the existing cost-per-million path (costh.y)', () => {
     const point = makePoint({ costh: { y: 1.23, roof: false } });
     expect(getNestedYValue(point, 'costh.y')).toBe(1.23);
   });
 
+  it('resolves the separate tokens-per-dollar path', () => {
+    const point = makePoint({ tokensPerDollarH: { y: 1_500_000, roof: false } });
+    expect(getNestedYValue(point, 'tokensPerDollarH.y')).toBe(1_500_000);
+  });
+
   it('returns 0 for missing paths', () => {
     const point = makePoint({});
     expect(getNestedYValue(point, 'nonexistent.y')).toBe(0);
   });
 });
+
+describe('formatInferenceTableNumber', () => {
+  it('groups large chart values with commas', () => {
+    expect(formatInferenceTableNumber(87_000)).toBe('87,000');
+    expect(formatInferenceTableNumber(125_500)).toBe('125,500');
+  });
+
+  it('preserves the requested fixed precision while grouping thousands', () => {
+    expect(formatInferenceTableNumber(125_500, 1)).toBe('125,500.0');
+  });
+
+  it('keeps the existing magnitude-based precision for smaller values', () => {
+    expect(formatInferenceTableNumber(12.34)).toBe('12.3');
+    expect(formatInferenceTableNumber(0.1234)).toBe('0.123');
+    expect(formatInferenceTableNumber(0.00123)).toBe('0.0012');
+  });
+});
diff --git a/packages/app/src/components/inference/ui/InferenceTable.tsx b/packages/app/src/components/inference/ui/InferenceTable.tsx
index 921a5bd37..8b7dc7bdd 100644
--- a/packages/app/src/components/inference/ui/InferenceTable.tsx
+++ b/packages/app/src/components/inference/ui/InferenceTable.tsx
@@ -16,13 +16,15 @@ interface InferenceTableProps {
   selectedYAxisMetric: string;
 }
 
-/** Format a number for table display — picks sensible precision based on magnitude. */
-function fmt(value: number, decimals?: number): string {
-  if (decimals !== undefined) return value.toFixed(decimals);
-  if (Math.abs(value) >= 100) return value.toFixed(0);
-  if (Math.abs(value) >= 1) return value.toFixed(1);
-  if (Math.abs(value) >= 0.01) return value.toFixed(3);
-  return value.toFixed(4);
+/** Format a number for table display — picks sensible precision and groups thousands. */
+export function formatInferenceTableNumber(value: number, decimals?: number): string {
+  const fixedDecimals =
+    decimals ??
+    (Math.abs(value) >= 100 ? 0 : Math.abs(value) >= 1 ? 1 : Math.abs(value) >= 0.01 ? 3 : 4);
+  return new Intl.NumberFormat('en-US', {
+    minimumFractionDigits: fixedDecimals,
+    maximumFractionDigits: fixedDecimals,
+  }).format(value);
 }
 
 export default function InferenceTable({
@@ -70,38 +72,24 @@ export default function InferenceTable({
       {
         header: yLabel,
         align: 'right',
-        cell: (row) => fmt(yPath ? getNestedYValue(row, yPath) : row.y),
+        cell: (row) => formatInferenceTableNumber(yPath ? getNestedYValue(row, yPath) : row.y),
         sortValue: (row) => (yPath ? getNestedYValue(row, yPath) : row.y),
         className: 'tabular-nums',
       },
       {
         header: xLabel,
         align: 'right',
-        cell: (row) => fmt(row.x),
+        cell: (row) => formatInferenceTableNumber(row.x),
         sortValue: (row) => row.x,
         className: 'tabular-nums',
       },
       {
         header: 'Throughput/Chip (tok/s)',
         align: 'right',
-        cell: (row) => fmt(row.tput_per_gpu ?? 0, 1),
+        cell: (row) => formatInferenceTableNumber(row.tput_per_gpu ?? 0, 1),
         sortValue: (row) => row.tput_per_gpu ?? 0,
         className: 'tabular-nums',
       },
-      {
-        header: 'Median TTFT (ms)',
-        align: 'right',
-        cell: (row) => fmt((row.median_ttft ?? 0) * 1000, 0),
-        sortValue: (row) => row.median_ttft ?? 0,
-        className: 'tabular-nums',
-      },
-      {
-        header: 'Median Interactivity (tok/s)',
-        align: 'right',
-        cell: (row) => fmt(row.median_intvty ?? 0, 1),
-        sortValue: (row) => row.median_intvty ?? 0,
-        className: 'tabular-nums',
-      },
     ],
     [yPath, yLabel, xLabel],
   );
diff --git a/packages/app/src/components/inference/ui/LegendPointsDialog.tsx b/packages/app/src/components/inference/ui/LegendPointsDialog.tsx
index 5bfdfc249..3be050f39 100644
--- a/packages/app/src/components/inference/ui/LegendPointsDialog.tsx
+++ b/packages/app/src/components/inference/ui/LegendPointsDialog.tsx
@@ -10,6 +10,8 @@ import {
   DialogHeader,
   DialogTitle,
 } from '@/components/ui/dialog';
+import { useEphemeralUrlState } from '@/hooks/useUrlState';
+import { rememberChartStateInUrl } from '@/lib/url-state';
 import { cn } from '@/lib/utils';
 
 import {
@@ -24,7 +26,7 @@ export interface LegendPointsDialogProps {
   onOpenChange: (open: boolean) => void;
   /** Series label, e.g. "B300 (vLLM)". */
   title: string;
-  /** Context line, e.g. "DeepSeek V4 Pro · Agentic Traces". */
+  /** Context line, e.g. "DeepSeek V4 Pro · Agentic". */
   subtitle: string;
   /** Legend swatch color for this series (overlayRunColor for overlay runs). */
   accentColor: string;
@@ -64,6 +66,7 @@ export default function LegendPointsDialog({
   onRowClick,
 }: LegendPointsDialogProps) {
   const [sort, setSort] = useState<{ key: LegendPointsSortKey; dir: 'asc' | 'desc' } | null>(null);
+  const ephemeralUrlState = useEphemeralUrlState();
 
   const hasOffload = rows.some((r) => r.offload !== null);
   const columns = useMemo(
@@ -181,7 +184,15 @@ export default function LegendPointsDialog({
                   key={row.key}
                   href={row.href}
                   {...(row.isExternal ? { target: '_blank', rel: 'noopener noreferrer' } : {})}
-                  onClick={() => onRowClick?.(row)}
+                  onClick={() => {
+                    // In-app detail links are full-document navigations, so the
+                    // chart state has to be written into this history entry
+                    // before we leave or Back lands on a default chart. Skipped
+                    // in ephemeral scopes (/model embeds): the store holds the
+                    // primary dashboard's state there, not this chart's.
+                    if (!row.isExternal && !ephemeralUrlState) rememberChartStateInUrl();
+                    onRowClick?.(row);
+                  }}
                   className="col-span-full grid grid-cols-subgrid items-center rounded-sm hover:bg-accent whitespace-nowrap"
                 >
                   {renderCells(row)}
@@ -200,7 +211,7 @@ export default function LegendPointsDialog({
           </div>
         )}
 
-        <p className="text-[10px] text-muted-foreground/70 leading-tight">
+        <p className="text-3xs text-muted-foreground/70 leading-tight">
           {isOverlay
             ? 'Unofficial overlay points have no stored benchmark records — metrics only, no detail links.'
             : 'Click a row for the point detail — agentic points open the trace detail page, fixed-seq points open the GitHub Actions run.'}{' '}
diff --git a/packages/app/src/components/inference/ui/ModelArchitectureDiagram.test.tsx b/packages/app/src/components/inference/ui/ModelArchitectureDiagram.test.tsx
new file mode 100644
index 000000000..3636fb491
--- /dev/null
+++ b/packages/app/src/components/inference/ui/ModelArchitectureDiagram.test.tsx
@@ -0,0 +1,144 @@
+// @vitest-environment jsdom
+
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { Model } from '@/lib/data-mappings';
+
+const mocks = vi.hoisted(() => ({
+  renderDiagram: vi.fn(),
+  theme: { resolved: 'light' },
+}));
+
+vi.mock('./model-architecture-diagram-renderer', () => ({
+  renderDiagram: mocks.renderDiagram,
+}));
+vi.mock('next-themes', () => ({
+  useTheme: () => ({ resolvedTheme: mocks.theme.resolved }),
+}));
+vi.mock('next/link', () => ({
+  default: ({ children, ...props }: React.AnchorHTMLAttributes<HTMLAnchorElement>) => (
+    <a {...props}>{children}</a>
+  ),
+}));
+vi.mock('@/lib/analytics', () => ({ track: vi.fn() }));
+
+import ModelArchitectureDiagram from './ModelArchitectureDiagram';
+
+declare global {
+  var IS_REACT_ACT_ENVIRONMENT: boolean;
+}
+
+globalThis.IS_REACT_ACT_ENVIRONMENT = true;
+
+class ResizeObserverStub {
+  static instances: ResizeObserverStub[] = [];
+
+  target: Element | null = null;
+
+  private readonly callback: ResizeObserverCallback;
+
+  constructor(callback: ResizeObserverCallback) {
+    this.callback = callback;
+    ResizeObserverStub.instances.push(this);
+  }
+
+  observe(target: Element) {
+    this.target = target;
+  }
+
+  disconnect() {}
+
+  trigger() {
+    if (!this.target) throw new Error('ResizeObserver has no observed target');
+    this.callback(
+      [
+        {
+          target: this.target,
+          contentRect: { width: observedWidth } as DOMRectReadOnly,
+        } as ResizeObserverEntry,
+      ],
+      this as unknown as ResizeObserver,
+    );
+  }
+}
+
+let container: HTMLDivElement;
+let root: Root;
+let observedWidth: number;
+let clientWidthSpy: { mockRestore: () => void };
+
+function render(model: Model = Model.DeepSeek_R1) {
+  act(() => root.render(<ModelArchitectureDiagram model={model} />));
+}
+
+function expand() {
+  const toggle = container.querySelector<HTMLButtonElement>(
+    '[data-testid="model-architecture-toggle"]',
+  );
+  if (!toggle) throw new Error('Architecture toggle did not render');
+  act(() => toggle.click());
+}
+
+beforeEach(() => {
+  observedWidth = 500;
+  mocks.theme.resolved = 'light';
+  mocks.renderDiagram.mockReset();
+  ResizeObserverStub.instances = [];
+  vi.stubGlobal('ResizeObserver', ResizeObserverStub);
+  clientWidthSpy = vi
+    .spyOn(HTMLElement.prototype, 'clientWidth', 'get')
+    .mockImplementation(() => observedWidth);
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  clientWidthSpy.mockRestore();
+  vi.unstubAllGlobals();
+});
+
+describe('ModelArchitectureDiagram rendering', () => {
+  it('renders once on initial expansion and ignores the observer callback for the same width', () => {
+    render();
+    expand();
+
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(1);
+    expect(ResizeObserverStub.instances).toHaveLength(1);
+
+    act(() => ResizeObserverStub.instances[0].trigger());
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(1);
+
+    observedWidth = 460;
+    act(() => ResizeObserverStub.instances[0].trigger());
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(2);
+
+    act(() => ResizeObserverStub.instances[0].trigger());
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(2);
+  });
+
+  it('rerenders when theme, expanded blocks, or model changes', () => {
+    render();
+    expand();
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(1);
+
+    mocks.theme.resolved = 'dark';
+    render();
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(2);
+
+    const toggleBlock = mocks.renderDiagram.mock.calls.at(-1)?.[4] as
+      | ((blockId: string) => void)
+      | undefined;
+    expect(toggleBlock).toBeTypeOf('function');
+    act(() => toggleBlock?.('experts'));
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(3);
+    expect(mocks.renderDiagram.mock.calls.at(-1)?.[3]).toEqual(new Set(['experts']));
+
+    render(Model.GptOss);
+    expect(mocks.renderDiagram).toHaveBeenCalledTimes(4);
+  });
+});
diff --git a/packages/app/src/components/inference/ui/ModelArchitectureDiagram.tsx b/packages/app/src/components/inference/ui/ModelArchitectureDiagram.tsx
index 8cf2de0f2..b0d45b08d 100644
--- a/packages/app/src/components/inference/ui/ModelArchitectureDiagram.tsx
+++ b/packages/app/src/components/inference/ui/ModelArchitectureDiagram.tsx
@@ -1,2271 +1,55 @@
 'use client';
 
 import { track } from '@/lib/analytics';
-import * as d3 from 'd3';
 import { ChevronDown, ChevronUp, ExternalLink } from 'lucide-react';
 import Link from 'next/link';
 import { useCallback, useEffect, useRef, useState } from 'react';
 import { useTheme } from 'next-themes';
 
+import { getModelReleaseDate } from '@semianalysisai/inferencex-constants';
+
 import { Badge } from '@/components/ui/badge';
 import type { Model } from '@/lib/data-mappings';
 import {
-  type ArchSubBlock,
-  type SubBlockFlow,
   type ModelArchitecture,
-  formatContextWindow,
   formatParamCount,
-  expertRouterSummary,
-  getAttentionLabel,
-  getAttentionSubBlocks,
-  getFFNSubBlocks,
-  denseLayerAttentionLabel,
-  ffnVariantLabel,
-  getHybridAttentionSubBlocks,
   getModelArchitecture,
-  sharedExpertCount,
 } from '@/lib/model-architectures';
+import { renderDiagram } from './model-architecture-diagram-renderer';
 
 interface ModelArchitectureDiagramProps {
   model: Model;
   className?: string;
+  /**
+   * `drawer` (default) renders the collapsible bar used inside dashboard
+   * cards. `inline` renders the diagram always-expanded with no toggle —
+   * used by the `/model/[slug]` pages where the architecture is the content,
+   * not an aside.
+   */
+  variant?: 'drawer' | 'inline';
 }
 
-/** Block color definitions for light/dark themes */
-const BLOCK_COLORS = {
-  embedding: { light: '#dbeafe', dark: '#1e3a5f', stroke: '#3b82f6' },
-  attention: { light: '#fef3c7', dark: '#422006', stroke: '#d97706' },
-  ffn: { light: '#d1fae5', dark: '#064e3b', stroke: '#059669' },
-  output: { light: '#e0e7ff', dark: '#1e1b4b', stroke: '#6366f1' },
-  norm: { light: '#f1f5f9', dark: '#1e293b', stroke: '#94a3b8' },
-  specs: { light: '#f8fafc', dark: '#1a1c1e', stroke: '#cbd5e1' },
-  router: { light: '#fce7f3', dark: '#500724', stroke: '#db2777' },
-  expert: { light: '#f3e8ff', dark: '#3b0764', stroke: '#9333ea' },
-  expertActive: { light: '#e9d5ff', dark: '#581c87', stroke: '#7c3aed' },
-};
-
-/** Sub-block color definitions for expanded drill-down views */
-const SUB_BLOCK_COLORS = {
-  projection: { light: '#e0f2fe', dark: '#0c4a6e', stroke: '#0284c7' },
-  activation: { light: '#dcfce7', dark: '#14532d', stroke: '#16a34a' },
-  operation: { light: '#f3f4f6', dark: '#1f2937', stroke: '#6b7280' },
-  attention: { light: '#fef9c3', dark: '#422006', stroke: '#ca8a04' },
-};
-
-function getColor(type: keyof typeof BLOCK_COLORS, isDark: boolean) {
-  const c = BLOCK_COLORS[type];
-  return { fill: isDark ? c.dark : c.light, stroke: c.stroke };
+interface ArchitectureContentProps {
+  model: Model;
+  arch: ModelArchitecture;
+  isExpanded: boolean;
 }
 
-function getSubBlockColor(type: ArchSubBlock['type'], isDark: boolean) {
-  const c = SUB_BLOCK_COLORS[type];
-  return { fill: isDark ? c.dark : c.light, stroke: c.stroke };
+interface DiagramRenderState {
+  width: number;
+  arch: ModelArchitecture;
+  isDark: boolean;
+  expandedBlocks: Set<string>;
+  onBlockClick: (blockId: string) => void;
 }
 
-/**
- * Renders the D3-based SVG architecture diagram.
- * Supports both dense (Llama) and MoE (DeepSeek R1) architectures.
- * For MoE models with denseFFNLayers, renders a separate dense transformer block.
- * MLA attention blocks are NOT expandable (shown as static blocks).
- */
-function renderDiagram(
-  svgEl: SVGSVGElement,
-  arch: ModelArchitecture,
-  isDark: boolean,
-  expandedBlocks: Set<string>,
-  onBlockClick: (blockId: string) => void,
-) {
-  const svg = d3.select(svgEl);
-  svg.selectAll('*').remove();
-
-  const containerWidth = svgEl.parentElement?.clientWidth || 600;
-  const width = Math.min(containerWidth, 640);
-
-  // Theme colors
-  const fg = isDark ? '#eaebec' : '#131416';
-  const mutedFg = isDark ? '#9ca3af' : '#6b7280';
-  const borderColor = isDark ? '#374151' : '#d1d5db';
-  const bgSubtle = isDark ? '#1f2937' : '#f9fafb';
-  const expandedBg = isDark ? '#111827' : '#f1f5f9';
-
-  // Layout constants
-  const pad = { top: 20, right: 16, bottom: 16, left: 16 };
-  const bw = width - pad.left - pad.right;
-  const innerW = bw - 32;
-  const innerX = pad.left + 16;
-  const cx = width / 2;
-  const blockH = 52;
-  const smallH = 32;
-  const arrowH = 20;
-  const circleR = 9;
-  const mergeGap = 36;
-  const residLeftX = pad.left + 6;
-  const collapsedTxH = 64;
-
-  // Expert grid constants
-  const expertSize = 36;
-  const expertGap = 6;
-  const expertGridH = 70;
-
-  // Sub-block layout constants
-  const subBlockH = 34;
-  const subArrowH = 12;
-  const subPadY = 10;
-  // Gap between a block and its drill-down flow so the dashed expansion rect
-  // reads as separate from the block above it (not overlapping its border).
-  const drillGap = 8;
-
-  // Architecture flags
-  const isMoE = arch.architectureType === 'moe';
-  // Shared experts are counted inside `numExperts`; read the model's own count
-  // rather than assuming the DeepSeek-style single shared expert.
-  const sharedExperts = sharedExpertCount(arch);
-  const ffnVariant = ffnVariantLabel(arch);
-  const hasDenseLayers = isMoE && (arch.denseFFNLayers ?? 0) > 0;
-  const denseLayerCount = arch.denseFFNLayers ?? 0;
-  const moeLayerCount = (arch.numLayers ?? 0) - denseLayerCount;
-  // MLA and AlternatingSinkGQA attention types are NOT expandable.
-  // Models can also explicitly opt out via attentionExpandable: false.
-  const isAttnExpandable =
-    arch.attentionExpandable !== false &&
-    arch.attentionType !== 'MLA' &&
-    arch.attentionType !== 'AlternatingSinkGQA';
-  const hasAlternatingLayers = Boolean(arch.alternatingLayers && arch.alternatingLayers.length > 0);
-  const alternatingSpecs = arch.alternatingLayers ?? [];
-
-  // Get sub-blocks data (only used for expandable attention types)
-  const attnFlow = isAttnExpandable ? getAttentionSubBlocks(arch) : null;
-  const ffnFlow = getFFNSubBlocks(arch);
-  const denseFFNFlow = hasDenseLayers ? getFFNSubBlocks(arch, { useDenseFFNDim: true }) : null;
-
-  const attnExpanded = isAttnExpandable && expandedBlocks.has('attention');
-  const ffnExpanded = expandedBlocks.has(isMoE ? 'experts' : 'ffn');
-  const txExpanded = expandedBlocks.has('transformer');
-  const denseTxExpanded = expandedBlocks.has('denseTransformer');
-  const denseAttnExpanded = isAttnExpandable && expandedBlocks.has('denseAttention');
-  const denseFFNExpanded = expandedBlocks.has('denseFFN');
-
-  // Hash-routed MoE prefix block (DeepSeek V4: first N layers route by token id).
-  const hasHashBlock = isMoE && (arch.hashRoutedLayers ?? 0) > 0;
-  const hashBlockExpanded = hasHashBlock && expandedBlocks.has('hashBlock');
-  const hashExpertsExpanded = hasHashBlock && expandedBlocks.has('hashExperts');
-
-  // Alternating block expand states (for models with alternating attention like gpt-oss)
-  const altBlockExpanded = [
-    hasAlternatingLayers && expandedBlocks.has('altBlock0'),
-    hasAlternatingLayers && expandedBlocks.has('altBlock1'),
-  ];
-  const altExpertsExpanded = [
-    hasAlternatingLayers && expandedBlocks.has('altExperts0'),
-    hasAlternatingLayers && expandedBlocks.has('altExperts1'),
-  ];
-
-  // Hybrid models (DeepSeek V4) expose an expandable attention drill-down inside
-  // each alternating block, revealing the sliding-window branch as an explicit
-  // block alongside the compressed branch. gpt-oss (AlternatingSinkGQA) keeps a
-  // static attention block. `getHybridAttentionSubBlocks` draws the specific
-  // local/compressed CSA-HCA flow, so hybrids built from other layer types
-  // (Kimi K3: KDA + gated MLA) opt out via alternatingAttentionExpandable and
-  // keep static attention blocks too. Note this is independent of
-  // `attentionExpandable`, which governs the single-block (non-alternating)
-  // drill-down — V4 disables that one while keeping this one.
-  const altAttnExpandable =
-    hasAlternatingLayers &&
-    arch.attentionType === 'Hybrid' &&
-    arch.alternatingAttentionExpandable !== false;
-  const altAttnExpanded = [
-    altAttnExpandable && expandedBlocks.has('altAttention0'),
-    altAttnExpandable && expandedBlocks.has('altAttention1'),
-  ];
-  const altAttnFlow: (SubBlockFlow | null)[] = altAttnExpandable
-    ? [0, 1].map((i) =>
-        alternatingSpecs[i] ? getHybridAttentionSubBlocks(arch, alternatingSpecs[i]) : null,
-      )
-    : [null, null];
-
-  // Calculate flow height for either sequential or parallel layouts
-  function getFlowHeight(flow: SubBlockFlow, hasLabel: boolean): number {
-    if (flow.layout === 'sequential') {
-      return (
-        flow.blocks.length * subBlockH +
-        Math.max(0, flow.blocks.length - 1) * subArrowH +
-        subPadY * 2 +
-        (hasLabel ? 24 : 0)
-      );
-    }
-    if (flow.layout === 'threeWay') {
-      const maxRows = Math.max(flow.leftPath.length, flow.middlePath.length, flow.rightPath.length);
-      const pathLabelsH = flow.leftLabel || flow.middleLabel || flow.rightLabel ? 16 : 0;
-      const splitH = subArrowH;
-      const hasIntermediate = flow.intermediateMergeBlocks.length > 0;
-      return (
-        subPadY * 2 +
-        (hasLabel ? 24 : 0) +
-        pathLabelsH +
-        splitH +
-        maxRows * subBlockH +
-        Math.max(0, maxRows - 1) * subArrowH +
-        (subArrowH + 4) +
-        (hasIntermediate
-          ? flow.intermediateMergeBlocks.length * subBlockH +
-            Math.max(0, flow.intermediateMergeBlocks.length - 1) * subArrowH +
-            (subArrowH + 4)
-          : 0) +
-        flow.finalMergeBlocks.length * subBlockH +
-        Math.max(0, flow.finalMergeBlocks.length - 1) * subArrowH
-      );
-    }
-    const maxRows = Math.max(flow.leftPath.length, flow.rightPath.length);
-    const pathLabelsH = flow.leftLabel || flow.rightLabel ? 16 : 0;
-    const splitH = subArrowH;
-    return (
-      subPadY * 2 +
-      (hasLabel ? 24 : 0) +
-      pathLabelsH +
-      splitH +
-      maxRows * subBlockH +
-      Math.max(0, maxRows - 1) * subArrowH +
-      (subArrowH + 4) +
-      flow.mergeBlocks.length * subBlockH +
-      Math.max(0, flow.mergeBlocks.length - 1) * subArrowH
-    );
-  }
-
-  // Calculate expanded section heights
-  const attnExpandedH = attnExpanded && attnFlow ? getFlowHeight(attnFlow, false) : 0;
-  const ffnExpandedH = ffnExpanded ? getFlowHeight(ffnFlow, true) : 0;
-  const denseAttnExpandedH = denseAttnExpanded && attnFlow ? getFlowHeight(attnFlow, false) : 0;
-  const denseFFNExpandedH =
-    denseFFNExpanded && denseFFNFlow ? getFlowHeight(denseFFNFlow, true) : 0;
-  const altExpertsExpandedH = [
-    altExpertsExpanded[0] ? getFlowHeight(ffnFlow, true) : 0,
-    altExpertsExpanded[1] ? getFlowHeight(ffnFlow, true) : 0,
-  ];
-  const altAttnExpandedH = [
-    altAttnExpanded[0] && altAttnFlow[0] ? getFlowHeight(altAttnFlow[0], false) : 0,
-    altAttnExpanded[1] && altAttnFlow[1] ? getFlowHeight(altAttnFlow[1], false) : 0,
-  ];
-  const hashExpertsExpandedH = hashExpertsExpanded ? getFlowHeight(ffnFlow, true) : 0;
-
-  // Compute vertical positions
-  let y = pad.top;
-
-  // Title
-  const titleY = y;
-  y += 44;
-
-  // Embedding
-  const embedY = y;
-  y += blockH + arrowH;
-
-  // === DENSE TRANSFORMER BLOCK (for MoE models with initial dense layers) ===
-  let denseTxStart = 0;
-  let denseNorm1Y = 0;
-  let denseAttnY = 0;
-  let denseAttnExpandedStartY = 0;
-  let denseMerge1Y = 0;
-  let denseNorm2Y = 0;
-  let denseFFNBlockY = 0;
-  let denseFFNExpandedStartY = 0;
-  let denseMerge2Y = 0;
-  let denseTxEnd = 0;
-
-  if (hasDenseLayers) {
-    denseTxStart = y;
-    if (denseTxExpanded) {
-      y += 14;
-
-      denseNorm1Y = y;
-      y += smallH + arrowH;
-
-      denseAttnY = y;
-      y += blockH;
-
-      if (denseAttnExpanded) y += drillGap;
-      denseAttnExpandedStartY = y;
-      if (denseAttnExpanded) {
-        y += denseAttnExpandedH;
-      }
-
-      y += 4;
-      denseMerge1Y = y + mergeGap / 2;
-      y += mergeGap;
-
-      y += arrowH;
-      denseNorm2Y = y;
-      y += smallH + arrowH;
-
-      denseFFNBlockY = y;
-      y += blockH;
-
-      denseFFNExpandedStartY = y;
-      if (denseFFNExpanded) {
-        y += denseFFNExpandedH;
-      }
-
-      denseMerge2Y = y + mergeGap / 2;
-      y += mergeGap;
-
-      y += 14;
-    } else {
-      y += collapsedTxH;
-    }
-    denseTxEnd = y;
-    y += arrowH;
-  }
-
-  // === HASH-ROUTED MoE PREFIX BLOCK (DeepSeek V4: first N layers use hash routing) ===
-  let hashTxStart = 0;
-  let hashNorm1Y = 0;
-  let hashAttnY = 0;
-  let hashMerge1Y = 0;
-  let hashNorm2Y = 0;
-  let hashExpertY = 0;
-  let hashFFNExpandedStartY = 0;
-  let hashMerge2Y = 0;
-  let hashTxEnd = 0;
-
-  if (hasHashBlock) {
-    hashTxStart = y;
-    if (hashBlockExpanded) {
-      y += 14;
-
-      hashNorm1Y = y;
-      y += smallH + arrowH;
-
-      hashAttnY = y;
-      y += blockH;
-
-      y += 4;
-      hashMerge1Y = y + mergeGap / 2;
-      y += mergeGap;
-
-      y += arrowH;
-      hashNorm2Y = y;
-      y += smallH + arrowH;
-
-      // Hash Router + Expert grid (drawExpertGrid lays out router above eY)
-      y += blockH + arrowH;
-      hashExpertY = y;
-      y += expertGridH;
-
-      hashFFNExpandedStartY = y;
-      if (hashExpertsExpanded) {
-        y += hashExpertsExpandedH;
-      }
-
-      hashMerge2Y = y + mergeGap / 2;
-      y += mergeGap;
-
-      y += 14;
-    } else {
-      y += collapsedTxH;
-    }
-    hashTxEnd = y;
-    y += arrowH;
-  }
-
-  // === ALTERNATING TRANSFORMER BLOCKS (for models like gpt-oss with alternating attention) ===
-  const altBlockStart = [0, 0];
-  const altBlockEnd = [0, 0];
-  const altNorm1Y = [0, 0];
-  const altAttnY = [0, 0];
-  const altAttnExpandedStartY = [0, 0];
-  const altMerge1Y = [0, 0];
-  const altNorm2Y = [0, 0];
-  const altRouterY = [0, 0];
-  const altExpertY = [0, 0];
-  const altFFNExpandedStartY = [0, 0];
-  const altMerge2Y = [0, 0];
-  let altIndicatorY = 0;
-
-  // Main transformer container (used when NOT hasAlternatingLayers)
-  const txStart = hasAlternatingLayers ? 0 : y;
-  let attnY = 0;
-  let attnExpandedStartY = 0;
-  let merge1Y = 0;
-  let norm1Y = 0;
-  let _routerY = 0;
-  let expertY = 0;
-  let ffnY = 0;
-  let ffnExpandedStartY = 0;
-  let merge2Y = 0;
-  let norm2Y = 0;
-  let txEnd = 0; // oxlint-disable-line no-useless-assignment -- overwritten when txExpanded
-
-  if (hasAlternatingLayers) {
-    // Layout two separate transformer blocks with alternating indicator between them
-    for (let bi = 0; bi < 2; bi++) {
-      altBlockStart[bi] = y;
-      const isExp = altBlockExpanded[bi];
-      const isExpExperts = altExpertsExpanded[bi];
-
-      if (isExp) {
-        y += 14;
-
-        altNorm1Y[bi] = y;
-        y += smallH + arrowH;
-
-        altAttnY[bi] = y;
-        y += blockH;
-
-        // Expanded hybrid-attention sub-blocks (sliding window + compressed)
-        if (altAttnExpanded[bi]) y += drillGap;
-        altAttnExpandedStartY[bi] = y;
-        if (altAttnExpanded[bi]) {
-          y += altAttnExpandedH[bi];
-        }
-
-        y += 4;
-        altMerge1Y[bi] = y + mergeGap / 2;
-        y += mergeGap;
-
-        y += arrowH;
-        altNorm2Y[bi] = y;
-        y += smallH + arrowH;
-
-        // MoE Router
-        altRouterY[bi] = y;
-        y += blockH + arrowH;
-
-        // Expert grid
-        altExpertY[bi] = y;
-        y += expertGridH;
-
-        // Expanded expert FFN sub-blocks
-        altFFNExpandedStartY[bi] = y;
-        if (isExpExperts) {
-          y += altExpertsExpandedH[bi];
-        }
-
-        altMerge2Y[bi] = y + mergeGap / 2;
-        y += mergeGap;
-
-        y += 14;
-      } else {
-        y += collapsedTxH;
-      }
-      altBlockEnd[bi] = y;
-
-      // Add alternating indicator between the two blocks
-      if (bi === 0) {
-        y += 6;
-        altIndicatorY = y + 10;
-        y += 24;
-      }
-    }
-  } else if (txExpanded) {
-    y += 14;
-
-    // Pre-LN: RMSNorm 1 before attention
-    norm1Y = y;
-    y += smallH + arrowH;
-
-    // Attention
-    attnY = y;
-    y += blockH;
-
-    // Expanded attention sub-blocks (only for non-MLA, non-AlternatingSinkGQA)
-    if (attnExpanded) y += drillGap;
-    attnExpandedStartY = y;
-    if (attnExpanded) {
-      y += attnExpandedH;
-    }
-
-    y += 4;
-    merge1Y = y + mergeGap / 2;
-    y += mergeGap;
-
-    // Pre-LN: RMSNorm 2 before FFN
-    y += arrowH;
-    norm2Y = y;
-    y += smallH + arrowH;
-
-    if (isMoE) {
-      // Router
-      _routerY = y;
-      y += blockH + arrowH;
-
-      // Expert grid
-      expertY = y;
-      y += expertGridH;
-    } else {
-      // Dense FFN
-      ffnY = y;
-      y += blockH;
-    }
-
-    // Expanded FFN sub-blocks (shared by MoE and dense paths)
-    ffnExpandedStartY = y;
-    if (ffnExpanded) {
-      y += ffnExpandedH;
-    }
-
-    merge2Y = y + mergeGap / 2;
-    y += mergeGap;
-
-    y += 14;
-  } else {
-    y += collapsedTxH;
-  }
-  txEnd = y;
-  y += arrowH;
-
-  // Final RMSNorm
-  const finalNormY = y;
-  y += smallH + arrowH;
-
-  // Output head
-  const outputY = y;
-  y += blockH + 16;
-
-  // Specs bar
-  const specsY = y;
-  const specsH = 56;
-  y += specsH;
-
-  const totalH = y + pad.bottom;
-
-  // Set SVG size
-  svg.attr('width', width).attr('height', totalH).attr('viewBox', `0 0 ${width} ${totalH}`);
-
-  // Defs (arrow markers)
-  const defs = svg.append('defs');
-  defs
-    .append('marker')
-    .attr('id', 'arch-arrow')
-    .attr('viewBox', '0 0 10 10')
-    .attr('refX', 5)
-    .attr('refY', 5)
-    .attr('markerWidth', 5)
-    .attr('markerHeight', 5)
-    .attr('orient', 'auto-start-reverse')
-    .append('path')
-    .attr('d', 'M 0 0 L 10 5 L 0 10 z')
-    .attr('fill', mutedFg);
-
-  defs
-    .append('marker')
-    .attr('id', 'arch-arrow-sub')
-    .attr('viewBox', '0 0 10 10')
-    .attr('refX', 5)
-    .attr('refY', 5)
-    .attr('markerWidth', 4)
-    .attr('markerHeight', 4)
-    .attr('orient', 'auto-start-reverse')
-    .append('path')
-    .attr('d', 'M 0 0 L 10 5 L 0 10 z')
-    .attr('fill', mutedFg);
-
-  const bgG = svg.append('g');
-  const g = svg.append('g');
-
-  // --- Helpers ---
-  function drawArrow(fromY: number, toY: number) {
-    g.append('line')
-      .attr('x1', cx)
-      .attr('y1', fromY)
-      .attr('x2', cx)
-      .attr('y2', toY - 2)
-      .attr('stroke', mutedFg)
-      .attr('stroke-width', 1.5)
-      .attr('marker-end', 'url(#arch-arrow)');
-  }
-
-  // Draw a +, −, or × glyph from geometric strokes, perfectly centered at
-  // (gx, gy). Rendering these as lines instead of <text> sidesteps font
-  // baseline drift, which left the symbol sitting slightly low inside the
-  // merge / expand circles regardless of dy tuning.
-  function drawCircleGlyph(gx: number, gy: number, color: string, symbol: string) {
-    const arm = 5;
-    const seg = (x1: number, y1: number, x2: number, y2: number) =>
-      g
-        .append('line')
-        .attr('x1', x1)
-        .attr('y1', y1)
-        .attr('x2', x2)
-        .attr('y2', y2)
-        .attr('stroke', color)
-        .attr('stroke-width', 2)
-        .attr('stroke-linecap', 'round')
-        .style('pointer-events', 'none');
-    if (symbol === '×') {
-      seg(gx - arm, gy - arm, gx + arm, gy + arm);
-      seg(gx - arm, gy + arm, gx + arm, gy - arm);
-    } else {
-      seg(gx - arm, gy, gx + arm, gy); // horizontal arm (+ and −)
-      if (symbol !== '−') seg(gx, gy - arm, gx, gy + arm); // vertical arm (+ only)
-    }
-  }
-
-  // Models with hyper-connections (mHC) replace the plain residual add with a
-  // multi-stream mixer; when present, residual merges render as an "mHC ×N" pill
-  // instead of a "+" circle. N = number of parallel residual streams.
-  const hcStreams = arch.hyperConnections ?? 0;
-  const isHyperConn = hcStreams > 1;
-
-  function drawResidualBypass(branchY: number, mergeY: number) {
-    // Tap the residual from the input stream ABOVE the norm (in the arrow gap)
-    // so the horizontal connector doesn't run across the RMSNorm block.
-    const tapY = branchY - arrowH / 2;
-    // Keep the node's vertical half-height = circleR so the spine arrows (drawn
-    // by callers to mergeY ± circleR) still meet its top/bottom edges.
-    const nodeHalfW = isHyperConn ? 25 : circleR;
-    bgG
-      .append('path')
-      .attr(
-        'd',
-        `M ${cx} ${tapY} L ${residLeftX} ${tapY} L ${residLeftX} ${mergeY} L ${cx - nodeHalfW} ${mergeY}`,
-      )
-      .attr('fill', 'none')
-      .attr('stroke', mutedFg)
-      .attr('stroke-width', 1.5)
-      .attr('opacity', 0.6);
-    if (isHyperConn) {
-      g.append('rect')
-        .attr('x', cx - nodeHalfW)
-        .attr('y', mergeY - circleR)
-        .attr('width', nodeHalfW * 2)
-        .attr('height', circleR * 2)
-        .attr('rx', circleR)
-        .attr('fill', bgSubtle)
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1.5);
-      g.append('text')
-        .attr('x', cx)
-        .attr('y', mergeY)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', fg)
-        .attr('font-size', '8.5px')
-        .attr('font-weight', 700)
-        .attr('font-family', 'inherit')
-        .text(`mHC ×${hcStreams}`);
-    } else {
-      g.append('circle')
-        .attr('cx', cx)
-        .attr('cy', mergeY)
-        .attr('r', circleR)
-        .attr('fill', bgSubtle)
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1.5);
-      drawCircleGlyph(cx, mergeY, fg, '+');
-    }
-  }
-
-  function drawBlock(
-    x: number,
-    by: number,
-    w: number,
-    h: number,
-    type: keyof typeof BLOCK_COLORS,
-    mainText: string,
-    subText?: string,
-  ) {
-    const c = getColor(type, isDark);
-    g.append('rect')
-      .attr('x', x)
-      .attr('y', by)
-      .attr('width', w)
-      .attr('height', h)
-      .attr('rx', 8)
-      .attr('fill', c.fill)
-      .attr('stroke', c.stroke)
-      .attr('stroke-width', 1.5);
-
-    g.append('text')
-      .attr('x', x + w / 2)
-      .attr('y', by + h / 2 - (subText ? 7 : 0))
-      .attr('text-anchor', 'middle')
-      .attr('dy', '0.35em')
-      .attr('fill', fg)
-      .attr('font-size', '13px')
-      .attr('font-weight', 600)
-      .attr('font-family', 'inherit')
-      .text(mainText);
-
-    if (subText) {
-      g.append('text')
-        .attr('x', x + w / 2)
-        .attr('y', by + h / 2 + 10)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', '11px')
-        .attr('font-family', 'inherit')
-        .text(subText);
-    }
-  }
-
-  function drawExpandableBlock(
-    x: number,
-    by: number,
-    w: number,
-    h: number,
-    type: keyof typeof BLOCK_COLORS,
-    mainText: string,
-    subText: string | undefined,
-    isBlockExpanded: boolean,
-    blockId: string,
-  ) {
-    const c = getColor(type, isDark);
-    g.append('rect')
-      .attr('x', x)
-      .attr('y', by)
-      .attr('width', w)
-      .attr('height', h)
-      .attr('rx', 8)
-      .attr('fill', c.fill)
-      .attr('stroke', c.stroke)
-      .attr('stroke-width', isBlockExpanded ? 2.5 : 1.5);
-
-    g.append('text')
-      .attr('x', x + w / 2 - 8)
-      .attr('y', by + h / 2 - (subText ? 7 : 0))
-      .attr('text-anchor', 'middle')
-      .attr('dy', '0.35em')
-      .attr('fill', fg)
-      .attr('font-size', '13px')
-      .attr('font-weight', 600)
-      .attr('font-family', 'inherit')
-      .style('pointer-events', 'none')
-      .text(mainText);
-
-    if (subText) {
-      g.append('text')
-        .attr('x', x + w / 2 - 8)
-        .attr('y', by + h / 2 + 10)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', '11px')
-        .attr('font-family', 'inherit')
-        .style('pointer-events', 'none')
-        .text(subText);
-    }
-
-    const iconX = x + w - 22;
-    const iconY = by + h / 2;
-
-    g.append('circle')
-      .attr('cx', iconX)
-      .attr('cy', iconY)
-      .attr('r', 10)
-      .attr('fill', isDark ? 'rgba(255,255,255,0.06)' : 'rgba(0,0,0,0.04)')
-      .attr('stroke', c.stroke)
-      .attr('stroke-width', 1)
-      .style('pointer-events', 'none');
-
-    drawCircleGlyph(iconX, iconY, c.stroke, isBlockExpanded ? '\u2212' : '+');
-
-    g.append('rect')
-      .attr('x', x)
-      .attr('y', by)
-      .attr('width', w)
-      .attr('height', h)
-      .attr('rx', 8)
-      .attr('fill', 'transparent')
-      .style('cursor', 'pointer')
-      .attr('data-testid', `expand-${blockId}`)
-      .on('click', () => {
-        onBlockClick(blockId);
-      });
-  }
-
-  const DEFAULT_SUB_BLOCK_FONT_SIZE = { name: '11px', detail: '9px' };
-
-  /** Draw a single sub-block at the given position */
-  function drawSingleSubBlock(
-    block: ArchSubBlock,
-    bx: number,
-    by: number,
-    subBw: number,
-    fontSize: { name: string; detail: string } = DEFAULT_SUB_BLOCK_FONT_SIZE,
-  ) {
-    const color = getSubBlockColor(block.type, isDark);
-    g.append('rect')
-      .attr('x', bx)
-      .attr('y', by)
-      .attr('width', subBw)
-      .attr('height', subBlockH)
-      .attr('rx', 6)
-      .attr('fill', color.fill)
-      .attr('stroke', color.stroke)
-      .attr('stroke-width', 1);
-
-    g.append('text')
-      .attr('x', bx + subBw / 2)
-      .attr('y', by + subBlockH / 2 - (block.detail ? 5 : 0))
-      .attr('text-anchor', 'middle')
-      .attr('dy', '0.35em')
-      .attr('fill', fg)
-      .attr('font-size', fontSize.name)
-      .attr('font-weight', 500)
-      .attr('font-family', 'inherit')
-      .text(block.name);
-
-    if (block.detail) {
-      g.append('text')
-        .attr('x', bx + subBw / 2)
-        .attr('y', by + subBlockH / 2 + 8)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', fontSize.detail)
-        .attr('font-family', 'inherit')
-        .text(block.detail);
-    }
-  }
-
-  /** Draw parallel sub-block flow (two columns converging into merge blocks) */
-  function drawParallelFlow(
-    flow: Extract<SubBlockFlow, { layout: 'parallel' }>,
-    startY: number,
-    x: number,
-    w: number,
-    label?: string,
-  ) {
-    let sy = startY;
-    const flowH = getFlowHeight(flow, Boolean(label));
-
-    g.append('rect')
-      .attr('x', x + 4)
-      .attr('y', sy)
-      .attr('width', w - 8)
-      .attr('height', flowH)
-      .attr('rx', 8)
-      .attr('fill', expandedBg)
-      .attr('stroke', borderColor)
-      .attr('stroke-width', 1)
-      .attr('stroke-dasharray', '4,3');
-
-    sy += subPadY;
-
-    if (label) {
-      g.append('text')
-        .attr('x', x + w / 2)
-        .attr('y', sy + 8)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', '10px')
-        .attr('font-weight', 600)
-        .attr('font-family', 'inherit')
-        .attr('letter-spacing', '0.5px')
-        .text(label.toUpperCase());
-      sy += 24;
-    }
-
-    const colGap = 10;
-    const innerPadCol = 12;
-    const colW = (w - 2 * innerPadCol - colGap) / 2;
-    const leftX = x + innerPadCol;
-    const rightX = x + innerPadCol + colW + colGap;
-    const leftCx = leftX + colW / 2;
-    const rightCx = rightX + colW / 2;
-    const mergeCx = x + w / 2;
-
-    if (flow.leftLabel || flow.rightLabel) {
-      if (flow.leftLabel) {
-        g.append('text')
-          .attr('x', leftCx)
-          .attr('y', sy + 6)
-          .attr('text-anchor', 'middle')
-          .attr('dy', '0.35em')
-          .attr('fill', mutedFg)
-          .attr('font-size', '9px')
-          .attr('font-weight', 600)
-          .attr('font-family', 'inherit')
-          .text(flow.leftLabel);
-      }
-      if (flow.rightLabel) {
-        g.append('text')
-          .attr('x', rightCx)
-          .attr('y', sy + 6)
-          .attr('text-anchor', 'middle')
-          .attr('dy', '0.35em')
-          .attr('fill', mutedFg)
-          .attr('font-size', '9px')
-          .attr('font-weight', 600)
-          .attr('font-family', 'inherit')
-          .text(flow.rightLabel);
-      }
-      sy += 16;
-    }
-
-    const splitTopY = sy;
-    const splitMidY = sy + subArrowH / 2;
-    sy += subArrowH;
-
-    const parallelStartY = sy;
-    const colFontSize = { name: '10px', detail: '8px' };
-
-    const maxRows = Math.max(flow.leftPath.length, flow.rightPath.length);
-    const rowsH = (n: number) => n * subBlockH + Math.max(0, n - 1) * subArrowH;
-    const colAreaH = rowsH(maxRows);
-    // Vertically center each column within the shared column area so an unequal
-    // split (e.g. 1 local vs 2 compressed) reads as an intentional branch merge
-    // rather than leaving the shorter column's connector dangling as a long
-    // unattached line beside the taller column.
-    const leftStartY = parallelStartY + (colAreaH - rowsH(flow.leftPath.length)) / 2;
-    const rightStartY = parallelStartY + (colAreaH - rowsH(flow.rightPath.length)) / 2;
-
-    g.append('line')
-      .attr('x1', mergeCx)
-      .attr('y1', splitTopY)
-      .attr('x2', mergeCx)
-      .attr('y2', splitMidY)
-      .attr('stroke', mutedFg)
-      .attr('stroke-width', 1);
-
-    g.append('path')
-      .attr('d', `M ${mergeCx} ${splitMidY} L ${leftCx} ${splitMidY} L ${leftCx} ${leftStartY - 2}`)
-      .attr('fill', 'none')
-      .attr('stroke', mutedFg)
-      .attr('stroke-width', 1)
-      .attr('marker-end', 'url(#arch-arrow-sub)');
-
-    g.append('path')
-      .attr(
-        'd',
-        `M ${mergeCx} ${splitMidY} L ${rightCx} ${splitMidY} L ${rightCx} ${rightStartY - 2}`,
-      )
-      .attr('fill', 'none')
-      .attr('stroke', mutedFg)
-      .attr('stroke-width', 1)
-      .attr('marker-end', 'url(#arch-arrow-sub)');
-
-    let lsy = leftStartY;
-    for (let i = 0; i < flow.leftPath.length; i++) {
-      drawSingleSubBlock(flow.leftPath[i], leftX, lsy, colW, colFontSize);
-      lsy += subBlockH;
-      if (i < flow.leftPath.length - 1) {
-        g.append('line')
-          .attr('x1', leftCx)
-          .attr('y1', lsy + 1)
-          .attr('x2', leftCx)
-          .attr('y2', lsy + subArrowH - 2)
-          .attr('stroke', mutedFg)
-          .attr('stroke-width', 1)
-          .attr('marker-end', 'url(#arch-arrow-sub)');
-        lsy += subArrowH;
-      }
-    }
-    const leftEndY = lsy;
-
-    let rsy = rightStartY;
-    for (let i = 0; i < flow.rightPath.length; i++) {
-      drawSingleSubBlock(flow.rightPath[i], rightX, rsy, colW, colFontSize);
-      rsy += subBlockH;
-      if (i < flow.rightPath.length - 1) {
-        g.append('line')
-          .attr('x1', rightCx)
-          .attr('y1', rsy + 1)
-          .attr('x2', rightCx)
-          .attr('y2', rsy + subArrowH - 2)
-          .attr('stroke', mutedFg)
-          .attr('stroke-width', 1)
-          .attr('marker-end', 'url(#arch-arrow-sub)');
-        rsy += subArrowH;
-      }
-    }
-    const rightEndY = rsy;
-
-    const mergeStartY = parallelStartY + colAreaH + subArrowH + 4;
-
-    const subInnerXLocal = x + 16;
-    const subInnerWLocal = w - 40;
-
-    const firstIsCircle = flow.mergeBlocks[0]?.circleSymbol;
-
-    if (firstIsCircle) {
-      const circleCy = mergeStartY + subBlockH / 2;
-      g.append('path')
-        .attr(
-          'd',
-          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${circleCy} L ${mergeCx - circleR - 2} ${circleCy}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-      g.append('path')
-        .attr(
-          'd',
-          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${circleCy} L ${mergeCx + circleR + 2} ${circleCy}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-    } else {
-      const convergeHorizY = Math.max(leftEndY, rightEndY) + (subArrowH + 4) / 2;
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${convergeHorizY} L ${mergeCx} ${convergeHorizY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${convergeHorizY} L ${mergeCx} ${convergeHorizY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('line')
-        .attr('x1', mergeCx)
-        .attr('y1', convergeHorizY)
-        .attr('x2', mergeCx)
-        .attr('y2', mergeStartY - 2)
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1)
-        .attr('marker-end', 'url(#arch-arrow-sub)');
-    }
-
-    let msy = mergeStartY;
-    for (let i = 0; i < flow.mergeBlocks.length; i++) {
-      const block = flow.mergeBlocks[i];
-      if (block.circleSymbol) {
-        const circleCy = msy + subBlockH / 2;
-        const symbolR = circleR + 2;
-        g.append('circle')
-          .attr('cx', mergeCx)
-          .attr('cy', circleCy)
-          .attr('r', symbolR)
-          .attr('fill', bgSubtle)
-          .attr('stroke', mutedFg)
-          .attr('stroke-width', 1.5);
-        drawCircleGlyph(mergeCx, circleCy, fg, block.circleSymbol);
-      } else {
-        drawSingleSubBlock(block, subInnerXLocal, msy, subInnerWLocal);
-      }
-      msy += subBlockH;
-      if (i < flow.mergeBlocks.length - 1) {
-        g.append('line')
-          .attr('x1', mergeCx)
-          .attr('y1', msy + 1)
-          .attr('x2', mergeCx)
-          .attr('y2', msy + subArrowH - 2)
-          .attr('stroke', mutedFg)
-          .attr('stroke-width', 1)
-          .attr('marker-end', 'url(#arch-arrow-sub)');
-        msy += subArrowH;
-      }
-    }
-  }
-
-  /** Draw three-way parallel flow (3 columns: left+middle converge first, then merge with right) */
-  function drawThreeWayFlow(
-    flow: Extract<SubBlockFlow, { layout: 'threeWay' }>,
-    startY: number,
-    x: number,
-    w: number,
-    _label?: string,
-  ) {
-    let sy = startY;
-    const flowH = getFlowHeight(flow, false);
-
-    g.append('rect')
-      .attr('x', x + 4)
-      .attr('y', sy)
-      .attr('width', w - 8)
-      .attr('height', flowH)
-      .attr('rx', 8)
-      .attr('fill', expandedBg)
-      .attr('stroke', borderColor)
-      .attr('stroke-width', 1)
-      .attr('stroke-dasharray', '4,3');
-
-    sy += subPadY;
-
-    const colGap = 6;
-    const innerPadCol = 8;
-    const colW = (w - 2 * innerPadCol - 2 * colGap) / 3;
-    const leftX = x + innerPadCol;
-    const middleX = x + innerPadCol + colW + colGap;
-    const rightX = x + innerPadCol + 2 * (colW + colGap);
-    const leftCx = leftX + colW / 2;
-    const middleCx = middleX + colW / 2;
-    const rightCx = rightX + colW / 2;
-    const mergeCxLocal = x + w / 2;
-    const qkMergeCx = (leftCx + middleCx) / 2;
-
-    if (flow.leftLabel || flow.middleLabel || flow.rightLabel) {
-      const labels = [
-        { label: flow.leftLabel, cx: leftCx },
-        { label: flow.middleLabel, cx: middleCx },
-        { label: flow.rightLabel, cx: rightCx },
-      ];
-      for (const { label: lbl, cx: lcx } of labels) {
-        if (lbl) {
-          g.append('text')
-            .attr('x', lcx)
-            .attr('y', sy + 6)
-            .attr('text-anchor', 'middle')
-            .attr('dy', '0.35em')
-            .attr('fill', mutedFg)
-            .attr('font-size', '9px')
-            .attr('font-weight', 600)
-            .attr('font-family', 'inherit')
-            .text(lbl);
-        }
-      }
-      sy += 16;
-    }
-
-    const splitTopY = sy;
-    const splitMidY = sy + subArrowH / 2;
-    sy += subArrowH;
-
-    const parallelStartY = sy;
-    const colFontSize = { name: '9px', detail: '7.5px' };
-
-    g.append('line')
-      .attr('x1', mergeCxLocal)
-      .attr('y1', splitTopY)
-      .attr('x2', mergeCxLocal)
-      .attr('y2', splitMidY)
-      .attr('stroke', mutedFg)
-      .attr('stroke-width', 1);
-
-    for (const colCx of [leftCx, middleCx, rightCx]) {
-      g.append('path')
-        .attr(
-          'd',
-          `M ${mergeCxLocal} ${splitMidY} L ${colCx} ${splitMidY} L ${colCx} ${parallelStartY - 2}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1)
-        .attr('marker-end', 'url(#arch-arrow-sub)');
-    }
-
-    const drawColumn = (path: ArchSubBlock[], colX: number, colCx: number) => {
-      let csy = parallelStartY;
-      for (let i = 0; i < path.length; i++) {
-        drawSingleSubBlock(path[i], colX, csy, colW, colFontSize);
-        csy += subBlockH;
-        if (i < path.length - 1) {
-          g.append('line')
-            .attr('x1', colCx)
-            .attr('y1', csy + 1)
-            .attr('x2', colCx)
-            .attr('y2', csy + subArrowH - 2)
-            .attr('stroke', mutedFg)
-            .attr('stroke-width', 1)
-            .attr('marker-end', 'url(#arch-arrow-sub)');
-          csy += subArrowH;
-        }
-      }
-      return csy;
-    };
-
-    const leftEndY = drawColumn(flow.leftPath, leftX, leftCx);
-    const middleEndY = drawColumn(flow.middlePath, middleX, middleCx);
-    const rightEndY = drawColumn(flow.rightPath, rightX, rightCx);
-
-    const maxRows = Math.max(flow.leftPath.length, flow.middlePath.length, flow.rightPath.length);
-    const hasIntermediate = flow.intermediateMergeBlocks.length > 0;
-    let finalMergeStartY: number;
-
-    if (hasIntermediate) {
-      const intermediateStartY =
-        parallelStartY + maxRows * subBlockH + Math.max(0, maxRows - 1) * subArrowH + subArrowH + 4;
-
-      const qkConvergeY = Math.max(leftEndY, middleEndY) + (subArrowH + 4) / 2;
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${qkConvergeY} L ${qkMergeCx} ${qkConvergeY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${middleCx} ${middleEndY + 1} L ${middleCx} ${qkConvergeY} L ${qkMergeCx} ${qkConvergeY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('line')
-        .attr('x1', qkMergeCx)
-        .attr('y1', qkConvergeY)
-        .attr('x2', qkMergeCx)
-        .attr('y2', intermediateStartY - 2)
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1)
-        .attr('marker-end', 'url(#arch-arrow-sub)');
-
-      const intermediateBlockW = middleX + colW - leftX;
-      let imsy = intermediateStartY;
-      for (let i = 0; i < flow.intermediateMergeBlocks.length; i++) {
-        drawSingleSubBlock(
-          flow.intermediateMergeBlocks[i],
-          leftX,
-          imsy,
-          intermediateBlockW,
-          colFontSize,
-        );
-        imsy += subBlockH;
-        if (i < flow.intermediateMergeBlocks.length - 1) {
-          g.append('line')
-            .attr('x1', qkMergeCx)
-            .attr('y1', imsy + 1)
-            .attr('x2', qkMergeCx)
-            .attr('y2', imsy + subArrowH - 2)
-            .attr('stroke', mutedFg)
-            .attr('stroke-width', 1)
-            .attr('marker-end', 'url(#arch-arrow-sub)');
-          imsy += subArrowH;
-        }
-      }
-
-      finalMergeStartY =
-        intermediateStartY +
-        flow.intermediateMergeBlocks.length * subBlockH +
-        Math.max(0, flow.intermediateMergeBlocks.length - 1) * subArrowH +
-        subArrowH +
-        4;
-
-      const finalConvergeY = Math.max(imsy, rightEndY) + (subArrowH + 4) / 2;
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${qkMergeCx} ${imsy + 1} L ${qkMergeCx} ${finalConvergeY} L ${mergeCxLocal} ${finalConvergeY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${finalConvergeY} L ${mergeCxLocal} ${finalConvergeY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('line')
-        .attr('x1', mergeCxLocal)
-        .attr('y1', finalConvergeY)
-        .attr('x2', mergeCxLocal)
-        .attr('y2', finalMergeStartY - 2)
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1)
-        .attr('marker-end', 'url(#arch-arrow-sub)');
-    } else {
-      finalMergeStartY =
-        parallelStartY + maxRows * subBlockH + Math.max(0, maxRows - 1) * subArrowH + subArrowH + 4;
-
-      const allEndY = Math.max(leftEndY, middleEndY, rightEndY);
-      const convergeY = allEndY + (subArrowH + 4) / 2;
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${convergeY} L ${mergeCxLocal} ${convergeY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('line')
-        .attr('x1', middleCx)
-        .attr('y1', middleEndY + 1)
-        .attr('x2', middleCx)
-        .attr('y2', convergeY)
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-      if (Math.abs(middleCx - mergeCxLocal) > 1) {
-        g.append('line')
-          .attr('x1', middleCx)
-          .attr('y1', convergeY)
-          .attr('x2', mergeCxLocal)
-          .attr('y2', convergeY)
-          .attr('stroke', mutedFg)
-          .attr('stroke-width', 1);
-      }
-
-      g.append('path')
-        .attr(
-          'd',
-          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${convergeY} L ${mergeCxLocal} ${convergeY}`,
-        )
-        .attr('fill', 'none')
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1);
-
-      g.append('line')
-        .attr('x1', mergeCxLocal)
-        .attr('y1', convergeY)
-        .attr('x2', mergeCxLocal)
-        .attr('y2', finalMergeStartY - 2)
-        .attr('stroke', mutedFg)
-        .attr('stroke-width', 1)
-        .attr('marker-end', 'url(#arch-arrow-sub)');
-    }
-
-    const subInnerXLocal = x + 16;
-    const subInnerWLocal = w - 40;
-    let fmsy = finalMergeStartY;
-    for (let i = 0; i < flow.finalMergeBlocks.length; i++) {
-      drawSingleSubBlock(flow.finalMergeBlocks[i], subInnerXLocal, fmsy, subInnerWLocal);
-      fmsy += subBlockH;
-      if (i < flow.finalMergeBlocks.length - 1) {
-        g.append('line')
-          .attr('x1', mergeCxLocal)
-          .attr('y1', fmsy + 1)
-          .attr('x2', mergeCxLocal)
-          .attr('y2', fmsy + subArrowH - 2)
-          .attr('stroke', mutedFg)
-          .attr('stroke-width', 1)
-          .attr('marker-end', 'url(#arch-arrow-sub)');
-        fmsy += subArrowH;
-      }
-    }
-  }
-
-  /** Dispatch to correct rendering based on flow layout */
-  function drawFlow(flow: SubBlockFlow, startY: number, x: number, w: number, label?: string) {
-    if (flow.layout === 'threeWay') {
-      drawThreeWayFlow(flow, startY, x, w, label);
-    } else if (flow.layout === 'parallel') {
-      drawParallelFlow(flow, startY, x, w, label);
-    }
-  }
-
-  /** Draw a collapsed transformer block (compact summary with expand icon) */
-  function drawCollapsedTransformerBlock(
-    x: number,
-    by: number,
-    w: number,
-    h: number,
-    label: string,
-    subtitle: string,
-    blockId: string,
-  ) {
-    g.append('rect')
-      .attr('x', x)
-      .attr('y', by)
-      .attr('width', w)
-      .attr('height', h)
-      .attr('rx', 10)
-      .attr('fill', bgSubtle)
-      .attr('stroke', borderColor)
-      .attr('stroke-width', 2)
-      .attr('stroke-dasharray', '6,3');
-
-    g.append('text')
-      .attr('x', x + w / 2 - 8)
-      .attr('y', by + h / 2 - 8)
-      .attr('text-anchor', 'middle')
-      .attr('dy', '0.35em')
-      .attr('fill', fg)
-      .attr('font-size', '13px')
-      .attr('font-weight', 600)
-      .attr('font-family', 'inherit')
-      .style('pointer-events', 'none')
-      .text(label);
-
-    g.append('text')
-      .attr('x', x + w / 2 - 8)
-      .attr('y', by + h / 2 + 10)
-      .attr('text-anchor', 'middle')
-      .attr('dy', '0.35em')
-      .attr('fill', mutedFg)
-      .attr('font-size', '11px')
-      .attr('font-family', 'inherit')
-      .style('pointer-events', 'none')
-      .text(subtitle);
-
-    const iconX = x + w - 22;
-    const iconY = by + h / 2;
-    g.append('circle')
-      .attr('cx', iconX)
-      .attr('cy', iconY)
-      .attr('r', 10)
-      .attr('fill', isDark ? 'rgba(255,255,255,0.06)' : 'rgba(0,0,0,0.04)')
-      .attr('stroke', borderColor)
-      .attr('stroke-width', 1)
-      .style('pointer-events', 'none');
-    drawCircleGlyph(iconX, iconY, mutedFg, '+');
-
-    g.append('rect')
-      .attr('x', x)
-      .attr('y', by)
-      .attr('width', w)
-      .attr('height', h)
-      .attr('rx', 10)
-      .attr('fill', 'transparent')
-      .style('cursor', 'pointer')
-      .attr('data-testid', `expand-${blockId}`)
-      .on('click', () => {
-        onBlockClick(blockId);
-      });
-  }
-
-  // === TITLE ===
-  const modelName = arch.developer || '';
-  const paramsSummary = [
-    `${formatParamCount(arch.totalParams)} total`,
-    isMoE ? `${formatParamCount(arch.activeParams)} active` : null,
-    arch.contextWindow ? `${formatContextWindow(arch.contextWindow)} context` : null,
-  ]
-    .filter(Boolean)
-    .join(' \u00B7 ');
-
-  g.append('text')
-    .attr('x', cx)
-    .attr('y', titleY + 14)
-    .attr('text-anchor', 'middle')
-    .attr('fill', fg)
-    .attr('font-size', '15px')
-    .attr('font-weight', 700)
-    .attr('font-family', 'inherit')
-    .text(modelName);
-
-  g.append('text')
-    .attr('x', cx)
-    .attr('y', titleY + 32)
-    .attr('text-anchor', 'middle')
-    .attr('fill', mutedFg)
-    .attr('font-size', '11px')
-    .attr('font-family', 'inherit')
-    .text(paramsSummary);
-
-  // === EMBEDDING ===
-  const embedSub = [
-    arch.hiddenSize ? `d = ${arch.hiddenSize.toLocaleString()}` : null,
-    arch.vocabSize ? `vocab = ${arch.vocabSize.toLocaleString()}` : null,
-  ]
-    .filter(Boolean)
-    .join('  \u00B7  ');
-  drawBlock(pad.left, embedY, bw, blockH, 'embedding', 'Token Embedding', embedSub || undefined);
-
-  // A block's incoming arrow should land on its first RMSNorm when the block is
-  // expanded (so there's a continuous line into the norm, through the container
-  // border), or on the block's top when collapsed.
-  const denseEntryY = denseTxExpanded ? denseNorm1Y : denseTxStart;
-  const hashEntryY = hashBlockExpanded ? hashNorm1Y : hashTxStart;
-  const altEntryY = [
-    altBlockExpanded[0] ? altNorm1Y[0] : altBlockStart[0],
-    altBlockExpanded[1] ? altNorm1Y[1] : altBlockStart[1],
-  ];
-  const mainEntryY = txExpanded ? norm1Y : txStart;
-
-  drawArrow(
-    embedY + blockH,
-    hasDenseLayers
-      ? denseEntryY
-      : hasHashBlock
-        ? hashEntryY
-        : hasAlternatingLayers
-          ? altEntryY[0]
-          : mainEntryY,
-  );
-
-  // === DENSE TRANSFORMER BLOCK (for MoE models with initial dense layers) ===
-  if (hasDenseLayers) {
-    if (denseTxExpanded) {
-      // Container
-      g.append('rect')
-        .attr('x', pad.left - 4)
-        .attr('y', denseTxStart)
-        .attr('width', bw + 8)
-        .attr('height', denseTxEnd - denseTxStart)
-        .attr('rx', 10)
-        .attr('fill', 'none')
-        .attr('stroke', borderColor)
-        .attr('stroke-width', 2)
-        .attr('stroke-dasharray', '6,3');
-
-      // Dense layer badge (clickable to collapse)
-      const denseLabel = `\u2212 \u00D7${denseLayerCount} dense layers`;
-      const denseBadgeW = denseLabel.length * 7 + 16;
-      g.append('rect')
-        .attr('x', width - pad.right - denseBadgeW - 4)
-        .attr('y', denseTxStart - 11)
-        .attr('width', denseBadgeW)
-        .attr('height', 22)
-        .attr('rx', 11)
-        .attr('fill', bgSubtle)
-        .attr('stroke', borderColor)
-        .attr('stroke-width', 1);
-      g.append('text')
-        .attr('x', width - pad.right - denseBadgeW / 2 - 4)
-        .attr('y', denseTxStart)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', '11px')
-        .attr('font-weight', 600)
-        .attr('font-family', 'inherit')
-        .text(denseLabel);
-      g.append('rect')
-        .attr('x', width - pad.right - denseBadgeW - 4)
-        .attr('y', denseTxStart - 11)
-        .attr('width', denseBadgeW)
-        .attr('height', 22)
-        .attr('rx', 11)
-        .attr('fill', 'transparent')
-        .style('cursor', 'pointer')
-        .attr('data-testid', 'collapse-denseTransformer')
-        .on('click', () => onBlockClick('denseTransformer'));
-
-      // RMSNorm 1 (Pre-LN)
-      drawBlock(innerX, denseNorm1Y, innerW, smallH, 'norm', 'RMSNorm');
-      drawArrow(denseNorm1Y + smallH, denseAttnY);
-
-      // Attention (expandable only for non-MLA types)
-      const denseAttnLabel = denseLayerAttentionLabel(arch);
-      const denseHeadSub = arch.numHeads ? `${arch.numHeads} heads` : undefined;
-      if (isAttnExpandable) {
-        drawExpandableBlock(
-          innerX,
-          denseAttnY,
-          innerW,
-          blockH,
-          'attention',
-          denseAttnLabel,
-          denseHeadSub,
-          denseAttnExpanded,
-          'denseAttention',
-        );
-        if (denseAttnExpanded && attnFlow) {
-          drawFlow(attnFlow, denseAttnExpandedStartY, innerX, innerW);
-        }
-      } else {
-        drawBlock(innerX, denseAttnY, innerW, blockH, 'attention', denseAttnLabel, denseHeadSub);
-      }
-
-      const denseAttnBottom = denseAttnExpanded
-        ? denseAttnExpandedStartY + denseAttnExpandedH + 4
-        : denseAttnY + blockH + 4;
-      drawArrow(denseAttnBottom, denseMerge1Y - circleR);
-      drawResidualBypass(denseNorm1Y, denseMerge1Y);
-      drawArrow(denseMerge1Y + circleR, denseNorm2Y);
-
-      // RMSNorm 2 (Pre-LN)
-      drawBlock(innerX, denseNorm2Y, innerW, smallH, 'norm', 'RMSNorm');
-      drawArrow(denseNorm2Y + smallH, denseFFNBlockY);
-
-      // Dense FFN (expandable)
-      const denseFFNSub = arch.denseFFNDim
-        ? `intermediate = ${arch.denseFFNDim.toLocaleString()}`
-        : 'Dense Feed-Forward';
-      drawExpandableBlock(
-        innerX,
-        denseFFNBlockY,
-        innerW,
-        blockH,
-        'ffn',
-        'Dense FFN',
-        denseFFNSub,
-        denseFFNExpanded,
-        'denseFFN',
-      );
-
-      if (denseFFNExpanded && denseFFNFlow) {
-        drawFlow(denseFFNFlow, denseFFNExpandedStartY, innerX, innerW, `${ffnVariant} FFN`);
-      }
-
-      const denseFFNBottom = denseFFNExpanded
-        ? denseFFNExpandedStartY + denseFFNExpandedH + 4
-        : denseFFNBlockY + blockH + 4;
-      drawArrow(denseFFNBottom, denseMerge2Y - circleR);
-      drawResidualBypass(denseNorm2Y, denseMerge2Y);
-    } else {
-      // Collapsed dense transformer block
-      const denseSub = `\u00D7${denseLayerCount} dense layers${arch.denseFFNDim ? ` \u00B7 FFN = ${arch.denseFFNDim.toLocaleString()}` : ''}`;
-      drawCollapsedTransformerBlock(
-        pad.left,
-        denseTxStart,
-        bw,
-        collapsedTxH,
-        'Dense Transformer Block',
-        denseSub,
-        'denseTransformer',
-      );
-    }
-
-    // Arrow from dense block to next block
-    drawArrow(
-      denseTxEnd,
-      hasHashBlock ? hashEntryY : hasAlternatingLayers ? altEntryY[0] : mainEntryY,
-    );
-  }
-
-  // Compute labels
-  const attnLabel = getAttentionLabel(arch.attentionType);
-  const mainLayerCount = hasDenseLayers ? moeLayerCount : arch.numLayers;
-  const layerSuffix = hasDenseLayers ? ' MoE layers' : ' layers';
-  const layerLabel = mainLayerCount ? `\u00D7${mainLayerCount}${layerSuffix}` : 'Transformer Block';
-
-  /** Helper: draw a MoE expert grid (used in both alternating and standard blocks) */
-  function drawExpertGrid(
-    eY: number,
-    isExpExpanded: boolean,
-    expandedH: number,
-    expandedStartY: number,
-    n2Y: number,
-    m2Y: number,
-    expertBlockId: string,
-    routerLabel = 'MoE Router',
-    routerSubOverride?: string,
-  ) {
-    const routerSub = routerSubOverride ?? expertRouterSummary(arch);
-    const rY = n2Y + smallH + arrowH;
-    drawBlock(innerX, rY, innerW, blockH, 'router', routerLabel, routerSub);
-    drawArrow(rY + blockH, rY + blockH + arrowH);
-
-    const ec = getColor('expert', isDark);
-    const eac = getColor('expertActive', isDark);
-
-    g.append('rect')
-      .attr('x', innerX)
-      .attr('y', eY)
-      .attr('width', innerW)
-      .attr('height', expertGridH)
-      .attr('rx', 8)
-      .attr('fill', ec.fill)
-      .attr('stroke', ec.stroke)
-      .attr('stroke-width', isExpExpanded ? 2.5 : 1)
-      .attr('stroke-dasharray', '4,3');
-
-    const numActive = arch.activeExperts || 2;
-    const numShow = Math.min(arch.numExperts || 8, 6);
-    const showEllipsis = (arch.numExperts || 0) > numShow;
-    const totalBoxes = numShow + (showEllipsis ? 2 : 0);
-    const totalBoxWidth = totalBoxes * expertSize + (totalBoxes - 1) * expertGap;
-    let ex = cx - totalBoxWidth / 2;
-    const ey = eY + (expertGridH - expertSize) / 2;
-
-    for (let i = 0; i < numShow; i++) {
-      const isActive = i < numActive;
-      g.append('rect')
-        .attr('x', ex)
-        .attr('y', ey)
-        .attr('width', expertSize)
-        .attr('height', expertSize)
-        .attr('rx', 6)
-        .attr('fill', isActive ? eac.fill : bgSubtle)
-        .attr('stroke', isActive ? eac.stroke : borderColor)
-        .attr('stroke-width', isActive ? 1.5 : 1)
-        .attr('opacity', isActive ? 1 : 0.6)
-        .style('pointer-events', 'none');
-      g.append('text')
-        .attr('x', ex + expertSize / 2)
-        .attr('y', ey + expertSize / 2)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', isActive ? fg : mutedFg)
-        .attr('font-size', '9px')
-        .attr('font-weight', isActive ? 600 : 400)
-        .attr('font-family', 'inherit')
-        .style('pointer-events', 'none')
-        .text(`E${i + 1}`);
-      ex += expertSize + expertGap;
-    }
-
-    if (showEllipsis) {
-      g.append('text')
-        .attr('x', ex + expertSize / 2)
-        .attr('y', ey + expertSize / 2)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', '14px')
-        .attr('font-weight', 700)
-        .style('pointer-events', 'none')
-        .text('\u00B7\u00B7\u00B7');
-      ex += expertSize + expertGap;
-      const countText = `\u00D7${arch.numExperts}`;
-      g.append('rect')
-        .attr('x', ex)
-        .attr('y', ey + 3)
-        .attr('width', expertSize)
-        .attr('height', expertSize - 6)
-        .attr('rx', 4)
-        .attr('fill', bgSubtle)
-        .attr('stroke', borderColor)
-        .attr('stroke-width', 1)
-        .style('pointer-events', 'none');
-      g.append('text')
-        .attr('x', ex + expertSize / 2)
-        .attr('y', ey + expertSize / 2)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', '9px')
-        .attr('font-weight', 600)
-        .attr('font-family', 'inherit')
-        .style('pointer-events', 'none')
-        .text(countText);
-    }
-
-    const expIconX = innerX + innerW - 18;
-    const expIconY = eY + expertGridH / 2;
-    g.append('circle')
-      .attr('cx', expIconX)
-      .attr('cy', expIconY)
-      .attr('r', 10)
-      .attr('fill', isDark ? 'rgba(255,255,255,0.06)' : 'rgba(0,0,0,0.04)')
-      .attr('stroke', ec.stroke)
-      .attr('stroke-width', 1)
-      .style('pointer-events', 'none');
-    drawCircleGlyph(expIconX, expIconY, ec.stroke, isExpExpanded ? '\u2212' : '+');
-
-    g.append('rect')
-      .attr('x', innerX)
-      .attr('y', eY)
-      .attr('width', innerW)
-      .attr('height', expertGridH)
-      .attr('rx', 8)
-      .attr('fill', 'transparent')
-      .style('cursor', 'pointer')
-      .attr('data-testid', `expand-${expertBlockId}`)
-      .on('click', () => {
-        onBlockClick(expertBlockId);
-      });
-
-    if (isExpExpanded) {
-      drawFlow(ffnFlow, expandedStartY, innerX, innerW, `Expert FFN (${ffnVariant})`);
-    }
-
-    const expertBottom = isExpExpanded ? expandedStartY + expandedH : eY + expertGridH;
-    drawArrow(expertBottom, m2Y - circleR);
-    drawResidualBypass(n2Y, m2Y);
-  }
-
-  // === HASH-ROUTED MoE PREFIX BLOCK (DeepSeek V4: first N layers route by token id) ===
-  if (hasHashBlock) {
-    if (hashBlockExpanded) {
-      // Container
-      g.append('rect')
-        .attr('x', pad.left - 4)
-        .attr('y', hashTxStart)
-        .attr('width', bw + 8)
-        .attr('height', hashTxEnd - hashTxStart)
-        .attr('rx', 10)
-        .attr('fill', 'none')
-        .attr('stroke', borderColor)
-        .attr('stroke-width', 2)
-        .attr('stroke-dasharray', '6,3');
-
-      // Collapse badge
-      const hashBadge = `− ×${arch.hashRoutedLayers} layers`;
-      const hashBadgeW = hashBadge.length * 7 + 16;
-      g.append('rect')
-        .attr('x', width - pad.right - hashBadgeW - 4)
-        .attr('y', hashTxStart - 11)
-        .attr('width', hashBadgeW)
-        .attr('height', 22)
-        .attr('rx', 11)
-        .attr('fill', bgSubtle)
-        .attr('stroke', borderColor)
-        .attr('stroke-width', 1);
-      g.append('text')
-        .attr('x', width - pad.right - hashBadgeW / 2 - 4)
-        .attr('y', hashTxStart)
-        .attr('text-anchor', 'middle')
-        .attr('dy', '0.35em')
-        .attr('fill', mutedFg)
-        .attr('font-size', '11px')
-        .attr('font-weight', 600)
-        .attr('font-family', 'inherit')
-        .text(hashBadge);
-      g.append('rect')
-        .attr('x', width - pad.right - hashBadgeW - 4)
-        .attr('y', hashTxStart - 11)
-        .attr('width', hashBadgeW)
-        .attr('height', 22)
-        .attr('rx', 11)
-        .attr('fill', 'transparent')
-        .style('cursor', 'pointer')
-        .attr('data-testid', 'collapse-hashBlock')
-        .on('click', () => onBlockClick('hashBlock'));
-
-      // RMSNorm 1
-      drawBlock(innerX, hashNorm1Y, innerW, smallH, 'norm', 'RMSNorm');
-      drawArrow(hashNorm1Y + smallH, hashAttnY);
-
-      // Attention — static hybrid block (same attention stack as the rest)
-      const hashHeadSub = [
-        arch.numHeads ? `${arch.numHeads} heads` : null,
-        arch.numKVHeads ? `${arch.numKVHeads} KV heads` : null,
-      ]
-        .filter(Boolean)
-        .join('  ·  ');
-      drawBlock(
-        innerX,
-        hashAttnY,
-        innerW,
-        blockH,
-        'attention',
-        attnLabel,
-        hashHeadSub || undefined,
-      );
-
-      drawArrow(hashAttnY + blockH + 4, hashMerge1Y - circleR);
-      drawResidualBypass(hashNorm1Y, hashMerge1Y);
-      drawArrow(hashMerge1Y + circleR, hashNorm2Y);
-
-      // RMSNorm 2
-      drawBlock(innerX, hashNorm2Y, innerW, smallH, 'norm', 'RMSNorm');
-      drawArrow(hashNorm2Y + smallH, hashNorm2Y + smallH + arrowH);
-
-      // Hash Router + Expert grid (token-id → fixed experts, not a learned gate)
-      const hashShared = sharedExpertCount(arch);
-      const hashRoutedCount = (arch.numExperts ?? 0) - hashShared;
-      const hashRouterSub = `token-id → ${arch.activeExperts} of ${hashRoutedCount}${hashShared > 0 ? ` + ${hashShared} shared` : ''}`;
-      drawExpertGrid(
-        hashExpertY,
-        hashExpertsExpanded,
-        hashExpertsExpandedH,
-        hashFFNExpandedStartY,
-        hashNorm2Y,
-        hashMerge2Y,
-        'hashExperts',
-        'Hash Router',
-        hashRouterSub,
-      );
-    } else {
-      const hashSub = `×${arch.hashRoutedLayers} first layers · token-id → experts`;
-      drawCollapsedTransformerBlock(
-        pad.left,
-        hashTxStart,
-        bw,
-        collapsedTxH,
-        'Hash-Routed MoE',
-        hashSub,
-        'hashBlock',
-      );
-    }
-
-    // Arrow from the hash block to the first alternating block (or main transformer)
-    drawArrow(hashTxEnd, hasAlternatingLayers ? altEntryY[0] : mainEntryY);
-  }
-
-  // === ALTERNATING TRANSFORMER BLOCKS (gpt-oss style) ===
-  if (hasAlternatingLayers) {
-    for (let bi = 0; bi < 2; bi++) {
-      const spec = alternatingSpecs[bi];
-      const blockId = `altBlock${bi}`;
-      const expertsId = `altExperts${bi}`;
-      const isExp = altBlockExpanded[bi];
-      const isExpExperts = altExpertsExpanded[bi];
-
-      if (isExp) {
-        // === EXPANDED ALTERNATING BLOCK ===
-        g.append('rect')
-          .attr('x', pad.left - 4)
-          .attr('y', altBlockStart[bi])
-          .attr('width', bw + 8)
-          .attr('height', altBlockEnd[bi] - altBlockStart[bi])
-          .attr('rx', 10)
-          .attr('fill', 'none')
-          .attr('stroke', borderColor)
-          .attr('stroke-width', 2)
-          .attr('stroke-dasharray', '6,3');
-
-        // Collapse badge
-        const badgeLabel = `\u2212 \u00D7${spec.count} layers`;
-        const badgeW = badgeLabel.length * 7 + 16;
-        g.append('rect')
-          .attr('x', width - pad.right - badgeW - 4)
-          .attr('y', altBlockStart[bi] - 11)
-          .attr('width', badgeW)
-          .attr('height', 22)
-          .attr('rx', 11)
-          .attr('fill', bgSubtle)
-          .attr('stroke', borderColor)
-          .attr('stroke-width', 1);
-        g.append('text')
-          .attr('x', width - pad.right - badgeW / 2 - 4)
-          .attr('y', altBlockStart[bi])
-          .attr('text-anchor', 'middle')
-          .attr('dy', '0.35em')
-          .attr('fill', mutedFg)
-          .attr('font-size', '11px')
-          .attr('font-weight', 600)
-          .attr('font-family', 'inherit')
-          .text(badgeLabel);
-        g.append('rect')
-          .attr('x', width - pad.right - badgeW - 4)
-          .attr('y', altBlockStart[bi] - 11)
-          .attr('width', badgeW)
-          .attr('height', 22)
-          .attr('rx', 11)
-          .attr('fill', 'transparent')
-          .style('cursor', 'pointer')
-          .attr('data-testid', `collapse-${blockId}`)
-          .on('click', () => onBlockClick(blockId));
-
-        // RMSNorm 1
-        drawBlock(innerX, altNorm1Y[bi], innerW, smallH, 'norm', 'RMSNorm');
-        drawArrow(altNorm1Y[bi] + smallH, altAttnY[bi]);
-
-        // Attention — expandable hybrid drill-down (DeepSeek V4 CSA/HCA) or a
-        // static block (gpt-oss AlternatingSinkGQA).
-        const headSub = [
-          arch.numHeads ? `${arch.numHeads} heads` : null,
-          arch.numKVHeads ? `${arch.numKVHeads} KV heads` : null,
-        ]
-          .filter(Boolean)
-          .join('  \u00B7  ');
-        // Sliding-window note is per layer-type: a hybrid model (e.g. DeepSeek
-        // V4) carries the window on every attention variant, whereas gpt-oss
-        // only puts it on its sliding block. Drive it off the spec, not bi.
-        const specWindow = spec.slidingWindow;
-        const attnSub = specWindow
-          ? `${headSub}${headSub ? '  \u00B7  ' : ''}window=${specWindow}`
-          : headSub || undefined;
-        if (altAttnExpandable) {
-          drawExpandableBlock(
-            innerX,
-            altAttnY[bi],
-            innerW,
-            blockH,
-            'attention',
-            spec.label,
-            attnSub,
-            altAttnExpanded[bi],
-            `altAttention${bi}`,
-          );
-          const flow = altAttnFlow[bi];
-          if (altAttnExpanded[bi] && flow) {
-            drawFlow(flow, altAttnExpandedStartY[bi], innerX, innerW);
-          }
-        } else {
-          drawBlock(innerX, altAttnY[bi], innerW, blockH, 'attention', spec.label, attnSub);
-        }
-
-        const aBottom = altAttnExpanded[bi]
-          ? altAttnExpandedStartY[bi] + altAttnExpandedH[bi] + 4
-          : altAttnY[bi] + blockH + 4;
-        drawArrow(aBottom, altMerge1Y[bi] - circleR);
-        drawResidualBypass(altNorm1Y[bi], altMerge1Y[bi]);
-        drawArrow(altMerge1Y[bi] + circleR, altNorm2Y[bi]);
-
-        // RMSNorm 2
-        drawBlock(innerX, altNorm2Y[bi], innerW, smallH, 'norm', 'RMSNorm');
-        drawArrow(altNorm2Y[bi] + smallH, altNorm2Y[bi] + smallH + arrowH);
-
-        // MoE Router + Expert Grid
-        drawExpertGrid(
-          altExpertY[bi],
-          isExpExperts,
-          altExpertsExpandedH[bi],
-          altFFNExpandedStartY[bi],
-          altNorm2Y[bi],
-          altMerge2Y[bi],
-          expertsId,
-        );
-      } else {
-        // === COLLAPSED ALTERNATING BLOCK ===
-        const collapsedSub = `\u00D7${spec.count} layers \u00B7 Top-${arch.activeExperts}/${arch.numExperts} MoE`;
-        drawCollapsedTransformerBlock(
-          pad.left,
-          altBlockStart[bi],
-          bw,
-          collapsedTxH,
-          spec.label,
-          collapsedSub,
-          blockId,
-        );
-      }
-
-      // Draw alternating indicator between the two blocks
-      if (bi === 0) {
-        // Arrow from block 0 end through indicator to block 1 start (drawn first, behind text)
-        drawArrow(altBlockEnd[0] + 2, altEntryY[1]);
-
-        // Opaque background rect behind the label so it doesn't overlap the arrow
-        const cardBg = isDark ? '#131416' : '#eaebec';
-        const labelText = `\u21C5 ${arch.alternatingNote ?? 'alternating every layer'}`;
-        const labelPadX = 6;
-        const labelPadY = 4;
-        const labelFontSize = 10;
-        const estLabelW = labelText.length * labelFontSize * 0.55 + labelPadX * 2;
-        const estLabelH = labelFontSize + labelPadY * 2;
-        g.append('rect')
-          .attr('x', cx - estLabelW / 2)
-          .attr('y', altIndicatorY - estLabelH / 2)
-          .attr('width', estLabelW)
-          .attr('height', estLabelH)
-          .attr('rx', 4)
-          .attr('fill', cardBg);
-
-        g.append('text')
-          .attr('x', cx)
-          .attr('y', altIndicatorY)
-          .attr('text-anchor', 'middle')
-          .attr('dy', '0.35em')
-          .attr('fill', mutedFg)
-          .attr('font-size', `${labelFontSize}px`)
-          .attr('font-weight', 500)
-          .attr('font-family', 'inherit')
-          .attr('data-testid', 'alternating-indicator')
-          .text(labelText);
-      }
-    }
-
-    // Arrow from last alternating block to final norm
-    drawArrow(altBlockEnd[1], finalNormY);
-  } else if (txExpanded) {
-    // === MAIN TRANSFORMER CONTAINER (expanded, non-alternating) ===
-    g.append('rect')
-      .attr('x', pad.left - 4)
-      .attr('y', txStart)
-      .attr('width', bw + 8)
-      .attr('height', txEnd - txStart)
-      .attr('rx', 10)
-      .attr('fill', 'none')
-      .attr('stroke', borderColor)
-      .attr('stroke-width', 2)
-      .attr('stroke-dasharray', '6,3');
-
-    // Layer badge (clickable to collapse)
-    const collapseBadgeLabel = `\u2212 ${layerLabel}`;
-    const badgeW = collapseBadgeLabel.length * 7 + 16;
-    g.append('rect')
-      .attr('x', width - pad.right - badgeW - 4)
-      .attr('y', txStart - 11)
-      .attr('width', badgeW)
-      .attr('height', 22)
-      .attr('rx', 11)
-      .attr('fill', bgSubtle)
-      .attr('stroke', borderColor)
-      .attr('stroke-width', 1);
-    g.append('text')
-      .attr('x', width - pad.right - badgeW / 2 - 4)
-      .attr('y', txStart)
-      .attr('text-anchor', 'middle')
-      .attr('dy', '0.35em')
-      .attr('fill', mutedFg)
-      .attr('font-size', '11px')
-      .attr('font-weight', 600)
-      .attr('font-family', 'inherit')
-      .text(collapseBadgeLabel);
-    g.append('rect')
-      .attr('x', width - pad.right - badgeW - 4)
-      .attr('y', txStart - 11)
-      .attr('width', badgeW)
-      .attr('height', 22)
-      .attr('rx', 11)
-      .attr('fill', 'transparent')
-      .style('cursor', 'pointer')
-      .attr('data-testid', 'collapse-transformer')
-      .on('click', () => onBlockClick('transformer'));
-
-    // === RMSNORM 1 (before attention — Pre-LN) ===
-    drawBlock(innerX, norm1Y, innerW, smallH, 'norm', 'RMSNorm');
-    drawArrow(norm1Y + smallH, attnY);
-
-    // === ATTENTION ===
-    const headSub = [
-      arch.numHeads ? `${arch.numHeads} heads` : null,
-      arch.numKVHeads ? `${arch.numKVHeads} KV heads` : null,
-    ]
-      .filter(Boolean)
-      .join('  \u00B7  ');
-
-    if (isAttnExpandable) {
-      drawExpandableBlock(
-        innerX,
-        attnY,
-        innerW,
-        blockH,
-        'attention',
-        attnLabel,
-        headSub || undefined,
-        attnExpanded,
-        'attention',
-      );
-      if (attnExpanded && attnFlow) {
-        drawFlow(attnFlow, attnExpandedStartY, innerX, innerW);
-      }
-    } else {
-      // MLA: non-expandable static block
-      drawBlock(innerX, attnY, innerW, blockH, 'attention', attnLabel, headSub || undefined);
-    }
-
-    const attnBottom = attnExpanded ? attnExpandedStartY + attnExpandedH + 4 : attnY + blockH + 4;
-    drawArrow(attnBottom, merge1Y - circleR);
-    drawResidualBypass(norm1Y, merge1Y);
-    drawArrow(merge1Y + circleR, norm2Y);
-
-    // === RMSNORM 2 (before FFN — Pre-LN) ===
-    drawBlock(innerX, norm2Y, innerW, smallH, 'norm', 'RMSNorm');
-    drawArrow(norm2Y + smallH, norm2Y + smallH + arrowH);
-
-    if (isMoE) {
-      // === ROUTER + EXPERTS (via helper) ===
-      drawExpertGrid(
-        expertY,
-        ffnExpanded,
-        ffnExpandedH,
-        ffnExpandedStartY,
-        norm2Y,
-        merge2Y,
-        'experts',
-      );
-    } else {
-      // === FFN (expandable) for dense models ===
-      const ffnSub = arch.ffnDim
-        ? `intermediate = ${arch.ffnDim.toLocaleString()}`
-        : `${ffnVariant} activation`;
-      drawExpandableBlock(
-        innerX,
-        ffnY,
-        innerW,
-        blockH,
-        'ffn',
-        'Feed-Forward Network',
-        ffnSub,
-        ffnExpanded,
-        'ffn',
-      );
-
-      if (ffnExpanded) {
-        drawFlow(ffnFlow, ffnExpandedStartY, innerX, innerW, `${ffnVariant} Details`);
-      }
-
-      const ffnBottom = ffnExpanded ? ffnExpandedStartY + ffnExpandedH : ffnY + blockH;
-      drawArrow(ffnBottom, merge2Y - circleR);
-      drawResidualBypass(norm2Y, merge2Y);
-    }
-  } else {
-    // === MAIN TRANSFORMER (collapsed) ===
-    const collapsedLabel = isMoE ? 'MoE Transformer Block' : 'Dense Transformer Block';
-    const collapsedSub = isMoE
-      ? `${layerLabel} \u00B7 ${attnLabel} \u00B7 Top-${arch.activeExperts}/${arch.numExperts}`
-      : `${layerLabel} \u00B7 ${attnLabel}`;
-    drawCollapsedTransformerBlock(
-      pad.left,
-      txStart,
-      bw,
-      collapsedTxH,
-      collapsedLabel,
-      collapsedSub,
-      'transformer',
-    );
-  }
-
-  if (!hasAlternatingLayers) {
-    // Arrow from transformer container to final norm
-    drawArrow(txEnd, finalNormY);
-  }
-
-  // === FINAL RMSNORM ===
-  drawBlock(pad.left, finalNormY, bw, smallH, 'norm', 'RMSNorm');
-  drawArrow(finalNormY + smallH, outputY);
-
-  // === OUTPUT HEAD ===
-  const outSub = arch.vocabSize ? `vocab = ${arch.vocabSize.toLocaleString()}` : undefined;
-  drawBlock(pad.left, outputY, bw, blockH, 'output', 'Output Head (LM Head)', outSub);
-
-  // === SPECS BAR ===
-  const specItems: { label: string; value: string }[] = [
-    { label: 'Type', value: isMoE ? 'MoE' : 'Dense' },
-    {
-      label: 'Layers',
-      value: hasDenseLayers
-        ? `${denseLayerCount}D + ${moeLayerCount}M`
-        : arch.numLayers
-          ? `${arch.numLayers}`
-          : '\u2014',
-    },
-    {
-      label: 'Attention',
-      value: arch.attentionType === 'AlternatingSinkGQA' ? 'Sink/Full GQA' : arch.attentionType,
-    },
-    {
-      label: 'Context',
-      value: arch.contextWindow ? formatContextWindow(arch.contextWindow) : '\u2014',
-    },
-    ...(isMoE
-      ? [
-          {
-            label: 'Experts',
-            // Active per token = routed top-k + the always-on shared experts, so
-            // show "6+1/385" (not "6/385"): the shared experts are active too.
-            // Count them from the model (Kimi K3 has 2), never assume one.
-            value: `${arch.activeExperts}${sharedExperts > 0 ? `+${sharedExperts}` : ''}/${arch.numExperts}`,
-          },
-        ]
-      : []),
-  ];
-
-  const specColW = bw / specItems.length;
-
-  g.append('rect')
-    .attr('x', pad.left)
-    .attr('y', specsY)
-    .attr('width', bw)
-    .attr('height', specsH)
-    .attr('rx', 8)
-    .attr('fill', bgSubtle)
-    .attr('stroke', borderColor)
-    .attr('stroke-width', 1);
-
-  specItems.forEach((spec, i) => {
-    const sx = pad.left + i * specColW + specColW / 2;
-
-    g.append('text')
-      .attr('x', sx)
-      .attr('y', specsY + 18)
-      .attr('text-anchor', 'middle')
-      .attr('fill', mutedFg)
-      .attr('font-size', '10px')
-      .attr('font-weight', 500)
-      .attr('font-family', 'inherit')
-      .text(spec.label);
-
-    g.append('text')
-      .attr('x', sx)
-      .attr('y', specsY + 38)
-      .attr('text-anchor', 'middle')
-      .attr('fill', fg)
-      .attr('font-size', '14px')
-      .attr('font-weight', 700)
-      .attr('font-family', 'inherit')
-      .text(spec.value);
-
-    if (i < specItems.length - 1) {
-      g.append('line')
-        .attr('x1', pad.left + (i + 1) * specColW)
-        .attr('y1', specsY + 8)
-        .attr('x2', pad.left + (i + 1) * specColW)
-        .attr('y2', specsY + specsH - 8)
-        .attr('stroke', borderColor)
-        .attr('stroke-width', 1);
-    }
-  });
-}
-
-export default function ModelArchitectureDiagram({
-  model,
-  className = '',
-}: ModelArchitectureDiagramProps) {
-  const [isExpanded, setIsExpanded] = useState(false);
+function ArchitectureContent({ model, arch, isExpanded }: ArchitectureContentProps) {
   const [expandedBlocks, setExpandedBlocks] = useState<Set<string>>(new Set());
   const svgRef = useRef<SVGSVGElement>(null);
   const containerRef = useRef<HTMLDivElement>(null);
+  const lastRenderRef = useRef<DiagramRenderState | null>(null);
   const { resolvedTheme } = useTheme();
-  const arch = getModelArchitecture(model);
+  const releaseDate = getModelReleaseDate(model);
 
   const toggleBlock = useCallback(
     (blockId: string) => {
@@ -2283,43 +67,161 @@ export default function ModelArchitectureDiagram({
     [model],
   );
 
-  useEffect(() => {
-    if (isExpanded && svgRef.current && arch) {
-      renderDiagram(
-        svgRef.current,
-        arch,
-        resolvedTheme === 'dark' || resolvedTheme === 'minecraft',
-        expandedBlocks,
-        toggleBlock,
-      );
-    }
-  }, [isExpanded, arch, resolvedTheme, model, expandedBlocks, toggleBlock]);
+  const renderIfChanged = useCallback(() => {
+    const svg = svgRef.current;
+    const container = containerRef.current;
+    if (!isExpanded || !svg || !container) return;
+
+    // Match the renderer's effective width so ResizeObserver notifications that
+    // cannot change the SVG (including its initial callback) stay no-ops.
+    const width = Math.min(container.clientWidth || 600, 640);
+    const isDark = resolvedTheme === 'dark' || resolvedTheme === 'minecraft';
+    const previous = lastRenderRef.current;
+    if (
+      previous?.width === width &&
+      previous.arch === arch &&
+      previous.isDark === isDark &&
+      previous.expandedBlocks === expandedBlocks &&
+      previous.onBlockClick === toggleBlock
+    ) {
+      return;
+    }
+
+    renderDiagram(svg, arch, isDark, expandedBlocks, toggleBlock);
+    lastRenderRef.current = { width, arch, isDark, expandedBlocks, onBlockClick: toggleBlock };
+  }, [isExpanded, arch, resolvedTheme, expandedBlocks, toggleBlock]);
 
   useEffect(() => {
-    if (!isExpanded || !containerRef.current || !arch) return;
-
-    const observer = new ResizeObserver(() => {
-      if (svgRef.current && arch) {
-        renderDiagram(
-          svgRef.current,
-          arch,
-          resolvedTheme === 'dark' || resolvedTheme === 'minecraft',
-          expandedBlocks,
-          toggleBlock,
-        );
-      }
-    });
+    if (!isExpanded || !containerRef.current) {
+      lastRenderRef.current = null;
+      return;
+    }
 
+    renderIfChanged();
+    const observer = new ResizeObserver(renderIfChanged);
     observer.observe(containerRef.current);
     return () => observer.disconnect();
-  }, [isExpanded, arch, resolvedTheme, expandedBlocks, toggleBlock]);
+  }, [isExpanded, renderIfChanged]);
 
-  useEffect(() => {
-    setExpandedBlocks(new Set());
-  }, [model]);
+  return (
+    <div
+      id="architecture-content"
+      className={`overflow-hidden transition-all duration-200 ease-in-out ${
+        isExpanded ? 'max-h-[4000px] opacity-100' : 'max-h-0 opacity-0'
+      }`}
+    >
+      <div ref={containerRef} className="px-4 pb-4">
+        <svg ref={svgRef} className="w-full" data-testid="model-architecture-svg" />
+        {/* The drill-down only renders while its parent block is expanded, so gate
+              the caption on the parent too — collapsing the parent leaves the child
+              id in expandedBlocks (state is restored on re-expand), and the caption
+              must not outlive the drawing it explains. */}
+        {/* Mirrors `altAttnExpandable` in renderDiagram: the caption describes the
+              CSA/HCA local-vs-compressed drill-down, so hybrids that opt out of that
+              drill-down must not show it either. */}
+        {arch.attentionType === 'Hybrid' &&
+          arch.alternatingAttentionExpandable !== false &&
+          [0, 1].some(
+            (i) => expandedBlocks.has(`altBlock${i}`) && expandedBlocks.has(`altAttention${i}`),
+          ) && (
+            <p
+              className="mt-2 text-2xs leading-snug text-muted-foreground"
+              data-testid="hybrid-attention-note"
+            >
+              <span className="font-medium text-foreground">Local</span> and{' '}
+              <span className="font-medium text-foreground">Compressed</span> are two KV sources,
+              not two separate attentions: each query attends in a{' '}
+              <span className="font-medium text-foreground">single softmax</span> to the union of
+              sliding-window + selected compressed keys, with a learnable per-head attention sink.
+            </p>
+          )}
+        {(arch.hyperConnections ?? 0) > 1 &&
+          ['altBlock0', 'altBlock1', 'hashBlock', 'transformer', 'denseTransformer'].some((id) =>
+            expandedBlocks.has(id),
+          ) && (
+            <p className="mt-2 text-2xs leading-snug text-muted-foreground" data-testid="mhc-note">
+              <span className="font-medium text-foreground">
+                Hyper-Connections (mHC ×{arch.hyperConnections})
+              </span>{' '}
+              replace each residual with {arch.hyperConnections} parallel streams combined by
+              learned, Sinkhorn-normalized weights — read ({arch.hyperConnections}→1), output, and a{' '}
+              {arch.hyperConnections}×{arch.hyperConnections} stream mix — shown as the mHC ×
+              {arch.hyperConnections} nodes.
+            </p>
+          )}
+        {arch.features && arch.features.length > 0 && (
+          <div className="mt-3 pt-3 border-t border-border/50">
+            <div className="flex flex-wrap gap-1.5 items-center">
+              <span className="text-xs text-muted-foreground mr-1">Features:</span>
+              {arch.features.map((feature) => (
+                <Badge key={feature} variant="secondary" className="text-xs py-0">
+                  {feature}
+                </Badge>
+              ))}
+              {arch.sourceUrl && (
+                <Link
+                  href={arch.sourceUrl}
+                  target="_blank"
+                  rel="noopener noreferrer"
+                  className="text-xs text-muted-foreground hover:text-foreground flex items-center gap-0.5 ml-auto"
+                  onClick={() =>
+                    track('model_architecture_source_clicked', { url: arch.sourceUrl! })
+                  }
+                >
+                  Source <ExternalLink className="size-3" />
+                </Link>
+              )}
+            </div>
+          </div>
+        )}
+        {arch.developer && releaseDate && (
+          <p className="text-xs text-muted-foreground mt-2">
+            Released by {arch.developer} on{' '}
+            {new Date(releaseDate).toLocaleDateString('en-US', {
+              year: 'numeric',
+              month: 'short',
+              day: 'numeric',
+            })}
+          </p>
+        )}
+      </div>
+    </div>
+  );
+}
+
+export default function ModelArchitectureDiagram({
+  model,
+  className = '',
+  variant = 'drawer',
+}: ModelArchitectureDiagramProps) {
+  const [isExpanded, setIsExpanded] = useState(false);
+  const arch = getModelArchitecture(model);
 
   if (!arch) return null;
 
+  if (variant === 'inline') {
+    return (
+      <div
+        className={`rounded-lg border border-border/50 bg-muted/30 overflow-hidden ${className}`}
+        data-testid="model-architecture-inline"
+      >
+        <div className="px-4 py-2 flex items-center gap-2">
+          <span className="text-sm font-medium">Model Architecture</span>
+          <Badge variant="outline" className="text-xs py-0">
+            {arch.architectureType === 'moe' ? 'MoE' : 'Dense'}
+          </Badge>
+          <Badge variant="outline" className="text-xs py-0">
+            {arch.attentionType === 'AlternatingSinkGQA' ? 'Sink/Full GQA' : arch.attentionType}
+          </Badge>
+          <Badge variant="outline" className="text-xs py-0">
+            {formatParamCount(arch.totalParams)}
+          </Badge>
+        </div>
+        <ArchitectureContent key={model} model={model} arch={arch} isExpanded />
+      </div>
+    );
+  }
+
   const handleToggle = () => {
     const newState = !isExpanded;
     setIsExpanded(newState);
@@ -2367,92 +269,7 @@ export default function ModelArchitectureDiagram({
           <ChevronDown className="size-4 shrink-0 text-muted-foreground" />
         )}
       </button>
-
-      <div
-        id="architecture-content"
-        className={`overflow-hidden transition-all duration-200 ease-in-out ${
-          isExpanded ? 'max-h-[4000px] opacity-100' : 'max-h-0 opacity-0'
-        }`}
-      >
-        <div ref={containerRef} className="px-4 pb-4">
-          <svg ref={svgRef} className="w-full" data-testid="model-architecture-svg" />
-          {/* The drill-down only renders while its parent block is expanded, so gate
-              the caption on the parent too — collapsing the parent leaves the child
-              id in expandedBlocks (state is restored on re-expand), and the caption
-              must not outlive the drawing it explains. */}
-          {/* Mirrors `altAttnExpandable` in renderDiagram: the caption describes the
-              CSA/HCA local-vs-compressed drill-down, so hybrids that opt out of that
-              drill-down must not show it either. */}
-          {arch.attentionType === 'Hybrid' &&
-            arch.alternatingAttentionExpandable !== false &&
-            [0, 1].some(
-              (i) => expandedBlocks.has(`altBlock${i}`) && expandedBlocks.has(`altAttention${i}`),
-            ) && (
-              <p
-                className="mt-2 text-[11px] leading-snug text-muted-foreground"
-                data-testid="hybrid-attention-note"
-              >
-                <span className="font-medium text-foreground">Local</span> and{' '}
-                <span className="font-medium text-foreground">Compressed</span> are two KV sources,
-                not two separate attentions: each query attends in a{' '}
-                <span className="font-medium text-foreground">single softmax</span> to the union of
-                sliding-window + selected compressed keys, with a learnable per-head attention sink.
-              </p>
-            )}
-          {(arch.hyperConnections ?? 0) > 1 &&
-            ['altBlock0', 'altBlock1', 'hashBlock', 'transformer', 'denseTransformer'].some((id) =>
-              expandedBlocks.has(id),
-            ) && (
-              <p
-                className="mt-2 text-[11px] leading-snug text-muted-foreground"
-                data-testid="mhc-note"
-              >
-                <span className="font-medium text-foreground">
-                  Hyper-Connections (mHC ×{arch.hyperConnections})
-                </span>{' '}
-                replace each residual with {arch.hyperConnections} parallel streams combined by
-                learned, Sinkhorn-normalized weights — read ({arch.hyperConnections}→1), output, and
-                a {arch.hyperConnections}×{arch.hyperConnections} stream mix — shown as the mHC ×
-                {arch.hyperConnections} nodes.
-              </p>
-            )}
-          {arch.features && arch.features.length > 0 && (
-            <div className="mt-3 pt-3 border-t border-border/50">
-              <div className="flex flex-wrap gap-1.5 items-center">
-                <span className="text-xs text-muted-foreground mr-1">Features:</span>
-                {arch.features.map((feature) => (
-                  <Badge key={feature} variant="secondary" className="text-xs py-0">
-                    {feature}
-                  </Badge>
-                ))}
-                {arch.sourceUrl && (
-                  <Link
-                    href={arch.sourceUrl}
-                    target="_blank"
-                    rel="noopener noreferrer"
-                    className="text-xs text-muted-foreground hover:text-foreground flex items-center gap-0.5 ml-auto"
-                    onClick={() =>
-                      track('model_architecture_source_clicked', { url: arch.sourceUrl! })
-                    }
-                  >
-                    Source <ExternalLink className="size-3" />
-                  </Link>
-                )}
-              </div>
-            </div>
-          )}
-          {arch.developer && arch.releaseDate && (
-            <p className="text-xs text-muted-foreground mt-2">
-              Released by {arch.developer} on{' '}
-              {new Date(arch.releaseDate).toLocaleDateString('en-US', {
-                year: 'numeric',
-                month: 'short',
-                day: 'numeric',
-              })}
-            </p>
-          )}
-        </div>
-      </div>
+      <ArchitectureContent key={model} model={model} arch={arch} isExpanded={isExpanded} />
     </div>
   );
 }
diff --git a/packages/app/src/components/inference/ui/ModelArchitectureInfoLink.tsx b/packages/app/src/components/inference/ui/ModelArchitectureInfoLink.tsx
new file mode 100644
index 000000000..91266b615
--- /dev/null
+++ b/packages/app/src/components/inference/ui/ModelArchitectureInfoLink.tsx
@@ -0,0 +1,81 @@
+'use client';
+
+import Link from 'next/link';
+
+import { track } from '@/lib/analytics';
+import { Badge } from '@/components/ui/badge';
+import { TooltipContent, TooltipRoot, TooltipTrigger } from '@/components/ui/tooltip';
+import { getModelSlugEntryForDisplayName } from '@/lib/compare-slug';
+import { formatParamCount, getModelArchitecture } from '@/lib/model-architectures';
+import { type Model, getModelLabel } from '@/lib/data-mappings';
+
+/**
+ * Compact replacement for the old full-width "Learn more about the … architecture"
+ * banner row: an icon link that sits beside the closed model-selector trigger and
+ * deep-links to the model's `/model/[slug]` page. The former banner copy and the
+ * architecture badges (MoE/Dense, attention type, param count) move into the
+ * tooltip so the dashboard no longer spends a whole row of vertical space on them.
+ * Renders nothing for models without a public slug (hidden models).
+ */
+export function ModelArchitectureInfoLink({
+  model,
+  locale,
+}: {
+  model: Model;
+  locale: 'en' | 'zh';
+}) {
+  const entry = getModelSlugEntryForDisplayName(model);
+  if (!entry) return null;
+  const arch = getModelArchitecture(model);
+  const label = getModelLabel(model);
+  const text =
+    locale === 'zh' ? `了解 ${label} 模型架构` : `Learn more about the ${label} architecture`;
+  return (
+    <TooltipRoot>
+      <TooltipTrigger asChild>
+        <Link
+          href={`/model/${entry.slug}`}
+          aria-label={text}
+          data-testid="model-architecture-link"
+          className="border-input flex size-9 shrink-0 items-center justify-center rounded-md border bg-transparent text-muted-foreground shadow-xs transition-colors hover:bg-muted/50 hover:text-foreground"
+          onClick={() => track('model_architecture_link_clicked', { model, slug: entry.slug })}
+        >
+          <svg
+            className="size-4"
+            viewBox="0 0 24 24"
+            fill="none"
+            stroke="currentColor"
+            strokeWidth="2"
+          >
+            <rect x="3" y="3" width="18" height="18" rx="2" />
+            <line x1="3" y1="9" x2="21" y2="9" />
+            <line x1="9" y1="9" x2="9" y2="21" />
+          </svg>
+        </Link>
+      </TooltipTrigger>
+      <TooltipContent
+        side="top"
+        collisionPadding={10}
+        className="z-[130]"
+        data-testid="model-architecture-tooltip"
+      >
+        <span className="flex flex-wrap items-center gap-1.5">
+          {text}
+          {arch && (
+            <span className="flex items-center gap-1.5">
+              <Badge variant="outline" className="text-xs py-0">
+                {arch.architectureType === 'moe' ? 'MoE' : 'Dense'}
+              </Badge>
+              <Badge variant="outline" className="text-xs py-0">
+                {arch.attentionType === 'AlternatingSinkGQA' ? 'Sink/Full GQA' : arch.attentionType}
+              </Badge>
+              <Badge variant="outline" className="text-xs py-0">
+                {formatParamCount(arch.totalParams)}
+              </Badge>
+            </span>
+          )}
+        </span>
+      </TooltipContent>
+    </TooltipRoot>
+  );
+}
diff --git a/packages/app/src/components/inference/ui/NormalizedInteractivityHelpLink.tsx b/packages/app/src/components/inference/ui/NormalizedInteractivityHelpLink.tsx
new file mode 100644
index 000000000..d6eedbc56
--- /dev/null
+++ b/packages/app/src/components/inference/ui/NormalizedInteractivityHelpLink.tsx
@@ -0,0 +1,29 @@
+'use client';
+
+import { CircleHelp } from 'lucide-react';
+
+import { track } from '@/lib/analytics';
+import type { Locale } from '@/lib/i18n';
+
+const STRINGS = {
+  en: 'What does E2E Normalized Interactivity mean?',
+  zh: '什么是端到端归一化交互性？',
+} as const;
+
+export function NormalizedInteractivityHelpLink({ locale }: { locale: Locale }) {
+  const label = STRINGS[locale];
+  const href = `${locale === 'zh' ? '/zh' : ''}/about#faq-normalized-interactivity`;
+
+  return (
+    <a
+      href={href}
+      data-testid="normalized-interactivity-faq-link"
+      aria-label={label}
+      title={label}
+      onClick={() => track('interactivity_normalized_faq_opened', { locale })}
+      className="no-export absolute right-1 top-1/2 z-10 inline-flex size-6 -translate-y-1/2 items-center justify-center rounded-full text-muted-foreground/55 transition-colors hover:bg-muted hover:text-brand focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-brand focus-visible:ring-offset-1"
+    >
+      <CircleHelp aria-hidden="true" className="size-3.5" strokeWidth={1.75} />
+    </a>
+  );
+}
diff --git a/packages/app/src/components/inference/ui/OffloadHaloLegendKey.tsx b/packages/app/src/components/inference/ui/OffloadHaloLegendKey.tsx
index c7e38e10b..29872fc37 100644
--- a/packages/app/src/components/inference/ui/OffloadHaloLegendKey.tsx
+++ b/packages/app/src/components/inference/ui/OffloadHaloLegendKey.tsx
@@ -5,15 +5,15 @@ export const OFFLOAD_HALO_STROKE_WIDTH = 1.5;
 export const OFFLOAD_HALO_DASHARRAY = '3 2';
 
 /**
- * Legend key for the dashed ring drawn around agentic points that use KV-cache
- * offload. This stays outside `no-export` content so the chart export clone
- * carries the same explanation into downloaded PNGs.
+ * Key for the dashed ring drawn around agentic points that use KV-cache
+ * offload. Rendered in the axis-metric info footer below the chart (the
+ * footer is `no-export`, so downloaded PNGs carry only the halo itself).
  */
 export function OffloadHaloLegendKey() {
   return (
     <div
       data-testid="offload-halo-key"
-      className="mt-2 flex w-full items-center gap-2 px-1 pr-2 text-xs text-muted-foreground"
+      className="flex w-full items-center gap-2 px-1 pr-2 text-xs text-muted-foreground"
     >
       <svg
         width="16"
diff --git a/packages/app/src/components/inference/ui/QuickFiltersDialog.tsx b/packages/app/src/components/inference/ui/QuickFiltersDialog.tsx
new file mode 100644
index 000000000..86f24fad5
--- /dev/null
+++ b/packages/app/src/components/inference/ui/QuickFiltersDialog.tsx
@@ -0,0 +1,290 @@
+'use client';
+
+import { ListFilter } from 'lucide-react';
+import { Button } from '@/components/ui/button';
+import { Switch } from '@/components/ui/switch';
+import type { DeploymentMode, SpecMode } from '@/components/inference/types';
+import { FRAMEWORK_FAMILIES } from '@/components/inference/utils/quickFilters';
+
+import {
+  useInferenceActions,
+  useInferenceData,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
+import {
+  Dialog,
+  DialogClose,
+  DialogContent,
+  DialogDescription,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from '@/components/ui/dialog';
+import { track } from '@/lib/analytics';
+import { Sequence } from '@/lib/data-mappings';
+import { useLocale } from '@/lib/use-locale';
+import { cn } from '@/lib/utils';
+
+const STRINGS = {
+  en: {
+    title: 'Quick Filters',
+    description:
+      'Narrow the chart by chip vendor, serving framework, deployment mode, and speculative decoding. Selecting none in a group shows all.',
+    agenticDescription:
+      'Narrow the chart by chip vendor, serving framework, and deployment mode. Selecting none in a group shows all.',
+    selected: 'selected',
+    bestPerSku: 'Best per SKU',
+    bestPerSkuHint: 'Show only the best configuration for each chip',
+    vendor: 'Vendor',
+    framework: 'Framework',
+    deployment: 'Deployment',
+    singleNode: 'Single-node',
+    multiNode: 'Multi-node',
+    disaggregated: 'Disaggregated',
+    specDecoding: 'Spec Decoding',
+    noData: 'No data for the current selection',
+    clear: 'Clear filters',
+    done: 'Done',
+  },
+  zh: {
+    title: '快捷筛选',
+    description: '按芯片厂商、推理框架、部署模式和投机解码筛选图表。某组不选则显示全部。',
+    agenticDescription: '按芯片厂商、推理框架和部署模式筛选图表。某组不选则显示全部。',
+    selected: '项已选',
+    bestPerSku: '每个 SKU 仅显示最佳配置',
+    bestPerSkuHint: '每款芯片只显示表现最佳的配置',
+    vendor: '厂商',
+    framework: '框架',
+    deployment: '部署模式',
+    singleNode: '单节点',
+    multiNode: '多节点聚合',
+    disaggregated: '分离式',
+    specDecoding: '投机解码',
+    noData: '当前选择无可用数据',
+    clear: '清除筛选',
+    done: '完成',
+  },
+} as const;
+
+const VENDORS = [
+  { value: 'NVIDIA', label: 'NVIDIA' },
+  { value: 'AMD', label: 'AMD' },
+] as const;
+const DEPLOYMENT_MODES: DeploymentMode[] = ['single-node', 'multi-node', 'disagg'];
+const SPEC_MODES: { value: SpecMode; label: string }[] = [
+  { value: 'mtp', label: 'MTP' },
+  { value: 'stp', label: 'STP' },
+];
+
+function toggleValue<T extends string>(current: T[], value: T): T[] {
+  return current.includes(value) ? current.filter((item) => item !== value) : [...current, value];
+}
+
+export function QuickFiltersDialog({
+  open,
+  onOpenChange,
+  bestPerSku,
+}: {
+  open: boolean;
+  onOpenChange: (open: boolean) => void;
+  /** Optional Best per SKU control (scatter chart only) — the chart owns the
+   * toggle logic because overlay mode manages a temporary unified selection. */
+  bestPerSku?: { checked: boolean; onCheckedChange: (checked: boolean) => void };
+}) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const {
+    setQuickFilterVendors,
+    setQuickFilterFrameworks,
+    setQuickFilterDeployment,
+    setQuickFilterSpec,
+  } = useInferenceActions();
+  const { selectedSequence, quickFilters } = useInferenceFilters();
+  const { availableQuickFilters } = useInferenceData();
+  const isAgentic = selectedSequence === Sequence.AgenticTraces;
+
+  const frameworkOptions = FRAMEWORK_FAMILIES.filter(
+    (framework) =>
+      availableQuickFilters.frameworks.includes(framework.key) ||
+      quickFilters.frameworks.includes(framework.key),
+  ).map((framework) => ({
+    value: framework.key,
+    label: framework.label,
+    available: availableQuickFilters.frameworks.includes(framework.key),
+  }));
+
+  const groups: {
+    key: 'vendor' | 'framework' | 'deployment' | 'spec';
+    label: string;
+    options: readonly { value: string; label: string; available: boolean }[];
+    selected: readonly string[];
+  }[] = [
+    {
+      key: 'vendor',
+      label: t.vendor,
+      options: VENDORS.map((option) => ({
+        ...option,
+        available: availableQuickFilters.vendors.includes(option.value),
+      })),
+      selected: quickFilters.vendors,
+    },
+    ...(frameworkOptions.length > 0
+      ? [
+          {
+            key: 'framework' as const,
+            label: t.framework,
+            options: frameworkOptions,
+            selected: quickFilters.frameworks,
+          },
+        ]
+      : []),
+    {
+      key: 'deployment',
+      label: t.deployment,
+      options: DEPLOYMENT_MODES.map((value) => ({
+        value,
+        label:
+          value === 'single-node'
+            ? t.singleNode
+            : value === 'multi-node'
+              ? t.multiNode
+              : t.disaggregated,
+        available: availableQuickFilters.deployment.includes(value),
+      })),
+      selected: quickFilters.deployment,
+    },
+    ...(isAgentic
+      ? []
+      : [
+          {
+            key: 'spec' as const,
+            label: t.specDecoding,
+            options: SPEC_MODES.map((option) => ({
+              ...option,
+              available: availableQuickFilters.spec.includes(option.value),
+            })),
+            selected: quickFilters.spec,
+          },
+        ]),
+  ];
+
+  const selectedCount = groups.reduce((count, group) => count + group.selected.length, 0);
+
+  const handleToggle = (
+    category: 'vendor' | 'framework' | 'deployment' | 'spec',
+    value: string,
+  ) => {
+    const wasActive =
+      category === 'vendor'
+        ? quickFilters.vendors.includes(value)
+        : category === 'framework'
+          ? quickFilters.frameworks.includes(value)
+          : category === 'deployment'
+            ? quickFilters.deployment.includes(value as DeploymentMode)
+            : quickFilters.spec.includes(value as SpecMode);
+    if (category === 'vendor') setQuickFilterVendors(toggleValue(quickFilters.vendors, value));
+    else if (category === 'framework')
+      setQuickFilterFrameworks(toggleValue(quickFilters.frameworks, value));
+    else if (category === 'deployment')
+      setQuickFilterDeployment(toggleValue(quickFilters.deployment, value as DeploymentMode));
+    else setQuickFilterSpec(toggleValue(quickFilters.spec, value as SpecMode));
+    track('inference_quick_filter_toggled', { category, value, active: !wasActive });
+  };
+
+  const clearFilters = () => {
+    setQuickFilterVendors([]);
+    setQuickFilterFrameworks([]);
+    setQuickFilterDeployment([]);
+    setQuickFilterSpec([]);
+    track('inference_quick_filters_cleared', { source: 'dialog' });
+  };
+
+  return (
+    <Dialog open={open} onOpenChange={onOpenChange}>
+      <DialogContent
+        className="max-h-[85vh] overflow-y-auto sm:max-w-xl"
+        data-testid="quick-filters-dialog"
+      >
+        <DialogHeader>
+          <DialogTitle className="flex items-center gap-2">
+            <ListFilter className="size-4 text-brand" />
+            {t.title}
+            {selectedCount > 0 && (
+              <span
+                className="rounded-full bg-brand/10 px-2 py-0.5 text-2xs font-medium text-brand"
+                data-testid="quick-filters-selected-count"
+              >
+                {selectedCount} {t.selected}
+              </span>
+            )}
+          </DialogTitle>
+          <DialogDescription>{isAgentic ? t.agenticDescription : t.description}</DialogDescription>
+        </DialogHeader>
+
+        <div className="divide-y divide-border rounded-md border">
+          {bestPerSku && (
+            <section className="flex items-center justify-between gap-3 px-3 py-3">
+              <div>
+                <h3 className="text-xs font-semibold text-muted-foreground">{t.bestPerSku}</h3>
+                <p className="mt-0.5 text-2xs text-muted-foreground/70">{t.bestPerSkuHint}</p>
+              </div>
+              <Switch
+                data-testid="quick-filter-best-per-sku"
+                checked={bestPerSku.checked}
+                onCheckedChange={bestPerSku.onCheckedChange}
+                aria-label={t.bestPerSku}
+              />
+            </section>
+          )}
+          {groups.map((group) => (
+            <section
+              key={group.key}
+              className="grid gap-2 px-3 py-3 sm:grid-cols-[7rem_1fr] sm:items-start"
+            >
+              <h3 className="pt-1 text-xs font-semibold text-muted-foreground">{group.label}</h3>
+              <div className="flex flex-wrap gap-1.5">
+                {group.options.map((option) => {
+                  const active = group.selected.includes(option.value);
+                  const disabled = !option.available && !active;
+                  return (
+                    <Button
+                      key={option.value}
+                      type="button"
+                      size="sm"
+                      variant={active ? 'default' : 'outline'}
+                      aria-pressed={active}
+                      disabled={disabled}
+                      title={disabled ? t.noData : undefined}
+                      className={cn(
+                        'h-7 rounded-full px-3 text-xs',
+                        active && 'bg-brand hover:bg-brand/90',
+                      )}
+                      data-testid={`quick-filter-${group.key}-${option.value}`}
+                      onClick={() => handleToggle(group.key, option.value)}
+                    >
+                      {option.label}
+                    </Button>
+                  );
+                })}
+              </div>
+            </section>
+          ))}
+        </div>
+
+        <DialogFooter className="flex-row justify-between sm:justify-between">
+          <Button
+            type="button"
+            variant="ghost"
+            disabled={selectedCount === 0}
+            onClick={clearFilters}
+          >
+            {t.clear}
+          </Button>
+          <DialogClose asChild>
+            <Button type="button">{t.done}</Button>
+          </DialogClose>
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
+  );
+}
diff --git a/packages/app/src/components/inference/ui/ScatterGraph.decoration.test.tsx b/packages/app/src/components/inference/ui/ScatterGraph.decoration.test.tsx
index 57320c91c..f70037913 100644
--- a/packages/app/src/components/inference/ui/ScatterGraph.decoration.test.tsx
+++ b/packages/app/src/components/inference/ui/ScatterGraph.decoration.test.tsx
@@ -1,79 +1,26 @@
 // @vitest-environment jsdom
-/**
- * Integration tests for the ScatterGraph toggle decoration path.
- *
- * Core behavior under test: legend hw toggles, precision toggles, and color
- * changes restyle the existing SVG (opacity / fill / shape) WITHOUT tearing
- * down and rebuilding the chart structure — the ~300ms main-thread long task
- * behind the failing field INP. A rebuild is detected via a spy on
- * setupChartStructure, which every full chart render must call.
- */
-import { act, createElement, useReducer } from 'react';
-import { createRoot } from 'react-dom/client';
-import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
-import * as d3 from 'd3';
-
-import { setupChartStructure } from '@/lib/d3-chart/chart-setup';
-import type { ChartDefinition, InferenceData } from '@/components/inference/types';
-
-import { computeToggle } from '@/hooks/useTogglableSet';
-// ── Module mocks ───────────────────────────────────────────────────────────
-vi.mock('@/lib/d3-chart/chart-setup', { spy: true });
-vi.mock('@/lib/analytics', () => ({ track: vi.fn() }));
-vi.mock('next-themes', () => ({ useTheme: () => ({ resolvedTheme: 'dark' }) }));
-// The legend is React-rendered (covered elsewhere) — keep the tree light.
-vi.mock('@/components/ui/chart-legend', () => ({
-  default: ({ keyIndicators }: { keyIndicators?: React.ReactNode }) => keyIndicators ?? null,
-}));
-
-const inferenceState = vi.hoisted(() => ({ current: {} as Record<string, unknown> }));
-vi.mock('@/components/inference/InferenceContext', () => ({
-  useInference: () => inferenceState.current,
-}));
-
-const overlayState = vi.hoisted(() => ({ current: {} as Record<string, unknown> }));
-vi.mock('@/components/unofficial-run-provider', () => ({
-  useUnofficialRun: () => overlayState.current,
-}));
-
-// ScatterGraph calls useTraceAvailability (a useQuery) for the agentic "View
-// charts" tooltip button. Stub it so these decoration tests don't need a
-// QueryClientProvider — trace presence is irrelevant to the toggle path.
-vi.mock('@/hooks/api/use-trace-availability', () => ({
-  useTraceAvailability: () => ({ data: undefined }),
-}));
-
-import ScatterGraph, { pointLabelText } from './ScatterGraph';
-
-// ── Environment stubs ────────────────────────────────────────────────────────
-class MockResizeObserver {
-  observe() {}
-  unobserve() {}
-  disconnect() {}
-}
-
-// ── Fixtures ─────────────────────────────────────────────────────────────────
-const point = (hwKey: string, precision: string, x: number, y: number, tp: number): InferenceData =>
-  ({ hwKey, precision, x, y, tp, conc: 16, framework: 'vllm' }) as unknown as InferenceData;
-
-// h100 owns both axis extremes so hiding b200 / showing fp4 keeps the niced
-// domains identical — exactly the toggle case that must not rebuild.
-const POINTS: InferenceData[] = [
-  point('h100', 'fp8', 1, 1, 1),
-  point('h100', 'fp8', 100, 1000, 8),
-  point('h100', 'fp4', 40, 400, 4),
-  point('b200', 'fp8', 50, 500, 4),
-  point('b200', 'fp8', 60, 600, 8),
-];
-
-const HARDWARE_CONFIG = {
-  h100: { name: 'H100', label: 'H100', gpu: 'H100' },
-  b200: { name: 'B200', label: 'B200', gpu: 'B200' },
-};
-
-const CHART_DEFINITION = { chartType: 'interactivity' } as unknown as ChartDefinition;
-
-const noop = () => {};
+import { act } from 'react';
+import { describe, expect, it, vi } from 'vitest';
+
+import type { InferenceData } from '@/components/inference/types';
+
+import {
+  POINTS,
+  baseInferenceState,
+  dotGroups,
+  inferenceState,
+  mountChart,
+  point,
+  rebuildCount,
+  traceAvailabilityState,
+  availabilityState,
+} from './ScatterGraph.test-harness';
+import { pointLabelText } from './point-label';
+
+const datedPoint = (date: string, x: number, y: number) => ({
+  ...point('h100', 'fp8', x, y, 8),
+  date,
+});
 
 describe('pointLabelText', () => {
   it('keeps decode mode out of mixed agentic point labels', () => {
@@ -83,9 +30,18 @@ describe('pointLabelText', () => {
     const mtp = { ...standard, spec_decoding: 'mtp' };
     const eagle = { ...standard, spec_decoding: 'eagle' };
 
-    expect(pointLabelText(standard, false)).toBe('8\nC=16');
-    expect(pointLabelText(mtp, false)).toBe('8\nC=16');
-    expect(pointLabelText(eagle, false)).toBe('8\nC=16');
+    expect(pointLabelText(standard, false, true)).toBe('8\nC=16');
+    expect(pointLabelText(mtp, false, true)).toBe('8\nC=16');
+    expect(pointLabelText(eagle, false, true)).toBe('8\nC=16');
+  });
+
+  it('omits concurrency unless the advanced toggle enables it', () => {
+    const standard = point('h100', 'fp8', 1, 1, 8);
+    standard.benchmark_type = 'agentic_traces';
+    standard.spec_decoding = 'none';
+
+    expect(pointLabelText(standard, false, false)).toBe('8');
+    expect(pointLabelText(standard, false, true)).toBe('8\nC=16');
   });
 
   it('keeps fixed-sequence labels unchanged', () => {
@@ -93,137 +49,8 @@ describe('pointLabelText', () => {
     fixed.benchmark_type = 'single_turn';
     fixed.spec_decoding = 'mtp';
 
-    expect(pointLabelText(fixed, false)).toBe('8\nC=16');
-  });
-});
-
-function baseInferenceState() {
-  return {
-    activeHwTypes: new Set(['h100', 'b200']),
-    hardwareConfig: HARDWARE_CONFIG,
-    toggleHwType: noop,
-    removeHwType: noop,
-    hwTypesWithData: new Set(['h100', 'b200']),
-    resolveComparisonSelection: (proposed: Set<string>) => ({
-      result: proposed,
-      keptGroup: null,
-      droppedGroups: [],
-    }),
-    toggleComparisonSelection: (prev: Set<string>, item: string, allItems: Set<string>) =>
-      computeToggle(prev, item, allItems),
-    selectedPrecisions: ['fp8'],
-    selectedYAxisMetric: 'y',
-    quickFilters: { vendors: [], frameworks: [], deployment: [], spec: [] },
-    availableQuickFilters: { vendors: [], frameworks: [], deployment: [], spec: [] },
-    availableRuns: null,
-    selectedRunId: '',
-    hideNonOptimal: false,
-    setHideNonOptimal: noop,
-    hidePointLabels: false,
-    setHidePointLabels: noop,
-    selectAllHwTypes: noop,
-    highContrast: false,
-    setHighContrast: noop,
-    logScale: false,
-    setLogScale: noop,
-    scaleType: 'auto',
-    isLegendExpanded: false,
-    setIsLegendExpanded: noop,
-    useAdvancedLabels: false,
-    setUseAdvancedLabels: noop,
-    showGradientLabels: false,
-    setShowGradientLabels: noop,
-    showLineLabels: false,
-    setShowLineLabels: noop,
-    showSpeedOverlay: false,
-    setShowSpeedOverlay: noop,
-    showMinecraftOverlay: false,
-    setShowMinecraftOverlay: noop,
-  };
-}
-
-function baseOverlayState() {
-  return {
-    isUnofficialRun: false,
-    activeOverlayHwTypes: new Set<string>(),
-    setActiveOverlayHwTypes: noop,
-    allOverlayHwTypes: new Set<string>(),
-    toggleOverlayHwType: noop,
-    resetOverlayHwTypes: noop,
-    localOfficialOverride: null,
-    setLocalOfficialOverride: noop,
-    runIndexByUrl: {},
-    unofficialRunInfos: [],
-  };
-}
-
-// ── Harness ──────────────────────────────────────────────────────────────────
-function mountChart(props?: Partial<Parameters<typeof ScatterGraph>[0]>) {
-  let forceUpdate: () => void = noop;
-  function Harness() {
-    // ScatterGraph and D3Chart are React.memo'd; mocked context hooks bypass
-    // React's context subscription, so re-renders are driven through a
-    // version-bumped caption prop.
-    const [version, bump] = useReducer((v: number) => v + 1, 0);
-    forceUpdate = bump;
-    return createElement(ScatterGraph, {
-      chartId: 'chart-test',
-      modelLabel: 'DeepSeek-R1-0528',
-      data: POINTS,
-      xLabel: 'Interactivity (tok/s/user)',
-      yLabel: 'Output Throughput per GPU',
-      chartDefinition: CHART_DEFINITION,
-      transitionDuration: 0,
-      caption: `v${version}`,
-      ...props,
-    });
-  }
-
-  const container = document.createElement('div');
-  document.body.append(container);
-  const root = createRoot(container);
-  act(() => {
-    root.render(createElement(Harness));
+    expect(pointLabelText(fixed, false, true)).toBe('8\nC=16');
   });
-  return {
-    container,
-    rerender: () => act(() => forceUpdate()),
-    unmount: () => {
-      act(() => root.unmount());
-      container.remove();
-    },
-  };
-}
-
-const dotGroups = (container: HTMLElement, hwKey?: string) =>
-  [...container.querySelectorAll<SVGGElement>('.dot-group')].filter(
-    (n) => !hwKey || n.dataset.hwKey === hwKey,
-  );
-
-const rebuildCount = () => vi.mocked(setupChartStructure).mock.calls.length;
-
-beforeEach(() => {
-  vi.stubGlobal('ResizeObserver', MockResizeObserver);
-  // Charts measure their container; jsdom reports 0 — give them real space.
-  vi.spyOn(HTMLElement.prototype, 'getBoundingClientRect').mockReturnValue({
-    width: 800,
-    height: 600,
-    top: 0,
-    left: 0,
-    right: 800,
-    bottom: 600,
-    x: 0,
-    y: 0,
-    toJSON: () => ({}),
-  } as DOMRect);
-  inferenceState.current = baseInferenceState();
-  overlayState.current = baseOverlayState();
-  vi.mocked(setupChartStructure).mockClear();
-});
-
-afterEach(() => {
-  vi.unstubAllGlobals();
-  vi.restoreAllMocks();
 });
 
 describe('ScatterGraph toggle decoration', () => {
@@ -236,6 +63,24 @@ describe('ScatterGraph toggle decoration', () => {
     unmount();
   });
 
+  it('checks log availability for fixed-sequence and agentic official points', () => {
+    const fixed = {
+      ...point('h100', 'fp8', 10, 100, 1),
+      id: 96255,
+      benchmark_type: 'single_turn',
+    } as InferenceData;
+    const agentic = {
+      ...point('h100', 'fp8', 20, 200, 2),
+      id: 206885,
+      benchmark_type: 'agentic_traces',
+    } as InferenceData;
+    const { unmount } = mountChart({ data: [fixed, agentic] });
+
+    expect(availabilityState.logIds).toEqual([96255, 206885]);
+    expect(availabilityState.traceIds).toEqual([206885]);
+    unmount();
+  });
+
   it('keeps speculative decoding out of point decorations and shows only KV offload', () => {
     const standard = {
       ...point('h100', 'fp8', 1, 1, 1),
@@ -275,8 +120,128 @@ describe('ScatterGraph toggle decoration', () => {
     expect(groups[1].querySelector('.offload-halo')).toBeNull();
     expect(groups[2].querySelector('.offload-halo')).not.toBeNull();
     expect(container.querySelector('[data-testid="spec-decode-marker-key"]')).toBeNull();
-    expect(container.querySelector('[data-testid="offload-halo-key"]')).not.toBeNull();
-    expect(container.querySelector('[data-testid="agentic-optimization-note"]')).not.toBeNull();
+    // The KV-offload key and optimization note moved to the axis-metric info
+    // footer rendered by ChartDisplay, so the chart itself carries neither.
+    expect(container.querySelector('[data-testid="offload-halo-key"]')).toBeNull();
+    expect(container.querySelector('[data-testid="agentic-optimization-note"]')).toBeNull();
+    unmount();
+  });
+
+  it('reads current trace availability without changing metric identity', () => {
+    const agenticPoint = {
+      ...point('h100', 'fp8', 20, 200, 2),
+      id: 42,
+      benchmark_type: 'agentic_traces',
+    } as InferenceData;
+    inferenceState.current = {
+      ...baseInferenceState(),
+      selectedSequence: 'agentic-traces',
+    };
+    const { container, rerender, unmount } = mountChart({ data: [agenticPoint] });
+    const buildsAfterMount = rebuildCount();
+    const originalDot = container.querySelector<SVGGElement>('.dot-group')!;
+    const dot = container.querySelector<SVGGElement>('.dot-group')!;
+
+    act(() =>
+      dot.dispatchEvent(new MouseEvent('click', { bubbles: true, clientX: 20, clientY: 20 })),
+    );
+    expect(document.querySelector('[data-action="view-charts"]')).toBeNull();
+
+    traceAvailabilityState.current = { 42: true };
+    rerender();
+    expect(rebuildCount()).toBe(buildsAfterMount);
+    expect(container.querySelector<SVGGElement>('.dot-group')).toBe(originalDot);
+    act(() =>
+      dot.dispatchEvent(new MouseEvent('click', { bubbles: true, clientX: 20, clientY: 20 })),
+    );
+
+    expect(document.querySelector('[data-action="view-charts"]')?.getAttribute('href')).toBe(
+      '/inference/agentic/42',
+    );
+    unmount();
+  });
+
+  it('does not measure point-label collisions while point labels are hidden', () => {
+    inferenceState.current = {
+      ...baseInferenceState(),
+      showPointLabels: false,
+    };
+    let pointLabelMeasurements = 0;
+    vi.spyOn(
+      SVGElement.prototype as unknown as { getBBox: () => DOMRect },
+      'getBBox',
+    ).mockImplementation(function (this: SVGElement) {
+      if (this.classList.contains('point-label')) pointLabelMeasurements++;
+      return {
+        x: 0,
+        y: 0,
+        width: 48,
+        height: 12,
+        top: 0,
+        right: 48,
+        bottom: 12,
+        left: 0,
+        toJSON: () => ({}),
+      } as DOMRect;
+    });
+
+    const { unmount } = mountChart();
+
+    expect(pointLabelMeasurements).toBe(0);
+    unmount();
+  });
+
+  it('keeps point-label visibility synchronized while the chart is empty', () => {
+    const chartProps = { data: [] as InferenceData[] };
+    inferenceState.current = {
+      ...baseInferenceState(),
+      showPointLabels: false,
+    };
+    const { container, rerender, unmount } = mountChart(chartProps);
+
+    inferenceState.current = {
+      ...inferenceState.current,
+      showPointLabels: true,
+    };
+    rerender();
+    chartProps.data = POINTS;
+    rerender();
+    const label = container.querySelector<SVGTextElement>('.point-label');
+    expect(label?.style.display).toBe('');
+
+    inferenceState.current = {
+      ...inferenceState.current,
+      showPointLabels: false,
+    };
+    rerender();
+    expect(label?.style.display).toBe('none');
+    unmount();
+  });
+
+  it('rechecks point-label collisions when line-label obstacles change', () => {
+    inferenceState.current = {
+      ...baseInferenceState(),
+      showPointLabels: true,
+      showLineLabels: false,
+    };
+    const getBBox = vi.spyOn(
+      SVGElement.prototype as unknown as { getBBox: () => DOMRect },
+      'getBBox',
+    );
+    const { rerender, unmount } = mountChart();
+    getBBox.mockClear();
+
+    inferenceState.current = {
+      ...inferenceState.current,
+      showLineLabels: true,
+    };
+    rerender();
+
+    expect(
+      getBBox.mock.contexts.some(
+        (element) => element instanceof SVGElement && element.classList.contains('point-label'),
+      ),
+    ).toBe(true);
     unmount();
   });
 
@@ -353,165 +318,98 @@ describe('ScatterGraph toggle decoration', () => {
     unmount();
   });
 
-  it('rebuilds when the scale domain actually changes (data refresh path intact)', () => {
-    const { rerender, unmount } = mountChart();
+  it('updates scale-dependent geometry without rebuilding chart structure', () => {
+    const { container, rerender, unmount } = mountChart();
     const buildsAfterMount = rebuildCount();
+    const b200Dot = dotGroups(container, 'b200')[0];
+    const transformBefore = b200Dot.getAttribute('transform');
 
-    // Hiding the hw that owns the axis extremes changes the visible domain —
-    // axes must rescale, which is a legitimate full render.
     inferenceState.current = {
       ...inferenceState.current,
       activeHwTypes: new Set(['b200']),
     };
     rerender();
 
-    expect(rebuildCount()).toBeGreaterThan(buildsAfterMount);
-    unmount();
-  });
-
-  it('animates rooflines together with dots on a domain-changing toggle', () => {
-    // Real transition duration: the rebuild restores each surviving element to
-    // its old position/path and schedules a "data-update" transition that only
-    // starts on the next timer tick. Regression: the decoration effect used to
-    // re-apply final roofline `d` attrs in the same commit, so the curve
-    // teleported to its destination while the dots animated.
-    const { container, rerender, unmount } = mountChart({ transitionDuration: 750 });
-
-    const b200Roofline = () =>
-      container.querySelector<SVGPathElement>('.roofline-path[data-hw-key="b200"]')!;
-    const b200Dot = () => dotGroups(container, 'b200')[0];
-    const dBefore = b200Roofline().getAttribute('d');
-    const dotTransformBefore = b200Dot().getAttribute('transform');
-    expect(dBefore).toBeTruthy();
-
-    // Hide the extreme-owning hw → domains shrink → full render + animation.
-    inferenceState.current = {
-      ...inferenceState.current,
-      activeHwTypes: new Set(['b200']),
-    };
-    rerender();
-
-    // At commit end the transitions are scheduled but have not ticked: both
-    // the roofline path and the dots must still sit at their OLD coordinates,
-    // each with a pending transition toward the new ones.
-    expect(b200Roofline().getAttribute('d')).toBe(dBefore);
-    expect(b200Dot().getAttribute('transform')).toBe(dotTransformBefore);
-    expect((b200Roofline() as unknown as { __transition?: object }).__transition).toBeTruthy();
-    expect((b200Dot() as unknown as { __transition?: object }).__transition).toBeTruthy();
-
-    // jsdom can't run the SVG transform interpolator (no transform.baseVal),
-    // so cancel the scheduled transitions before teardown.
-    d3.select(container).selectAll('.dot-group, .roofline-path').interrupt('data-update');
+    expect(rebuildCount()).toBe(buildsAfterMount);
+    expect(dotGroups(container, 'b200')[0]).toBe(b200Dot);
+    expect(b200Dot.getAttribute('transform')).not.toBe(transformBefore);
     unmount();
   });
 
-  it('keeps unofficial-run overlay markers rendered through official toggles', () => {
-    const overlayPoints = [point('h100', 'fp8', 30, 300, 2), point('h100', 'fp8', 35, 350, 4)].map(
-      (p) => ({ ...p, run_url: 'https://github.com/o/r/actions/runs/123' }),
-    );
-    overlayState.current = {
-      ...baseOverlayState(),
-      isUnofficialRun: true,
-      activeOverlayHwTypes: new Set(['h100']),
-      allOverlayHwTypes: new Set(['h100']),
-      runIndexByUrl: { 'https://github.com/o/r/actions/runs/123': 0 },
-      unofficialRunInfos: [
-        { id: '123', branch: 'test-branch', url: 'https://github.com/o/r/actions/runs/123' },
-      ],
+  it('keeps identical configs from different dates distinct during metric updates', () => {
+    const chartProps = {
+      data: [datedPoint('2026-05-15', 20, 200), datedPoint('2026-06-15', 80, 800)],
+      xExtentOverride: [0, 100] as [number, number],
+      yExtentOverride: [0, 1000] as [number, number],
+      transitionDuration: 0,
     };
-    const { container, rerender, unmount } = mountChart({
-      overlayData: {
-        data: overlayPoints,
-        hardwareConfig: HARDWARE_CONFIG,
-      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
-    });
+    const { container, rerender, unmount } = mountChart(chartProps);
+    const groups = dotGroups(container);
     const buildsAfterMount = rebuildCount();
 
-    expect(container.querySelectorAll('.unofficial-overlay-pt')).toHaveLength(2);
-    expect(container.querySelectorAll('.overlay-roofline-path').length).toBeGreaterThan(0);
+    expect(groups).toHaveLength(2);
+    expect(groups[0].getAttribute('transform')).not.toBe(groups[1].getAttribute('transform'));
 
-    // Toggling an official hw must not rebuild or disturb overlay markers.
-    inferenceState.current = {
-      ...inferenceState.current,
-      activeHwTypes: new Set(['h100']),
-    };
+    chartProps.data = [datedPoint('2026-05-15', 30, 300), datedPoint('2026-06-15', 70, 700)];
     rerender();
 
-    expect(container.querySelectorAll('.unofficial-overlay-pt')).toHaveLength(2);
+    const updatedGroups = dotGroups(container);
+    const pointsByDate = new Map(
+      updatedGroups.map((group) => {
+        const datum = (group as SVGGElement & { __data__: InferenceData }).__data__;
+        return [datum.date, { datum, transform: group.getAttribute('transform') }] as const;
+      }),
+    );
+    expect(updatedGroups).toEqual(groups);
+    expect(pointsByDate.size).toBe(2);
+    expect(pointsByDate.get('2026-05-15')?.datum).toMatchObject({ x: 30, y: 300 });
+    expect(pointsByDate.get('2026-06-15')?.datum).toMatchObject({ x: 70, y: 700 });
+    expect(pointsByDate.get('2026-05-15')?.transform).not.toBe(
+      pointsByDate.get('2026-06-15')?.transform,
+    );
     expect(rebuildCount()).toBe(buildsAfterMount);
     unmount();
   });
 
-  it('keeps speculative decoding out of unofficial-run point decorations', () => {
-    const runUrl = 'https://github.com/o/r/actions/runs/123';
-    const overlayPoints = [
-      {
-        ...point('h100', 'fp8', 30, 300, 2),
-        benchmark_type: 'agentic_traces',
-        spec_decoding: 'mtp',
-        offload_mode: 'on',
-        run_url: runUrl,
-      } as InferenceData,
-      {
-        ...point('h100', 'fp8', 35, 350, 4),
-        benchmark_type: 'agentic_traces',
-        spec_decoding: 'none',
-        offload_mode: 'off',
-        run_url: runUrl,
-      } as InferenceData,
-    ];
-    overlayState.current = {
-      ...baseOverlayState(),
-      isUnofficialRun: true,
-      activeOverlayHwTypes: new Set(['h100']),
-      allOverlayHwTypes: new Set(['h100']),
-      runIndexByUrl: { [runUrl]: 0 },
-      unofficialRunInfos: [{ id: '123', branch: 'test-branch', url: runUrl }],
+  it('refreshes bound metadata when a single-date series advances at fixed coordinates', () => {
+    const chartProps = {
+      data: [datedPoint('2026-05-15', 20, 200)],
+      xExtentOverride: [0, 100] as [number, number],
+      yExtentOverride: [0, 1000] as [number, number],
+      transitionDuration: 0,
     };
+    const { container, rerender, unmount } = mountChart(chartProps);
+    const group = dotGroups(container)[0];
+    const buildsAfterMount = rebuildCount();
 
-    const { container, unmount } = mountChart({
-      overlayData: {
-        data: overlayPoints,
-        hardwareConfig: HARDWARE_CONFIG,
-      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
-    });
-    const groups = [...container.querySelectorAll<SVGGElement>('.unofficial-overlay-pt')];
+    chartProps.data = [datedPoint('2026-06-15', 20, 200)];
+    rerender();
 
-    expect(groups[0].querySelector('.spec-decode-marker')).toBeNull();
-    expect(groups[0].querySelector('.offload-halo')).not.toBeNull();
-    expect(groups[1].querySelector('.spec-decode-marker')).toBeNull();
-    expect(groups[1].querySelector('.offload-halo')).toBeNull();
+    const updatedGroup = dotGroups(container)[0];
+    const datum = (updatedGroup as SVGGElement & { __data__: InferenceData }).__data__;
+    expect(updatedGroup).toBe(group);
+    expect(datum.date).toBe('2026-06-15');
+    expect(rebuildCount()).toBe(buildsAfterMount);
     unmount();
   });
 
-  it('applies quick filters to unofficial-run overlay markers', () => {
-    const overlayPoints = [point('h100', 'fp8', 30, 300, 2), point('h100', 'fp8', 35, 350, 4)].map(
-      (p) => ({ ...p, run_url: 'https://github.com/o/r/actions/runs/123' }),
-    );
-    overlayState.current = {
-      ...baseOverlayState(),
-      isUnofficialRun: true,
-      activeOverlayHwTypes: new Set(['h100']),
-      allOverlayHwTypes: new Set(['h100']),
-      runIndexByUrl: { 'https://github.com/o/r/actions/runs/123': 0 },
-      unofficialRunInfos: [
-        { id: '123', branch: 'test-branch', url: 'https://github.com/o/r/actions/runs/123' },
-      ],
-    };
-    // Overlay points are all NVIDIA (h100); an AMD-only quick filter must hide them,
-    // exactly as it would the official points.
+  it('updates rooflines and dots in the same domain-change phase', () => {
+    const { container, rerender, unmount } = mountChart({ transitionDuration: 0 });
+    const b200Roofline = container.querySelector<SVGPathElement>(
+      '.roofline-path[data-hw-key="b200"]',
+    )!;
+    const b200Dot = dotGroups(container, 'b200')[0];
+    const pathBefore = b200Roofline.getAttribute('d');
+    const transformBefore = b200Dot.getAttribute('transform');
+
     inferenceState.current = {
-      ...baseInferenceState(),
-      quickFilters: { vendors: ['AMD'], frameworks: [], deployment: [], spec: [] },
+      ...inferenceState.current,
+      activeHwTypes: new Set(['b200']),
     };
-    const { container, unmount } = mountChart({
-      overlayData: {
-        data: overlayPoints,
-        hardwareConfig: HARDWARE_CONFIG,
-      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
-    });
+    rerender();
 
-    expect(container.querySelectorAll('.unofficial-overlay-pt')).toHaveLength(0);
+    expect(b200Roofline.getAttribute('d')).not.toBe(pathBefore);
+    expect(b200Dot.getAttribute('transform')).not.toBe(transformBefore);
     unmount();
   });
 });
diff --git a/packages/app/src/components/inference/ui/ScatterGraph.overlay.test.tsx b/packages/app/src/components/inference/ui/ScatterGraph.overlay.test.tsx
new file mode 100644
index 000000000..889a4a82e
--- /dev/null
+++ b/packages/app/src/components/inference/ui/ScatterGraph.overlay.test.tsx
@@ -0,0 +1,228 @@
+// @vitest-environment jsdom
+import { act } from 'react';
+import { describe, expect, it, vi } from 'vitest';
+
+import type { InferenceData } from '@/components/inference/types';
+
+import {
+  type ScatterGraph,
+  HARDWARE_CONFIG,
+  baseInferenceState,
+  baseOverlayState,
+  inferenceState,
+  legendState,
+  mountChart,
+  overlayState,
+  point,
+  rebuildCount,
+} from './ScatterGraph.test-harness';
+
+describe('ScatterGraph unofficial overlays', () => {
+  it('keeps unofficial-run overlay markers rendered through official toggles', () => {
+    const overlayPoints = [point('h100', 'fp8', 30, 300, 2), point('h100', 'fp8', 35, 350, 4)].map(
+      (p) => ({ ...p, run_url: 'https://github.com/o/r/actions/runs/123' }),
+    );
+    overlayState.current = {
+      ...baseOverlayState(),
+      isUnofficialRun: true,
+      activeOverlayHwTypes: new Set(['h100']),
+      allOverlayHwTypes: new Set(['h100']),
+      runIndexByUrl: { 'https://github.com/o/r/actions/runs/123': 0 },
+      unofficialRunInfos: [
+        { id: '123', branch: 'test-branch', url: 'https://github.com/o/r/actions/runs/123' },
+      ],
+    };
+    const { container, rerender, unmount } = mountChart({
+      overlayData: {
+        data: overlayPoints,
+        hardwareConfig: HARDWARE_CONFIG,
+      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
+    });
+    const buildsAfterMount = rebuildCount();
+
+    expect(container.querySelectorAll('.unofficial-overlay-pt')).toHaveLength(2);
+    expect(container.querySelectorAll('.overlay-roofline-path').length).toBeGreaterThan(0);
+
+    // Toggling an official hw must not rebuild or disturb overlay markers.
+    inferenceState.current = {
+      ...inferenceState.current,
+      activeHwTypes: new Set(['h100']),
+    };
+    rerender();
+
+    expect(container.querySelectorAll('.unofficial-overlay-pt')).toHaveLength(2);
+    expect(rebuildCount()).toBe(buildsAfterMount);
+    unmount();
+  });
+
+  it('toggles official and overlay point labels through the selective display phase', () => {
+    const runUrl = 'https://github.com/o/r/actions/runs/123';
+    const overlayPoints = [
+      { ...point('h100', 'fp8', 30, 300, 2), run_url: runUrl },
+      { ...point('h100', 'fp8', 35, 350, 4), run_url: runUrl },
+    ];
+    inferenceState.current = {
+      ...baseInferenceState(),
+      showPointLabels: false,
+    };
+    overlayState.current = {
+      ...baseOverlayState(),
+      isUnofficialRun: true,
+      activeOverlayHwTypes: new Set(['h100']),
+      allOverlayHwTypes: new Set(['h100']),
+      runIndexByUrl: { [runUrl]: 0 },
+      unofficialRunInfos: [{ id: '123', branch: 'test-branch', url: runUrl }],
+    };
+    const { container, rerender, unmount } = mountChart({
+      overlayData: {
+        data: overlayPoints,
+        hardwareConfig: HARDWARE_CONFIG,
+      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
+    });
+    const buildsAfterMount = rebuildCount();
+    const officialLabel = container.querySelector<SVGTextElement>('.dot-group .point-label');
+    const overlayLabel = container.querySelector<SVGTextElement>(
+      '.unofficial-overlay-pt .overlay-label',
+    );
+
+    expect(officialLabel).not.toBeNull();
+    expect(overlayLabel).not.toBeNull();
+    expect(officialLabel!.style.display).toBe('none');
+    expect(overlayLabel!.style.display).toBe('none');
+    const mutationRecords: MutationRecord[] = [];
+    const observer = new MutationObserver((records) => mutationRecords.push(...records));
+    observer.observe(container.querySelector('svg')!, { attributes: true, subtree: true });
+
+    inferenceState.current = {
+      ...inferenceState.current,
+      showPointLabels: true,
+    };
+    rerender();
+
+    expect(officialLabel!.style.display).toBe('');
+    expect(overlayLabel!.style.display).toBe('');
+    mutationRecords.push(...observer.takeRecords());
+    observer.disconnect();
+    expect(mutationRecords.length).toBeGreaterThan(0);
+    expect(
+      mutationRecords.every(
+        ({ target }) =>
+          target instanceof Element && target.closest('.point-label, .overlay-label') !== null,
+      ),
+    ).toBe(true);
+    expect(rebuildCount()).toBe(buildsAfterMount);
+    unmount();
+  });
+
+  it('disables Best per SKU for overlay edits without applying a context selection', () => {
+    const setBestPerSku = vi.fn();
+    const runUrl = 'https://github.com/o/r/actions/runs/123';
+    const overlayPoints = [
+      { ...point('h100', 'fp8', 30, 300, 2), run_url: runUrl },
+      { ...point('h100', 'fp8', 35, 350, 4), run_url: runUrl },
+    ];
+    inferenceState.current = {
+      ...baseInferenceState(),
+      bestPerSku: true,
+      setBestPerSku,
+    };
+    overlayState.current = {
+      ...baseOverlayState(),
+      isUnofficialRun: true,
+      activeOverlayHwTypes: new Set(['h100']),
+      allOverlayHwTypes: new Set(['h100']),
+      runIndexByUrl: { [runUrl]: 0 },
+      unofficialRunInfos: [{ id: '123', branch: 'test-branch', url: runUrl }],
+    };
+
+    const { unmount } = mountChart({
+      overlayData: {
+        data: overlayPoints,
+        hardwareConfig: HARDWARE_CONFIG,
+      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
+    });
+    const officialItem = legendState.current!.legendItems.find(
+      (item: { hw: string }) => item.hw === 'h100',
+    );
+
+    act(() => officialItem.onClick());
+    expect(setBestPerSku).toHaveBeenLastCalledWith(false, { applySelection: false });
+
+    act(() => legendState.current!.onItemRemove('h100'));
+    expect(setBestPerSku).toHaveBeenLastCalledWith(false, { applySelection: false });
+    unmount();
+  });
+
+  it('keeps speculative decoding out of unofficial-run point decorations', () => {
+    const runUrl = 'https://github.com/o/r/actions/runs/123';
+    const overlayPoints = [
+      {
+        ...point('h100', 'fp8', 30, 300, 2),
+        benchmark_type: 'agentic_traces',
+        spec_decoding: 'mtp',
+        offload_mode: 'on',
+        run_url: runUrl,
+      } as InferenceData,
+      {
+        ...point('h100', 'fp8', 35, 350, 4),
+        benchmark_type: 'agentic_traces',
+        spec_decoding: 'none',
+        offload_mode: 'off',
+        run_url: runUrl,
+      } as InferenceData,
+    ];
+    overlayState.current = {
+      ...baseOverlayState(),
+      isUnofficialRun: true,
+      activeOverlayHwTypes: new Set(['h100']),
+      allOverlayHwTypes: new Set(['h100']),
+      runIndexByUrl: { [runUrl]: 0 },
+      unofficialRunInfos: [{ id: '123', branch: 'test-branch', url: runUrl }],
+    };
+
+    const { container, unmount } = mountChart({
+      overlayData: {
+        data: overlayPoints,
+        hardwareConfig: HARDWARE_CONFIG,
+      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
+    });
+    const groups = [...container.querySelectorAll<SVGGElement>('.unofficial-overlay-pt')];
+
+    expect(groups[0].querySelector('.spec-decode-marker')).toBeNull();
+    expect(groups[0].querySelector('.offload-halo')).not.toBeNull();
+    expect(groups[1].querySelector('.spec-decode-marker')).toBeNull();
+    expect(groups[1].querySelector('.offload-halo')).toBeNull();
+    unmount();
+  });
+
+  it('applies quick filters to unofficial-run overlay markers', () => {
+    const overlayPoints = [point('h100', 'fp8', 30, 300, 2), point('h100', 'fp8', 35, 350, 4)].map(
+      (p) => ({ ...p, run_url: 'https://github.com/o/r/actions/runs/123' }),
+    );
+    overlayState.current = {
+      ...baseOverlayState(),
+      isUnofficialRun: true,
+      activeOverlayHwTypes: new Set(['h100']),
+      allOverlayHwTypes: new Set(['h100']),
+      runIndexByUrl: { 'https://github.com/o/r/actions/runs/123': 0 },
+      unofficialRunInfos: [
+        { id: '123', branch: 'test-branch', url: 'https://github.com/o/r/actions/runs/123' },
+      ],
+    };
+    // Overlay points are all NVIDIA (h100); an AMD-only quick filter must hide them,
+    // exactly as it would the official points.
+    inferenceState.current = {
+      ...baseInferenceState(),
+      quickFilters: { vendors: ['AMD'], frameworks: [], deployment: [], spec: [] },
+    };
+    const { container, unmount } = mountChart({
+      overlayData: {
+        data: overlayPoints,
+        hardwareConfig: HARDWARE_CONFIG,
+      } as unknown as Parameters<typeof ScatterGraph>[0]['overlayData'],
+    });
+
+    expect(container.querySelectorAll('.unofficial-overlay-pt')).toHaveLength(0);
+    unmount();
+  });
+});
diff --git a/packages/app/src/components/inference/ui/ScatterGraph.test-harness.tsx b/packages/app/src/components/inference/ui/ScatterGraph.test-harness.tsx
new file mode 100644
index 000000000..952013187
--- /dev/null
+++ b/packages/app/src/components/inference/ui/ScatterGraph.test-harness.tsx
@@ -0,0 +1,254 @@
+import { act, createElement, useReducer } from 'react';
+import { createRoot } from 'react-dom/client';
+import { afterEach, beforeEach, vi } from 'vitest';
+
+import { setupChartStructure } from '@/lib/d3-chart/chart-setup';
+import type { ChartDefinition, InferenceData } from '@/components/inference/types';
+import { computeToggle } from '@/hooks/useTogglableSet';
+
+vi.mock('@/lib/d3-chart/chart-setup', { spy: true });
+vi.mock('@/lib/analytics', () => ({ track: vi.fn() }));
+vi.mock('next-themes', () => ({ useTheme: () => ({ resolvedTheme: 'dark' }) }));
+
+declare global {
+  var __scatterLegendState: { current: Record<string, any> | null };
+  var __scatterInferenceState: { current: Record<string, unknown> };
+  var __scatterOverlayState: { current: Record<string, unknown> };
+  var __scatterTraceAvailabilityState: { current: Record<number, boolean> | undefined };
+  var __scatterLogAvailabilityState: { current: Record<number, boolean> | undefined };
+}
+const availabilityIdsGlobal = vi.hoisted(() => ({
+  traceIds: [] as number[],
+  logIds: [] as number[],
+}));
+vi.hoisted(() => {
+  globalThis.__scatterLegendState = { current: null };
+  globalThis.__scatterInferenceState = { current: {} };
+  globalThis.__scatterOverlayState = { current: {} };
+  globalThis.__scatterTraceAvailabilityState = { current: undefined };
+  globalThis.__scatterLogAvailabilityState = { current: undefined };
+});
+
+vi.mock('@/components/ui/chart-legend', () => ({
+  default: (props: Record<string, any>) => {
+    globalThis.__scatterLegendState.current = props;
+    return props.keyIndicators ?? null;
+  },
+}));
+vi.mock('@/components/inference/InferenceContext', () => ({
+  useInferenceActions: () => globalThis.__scatterInferenceState.current,
+  useInferenceData: () => globalThis.__scatterInferenceState.current,
+  useInferenceDisplay: () => globalThis.__scatterInferenceState.current,
+  useInferenceFilters: () => globalThis.__scatterInferenceState.current,
+}));
+vi.mock('@/components/unofficial-run-provider', () => ({
+  useUnofficialRun: () => globalThis.__scatterOverlayState.current,
+}));
+
+export const legendState = globalThis.__scatterLegendState;
+export const inferenceState = globalThis.__scatterInferenceState;
+export const logAvailabilityState = globalThis.__scatterLogAvailabilityState;
+export const overlayState = globalThis.__scatterOverlayState;
+export const traceAvailabilityState = globalThis.__scatterTraceAvailabilityState;
+export const availabilityState = availabilityIdsGlobal;
+vi.mock('@/hooks/api/use-trace-availability', () => ({
+  useTraceAvailability: (ids: number[]) => {
+    availabilityIdsGlobal.traceIds = ids ?? [];
+    return { data: globalThis.__scatterTraceAvailabilityState.current, isPending: false };
+  },
+}));
+vi.mock('@/hooks/api/use-log-availability', () => ({
+  useLogAvailability: (ids: number[]) => {
+    availabilityIdsGlobal.logIds = ids ?? [];
+    return { data: globalThis.__scatterLogAvailabilityState.current, isPending: false };
+  },
+}));
+
+import ScatterGraph from './ScatterGraph';
+
+class MockResizeObserver {
+  observe() {}
+  unobserve() {}
+  disconnect() {}
+}
+const originalGetBBox = Object.getOwnPropertyDescriptor(SVGElement.prototype, 'getBBox');
+
+export const point = (
+  hwKey: string,
+  precision: string,
+  x: number,
+  y: number,
+  tp: number,
+): InferenceData =>
+  ({ hwKey, precision, x, y, tp, conc: 16, framework: 'vllm' }) as unknown as InferenceData;
+
+export const POINTS: InferenceData[] = [
+  point('h100', 'fp8', 1, 1, 1),
+  point('h100', 'fp8', 100, 1000, 8),
+  point('h100', 'fp4', 40, 400, 4),
+  point('b200', 'fp8', 50, 500, 4),
+  point('b200', 'fp8', 60, 600, 8),
+];
+
+export const HARDWARE_CONFIG = {
+  h100: { name: 'H100', label: 'H100', gpu: 'H100' },
+  b200: { name: 'B200', label: 'B200', gpu: 'B200' },
+};
+
+const CHART_DEFINITION = { chartType: 'interactivity' } as unknown as ChartDefinition;
+export const noop = () => {};
+
+export function baseInferenceState() {
+  return {
+    activeHwTypes: new Set(['h100', 'b200']),
+    hardwareConfig: HARDWARE_CONFIG,
+    toggleHwType: noop,
+    removeHwType: noop,
+    hwTypesWithData: new Set(['h100', 'b200']),
+    resolveComparisonSelection: (proposed: Set<string>) => ({
+      result: proposed,
+      keptGroup: null,
+      droppedGroups: [],
+    }),
+    toggleComparisonSelection: (prev: Set<string>, item: string, allItems: Set<string>) =>
+      computeToggle(prev, item, allItems),
+    selectedPrecisions: ['fp8'],
+    selectedYAxisMetric: 'y',
+    tokenRevenuePriceSource: 'normalized' as const,
+    tokenRevenuePricing: {
+      source: 'normalized' as const,
+      inputPerMillion: 1,
+      outputPerMillion: 1,
+    },
+    openRouterModelId: null,
+    openRouterPricingLoading: false,
+    openRouterPricingError: null,
+    setTokenRevenuePriceSource: noop,
+    quickFilters: { vendors: [], frameworks: [], deployment: [], spec: [] },
+    availableQuickFilters: { vendors: [], frameworks: [], deployment: [], spec: [] },
+    availableRuns: {},
+    selectedRunId: '',
+    hideNonOptimal: false,
+    setHideNonOptimal: noop,
+    hidePointLabels: false,
+    setHidePointLabels: noop,
+    selectAllHwTypes: noop,
+    highContrast: false,
+    setHighContrast: noop,
+    logScale: false,
+    setLogScale: noop,
+    scaleType: 'auto',
+    isLegendExpanded: false,
+    setIsLegendExpanded: noop,
+    useAdvancedLabels: false,
+    setUseAdvancedLabels: noop,
+    showConcurrencyLabels: false,
+    setShowConcurrencyLabels: noop,
+    showGradientLabels: false,
+    setShowGradientLabels: noop,
+    showLineLabels: false,
+    setShowLineLabels: noop,
+  };
+}
+
+export function baseOverlayState() {
+  return {
+    isUnofficialRun: false,
+    activeOverlayHwTypes: new Set<string>(),
+    allOverlayHwTypes: new Set<string>(),
+    localOfficialOverride: null,
+    setUnifiedOverlaySelection: noop,
+    resetOverlaySelection: noop,
+    runIndexByUrl: {},
+    unofficialRunInfos: [],
+  };
+}
+
+export function mountChart(props?: Partial<Parameters<typeof ScatterGraph>[0]>) {
+  let forceUpdate: () => void = noop;
+  function Harness() {
+    const [version, bump] = useReducer((value: number) => value + 1, 0);
+    forceUpdate = bump;
+    return createElement(ScatterGraph, {
+      chartId: 'chart-test',
+      modelLabel: 'DeepSeek-R1-0528',
+      data: POINTS,
+      xLabel: 'Interactivity (tok/s/user)',
+      yLabel: 'Output Throughput per GPU',
+      chartDefinition: CHART_DEFINITION,
+      transitionDuration: 0,
+      caption: `v${version}`,
+      ...props,
+    });
+  }
+
+  const container = document.createElement('div');
+  document.body.append(container);
+  const root = createRoot(container);
+  act(() => root.render(createElement(Harness)));
+  return {
+    container,
+    rerender: () => act(() => forceUpdate()),
+    unmount: () => {
+      act(() => root.unmount());
+      container.remove();
+    },
+  };
+}
+
+export const dotGroups = (container: HTMLElement, hwKey?: string) =>
+  [...container.querySelectorAll<SVGGElement>('.dot-group')].filter(
+    (node) => !hwKey || node.dataset.hwKey === hwKey,
+  );
+
+export const rebuildCount = () => vi.mocked(setupChartStructure).mock.calls.length;
+
+beforeEach(() => {
+  vi.stubGlobal('ResizeObserver', MockResizeObserver);
+  Object.defineProperty(SVGElement.prototype, 'getBBox', {
+    configurable: true,
+    value: () =>
+      ({
+        x: 0,
+        y: 0,
+        width: 48,
+        height: 12,
+        top: 0,
+        right: 48,
+        bottom: 12,
+        left: 0,
+        toJSON: () => ({}),
+      }) as DOMRect,
+  });
+  vi.spyOn(HTMLElement.prototype, 'getBoundingClientRect').mockReturnValue({
+    width: 800,
+    height: 600,
+    top: 0,
+    left: 0,
+    right: 800,
+    bottom: 600,
+    x: 0,
+    y: 0,
+    toJSON: () => ({}),
+  } as DOMRect);
+  inferenceState.current = baseInferenceState();
+  overlayState.current = baseOverlayState();
+  legendState.current = null;
+  traceAvailabilityState.current = undefined;
+  logAvailabilityState.current = undefined;
+  availabilityState.traceIds = [];
+  availabilityState.logIds = [];
+  vi.mocked(setupChartStructure).mockClear();
+});
+
+afterEach(() => {
+  vi.unstubAllGlobals();
+  vi.restoreAllMocks();
+  if (originalGetBBox) {
+    Object.defineProperty(SVGElement.prototype, 'getBBox', originalGetBBox);
+  } else {
+    Reflect.deleteProperty(SVGElement.prototype, 'getBBox');
+  }
+});
+
+export { ScatterGraph };
diff --git a/packages/app/src/components/inference/ui/ScatterGraph.tsx b/packages/app/src/components/inference/ui/ScatterGraph.tsx
index 063d007c2..ef8e5efe4 100644
--- a/packages/app/src/components/inference/ui/ScatterGraph.tsx
+++ b/packages/app/src/components/inference/ui/ScatterGraph.tsx
@@ -1,21 +1,43 @@
 'use client';
 
 import { track } from '@/lib/analytics';
+import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
+import { useEphemeralUrlState } from '@/hooks/useUrlState';
+import { rememberChartStateInUrl } from '@/lib/url-state';
 import * as d3 from 'd3';
+import dynamic from 'next/dynamic';
 import React, { useCallback, useEffect, useLayoutEffect, useMemo, useRef, useState } from 'react';
 
+import { SCATTER_RENDERED_EVENT } from '@/lib/nudges/agentic-point-coach-mark';
 import { GRADIENT_NUDGE_EVENT } from '@/lib/nudges/registry';
-import { useInference } from '@/components/inference/InferenceContext';
+import {
+  useInferenceActions,
+  useInferenceData,
+  useInferenceDisplay,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
 import { useTraceAvailability } from '@/hooks/api/use-trace-availability';
+import { useLogAvailability } from '@/hooks/api/use-log-availability';
 import { computeToggle } from '@/hooks/useTogglableSet';
-import { pointNearestX } from '@/components/inference/ui/line-label-anchor';
+import {
+  avoidPointLabelCollisions,
+  parallelismLabelBoxes,
+  placeLineLabels,
+  placeEndpointLineLabels,
+  updateRenderedLineLabels,
+  renderLineLabels,
+  type LineLabelPlacement,
+  type LineLabelSeries,
+} from '@/components/inference/ui/line-label-layer';
 import {
   labelOpacityForActiveState,
   labelOpacityForHover,
 } from '@/components/inference/ui/line-label-visibility';
 import ChartLegend from '@/components/ui/chart-legend';
+import { Button } from '@/components/ui/button';
 import { useUnofficialRun } from '@/components/unofficial-run-provider';
-import { getHardwareConfig, getModelSortIndex } from '@/lib/constants';
+import { OFFICIAL_PREVIEW_SERIES } from '@/components/official-preview-notice';
+import { getHardwareConfig, getModelSortIndex, hardwareKeyMatchesAnyBase } from '@/lib/constants';
 import {
   getChartWatermark,
   getPrecisionLabel,
@@ -25,6 +47,7 @@ import {
 } from '@/lib/data-mappings';
 import { matchKnownConfigIssues, pointMatchesIssue } from '@/lib/known-issues';
 import { useLocale } from '@/lib/use-locale';
+import { getLineLabelVendorIcon } from '@/lib/vendor-logos';
 import { formatNumber, getDisplayLabel, updateRepoUrl } from '@/lib/utils';
 import { D3Chart } from '@/lib/d3-chart/D3Chart';
 import type {
@@ -36,6 +59,31 @@ import type {
 } from '@/lib/d3-chart/D3Chart/types';
 import type { ContinuousScale } from '@/lib/d3-chart/types';
 import { computeTooltipPosition, syncPointShape } from '@/lib/d3-chart/layers/scatter-points';
+import {
+  EMPTY_PERF_RULER_STATE,
+  clampIsoX,
+  clearPerfRulers,
+  computeIsoXRulerGeometry,
+  computePerfRulerLabelLayouts,
+  deletePerfRuler,
+  intersectPathAtX,
+  isPerfRulerCurveVisible,
+  movePerfRulerIsoX,
+  nextPerfRulerState,
+  pathXExtent,
+  perfRulerCurveSet,
+  prunePerfRulers,
+  renderPerfRulers,
+  type PerfRulerEndInput,
+  type PerfRulerGeometry,
+  type PerfRulerRenderEntry,
+  type PerfRulerState,
+} from '@/lib/d3-chart/layers/perf-ruler';
+import {
+  attachOverlayXMarkerHandlers,
+  overlayMarkerPosition,
+  xMarkerPath,
+} from '@/lib/d3-chart/overlay-x-marker';
 import { useStableValue } from '@/hooks/useStableValue';
 import {
   overlayRooflineDasharray,
@@ -57,7 +105,6 @@ import {
   type ParetoDirection,
 } from '@/lib/chart-utils';
 import { canonicalParetoIntersection } from '@/components/inference/utils/canonicalFrontier';
-import { type RooflineDirection, getSpeedOverlayCorners } from '@/lib/speed-overlay';
 import type {
   ChartDefinition,
   ClippedInferenceData,
@@ -67,18 +114,16 @@ import type {
 import {
   generateOverlayTooltipContent,
   generateTooltipContent,
-  getPointLabel,
 } from '@/components/inference/utils/tooltipUtils';
-import { scatterPointConfigId } from '@/components/inference/utils/point-identity';
-import LegendPointsDialog from '@/components/inference/ui/LegendPointsDialog';
+import { QuickFiltersDialog } from '@/components/inference/ui/QuickFiltersDialog';
 import {
-  OFFLOAD_HALO_DASHARRAY,
-  OFFLOAD_HALO_RADIUS,
-  OFFLOAD_HALO_STROKE_WIDTH,
-  OffloadHaloLegendKey,
-} from '@/components/inference/ui/OffloadHaloLegendKey';
-import { AgenticOptimizationNote } from '@/components/inference/ui/AgenticOptimizationNote';
+  scatterPointConfigId,
+  scatterPointJoinId,
+} from '@/components/inference/utils/point-identity';
+import LegendPointsDialog from '@/components/inference/ui/LegendPointsDialog';
+import { renderOffloadHalo } from '@/components/inference/utils/offload-halo';
 import { buildLegendPointsRows } from '@/components/inference/utils/legend-points-table';
+import { pointLabelText } from './point-label';
 import {
   type ParetoPointLabel,
   getParetoLabel,
@@ -89,8 +134,7 @@ import {
 } from '@/components/inference/utils/paretoLabels';
 import {
   type KnownIssueAnnotation,
-  measureLegendRightInset,
-  renderKnownIssueAnnotations,
+  createKnownIssueLayer,
 } from '@/components/inference/utils/knownIssueAnnotations';
 import { matchesQuickFilters } from '@/components/inference/utils/quickFilters';
 import { bestSeriesPerSku } from '@/components/inference/utils/best-series-per-sku';
@@ -100,126 +144,13 @@ import {
   fitContinuationLabelBaseline,
 } from '@/components/inference/utils/overflowContinuations';
 
-// Greedy label-collision avoidance.
-// Each candidate is the y-position of the FIRST baseline (relative to point
-// center) which we apply via the first tspan's `dy` — later tspans cascade
-// down by 1.1em. We try above/below at primary and secondary offsets, and
-// hide the label if all four positions collide.
-function avoidLabelCollisions(
-  zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
-): void {
-  interface LabelInfo {
-    el: SVGTextElement;
-    firstTspan: SVGTSpanElement;
-    cx: number;
-    cy: number;
-    w: number;
-    nLines: number;
-    defaultFirstY: number;
-  }
-  const pending: Omit<LabelInfo, 'w'>[] = [];
-  const ASCENT = 9;
-  const DESCENT = 3;
-  const LINE_H = 11;
-
-  // Pass 1 — writes only: reset every label to its default position so prior
-  // positioning doesn't bias the measurement.
-  zoomGroup.selectAll<SVGGElement, unknown>('.dot-group').each(function () {
-    const labelEl = this.querySelector<SVGTextElement>('.point-label');
-    if (!labelEl) return;
-    if ((this as SVGGElement).style.opacity === '0') return;
-    const tspans = labelEl.querySelectorAll<SVGTSpanElement>('tspan');
-    if (tspans.length === 0) return;
-    const transform = (this as SVGGElement).getAttribute('transform') ?? '';
-    const m = transform.match(/translate\((?<tx>[^,]+),(?<ty>[^)]+)\)/u);
-    if (!m) return;
-    const cx = parseFloat(m[1]);
-    const cy = parseFloat(m[2]);
-    const nLines = tspans.length;
-    const defaultFirstY = -(8 + (nLines - 1) * LINE_H); // last baseline 8px above point
-    tspans[0].setAttribute('dy', `${defaultFirstY}px`);
-    labelEl.style.opacity = '1';
-    pending.push({
-      el: labelEl,
-      firstTspan: tspans[0],
-      cx,
-      cy,
-      nLines,
-      defaultFirstY,
-    });
-  });
-
-  // Pass 2 — reads only: measure after all writes so the whole batch costs a
-  // single forced layout instead of one per label.
-  const labels: LabelInfo[] = pending.map((lab) => ({ ...lab, w: lab.el.getBBox().width }));
-
-  labels.sort((a, b) => a.cx - b.cx);
-  const placed: { left: number; right: number; top: number; bottom: number }[] = [];
-  const pad = 2;
-
-  for (const lab of labels) {
-    const blockH = (lab.nLines - 1) * LINE_H + ASCENT + DESCENT;
-    const aboveFirstY = lab.defaultFirstY;
-    const belowFirstY = 14; // first baseline 14px below point center
-    const candidates = [
-      aboveFirstY,
-      belowFirstY,
-      aboveFirstY - blockH - 2,
-      belowFirstY + blockH + 2,
-    ];
-    let chosenY: number | null = null;
-    let chosenBox: { left: number; right: number; top: number; bottom: number } | null = null;
-    for (const firstY of candidates) {
-      const top = lab.cy + firstY - ASCENT - pad;
-      const bottom = lab.cy + firstY + (lab.nLines - 1) * LINE_H + DESCENT + pad;
-      const left = lab.cx - lab.w / 2 - pad;
-      const right = lab.cx + lab.w / 2 + pad;
-      const collides = placed.some(
-        (p) => !(right < p.left || left > p.right || bottom < p.top || top > p.bottom),
-      );
-      if (!collides) {
-        chosenY = firstY;
-        chosenBox = { left, right, top, bottom };
-        break;
-      }
-    }
-    if (chosenY !== null && chosenBox) {
-      lab.firstTspan.setAttribute('dy', `${chosenY}px`);
-      lab.el.style.opacity = '1';
-      placed.push(chosenBox);
-    } else {
-      lab.el.style.opacity = '0';
-    }
-  }
-}
-
-// X-shape path for overlay (unofficial) data points
-const X_SIZE = 5;
-const X_HOVER_SIZE = 7;
-const getXPath = (size: number) => {
-  const s = size * 0.7;
-  return `M ${-s} ${-s} L ${s} ${s} M ${s} ${-s} L ${-s} ${s}`;
-};
-
-/** Render the KV-offload halo around a point when applicable. */
-function renderOffloadHalo(
-  group: d3.Selection<SVGGElement, InferenceData, null, undefined>,
-  point: InferenceData,
-  stroke: string,
-): void {
-  group
-    .selectAll<SVGCircleElement, boolean>('.offload-halo')
-    .data(point.offload_mode === 'on' ? [true] : [])
-    .join('circle')
-    .attr('class', 'offload-halo')
-    .attr('r', OFFLOAD_HALO_RADIUS)
-    .attr('fill', 'none')
-    .attr('stroke', stroke)
-    .attr('stroke-width', OFFLOAD_HALO_STROKE_WIDTH)
-    .attr('stroke-dasharray', OFFLOAD_HALO_DASHARRAY)
-    .attr('opacity', 0.9)
-    .attr('pointer-events', 'none');
-}
+const FixedSequenceLogDialog = dynamic(
+  () =>
+    import('@/components/inference/log-viewer/fixed-sequence-log-dialog').then(
+      (module) => module.FixedSequenceLogDialog,
+    ),
+  { ssr: false },
+);
 
 const formatChangelogDescription = (desc: string | string[]): React.JSX.Element => {
   if (typeof desc === 'string') {
@@ -267,10 +198,6 @@ function groupPointsByDate(points: InferenceData[]): Map<string, InferenceData[]
 const optimalPointKey = (d: InferenceData): string =>
   `${d.hwKey}_${d.precision}_${d.date}-${d.x}-${d.y}`;
 
-/** Point label lines. Decode mode is shown by a point marker instead of text. */
-export const pointLabelText = (d: InferenceData, advanced: boolean): string =>
-  advanced ? `${getPointLabel(d)}\nC=${d.conc}` : `${d.tp}\nC=${d.conc}`;
-
 // Referentially stable "no overlay data" result (see processedOverlayData).
 const EMPTY_OVERLAY_DATA: InferenceData[] = [];
 const EMPTY_CLIPPED_DATA: ClippedInferenceData[] = [];
@@ -360,11 +287,9 @@ function visibleContinuationEndpoint(
   };
 }
 
-// Scale configs are recomputed from the visible points on every render, but a
-// legend / precision toggle usually leaves the actual domain untouched (x-min
-// is pinned at 0; extremes are owned by a handful of points). Comparing by
-// value lets those toggles keep the previous config object, so the chart
-// render effect doesn't tear down and rebuild the SVG for identical scales.
+// Scale configs are recomputed from visible points, but a legend or precision
+// toggle often leaves the domain unchanged. Comparing by value preserves the
+// previous config object and prevents a redundant metric-coordinate phase.
 interface ScaleConfigValue {
   type: 'log' | 'linear';
   domain: [number, number];
@@ -390,6 +315,16 @@ const hasNamedTransition = (node: Element, name: string): boolean => {
   return Object.values(schedules).some((schedule) => schedule?.name === name);
 };
 
+const currentZoomRenderContext = (svg: SVGSVGElement, ctx: RenderContext): RenderContext => {
+  const transform = d3.zoomTransform(svg);
+  if (transform.k === 1 && transform.x === 0 && transform.y === 0) return ctx;
+  return {
+    ...ctx,
+    xScale: transform.rescaleX(ctx.xScale as ContinuousScale),
+    yScale: transform.rescaleY(ctx.yScale as ContinuousScale),
+  };
+};
+
 // Derive a readable label from a hwKey using the HARDWARE_CONFIG source of truth.
 // `model` (display name) enables per-model suffix overrides (e.g. M3 MTP → EAGLE).
 const parseHwKeyToLabel = (hwKey: string, model?: string): { name: string; label: string } => {
@@ -398,16 +333,22 @@ const parseHwKeyToLabel = (hwKey: string, model?: string): { name: string; label
 };
 
 // Line-label text for a curve. When more than one precision is shown, each curve
-// is its own line, so append the precision (e.g. "B200 (vLLM) FP8") to keep the
-// FP4 and FP8 curves of the same hardware distinguishable.
+// is its own line, so place the precision between the GPU and engine (e.g.
+// "B200 FP8 (vLLM)") to keep the primary identifiers together.
 const lineLabelText = (
   hwKey: string,
   precision: string,
   includePrecision: boolean,
   model?: string,
 ): string => {
-  const base = parseHwKeyToLabel(hwKey, model).label;
-  return includePrecision ? `${base} ${getPrecisionLabel(precision as Precision)}` : base;
+  const config = getHardwareConfig(hwKey, model);
+  if (config.alwaysShowPrecision) {
+    return [getDisplayLabel(config), getPrecisionLabel(precision as Precision)].join(' ');
+  }
+  if (!includePrecision) return getDisplayLabel(config);
+  return [config.label, getPrecisionLabel(precision as Precision), config.suffix]
+    .filter(Boolean)
+    .join(' ');
 };
 
 const pointCountEn = (count: number) => `${count} ${count === 1 ? 'point' : 'points'}`;
@@ -416,15 +357,19 @@ const SCATTER_STRINGS = {
   en: {
     logScale: 'Log Scale',
     optimalOnly: 'Optimal Only',
-    bestPerSku: 'Best per SKU',
-    optimalInfo:
-      'On agentic, optimal points must be Pareto-optimal on the selected x-axis and also belong to the E2E Normalized Interactivity frontier.',
+    optimalInfo: 'Optimal points form the Pareto frontier for the selected axes.',
     labels: 'Labels',
     highContrast: 'High Contrast',
     parallelismLabels: 'Parallelism Labels',
+    concurrencyLabels: '# Concurrent Sessions',
     gradientLabels: 'Gradient Labels',
     lineLabels: 'Line Labels',
+    perfRuler: 'Perf Ruler',
+    perfRulerInfo:
+      'Click two curves to place a vertical ruler, then drag it to measure the performance multiple between them at any x value. Repeat to add more rulers (up to 8); hover a ruler and click × to delete it. Turning the toggle off clears all rulers.',
     resetFilter: 'Reset filter',
+    clearPerfRulers: (count: number) => `Clear rulers (${count})`,
+    quickFilters: (count: number) => (count > 0 ? `Quick Filters (${count})` : 'Quick Filters'),
     overflowMixed: (count: number) => `${pointCountEn(count)} clipped`,
     overflowCost: (count: number, limit: number) => `${pointCountEn(count)} > $${limit}/Mtok`,
     overflowLatency: (count: number, limit: number) => `${pointCountEn(count)} > ${limit}s TTFT`,
@@ -432,15 +377,19 @@ const SCATTER_STRINGS = {
   zh: {
     logScale: '对数缩放',
     optimalOnly: '仅最优',
-    bestPerSku: '每个 SKU 仅显示最佳配置',
-    optimalInfo:
-      '在智能体场景中，最优点既必须在当前横轴上满足 Pareto 最优，也必须属于端到端归一化交互性的 Pareto 前沿。',
+    optimalInfo: '最优点构成当前所选坐标轴的 Pareto 前沿。',
     labels: '标签',
     highContrast: '高对比度',
     parallelismLabels: '并行配置标签',
+    concurrencyLabels: '并发会话数',
     gradientLabels: '渐变标签',
     lineLabels: '曲线标签',
+    perfRuler: '性能标尺',
+    perfRulerInfo:
+      '先点击两条曲线放置垂直标尺，再拖动标尺，测量任意横坐标下两条曲线之间的性能倍数。重复操作可添加多把标尺（最多 8 把）；悬停标尺并点击 × 可删除该标尺。关闭开关将清除所有标尺。',
     resetFilter: '重置筛选',
+    clearPerfRulers: (count: number) => `清除标尺（${count}）`,
+    quickFilters: (count: number) => (count > 0 ? `快捷筛选（${count}）` : '快捷筛选'),
     overflowMixed: (count: number) => `${count} 个点已截断`,
     overflowCost: (count: number, limit: number) => `${count} 个点 > $${limit}/Mtok`,
     overflowLatency: (count: number, limit: number) => `${count} 个点 > ${limit}s TTFT`,
@@ -460,63 +409,70 @@ const ScatterGraph = React.memo(
     showAllHardwareTypes = false,
     hardwareConfigOverride,
     overlayData,
-    transitionDuration = 750,
+    transitionDuration = 300,
     niceAxes = true,
     pinLineLabels = false,
     xExtentOverride,
     yExtentOverride,
   }: ScatterGraphProps) => {
+    const {
+      hardwareConfig: contextHardwareConfig,
+      hwTypesWithData,
+      availableRuns,
+    } = useInferenceData();
     const {
       activeHwTypes,
       bestPerSku,
+      selectedPrecisions,
+      selectedRunId,
+      selectedSequence,
+      quickFilters,
+    } = useInferenceFilters();
+    const {
+      selectedYAxisMetric,
+      hideNonOptimal,
+      showPointLabels,
+      highContrast,
+      logScale,
+      scaleType,
+      isLegendExpanded,
+      useAdvancedLabels,
+      showConcurrencyLabels,
+      showGradientLabels,
+      showLineLabels,
+    } = useInferenceDisplay();
+    const {
       setBestPerSku,
-      hardwareConfig: contextHardwareConfig,
       toggleHwType,
       removeHwType,
-      hwTypesWithData,
       resolveComparisonSelection,
-      selectedPrecisions,
-      selectedYAxisMetric,
-      availableRuns,
-      selectedRunId,
-      hideNonOptimal,
       setHideNonOptimal,
-      showPointLabels,
       setShowPointLabels,
       selectAllHwTypes,
-      highContrast,
       setHighContrast,
-      logScale,
       setLogScale,
-      scaleType,
-      isLegendExpanded,
       setIsLegendExpanded,
-      useAdvancedLabels,
       setUseAdvancedLabels,
-      showGradientLabels,
+      setShowConcurrencyLabels,
       setShowGradientLabels,
-      showLineLabels,
       setShowLineLabels,
-      showSpeedOverlay,
-      setShowSpeedOverlay,
-      showMinecraftOverlay,
-      setShowMinecraftOverlay,
-      selectedSequence,
-      selectedModel,
-      quickFilters,
-      loading,
-    } = useInference();
+      setQuickFilterVendors,
+      setQuickFilterFrameworks,
+      setQuickFilterDeployment,
+      setQuickFilterSpec,
+    } = useInferenceActions();
     const locale = useLocale();
     const legendT = SCATTER_STRINGS[locale];
+    const ephemeralUrlState = useEphemeralUrlState();
     const costLimit = chartDefinition.y_cost_limit ?? 0;
     const latencyLimit = chartDefinition.y_latency_limit ?? 0;
 
     const {
       isUnofficialRun,
       activeOverlayHwTypes: providerActiveOverlayHwTypes,
-      setActiveOverlayHwTypes,
       localOfficialOverride,
-      setLocalOfficialOverride,
+      resetOverlaySelection,
+      setUnifiedOverlaySelection,
       runIndexByUrl,
       unofficialRunInfos,
     } = useUnofficialRun();
@@ -539,45 +495,14 @@ const ScatterGraph = React.memo(
       }
       return keys;
     }, [overlayData, selectedPrecisions, quickFilters]);
-    const overlayScopeKey = `${selectedModel}|${selectedSequence}|${selectedPrecisions.join(',')}`;
-    const previousOverlayScopeRef = useRef(overlayScopeKey);
-    // Scope seeding is preview-only. Official charts have no overlay lifecycle
-    // to commit the ref below, so treating their key changes as pending would
-    // bypass subsequent activeHwTypes legend toggles indefinitely.
-    const overlayScopeChanged =
-      Boolean(overlayData) && previousOverlayScopeRef.current !== overlayScopeKey;
-
-    const localOfficialOverrideIsStale = useMemo(() => {
-      if (localOfficialOverride === null) return false;
-      if (overlayScopeChanged) return true;
-      if (localOfficialOverride.size === 0) return false;
-      for (const key of localOfficialOverride) {
-        if (hwTypesWithData.has(key)) return false;
-      }
-      return true;
-    }, [localOfficialOverride, overlayScopeChanged, hwTypesWithData]);
     const rawOfficialHwTypes = useMemo(() => {
-      const source = overlayScopeChanged
-        ? hwTypesWithData
-        : localOfficialOverrideIsStale
-          ? activeHwTypes
-          : (localOfficialOverride ?? activeHwTypes);
+      const source = localOfficialOverride ?? activeHwTypes;
       return new Set([...source].filter((key) => hwTypesWithData.has(key)));
-    }, [
-      activeHwTypes,
-      hwTypesWithData,
-      localOfficialOverride,
-      localOfficialOverrideIsStale,
-      overlayScopeChanged,
-    ]);
+    }, [activeHwTypes, hwTypesWithData, localOfficialOverride]);
     const rawOverlayHwTypes = useMemo(
       () =>
-        overlayScopeChanged
-          ? scopedOverlayHwTypes
-          : new Set(
-              [...providerActiveOverlayHwTypes].filter((key) => scopedOverlayHwTypes.has(key)),
-            ),
-      [overlayScopeChanged, providerActiveOverlayHwTypes, scopedOverlayHwTypes],
+        new Set([...providerActiveOverlayHwTypes].filter((key) => scopedOverlayHwTypes.has(key))),
+      [providerActiveOverlayHwTypes, scopedOverlayHwTypes],
     );
     const allUnifiedHwTypes = useMemo(() => {
       const all = new Set(hwTypesWithData);
@@ -589,9 +514,6 @@ const ScatterGraph = React.memo(
       rawOverlayHwTypes.forEach((key) => combined.add(`overlay:${key}`));
       return combined;
     }, [rawOfficialHwTypes, rawOverlayHwTypes]);
-    // Preview mode is diagnostic: official and unofficial runs may use different
-    // engines, and all of them must remain comparable on the same graph. Keep the
-    // production cross-engine guard only when no unofficial overlay is present.
     const resolvedUnifiedSelection = useMemo(
       () =>
         overlayData
@@ -612,105 +534,108 @@ const ScatterGraph = React.memo(
       return { official, overlay };
     }, [resolvedUnifiedSelection]);
     const effectiveOfficialHwTypes = resolvedHwTypes.official;
-    // Official-only toggles must not rebuild the D3 overlay layer. Preserve the
-    // overlay Set identity when its contents did not change.
     const activeOverlayHwTypesRef = useRef(resolvedHwTypes.overlay);
     if (!setsEqual(activeOverlayHwTypesRef.current, resolvedHwTypes.overlay)) {
       activeOverlayHwTypesRef.current = resolvedHwTypes.overlay;
     }
     const activeOverlayHwTypes = activeOverlayHwTypesRef.current;
-    const mergeScopedOverlaySelection = useCallback(
-      (scopedSelection: Set<string>) => {
-        const merged = new Set(providerActiveOverlayHwTypes);
-        scopedOverlayHwTypes.forEach((key) => merged.delete(key));
-        scopedSelection.forEach((key) => merged.add(key));
-        return merged;
-      },
-      [providerActiveOverlayHwTypes, scopedOverlayHwTypes],
-    );
-
-    useEffect(() => {
-      if (!overlayData) return;
-      // Keep the scope pending while the official query is between scopes.
-      // ChartDisplay uses the same readiness rule before seeding the full
-      // mixed-engine override, so a rerender from overlay reconciliation cannot
-      // fall back to the production-filtered active set in the meantime.
-      if (!overlayScopeChanged || !loading || hwTypesWithData.size > 0) {
-        previousOverlayScopeRef.current = overlayScopeKey;
-      }
-      // ChartDisplay seeds the new preview scope with every eligible official
-      // series. Avoid replacing it with the production-filtered active set while
-      // that parent update is being committed.
-      if (localOfficialOverrideIsStale && !overlayScopeChanged) {
-        setLocalOfficialOverride(null);
-      } else if (
-        !overlayScopeChanged &&
-        localOfficialOverride !== null &&
-        !setsEqual(localOfficialOverride, effectiveOfficialHwTypes)
-      ) {
-        setLocalOfficialOverride(effectiveOfficialHwTypes);
-      }
-      const mergedOverlaySelection = mergeScopedOverlaySelection(activeOverlayHwTypes);
-      if (!setsEqual(providerActiveOverlayHwTypes, mergedOverlaySelection)) {
-        setActiveOverlayHwTypes(mergedOverlaySelection);
-      }
-    }, [
-      overlayData,
-      overlayScopeKey,
-      overlayScopeChanged,
-      loading,
-      hwTypesWithData,
-      localOfficialOverride,
-      localOfficialOverrideIsStale,
-      effectiveOfficialHwTypes,
-      activeOverlayHwTypes,
-      providerActiveOverlayHwTypes,
-      setLocalOfficialOverride,
-      setActiveOverlayHwTypes,
-      mergeScopedOverlaySelection,
-    ]);
 
     const commitUnifiedSelection = useCallback(
       (selection: Set<string>) => {
         const official = new Set<string>();
-        const overlay = new Set<string>();
+        const scopedOverlay = new Set<string>();
         for (const key of selection) {
-          if (key.startsWith('overlay:')) overlay.add(key.slice('overlay:'.length));
+          if (key.startsWith('overlay:')) scopedOverlay.add(key.slice('overlay:'.length));
           else official.add(key);
         }
-        setLocalOfficialOverride(official);
-        setActiveOverlayHwTypes(mergeScopedOverlaySelection(overlay));
+        const overlay = new Set(providerActiveOverlayHwTypes);
+        scopedOverlayHwTypes.forEach((key) => overlay.delete(key));
+        scopedOverlay.forEach((key) => overlay.add(key));
+        setUnifiedOverlaySelection(official, overlay);
       },
-      [setLocalOfficialOverride, setActiveOverlayHwTypes, mergeScopedOverlaySelection],
+      [providerActiveOverlayHwTypes, scopedOverlayHwTypes, setUnifiedOverlaySelection],
     );
-    const unifiedToggle = useCallback(
-      (key: string, isOverlay: boolean) => {
-        const prefixedKey = isOverlay ? `overlay:${key}` : key;
-        commitUnifiedSelection(
-          computeToggle(resolvedUnifiedSelection, prefixedKey, allUnifiedHwTypes),
+    const toggleOfficialHwType = useCallback(
+      (key: string) => {
+        // The unofficial run is a pinned comparison, not an official legend item.
+        // Soloing an official series must therefore leave every overlay selection unchanged.
+        setUnifiedOverlaySelection(
+          computeToggle(effectiveOfficialHwTypes, key, hwTypesWithData),
+          providerActiveOverlayHwTypes,
         );
       },
-      [resolvedUnifiedSelection, allUnifiedHwTypes, commitUnifiedSelection],
+      [
+        effectiveOfficialHwTypes,
+        hwTypesWithData,
+        providerActiveOverlayHwTypes,
+        setUnifiedOverlaySelection,
+      ],
     );
     const resetUnifiedSelection = useCallback(() => {
       selectAllHwTypes();
       if (!overlayData) {
-        setLocalOfficialOverride(null);
+        resetOverlaySelection();
         return;
       }
       commitUnifiedSelection(allUnifiedHwTypes);
     }, [
       selectAllHwTypes,
       overlayData,
-      setLocalOfficialOverride,
+      resetOverlaySelection,
       allUnifiedHwTypes,
       commitUnifiedSelection,
     ]);
 
     // When no overlay data, delegate to context's toggleHwType (preserves setActivePresetId)
     const handleToggleHwType = useCallback(
-      (key: string) => (overlayData ? unifiedToggle(key, false) : toggleHwType(key)),
-      [overlayData, unifiedToggle, toggleHwType],
+      (key: string) => {
+        if (!overlayData) {
+          toggleHwType(key);
+          return;
+        }
+        setBestPerSku(false, { applySelection: false });
+        toggleOfficialHwType(key);
+      },
+      [overlayData, setBestPerSku, toggleOfficialHwType, toggleHwType],
+    );
+
+    // Best per SKU lives in the Quick Filters dialog, but the handler stays
+    // here: overlay mode owns a temporary unified selection that must be
+    // recomputed (or restored) when the automatic mode flips.
+    const handleBestPerSkuChange = useCallback(
+      (checked: boolean) => {
+        setBestPerSku(checked);
+        if (overlayData) {
+          if (checked) {
+            const direction =
+              chartDefinition[`${selectedYAxisMetric}_roofline` as keyof ChartDefinition];
+            if (
+              direction === 'upper_right' ||
+              direction === 'upper_left' ||
+              direction === 'lower_left' ||
+              direction === 'lower_right'
+            ) {
+              const selection = bestSeriesPerSku(data, direction);
+              for (const key of bestSeriesPerSku(overlayData.data, direction)) {
+                selection.add(`overlay:${key}`);
+              }
+              commitUnifiedSelection(selection);
+            }
+          } else {
+            resetUnifiedSelection();
+          }
+        }
+        track('inference_best_per_sku_toggled', { enabled: checked });
+      },
+      [
+        setBestPerSku,
+        overlayData,
+        chartDefinition,
+        selectedYAxisMetric,
+        data,
+        commitUnifiedSelection,
+        resetUnifiedSelection,
+      ],
     );
 
     // Legend "X" (remove) — same overlay split as handleToggleHwType. With an
@@ -726,11 +651,12 @@ const ScatterGraph = React.memo(
           removeHwType(key);
           return;
         }
+        setBestPerSku(false, { applySelection: false });
         const next = new Set(resolvedUnifiedSelection);
         next.delete(key);
         commitUnifiedSelection(next);
       },
-      [overlayData, removeHwType, resolvedUnifiedSelection, commitUnifiedSelection],
+      [overlayData, setBestPerSku, removeHwType, resolvedUnifiedSelection, commitUnifiedSelection],
     );
 
     // --- Theme ---
@@ -841,7 +767,24 @@ const ScatterGraph = React.memo(
       return effectiveOfficialHwTypes;
     }, [showAllHardwareTypes, groupedData, effectiveOfficialHwTypes]);
 
-    const buildPointConfigId = useCallback(scatterPointConfigId, []);
+    const distinguishPointDates = useMemo(() => {
+      let firstDate: string | undefined;
+      let hasFirstPoint = false;
+      for (const point of data) {
+        const date = point.date || undefined;
+        if (!hasFirstPoint) {
+          firstDate = date;
+          hasFirstPoint = true;
+        } else if (date !== firstDate) {
+          return true;
+        }
+      }
+      return false;
+    }, [data]);
+    const buildPointId = useCallback(
+      (point: InferenceData) => scatterPointJoinId(point, distinguishPointDates),
+      [distinguishPointDates],
+    );
 
     // filteredData: visible points only (for scale domain calculation)
     const filteredData = useMemo(
@@ -993,7 +936,10 @@ const ScatterGraph = React.memo(
         activeOverlayHwTypes.has(p.hwKey as string),
       );
       const visiblePoints = [...filteredData, ...visibleOverlayPoints];
-      return matchKnownConfigIssues(modelLabel, visiblePoints).map((issue) => ({
+      const annotations: KnownIssueAnnotation[] = matchKnownConfigIssues(
+        modelLabel,
+        visiblePoints,
+      ).map((issue) => ({
         issue,
         label: parseHwKeyToLabel(issue.hwKey, modelLabel).label,
         color: getCssColor(resolveColor(issue.hwKey)),
@@ -1001,6 +947,30 @@ const ScatterGraph = React.memo(
           .filter((p) => pointMatchesIssue(issue, p))
           .map((p) => ({ x: p.x, y: p.y })),
       }));
+      // Official-preview notices intentionally follow only official data. An
+      // unofficial overlay is not an InferenceX publication. `filteredData`
+      // has already applied token-metric support, so the July Rubin notice is
+      // present only on output-token charts alongside its visible curve.
+      for (const previewConfig of OFFICIAL_PREVIEW_SERIES) {
+        const previewPoints = filteredData.filter((point) =>
+          hardwareKeyMatchesAnyBase(String(point.hwKey), previewConfig.baseGpuKeys),
+        );
+        if (previewPoints.length === 0) continue;
+
+        const hwKey = String(previewPoints[0]!.hwKey);
+        const previewCopy = previewConfig.strings[locale];
+        annotations.push({
+          preview: {
+            id: previewConfig.id,
+            summary: previewCopy.title,
+            detail: previewCopy.chartDetail,
+          },
+          label: getDisplayLabel(getHardwareConfig(hwKey, modelLabel)),
+          color: getCssColor(resolveColor(hwKey)),
+          points: previewPoints.map((point) => ({ x: point.x, y: point.y })),
+        });
+      }
+      return annotations;
     }, [
       modelLabel,
       filteredData,
@@ -1008,6 +978,7 @@ const ScatterGraph = React.memo(
       activeOverlayHwTypes,
       resolveColor,
       getCssColor,
+      locale,
     ]);
 
     const overlayRooflines = useMemo(() => {
@@ -1069,12 +1040,6 @@ const ScatterGraph = React.memo(
 
     // All official points for rendering (unfiltered — visibility via opacity)
     const pointsData = useMemo(() => Object.values(groupedData).flat(), [groupedData]);
-    const hasOffloadHalo = useMemo(
-      () =>
-        pointsData.some((point) => point.offload_mode === 'on') ||
-        processedOverlayData.some((point) => point.offload_mode === 'on'),
-      [pointsData, processedOverlayData],
-    );
     // Bulk presence lookup for agentic points: which ids have a stored
     // trace_replay blob → controls the "View charts" button in the pinned
     // tooltip. We deliberately don't fetch the histograms themselves here;
@@ -1083,14 +1048,41 @@ const ScatterGraph = React.memo(
     const agenticIds = useMemo(() => {
       const ids: number[] = [];
       for (const p of pointsData) {
-        if (p.benchmark_type === 'agentic_traces' && typeof p.id === 'number') ids.push(p.id);
+        if (p.benchmark_type === 'agentic_traces' && isPersistedBenchmarkId(p.id)) ids.push(p.id);
       }
       return ids;
     }, [pointsData]);
-    const { data: traceAvailability } = useTraceAvailability(agenticIds);
+    const { data: traceAvailability, isPending: isTraceAvailabilityPending } =
+      useTraceAvailability(agenticIds);
+
+    // Log availability applies to every persisted official point, including
+    // fixed-sequence runs.
+    const persistedPointIds = useMemo(
+      () => pointsData.flatMap((point) => (isPersistedBenchmarkId(point.id) ? [point.id] : [])),
+      [pointsData],
+    );
+    const { data: persistedLogAvailability } = useLogAvailability(persistedPointIds);
+    const [fixedLogPointId, setFixedLogPointId] = useState<number | null>(null);
 
     // --- Legend points table (per-series drill-down opened from the legend) ---
     const [pointsTableTarget, setPointsTableTarget] = useState<LegendPointsTarget | null>(null);
+    const [quickFiltersOpen, setQuickFiltersOpen] = useState(false);
+    const quickFilterCount =
+      quickFilters.vendors.length +
+      quickFilters.frameworks.length +
+      quickFilters.deployment.length +
+      (selectedSequence === Sequence.AgenticTraces ? 0 : quickFilters.spec.length);
+    const clearQuickFilters = useCallback(() => {
+      setQuickFilterVendors([]);
+      setQuickFilterFrameworks([]);
+      setQuickFilterDeployment([]);
+      setQuickFilterSpec([]);
+    }, [
+      setQuickFilterVendors,
+      setQuickFilterFrameworks,
+      setQuickFilterDeployment,
+      setQuickFilterSpec,
+    ]);
 
     const pointsTable = useMemo(() => {
       if (!pointsTableTarget) return null;
@@ -1110,7 +1102,7 @@ const ScatterGraph = React.memo(
           title: hwConfig ? getDisplayLabel(hwConfig) : hwKey,
           color: resolveColor(hwKey),
           isOverlay: false,
-          rows: buildLegendPointsRows(pts, false),
+          rows: buildLegendPointsRows(pts, false, locale),
         };
       }
       const { runIndex, runId, branch } = pointsTableTarget;
@@ -1127,7 +1119,7 @@ const ScatterGraph = React.memo(
         title: `✕ ${branch}`,
         color: overlayRunColor(runIndex),
         isOverlay: true,
-        rows: buildLegendPointsRows(pts, true),
+        rows: buildLegendPointsRows(pts, true, locale),
       };
     }, [
       pointsTableTarget,
@@ -1141,6 +1133,7 @@ const ScatterGraph = React.memo(
       processedOverlayData,
       runIndexByUrl,
       activeOverlayHwTypes,
+      locale,
     ]);
 
     // Gradient label data
@@ -1247,6 +1240,47 @@ const ScatterGraph = React.memo(
     }, [visiblePoints, isInputTputMetric, logScale, niceAxes, yExtentOverride]);
     const yScaleConfig = useStableValue(yScaleConfigRaw, isSameScaleConfig);
 
+    const dataIdentity = useMemo(
+      () =>
+        [
+          ...pointsData.map((point) => `official:${buildPointId(point)}`),
+          ...processedOverlayData.map((point) => `overlay:${scatterPointConfigId(point)}`),
+        ]
+          .toSorted()
+          .join('|'),
+      [pointsData, processedOverlayData, buildPointId],
+    );
+    const metricIdentity = useMemo(
+      () =>
+        [
+          useAdvancedLabels ? 'advanced-labels' : 'basic-labels',
+          showConcurrencyLabels ? 'conc-labels' : 'no-conc-labels',
+          selectedYAxisMetric,
+          `${xScaleConfig.type}:${xScaleConfig.domain.join(',')}`,
+          `${yScaleConfig.type}:${yScaleConfig.domain.join(',')}`,
+          ...pointsData.map(
+            (point) =>
+              `${buildPointId(point)}:${point.date ?? ''}:${point.id ?? ''}:${point.run_url ?? ''}:${point.x}:${point.y}`,
+          ),
+          ...processedOverlayData.map(
+            (point) =>
+              `overlay:${scatterPointConfigId(point)}:${point.date ?? ''}:${point.id ?? ''}:${point.run_url ?? ''}:${point.x}:${point.y}`,
+          ),
+        ]
+          .toSorted()
+          .join('|'),
+      [
+        selectedYAxisMetric,
+        useAdvancedLabels,
+        showConcurrencyLabels,
+        xScaleConfig,
+        yScaleConfig,
+        pointsData,
+        processedOverlayData,
+        buildPointId,
+      ],
+    );
+
     // --- Axis configs ---
     const xAxisConfig = useMemo(
       () => ({
@@ -1310,6 +1344,8 @@ const ScatterGraph = React.memo(
       getCssColor,
       resolveColor,
       knownIssueAnnotations,
+      traceAvailability,
+      logAvailability: persistedLogAvailability,
     });
     interactionRef.current = {
       isPointVisible,
@@ -1320,11 +1356,369 @@ const ScatterGraph = React.memo(
       getCssColor,
       resolveColor,
       knownIssueAnnotations,
+      traceAvailability,
+      logAvailability: persistedLogAvailability,
     };
 
+    // --- Perf ruler (opt-in: click two curves, drag the ruler to any iso-x) ---
+    // Curve-to-curve ISO-X semantics: each measurement is two CURVES
+    // (rendered roofline path class tokens) plus a freely chosen iso-x
+    // stored in DATA space (xScale.invert of the click), so measurements
+    // survive zoom and metric changes. BOTH ruler ends are interpolated on
+    // the curves' rendered paths at the iso-x — neither end needs to be a
+    // data point. Multiple rulers accumulate (capped in the pure module);
+    // completing one immediately allows starting the next.
+    const [perfRulerMode, setPerfRulerMode] = useState(false);
+    const [perfRulerState, setPerfRulerState] = useState<PerfRulerState>(EMPTY_PERF_RULER_STATE);
+    // Draw passes read mode/state through refs so toggling off clears the
+    // rulers in the same pre-paint layout pass — lines/labels must never
+    // linger a frame after the switch flips (Bugbot report on PR #853).
+    const perfRulerModeRef = useRef(perfRulerMode);
+    perfRulerModeRef.current = perfRulerMode;
+    const perfRulerStateRef = useRef(perfRulerState);
+    perfRulerStateRef.current = perfRulerState;
+    // Live per-ruler iso-x overrides for draw passes. Dragging a ruler
+    // updates its entry once per animation frame WITHOUT touching React
+    // state (the drag commits on end), so the ruler follows the pointer
+    // with one redraw per frame. Cleared when committed state changes (an
+    // effect, not per render, so an unrelated re-render mid-drag cannot
+    // snap the dragged ruler back).
+    const perfRulerLiveIsoXRef = useRef(new Map<number, number>());
+    useLayoutEffect(() => {
+      perfRulerLiveIsoXRef.current.clear();
+    }, [perfRulerState]);
+
+    // Scales/group from the most recent draw pass — click and drag handlers
+    // convert between pixel and data space with the exact scales the chart
+    // is currently drawn with.
+    const perfRulerDrawCtxRef = useRef<{
+      zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>;
+      xScale: ContinuousScale;
+      yScale: ContinuousScale;
+      width: number;
+      height: number;
+    } | null>(null);
+    // Forward ref: the drag frame needs to redraw, but drawPerfRuler is
+    // defined below (it also attaches the drag behavior — benign cycle).
+    const drawPerfRulerRef = useRef<
+      | ((
+          zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+          xScale: ContinuousScale,
+          yScale: ContinuousScale,
+          width: number,
+          height: number,
+        ) => void)
+      | null
+    >(null);
+
+    // Completion clamp for nextPerfRulerState: the stored iso-x must lie
+    // inside the curve pair's overlapping x range, or the completed ruler
+    // could render nowhere — no line and no drag handle to recover it.
+    // Data → pixel through the exact scales/paths the chart is currently
+    // drawn with, clamp, and back. Null (missing paths or disjoint spans)
+    // rejects the measurement and keeps the draft anchored.
+    const clampPerfRulerIsoXToOverlap = useCallback(
+      (curveA: string, curveB: string, isoX: number): number | null => {
+        const ctx = perfRulerDrawCtxRef.current;
+        if (!ctx) return null;
+        const [nodeA, nodeB] = [curveA, curveB].map((cls) =>
+          ctx.zoomGroup.select<SVGPathElement>(`.${CSS.escape(cls)}`).node(),
+        );
+        if (!nodeA || !nodeB || typeof nodeA.getPointAtLength !== 'function') return null;
+        const extentA = pathXExtent(nodeA);
+        const extentB = pathXExtent(nodeB);
+        if (!extentA || !extentB) return null;
+        const clamped = clampIsoX(Number(ctx.xScale(isoX)), extentA, extentB);
+        return clamped === null ? null : Number(ctx.xScale.invert(clamped));
+      },
+      [],
+    );
+
+    // Curve click (widened hit strokes): iso-x is the click's x pixel
+    // through the CURRENT rendered x scale, stored in data space.
+    const handlePerfRulerCurveClick = useCallback(
+      (curve: string, pixelX: number) => {
+        const ctx = perfRulerDrawCtxRef.current;
+        if (!ctx) return;
+        track('latency_perf_ruler_curve_clicked', { curve });
+        const isoX = Number(ctx.xScale.invert(pixelX));
+        setPerfRulerState((prev) =>
+          nextPerfRulerState(prev, { curve, isoX }, clampPerfRulerIsoXToOverlap),
+        );
+        chartRef.current?.dismissTooltip();
+        chartRef.current?.hideTooltip();
+      },
+      [clampPerfRulerIsoXToOverlap],
+    );
+    const perfRulerCurveClickRef = useRef(handlePerfRulerCurveClick);
+    perfRulerCurveClickRef.current = handlePerfRulerCurveClick;
+
+    // Points sit on curves: a ruler-mode click on a data point behaves like
+    // clicking the point's curve at that point's x. Candidates cover the
+    // single- vs multi-date roofline class variants; the first one present
+    // in the DOM wins. Ruler-mode clicks measure INSTEAD of pinning the
+    // tooltip, so drop the pin the shared click handler applied just before
+    // this callback ran.
+    const handlePerfRulerPointClick = useCallback(
+      (point: InferenceData, source: 'official' | 'overlay') => {
+        const ctx = perfRulerDrawCtxRef.current;
+        if (!ctx) return;
+        const base = `${String(point.hwKey)}_${point.precision}`;
+        const candidates =
+          source === 'overlay'
+            ? [
+                `overlay-roofline-${base}_run${overlayRunIndex(point.run_url ?? null, runIndexByUrl)}`,
+              ]
+            : [`roofline-${base}`, `roofline-${base}__${point.date}`];
+        const curve = candidates.find(
+          (cls) => !ctx.zoomGroup.select(`.${CSS.escape(cls)}`).empty(),
+        );
+        // Single-point series render no roofline path — nothing to measure.
+        if (!curve) return;
+        setPerfRulerState((prev) =>
+          nextPerfRulerState(prev, { curve, isoX: point.x }, clampPerfRulerIsoXToOverlap),
+        );
+        chartRef.current?.dismissTooltip();
+        chartRef.current?.hideTooltip();
+      },
+      [runIndexByUrl, clampPerfRulerIsoXToOverlap],
+    );
+
+    // Read by long-lived D3 click closures (official tooltip config + overlay
+    // marker handlers) — same refs-over-closures rule as interactionRef.
+    const perfRulerRef = useRef({ mode: perfRulerMode, onPointClick: handlePerfRulerPointClick });
+    perfRulerRef.current = { mode: perfRulerMode, onPointClick: handlePerfRulerPointClick };
+
+    // Turning the toggle off clears ALL rulers and any in-progress
+    // selection (documented in the info tooltip — simple and predictable;
+    // the switch handler also clears synchronously, this covers
+    // programmatic mode changes). `clearPerfRulers` bails out with the same
+    // reference when there is nothing to clear.
+    useEffect(() => {
+      if (!perfRulerMode) setPerfRulerState(clearPerfRulers);
+    }, [perfRulerMode]);
+
+    // Invisible widened hit strokes over every rendered roofline path
+    // (official AND overlay) make the curves themselves clickable in ruler
+    // mode. The hit layer sits directly ABOVE `.rooflines-layer` (so the
+    // visible 2.5px strokes never shadow the hit strokes) but BELOW the
+    // dot-groups and overlay markers, so point hover/click behavior is
+    // untouched. The layer only exists while ruler mode is on — with the
+    // mode off there is nothing to intercept hovers, clicks, or zoom.
+    const syncPerfRulerHitPaths = useCallback(
+      (zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>) => {
+        let hitLayer = zoomGroup.select<SVGGElement>('.perf-ruler-hits');
+        const rooflinesLayerNode = zoomGroup.select<SVGGElement>('.rooflines-layer').node();
+        if (!perfRulerModeRef.current || !rooflinesLayerNode) {
+          hitLayer.remove();
+          return;
+        }
+        if (hitLayer.empty()) {
+          const node = document.createElementNS('http://www.w3.org/2000/svg', 'g');
+          node.setAttribute('class', 'perf-ruler-hits');
+          rooflinesLayerNode.after(node);
+          hitLayer = d3.select(node) as typeof hitLayer;
+        }
+        interface HitEntry {
+          curve: string;
+          d: string;
+        }
+        const entries: HitEntry[] = [];
+        zoomGroup
+          .selectAll<SVGPathElement, unknown>('.roofline-path, .overlay-roofline-path')
+          .each(function () {
+            // The identity token is the curve-specific class, e.g.
+            // `roofline-H100_fp8`, `roofline-H100_fp8__2026-01-01`, or
+            // `overlay-roofline-H100_fp8_run0`.
+            const curve = [...this.classList].find(
+              (cls) => cls !== 'roofline-path' && cls !== 'overlay-roofline-path',
+            );
+            const d = this.getAttribute('d');
+            // Curves hidden by legend/precision filters keep their paths at
+            // opacity 0 — invisible curves must not be clickable.
+            if (!curve || !d || !isPerfRulerCurveVisible(this.style.opacity)) return;
+            entries.push({ curve, d });
+          });
+        const selected = perfRulerCurveSet(perfRulerStateRef.current);
+        hitLayer
+          .selectAll<SVGPathElement, HitEntry>('.perf-ruler-hit')
+          .data(entries, (e) => e.curve)
+          .join('path')
+          .attr('class', 'perf-ruler-hit')
+          .attr('fill', 'none')
+          .attr('d', (e) => e.d)
+          .attr('stroke', 'var(--primary)')
+          // Selected curves get a faint halo as feedback; unselected hit
+          // strokes are fully transparent (`pointer-events: stroke` still
+          // hit-tests the invisible stroke geometry).
+          .attr('stroke-opacity', (e) => (selected.has(e.curve) ? 0.18 : 0))
+          .attr('stroke-width', 13)
+          .style('pointer-events', 'stroke')
+          .style('cursor', 'crosshair')
+          .on('click', (event: MouseEvent, e: HitEntry) => {
+            event.stopPropagation();
+            const [pixelX] = d3.pointer(event, zoomGroup.node());
+            perfRulerCurveClickRef.current(e.curve, pixelX);
+          });
+      },
+      [],
+    );
+
+    // Horizontal drag on a ruler line — the primary way to fine-tune that
+    // ruler's iso-x. The drag handle's datum (set by the render join)
+    // identifies WHICH ruler is dragged. rAF-throttled: each frame clamps
+    // the pointer x to that ruler's curve pair's overlapping x range,
+    // updates its live iso-x, and redraws; the value commits to React state
+    // on drag end. d3.drag stops mousedown propagation itself, so dragging
+    // a ruler never pans the chart.
+    const perfRulerDragTargetRef = useRef<{ id: number; pixelX: number } | null>(null);
+    const perfRulerDragFrameRef = useRef<number | null>(null);
+    const applyPerfRulerDragFrame = useCallback(() => {
+      const ctx = perfRulerDrawCtxRef.current;
+      const target = perfRulerDragTargetRef.current;
+      if (!ctx || !target) return;
+      const ruler = perfRulerStateRef.current.rulers.find((r) => r.id === target.id);
+      if (!ruler) return;
+      const [nodeA, nodeB] = [ruler.curveA, ruler.curveB].map((cls) =>
+        ctx.zoomGroup.select<SVGPathElement>(`.${CSS.escape(cls)}`).node(),
+      );
+      if (!nodeA || !nodeB || typeof nodeA.getPointAtLength !== 'function') return;
+      const extentA = pathXExtent(nodeA);
+      const extentB = pathXExtent(nodeB);
+      if (!extentA || !extentB) return;
+      const clamped = clampIsoX(target.pixelX, extentA, extentB);
+      if (clamped === null) return;
+      perfRulerLiveIsoXRef.current.set(target.id, Number(ctx.xScale.invert(clamped)));
+      drawPerfRulerRef.current?.(ctx.zoomGroup, ctx.xScale, ctx.yScale, ctx.width, ctx.height);
+    }, []);
+    const applyPerfRulerDragFrameRef = useRef(applyPerfRulerDragFrame);
+    applyPerfRulerDragFrameRef.current = applyPerfRulerDragFrame;
+
+    const perfRulerDrag = useMemo(
+      () =>
+        d3
+          .drag<SVGLineElement, PerfRulerRenderEntry>()
+          .on(
+            'drag',
+            (
+              event: d3.D3DragEvent<SVGLineElement, PerfRulerRenderEntry, unknown>,
+              entry: PerfRulerRenderEntry,
+            ) => {
+              perfRulerDragTargetRef.current = { id: entry.id, pixelX: event.x };
+              if (perfRulerDragFrameRef.current === null) {
+                perfRulerDragFrameRef.current = requestAnimationFrame(() => {
+                  perfRulerDragFrameRef.current = null;
+                  applyPerfRulerDragFrameRef.current();
+                });
+              }
+            },
+          )
+          .on('end', () => {
+            // Flush the pending frame, then commit this ruler's iso-x.
+            if (perfRulerDragFrameRef.current !== null) {
+              cancelAnimationFrame(perfRulerDragFrameRef.current);
+              perfRulerDragFrameRef.current = null;
+              applyPerfRulerDragFrameRef.current();
+            }
+            const target = perfRulerDragTargetRef.current;
+            perfRulerDragTargetRef.current = null;
+            if (!target) return;
+            const isoX = perfRulerLiveIsoXRef.current.get(target.id);
+            if (isoX === undefined) return;
+            setPerfRulerState((prev) => movePerfRulerIsoX(prev, target.id, isoX));
+          }),
+      [],
+    );
+
+    // Draw (or clear) the hit strokes and all rulers for the current draw
+    // pass. Stable callback so the custom layer can live outside the
+    // perf-ruler state dependencies — render/zoom passes read mode, rulers,
+    // and live iso-x values through refs. Gated on the mode ref so a
+    // toggle-off clears everything in the very next pre-paint pass.
+    const drawPerfRuler = useCallback(
+      (
+        zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+        xScale: ContinuousScale,
+        yScale: ContinuousScale,
+        width: number,
+        height: number,
+      ) => {
+        perfRulerDrawCtxRef.current = { zoomGroup, xScale, yScale, width, height };
+        syncPerfRulerHitPaths(zoomGroup);
+        const state = perfRulerStateRef.current;
+        const entries: PerfRulerRenderEntry[] = [];
+        if (perfRulerModeRef.current && state.rulers.length > 0) {
+          // Both ends of each ruler are interpolated on the RENDERED paths —
+          // the drawn curve is the measurement surface, so the ruler matches
+          // it exactly and stays correct under zoom (rooflines redraw before
+          // this layer runs). When either of a ruler's curves does not span
+          // its iso-x, has no rendered path, or is hidden (legend/precision
+          // toggles keep the path in the DOM at opacity 0), that ruler draws
+          // nothing this frame but KEEPS its state — per ruler, others stay
+          // visible; zoom/drag/legend re-toggles that restore both visible
+          // intersections bring it back.
+          const geometries: (PerfRulerGeometry | null)[] = state.rulers.map((ruler) => {
+            const isoX = perfRulerLiveIsoXRef.current.get(ruler.id) ?? ruler.isoX;
+            const pixelX = xScale(isoX);
+            const ends: PerfRulerEndInput[] = [];
+            for (const cls of [ruler.curveA, ruler.curveB]) {
+              const node = zoomGroup.select<SVGPathElement>(`.${CSS.escape(cls)}`).node();
+              if (!node || typeof node.getPointAtLength !== 'function') break;
+              if (!isPerfRulerCurveVisible(node.style.opacity)) break;
+              const hit = intersectPathAtX(node, pixelX);
+              if (!hit) break;
+              ends.push({ py: hit.y, rawY: yScale.invert(hit.y) });
+            }
+            return ends.length === 2 ? computeIsoXRulerGeometry(pixelX, ends[0], ends[1]) : null;
+          });
+          // Lay out all labels together so overlapping labels nudge apart.
+          const layouts = computePerfRulerLabelLayouts(geometries, {
+            chartWidth: width,
+            chartHeight: height,
+          });
+          for (const [index, ruler] of state.rulers.entries()) {
+            const geometry = geometries[index];
+            const layout = layouts[index];
+            if (geometry && layout) entries.push({ id: ruler.id, geometry, layout });
+          }
+        }
+        renderPerfRulers(zoomGroup, entries, {
+          color: 'var(--primary)',
+          halo: 'var(--background)',
+          onDelete: (id) => {
+            track('latency_perf_ruler_deleted');
+            setPerfRulerState((prev) => deletePerfRuler(prev, id));
+          },
+        });
+        // (Re)attach the horizontal drag behavior to the (possibly fresh)
+        // drag handles the render pass just joined — their datum (the
+        // render entry) tells the drag which ruler it moves.
+        const dragHandles = zoomGroup.selectAll<SVGLineElement, PerfRulerRenderEntry>(
+          '.perf-ruler .pr-drag',
+        );
+        if (!dragHandles.empty()) dragHandles.call(perfRulerDrag);
+      },
+      [syncPerfRulerHitPaths, perfRulerDrag],
+    );
+    drawPerfRulerRef.current = drawPerfRuler;
+
     // Render context from the last D3 render — lets the decoration effect
     // restyle with the same layout/scales the chart was drawn with.
     const lastRenderCtxRef = useRef<RenderContext | null>(null);
+    const labelDisplayRef = useRef({ showPointLabels, showGradientLabels, showLineLabels });
+    labelDisplayRef.current = { showPointLabels, showGradientLabels, showLineLabels };
+    const pointLabelsVisible = showPointLabels && !showGradientLabels;
+    const lastPointLabelsVisibleRef = useRef(pointLabelsVisible);
+    const lastShowLineLabelsRef = useRef(showLineLabels);
+    const lastShowGradientLabelsRef = useRef(showGradientLabels);
+
+    const getDisplaySelection = useCallback(() => {
+      const svg = chartRef.current?.getSvgElement?.();
+      const ctx = lastRenderCtxRef.current;
+      if (!svg || !ctx) return null;
+      const zoomGroup = d3.select(svg).select<SVGGElement>('.zoom-group');
+      return zoomGroup.empty() ? null : { svg, ctx, zoomGroup };
+    }, []);
 
     // Hover dimming animates via the inline `transition: opacity 150ms ease`
     // the render path puts on dots, rooflines, and labels — a single style
@@ -1421,10 +1815,19 @@ const ScatterGraph = React.memo(
               d3.axisLeft(newYS).ticks(10).tickFormat(logTickFormat(newYS)) as any,
             );
           }
-          avoidLabelCollisions(ctx.layout.zoomGroup);
+          if (showPointLabels && !showGradientLabels) {
+            avoidPointLabelCollisions(ctx.layout.zoomGroup);
+          }
         },
       }),
-      [zoomResetEventName, eventPrefix, xScaleConfig._isLog, yScaleConfig.type],
+      [
+        zoomResetEventName,
+        eventPrefix,
+        xScaleConfig._isLog,
+        yScaleConfig.type,
+        showPointLabels,
+        showGradientLabels,
+      ],
     );
 
     // --- Tooltip config ---
@@ -1440,7 +1843,13 @@ const ScatterGraph = React.memo(
             selectedYAxisMetric,
             hardwareConfig,
             runUrl: d.run_url ? updateRepoUrl(d.run_url) : undefined,
-            hasTrace: typeof d.id === 'number' ? traceAvailability?.[d.id] === true : false,
+            hasTrace:
+              d.benchmark_type === 'agentic_traces' && isPersistedBenchmarkId(d.id)
+                ? (interactionRef.current.traceAvailability?.[d.id] ?? false)
+                : false,
+            hasLog: isPersistedBenchmarkId(d.id)
+              ? (interactionRef.current.logAvailability?.[d.id] ?? false)
+              : false,
             locale,
           }),
         getRulerX: (d: InferenceData, xScale: any) => (xScale as ContinuousScale)(d.x),
@@ -1456,6 +1865,21 @@ const ScatterGraph = React.memo(
             getShapeKeyForPrecision(d.precision, interactionRef.current.selectedPrecisions),
           ),
         onPointClick: (d: InferenceData) => {
+          // Perf ruler mode: the click acts as a click on the point's CURVE
+          // at the point's x INSTEAD of pinning the tooltip (the handler
+          // drops the pin applied by the shared click path; the hover
+          // tooltip stays available).
+          const ruler = perfRulerRef.current;
+          if (ruler.mode) {
+            track('latency_data_point_clicked', {
+              hw: String(d.hwKey),
+              x: d.x,
+              y: d.y,
+              perfRuler: true,
+            });
+            ruler.onPointClick(d, 'official');
+            return;
+          }
           track('latency_data_point_clicked', { hw: String(d.hwKey), x: d.x, y: d.y });
           const tooltipEl = chartRef.current?.getTooltipElement();
           if (!tooltipEl) return;
@@ -1463,9 +1887,15 @@ const ScatterGraph = React.memo(
           // ── Summary-page actions ──────────────────────────────────────────
           // ── "View charts" real link (supports browser open-in-new-tab) ───
           const viewBtn = tooltipEl.querySelector('[data-action="view-charts"]');
-          if (viewBtn && typeof d.id === 'number') {
+          if (viewBtn && isPersistedBenchmarkId(d.id)) {
             viewBtn.addEventListener('click', (btnEvent) => {
               btnEvent.stopPropagation();
+              // Full-document navigation: stamp the chart state onto THIS
+              // history entry first, or Back returns to a bare /inference that
+              // rebuilds from defaults. Skipped in ephemeral scopes (/model
+              // embeds): the store holds the primary dashboard's state there,
+              // not this chart's.
+              if (!ephemeralUrlState) rememberChartStateInUrl();
               track('latency_view_charts_opened', {
                 id: d.id,
                 hwKey: String(d.hwKey),
@@ -1473,6 +1903,23 @@ const ScatterGraph = React.memo(
               });
             });
           }
+          const logsBtn = tooltipEl.querySelector('[data-action="view-logs"]');
+          if (logsBtn && typeof d.id === 'number') {
+            logsBtn.addEventListener('click', (btnEvent) => {
+              btnEvent.stopPropagation();
+              if (d.benchmark_type !== 'agentic_traces') {
+                btnEvent.preventDefault();
+                setFixedLogPointId(d.id!);
+                chartRef.current?.dismissTooltip();
+              }
+              track('latency_view_logs_opened', {
+                id: d.id,
+                hwKey: String(d.hwKey),
+                conc: d.conc,
+                benchmarkType: d.benchmark_type ?? 'single_turn',
+              });
+            });
+          }
         },
         attachToLayer: 1, // scatter layer is index 1 (after rooflines at 0)
       }),
@@ -1481,11 +1928,9 @@ const ScatterGraph = React.memo(
         yLabel,
         selectedYAxisMetric,
         hardwareConfig,
-        // selectedPrecisions is read via interactionRef.current in the hover
-        // handlers, so it isn't a dep. traceAvailability IS read directly in the
-        // tooltip content closure (the "View charts" button), so rebuild the
-        // config when the presence fetch resolves.
-        traceAvailability,
+        // selectedPrecisions and traceAvailability are read through
+        // interactionRef.current so long-lived D3 handlers always observe the
+        // latest state without turning tooltip-only data into chart identity.
         locale,
       ],
     );
@@ -1496,6 +1941,7 @@ const ScatterGraph = React.memo(
       const rooflineLayer: CustomLayerConfig = {
         type: 'custom',
         key: 'rooflines',
+        displayIdentity: `${showGradientLabels}:${showLineLabels}`,
         render: (zoomGroup, ctx) => {
           // Visibility / colors come from the interaction ref so this closure
           // stays correct between layer recreations (toggles restyle via the
@@ -1732,317 +2178,159 @@ const ScatterGraph = React.memo(
           }
 
           // ── Line labels (run name along each roofline) ──
-          interface LineLabel {
-            key: string;
-            hw: string;
-            label: string;
-            color: string;
-            x: number;
-            y: number;
-            visible: boolean;
-          }
-          const lineLabels: LineLabel[] = [];
-
+          let lineLabels: LineLabelPlacement[] = [];
           if (showLineLabels) {
-            const isInteractivity = chartDefinition.chartType === 'interactivity';
-            // With >1 precision selected each precision is its own curve, so label
-            // every curve and include the precision in the text.
             const multiPrecision = ir.selectedPrecisions.length > 1;
-            const LABEL_H = 18;
-            const LABEL_W = 120; // approximate label width for overlap check
-
-            if (isInteractivity) {
-              // Greedy placement: try top-left → midpoint → right-side → hide
-              const placed: { x: number; y: number }[] = [];
-
-              const collides = (cx: number, cy: number) =>
-                placed.some((p) => Math.abs(p.y - cy) < LABEL_H && Math.abs(p.x - cx) < LABEL_W);
-
-              // Deduplicate by group key — one label per curve. With a single
-              // precision that's one per hw; with multiple it's one per (hw,
-              // precision) so each precision curve keeps its own label.
-              const bestByGroup = new Map<string, (typeof entries)[0]>();
-              for (const e of entries) {
-                if (!e.visible) continue;
-                const groupKey = multiPrecision ? e.key : e.hw;
-                const prev = bestByGroup.get(groupKey);
-                if (!prev || e.points.length > prev.points.length) bestByGroup.set(groupKey, e);
+            const officialByGroup = new Map<string, (typeof entries)[number]>();
+            for (const entry of entries) {
+              if (!entry.visible) continue;
+              const groupKey = multiPrecision ? entry.key : entry.hw;
+              const previous = officialByGroup.get(groupKey);
+              if (!previous || entry.points.length > previous.points.length) {
+                officialByGroup.set(groupKey, entry);
               }
+            }
 
-              // Place one label per series. When pinned (replay), reuse a stored
-              // data-space anchor so the label tracks the same spot along its line
-              // as it animates; otherwise re-run greedy placement each render and
-              // hide on collision (the static chart's de-overlap behavior).
-              const anchors = lineLabelAnchorRef.current;
-              const placeLabel = (
-                key: string,
-                hw: string,
-                label: string,
-                color: string,
-                pts: InferenceData[],
-                keepVisibleOnCollision = false,
-              ) => {
-                const candidates = [
-                  pts[Math.min(1, pts.length - 1)], // near start
-                  pts[Math.floor(pts.length / 2)], // midpoint
-                  pts[Math.max(0, Math.floor((pts.length * 2) / 3))], // right-third
-                  pts.at(-1)!, // endpoint
-                ];
-                if (pinLineLabels) {
-                  let anchorX = anchors.get(key);
-                  if (anchorX === undefined) {
-                    // First sighting: pick the first non-colliding candidate
-                    // (endpoint as fallback) and remember its data-x for later
-                    // frames so the label no longer hops between candidates.
-                    let chosen = candidates.at(-1)!;
-                    for (const pt of candidates) {
-                      if (!collides(xScale(pt.x), yScale(pt.y))) {
-                        chosen = pt;
-                        break;
-                      }
-                    }
-                    anchorX = chosen.x;
-                    anchors.set(key, anchorX);
-                  }
-                  const pt = pointNearestX(pts, anchorX);
-                  const px = xScale(pt.x);
-                  const py = yScale(pt.y);
-                  placed.push({ x: px, y: py });
-                  // Stay visible across frames — positional stability is the goal
-                  // during animation, so we don't hide on transient collisions.
-                  lineLabels.push({ key, hw, label, color, x: px, y: py, visible: true });
-                  return;
-                }
-                for (const pt of candidates) {
-                  const px = xScale(pt.x);
-                  const py = yScale(pt.y);
-                  if (!collides(px, py)) {
-                    lineLabels.push({ key, hw, label, color, x: px, y: py, visible: true });
-                    placed.push({ x: px, y: py });
-                    return;
-                  }
-                }
-                // All candidates collide — hide this label.
-                const pt = pts[0];
-                const px = xScale(pt.x);
-                const py = yScale(pt.y);
-                lineLabels.push({
-                  key,
-                  hw,
+            const officialSeries: LineLabelSeries<InferenceData>[] = [
+              ...officialByGroup.values(),
+            ].map((entry) => ({
+              key: entry.key,
+              seriesId: entry.hw,
+              label: lineLabelText(entry.hw, entry.precision, multiPrecision, modelLabel),
+              color: ir.getCssColor(ir.resolveColor(entry.hw)),
+              points: entry.points,
+              keepVisibleOnCollision: entry.points.length === 1,
+            }));
+            const overlaySeries: LineLabelSeries<InferenceData>[] = Object.entries(
+              overlayRooflines,
+            ).flatMap(([overlayKey, group]) => {
+              if (!ir.activeOverlayHwTypes.has(group.hwKey)) return [];
+              const info = unofficialRunInfos[group.runIndex];
+              const precision = group.points[0]?.precision ?? '';
+              const label = info
+                ? multiPrecision
+                  ? `✕ ${info.branch || `run ${info.id}`} ${getPrecisionLabel(precision as Precision)}`
+                  : `✕ ${info.branch || `run ${info.id}`}`
+                : lineLabelText(group.hwKey, precision, multiPrecision, modelLabel);
+              return [
+                {
+                  key: `overlay-${overlayKey}`,
+                  seriesId: group.hwKey,
                   label,
-                  color,
-                  x: px,
-                  y: py,
-                  visible: keepVisibleOnCollision,
-                });
-                if (keepVisibleOnCollision) placed.push({ x: px, y: py });
-              };
-
-              // Sort entries by highest y-value first (top of chart) for priority
-              const sorted = [...bestByGroup.values()].toSorted((a, b) => {
-                const ay = yScale(a.points[0].y);
-                const by = yScale(b.points[0].y);
-                return ay - by; // smaller pixel y = higher on chart
-              });
-
-              for (const entry of sorted) {
-                placeLabel(
-                  entry.key,
-                  entry.hw,
-                  lineLabelText(entry.hw, entry.precision, multiPrecision, modelLabel),
-                  ir.getCssColor(ir.resolveColor(entry.hw)),
-                  entry.points,
-                  entry.points.length === 1,
-                );
-              }
-
-              // Also add hidden entries for any curve that wasn't placed (so the
-              // D3 data-join, keyed by series key, is clean).
-              const labeledKeys = new Set(lineLabels.map((l) => l.key));
-              for (const entry of entries) {
-                if (!labeledKeys.has(entry.key)) {
-                  lineLabels.push({
-                    key: entry.key,
-                    hw: entry.hw,
-                    label: lineLabelText(entry.hw, entry.precision, multiPrecision, modelLabel),
-                    color: ir.getCssColor(ir.resolveColor(entry.hw)),
-                    x: xScale(entry.points[0].x),
-                    y: yScale(entry.points[0].y),
-                    visible: false,
+                  color: overlayRunColor(group.runIndex),
+                  points: group.points,
+                },
+              ];
+            });
+            const labelSeries = [...officialSeries, ...overlaySeries];
+
+            lineLabels =
+              chartDefinition.chartType === 'interactivity'
+                ? placeLineLabels(labelSeries, xScale, yScale, {
+                    collisionWidth: 120,
+                    anchors: lineLabelAnchorRef.current,
+                    pinAnchors: pinLineLabels,
+                    obstacles: parallelismLabelBoxes(ctx.layout.zoomGroup.node()),
+                  })
+                : placeEndpointLineLabels(labelSeries, xScale, yScale, {
+                    nudge: !pinLineLabels,
                   });
-                  labeledKeys.add(entry.key);
-                }
-              }
 
-              // Overlay (unofficial run) rooflines also get line labels using the
-              // run-palette color so they match the legend swatches. The label
-              // text mirrors the overlay legend ("✕ <branch>" — falls back to the
-              // hw label if run metadata isn't available, e.g. legacy callers).
-              const overlayLabelText = (
-                runIndex: number,
-                hwKey: string,
-                precision: string,
-              ): string => {
-                const info = unofficialRunInfos[runIndex];
-                const base = info
-                  ? `✕ ${info.branch || `run ${info.id}`}`
-                  : parseHwKeyToLabel(hwKey, modelLabel).label;
-                return multiPrecision
-                  ? `${base} ${getPrecisionLabel(precision as Precision)}`
-                  : base;
-              };
-              const sortedOverlay = Object.entries(overlayRooflines)
-                .filter(([, group]) => ir.activeOverlayHwTypes.has(group.hwKey))
-                .toSorted(([, a], [, b]) => yScale(a.points[0].y) - yScale(b.points[0].y));
-
-              for (const [ovKey, group] of sortedOverlay) {
-                placeLabel(
-                  `overlay-${ovKey}`,
-                  group.hwKey,
-                  overlayLabelText(group.runIndex, group.hwKey, group.points[0]?.precision ?? ''),
-                  overlayRunColor(group.runIndex),
-                  group.points,
-                );
-              }
+            // Keep hidden data-join entries for precision/date curves that lost
+            // deduplication, preserving the chart's one-label-per-series identity.
+            const labeledKeys = new Set(lineLabels.map((label) => label.key));
+            for (const entry of entries) {
+              if (labeledKeys.has(entry.key)) continue;
+              lineLabels.push({
+                key: entry.key,
+                seriesId: entry.hw,
+                label: lineLabelText(entry.hw, entry.precision, multiPrecision, modelLabel),
+                color: ir.getCssColor(ir.resolveColor(entry.hw)),
+                x: xScale(entry.points[0].x),
+                y: yScale(entry.points[0].y),
+                visible: false,
+              });
+              labeledKeys.add(entry.key);
+            }
 
-              // Drop anchors for series no longer present so the map stays bounded
-              // and a re-appearing series gets a fresh, in-range anchor.
-              if (pinLineLabels) {
-                const live = new Set(lineLabels.map((l) => l.key));
-                // Deleting the current key during Map iteration is well-defined.
-                for (const k of anchors.keys()) {
-                  if (!live.has(k)) anchors.delete(k);
-                }
-              }
-            } else {
-              // TTFT / E2EL: endpoint labels, one per curve (per hw, or per
-              // (hw, precision) when multiple precisions are shown).
-              const seen = new Set<string>();
-              for (const entry of entries) {
-                if (!entry.visible) continue;
-                const groupKey = multiPrecision ? entry.key : entry.hw;
-                if (seen.has(groupKey)) continue;
-                seen.add(groupKey);
-                const pt = entry.points.at(-1)!;
-                lineLabels.push({
-                  key: entry.key,
-                  hw: entry.hw,
-                  label: lineLabelText(entry.hw, entry.precision, multiPrecision, modelLabel),
-                  color: ir.getCssColor(ir.resolveColor(entry.hw)),
-                  x: xScale(pt.x),
-                  y: yScale(pt.y),
-                  visible: true,
-                });
-              }
-              // Endpoint labels for overlay rooflines too (one per (hw, runIndex)),
-              // labeled with the run's branch name to mirror the overlay legend.
-              for (const [ovKey, group] of Object.entries(overlayRooflines)) {
-                if (!ir.activeOverlayHwTypes.has(group.hwKey)) continue;
-                const info = unofficialRunInfos[group.runIndex];
-                const branchOrHw = info
-                  ? `✕ ${info.branch || `run ${info.id}`}`
-                  : parseHwKeyToLabel(group.hwKey, modelLabel).label;
-                const labelText = multiPrecision
-                  ? `${branchOrHw} ${getPrecisionLabel((group.points[0]?.precision ?? '') as Precision)}`
-                  : branchOrHw;
-                const labelKey = `overlay-${ovKey}`;
-                const pt = group.points.at(-1)!;
-                lineLabels.push({
-                  key: labelKey,
-                  hw: group.hwKey,
-                  label: labelText,
-                  color: overlayRunColor(group.runIndex),
-                  x: xScale(pt.x),
-                  y: yScale(pt.y),
-                  visible: true,
-                });
-              }
-              // Pinned (replay): keep labels exactly at their endpoints, which
-              // already move smoothly with the line. The vertical de-overlap
-              // nudge below reshuffles positions as endpoints shift frame-to-
-              // frame, so skip it to preserve positional affinity.
-              const visible = lineLabels.filter((l) => l.visible);
-              if (visible.length > 1 && !pinLineLabels) {
-                const yRange = yScale.range();
-                const top = Math.min(yRange[0], yRange[1]) + LABEL_H;
-                const bottom = Math.max(yRange[0], yRange[1]) - LABEL_H;
-                visible.sort((a, b) => a.y - b.y);
-                for (let pass = 0; pass < 5; pass++) {
-                  for (let i = 1; i < visible.length; i++) {
-                    const overlap = visible[i - 1].y + LABEL_H - visible[i].y;
-                    if (overlap > 0) {
-                      const half = overlap / 2;
-                      visible[i - 1].y -= half;
-                      visible[i].y += half;
-                    }
-                  }
-                  for (const l of visible) {
-                    l.y = Math.max(top, Math.min(bottom, l.y));
-                  }
-                }
+            if (pinLineLabels) {
+              for (const key of lineLabelAnchorRef.current.keys()) {
+                if (!labeledKeys.has(key)) lineLabelAnchorRef.current.delete(key);
               }
             }
           }
 
-          const llSel = zoomGroup
-            .selectAll<SVGGElement, LineLabel>('.line-label')
-            .data(lineLabels, (d) => d.key)
-            .join(
-              (enter) => {
-                const g = enter
-                  .append('g')
-                  .attr('class', 'line-label')
-                  .style('pointer-events', 'none')
-                  .attr('transform', (d) => `translate(${d.x},${d.y})`);
-                g.append('rect')
-                  .attr('class', 'll-bg')
-                  .attr('rx', 4)
-                  .attr('ry', 4)
-                  .attr('opacity', 0.95);
-                g.append('text')
-                  .attr('class', 'll-text')
-                  .attr('text-anchor', 'start')
-                  .attr('dominant-baseline', 'central')
-                  .attr('fill', 'white')
-                  .attr('font-size', '10px')
-                  .attr('font-weight', '600');
-                return g;
-              },
-              (update) => update,
-              (exit) => exit.remove(),
-            )
-            .attr('data-line-key', (d) => d.key)
-            .attr('data-hw-key', (d) => d.hw)
-            // Persist the per-label visibility decision (one label per hw group;
-            // de-duplicated / collision-hidden curves are `false`) so the
-            // visibility-sync effects below don't re-show hidden duplicates that
-            // share the same `data-hw-key`. See GH #470.
-            .attr('data-visible', (d) => (d.visible ? '1' : '0'))
-            .attr('transform', (d) => `translate(${d.x + 8},${d.y - 14})`)
-            .style('transition', 'opacity 150ms ease')
-            .style('opacity', (d) => (d.visible ? 1 : 0));
-
-          // Two-pass text/bbox sizing — same batching rationale as the
-          // parallelism labels above.
-          llSel.each(function (d) {
-            d3.select(this).select<SVGTextElement>('.ll-text').text(d.label);
+          renderLineLabels(zoomGroup, lineLabels, {
+            seriesAttribute: 'data-hw-key',
+            iconFor: (label) => getLineLabelVendorIcon(label.seriesId),
+            configureGroup: (labelGroup, label) => {
+              labelGroup
+                .attr('data-visible', label.visible ? '1' : '0')
+                .select('.ll-bg')
+                .attr('opacity', 0.95);
+            },
+            configureText: (text, label) => {
+              const config = getHardwareConfig(label.seriesId, modelLabel);
+              const hardwareLabel = getDisplayLabel(config);
+              const isHardwareLabel =
+                label.label === hardwareLabel || label.label.startsWith(`${config.label} `);
+              const remainingLabel = isHardwareLabel ? label.label.slice(config.label.length) : '';
+              const engineLabel = config.suffix ? ` ${config.suffix}` : '';
+              const precisionLabel =
+                engineLabel && remainingLabel.endsWith(engineLabel)
+                  ? remainingLabel.slice(0, -engineLabel.length)
+                  : remainingLabel;
+              const segments = isHardwareLabel
+                ? [
+                    { className: 'll-gpu', text: config.label, fill: 'white', weight: '700' },
+                    ...(precisionLabel
+                      ? [
+                          {
+                            className: 'll-precision',
+                            text: precisionLabel,
+                            fill: 'white',
+                            weight: '600',
+                          },
+                        ]
+                      : []),
+                    ...(config.suffix
+                      ? [
+                          {
+                            className: 'll-engine',
+                            text: engineLabel,
+                            fill: '#d1d5db',
+                            weight: '400',
+                          },
+                        ]
+                      : []),
+                  ]
+                : [
+                    {
+                      className: 'll-plain',
+                      text: label.label,
+                      fill: 'white',
+                      weight: '600',
+                    },
+                  ];
+              text
+                .selectAll<SVGTSpanElement, (typeof segments)[number]>('tspan')
+                .data(segments, (segment) => segment.className)
+                .join('tspan')
+                .attr('class', (segment) => segment.className)
+                .attr('fill', (segment) => segment.fill)
+                .attr('font-weight', (segment) => segment.weight)
+                .text((segment) => segment.text);
+            },
           });
-          const llMeasured: { node: SVGGElement; d: LineLabel; bbox: DOMRect }[] = [];
-          llSel.each(function (d) {
-            const text = this.querySelector<SVGTextElement>('.ll-text');
-            if (text) llMeasured.push({ node: this, d, bbox: text.getBBox() });
+        },
+        onDisplayUpdate: (zoomGroup, ctx) => {
+          const transform = d3.zoomTransform(ctx.layout.svg.node()!);
+          rooflineLayer.render?.(zoomGroup, {
+            ...ctx,
+            xScale: transform.rescaleX(ctx.xScale as ContinuousScale),
+            yScale: transform.rescaleY(ctx.yScale as ContinuousScale),
           });
-          for (const { node, d, bbox } of llMeasured) {
-            const px = 5;
-            const py = 3;
-            d3.select(node)
-              .select('.ll-bg')
-              .attr('x', bbox.x - px)
-              .attr('y', bbox.y - py)
-              .attr('width', bbox.width + px * 2)
-              .attr('height', bbox.height + py * 2)
-              .attr('fill', d.color);
-          }
+          zoomGroup.selectAll('.line-label').raise();
         },
         onZoom: (zoomGroup, ctx) => {
           const ir = interactionRef.current;
@@ -2120,162 +2408,74 @@ const ScatterGraph = React.memo(
             });
           }
 
-          // Update line label positions on zoom
+          // Update line label positions on zoom with the same placement
+          // primitives used for the initial render.
           if (showLineLabels) {
-            const isInteractivity = chartDefinition.chartType === 'interactivity';
             const multiPrecision = ir.selectedPrecisions.length > 1;
-            const LABEL_H = 18;
-            const LABEL_W = 120;
-
-            if (isInteractivity) {
-              // Deduplicate by group key — one curve per hw, or per (hw, precision)
-              // when multiple precisions are shown (mirrors the static render).
-              const bestByGroup = new Map<string, [string, InferenceData[]]>();
-              for (const [key, pts] of Object.entries(rooflines)) {
-                const hw = key.split('_').slice(0, -1).join('_');
-                const prec = key.split('_').pop()!;
-                if (!ir.effectiveActiveHwTypes.has(hw) || !ir.selectedPrecisions.includes(prec))
-                  continue;
-                const groupKey = multiPrecision ? key : hw;
-                const prev = bestByGroup.get(groupKey);
-                if (!prev || pts.length > prev[1].length) bestByGroup.set(groupKey, [key, pts]);
+            const bestByGroup = new Map<
+              string,
+              { key: string; seriesId: string; points: InferenceData[] }
+            >();
+            for (const [key, points] of Object.entries(rooflines)) {
+              const hardware = key.split('_').slice(0, -1).join('_');
+              const precision = key.split('_').pop()!;
+              if (
+                !ir.effectiveActiveHwTypes.has(hardware) ||
+                !ir.selectedPrecisions.includes(precision)
+              ) {
+                continue;
               }
-              const visibleEntries = [...bestByGroup.values()].toSorted(
-                ([, a], [, b]) => newYScale(a[0].y) - newYScale(b[0].y),
-              );
-              const overlayVisible = Object.entries(overlayRooflines)
-                .filter(([, group]) => ir.activeOverlayHwTypes.has(group.hwKey))
-                .toSorted(([, a], [, b]) => newYScale(a.points[0].y) - newYScale(b.points[0].y));
-
-              const zoomResults = new Map<string, { x: number; y: number; vis: boolean }>();
-
-              if (pinLineLabels) {
-                // Pinned (replay): keep each label on its stored data-space anchor
-                // under the zoomed scales instead of re-running greedy placement, so
-                // a zoom mid-replay preserves the same positional affinity as the
-                // render path. Always visible — positional stability is the goal.
-                const anchors = lineLabelAnchorRef.current;
-                const pinTo = (key: string, pts: InferenceData[]) => {
-                  const anchorX = anchors.get(key);
-                  const pt = anchorX === undefined ? pts.at(-1)! : pointNearestX(pts, anchorX);
-                  zoomResults.set(key, { x: newXScale(pt.x), y: newYScale(pt.y), vis: true });
-                };
-                for (const [key, pts] of visibleEntries) pinTo(key, pts);
-                for (const [ovKey, group] of overlayVisible)
-                  pinTo(`overlay-${ovKey}`, group.points);
-              } else {
-                // Re-run greedy placement with zoomed scales (static chart). Overlay
-                // rooflines share the same `placed` array so they stay non-
-                // overlapping with the official labels post-zoom.
-                const placed: { x: number; y: number }[] = [];
-                const collides = (cx: number, cy: number) =>
-                  placed.some((p) => Math.abs(p.y - cy) < LABEL_H && Math.abs(p.x - cx) < LABEL_W);
-                const greedyPlace = (
-                  key: string,
-                  pts: InferenceData[],
-                  keepVisibleOnCollision = false,
-                ) => {
-                  const candidates = [
-                    pts[Math.min(1, pts.length - 1)],
-                    pts[Math.floor(pts.length / 2)],
-                    pts[Math.max(0, Math.floor((pts.length * 2) / 3))],
-                    pts.at(-1)!,
-                  ];
-                  for (const pt of candidates) {
-                    const px = newXScale(pt.x);
-                    const py = newYScale(pt.y);
-                    if (!collides(px, py)) {
-                      zoomResults.set(key, { x: px, y: py, vis: true });
-                      placed.push({ x: px, y: py });
-                      return;
-                    }
-                  }
-                  const px = newXScale(pts[0].x);
-                  const py = newYScale(pts[0].y);
-                  zoomResults.set(key, {
-                    x: px,
-                    y: py,
-                    vis: keepVisibleOnCollision,
+              const pointsByDate = groupPointsByDate(points);
+              const singleDate = pointsByDate.size === 1;
+              for (const [date, datePoints] of pointsByDate) {
+                const entryKey = singleDate ? key : `${key}__${date}`;
+                const groupKey = multiPrecision ? entryKey : hardware;
+                const previous = bestByGroup.get(groupKey);
+                if (!previous || datePoints.length > previous.points.length) {
+                  bestByGroup.set(groupKey, {
+                    key: entryKey,
+                    seriesId: hardware,
+                    points: datePoints,
                   });
-                  if (keepVisibleOnCollision) placed.push({ x: px, y: py });
-                };
-                for (const [key, pts] of visibleEntries) greedyPlace(key, pts, pts.length === 1);
-                for (const [ovKey, group] of overlayVisible)
-                  greedyPlace(`overlay-${ovKey}`, group.points);
-              }
-
-              zoomGroup.selectAll<SVGGElement, unknown>('.line-label').each(function () {
-                const el = d3.select(this);
-                const k = el.attr('data-line-key');
-                const zl = zoomResults.get(k);
-                if (zl) {
-                  el.attr('transform', `translate(${zl.x + 8},${zl.y - 14})`);
-                  el.style('opacity', zl.vis ? 1 : 0);
-                } else {
-                  el.style('opacity', 0);
-                }
-              });
-            } else {
-              // TTFT / E2EL: endpoint with bidirectional nudge, one per hw
-              interface ZoomLabel {
-                key: string;
-                x: number;
-                y: number;
-              }
-              const zoomLabels: ZoomLabel[] = [];
-              const seen = new Set<string>();
-              Object.entries(rooflines).forEach(([key, pts]) => {
-                const hw = key.split('_').slice(0, -1).join('_');
-                const prec = key.split('_').pop()!;
-                if (!ir.effectiveActiveHwTypes.has(hw) || !ir.selectedPrecisions.includes(prec))
-                  return;
-                const groupKey = multiPrecision ? key : hw;
-                if (seen.has(groupKey)) return;
-                seen.add(groupKey);
-                const pt = pts.at(-1)!;
-                zoomLabels.push({ key, x: newXScale(pt.x), y: newYScale(pt.y) });
-              });
-              // Overlay rooflines: per-(hw, runIndex) endpoint labels.
-              for (const [ovKey, group] of Object.entries(overlayRooflines)) {
-                if (!ir.activeOverlayHwTypes.has(group.hwKey)) continue;
-                const pt = group.points.at(-1)!;
-                zoomLabels.push({
-                  key: `overlay-${ovKey}`,
-                  x: newXScale(pt.x),
-                  y: newYScale(pt.y),
-                });
-              }
-              // Skip the vertical de-overlap nudge while pinned (replay): the
-              // endpoints already move smoothly with the lines, and nudging
-              // reshuffles positions frame-to-frame (mirrors the render path).
-              if (zoomLabels.length > 1 && !pinLineLabels) {
-                const yRange = newYScale.range();
-                const top = Math.min(yRange[0], yRange[1]) + LABEL_H;
-                const bottom = Math.max(yRange[0], yRange[1]) - LABEL_H;
-                zoomLabels.sort((a, b) => a.y - b.y);
-                for (let pass = 0; pass < 5; pass++) {
-                  for (let i = 1; i < zoomLabels.length; i++) {
-                    const overlap = zoomLabels[i - 1].y + LABEL_H - zoomLabels[i].y;
-                    if (overlap > 0) {
-                      const half = overlap / 2;
-                      zoomLabels[i - 1].y -= half;
-                      zoomLabels[i].y += half;
-                    }
-                  }
-                  for (const l of zoomLabels) {
-                    l.y = Math.max(top, Math.min(bottom, l.y));
-                  }
-                }
-              }
-              for (const zl of zoomLabels) {
-                const labelGroup = zoomGroup.select<SVGGElement>(
-                  `.line-label[data-line-key="${zl.key}"]`,
-                );
-                if (!labelGroup.empty()) {
-                  labelGroup.attr('transform', `translate(${zl.x + 8},${zl.y - 14})`);
                 }
               }
             }
+            const officialSeries: LineLabelSeries<InferenceData>[] = [...bestByGroup.values()].map(
+              (entry) => ({
+                ...entry,
+                label: '',
+                color: '',
+                keepVisibleOnCollision: entry.points.length === 1,
+              }),
+            );
+            const overlaySeries: LineLabelSeries<InferenceData>[] = Object.entries(
+              overlayRooflines,
+            ).flatMap(([overlayKey, group]) =>
+              ir.activeOverlayHwTypes.has(group.hwKey)
+                ? [
+                    {
+                      key: `overlay-${overlayKey}`,
+                      seriesId: group.hwKey,
+                      label: '',
+                      color: '',
+                      points: group.points,
+                    },
+                  ]
+                : [],
+            );
+            const labelSeries = [...officialSeries, ...overlaySeries];
+            const zoomLabels =
+              chartDefinition.chartType === 'interactivity'
+                ? placeLineLabels(labelSeries, newXScale, newYScale, {
+                    collisionWidth: 120,
+                    anchors: lineLabelAnchorRef.current,
+                    pinAnchors: pinLineLabels,
+                    obstacles: parallelismLabelBoxes(zoomGroup.node()),
+                  })
+                : placeEndpointLineLabels(labelSeries, newXScale, newYScale, {
+                    nudge: !pinLineLabels,
+                  });
+            updateRenderedLineLabels(zoomGroup, zoomLabels);
           }
         },
       };
@@ -2297,17 +2497,21 @@ const ScatterGraph = React.memo(
           getOpacity: (d) => (interactionRef.current.isPointVisible(d) ? 1 : 0),
           getPointerEvents: (d) => (interactionRef.current.isPointVisible(d) ? 'auto' : 'none'),
           hideLabels: !showPointLabels || showGradientLabels,
-          // Keep the concurrency (C=) annotation from the agentx scatter labels.
-          getLabelText: (d) => pointLabelText(d, useAdvancedLabels),
+          // Concurrency (C=) is appended only when the advanced
+          // "# Concurrent Sessions" toggle is on.
+          getLabelText: (d) => pointLabelText(d, useAdvancedLabels, showConcurrencyLabels),
           foreground: 'var(--foreground)',
           dataAttrs: {
             'hw-key': (d) => String(d.hwKey),
             precision: (d) => d.precision,
+            // Lets the agentic coach mark pick an anchor out of the DOM
+            // without knowing anything about React state.
+            'benchmark-type': (d) => d.benchmark_type ?? '',
           },
           getShapeKey: (d) =>
             getShapeKeyForPrecision(d.precision, interactionRef.current.selectedPrecisions),
         },
-        keyFn: buildPointConfigId,
+        keyFn: buildPointId,
       };
 
       // ── Layer 2: Overlay (rooflines + X-shape points) ──
@@ -2315,6 +2519,7 @@ const ScatterGraph = React.memo(
         ? {
             type: 'custom',
             key: 'overlay',
+            displayIdentity: `${showPointLabels}:${showGradientLabels}`,
             render: (zoomGroup, ctx) => {
               const xScale = ctx.xScale as ContinuousScale;
               const yScale = ctx.yScale as ContinuousScale;
@@ -2378,7 +2583,7 @@ const ScatterGraph = React.memo(
                       d3.select(this)
                         .append('path')
                         .attr('class', 'visible-shape overlay-x')
-                        .attr('d', getXPath(X_SIZE))
+                        .attr('d', xMarkerPath(5, 0.7))
                         .attr('fill', 'none')
                         .attr('stroke-width', 2.5)
                         .attr('stroke-linecap', 'round')
@@ -2418,11 +2623,13 @@ const ScatterGraph = React.memo(
               // Labels
               const showLabels = showPointLabels && !showGradientLabels;
               overlayPoints.each(function (d) {
-                const lines = showLabels ? pointLabelText(d, useAdvancedLabels).split('\n') : [];
+                const lines = pointLabelText(d, useAdvancedLabels, showConcurrencyLabels).split(
+                  '\n',
+                );
                 const text = d3
                   .select(this)
                   .selectAll<SVGTextElement, boolean>('.overlay-label')
-                  .data(showLabels ? [true] : [])
+                  .data([true])
                   .join('text')
                   .attr('class', 'overlay-label')
                   .attr('text-anchor', 'middle')
@@ -2439,87 +2646,82 @@ const ScatterGraph = React.memo(
                   .attr('dy', (_l, i) => (i === 0 ? `${firstDy}em` : '1.1em'))
                   .text((l) => l);
               });
+              overlayPoints
+                .selectAll('.overlay-label')
+                .style('display', showLabels ? '' : 'none')
+                .style('opacity', showLabels ? 1 : 0);
 
-              // Overlay tooltip handlers
+              // Overlay marker content/coordinates stay chart-local; the shared
+              // helper owns the hover, pin, ruler, and marker-state lifecycle.
               const svgNode = ctx.layout.svg.node()!;
               const container = svgNode.parentElement as HTMLDivElement;
               const tooltip = d3.select(ctx.tooltipElement);
-
-              const createOverlayConfig = (d: InferenceData, pinned: boolean) => ({
-                data: d,
-                isPinned: pinned,
-                xLabel,
-                yLabel,
-                selectedYAxisMetric,
-                hardwareConfig: overlayData.hardwareConfig,
-                overlayData,
-                locale,
-              });
-
-              overlayPoints
-                .on('mouseenter', function (_event, d) {
-                  if (chartRef.current?.isPinned()) return;
-                  const shape = d3.select(this).select('.overlay-x');
-                  shape.attr('d', getXPath(X_HOVER_SIZE)).attr('stroke-width', 3.5);
-                  tooltip
-                    .style('opacity', 1)
-                    .style('display', 'block')
-                    .style('pointer-events', 'none');
-
-                  // Position rulers
-                  const rulerGroup = zoomGroup.select('.ruler-group');
-                  const vRuler = zoomGroup.select('.vertical-ruler');
-                  const hRuler = zoomGroup.select('.horizontal-ruler');
-                  const ct = d3.zoomTransform(svgNode);
-                  const curX = ct.rescaleX(xScale);
-                  const curY = ct.rescaleY(yScale);
-                  rulerGroup.style('display', 'block');
-                  vRuler.attr('x1', curX(d.x)).attr('x2', curX(d.x));
-                  hRuler.attr('y1', curY(d.y)).attr('y2', curY(d.y));
-
-                  tooltip.html(generateOverlayTooltipContent(createOverlayConfig(d, false)));
-                })
-                .on('mousemove', (event) => {
-                  if (chartRef.current?.isPinned()) return;
-                  const [mx, my] = d3.pointer(event, container);
-                  const pos = computeTooltipPosition(mx, my, tooltip, container);
-                  tooltip.style('left', `${pos.left}px`).style('top', `${pos.top}px`);
-                })
-                .on('mouseleave', function () {
-                  if (chartRef.current?.isPinned()) return;
-                  const shape = d3.select(this).select('.overlay-x');
-                  shape.attr('d', getXPath(X_SIZE)).attr('stroke-width', 2.5);
-                  tooltip.style('opacity', 0).style('display', 'none');
-                  zoomGroup.select('.ruler-group').style('display', 'none');
-                })
-                .on('click', (event, d) => {
-                  event.stopPropagation();
-                  const [mx, my] = d3.pointer(event, container);
-                  tooltip.html(generateOverlayTooltipContent(createOverlayConfig(d, true)));
-                  const pos = computeTooltipPosition(mx, my, tooltip, container);
-                  tooltip
-                    .style('left', `${pos.left}px`)
-                    .style('top', `${pos.top}px`)
-                    .style('opacity', 1)
-                    .style('display', 'block')
-                    .style('pointer-events', 'auto');
-
-                  // Position rulers at clicked point
-                  const ct = d3.zoomTransform(svgNode);
-                  const curX = ct.rescaleX(xScale);
-                  const curY = ct.rescaleY(yScale);
-                  zoomGroup.select('.ruler-group').style('display', 'block');
-                  zoomGroup.select('.vertical-ruler').attr('x1', curX(d.x)).attr('x2', curX(d.x));
-                  zoomGroup.select('.horizontal-ruler').attr('y1', curY(d.y)).attr('y2', curY(d.y));
-
-                  chartRef.current?.pinTooltip(d, true);
+              const showRulers = (point: InferenceData, marker: SVGGElement) => {
+                const position = overlayMarkerPosition(marker) ?? {
+                  x: xScale(point.x),
+                  y: yScale(point.y),
+                };
+                zoomGroup.select('.ruler-group').style('display', 'block');
+                zoomGroup.select('.vertical-ruler').attr('x1', position.x).attr('x2', position.x);
+                zoomGroup.select('.horizontal-ruler').attr('y1', position.y).attr('y2', position.y);
+              };
+              attachOverlayXMarkerHandlers(overlayPoints, {
+                markerSelector: '.overlay-x',
+                normalPath: xMarkerPath(5, 0.7),
+                hoverPath: xMarkerPath(7, 0.7),
+                tooltip,
+                handle: chartRef.current,
+                content: (point, pinned) =>
+                  generateOverlayTooltipContent({
+                    data: point,
+                    isPinned: pinned,
+                    xLabel,
+                    yLabel,
+                    selectedYAxisMetric,
+                    hardwareConfig: overlayData.hardwareConfig,
+                    overlayData,
+                    locale,
+                  }),
+                position: (event) => {
+                  const [mouseX, mouseY] = d3.pointer(event, container);
+                  return computeTooltipPosition(mouseX, mouseY, tooltip, container);
+                },
+                rulers: {
+                  show: showRulers,
+                  hide: () => zoomGroup.select('.ruler-group').style('display', 'none'),
+                },
+                onClick: (point) => {
+                  // Overlay points participate in the perf ruler too — a
+                  // ruler-mode click acts as a click on the overlay curve at
+                  // the point's x instead of keeping the pinned tooltip (see
+                  // AGENTS.md "Unofficial Run Support").
+                  const ruler = perfRulerRef.current;
+                  if (ruler.mode) {
+                    track('latency_data_point_clicked', {
+                      hw: String(point.hwKey),
+                      x: point.x,
+                      y: point.y,
+                      overlay: true,
+                      perfRuler: true,
+                    });
+                    ruler.onPointClick(point, 'overlay');
+                    return;
+                  }
                   track('latency_data_point_clicked', {
-                    hw: String(d.hwKey),
-                    x: d.x,
-                    y: d.y,
+                    hw: String(point.hwKey),
+                    x: point.x,
+                    y: point.y,
                     overlay: true,
                   });
-                });
+                },
+              });
+            },
+            onDisplayUpdate: (zoomGroup) => {
+              const showLabels = showPointLabels && !showGradientLabels;
+              zoomGroup
+                .selectAll('.unofficial-overlay-pt .overlay-label')
+                .style('display', showLabels ? '' : 'none')
+                .style('opacity', showLabels ? 1 : 0);
             },
             onZoom: (zoomGroup, ctx) => {
               const newXScale = ctx.newXScale as ContinuousScale;
@@ -2546,105 +2748,6 @@ const ScatterGraph = React.memo(
           }
         : null;
 
-      const speedOverlayLayer: CustomLayerConfig = {
-        type: 'custom',
-        key: 'speed-overlay',
-        render: (_zoomGroup, ctx) => {
-          const { g } = ctx.layout;
-          g.selectAll('.speed-overlay').remove();
-          if (!showSpeedOverlay && !showMinecraftOverlay) return;
-          const w = ctx.width;
-          const h = ctx.height;
-          const SIZE = 78;
-          const PAD = 8;
-          const STACK_GAP = 4;
-          const rooflineKey = `${selectedYAxisMetric}_roofline` as keyof ChartDefinition;
-          const dir = chartDefinition[rooflineKey] as RooflineDirection | undefined;
-          const { busTop, busLeft } = getSpeedOverlayCorners(dir);
-          const layer = g.append('g').attr('class', 'speed-overlay').attr('pointer-events', 'none');
-
-          // Each enabled "pair" stacks horizontally inward from the chart corner so
-          // the second pair sits next to (not on top of) the first one when both
-          // toggles are on. The bus-side stays anchored to the batch corner; the
-          // car-side stays anchored to the interactive corner. Pair items can have
-          // independent slow/fast sizes so the donkey can be visually heavier than
-          // the elytra without affecting the bus/car pair.
-          interface OverlayPair {
-            id: string;
-            slowSrc: string;
-            fastSrc: string;
-            slowSize: number;
-            fastSize: number;
-          }
-          const enabledPairs: OverlayPair[] = [];
-          if (showSpeedOverlay) {
-            enabledPairs.push({
-              id: 'speed',
-              slowSrc: '/decorative/bus.png',
-              fastSrc: '/decorative/racing-car.png',
-              slowSize: SIZE,
-              fastSize: SIZE,
-            });
-          }
-          if (showMinecraftOverlay) {
-            // donkey-chest.png — Chested_Donkey_JE5 from minecraft.wiki/w/Donkey,
-            //   rendered 50% larger than the other overlay icons (1.5× SIZE).
-            // elytra.png — ElytraNew sprite (front-facing both wings) from the
-            //   Minecraft Fandom wiki at 160×160 pixel-art.
-            enabledPairs.push({
-              id: 'minecraft',
-              slowSrc: '/decorative/donkey-chest.png',
-              fastSrc: '/decorative/elytra.png',
-              slowSize: Math.round(SIZE * 1.5),
-              fastSize: SIZE,
-            });
-          }
-
-          const slowCornerName = `${busTop ? 'top' : 'bottom'}-${busLeft ? 'left' : 'right'}`;
-          const fastCornerName = `${busTop ? 'bottom' : 'top'}-${busLeft ? 'right' : 'left'}`;
-          let slowInward = 0;
-          let fastInward = 0;
-          enabledPairs.forEach((pair) => {
-            const slowX = busLeft ? PAD + slowInward : w - pair.slowSize - PAD - slowInward;
-            const slowY = busTop ? PAD : h - pair.slowSize - PAD;
-            const fastX = busLeft ? w - pair.fastSize - PAD - fastInward : PAD + fastInward;
-            const fastY = busTop ? h - pair.fastSize - PAD : PAD;
-            layer
-              .append('image')
-              .attr('class', `speed-overlay-slow speed-overlay-${pair.id}-slow`)
-              .attr('data-testid', `speed-overlay-${pair.id}-slow`)
-              .attr('data-corner', slowCornerName)
-              .attr('href', pair.slowSrc)
-              .attr('x', slowX)
-              .attr('y', slowY)
-              .attr('width', pair.slowSize)
-              .attr('height', pair.slowSize)
-              .attr('opacity', 0.85);
-            layer
-              .append('image')
-              .attr('class', `speed-overlay-fast speed-overlay-${pair.id}-fast`)
-              .attr('data-testid', `speed-overlay-${pair.id}-fast`)
-              .attr('data-corner', fastCornerName)
-              .attr('href', pair.fastSrc)
-              .attr('x', fastX)
-              .attr('y', fastY)
-              .attr('width', pair.fastSize)
-              .attr('height', pair.fastSize)
-              .attr('opacity', 0.85);
-            slowInward += pair.slowSize + STACK_GAP;
-            fastInward += pair.fastSize + STACK_GAP;
-          });
-
-          // Backwards-compatible aliases so existing E2E tests (speed-overlay.cy.ts)
-          // can still find the bus/car pair via `[data-testid="speed-overlay-bus"]`
-          // and `[data-testid="speed-overlay-car"]`.
-          if (showSpeedOverlay) {
-            layer.select('.speed-overlay-speed-slow').attr('data-testid', 'speed-overlay-bus');
-            layer.select('.speed-overlay-speed-fast').attr('data-testid', 'speed-overlay-car');
-          }
-        },
-      };
-
       // ── Intentional clipping: interpolated dashed Pareto continuation + arrow ──
       const drawOverflowContinuations = (
         zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
@@ -2799,56 +2902,56 @@ const ScatterGraph = React.memo(
           ),
       };
 
-      // ── Known-issue annotations: warning box + arrow to the affected line ──
-      const drawKnownIssues = (
-        ctx: RenderContext,
-        xScale: ContinuousScale,
-        yScale: ContinuousScale,
-      ) => {
-        // Annotations / colors via the interaction ref: they change with the
-        // visible series, and the decoration effect re-runs this layer on
-        // toggles without recreating it.
-        const ir = interactionRef.current;
-        renderKnownIssueAnnotations(ctx.layout.g, ctx.layout.defs, {
+      // Annotations/colors resolve through the interaction ref so visibility
+      // toggles can redraw this layer without recreating the chart.
+      const knownIssueLayer = createKnownIssueLayer(() => {
+        const current = interactionRef.current;
+        return {
           chartId,
-          width: ctx.width,
-          height: ctx.height,
-          xScale,
-          yScale,
-          annotations: ir.knownIssueAnnotations,
-          // Only measure the legend overlap when there are boxes to place —
-          // this runs on every zoom frame, and the measurement forces layout.
-          rightInset:
-            ir.knownIssueAnnotations.length === 0
-              ? 0
-              : measureLegendRightInset(
-                  chartId,
-                  ctx.layout.svg.node(),
-                  ctx.layout.margin.left,
-                  ctx.width,
-                ),
-          background: ir.getCssColor('--background'),
-          foreground: ir.getCssColor('--foreground'),
-          mutedForeground: ir.getCssColor('--muted-foreground'),
-          onLinkClick: (a) =>
+          annotations: current.knownIssueAnnotations,
+          background: current.getCssColor('--background'),
+          foreground: current.getCssColor('--foreground'),
+          mutedForeground: current.getCssColor('--muted-foreground'),
+          onLinkClick: (annotation) =>
+            annotation.issue &&
             track('inference_known_issue_clicked', {
-              hwKey: a.issue.hwKey,
-              issue: a.issue.issueRef,
+              hwKey: annotation.issue.hwKey,
+              issue: annotation.issue.issueRef,
             }),
-        });
-      };
-      const knownIssueLayer: CustomLayerConfig = {
+        };
+      });
+
+      // ── Perf ruler (opt-in iso-x measurement between two curves) ──
+      // Mode and measurement are read through refs, so this layer needs
+      // no perf-ruler dependencies: full re-renders redraw it after the data
+      // phase (the iso-x lives in data space, so it survives metric/display
+      // re-renders), and onZoom keeps it glued to the curves during pan/zoom.
+      // It lives inside the zoomGroup, so it is clipped and PNG-exported like
+      // any other mark.
+      const perfRulerLayer: CustomLayerConfig = {
         type: 'custom',
-        key: 'known-issues',
-        render: (_zoomGroup, ctx) =>
-          drawKnownIssues(ctx, ctx.xScale as ContinuousScale, ctx.yScale as ContinuousScale),
-        onZoom: (_zoomGroup, ctx) =>
-          drawKnownIssues(ctx, ctx.newXScale as ContinuousScale, ctx.newYScale as ContinuousScale),
+        key: 'perf-ruler',
+        render: (zoomGroup, ctx) =>
+          drawPerfRuler(
+            zoomGroup,
+            (ctx.renderedXScale ?? ctx.xScale) as ContinuousScale,
+            (ctx.renderedYScale ?? ctx.yScale) as ContinuousScale,
+            ctx.width,
+            ctx.height,
+          ),
+        onZoom: (zoomGroup, ctx) =>
+          drawPerfRuler(
+            zoomGroup,
+            ctx.newXScale as ContinuousScale,
+            ctx.newYScale as ContinuousScale,
+            ctx.width,
+            ctx.height,
+          ),
       };
 
       const result: LayerConfig<InferenceData>[] = [rooflineLayer, scatterLayer];
       if (overlayLayer) result.push(overlayLayer);
-      result.push(overflowContinuationLayer, speedOverlayLayer, knownIssueLayer);
+      result.push(overflowContinuationLayer, perfRulerLayer, knownIssueLayer);
       return result;
       // Interaction state (visibility, colors, precision shapes, known-issue
       // annotations) is deliberately NOT a dependency: layer closures read it
@@ -2861,14 +2964,13 @@ const ScatterGraph = React.memo(
       showGradientLabels,
       showLineLabels,
       pinLineLabels,
-      showSpeedOverlay,
-      showMinecraftOverlay,
       gradientColorByPoint,
       chartId,
       pointsData,
       showPointLabels,
       useAdvancedLabels,
-      buildPointConfigId,
+      showConcurrencyLabels,
+      buildPointId,
       overlayData,
       processedOverlayData,
       overlayRooflines,
@@ -2882,6 +2984,7 @@ const ScatterGraph = React.memo(
       selectedYAxisMetric,
       chartDefinition,
       locale,
+      drawPerfRuler,
     ]);
 
     // Layers handle for the decoration effect — lets it re-run individual
@@ -2905,17 +3008,21 @@ const ScatterGraph = React.memo(
           renderOffloadHalo(d3.select(this), d, 'var(--foreground)');
         });
 
-        avoidLabelCollisions(zoomGroup);
+        avoidPointLabelCollisions(zoomGroup);
 
+        // Tell the nudge engine the chart has painted, so an anchored coach
+        // mark can retry resolving a point to point at. This fires once per
+        // full render, never per zoom frame.
+        window.dispatchEvent(new CustomEvent(SCATTER_RENDERED_EVENT));
         // Log tick formatting on initial render
         if (xScaleConfig._isLog) {
-          const xScale = ctx.xScale as d3.ScaleLogarithmic<number, number>;
+          const xScale = (ctx.renderedXScale ?? ctx.xScale) as d3.ScaleLogarithmic<number, number>;
           ctx.layout.xAxisGroup.call(
             d3.axisBottom(xScale).ticks(10).tickFormat(logTickFormat(xScale)) as any,
           );
         }
         if (yScaleConfig.type === 'log') {
-          const yScale = ctx.yScale as d3.ScaleLogarithmic<number, number>;
+          const yScale = (ctx.renderedYScale ?? ctx.yScale) as d3.ScaleLogarithmic<number, number>;
           ctx.layout.yAxisGroup.call(
             d3.axisLeft(yScale).ticks(10).tickFormat(logTickFormat(yScale)) as any,
           );
@@ -2933,50 +3040,40 @@ const ScatterGraph = React.memo(
 
     // --- Side effects ---
 
-    // Toggle decoration: restyle the existing DOM when visibility or colors
-    // change (legend hw toggles, precision toggles, optimal-only, high
-    // contrast, theme). This is the cheap "Effect 4" display-toggle path from
-    // docs/d3-charts.md — the full chart rebuild only runs when data or scale
-    // domains actually change.
+    // ScatterGraph has more independent display controls than the generic
+    // D3 scatter layer. Keep their mutation scopes separate so a label-only
+    // toggle never restamps every point shape, visibility flag, or trace
+    // attribute.
     //
-    // This effect can run in the same commit as (right after) the full render
-    // effect, while the renderer's old→new "data-update" entrance transitions
-    // are scheduled but not yet started. It therefore NEVER writes the
-    // attributes those transitions animate — dot-group `transform` and
-    // roofline `d` — or a freshly rebuilt roofline would start its transition
-    // already at the destination and teleport while the dots animate.
+    // Mark styling: visibility, palette, and precision shape. This effect can
+    // run immediately after a full render while the renderer's old-to-new
+    // entrance transitions are only scheduled, so it never writes the
+    // animated dot-group `transform` or roofline `d` attributes.
     useLayoutEffect(() => {
-      const svg = chartRef.current?.getSvgElement?.();
-      const ctx = lastRenderCtxRef.current;
-      if (!svg || !ctx) return;
+      const display = getDisplaySelection();
+      if (!display) return;
+      const { zoomGroup } = display;
       const ir = interactionRef.current;
-      const zoomGroup = d3.select(svg).select<SVGGElement>('.zoom-group');
-      if (zoomGroup.empty()) return;
 
-      // Dots: visibility, vendor recolor, and precision shape.
-      // Hand-rolled rather than a full renderScatterPoints pass so we skip
-      // re-writing label text on every point (the expensive part of the join)
-      // — and, critically, never touch the animated `transform`.
       zoomGroup.selectAll<SVGGElement, InferenceData>('.dot-group').each(function (d) {
-        const sel = d3.select(this);
+        const point = d3.select(this);
         const visible = ir.isPointVisible(d);
-        sel.style('opacity', visible ? 1 : 0).style('pointer-events', visible ? 'auto' : 'none');
+        point.style('opacity', visible ? 1 : 0).style('pointer-events', visible ? 'auto' : 'none');
         const color =
           (showGradientLabels && gradientColorByPoint.get(d)) ||
           ir.getCssColor(ir.resolveColor(d.hwKey as string));
         syncPointShape(
-          sel as unknown as d3.Selection<SVGGElement, unknown, null, undefined>,
+          point as unknown as d3.Selection<SVGGElement, unknown, null, undefined>,
           getShapeKeyForPrecision(d.precision, ir.selectedPrecisions),
           color,
         );
         // A precision toggle may replace and append the visible SVG shape.
         // Keep the offload halo above that shape after the swap.
-        sel.selectAll('.offload-halo').raise();
+        point.selectAll('.offload-halo').raise();
       });
 
-      // Overlay X markers: Optimal Only visibility (mirrors the official dot
-      // loop above — the overlay layer render applies the same predicate, but
-      // a toggle flip must restyle existing DOM without a chart rebuild).
+      // Overlay points keep their X marker and run-derived color. Only their
+      // visibility follows Optimal Only and overlay hardware selection.
       zoomGroup.selectAll<SVGGElement, InferenceData>('.unofficial-overlay-pt').each(function (d) {
         const visible = ir.isOverlayPointVisible(d);
         d3.select(this)
@@ -2984,21 +3081,19 @@ const ScatterGraph = React.memo(
           .style('pointer-events', visible ? 'auto' : 'none');
       });
 
-      // Rooflines: visibility + solid-stroke recolor as direct writes (never
-      // `d`). Gradient strokes keep their url(#…) reference — gradient stop
-      // colors come from the fixed parallelism palette and don't change with
-      // the active set.
+      // Rooflines: visibility and solid-stroke recolor as direct writes. Keep
+      // gradient url references intact and never touch animated path geometry.
       zoomGroup.selectAll<SVGPathElement, unknown>('.roofline-path').each(function () {
         const hw = this.dataset.hwKey;
         const precision = this.dataset.precision;
         if (!hw || !precision) return;
-        const el = d3.select(this);
+        const roofline = d3.select(this);
         const visible =
           ir.effectiveActiveHwTypes.has(hw) && ir.selectedPrecisions.includes(precision);
-        el.style('opacity', visible ? 1 : 0);
-        const stroke = el.attr('stroke');
+        roofline.style('opacity', visible ? 1 : 0);
+        const stroke = roofline.attr('stroke');
         if (stroke && !stroke.startsWith('url(')) {
-          el.attr('stroke', ir.getCssColor(ir.resolveColor(hw)));
+          roofline.attr('stroke', ir.getCssColor(ir.resolveColor(hw)));
         }
       });
 
@@ -3021,8 +3116,6 @@ const ScatterGraph = React.memo(
         .selectAll<SVGTextElement, unknown>('.overflow-continuation-label')
         .attr('stroke', ir.getCssColor('--background'));
 
-      // Parallelism / line labels: visibility via data attributes (mirrors
-      // handleLegendHoverEnd). Placement-level updates happen below.
       zoomGroup
         .selectAll<SVGGElement, unknown>('.parallelism-label, .line-label')
         .style('opacity', function () {
@@ -3032,41 +3125,99 @@ const ScatterGraph = React.memo(
             ir.selectedPrecisions,
           );
         });
+    }, [
+      getDisplaySelection,
+      isPointVisible,
+      isOverlayPointVisible,
+      effectiveActiveHwTypes,
+      selectedPrecisions,
+      activeOverlayHwTypes,
+      getCssColor,
+      resolveColor,
+      showGradientLabels,
+      gradientColorByPoint,
+    ]);
 
-      // Label placement (greedy collision layout) depends on the visible set,
-      // so when labels are shown, re-run the rooflines layer render — UNLESS
-      // an entrance transition is still pending/running, because that render
-      // also rewrites roofline `d` and would defeat the animation. In that
-      // case the in-flight render was produced with current interaction state
-      // anyway; the direct writes above keep visibility correct.
+    // Trace presence resolves asynchronously after the points render. Stamp
+    // only the tooltip metadata it owns, without restyling any chart marks.
+    useLayoutEffect(() => {
+      const display = getDisplaySelection();
+      if (!display) return;
+      display.zoomGroup.selectAll<SVGGElement, InferenceData>('.dot-group').each(function (d) {
+        const pointId = isPersistedBenchmarkId(d.id) ? d.id : null;
+        const persistedAgenticPoint = d.benchmark_type === 'agentic_traces' && pointId !== null;
+        d3.select(this)
+          .attr(
+            'data-has-trace',
+            pointId !== null && traceAvailability?.[pointId] === true ? 'true' : null,
+          )
+          .attr(
+            'data-trace-availability',
+            persistedAgenticPoint ? (isTraceAvailabilityPending ? 'pending' : 'resolved') : null,
+          );
+      });
+    }, [getDisplaySelection, dataIdentity, traceAvailability, isTraceAvailabilityPending]);
+
+    // Label-mode controls own point-label visibility and collision placement.
+    // Line and gradient labels are collision obstacles whose custom layer
+    // renders once through D3's display phase. Replaying that layer before the
+    // point-label collision pass preserves its stable placement behavior
+    // without restyling unrelated point marks. Overlay label visibility remains
+    // owned by the overlay custom layer's matching selective display callback.
+    useLayoutEffect(() => {
+      const visibilityChanged = lastPointLabelsVisibleRef.current !== pointLabelsVisible;
+      const lineLabelsChanged = lastShowLineLabelsRef.current !== showLineLabels;
+      const gradientLabelsChanged = lastShowGradientLabelsRef.current !== showGradientLabels;
+      if (!visibilityChanged && !lineLabelsChanged && !gradientLabelsChanged) return;
+      lastPointLabelsVisibleRef.current = pointLabelsVisible;
+      lastShowLineLabelsRef.current = showLineLabels;
+      lastShowGradientLabelsRef.current = showGradientLabels;
+      const display = getDisplaySelection();
+      if (!display) return;
+      const { svg, ctx, zoomGroup } = display;
+      if (visibilityChanged) {
+        zoomGroup
+          .selectAll<SVGTextElement, unknown>('.dot-group .point-label')
+          .style('display', pointLabelsVisible ? '' : 'none')
+          .style('opacity', pointLabelsVisible ? 1 : 0);
+      }
+      if ((lineLabelsChanged || gradientLabelsChanged) && (showLineLabels || showGradientLabels)) {
+        const entranceInFlight = zoomGroup
+          .selectAll<SVGPathElement, unknown>('.roofline-path')
+          .nodes()
+          .some((node) => hasNamedTransition(node, 'data-update'));
+        const rooflineLayer = layersRef.current.find((layer) => layer.key === 'rooflines');
+        if (!entranceInFlight && rooflineLayer?.type === 'custom' && rooflineLayer.render) {
+          rooflineLayer.render(zoomGroup, currentZoomRenderContext(svg, ctx));
+        }
+      }
+      if (pointLabelsVisible) avoidPointLabelCollisions(zoomGroup);
+    }, [getDisplaySelection, pointLabelsVisible, showGradientLabels, showLineLabels]);
+
+    // Visibility and palette changes can alter label collision placement and
+    // line-label colors. Label mode changes are handled by each custom layer's
+    // display identity, so they do not also enter this pass.
+    useLayoutEffect(() => {
+      const display = getDisplaySelection();
+      if (!display) return;
+      const labelDisplay = labelDisplayRef.current;
+      const { svg, ctx, zoomGroup } = display;
       const entranceInFlight = zoomGroup
         .selectAll<SVGPathElement, unknown>('.roofline-path')
         .nodes()
         .some((node) => hasNamedTransition(node, 'data-update'));
 
-      // Current (possibly zoomed) scales for layer re-renders — same scales
-      // the zoom handler would use.
-      const t = d3.zoomTransform(svg);
-      const zoomed = t.k !== 1 || t.x !== 0 || t.y !== 0;
-      const xScale = zoomed ? t.rescaleX(ctx.xScale as ContinuousScale) : ctx.xScale;
-      const yScale = zoomed ? t.rescaleY(ctx.yScale as ContinuousScale) : ctx.yScale;
-      const decorationCtx: RenderContext = { ...ctx, xScale, yScale };
-
-      const layerByKey = (key: string) => layersRef.current.find((l) => l.key === key);
-      if ((showGradientLabels || showLineLabels) && !entranceInFlight) {
-        const rooflineLayer = layerByKey('rooflines');
+      if ((labelDisplay.showGradientLabels || labelDisplay.showLineLabels) && !entranceInFlight) {
+        const rooflineLayer = layersRef.current.find((layer) => layer.key === 'rooflines');
         if (rooflineLayer?.type === 'custom' && rooflineLayer.render) {
-          rooflineLayer.render(zoomGroup, decorationCtx);
+          rooflineLayer.render(zoomGroup, currentZoomRenderContext(svg, ctx));
         }
       }
-
-      // Known-issue annotations follow the visible series; their layer writes
-      // no animated attributes, so re-rendering is always safe.
-      const knownIssueLayer = layerByKey('known-issues');
-      if (knownIssueLayer?.type === 'custom' && knownIssueLayer.render) {
-        knownIssueLayer.render(zoomGroup, decorationCtx);
+      if (labelDisplay.showPointLabels && !labelDisplay.showGradientLabels) {
+        avoidPointLabelCollisions(zoomGroup);
       }
     }, [
+      getDisplaySelection,
       isPointVisible,
       isOverlayPointVisible,
       effectiveActiveHwTypes,
@@ -3074,12 +3225,22 @@ const ScatterGraph = React.memo(
       activeOverlayHwTypes,
       getCssColor,
       resolveColor,
-      knownIssueAnnotations,
-      showGradientLabels,
-      showLineLabels,
-      gradientColorByPoint,
     ]);
 
+    // Known-issue annotations have their own data and theme inputs. Re-render
+    // only that layer when either changes.
+    useLayoutEffect(() => {
+      const display = getDisplaySelection();
+      if (!display) return;
+      const knownIssueLayer = layersRef.current.find((layer) => layer.key === 'known-issues');
+      if (knownIssueLayer?.type === 'custom' && knownIssueLayer.render) {
+        knownIssueLayer.render(
+          display.zoomGroup,
+          currentZoomRenderContext(display.svg, display.ctx),
+        );
+      }
+    }, [getDisplaySelection, knownIssueAnnotations, getCssColor, resolveColor]);
+
     // D3 custom layers are keyed additions, so removing the overlay layer from
     // the config does not delete DOM that the previous render created. Clear
     // those marks explicitly when the last unofficial run is dismissed.
@@ -3092,6 +3253,50 @@ const ScatterGraph = React.memo(
         .remove();
     }, [overlayData]);
 
+    // Perf-ruler decorations: refresh the curve hit strokes and redraw all
+    // rulers whenever the mode, ruler state, visibility filters, or
+    // underlying data change. Narrow mutation scope — only the hit layer
+    // and the ruler groups are touched; the chart itself never rebuilds for
+    // a ruler interaction. Redraws use the currently applied zoom transform
+    // (mode, rulers, and live iso-x are read through refs in drawPerfRuler).
+    //
+    // Ordering matters twice here — this effect is deliberately declared
+    // AFTER both (a) the mark-styling effect that writes curve opacities
+    // from the legend/precision filters, so hiding a curve clears the ruler,
+    // label, arrow, and halos in the SAME commit paint (no lingering frame),
+    // and (b) the overlay-cleanup effect above, so a dismissed overlay's
+    // paths are already gone when the ruler re-resolves.
+    useLayoutEffect(() => {
+      const display = getDisplaySelection();
+      if (!display) return;
+      const zoomCtx = currentZoomRenderContext(display.svg, display.ctx);
+      drawPerfRuler(
+        display.zoomGroup,
+        zoomCtx.xScale as ContinuousScale,
+        zoomCtx.yScale as ContinuousScale,
+        display.ctx.width,
+        display.ctx.height,
+      );
+      // Hidden-but-present curves KEEP their rulers (a legend re-toggle
+      // brings a hidden ruler back); curves whose paths left the DOM
+      // entirely are truly gone from the data, so prune each ruler (and the
+      // draft) that references one. `prunePerfRulers` bails out with the
+      // same reference when nothing changed.
+      setPerfRulerState((prev) =>
+        prunePerfRulers(prev, (cls) => !display.zoomGroup.select(`.${CSS.escape(cls)}`).empty()),
+      );
+    }, [
+      getDisplaySelection,
+      perfRulerMode,
+      perfRulerState,
+      drawPerfRuler,
+      dataIdentity,
+      effectiveActiveHwTypes,
+      selectedPrecisions,
+      hideNonOptimal,
+      overlayData,
+    ]);
+
     // Dismiss tooltip on filter changes
     useEffect(() => {
       chartRef.current?.dismissTooltip();
@@ -3136,8 +3341,26 @@ const ScatterGraph = React.memo(
               <p className="text-xs">
                 Please change the model, sequence, precision, date range or chip selection.
               </p>
+              <Button
+                type="button"
+                size="sm"
+                variant="outline"
+                className="mt-4"
+                data-testid="scatter-empty-quick-filters"
+                onClick={() => {
+                  setQuickFiltersOpen(true);
+                  track('inference_quick_filters_dialog_opened', { source: 'scatter_empty' });
+                }}
+              >
+                {legendT.quickFilters(quickFilterCount)}
+              </Button>
             </div>
           </div>
+          <QuickFiltersDialog
+            open={quickFiltersOpen}
+            onOpenChange={setQuickFiltersOpen}
+            bestPerSku={{ checked: bestPerSku, onCheckedChange: handleBestPerSkuChange }}
+          />
         </div>
       );
     }
@@ -3151,6 +3374,8 @@ const ScatterGraph = React.memo(
           // changed" rebuilds; scale domains come from x/yScaleConfig (computed
           // from the visible points), and visibility is applied via opacity.
           data={pointsData}
+          dataIdentity={dataIdentity}
+          metricIdentity={metricIdentity}
           margin={CHART_MARGIN}
           watermark={getChartWatermark(isUnofficialRun)}
           testId="scatter-graph"
@@ -3283,43 +3508,13 @@ const ScatterGraph = React.memo(
                 track('latency_legend_expanded', { expanded });
               }}
               switches={[
-                {
-                  id: 'scatter-best-per-sku',
-                  label: legendT.bestPerSku,
-                  checked: bestPerSku,
-                  onCheckedChange: (checked: boolean) => {
-                    setBestPerSku(checked);
-                    if (overlayData) {
-                      if (checked) {
-                        const direction =
-                          chartDefinition[
-                            `${selectedYAxisMetric}_roofline` as keyof ChartDefinition
-                          ];
-                        if (
-                          direction === 'upper_right' ||
-                          direction === 'upper_left' ||
-                          direction === 'lower_left' ||
-                          direction === 'lower_right'
-                        ) {
-                          const selection = bestSeriesPerSku(data, direction);
-                          for (const key of bestSeriesPerSku(overlayData.data, direction)) {
-                            selection.add(`overlay:${key}`);
-                          }
-                          commitUnifiedSelection(selection);
-                        }
-                      } else {
-                        resetUnifiedSelection();
-                      }
-                    }
-                    track('inference_best_per_sku_toggled', { enabled: checked });
-                  },
-                },
                 ...(selectedYAxisMetric === 'y_inputTputPerGpu'
                   ? []
                   : [
                       {
                         id: 'scatter-log-scale',
                         label: legendT.logScale,
+                        advanced: true,
                         checked: logScale,
                         onCheckedChange: (checked: boolean) => {
                           setLogScale(checked);
@@ -3345,6 +3540,7 @@ const ScatterGraph = React.memo(
                 {
                   id: 'scatter-point-labels',
                   label: legendT.labels,
+                  advanced: true,
                   checked: showPointLabels,
                   onCheckedChange: (checked: boolean) => {
                     setShowPointLabels(checked);
@@ -3354,6 +3550,7 @@ const ScatterGraph = React.memo(
                 {
                   id: 'scatter-high-contrast',
                   label: legendT.highContrast,
+                  advanced: true,
                   checked: highContrast,
                   onCheckedChange: (checked: boolean) => {
                     setHighContrast(checked);
@@ -3363,6 +3560,7 @@ const ScatterGraph = React.memo(
                 {
                   id: 'scatter-parallelism-labels',
                   label: legendT.parallelismLabels,
+                  advanced: true,
                   checked: useAdvancedLabels,
                   onCheckedChange: (checked: boolean) => {
                     setUseAdvancedLabels(checked);
@@ -3400,6 +3598,7 @@ const ScatterGraph = React.memo(
                 {
                   id: 'scatter-line-labels',
                   label: legendT.lineLabels,
+                  advanced: true,
                   checked: showLineLabels,
                   onCheckedChange: (checked: boolean) => {
                     setShowLineLabels(checked);
@@ -3407,57 +3606,88 @@ const ScatterGraph = React.memo(
                   },
                 },
                 {
-                  id: 'scatter-speed-overlay',
-                  label: 'Bus / Race Car',
+                  id: 'scatter-perf-ruler',
+                  label: legendT.perfRuler,
                   advanced: true,
-                  checked: showSpeedOverlay,
+                  checked: perfRulerMode,
+                  infoTooltip: legendT.perfRulerInfo,
                   onCheckedChange: (checked: boolean) => {
-                    setShowSpeedOverlay(checked);
-                    track('latency_speed_overlay_toggled', { enabled: checked });
+                    setPerfRulerMode(checked);
+                    // Toggle-off clears ALL rulers in the same state batch —
+                    // the pre-paint decoration effect then removes the rulers
+                    // and the curve hit strokes before the next frame (no
+                    // lingering lines after toggle-off).
+                    if (!checked) setPerfRulerState(clearPerfRulers);
+                    track('latency_perf_ruler_toggled', { enabled: checked });
                   },
                 },
                 {
-                  id: 'scatter-minecraft-overlay',
-                  label: 'Donkey / Elytra',
+                  id: 'scatter-concurrency-labels',
+                  label: legendT.concurrencyLabels,
                   advanced: true,
-                  checked: showMinecraftOverlay,
+                  checked: showConcurrencyLabels,
                   onCheckedChange: (checked: boolean) => {
-                    setShowMinecraftOverlay(checked);
-                    track('latency_minecraft_overlay_toggled', { enabled: checked });
+                    setShowConcurrencyLabels(checked);
+                    track('latency_concurrency_labels_toggled', { enabled: checked });
+                    // Concurrency is a point-label annotation; turning it on is
+                    // pointless if labels are hidden, so auto-enable Labels.
+                    if (checked && !showPointLabels) setShowPointLabels(true);
                   },
                 },
               ]}
               onAdvancedExpandedChange={(expanded) => {
                 track('latency_advanced_controls_toggled', { expanded });
               }}
-              actions={
-                effectiveOfficialHwTypes.size < hwTypesWithData.size ||
-                activeOverlayHwTypes.size < scopedOverlayHwTypes.size
+              actions={[
+                ...(effectiveOfficialHwTypes.size < hwTypesWithData.size ||
+                activeOverlayHwTypes.size < scopedOverlayHwTypes.size ||
+                quickFilterCount > 0
                   ? [
                       {
                         id: 'scatter-reset-filter',
                         label: legendT.resetFilter,
                         onClick: () => {
                           resetUnifiedSelection();
+                          clearQuickFilters();
                           track('latency_legend_filter_reset');
                         },
                       },
                     ]
-                  : []
-              }
+                  : []),
+                ...(perfRulerState.rulers.length > 0
+                  ? [
+                      {
+                        id: 'scatter-clear-perf-rulers',
+                        label: legendT.clearPerfRulers(perfRulerState.rulers.length),
+                        onClick: () => {
+                          track('latency_perf_ruler_cleared', {
+                            count: perfRulerState.rulers.length,
+                          });
+                          setPerfRulerState(clearPerfRulers);
+                        },
+                      },
+                    ]
+                  : []),
+                {
+                  id: 'scatter-quick-filters',
+                  label: legendT.quickFilters(quickFilterCount),
+                  onClick: () => {
+                    setQuickFiltersOpen(true);
+                    track('inference_quick_filters_dialog_opened', { source: 'scatter_legend' });
+                  },
+                },
+              ]}
               precisionIndicators={selectedPrecisions}
-              keyIndicators={
-                hasOffloadHalo || selectedSequence === Sequence.AgenticTraces ? (
-                  <>
-                    {hasOffloadHalo && <OffloadHaloLegendKey />}
-                    {selectedSequence === Sequence.AgenticTraces && <AgenticOptimizationNote />}
-                  </>
-                ) : undefined
-              }
+              hideAtomFootnote
               enableTooltips={true}
             />
           }
         />
+        <QuickFiltersDialog
+          open={quickFiltersOpen}
+          onOpenChange={setQuickFiltersOpen}
+          bestPerSku={{ checked: bestPerSku, onCheckedChange: handleBestPerSkuChange }}
+        />
         {pointsTable && (
           <LegendPointsDialog
             open
@@ -3478,6 +3708,14 @@ const ScatterGraph = React.memo(
             }
           />
         )}
+        {fixedLogPointId === null ? null : (
+          <FixedSequenceLogDialog
+            pointId={fixedLogPointId}
+            onOpenChange={(open) => {
+              if (!open) setFixedLogPointId(null);
+            }}
+          />
+        )}
       </>
     );
   },
diff --git a/packages/app/src/components/inference/ui/TrendChart.tsx b/packages/app/src/components/inference/ui/TrendChart.tsx
index fcbfd45c5..25b232697 100644
--- a/packages/app/src/components/inference/ui/TrendChart.tsx
+++ b/packages/app/src/components/inference/ui/TrendChart.tsx
@@ -173,9 +173,9 @@ const TrendChart = React.memo(
 
     // Apply logTickFormat to the y-axis after render (needs the built scale)
     const applyLogTickFormat = useCallback(
-      (ctx: Pick<RenderContext, 'layout' | 'yScale'>) => {
+      (ctx: Pick<RenderContext, 'layout' | 'yScale' | 'renderedYScale'>) => {
         if (!logScale) return;
-        const scale = ctx.yScale as d3.ScaleLogarithmic<number, number>;
+        const scale = (ctx.renderedYScale ?? ctx.yScale) as d3.ScaleLogarithmic<number, number>;
         const fmt = logTickFormat(scale);
         ctx.layout.yAxisGroup.call(d3.axisLeft(scale).ticks(8).tickFormat(fmt));
       },
@@ -229,7 +229,7 @@ const TrendChart = React.memo(
         rulerType: 'crosshair' as const,
         content: (d: PreparedPoint, isPinned: boolean) =>
           `<div class="rounded-md border bg-background/95 px-3 py-2 text-xs shadow-md backdrop-blur-sm" style="min-width: 160px; user-select: ${isPinned ? 'text' : 'none'};">
-            ${isPinned ? '<div class="text-muted-foreground text-[10px] mb-1 italic">Click elsewhere to dismiss</div>' : ''}
+            ${isPinned ? '<div class="text-muted-foreground text-3xs mb-1 italic">Click elsewhere to dismiss</div>' : ''}
             <div class="font-semibold mb-1" style="color: ${getPointConfig(d)?.color ?? '#888'}">${getPointConfig(d)?.label ?? ''}</div>
             <div class="text-muted-foreground">${d.raw.date}</div>
             <div class="mt-1 font-medium">${yLabel}: ${formatLargeNumber(d.value)}</div>
diff --git a/packages/app/src/components/inference/ui/UnofficialChartDisplay.tsx b/packages/app/src/components/inference/ui/UnofficialChartDisplay.tsx
deleted file mode 100644
index f18903ea6..000000000
--- a/packages/app/src/components/inference/ui/UnofficialChartDisplay.tsx
+++ /dev/null
@@ -1,211 +0,0 @@
-'use client';
-
-import { useEffect, useMemo } from 'react';
-
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
-import { useInference } from '@/components/inference/InferenceContext';
-import type {
-  ChartDefinition,
-  HardwareConfig,
-  InferenceData,
-  RenderableGraph,
-  YAxisMetricKey,
-} from '@/components/inference/types';
-import { Badge } from '@/components/ui/badge';
-import { Card } from '@/components/ui/card';
-import { Skeleton } from '@/components/ui/skeleton';
-import { useUnofficialRun } from '@/components/unofficial-run-provider';
-import { getHardwareConfig } from '@/lib/constants';
-
-import ScatterGraph from './ScatterGraph';
-
-export function UnofficialChartDisplay() {
-  const { unofficialChartData, unofficialRunInfo, loading, error, availableModelsAndSequences } =
-    useUnofficialRun();
-  const {
-    selectedModel,
-    selectedSequence,
-    selectedPrecisions,
-    selectedYAxisMetric,
-    setSelectedModel,
-    setSelectedSequence,
-  } = useInference();
-
-  // Auto-select a model/sequence that has data when unofficial run loads
-  useEffect(() => {
-    if (availableModelsAndSequences.length > 0 && unofficialChartData) {
-      // Check if current selection has data
-      const currentHasData = availableModelsAndSequences.some(
-        (item) => item.model === selectedModel && item.sequence === selectedSequence,
-      );
-
-      if (!currentHasData) {
-        // Find the first available combination
-        const first = availableModelsAndSequences[0];
-        setSelectedModel(first.model);
-        setSelectedSequence(first.sequence);
-      }
-    }
-  }, [
-    availableModelsAndSequences,
-    unofficialChartData,
-    selectedModel,
-    selectedSequence,
-    setSelectedModel,
-    setSelectedSequence,
-  ]);
-
-  // Generate the key to look up unofficial data
-  const dataKey = useMemo(
-    () => `${selectedModel}_${selectedSequence}`,
-    [selectedModel, selectedSequence],
-  );
-
-  // Create graphs with hardware config for unofficial data
-  interface UnofficialGraph extends RenderableGraph {
-    hardwareConfig: HardwareConfig;
-  }
-
-  const graphs: UnofficialGraph[] = useMemo(() => {
-    if (!unofficialChartData || !unofficialChartData[dataKey]) {
-      return [];
-    }
-
-    const chartData = unofficialChartData[dataKey];
-
-    return (chartDefinitions as ChartDefinition[]).map((chartDef) => {
-      const dataForChart = chartDef.chartType === 'e2e' ? chartData.e2e : chartData.interactivity;
-
-      const metricKey = selectedYAxisMetric.replace('y_', '') as YAxisMetricKey;
-
-      const processedData =
-        dataForChart.data.length > 0 && metricKey in dataForChart.data[0]
-          ? dataForChart.data.map((d: InferenceData) => {
-              const yValue = (d[metricKey] as { y: number })?.y || d.y;
-              const roof = (d[metricKey] as { roof: boolean })?.roof ?? false;
-
-              return {
-                ...d,
-                y: yValue,
-                roof,
-              };
-            })
-          : [];
-
-      const yLabelKey = `${selectedYAxisMetric}_label` as keyof ChartDefinition;
-      const dynamicYLabel = chartDef[yLabelKey];
-
-      return {
-        model: selectedModel,
-        sequence: selectedSequence,
-        chartDefinition: {
-          ...chartDef,
-          y_label: dynamicYLabel === null ? undefined : String(dynamicYLabel),
-        },
-        data: processedData,
-        hardwareConfig: Object.fromEntries(
-          Object.entries(dataForChart.gpus || {}).map(([k, v]) => [
-            k,
-            { ...getHardwareConfig(k, selectedModel), ...v },
-          ]),
-        ),
-      };
-    });
-  }, [unofficialChartData, dataKey, selectedYAxisMetric, selectedModel, selectedSequence]);
-
-  if (loading) {
-    return (
-      <section>
-        <Card>
-          <div className="flex items-center gap-2 mb-4">
-            <Skeleton className="h-6 w-32" />
-            <Skeleton className="h-6 w-48" />
-          </div>
-          <Skeleton className="h-[400px] w-full" />
-        </Card>
-      </section>
-    );
-  }
-
-  if (error) {
-    return (
-      <section>
-        <Card>
-          <div className="text-red-600 p-4">
-            <h3 className="font-semibold">Error loading unofficial run data</h3>
-            <p className="text-sm">{error}</p>
-          </div>
-        </Card>
-      </section>
-    );
-  }
-
-  if (!unofficialChartData || graphs.every((g) => g.data.length === 0)) {
-    return (
-      <section>
-        <Card>
-          <div className="text-muted-foreground p-4">
-            <h3 className="font-semibold">No data available</h3>
-            <p className="text-sm">
-              No benchmark data found for the selected model ({selectedModel}) and sequence (
-              {selectedSequence}) in this unofficial run.
-            </p>
-            <p className="text-sm mt-2">
-              Available data keys: {Object.keys(unofficialChartData || {}).join(', ') || 'none'}
-            </p>
-          </div>
-        </Card>
-      </section>
-    );
-  }
-
-  return (
-    <>
-      {graphs.map((graph, graphIndex) => (
-        <section key={graphIndex}>
-          <figure>
-            <Card>
-              <ScatterGraph
-                chartId={`unofficial-chart-${graphIndex}`}
-                modelLabel={graph.model}
-                data={graph.data}
-                xLabel={graph.chartDefinition.x_label}
-                yLabel={`${
-                  graph.chartDefinition[
-                    `${selectedYAxisMetric}_label` as keyof typeof graph.chartDefinition
-                  ]
-                }`}
-                chartDefinition={graph.chartDefinition}
-                showAllHardwareTypes={true}
-                hardwareConfigOverride={graph.hardwareConfig}
-                caption={
-                  <>
-                    <div className="flex items-center gap-2 mb-2">
-                      <Badge variant="destructive" className="text-lg px-3 py-1">
-                        NON-OFFICIAL
-                      </Badge>
-                      <span className="text-sm text-muted-foreground">
-                        Branch: {unofficialRunInfo?.branch}
-                      </span>
-                    </div>
-                    <h2 className="text-lg font-semibold">
-                      {
-                        graph.chartDefinition[
-                          `${selectedYAxisMetric}_title` as keyof typeof graph.chartDefinition
-                        ]
-                      }{' '}
-                      {graph.chartDefinition.heading}
-                    </h2>
-                    <p className="text-sm text-muted-foreground mb-2">
-                      {graph.model} • {selectedPrecisions.join(', ')} • {graph.sequence}
-                    </p>
-                  </>
-                }
-              />
-            </Card>
-          </figure>
-        </section>
-      ))}
-    </>
-  );
-}
diff --git a/packages/app/src/components/inference/ui/WorkflowInfoDisplay.tsx b/packages/app/src/components/inference/ui/WorkflowInfoDisplay.tsx
index 8458d5976..1a2f25955 100644
--- a/packages/app/src/components/inference/ui/WorkflowInfoDisplay.tsx
+++ b/packages/app/src/components/inference/ui/WorkflowInfoDisplay.tsx
@@ -14,9 +14,12 @@ import {
 } from '@/components/ui/select';
 import { updateRepoUrl } from '@/lib/utils';
 
-import { useGlobalFilters } from '@/components/GlobalFilterContext';
-import { useInference } from '@/components/inference/InferenceContext';
-import type { WorkflowInfo } from '@/components/inference/types';
+import { useGlobalFilterSelection } from '@/components/GlobalFilterContext';
+import {
+  useInferenceActions,
+  useInferenceData,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
 import {
   formatChangelogDescription,
   formatConfigKeys,
@@ -48,25 +51,15 @@ function RunConclusionDot({ conclusion }: { conclusion: string | null }) {
   );
 }
 
-export default function WorkflowInfoDisplay({
-  workflowInfo,
-}: {
-  workflowInfo: WorkflowInfo[] | null;
-}) {
-  const {
-    selectedRunDate,
-    setSelectedRunDate,
-    availableDates,
-    availableRuns,
-    selectedRunId,
-    setSelectedRunId,
-    isCheckingAvailableDates,
-  } = useInference();
+export default function WorkflowInfoDisplay() {
+  const { selectedRunDate, selectedRunId } = useInferenceFilters();
+  const { availableDates, availableRuns, isCheckingAvailableDates } = useInferenceData();
+  const { setSelectedRunDate, setSelectedRunId } = useInferenceActions();
 
-  const { effectivePrecisions } = useGlobalFilters();
+  const { effectivePrecisions } = useGlobalFilterSelection();
 
   // Navigation functions for runs
-  const runIds = availableRuns ? Object.keys(availableRuns) : [];
+  const runIds = Object.keys(availableRuns);
   const currentRunIndex = runIds.indexOf(selectedRunId);
 
   const canGoPreviousRun = () => currentRunIndex > 0;
@@ -90,7 +83,7 @@ export default function WorkflowInfoDisplay({
     }
   };
 
-  if (!workflowInfo || workflowInfo.length === 0 || !workflowInfo[0]) {
+  if (runIds.length === 0) {
     return (
       <div className="flex flex-col lg:flex-row gap-2 text-muted-foreground">
         <DatePicker
@@ -104,7 +97,7 @@ export default function WorkflowInfoDisplay({
   }
 
   const changelog = (() => {
-    const raw = availableRuns ? availableRuns[selectedRunId]?.changelog || null : null;
+    const raw = availableRuns[selectedRunId]?.changelog || null;
     if (!raw) return null;
     // Filter config_keys by selected precisions, drop entries with no matching keys
     const filtered = raw.entries
@@ -121,17 +114,13 @@ export default function WorkflowInfoDisplay({
 
   return (
     <div className="flex flex-wrap gap-2 lg:gap-4 text-muted-foreground">
-      {/* <div className="flex items-center gap-2">
-        <CalendarRange size={16} />
-        <strong>Run Date:</strong> {workflowInfo[0].run_date} UTC
-      </div> */}
       <DatePicker
         date={selectedRunDate}
         onChange={(date) => setSelectedRunDate(date)}
         availableDates={availableDates}
         isCheckingAvailableDates={isCheckingAvailableDates}
       />
-      {availableRuns ? (
+      {runIds.length > 0 ? (
         <div className="flex items-center gap-0.5">
           <Button
             variant="ghost"
diff --git a/packages/app/src/components/inference/ui/line-label-anchor.ts b/packages/app/src/components/inference/ui/line-label-anchor.ts
index 89a7cfcb1..1af00c008 100644
--- a/packages/app/src/components/inference/ui/line-label-anchor.ts
+++ b/packages/app/src/components/inference/ui/line-label-anchor.ts
@@ -1,5 +1,3 @@
-import type { InferenceData } from '@/components/inference/types';
-
 /**
  * Find the point on a polyline whose x is closest to a target data-space x.
  *
@@ -11,7 +9,7 @@ import type { InferenceData } from '@/components/inference/types';
  *
  * `pts` is assumed non-empty (callers guard with `pts.length >= 2`).
  */
-export const pointNearestX = (pts: InferenceData[], targetX: number): InferenceData => {
+export const pointNearestX = <T extends { x: number }>(pts: readonly T[], targetX: number): T => {
   let best = pts[0];
   let bestDist = Math.abs(pts[0].x - targetX);
   for (let i = 1; i < pts.length; i++) {
diff --git a/packages/app/src/components/inference/ui/line-label-layer.test.ts b/packages/app/src/components/inference/ui/line-label-layer.test.ts
new file mode 100644
index 000000000..451881c2a
--- /dev/null
+++ b/packages/app/src/components/inference/ui/line-label-layer.test.ts
@@ -0,0 +1,117 @@
+import { describe, expect, it } from 'vitest';
+import * as d3 from 'd3';
+
+import {
+  firstNonCollidingRect,
+  placeEndpointLineLabels,
+  placeLineLabels,
+  rectsOverlap,
+  type LineLabelSeries,
+} from './line-label-layer';
+
+interface Point {
+  x: number;
+  y: number;
+}
+
+const series = (
+  key: string,
+  points: Point[],
+  keepVisibleOnCollision = false,
+): LineLabelSeries<Point> => ({
+  key,
+  seriesId: key,
+  label: key,
+  color: '#000',
+  points,
+  keepVisibleOnCollision,
+});
+
+const identity = (value: number) => value;
+
+describe('line-label collision primitives', () => {
+  it('treats edge-separated rectangles as non-overlapping', () => {
+    const left = { left: 0, right: 10, top: 0, bottom: 10 };
+    const right = { left: 11, right: 20, top: 0, bottom: 10 };
+
+    expect(rectsOverlap(left, right)).toBe(false);
+    expect(firstNonCollidingRect([left, right], [left])).toBe(1);
+  });
+
+  it('returns null when every candidate intersects a placed rectangle', () => {
+    const placed = [{ left: 0, right: 10, top: 0, bottom: 10 }];
+    const candidates = [
+      { left: 2, right: 4, top: 2, bottom: 4 },
+      { left: 8, right: 12, top: 8, bottom: 12 },
+    ];
+
+    expect(firstNonCollidingRect(candidates, placed)).toBeNull();
+  });
+});
+
+describe('line-label placement', () => {
+  it('uses later candidates when the preferred anchor collides', () => {
+    const labels = placeLineLabels(
+      [
+        series('first', [
+          { x: 0, y: 0 },
+          { x: 10, y: 10 },
+        ]),
+        series('second', [
+          { x: 0, y: 0 },
+          { x: 20, y: 20 },
+          { x: 40, y: 40 },
+        ]),
+      ],
+      identity,
+      identity,
+      { collisionWidth: 15, collisionHeight: 15 },
+    );
+
+    expect(labels.find((label) => label.key === 'first')).toMatchObject({ x: 10, y: 10 });
+    expect(labels.find((label) => label.key === 'second')).toMatchObject({
+      x: 40,
+      y: 40,
+      visible: true,
+    });
+  });
+
+  it('keeps pinned data-space anchors stable while scales change', () => {
+    const anchors = new Map<string, number>();
+    const input = [
+      series('run', [
+        { x: 10, y: 20 },
+        { x: 30, y: 40 },
+        { x: 50, y: 60 },
+      ]),
+    ];
+    const initial = placeLineLabels(input, identity, identity, {
+      collisionWidth: 100,
+      anchors,
+      pinAnchors: true,
+    });
+    const zoomed = placeLineLabels(
+      input,
+      (value) => value * 2,
+      (value) => value * 3,
+      { collisionWidth: 100, anchors, pinAnchors: true },
+    );
+
+    expect(anchors.get('run')).toBe(30);
+    expect(initial[0]).toMatchObject({ x: 30, y: 40, visible: true });
+    expect(zoomed[0]).toMatchObject({ x: 60, y: 120, visible: true });
+  });
+
+  it('nudges endpoint labels into the scale range without changing identities', () => {
+    const yScale = d3.scaleLinear().domain([0, 100]).range([100, 0]);
+    const labels = placeEndpointLineLabels(
+      [series('a', [{ x: 1, y: 50 }]), series('b', [{ x: 2, y: 51 }])],
+      identity,
+      yScale,
+    );
+
+    expect(labels.map((label) => label.key).toSorted()).toEqual(['a', 'b']);
+    expect(Math.abs(labels[0].y - labels[1].y)).toBeGreaterThanOrEqual(17.9);
+    expect(labels.every((label) => label.y >= 18 && label.y <= 82)).toBe(true);
+  });
+});
diff --git a/packages/app/src/components/inference/ui/line-label-layer.ts b/packages/app/src/components/inference/ui/line-label-layer.ts
new file mode 100644
index 000000000..bda7ff113
--- /dev/null
+++ b/packages/app/src/components/inference/ui/line-label-layer.ts
@@ -0,0 +1,486 @@
+import * as d3 from 'd3';
+
+import { pointNearestX } from '@/components/inference/ui/line-label-anchor';
+
+export interface CartesianPoint {
+  x: number;
+  y: number;
+}
+
+export interface LineLabelSeries<TPoint extends CartesianPoint> {
+  key: string;
+  seriesId: string;
+  label: string;
+  color: string;
+  points: readonly TPoint[];
+  keepVisibleOnCollision?: boolean;
+}
+
+export interface LineLabelPlacement {
+  key: string;
+  seriesId: string;
+  label: string;
+  color: string;
+  x: number;
+  y: number;
+  visible: boolean;
+}
+
+export interface RectBounds {
+  left: number;
+  right: number;
+  top: number;
+  bottom: number;
+}
+
+export function rectsOverlap(a: RectBounds, b: RectBounds): boolean {
+  return !(a.right < b.left || a.left > b.right || a.bottom < b.top || a.top > b.bottom);
+}
+
+export function firstNonCollidingRect(
+  candidates: readonly RectBounds[],
+  placed: readonly RectBounds[],
+): number | null {
+  for (let i = 0; i < candidates.length; i++) {
+    if (!placed.some((other) => rectsOverlap(candidates[i], other))) return i;
+  }
+  return null;
+}
+
+/** A drawn pill box, as centre point + half-width, for overlap tests. */
+export interface PlacedBox {
+  x: number;
+  y: number;
+  halfW: number;
+}
+
+/**
+ * Centre + half-width of every visible parallelism pill inside `root`.
+ *
+ * Line labels are placed greedily against a list of boxes they must not sit on
+ * top of. The pills are drawn before the label layers run, so seeding that
+ * list from the DOM is what keeps a run name off a "TP8 / PP2" chip.
+ */
+export function parallelismLabelBoxes(root: SVGGElement | null): PlacedBox[] {
+  const boxes: PlacedBox[] = [];
+  if (!root) return boxes;
+  for (const node of root.querySelectorAll<SVGGElement>('.parallelism-label')) {
+    // Hidden pills stay in the DOM but must not push labels around.
+    if (node.style.opacity === '0') continue;
+    const match = /translate\((?<tx>[^,]+),(?<ty>[^)]+)\)/u.exec(
+      node.getAttribute('transform') ?? '',
+    );
+    if (!match?.groups) continue;
+    const x = Number(match.groups.tx);
+    const y = Number(match.groups.ty);
+    if (!Number.isFinite(x) || !Number.isFinite(y)) continue;
+    const width = Number(node.querySelector('.pl-bg')?.getAttribute('width') ?? 0);
+    boxes.push({ x, y, halfW: (Number.isFinite(width) ? width : 0) / 2 });
+  }
+  return boxes;
+}
+
+/**
+ * Bounding boxes of the pills drawn over the plot: run-name labels
+ * (`.line-label`) and parallelism labels (`.parallelism-label`).
+ *
+ * Both are positioned by the roofline layer, which always runs before point
+ * label collision avoidance, so reading the DOM gives their final placement
+ * for the current frame. The group carries a `translate(...)` and the
+ * `.ll-bg`/`.pl-bg` rect carries an offset sized to its text; summing the two
+ * puts the box in the same space as a point label's centre.
+ */
+function pillObstacles(
+  zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+): RectBounds[] {
+  const boxes: RectBounds[] = [];
+  zoomGroup.selectAll<SVGGElement, unknown>('.line-label, .parallelism-label').each(function () {
+    // A hidden pill is still in the DOM but covers nothing.
+    if (this.style.opacity === '0') return;
+    const match = /translate\((?<tx>[^,]+),(?<ty>[^)]+)\)/u.exec(
+      this.getAttribute('transform') ?? '',
+    );
+    const background = this.querySelector<SVGRectElement>('.ll-bg, .pl-bg');
+    if (!match?.groups || !background) return;
+    const tx = Number(match.groups.tx);
+    const ty = Number(match.groups.ty);
+    const x = Number(background.getAttribute('x'));
+    const y = Number(background.getAttribute('y'));
+    const width = Number(background.getAttribute('width'));
+    const height = Number(background.getAttribute('height'));
+    if (![tx, ty, x, y, width, height].every((value) => Number.isFinite(value))) return;
+    boxes.push({ left: tx + x, top: ty + y, right: tx + x + width, bottom: ty + y + height });
+  });
+  return boxes;
+}
+
+function lineCandidates<TPoint extends CartesianPoint>(points: readonly TPoint[]): TPoint[] {
+  return [
+    points[Math.min(1, points.length - 1)],
+    points[Math.floor(points.length / 2)],
+    points[Math.max(0, Math.floor((points.length * 2) / 3))],
+    points.at(-1)!,
+  ];
+}
+
+export function placeLineLabels<TPoint extends CartesianPoint>(
+  series: readonly LineLabelSeries<TPoint>[],
+  xScale: (value: number) => number,
+  yScale: (value: number) => number,
+  options: {
+    collisionWidth: number;
+    collisionHeight?: number;
+    anchors?: Map<string, number>;
+    pinAnchors?: boolean;
+    /** Pill boxes already on the plot that labels must not cover. */
+    obstacles?: readonly PlacedBox[];
+  },
+): LineLabelPlacement[] {
+  const collisionHeight = options.collisionHeight ?? 18;
+  const placed: PlacedBox[] = [...(options.obstacles ?? [])];
+  const result: LineLabelPlacement[] = [];
+  const sorted = [...series].toSorted(
+    (a, b) => yScale(a.points[0]?.y ?? 0) - yScale(b.points[0]?.y ?? 0),
+  );
+
+  const labelHalfWidth = options.collisionWidth / 2;
+  const collides = (x: number, y: number) =>
+    placed.some(
+      (other) =>
+        Math.abs(other.y - y) < collisionHeight &&
+        Math.abs(other.x - x) < other.halfW + labelHalfWidth,
+    );
+
+  for (const entry of sorted) {
+    if (entry.points.length === 0) continue;
+    const candidates = lineCandidates(entry.points);
+
+    if (options.pinAnchors && options.anchors) {
+      let anchorX = options.anchors.get(entry.key);
+      if (anchorX === undefined) {
+        const candidate = candidates.find((point) => !collides(xScale(point.x), yScale(point.y)));
+        anchorX = (candidate ?? candidates.at(-1)!).x;
+        options.anchors.set(entry.key, anchorX);
+      }
+      const point = pointNearestX(entry.points, anchorX);
+      const x = xScale(point.x);
+      const y = yScale(point.y);
+      placed.push({ x, y, halfW: labelHalfWidth });
+      result.push({
+        key: entry.key,
+        seriesId: entry.seriesId,
+        label: entry.label,
+        color: entry.color,
+        x,
+        y,
+        visible: true,
+      });
+      continue;
+    }
+
+    const candidate = candidates.find((point) => !collides(xScale(point.x), yScale(point.y)));
+    if (candidate) {
+      const x = xScale(candidate.x);
+      const y = yScale(candidate.y);
+      placed.push({ x, y, halfW: labelHalfWidth });
+      result.push({
+        key: entry.key,
+        seriesId: entry.seriesId,
+        label: entry.label,
+        color: entry.color,
+        x,
+        y,
+        visible: true,
+      });
+      continue;
+    }
+
+    const fallback = entry.points[0];
+    const x = xScale(fallback.x);
+    const y = yScale(fallback.y);
+    const visible = entry.keepVisibleOnCollision === true;
+    if (visible) placed.push({ x, y, halfW: labelHalfWidth });
+    result.push({
+      key: entry.key,
+      seriesId: entry.seriesId,
+      label: entry.label,
+      color: entry.color,
+      x,
+      y,
+      visible,
+    });
+  }
+
+  return result;
+}
+
+export function placeEndpointLineLabels<TPoint extends CartesianPoint>(
+  series: readonly LineLabelSeries<TPoint>[],
+  xScale: (value: number) => number,
+  yScale: (value: number) => number,
+  options: { collisionHeight?: number; nudge?: boolean } = {},
+): LineLabelPlacement[] {
+  const collisionHeight = options.collisionHeight ?? 18;
+  const labels = series.flatMap((entry) => {
+    const point = entry.points.at(-1);
+    return point
+      ? [
+          {
+            key: entry.key,
+            seriesId: entry.seriesId,
+            label: entry.label,
+            color: entry.color,
+            x: xScale(point.x),
+            y: yScale(point.y),
+            visible: true,
+          },
+        ]
+      : [];
+  });
+
+  if (labels.length < 2 || options.nudge === false) return labels;
+
+  const range = yScaleRange(yScale);
+  if (!range) return labels;
+  const top = Math.min(range[0], range[1]) + collisionHeight;
+  const bottom = Math.max(range[0], range[1]) - collisionHeight;
+  labels.sort((a, b) => a.y - b.y);
+  for (let pass = 0; pass < 5; pass++) {
+    for (let i = 1; i < labels.length; i++) {
+      const overlap = labels[i - 1].y + collisionHeight - labels[i].y;
+      if (overlap <= 0) continue;
+      const half = overlap / 2;
+      labels[i - 1].y -= half;
+      labels[i].y += half;
+    }
+    for (const label of labels) label.y = Math.max(top, Math.min(bottom, label.y));
+  }
+  return labels;
+}
+
+function yScaleRange(scale: (value: number) => number): [number, number] | null {
+  const withRange = scale as ((value: number) => number) & { range?: () => unknown[] };
+  const range = withRange.range?.();
+  return range && range.length >= 2 && typeof range[0] === 'number' && typeof range[1] === 'number'
+    ? [range[0], range[1]]
+    : null;
+}
+
+/** Full-color icon rendered on a white chip at the left edge of a pill. */
+export interface LineLabelIconSpec {
+  href: string;
+  width: number;
+  height: number;
+}
+
+/** Gap between the icon and the label text. */
+const ICON_TEXT_GAP = 4;
+
+/** Horizontal space an icon occupies to the left of the label text. */
+function iconSpace(icon: LineLabelIconSpec | undefined): number {
+  return icon ? icon.width + ICON_TEXT_GAP : 0;
+}
+
+export function renderLineLabels(
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  labels: readonly LineLabelPlacement[],
+  options: {
+    seriesAttribute: string;
+    opacity?: number;
+    offsetX?: number;
+    offsetY?: number;
+    /** Optional full-color icon (e.g. vendor mark) per label. */
+    iconFor?: (label: LineLabelPlacement) => LineLabelIconSpec | undefined;
+    configureText?: (
+      text: d3.Selection<SVGTextElement, LineLabelPlacement, null, undefined>,
+      label: LineLabelPlacement,
+    ) => void;
+    configureGroup?: (
+      labelGroup: d3.Selection<SVGGElement, LineLabelPlacement, null, undefined>,
+      label: LineLabelPlacement,
+    ) => void;
+  },
+): void {
+  const opacity = options.opacity ?? 1;
+  const offsetX = options.offsetX ?? 8;
+  const offsetY = options.offsetY ?? -14;
+  const selection = group
+    .selectAll<SVGGElement, LineLabelPlacement>('.line-label')
+    .data(labels, (label) => label.key)
+    .join(
+      (enter) => {
+        const labelGroup = enter
+          .append('g')
+          .attr('class', 'line-label')
+          .style('pointer-events', 'none');
+        labelGroup.append('rect').attr('class', 'll-bg').attr('rx', 4).attr('ry', 4);
+        labelGroup
+          .append('text')
+          .attr('class', 'll-text')
+          .attr('text-anchor', 'start')
+          .attr('dominant-baseline', 'central')
+          .attr('fill', 'white')
+          .attr('font-size', '10px')
+          .attr('font-weight', '600');
+        return labelGroup;
+      },
+      (update) => update,
+      (exit) => exit.remove(),
+    )
+    .attr('data-line-key', (label) => label.key)
+    .attr(options.seriesAttribute, (label) => label.seriesId)
+    .attr('transform', (label) => `translate(${label.x + offsetX},${label.y + offsetY})`)
+    .style('transition', 'opacity 150ms ease')
+    .style('opacity', (label) => (label.visible ? opacity : 0));
+
+  selection.each(function (label) {
+    const labelGroup = d3.select<SVGGElement, LineLabelPlacement>(this);
+    options.configureGroup?.(labelGroup, label);
+    const text = labelGroup.select<SVGTextElement>('.ll-text');
+    // Shift the text right to leave room for the icon chip; the background
+    // sizing pass below expands the pill back over that space.
+    const space = iconSpace(options.iconFor?.(label));
+    if (space > 0) text.attr('x', space);
+    else text.attr('x', null);
+    if (options.configureText) options.configureText(text, label);
+    else text.text(label.label);
+  });
+
+  const measured: { node: SVGGElement; label: LineLabelPlacement; bbox: DOMRect }[] = [];
+  selection.each(function (label) {
+    const text = this.querySelector<SVGTextElement>('.ll-text');
+    if (text) measured.push({ node: this, label, bbox: text.getBBox() });
+  });
+  for (const { node, label, bbox } of measured) {
+    const labelGroup = d3.select(node);
+    const icon = options.iconFor?.(label);
+    const space = iconSpace(icon);
+    labelGroup
+      .select('.ll-bg')
+      .attr('x', bbox.x - space - 5)
+      .attr('y', bbox.y - 3)
+      .attr('width', bbox.width + space + 10)
+      .attr('height', bbox.height + 6)
+      .attr('fill', label.color);
+
+    // Full-color mark drawn directly on the pill — transparent background, so
+    // the icon shares the label's own fill shade (including gradient fills).
+    labelGroup
+      .selectAll<SVGImageElement, LineLabelIconSpec>('.ll-logo')
+      .data(icon ? [icon] : [])
+      .join('image')
+      .attr('class', 'll-logo')
+      .attr('href', (d) => d.href)
+      .attr('preserveAspectRatio', 'xMidYMid meet')
+      .attr('x', bbox.x - space)
+      .attr('y', (d) => bbox.y + bbox.height / 2 - d.height / 2)
+      .attr('width', (d) => d.width)
+      .attr('height', (d) => d.height);
+  }
+}
+
+export function updateRenderedLineLabels(
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  labels: readonly LineLabelPlacement[],
+  options: { opacity?: number; offsetX?: number; offsetY?: number } = {},
+): void {
+  const byKey = new Map(labels.map((label) => [label.key, label]));
+  const opacity = options.opacity ?? 1;
+  const offsetX = options.offsetX ?? 8;
+  const offsetY = options.offsetY ?? -14;
+  group.selectAll<SVGGElement, unknown>('.line-label').each(function () {
+    const element = d3.select(this);
+    const label = byKey.get(element.attr('data-line-key'));
+    if (!label) {
+      element.style('opacity', 0);
+      return;
+    }
+    element
+      .attr('transform', `translate(${label.x + offsetX},${label.y + offsetY})`)
+      .style('opacity', label.visible ? opacity : 0);
+  });
+}
+
+export function avoidPointLabelCollisions(
+  zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+): void {
+  interface LabelInfo {
+    element: SVGTextElement;
+    firstTspan: SVGTSpanElement;
+    centerX: number;
+    centerY: number;
+    width: number;
+    lineCount: number;
+    defaultFirstY: number;
+  }
+  const pending: Omit<LabelInfo, 'width'>[] = [];
+  const ascent = 9;
+  const descent = 3;
+  const lineHeight = 11;
+  const padding = 2;
+
+  zoomGroup.selectAll<SVGGElement, unknown>('.dot-group').each(function () {
+    const label = this.querySelector<SVGTextElement>('.point-label');
+    if (
+      !label ||
+      label.style.display === 'none' ||
+      label.style.visibility === 'hidden' ||
+      label.style.opacity === '0' ||
+      this.style.opacity === '0'
+    ) {
+      return;
+    }
+    const tspans = label.querySelectorAll<SVGTSpanElement>('tspan');
+    if (tspans.length === 0) return;
+    const transform = this.getAttribute('transform') ?? '';
+    const match = transform.match(/translate\((?<tx>[^,]+),(?<ty>[^)]+)\)/u);
+    if (!match) return;
+    const lineCount = tspans.length;
+    const defaultFirstY = -(8 + (lineCount - 1) * lineHeight);
+    tspans[0].setAttribute('dy', `${defaultFirstY}px`);
+    label.style.opacity = '1';
+    pending.push({
+      element: label,
+      firstTspan: tspans[0],
+      centerX: Number.parseFloat(match[1]),
+      centerY: Number.parseFloat(match[2]),
+      lineCount,
+      defaultFirstY,
+    });
+  });
+
+  const labels: LabelInfo[] = pending.map((label) => ({
+    ...label,
+    width: label.element.getBBox().width,
+  }));
+  labels.sort((a, b) => a.centerX - b.centerX);
+  // Seed with the run-name and parallelism pills so a point label that would
+  // sit underneath one is pushed to its next candidate slot (or hidden) rather
+  // than being drawn through it. Point-label-vs-point-label behaviour below is
+  // unchanged; the pills simply occupy space before the first label is placed.
+  const placed: RectBounds[] = pillObstacles(zoomGroup);
+
+  for (const label of labels) {
+    const blockHeight = (label.lineCount - 1) * lineHeight + ascent + descent;
+    const firstBaselines = [
+      label.defaultFirstY,
+      14,
+      label.defaultFirstY - blockHeight - 2,
+      14 + blockHeight + 2,
+    ];
+    const boxes = firstBaselines.map((firstY) => ({
+      left: label.centerX - label.width / 2 - padding,
+      right: label.centerX + label.width / 2 + padding,
+      top: label.centerY + firstY - ascent - padding,
+      bottom: label.centerY + firstY + (label.lineCount - 1) * lineHeight + descent + padding,
+    }));
+    const index = firstNonCollidingRect(boxes, placed);
+    if (index === null) {
+      label.element.style.opacity = '0';
+      continue;
+    }
+    label.firstTspan.setAttribute('dy', `${firstBaselines[index]}px`);
+    label.element.style.opacity = '1';
+    placed.push(boxes[index]);
+  }
+}
diff --git a/packages/app/src/components/inference/ui/model-architecture-diagram-renderer.ts b/packages/app/src/components/inference/ui/model-architecture-diagram-renderer.ts
new file mode 100644
index 000000000..d6893f059
--- /dev/null
+++ b/packages/app/src/components/inference/ui/model-architecture-diagram-renderer.ts
@@ -0,0 +1,2316 @@
+import * as d3 from 'd3';
+
+import {
+  type ArchSubBlock,
+  type SubBlockFlow,
+  type ModelArchitecture,
+  formatContextWindow,
+  formatParamCount,
+  expertRouterSummary,
+  getAttentionLabel,
+  getAttentionSubBlocks,
+  getFFNSubBlocks,
+  denseLayerAttentionLabel,
+  ffnVariantLabel,
+  getHybridAttentionSubBlocks,
+  sharedExpertCount,
+} from '@/lib/model-architectures';
+/** Block color definitions for light/dark themes */
+const BLOCK_COLORS = {
+  embedding: { light: '#dbeafe', dark: '#1e3a5f', stroke: '#3b82f6' },
+  attention: { light: '#fef3c7', dark: '#422006', stroke: '#d97706' },
+  ffn: { light: '#d1fae5', dark: '#064e3b', stroke: '#059669' },
+  output: { light: '#e0e7ff', dark: '#1e1b4b', stroke: '#6366f1' },
+  norm: { light: '#f1f5f9', dark: '#1e293b', stroke: '#94a3b8' },
+  specs: { light: '#f8fafc', dark: '#1a1c1e', stroke: '#cbd5e1' },
+  router: { light: '#fce7f3', dark: '#500724', stroke: '#db2777' },
+  expert: { light: '#f3e8ff', dark: '#3b0764', stroke: '#9333ea' },
+  expertActive: { light: '#e9d5ff', dark: '#581c87', stroke: '#7c3aed' },
+};
+
+/** Sub-block color definitions for expanded drill-down views */
+const SUB_BLOCK_COLORS = {
+  projection: { light: '#e0f2fe', dark: '#0c4a6e', stroke: '#0284c7' },
+  activation: { light: '#dcfce7', dark: '#14532d', stroke: '#16a34a' },
+  operation: { light: '#f3f4f6', dark: '#1f2937', stroke: '#6b7280' },
+  attention: { light: '#fef9c3', dark: '#422006', stroke: '#ca8a04' },
+};
+
+function getColor(type: keyof typeof BLOCK_COLORS, isDark: boolean) {
+  const c = BLOCK_COLORS[type];
+  return { fill: isDark ? c.dark : c.light, stroke: c.stroke };
+}
+
+function getSubBlockColor(type: ArchSubBlock['type'], isDark: boolean) {
+  const c = SUB_BLOCK_COLORS[type];
+  return { fill: isDark ? c.dark : c.light, stroke: c.stroke };
+}
+
+interface SubBlockLayoutMetrics {
+  blockHeight: number;
+  arrowHeight: number;
+  paddingY: number;
+}
+
+/** Pure height calculation shared by every sequential/branched drill-down. */
+export function getFlowHeight(
+  flow: SubBlockFlow,
+  hasLabel: boolean,
+  metrics: SubBlockLayoutMetrics,
+): number {
+  const { blockHeight, arrowHeight, paddingY } = metrics;
+  if (flow.layout === 'sequential') {
+    return (
+      flow.blocks.length * blockHeight +
+      Math.max(0, flow.blocks.length - 1) * arrowHeight +
+      paddingY * 2 +
+      (hasLabel ? 24 : 0)
+    );
+  }
+  if (flow.layout === 'threeWay') {
+    const maxRows = Math.max(flow.leftPath.length, flow.middlePath.length, flow.rightPath.length);
+    const pathLabelsHeight = flow.leftLabel || flow.middleLabel || flow.rightLabel ? 16 : 0;
+    const hasIntermediate = flow.intermediateMergeBlocks.length > 0;
+    return (
+      paddingY * 2 +
+      (hasLabel ? 24 : 0) +
+      pathLabelsHeight +
+      arrowHeight +
+      maxRows * blockHeight +
+      Math.max(0, maxRows - 1) * arrowHeight +
+      (arrowHeight + 4) +
+      (hasIntermediate
+        ? flow.intermediateMergeBlocks.length * blockHeight +
+          Math.max(0, flow.intermediateMergeBlocks.length - 1) * arrowHeight +
+          (arrowHeight + 4)
+        : 0) +
+      flow.finalMergeBlocks.length * blockHeight +
+      Math.max(0, flow.finalMergeBlocks.length - 1) * arrowHeight
+    );
+  }
+  const maxRows = Math.max(flow.leftPath.length, flow.rightPath.length);
+  const pathLabelsHeight = flow.leftLabel || flow.rightLabel ? 16 : 0;
+  return (
+    paddingY * 2 +
+    (hasLabel ? 24 : 0) +
+    pathLabelsHeight +
+    arrowHeight +
+    maxRows * blockHeight +
+    Math.max(0, maxRows - 1) * arrowHeight +
+    (arrowHeight + 4) +
+    flow.mergeBlocks.length * blockHeight +
+    Math.max(0, flow.mergeBlocks.length - 1) * arrowHeight
+  );
+}
+
+/** Derive architecture variants, expansion state, flows, and their pure heights. */
+export function deriveArchitectureVariant(
+  arch: ModelArchitecture,
+  expandedBlocks: ReadonlySet<string>,
+  metrics: SubBlockLayoutMetrics,
+) {
+  const isMoE = arch.architectureType === 'moe';
+  const sharedExperts = sharedExpertCount(arch);
+  const ffnVariant = ffnVariantLabel(arch);
+  const hasDenseLayers = isMoE && (arch.denseFFNLayers ?? 0) > 0;
+  const denseLayerCount = arch.denseFFNLayers ?? 0;
+  const moeLayerCount = (arch.numLayers ?? 0) - denseLayerCount;
+  const isAttnExpandable =
+    arch.attentionExpandable !== false &&
+    arch.attentionType !== 'MLA' &&
+    arch.attentionType !== 'AlternatingSinkGQA';
+  const hasAlternatingLayers = Boolean(arch.alternatingLayers?.length);
+  const alternatingSpecs = arch.alternatingLayers ?? [];
+  const attnFlow = isAttnExpandable ? getAttentionSubBlocks(arch) : null;
+  const ffnFlow = getFFNSubBlocks(arch);
+  const denseFFNFlow = hasDenseLayers ? getFFNSubBlocks(arch, { useDenseFFNDim: true }) : null;
+  const attnExpanded = isAttnExpandable && expandedBlocks.has('attention');
+  const ffnExpanded = expandedBlocks.has(isMoE ? 'experts' : 'ffn');
+  const txExpanded = expandedBlocks.has('transformer');
+  const denseTxExpanded = expandedBlocks.has('denseTransformer');
+  const denseAttnExpanded = isAttnExpandable && expandedBlocks.has('denseAttention');
+  const denseFFNExpanded = expandedBlocks.has('denseFFN');
+  const hasHashBlock = isMoE && (arch.hashRoutedLayers ?? 0) > 0;
+  const hashBlockExpanded = hasHashBlock && expandedBlocks.has('hashBlock');
+  const hashExpertsExpanded = hasHashBlock && expandedBlocks.has('hashExperts');
+  const altBlockExpanded = [
+    hasAlternatingLayers && expandedBlocks.has('altBlock0'),
+    hasAlternatingLayers && expandedBlocks.has('altBlock1'),
+  ];
+  const altExpertsExpanded = [
+    hasAlternatingLayers && expandedBlocks.has('altExperts0'),
+    hasAlternatingLayers && expandedBlocks.has('altExperts1'),
+  ];
+  const altAttnExpandable =
+    hasAlternatingLayers &&
+    arch.attentionType === 'Hybrid' &&
+    arch.alternatingAttentionExpandable !== false;
+  const altAttnExpanded = [
+    altAttnExpandable && expandedBlocks.has('altAttention0'),
+    altAttnExpandable && expandedBlocks.has('altAttention1'),
+  ];
+  const altAttnFlow: (SubBlockFlow | null)[] = altAttnExpandable
+    ? [0, 1].map((index) =>
+        alternatingSpecs[index] ? getHybridAttentionSubBlocks(arch, alternatingSpecs[index]) : null,
+      )
+    : [null, null];
+  const attnExpandedH = attnExpanded && attnFlow ? getFlowHeight(attnFlow, false, metrics) : 0;
+  const ffnExpandedH = ffnExpanded ? getFlowHeight(ffnFlow, true, metrics) : 0;
+  const denseAttnExpandedH =
+    denseAttnExpanded && attnFlow ? getFlowHeight(attnFlow, false, metrics) : 0;
+  const denseFFNExpandedH =
+    denseFFNExpanded && denseFFNFlow ? getFlowHeight(denseFFNFlow, true, metrics) : 0;
+  const altExpertsExpandedH = altExpertsExpanded.map((expanded) =>
+    expanded ? getFlowHeight(ffnFlow, true, metrics) : 0,
+  );
+  const altAttnExpandedH = altAttnExpanded.map((expanded, index) => {
+    const flow = altAttnFlow[index];
+    return expanded && flow ? getFlowHeight(flow, false, metrics) : 0;
+  });
+  const hashExpertsExpandedH = hashExpertsExpanded ? getFlowHeight(ffnFlow, true, metrics) : 0;
+
+  return {
+    isMoE,
+    sharedExperts,
+    ffnVariant,
+    hasDenseLayers,
+    denseLayerCount,
+    moeLayerCount,
+    isAttnExpandable,
+    hasAlternatingLayers,
+    alternatingSpecs,
+    attnFlow,
+    ffnFlow,
+    denseFFNFlow,
+    attnExpanded,
+    ffnExpanded,
+    txExpanded,
+    denseTxExpanded,
+    denseAttnExpanded,
+    denseFFNExpanded,
+    hasHashBlock,
+    hashBlockExpanded,
+    hashExpertsExpanded,
+    altBlockExpanded,
+    altExpertsExpanded,
+    altAttnExpandable,
+    altAttnExpanded,
+    altAttnFlow,
+    attnExpandedH,
+    ffnExpandedH,
+    denseAttnExpandedH,
+    denseFFNExpandedH,
+    altExpertsExpandedH,
+    altAttnExpandedH,
+    hashExpertsExpandedH,
+  };
+}
+
+/**
+ * Renders the D3-based SVG architecture diagram.
+ * Supports both dense (Llama) and MoE (DeepSeek R1) architectures.
+ * For MoE models with denseFFNLayers, renders a separate dense transformer block.
+ * MLA attention blocks are NOT expandable (shown as static blocks).
+ */
+export function renderDiagram(
+  svgEl: SVGSVGElement,
+  arch: ModelArchitecture,
+  isDark: boolean,
+  expandedBlocks: Set<string>,
+  onBlockClick: (blockId: string) => void,
+) {
+  const svg = d3.select(svgEl);
+  svg.selectAll('*').remove();
+
+  const containerWidth = svgEl.parentElement?.clientWidth || 600;
+  const width = Math.min(containerWidth, 640);
+
+  // Theme colors
+  const fg = isDark ? '#eaebec' : '#131416';
+  const mutedFg = isDark ? '#9ca3af' : '#6b7280';
+  const borderColor = isDark ? '#374151' : '#d1d5db';
+  const bgSubtle = isDark ? '#1f2937' : '#f9fafb';
+  const expandedBg = isDark ? '#111827' : '#f1f5f9';
+
+  // Layout constants
+  const pad = { top: 20, right: 16, bottom: 16, left: 16 };
+  const bw = width - pad.left - pad.right;
+  const innerW = bw - 32;
+  const innerX = pad.left + 16;
+  const cx = width / 2;
+  const blockH = 52;
+  const smallH = 32;
+  const arrowH = 20;
+  const circleR = 9;
+  const mergeGap = 36;
+  const residLeftX = pad.left + 6;
+  const collapsedTxH = 64;
+
+  // Expert grid constants
+  const expertSize = 36;
+  const expertGap = 6;
+  const expertGridH = 70;
+
+  // Sub-block layout constants
+  const subBlockH = 34;
+  const subArrowH = 12;
+  const subPadY = 10;
+  // Gap between a block and its drill-down flow so the dashed expansion rect
+  // reads as separate from the block above it (not overlapping its border).
+  const drillGap = 8;
+
+  const {
+    isMoE,
+    sharedExperts,
+    ffnVariant,
+    hasDenseLayers,
+    denseLayerCount,
+    moeLayerCount,
+    isAttnExpandable,
+    hasAlternatingLayers,
+    alternatingSpecs,
+    attnFlow,
+    ffnFlow,
+    denseFFNFlow,
+    attnExpanded,
+    ffnExpanded,
+    txExpanded,
+    denseTxExpanded,
+    denseAttnExpanded,
+    denseFFNExpanded,
+    hasHashBlock,
+    hashBlockExpanded,
+    hashExpertsExpanded,
+    altBlockExpanded,
+    altExpertsExpanded,
+    altAttnExpandable,
+    altAttnExpanded,
+    altAttnFlow,
+    attnExpandedH,
+    ffnExpandedH,
+    denseAttnExpandedH,
+    denseFFNExpandedH,
+    altExpertsExpandedH,
+    altAttnExpandedH,
+    hashExpertsExpandedH,
+  } = deriveArchitectureVariant(arch, expandedBlocks, {
+    blockHeight: subBlockH,
+    arrowHeight: subArrowH,
+    paddingY: subPadY,
+  });
+
+  // Compute vertical positions
+  let y = pad.top;
+
+  // Title
+  const titleY = y;
+  y += 44;
+
+  // Embedding
+  const embedY = y;
+  y += blockH + arrowH;
+
+  // === DENSE TRANSFORMER BLOCK (for MoE models with initial dense layers) ===
+  let denseTxStart = 0;
+  let denseNorm1Y = 0;
+  let denseAttnY = 0;
+  let denseAttnExpandedStartY = 0;
+  let denseMerge1Y = 0;
+  let denseNorm2Y = 0;
+  let denseFFNBlockY = 0;
+  let denseFFNExpandedStartY = 0;
+  let denseMerge2Y = 0;
+  let denseTxEnd = 0;
+
+  if (hasDenseLayers) {
+    denseTxStart = y;
+    if (denseTxExpanded) {
+      y += 14;
+
+      denseNorm1Y = y;
+      y += smallH + arrowH;
+
+      denseAttnY = y;
+      y += blockH;
+
+      if (denseAttnExpanded) y += drillGap;
+      denseAttnExpandedStartY = y;
+      if (denseAttnExpanded) {
+        y += denseAttnExpandedH;
+      }
+
+      y += 4;
+      denseMerge1Y = y + mergeGap / 2;
+      y += mergeGap;
+
+      y += arrowH;
+      denseNorm2Y = y;
+      y += smallH + arrowH;
+
+      denseFFNBlockY = y;
+      y += blockH;
+
+      denseFFNExpandedStartY = y;
+      if (denseFFNExpanded) {
+        y += denseFFNExpandedH;
+      }
+
+      denseMerge2Y = y + mergeGap / 2;
+      y += mergeGap;
+
+      y += 14;
+    } else {
+      y += collapsedTxH;
+    }
+    denseTxEnd = y;
+    y += arrowH;
+  }
+
+  // === HASH-ROUTED MoE PREFIX BLOCK (DeepSeek V4: first N layers use hash routing) ===
+  let hashTxStart = 0;
+  let hashNorm1Y = 0;
+  let hashAttnY = 0;
+  let hashMerge1Y = 0;
+  let hashNorm2Y = 0;
+  let hashExpertY = 0;
+  let hashFFNExpandedStartY = 0;
+  let hashMerge2Y = 0;
+  let hashTxEnd = 0;
+
+  if (hasHashBlock) {
+    hashTxStart = y;
+    if (hashBlockExpanded) {
+      y += 14;
+
+      hashNorm1Y = y;
+      y += smallH + arrowH;
+
+      hashAttnY = y;
+      y += blockH;
+
+      y += 4;
+      hashMerge1Y = y + mergeGap / 2;
+      y += mergeGap;
+
+      y += arrowH;
+      hashNorm2Y = y;
+      y += smallH + arrowH;
+
+      // Hash Router + Expert grid (drawExpertGrid lays out router above eY)
+      y += blockH + arrowH;
+      hashExpertY = y;
+      y += expertGridH;
+
+      hashFFNExpandedStartY = y;
+      if (hashExpertsExpanded) {
+        y += hashExpertsExpandedH;
+      }
+
+      hashMerge2Y = y + mergeGap / 2;
+      y += mergeGap;
+
+      y += 14;
+    } else {
+      y += collapsedTxH;
+    }
+    hashTxEnd = y;
+    y += arrowH;
+  }
+
+  // === ALTERNATING TRANSFORMER BLOCKS (for models like gpt-oss with alternating attention) ===
+  const altBlockStart = [0, 0];
+  const altBlockEnd = [0, 0];
+  const altNorm1Y = [0, 0];
+  const altAttnY = [0, 0];
+  const altAttnExpandedStartY = [0, 0];
+  const altMerge1Y = [0, 0];
+  const altNorm2Y = [0, 0];
+  const altRouterY = [0, 0];
+  const altExpertY = [0, 0];
+  const altFFNExpandedStartY = [0, 0];
+  const altMerge2Y = [0, 0];
+  let altIndicatorY = 0;
+
+  // Main transformer container (used when NOT hasAlternatingLayers)
+  const txStart = hasAlternatingLayers ? 0 : y;
+  let attnY = 0;
+  let attnExpandedStartY = 0;
+  let merge1Y = 0;
+  let norm1Y = 0;
+  const _routerY = 0;
+  let expertY = 0;
+  let ffnY = 0;
+  let ffnExpandedStartY = 0;
+  let merge2Y = 0;
+  let norm2Y = 0;
+  let txEnd = 0; // oxlint-disable-line no-useless-assignment -- overwritten when txExpanded
+
+  if (hasAlternatingLayers) {
+    // Layout two separate transformer blocks with alternating indicator between them
+    for (let bi = 0; bi < 2; bi++) {
+      altBlockStart[bi] = y;
+      const isExp = altBlockExpanded[bi];
+      const isExpExperts = altExpertsExpanded[bi];
+
+      if (isExp) {
+        y += 14;
+
+        altNorm1Y[bi] = y;
+        y += smallH + arrowH;
+
+        altAttnY[bi] = y;
+        y += blockH;
+
+        // Expanded hybrid-attention sub-blocks (sliding window + compressed)
+        if (altAttnExpanded[bi]) y += drillGap;
+        altAttnExpandedStartY[bi] = y;
+        if (altAttnExpanded[bi]) {
+          y += altAttnExpandedH[bi];
+        }
+
+        y += 4;
+        altMerge1Y[bi] = y + mergeGap / 2;
+        y += mergeGap;
+
+        y += arrowH;
+        altNorm2Y[bi] = y;
+        y += smallH + arrowH;
+
+        // MoE Router
+        altRouterY[bi] = y;
+        y += blockH + arrowH;
+
+        // Expert grid
+        altExpertY[bi] = y;
+        y += expertGridH;
+
+        // Expanded expert FFN sub-blocks
+        altFFNExpandedStartY[bi] = y;
+        if (isExpExperts) {
+          y += altExpertsExpandedH[bi];
+        }
+
+        altMerge2Y[bi] = y + mergeGap / 2;
+        y += mergeGap;
+
+        y += 14;
+      } else {
+        y += collapsedTxH;
+      }
+      altBlockEnd[bi] = y;
+
+      // Add alternating indicator between the two blocks
+      if (bi === 0) {
+        y += 6;
+        altIndicatorY = y + 10;
+        y += 24;
+      }
+    }
+  } else if (txExpanded) {
+    y += 14;
+
+    // Pre-LN: RMSNorm 1 before attention
+    norm1Y = y;
+    y += smallH + arrowH;
+
+    // Attention
+    attnY = y;
+    y += blockH;
+
+    // Expanded attention sub-blocks (only for non-MLA, non-AlternatingSinkGQA)
+    if (attnExpanded) y += drillGap;
+    attnExpandedStartY = y;
+    if (attnExpanded) {
+      y += attnExpandedH;
+    }
+
+    y += 4;
+    merge1Y = y + mergeGap / 2;
+    y += mergeGap;
+
+    // Pre-LN: RMSNorm 2 before FFN
+    y += arrowH;
+    norm2Y = y;
+    y += smallH + arrowH;
+
+    if (isMoE) {
+      // Router
+      y += blockH + arrowH;
+
+      // Expert grid
+      expertY = y;
+      y += expertGridH;
+    } else {
+      // Dense FFN
+      ffnY = y;
+      y += blockH;
+    }
+
+    // Expanded FFN sub-blocks (shared by MoE and dense paths)
+    ffnExpandedStartY = y;
+    if (ffnExpanded) {
+      y += ffnExpandedH;
+    }
+
+    merge2Y = y + mergeGap / 2;
+    y += mergeGap;
+
+    y += 14;
+  } else {
+    y += collapsedTxH;
+  }
+  txEnd = y;
+  y += arrowH;
+
+  // Final RMSNorm
+  const finalNormY = y;
+  y += smallH + arrowH;
+
+  // Output head
+  const outputY = y;
+  y += blockH + 16;
+
+  // Specs bar
+  const specsY = y;
+  const specsH = 56;
+  y += specsH;
+
+  const totalH = y + pad.bottom;
+
+  // Set SVG size
+  svg.attr('width', width).attr('height', totalH).attr('viewBox', `0 0 ${width} ${totalH}`);
+
+  // Defs (arrow markers)
+  const defs = svg.append('defs');
+  defs
+    .append('marker')
+    .attr('id', 'arch-arrow')
+    .attr('viewBox', '0 0 10 10')
+    .attr('refX', 5)
+    .attr('refY', 5)
+    .attr('markerWidth', 5)
+    .attr('markerHeight', 5)
+    .attr('orient', 'auto-start-reverse')
+    .append('path')
+    .attr('d', 'M 0 0 L 10 5 L 0 10 z')
+    .attr('fill', mutedFg);
+
+  defs
+    .append('marker')
+    .attr('id', 'arch-arrow-sub')
+    .attr('viewBox', '0 0 10 10')
+    .attr('refX', 5)
+    .attr('refY', 5)
+    .attr('markerWidth', 4)
+    .attr('markerHeight', 4)
+    .attr('orient', 'auto-start-reverse')
+    .append('path')
+    .attr('d', 'M 0 0 L 10 5 L 0 10 z')
+    .attr('fill', mutedFg);
+
+  const bgG = svg.append('g');
+  const g = svg.append('g');
+
+  // --- Helpers ---
+  function drawArrow(fromY: number, toY: number) {
+    g.append('line')
+      .attr('x1', cx)
+      .attr('y1', fromY)
+      .attr('x2', cx)
+      .attr('y2', toY - 2)
+      .attr('stroke', mutedFg)
+      .attr('stroke-width', 1.5)
+      .attr('marker-end', 'url(#arch-arrow)');
+  }
+
+  // Draw a +, −, or × glyph from geometric strokes, perfectly centered at
+  // (gx, gy). Rendering these as lines instead of <text> sidesteps font
+  // baseline drift, which left the symbol sitting slightly low inside the
+  // merge / expand circles regardless of dy tuning.
+  function drawCircleGlyph(gx: number, gy: number, color: string, symbol: string) {
+    const arm = 5;
+    const seg = (x1: number, y1: number, x2: number, y2: number) =>
+      g
+        .append('line')
+        .attr('x1', x1)
+        .attr('y1', y1)
+        .attr('x2', x2)
+        .attr('y2', y2)
+        .attr('stroke', color)
+        .attr('stroke-width', 2)
+        .attr('stroke-linecap', 'round')
+        .style('pointer-events', 'none');
+    if (symbol === '×') {
+      seg(gx - arm, gy - arm, gx + arm, gy + arm);
+      seg(gx - arm, gy + arm, gx + arm, gy - arm);
+    } else {
+      seg(gx - arm, gy, gx + arm, gy); // horizontal arm (+ and −)
+      if (symbol !== '−') seg(gx, gy - arm, gx, gy + arm); // vertical arm (+ only)
+    }
+  }
+
+  // Models with hyper-connections (mHC) replace the plain residual add with a
+  // multi-stream mixer; when present, residual merges render as an "mHC ×N" pill
+  // instead of a "+" circle. N = number of parallel residual streams.
+  const hcStreams = arch.hyperConnections ?? 0;
+  const isHyperConn = hcStreams > 1;
+
+  function drawResidualBypass(branchY: number, mergeY: number) {
+    // Tap the residual from the input stream ABOVE the norm (in the arrow gap)
+    // so the horizontal connector doesn't run across the RMSNorm block.
+    const tapY = branchY - arrowH / 2;
+    // Keep the node's vertical half-height = circleR so the spine arrows (drawn
+    // by callers to mergeY ± circleR) still meet its top/bottom edges.
+    const nodeHalfW = isHyperConn ? 25 : circleR;
+    bgG
+      .append('path')
+      .attr(
+        'd',
+        `M ${cx} ${tapY} L ${residLeftX} ${tapY} L ${residLeftX} ${mergeY} L ${cx - nodeHalfW} ${mergeY}`,
+      )
+      .attr('fill', 'none')
+      .attr('stroke', mutedFg)
+      .attr('stroke-width', 1.5)
+      .attr('opacity', 0.6);
+    if (isHyperConn) {
+      g.append('rect')
+        .attr('x', cx - nodeHalfW)
+        .attr('y', mergeY - circleR)
+        .attr('width', nodeHalfW * 2)
+        .attr('height', circleR * 2)
+        .attr('rx', circleR)
+        .attr('fill', bgSubtle)
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1.5);
+      g.append('text')
+        .attr('x', cx)
+        .attr('y', mergeY)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', fg)
+        .attr('font-size', '8.5px')
+        .attr('font-weight', 700)
+        .attr('font-family', 'inherit')
+        .text(`mHC ×${hcStreams}`);
+    } else {
+      g.append('circle')
+        .attr('cx', cx)
+        .attr('cy', mergeY)
+        .attr('r', circleR)
+        .attr('fill', bgSubtle)
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1.5);
+      drawCircleGlyph(cx, mergeY, fg, '+');
+    }
+  }
+
+  function drawBlock(
+    x: number,
+    by: number,
+    w: number,
+    h: number,
+    type: keyof typeof BLOCK_COLORS,
+    mainText: string,
+    subText?: string,
+  ) {
+    const c = getColor(type, isDark);
+    g.append('rect')
+      .attr('x', x)
+      .attr('y', by)
+      .attr('width', w)
+      .attr('height', h)
+      .attr('rx', 8)
+      .attr('fill', c.fill)
+      .attr('stroke', c.stroke)
+      .attr('stroke-width', 1.5);
+
+    g.append('text')
+      .attr('x', x + w / 2)
+      .attr('y', by + h / 2 - (subText ? 7 : 0))
+      .attr('text-anchor', 'middle')
+      .attr('dy', '0.35em')
+      .attr('fill', fg)
+      .attr('font-size', '13px')
+      .attr('font-weight', 600)
+      .attr('font-family', 'inherit')
+      .text(mainText);
+
+    if (subText) {
+      g.append('text')
+        .attr('x', x + w / 2)
+        .attr('y', by + h / 2 + 10)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', '11px')
+        .attr('font-family', 'inherit')
+        .text(subText);
+    }
+  }
+
+  function drawExpandableBlock(
+    x: number,
+    by: number,
+    w: number,
+    h: number,
+    type: keyof typeof BLOCK_COLORS,
+    mainText: string,
+    subText: string | undefined,
+    isBlockExpanded: boolean,
+    blockId: string,
+  ) {
+    const c = getColor(type, isDark);
+    g.append('rect')
+      .attr('x', x)
+      .attr('y', by)
+      .attr('width', w)
+      .attr('height', h)
+      .attr('rx', 8)
+      .attr('fill', c.fill)
+      .attr('stroke', c.stroke)
+      .attr('stroke-width', isBlockExpanded ? 2.5 : 1.5);
+
+    g.append('text')
+      .attr('x', x + w / 2 - 8)
+      .attr('y', by + h / 2 - (subText ? 7 : 0))
+      .attr('text-anchor', 'middle')
+      .attr('dy', '0.35em')
+      .attr('fill', fg)
+      .attr('font-size', '13px')
+      .attr('font-weight', 600)
+      .attr('font-family', 'inherit')
+      .style('pointer-events', 'none')
+      .text(mainText);
+
+    if (subText) {
+      g.append('text')
+        .attr('x', x + w / 2 - 8)
+        .attr('y', by + h / 2 + 10)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', '11px')
+        .attr('font-family', 'inherit')
+        .style('pointer-events', 'none')
+        .text(subText);
+    }
+
+    const iconX = x + w - 22;
+    const iconY = by + h / 2;
+
+    g.append('circle')
+      .attr('cx', iconX)
+      .attr('cy', iconY)
+      .attr('r', 10)
+      .attr('fill', isDark ? 'rgba(255,255,255,0.06)' : 'rgba(0,0,0,0.04)')
+      .attr('stroke', c.stroke)
+      .attr('stroke-width', 1)
+      .style('pointer-events', 'none');
+
+    drawCircleGlyph(iconX, iconY, c.stroke, isBlockExpanded ? '\u2212' : '+');
+
+    g.append('rect')
+      .attr('x', x)
+      .attr('y', by)
+      .attr('width', w)
+      .attr('height', h)
+      .attr('rx', 8)
+      .attr('fill', 'transparent')
+      .style('cursor', 'pointer')
+      .attr('data-testid', `expand-${blockId}`)
+      .on('click', () => {
+        onBlockClick(blockId);
+      });
+  }
+
+  const DEFAULT_SUB_BLOCK_FONT_SIZE = { name: '11px', detail: '9px' };
+
+  /** Draw a single sub-block at the given position */
+  function drawSingleSubBlock(
+    block: ArchSubBlock,
+    bx: number,
+    by: number,
+    subBw: number,
+    fontSize: { name: string; detail: string } = DEFAULT_SUB_BLOCK_FONT_SIZE,
+  ) {
+    const color = getSubBlockColor(block.type, isDark);
+    g.append('rect')
+      .attr('x', bx)
+      .attr('y', by)
+      .attr('width', subBw)
+      .attr('height', subBlockH)
+      .attr('rx', 6)
+      .attr('fill', color.fill)
+      .attr('stroke', color.stroke)
+      .attr('stroke-width', 1);
+
+    g.append('text')
+      .attr('x', bx + subBw / 2)
+      .attr('y', by + subBlockH / 2 - (block.detail ? 5 : 0))
+      .attr('text-anchor', 'middle')
+      .attr('dy', '0.35em')
+      .attr('fill', fg)
+      .attr('font-size', fontSize.name)
+      .attr('font-weight', 500)
+      .attr('font-family', 'inherit')
+      .text(block.name);
+
+    if (block.detail) {
+      g.append('text')
+        .attr('x', bx + subBw / 2)
+        .attr('y', by + subBlockH / 2 + 8)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', fontSize.detail)
+        .attr('font-family', 'inherit')
+        .text(block.detail);
+    }
+  }
+
+  /** Draw parallel sub-block flow (two columns converging into merge blocks) */
+  function drawParallelFlow(
+    flow: Extract<SubBlockFlow, { layout: 'parallel' }>,
+    startY: number,
+    x: number,
+    w: number,
+    label?: string,
+  ) {
+    let sy = startY;
+    const flowH = getFlowHeight(flow, Boolean(label), {
+      blockHeight: subBlockH,
+      arrowHeight: subArrowH,
+      paddingY: subPadY,
+    });
+
+    g.append('rect')
+      .attr('x', x + 4)
+      .attr('y', sy)
+      .attr('width', w - 8)
+      .attr('height', flowH)
+      .attr('rx', 8)
+      .attr('fill', expandedBg)
+      .attr('stroke', borderColor)
+      .attr('stroke-width', 1)
+      .attr('stroke-dasharray', '4,3');
+
+    sy += subPadY;
+
+    if (label) {
+      g.append('text')
+        .attr('x', x + w / 2)
+        .attr('y', sy + 8)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', '10px')
+        .attr('font-weight', 600)
+        .attr('font-family', 'inherit')
+        .attr('letter-spacing', '0.5px')
+        .text(label.toUpperCase());
+      sy += 24;
+    }
+
+    const colGap = 10;
+    const innerPadCol = 12;
+    const colW = (w - 2 * innerPadCol - colGap) / 2;
+    const leftX = x + innerPadCol;
+    const rightX = x + innerPadCol + colW + colGap;
+    const leftCx = leftX + colW / 2;
+    const rightCx = rightX + colW / 2;
+    const mergeCx = x + w / 2;
+
+    if (flow.leftLabel || flow.rightLabel) {
+      if (flow.leftLabel) {
+        g.append('text')
+          .attr('x', leftCx)
+          .attr('y', sy + 6)
+          .attr('text-anchor', 'middle')
+          .attr('dy', '0.35em')
+          .attr('fill', mutedFg)
+          .attr('font-size', '9px')
+          .attr('font-weight', 600)
+          .attr('font-family', 'inherit')
+          .text(flow.leftLabel);
+      }
+      if (flow.rightLabel) {
+        g.append('text')
+          .attr('x', rightCx)
+          .attr('y', sy + 6)
+          .attr('text-anchor', 'middle')
+          .attr('dy', '0.35em')
+          .attr('fill', mutedFg)
+          .attr('font-size', '9px')
+          .attr('font-weight', 600)
+          .attr('font-family', 'inherit')
+          .text(flow.rightLabel);
+      }
+      sy += 16;
+    }
+
+    const splitTopY = sy;
+    const splitMidY = sy + subArrowH / 2;
+    sy += subArrowH;
+
+    const parallelStartY = sy;
+    const colFontSize = { name: '10px', detail: '8px' };
+
+    const maxRows = Math.max(flow.leftPath.length, flow.rightPath.length);
+    const rowsH = (n: number) => n * subBlockH + Math.max(0, n - 1) * subArrowH;
+    const colAreaH = rowsH(maxRows);
+    // Vertically center each column within the shared column area so an unequal
+    // split (e.g. 1 local vs 2 compressed) reads as an intentional branch merge
+    // rather than leaving the shorter column's connector dangling as a long
+    // unattached line beside the taller column.
+    const leftStartY = parallelStartY + (colAreaH - rowsH(flow.leftPath.length)) / 2;
+    const rightStartY = parallelStartY + (colAreaH - rowsH(flow.rightPath.length)) / 2;
+
+    g.append('line')
+      .attr('x1', mergeCx)
+      .attr('y1', splitTopY)
+      .attr('x2', mergeCx)
+      .attr('y2', splitMidY)
+      .attr('stroke', mutedFg)
+      .attr('stroke-width', 1);
+
+    g.append('path')
+      .attr('d', `M ${mergeCx} ${splitMidY} L ${leftCx} ${splitMidY} L ${leftCx} ${leftStartY - 2}`)
+      .attr('fill', 'none')
+      .attr('stroke', mutedFg)
+      .attr('stroke-width', 1)
+      .attr('marker-end', 'url(#arch-arrow-sub)');
+
+    g.append('path')
+      .attr(
+        'd',
+        `M ${mergeCx} ${splitMidY} L ${rightCx} ${splitMidY} L ${rightCx} ${rightStartY - 2}`,
+      )
+      .attr('fill', 'none')
+      .attr('stroke', mutedFg)
+      .attr('stroke-width', 1)
+      .attr('marker-end', 'url(#arch-arrow-sub)');
+
+    let lsy = leftStartY;
+    for (let i = 0; i < flow.leftPath.length; i++) {
+      drawSingleSubBlock(flow.leftPath[i], leftX, lsy, colW, colFontSize);
+      lsy += subBlockH;
+      if (i < flow.leftPath.length - 1) {
+        g.append('line')
+          .attr('x1', leftCx)
+          .attr('y1', lsy + 1)
+          .attr('x2', leftCx)
+          .attr('y2', lsy + subArrowH - 2)
+          .attr('stroke', mutedFg)
+          .attr('stroke-width', 1)
+          .attr('marker-end', 'url(#arch-arrow-sub)');
+        lsy += subArrowH;
+      }
+    }
+    const leftEndY = lsy;
+
+    let rsy = rightStartY;
+    for (let i = 0; i < flow.rightPath.length; i++) {
+      drawSingleSubBlock(flow.rightPath[i], rightX, rsy, colW, colFontSize);
+      rsy += subBlockH;
+      if (i < flow.rightPath.length - 1) {
+        g.append('line')
+          .attr('x1', rightCx)
+          .attr('y1', rsy + 1)
+          .attr('x2', rightCx)
+          .attr('y2', rsy + subArrowH - 2)
+          .attr('stroke', mutedFg)
+          .attr('stroke-width', 1)
+          .attr('marker-end', 'url(#arch-arrow-sub)');
+        rsy += subArrowH;
+      }
+    }
+    const rightEndY = rsy;
+
+    const mergeStartY = parallelStartY + colAreaH + subArrowH + 4;
+
+    const subInnerXLocal = x + 16;
+    const subInnerWLocal = w - 40;
+
+    const firstIsCircle = flow.mergeBlocks[0]?.circleSymbol;
+
+    if (firstIsCircle) {
+      const circleCy = mergeStartY + subBlockH / 2;
+      g.append('path')
+        .attr(
+          'd',
+          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${circleCy} L ${mergeCx - circleR - 2} ${circleCy}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+      g.append('path')
+        .attr(
+          'd',
+          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${circleCy} L ${mergeCx + circleR + 2} ${circleCy}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+    } else {
+      const convergeHorizY = Math.max(leftEndY, rightEndY) + (subArrowH + 4) / 2;
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${convergeHorizY} L ${mergeCx} ${convergeHorizY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${convergeHorizY} L ${mergeCx} ${convergeHorizY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('line')
+        .attr('x1', mergeCx)
+        .attr('y1', convergeHorizY)
+        .attr('x2', mergeCx)
+        .attr('y2', mergeStartY - 2)
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1)
+        .attr('marker-end', 'url(#arch-arrow-sub)');
+    }
+
+    let msy = mergeStartY;
+    for (let i = 0; i < flow.mergeBlocks.length; i++) {
+      const block = flow.mergeBlocks[i];
+      if (block.circleSymbol) {
+        const circleCy = msy + subBlockH / 2;
+        const symbolR = circleR + 2;
+        g.append('circle')
+          .attr('cx', mergeCx)
+          .attr('cy', circleCy)
+          .attr('r', symbolR)
+          .attr('fill', bgSubtle)
+          .attr('stroke', mutedFg)
+          .attr('stroke-width', 1.5);
+        drawCircleGlyph(mergeCx, circleCy, fg, block.circleSymbol);
+      } else {
+        drawSingleSubBlock(block, subInnerXLocal, msy, subInnerWLocal);
+      }
+      msy += subBlockH;
+      if (i < flow.mergeBlocks.length - 1) {
+        g.append('line')
+          .attr('x1', mergeCx)
+          .attr('y1', msy + 1)
+          .attr('x2', mergeCx)
+          .attr('y2', msy + subArrowH - 2)
+          .attr('stroke', mutedFg)
+          .attr('stroke-width', 1)
+          .attr('marker-end', 'url(#arch-arrow-sub)');
+        msy += subArrowH;
+      }
+    }
+  }
+
+  /** Draw three-way parallel flow (3 columns: left+middle converge first, then merge with right) */
+  function drawThreeWayFlow(
+    flow: Extract<SubBlockFlow, { layout: 'threeWay' }>,
+    startY: number,
+    x: number,
+    w: number,
+    _label?: string,
+  ) {
+    let sy = startY;
+    const flowH = getFlowHeight(flow, false, {
+      blockHeight: subBlockH,
+      arrowHeight: subArrowH,
+      paddingY: subPadY,
+    });
+
+    g.append('rect')
+      .attr('x', x + 4)
+      .attr('y', sy)
+      .attr('width', w - 8)
+      .attr('height', flowH)
+      .attr('rx', 8)
+      .attr('fill', expandedBg)
+      .attr('stroke', borderColor)
+      .attr('stroke-width', 1)
+      .attr('stroke-dasharray', '4,3');
+
+    sy += subPadY;
+
+    const colGap = 6;
+    const innerPadCol = 8;
+    const colW = (w - 2 * innerPadCol - 2 * colGap) / 3;
+    const leftX = x + innerPadCol;
+    const middleX = x + innerPadCol + colW + colGap;
+    const rightX = x + innerPadCol + 2 * (colW + colGap);
+    const leftCx = leftX + colW / 2;
+    const middleCx = middleX + colW / 2;
+    const rightCx = rightX + colW / 2;
+    const mergeCxLocal = x + w / 2;
+    const qkMergeCx = (leftCx + middleCx) / 2;
+
+    if (flow.leftLabel || flow.middleLabel || flow.rightLabel) {
+      const labels = [
+        { label: flow.leftLabel, cx: leftCx },
+        { label: flow.middleLabel, cx: middleCx },
+        { label: flow.rightLabel, cx: rightCx },
+      ];
+      for (const { label: lbl, cx: lcx } of labels) {
+        if (lbl) {
+          g.append('text')
+            .attr('x', lcx)
+            .attr('y', sy + 6)
+            .attr('text-anchor', 'middle')
+            .attr('dy', '0.35em')
+            .attr('fill', mutedFg)
+            .attr('font-size', '9px')
+            .attr('font-weight', 600)
+            .attr('font-family', 'inherit')
+            .text(lbl);
+        }
+      }
+      sy += 16;
+    }
+
+    const splitTopY = sy;
+    const splitMidY = sy + subArrowH / 2;
+    sy += subArrowH;
+
+    const parallelStartY = sy;
+    const colFontSize = { name: '9px', detail: '7.5px' };
+
+    g.append('line')
+      .attr('x1', mergeCxLocal)
+      .attr('y1', splitTopY)
+      .attr('x2', mergeCxLocal)
+      .attr('y2', splitMidY)
+      .attr('stroke', mutedFg)
+      .attr('stroke-width', 1);
+
+    for (const colCx of [leftCx, middleCx, rightCx]) {
+      g.append('path')
+        .attr(
+          'd',
+          `M ${mergeCxLocal} ${splitMidY} L ${colCx} ${splitMidY} L ${colCx} ${parallelStartY - 2}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1)
+        .attr('marker-end', 'url(#arch-arrow-sub)');
+    }
+
+    const drawColumn = (path: ArchSubBlock[], colX: number, colCx: number) => {
+      let csy = parallelStartY;
+      for (let i = 0; i < path.length; i++) {
+        drawSingleSubBlock(path[i], colX, csy, colW, colFontSize);
+        csy += subBlockH;
+        if (i < path.length - 1) {
+          g.append('line')
+            .attr('x1', colCx)
+            .attr('y1', csy + 1)
+            .attr('x2', colCx)
+            .attr('y2', csy + subArrowH - 2)
+            .attr('stroke', mutedFg)
+            .attr('stroke-width', 1)
+            .attr('marker-end', 'url(#arch-arrow-sub)');
+          csy += subArrowH;
+        }
+      }
+      return csy;
+    };
+
+    const leftEndY = drawColumn(flow.leftPath, leftX, leftCx);
+    const middleEndY = drawColumn(flow.middlePath, middleX, middleCx);
+    const rightEndY = drawColumn(flow.rightPath, rightX, rightCx);
+
+    const maxRows = Math.max(flow.leftPath.length, flow.middlePath.length, flow.rightPath.length);
+    const hasIntermediate = flow.intermediateMergeBlocks.length > 0;
+    let finalMergeStartY: number;
+
+    if (hasIntermediate) {
+      const intermediateStartY =
+        parallelStartY + maxRows * subBlockH + Math.max(0, maxRows - 1) * subArrowH + subArrowH + 4;
+
+      const qkConvergeY = Math.max(leftEndY, middleEndY) + (subArrowH + 4) / 2;
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${qkConvergeY} L ${qkMergeCx} ${qkConvergeY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${middleCx} ${middleEndY + 1} L ${middleCx} ${qkConvergeY} L ${qkMergeCx} ${qkConvergeY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('line')
+        .attr('x1', qkMergeCx)
+        .attr('y1', qkConvergeY)
+        .attr('x2', qkMergeCx)
+        .attr('y2', intermediateStartY - 2)
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1)
+        .attr('marker-end', 'url(#arch-arrow-sub)');
+
+      const intermediateBlockW = middleX + colW - leftX;
+      let imsy = intermediateStartY;
+      for (let i = 0; i < flow.intermediateMergeBlocks.length; i++) {
+        drawSingleSubBlock(
+          flow.intermediateMergeBlocks[i],
+          leftX,
+          imsy,
+          intermediateBlockW,
+          colFontSize,
+        );
+        imsy += subBlockH;
+        if (i < flow.intermediateMergeBlocks.length - 1) {
+          g.append('line')
+            .attr('x1', qkMergeCx)
+            .attr('y1', imsy + 1)
+            .attr('x2', qkMergeCx)
+            .attr('y2', imsy + subArrowH - 2)
+            .attr('stroke', mutedFg)
+            .attr('stroke-width', 1)
+            .attr('marker-end', 'url(#arch-arrow-sub)');
+          imsy += subArrowH;
+        }
+      }
+
+      finalMergeStartY =
+        intermediateStartY +
+        flow.intermediateMergeBlocks.length * subBlockH +
+        Math.max(0, flow.intermediateMergeBlocks.length - 1) * subArrowH +
+        subArrowH +
+        4;
+
+      const finalConvergeY = Math.max(imsy, rightEndY) + (subArrowH + 4) / 2;
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${qkMergeCx} ${imsy + 1} L ${qkMergeCx} ${finalConvergeY} L ${mergeCxLocal} ${finalConvergeY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${finalConvergeY} L ${mergeCxLocal} ${finalConvergeY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('line')
+        .attr('x1', mergeCxLocal)
+        .attr('y1', finalConvergeY)
+        .attr('x2', mergeCxLocal)
+        .attr('y2', finalMergeStartY - 2)
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1)
+        .attr('marker-end', 'url(#arch-arrow-sub)');
+    } else {
+      finalMergeStartY =
+        parallelStartY + maxRows * subBlockH + Math.max(0, maxRows - 1) * subArrowH + subArrowH + 4;
+
+      const allEndY = Math.max(leftEndY, middleEndY, rightEndY);
+      const convergeY = allEndY + (subArrowH + 4) / 2;
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${leftCx} ${leftEndY + 1} L ${leftCx} ${convergeY} L ${mergeCxLocal} ${convergeY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('line')
+        .attr('x1', middleCx)
+        .attr('y1', middleEndY + 1)
+        .attr('x2', middleCx)
+        .attr('y2', convergeY)
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+      if (Math.abs(middleCx - mergeCxLocal) > 1) {
+        g.append('line')
+          .attr('x1', middleCx)
+          .attr('y1', convergeY)
+          .attr('x2', mergeCxLocal)
+          .attr('y2', convergeY)
+          .attr('stroke', mutedFg)
+          .attr('stroke-width', 1);
+      }
+
+      g.append('path')
+        .attr(
+          'd',
+          `M ${rightCx} ${rightEndY + 1} L ${rightCx} ${convergeY} L ${mergeCxLocal} ${convergeY}`,
+        )
+        .attr('fill', 'none')
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1);
+
+      g.append('line')
+        .attr('x1', mergeCxLocal)
+        .attr('y1', convergeY)
+        .attr('x2', mergeCxLocal)
+        .attr('y2', finalMergeStartY - 2)
+        .attr('stroke', mutedFg)
+        .attr('stroke-width', 1)
+        .attr('marker-end', 'url(#arch-arrow-sub)');
+    }
+
+    const subInnerXLocal = x + 16;
+    const subInnerWLocal = w - 40;
+    let fmsy = finalMergeStartY;
+    for (let i = 0; i < flow.finalMergeBlocks.length; i++) {
+      drawSingleSubBlock(flow.finalMergeBlocks[i], subInnerXLocal, fmsy, subInnerWLocal);
+      fmsy += subBlockH;
+      if (i < flow.finalMergeBlocks.length - 1) {
+        g.append('line')
+          .attr('x1', mergeCxLocal)
+          .attr('y1', fmsy + 1)
+          .attr('x2', mergeCxLocal)
+          .attr('y2', fmsy + subArrowH - 2)
+          .attr('stroke', mutedFg)
+          .attr('stroke-width', 1)
+          .attr('marker-end', 'url(#arch-arrow-sub)');
+        fmsy += subArrowH;
+      }
+    }
+  }
+
+  /** Dispatch to correct rendering based on flow layout */
+  function drawFlow(flow: SubBlockFlow, startY: number, x: number, w: number, label?: string) {
+    if (flow.layout === 'threeWay') {
+      drawThreeWayFlow(flow, startY, x, w, label);
+    } else if (flow.layout === 'parallel') {
+      drawParallelFlow(flow, startY, x, w, label);
+    }
+  }
+
+  /** Draw a collapsed transformer block (compact summary with expand icon) */
+  function drawCollapsedTransformerBlock(
+    x: number,
+    by: number,
+    w: number,
+    h: number,
+    label: string,
+    subtitle: string,
+    blockId: string,
+  ) {
+    g.append('rect')
+      .attr('x', x)
+      .attr('y', by)
+      .attr('width', w)
+      .attr('height', h)
+      .attr('rx', 10)
+      .attr('fill', bgSubtle)
+      .attr('stroke', borderColor)
+      .attr('stroke-width', 2)
+      .attr('stroke-dasharray', '6,3');
+
+    g.append('text')
+      .attr('x', x + w / 2 - 8)
+      .attr('y', by + h / 2 - 8)
+      .attr('text-anchor', 'middle')
+      .attr('dy', '0.35em')
+      .attr('fill', fg)
+      .attr('font-size', '13px')
+      .attr('font-weight', 600)
+      .attr('font-family', 'inherit')
+      .style('pointer-events', 'none')
+      .text(label);
+
+    g.append('text')
+      .attr('x', x + w / 2 - 8)
+      .attr('y', by + h / 2 + 10)
+      .attr('text-anchor', 'middle')
+      .attr('dy', '0.35em')
+      .attr('fill', mutedFg)
+      .attr('font-size', '11px')
+      .attr('font-family', 'inherit')
+      .style('pointer-events', 'none')
+      .text(subtitle);
+
+    const iconX = x + w - 22;
+    const iconY = by + h / 2;
+    g.append('circle')
+      .attr('cx', iconX)
+      .attr('cy', iconY)
+      .attr('r', 10)
+      .attr('fill', isDark ? 'rgba(255,255,255,0.06)' : 'rgba(0,0,0,0.04)')
+      .attr('stroke', borderColor)
+      .attr('stroke-width', 1)
+      .style('pointer-events', 'none');
+    drawCircleGlyph(iconX, iconY, mutedFg, '+');
+
+    g.append('rect')
+      .attr('x', x)
+      .attr('y', by)
+      .attr('width', w)
+      .attr('height', h)
+      .attr('rx', 10)
+      .attr('fill', 'transparent')
+      .style('cursor', 'pointer')
+      .attr('data-testid', `expand-${blockId}`)
+      .on('click', () => {
+        onBlockClick(blockId);
+      });
+  }
+
+  // === TITLE ===
+  const modelName = arch.developer || '';
+  const paramsSummary = [
+    `${formatParamCount(arch.totalParams)} total`,
+    isMoE ? `${formatParamCount(arch.activeParams)} active` : null,
+    arch.contextWindow ? `${formatContextWindow(arch.contextWindow)} context` : null,
+  ]
+    .filter(Boolean)
+    .join(' \u00B7 ');
+
+  g.append('text')
+    .attr('x', cx)
+    .attr('y', titleY + 14)
+    .attr('text-anchor', 'middle')
+    .attr('fill', fg)
+    .attr('font-size', '15px')
+    .attr('font-weight', 700)
+    .attr('font-family', 'inherit')
+    .text(modelName);
+
+  g.append('text')
+    .attr('x', cx)
+    .attr('y', titleY + 32)
+    .attr('text-anchor', 'middle')
+    .attr('fill', mutedFg)
+    .attr('font-size', '11px')
+    .attr('font-family', 'inherit')
+    .text(paramsSummary);
+
+  // === EMBEDDING ===
+  const embedSub = [
+    arch.hiddenSize ? `d = ${arch.hiddenSize.toLocaleString()}` : null,
+    arch.vocabSize ? `vocab = ${arch.vocabSize.toLocaleString()}` : null,
+  ]
+    .filter(Boolean)
+    .join('  \u00B7  ');
+  drawBlock(pad.left, embedY, bw, blockH, 'embedding', 'Token Embedding', embedSub || undefined);
+
+  // A block's incoming arrow should land on its first RMSNorm when the block is
+  // expanded (so there's a continuous line into the norm, through the container
+  // border), or on the block's top when collapsed.
+  const denseEntryY = denseTxExpanded ? denseNorm1Y : denseTxStart;
+  const hashEntryY = hashBlockExpanded ? hashNorm1Y : hashTxStart;
+  const altEntryY = [
+    altBlockExpanded[0] ? altNorm1Y[0] : altBlockStart[0],
+    altBlockExpanded[1] ? altNorm1Y[1] : altBlockStart[1],
+  ];
+  const mainEntryY = txExpanded ? norm1Y : txStart;
+
+  drawArrow(
+    embedY + blockH,
+    hasDenseLayers
+      ? denseEntryY
+      : hasHashBlock
+        ? hashEntryY
+        : hasAlternatingLayers
+          ? altEntryY[0]
+          : mainEntryY,
+  );
+
+  // === DENSE TRANSFORMER BLOCK (for MoE models with initial dense layers) ===
+  if (hasDenseLayers) {
+    if (denseTxExpanded) {
+      // Container
+      g.append('rect')
+        .attr('x', pad.left - 4)
+        .attr('y', denseTxStart)
+        .attr('width', bw + 8)
+        .attr('height', denseTxEnd - denseTxStart)
+        .attr('rx', 10)
+        .attr('fill', 'none')
+        .attr('stroke', borderColor)
+        .attr('stroke-width', 2)
+        .attr('stroke-dasharray', '6,3');
+
+      // Dense layer badge (clickable to collapse)
+      const denseLabel = `\u2212 \u00D7${denseLayerCount} dense layers`;
+      const denseBadgeW = denseLabel.length * 7 + 16;
+      g.append('rect')
+        .attr('x', width - pad.right - denseBadgeW - 4)
+        .attr('y', denseTxStart - 11)
+        .attr('width', denseBadgeW)
+        .attr('height', 22)
+        .attr('rx', 11)
+        .attr('fill', bgSubtle)
+        .attr('stroke', borderColor)
+        .attr('stroke-width', 1);
+      g.append('text')
+        .attr('x', width - pad.right - denseBadgeW / 2 - 4)
+        .attr('y', denseTxStart)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', '11px')
+        .attr('font-weight', 600)
+        .attr('font-family', 'inherit')
+        .text(denseLabel);
+      g.append('rect')
+        .attr('x', width - pad.right - denseBadgeW - 4)
+        .attr('y', denseTxStart - 11)
+        .attr('width', denseBadgeW)
+        .attr('height', 22)
+        .attr('rx', 11)
+        .attr('fill', 'transparent')
+        .style('cursor', 'pointer')
+        .attr('data-testid', 'collapse-denseTransformer')
+        .on('click', () => onBlockClick('denseTransformer'));
+
+      // RMSNorm 1 (Pre-LN)
+      drawBlock(innerX, denseNorm1Y, innerW, smallH, 'norm', 'RMSNorm');
+      drawArrow(denseNorm1Y + smallH, denseAttnY);
+
+      // Attention (expandable only for non-MLA types)
+      const denseAttnLabel = denseLayerAttentionLabel(arch);
+      const denseHeadSub = arch.numHeads ? `${arch.numHeads} heads` : undefined;
+      if (isAttnExpandable) {
+        drawExpandableBlock(
+          innerX,
+          denseAttnY,
+          innerW,
+          blockH,
+          'attention',
+          denseAttnLabel,
+          denseHeadSub,
+          denseAttnExpanded,
+          'denseAttention',
+        );
+        if (denseAttnExpanded && attnFlow) {
+          drawFlow(attnFlow, denseAttnExpandedStartY, innerX, innerW);
+        }
+      } else {
+        drawBlock(innerX, denseAttnY, innerW, blockH, 'attention', denseAttnLabel, denseHeadSub);
+      }
+
+      const denseAttnBottom = denseAttnExpanded
+        ? denseAttnExpandedStartY + denseAttnExpandedH + 4
+        : denseAttnY + blockH + 4;
+      drawArrow(denseAttnBottom, denseMerge1Y - circleR);
+      drawResidualBypass(denseNorm1Y, denseMerge1Y);
+      drawArrow(denseMerge1Y + circleR, denseNorm2Y);
+
+      // RMSNorm 2 (Pre-LN)
+      drawBlock(innerX, denseNorm2Y, innerW, smallH, 'norm', 'RMSNorm');
+      drawArrow(denseNorm2Y + smallH, denseFFNBlockY);
+
+      // Dense FFN (expandable)
+      const denseFFNSub = arch.denseFFNDim
+        ? `intermediate = ${arch.denseFFNDim.toLocaleString()}`
+        : 'Dense Feed-Forward';
+      drawExpandableBlock(
+        innerX,
+        denseFFNBlockY,
+        innerW,
+        blockH,
+        'ffn',
+        'Dense FFN',
+        denseFFNSub,
+        denseFFNExpanded,
+        'denseFFN',
+      );
+
+      if (denseFFNExpanded && denseFFNFlow) {
+        drawFlow(denseFFNFlow, denseFFNExpandedStartY, innerX, innerW, `${ffnVariant} FFN`);
+      }
+
+      const denseFFNBottom = denseFFNExpanded
+        ? denseFFNExpandedStartY + denseFFNExpandedH + 4
+        : denseFFNBlockY + blockH + 4;
+      drawArrow(denseFFNBottom, denseMerge2Y - circleR);
+      drawResidualBypass(denseNorm2Y, denseMerge2Y);
+    } else {
+      // Collapsed dense transformer block
+      const denseSub = `\u00D7${denseLayerCount} dense layers${arch.denseFFNDim ? ` \u00B7 FFN = ${arch.denseFFNDim.toLocaleString()}` : ''}`;
+      drawCollapsedTransformerBlock(
+        pad.left,
+        denseTxStart,
+        bw,
+        collapsedTxH,
+        'Dense Transformer Block',
+        denseSub,
+        'denseTransformer',
+      );
+    }
+
+    // Arrow from dense block to next block
+    drawArrow(
+      denseTxEnd,
+      hasHashBlock ? hashEntryY : hasAlternatingLayers ? altEntryY[0] : mainEntryY,
+    );
+  }
+
+  // Compute labels
+  const attnLabel = getAttentionLabel(arch.attentionType);
+  const mainLayerCount = hasDenseLayers ? moeLayerCount : arch.numLayers;
+  const layerSuffix = hasDenseLayers ? ' MoE layers' : ' layers';
+  const layerLabel = mainLayerCount ? `\u00D7${mainLayerCount}${layerSuffix}` : 'Transformer Block';
+
+  /** Helper: draw a MoE expert grid (used in both alternating and standard blocks) */
+  function drawExpertGrid(
+    eY: number,
+    isExpExpanded: boolean,
+    expandedH: number,
+    expandedStartY: number,
+    n2Y: number,
+    m2Y: number,
+    expertBlockId: string,
+    routerLabel = 'MoE Router',
+    routerSubOverride?: string,
+  ) {
+    const routerSub = routerSubOverride ?? expertRouterSummary(arch);
+    const rY = n2Y + smallH + arrowH;
+    drawBlock(innerX, rY, innerW, blockH, 'router', routerLabel, routerSub);
+    drawArrow(rY + blockH, rY + blockH + arrowH);
+
+    const ec = getColor('expert', isDark);
+    const eac = getColor('expertActive', isDark);
+
+    g.append('rect')
+      .attr('x', innerX)
+      .attr('y', eY)
+      .attr('width', innerW)
+      .attr('height', expertGridH)
+      .attr('rx', 8)
+      .attr('fill', ec.fill)
+      .attr('stroke', ec.stroke)
+      .attr('stroke-width', isExpExpanded ? 2.5 : 1)
+      .attr('stroke-dasharray', '4,3');
+
+    const numActive = arch.activeExperts || 2;
+    const numShow = Math.min(arch.numExperts || 8, 6);
+    const showEllipsis = (arch.numExperts || 0) > numShow;
+    const totalBoxes = numShow + (showEllipsis ? 2 : 0);
+    const totalBoxWidth = totalBoxes * expertSize + (totalBoxes - 1) * expertGap;
+    let ex = cx - totalBoxWidth / 2;
+    const ey = eY + (expertGridH - expertSize) / 2;
+
+    for (let i = 0; i < numShow; i++) {
+      const isActive = i < numActive;
+      g.append('rect')
+        .attr('x', ex)
+        .attr('y', ey)
+        .attr('width', expertSize)
+        .attr('height', expertSize)
+        .attr('rx', 6)
+        .attr('fill', isActive ? eac.fill : bgSubtle)
+        .attr('stroke', isActive ? eac.stroke : borderColor)
+        .attr('stroke-width', isActive ? 1.5 : 1)
+        .attr('opacity', isActive ? 1 : 0.6)
+        .style('pointer-events', 'none');
+      g.append('text')
+        .attr('x', ex + expertSize / 2)
+        .attr('y', ey + expertSize / 2)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', isActive ? fg : mutedFg)
+        .attr('font-size', '9px')
+        .attr('font-weight', isActive ? 600 : 400)
+        .attr('font-family', 'inherit')
+        .style('pointer-events', 'none')
+        .text(`E${i + 1}`);
+      ex += expertSize + expertGap;
+    }
+
+    if (showEllipsis) {
+      g.append('text')
+        .attr('x', ex + expertSize / 2)
+        .attr('y', ey + expertSize / 2)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', '14px')
+        .attr('font-weight', 700)
+        .style('pointer-events', 'none')
+        .text('\u00B7\u00B7\u00B7');
+      ex += expertSize + expertGap;
+      const countText = `\u00D7${arch.numExperts}`;
+      g.append('rect')
+        .attr('x', ex)
+        .attr('y', ey + 3)
+        .attr('width', expertSize)
+        .attr('height', expertSize - 6)
+        .attr('rx', 4)
+        .attr('fill', bgSubtle)
+        .attr('stroke', borderColor)
+        .attr('stroke-width', 1)
+        .style('pointer-events', 'none');
+      g.append('text')
+        .attr('x', ex + expertSize / 2)
+        .attr('y', ey + expertSize / 2)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', '9px')
+        .attr('font-weight', 600)
+        .attr('font-family', 'inherit')
+        .style('pointer-events', 'none')
+        .text(countText);
+    }
+
+    const expIconX = innerX + innerW - 18;
+    const expIconY = eY + expertGridH / 2;
+    g.append('circle')
+      .attr('cx', expIconX)
+      .attr('cy', expIconY)
+      .attr('r', 10)
+      .attr('fill', isDark ? 'rgba(255,255,255,0.06)' : 'rgba(0,0,0,0.04)')
+      .attr('stroke', ec.stroke)
+      .attr('stroke-width', 1)
+      .style('pointer-events', 'none');
+    drawCircleGlyph(expIconX, expIconY, ec.stroke, isExpExpanded ? '\u2212' : '+');
+
+    g.append('rect')
+      .attr('x', innerX)
+      .attr('y', eY)
+      .attr('width', innerW)
+      .attr('height', expertGridH)
+      .attr('rx', 8)
+      .attr('fill', 'transparent')
+      .style('cursor', 'pointer')
+      .attr('data-testid', `expand-${expertBlockId}`)
+      .on('click', () => {
+        onBlockClick(expertBlockId);
+      });
+
+    if (isExpExpanded) {
+      drawFlow(ffnFlow, expandedStartY, innerX, innerW, `Expert FFN (${ffnVariant})`);
+    }
+
+    const expertBottom = isExpExpanded ? expandedStartY + expandedH : eY + expertGridH;
+    drawArrow(expertBottom, m2Y - circleR);
+    drawResidualBypass(n2Y, m2Y);
+  }
+
+  // === HASH-ROUTED MoE PREFIX BLOCK (DeepSeek V4: first N layers route by token id) ===
+  if (hasHashBlock) {
+    if (hashBlockExpanded) {
+      // Container
+      g.append('rect')
+        .attr('x', pad.left - 4)
+        .attr('y', hashTxStart)
+        .attr('width', bw + 8)
+        .attr('height', hashTxEnd - hashTxStart)
+        .attr('rx', 10)
+        .attr('fill', 'none')
+        .attr('stroke', borderColor)
+        .attr('stroke-width', 2)
+        .attr('stroke-dasharray', '6,3');
+
+      // Collapse badge
+      const hashBadge = `− ×${arch.hashRoutedLayers} layers`;
+      const hashBadgeW = hashBadge.length * 7 + 16;
+      g.append('rect')
+        .attr('x', width - pad.right - hashBadgeW - 4)
+        .attr('y', hashTxStart - 11)
+        .attr('width', hashBadgeW)
+        .attr('height', 22)
+        .attr('rx', 11)
+        .attr('fill', bgSubtle)
+        .attr('stroke', borderColor)
+        .attr('stroke-width', 1);
+      g.append('text')
+        .attr('x', width - pad.right - hashBadgeW / 2 - 4)
+        .attr('y', hashTxStart)
+        .attr('text-anchor', 'middle')
+        .attr('dy', '0.35em')
+        .attr('fill', mutedFg)
+        .attr('font-size', '11px')
+        .attr('font-weight', 600)
+        .attr('font-family', 'inherit')
+        .text(hashBadge);
+      g.append('rect')
+        .attr('x', width - pad.right - hashBadgeW - 4)
+        .attr('y', hashTxStart - 11)
+        .attr('width', hashBadgeW)
+        .attr('height', 22)
+        .attr('rx', 11)
+        .attr('fill', 'transparent')
+        .style('cursor', 'pointer')
+        .attr('data-testid', 'collapse-hashBlock')
+        .on('click', () => onBlockClick('hashBlock'));
+
+      // RMSNorm 1
+      drawBlock(innerX, hashNorm1Y, innerW, smallH, 'norm', 'RMSNorm');
+      drawArrow(hashNorm1Y + smallH, hashAttnY);
+
+      // Attention — static hybrid block (same attention stack as the rest)
+      const hashHeadSub = [
+        arch.numHeads ? `${arch.numHeads} heads` : null,
+        arch.numKVHeads ? `${arch.numKVHeads} KV heads` : null,
+      ]
+        .filter(Boolean)
+        .join('  ·  ');
+      drawBlock(
+        innerX,
+        hashAttnY,
+        innerW,
+        blockH,
+        'attention',
+        attnLabel,
+        hashHeadSub || undefined,
+      );
+
+      drawArrow(hashAttnY + blockH + 4, hashMerge1Y - circleR);
+      drawResidualBypass(hashNorm1Y, hashMerge1Y);
+      drawArrow(hashMerge1Y + circleR, hashNorm2Y);
+
+      // RMSNorm 2
+      drawBlock(innerX, hashNorm2Y, innerW, smallH, 'norm', 'RMSNorm');
+      drawArrow(hashNorm2Y + smallH, hashNorm2Y + smallH + arrowH);
+
+      // Hash Router + Expert grid (token-id → fixed experts, not a learned gate)
+      const hashShared = sharedExpertCount(arch);
+      const hashRoutedCount = (arch.numExperts ?? 0) - hashShared;
+      const hashRouterSub = `token-id → ${arch.activeExperts} of ${hashRoutedCount}${hashShared > 0 ? ` + ${hashShared} shared` : ''}`;
+      drawExpertGrid(
+        hashExpertY,
+        hashExpertsExpanded,
+        hashExpertsExpandedH,
+        hashFFNExpandedStartY,
+        hashNorm2Y,
+        hashMerge2Y,
+        'hashExperts',
+        'Hash Router',
+        hashRouterSub,
+      );
+    } else {
+      const hashSub = `×${arch.hashRoutedLayers} first layers · token-id → experts`;
+      drawCollapsedTransformerBlock(
+        pad.left,
+        hashTxStart,
+        bw,
+        collapsedTxH,
+        'Hash-Routed MoE',
+        hashSub,
+        'hashBlock',
+      );
+    }
+
+    // Arrow from the hash block to the first alternating block (or main transformer)
+    drawArrow(hashTxEnd, hasAlternatingLayers ? altEntryY[0] : mainEntryY);
+  }
+
+  // === ALTERNATING TRANSFORMER BLOCKS (gpt-oss style) ===
+  if (hasAlternatingLayers) {
+    for (let bi = 0; bi < 2; bi++) {
+      const spec = alternatingSpecs[bi];
+      const blockId = `altBlock${bi}`;
+      const expertsId = `altExperts${bi}`;
+      const isExp = altBlockExpanded[bi];
+      const isExpExperts = altExpertsExpanded[bi];
+
+      if (isExp) {
+        // === EXPANDED ALTERNATING BLOCK ===
+        g.append('rect')
+          .attr('x', pad.left - 4)
+          .attr('y', altBlockStart[bi])
+          .attr('width', bw + 8)
+          .attr('height', altBlockEnd[bi] - altBlockStart[bi])
+          .attr('rx', 10)
+          .attr('fill', 'none')
+          .attr('stroke', borderColor)
+          .attr('stroke-width', 2)
+          .attr('stroke-dasharray', '6,3');
+
+        // Collapse badge
+        const badgeLabel = `\u2212 \u00D7${spec.count} layers`;
+        const badgeW = badgeLabel.length * 7 + 16;
+        g.append('rect')
+          .attr('x', width - pad.right - badgeW - 4)
+          .attr('y', altBlockStart[bi] - 11)
+          .attr('width', badgeW)
+          .attr('height', 22)
+          .attr('rx', 11)
+          .attr('fill', bgSubtle)
+          .attr('stroke', borderColor)
+          .attr('stroke-width', 1);
+        g.append('text')
+          .attr('x', width - pad.right - badgeW / 2 - 4)
+          .attr('y', altBlockStart[bi])
+          .attr('text-anchor', 'middle')
+          .attr('dy', '0.35em')
+          .attr('fill', mutedFg)
+          .attr('font-size', '11px')
+          .attr('font-weight', 600)
+          .attr('font-family', 'inherit')
+          .text(badgeLabel);
+        g.append('rect')
+          .attr('x', width - pad.right - badgeW - 4)
+          .attr('y', altBlockStart[bi] - 11)
+          .attr('width', badgeW)
+          .attr('height', 22)
+          .attr('rx', 11)
+          .attr('fill', 'transparent')
+          .style('cursor', 'pointer')
+          .attr('data-testid', `collapse-${blockId}`)
+          .on('click', () => onBlockClick(blockId));
+
+        // RMSNorm 1
+        drawBlock(innerX, altNorm1Y[bi], innerW, smallH, 'norm', 'RMSNorm');
+        drawArrow(altNorm1Y[bi] + smallH, altAttnY[bi]);
+
+        // Attention — expandable hybrid drill-down (DeepSeek V4 CSA/HCA) or a
+        // static block (gpt-oss AlternatingSinkGQA).
+        const headSub = [
+          arch.numHeads ? `${arch.numHeads} heads` : null,
+          arch.numKVHeads ? `${arch.numKVHeads} KV heads` : null,
+        ]
+          .filter(Boolean)
+          .join('  \u00B7  ');
+        // Sliding-window note is per layer-type: a hybrid model (e.g. DeepSeek
+        // V4) carries the window on every attention variant, whereas gpt-oss
+        // only puts it on its sliding block. Drive it off the spec, not bi.
+        const specWindow = spec.slidingWindow;
+        const attnSub = specWindow
+          ? `${headSub}${headSub ? '  \u00B7  ' : ''}window=${specWindow}`
+          : headSub || undefined;
+        if (altAttnExpandable) {
+          drawExpandableBlock(
+            innerX,
+            altAttnY[bi],
+            innerW,
+            blockH,
+            'attention',
+            spec.label,
+            attnSub,
+            altAttnExpanded[bi],
+            `altAttention${bi}`,
+          );
+          const flow = altAttnFlow[bi];
+          if (altAttnExpanded[bi] && flow) {
+            drawFlow(flow, altAttnExpandedStartY[bi], innerX, innerW);
+          }
+        } else {
+          drawBlock(innerX, altAttnY[bi], innerW, blockH, 'attention', spec.label, attnSub);
+        }
+
+        const aBottom = altAttnExpanded[bi]
+          ? altAttnExpandedStartY[bi] + altAttnExpandedH[bi] + 4
+          : altAttnY[bi] + blockH + 4;
+        drawArrow(aBottom, altMerge1Y[bi] - circleR);
+        drawResidualBypass(altNorm1Y[bi], altMerge1Y[bi]);
+        drawArrow(altMerge1Y[bi] + circleR, altNorm2Y[bi]);
+
+        // RMSNorm 2
+        drawBlock(innerX, altNorm2Y[bi], innerW, smallH, 'norm', 'RMSNorm');
+        drawArrow(altNorm2Y[bi] + smallH, altNorm2Y[bi] + smallH + arrowH);
+
+        // MoE Router + Expert Grid
+        drawExpertGrid(
+          altExpertY[bi],
+          isExpExperts,
+          altExpertsExpandedH[bi],
+          altFFNExpandedStartY[bi],
+          altNorm2Y[bi],
+          altMerge2Y[bi],
+          expertsId,
+        );
+      } else {
+        // === COLLAPSED ALTERNATING BLOCK ===
+        const collapsedSub = `\u00D7${spec.count} layers \u00B7 Top-${arch.activeExperts}/${arch.numExperts} MoE`;
+        drawCollapsedTransformerBlock(
+          pad.left,
+          altBlockStart[bi],
+          bw,
+          collapsedTxH,
+          spec.label,
+          collapsedSub,
+          blockId,
+        );
+      }
+
+      // Draw alternating indicator between the two blocks
+      if (bi === 0) {
+        // Arrow from block 0 end through indicator to block 1 start (drawn first, behind text)
+        drawArrow(altBlockEnd[0] + 2, altEntryY[1]);
+
+        // Opaque background rect behind the label so it doesn't overlap the arrow
+        const cardBg = isDark ? '#131416' : '#eaebec';
+        const labelText = `\u21C5 ${arch.alternatingNote ?? 'alternating every layer'}`;
+        const labelPadX = 6;
+        const labelPadY = 4;
+        const labelFontSize = 10;
+        const estLabelW = labelText.length * labelFontSize * 0.55 + labelPadX * 2;
+        const estLabelH = labelFontSize + labelPadY * 2;
+        g.append('rect')
+          .attr('x', cx - estLabelW / 2)
+          .attr('y', altIndicatorY - estLabelH / 2)
+          .attr('width', estLabelW)
+          .attr('height', estLabelH)
+          .attr('rx', 4)
+          .attr('fill', cardBg);
+
+        g.append('text')
+          .attr('x', cx)
+          .attr('y', altIndicatorY)
+          .attr('text-anchor', 'middle')
+          .attr('dy', '0.35em')
+          .attr('fill', mutedFg)
+          .attr('font-size', `${labelFontSize}px`)
+          .attr('font-weight', 500)
+          .attr('font-family', 'inherit')
+          .attr('data-testid', 'alternating-indicator')
+          .text(labelText);
+      }
+    }
+
+    // Arrow from last alternating block to final norm
+    drawArrow(altBlockEnd[1], finalNormY);
+  } else if (txExpanded) {
+    // === MAIN TRANSFORMER CONTAINER (expanded, non-alternating) ===
+    g.append('rect')
+      .attr('x', pad.left - 4)
+      .attr('y', txStart)
+      .attr('width', bw + 8)
+      .attr('height', txEnd - txStart)
+      .attr('rx', 10)
+      .attr('fill', 'none')
+      .attr('stroke', borderColor)
+      .attr('stroke-width', 2)
+      .attr('stroke-dasharray', '6,3');
+
+    // Layer badge (clickable to collapse)
+    const collapseBadgeLabel = `\u2212 ${layerLabel}`;
+    const badgeW = collapseBadgeLabel.length * 7 + 16;
+    g.append('rect')
+      .attr('x', width - pad.right - badgeW - 4)
+      .attr('y', txStart - 11)
+      .attr('width', badgeW)
+      .attr('height', 22)
+      .attr('rx', 11)
+      .attr('fill', bgSubtle)
+      .attr('stroke', borderColor)
+      .attr('stroke-width', 1);
+    g.append('text')
+      .attr('x', width - pad.right - badgeW / 2 - 4)
+      .attr('y', txStart)
+      .attr('text-anchor', 'middle')
+      .attr('dy', '0.35em')
+      .attr('fill', mutedFg)
+      .attr('font-size', '11px')
+      .attr('font-weight', 600)
+      .attr('font-family', 'inherit')
+      .text(collapseBadgeLabel);
+    g.append('rect')
+      .attr('x', width - pad.right - badgeW - 4)
+      .attr('y', txStart - 11)
+      .attr('width', badgeW)
+      .attr('height', 22)
+      .attr('rx', 11)
+      .attr('fill', 'transparent')
+      .style('cursor', 'pointer')
+      .attr('data-testid', 'collapse-transformer')
+      .on('click', () => onBlockClick('transformer'));
+
+    // === RMSNORM 1 (before attention — Pre-LN) ===
+    drawBlock(innerX, norm1Y, innerW, smallH, 'norm', 'RMSNorm');
+    drawArrow(norm1Y + smallH, attnY);
+
+    // === ATTENTION ===
+    const headSub = [
+      arch.numHeads ? `${arch.numHeads} heads` : null,
+      arch.numKVHeads ? `${arch.numKVHeads} KV heads` : null,
+    ]
+      .filter(Boolean)
+      .join('  \u00B7  ');
+
+    if (isAttnExpandable) {
+      drawExpandableBlock(
+        innerX,
+        attnY,
+        innerW,
+        blockH,
+        'attention',
+        attnLabel,
+        headSub || undefined,
+        attnExpanded,
+        'attention',
+      );
+      if (attnExpanded && attnFlow) {
+        drawFlow(attnFlow, attnExpandedStartY, innerX, innerW);
+      }
+    } else {
+      // MLA: non-expandable static block
+      drawBlock(innerX, attnY, innerW, blockH, 'attention', attnLabel, headSub || undefined);
+    }
+
+    const attnBottom = attnExpanded ? attnExpandedStartY + attnExpandedH + 4 : attnY + blockH + 4;
+    drawArrow(attnBottom, merge1Y - circleR);
+    drawResidualBypass(norm1Y, merge1Y);
+    drawArrow(merge1Y + circleR, norm2Y);
+
+    // === RMSNORM 2 (before FFN — Pre-LN) ===
+    drawBlock(innerX, norm2Y, innerW, smallH, 'norm', 'RMSNorm');
+    drawArrow(norm2Y + smallH, norm2Y + smallH + arrowH);
+
+    if (isMoE) {
+      // === ROUTER + EXPERTS (via helper) ===
+      drawExpertGrid(
+        expertY,
+        ffnExpanded,
+        ffnExpandedH,
+        ffnExpandedStartY,
+        norm2Y,
+        merge2Y,
+        'experts',
+      );
+    } else {
+      // === FFN (expandable) for dense models ===
+      const ffnSub = arch.ffnDim
+        ? `intermediate = ${arch.ffnDim.toLocaleString()}`
+        : `${ffnVariant} activation`;
+      drawExpandableBlock(
+        innerX,
+        ffnY,
+        innerW,
+        blockH,
+        'ffn',
+        'Feed-Forward Network',
+        ffnSub,
+        ffnExpanded,
+        'ffn',
+      );
+
+      if (ffnExpanded) {
+        drawFlow(ffnFlow, ffnExpandedStartY, innerX, innerW, `${ffnVariant} Details`);
+      }
+
+      const ffnBottom = ffnExpanded ? ffnExpandedStartY + ffnExpandedH : ffnY + blockH;
+      drawArrow(ffnBottom, merge2Y - circleR);
+      drawResidualBypass(norm2Y, merge2Y);
+    }
+  } else {
+    // === MAIN TRANSFORMER (collapsed) ===
+    const collapsedLabel = isMoE ? 'MoE Transformer Block' : 'Dense Transformer Block';
+    const collapsedSub = isMoE
+      ? `${layerLabel} \u00B7 ${attnLabel} \u00B7 Top-${arch.activeExperts}/${arch.numExperts}`
+      : `${layerLabel} \u00B7 ${attnLabel}`;
+    drawCollapsedTransformerBlock(
+      pad.left,
+      txStart,
+      bw,
+      collapsedTxH,
+      collapsedLabel,
+      collapsedSub,
+      'transformer',
+    );
+  }
+
+  if (!hasAlternatingLayers) {
+    // Arrow from transformer container to final norm
+    drawArrow(txEnd, finalNormY);
+  }
+
+  // === FINAL RMSNORM ===
+  drawBlock(pad.left, finalNormY, bw, smallH, 'norm', 'RMSNorm');
+  drawArrow(finalNormY + smallH, outputY);
+
+  // === OUTPUT HEAD ===
+  const outSub = arch.vocabSize ? `vocab = ${arch.vocabSize.toLocaleString()}` : undefined;
+  drawBlock(pad.left, outputY, bw, blockH, 'output', 'Output Head (LM Head)', outSub);
+
+  // === SPECS BAR ===
+  const specItems: { label: string; value: string }[] = [
+    { label: 'Type', value: isMoE ? 'MoE' : 'Dense' },
+    {
+      label: 'Layers',
+      value: hasDenseLayers
+        ? `${denseLayerCount}D + ${moeLayerCount}M`
+        : arch.numLayers
+          ? `${arch.numLayers}`
+          : '\u2014',
+    },
+    {
+      label: 'Attention',
+      value: arch.attentionType === 'AlternatingSinkGQA' ? 'Sink/Full GQA' : arch.attentionType,
+    },
+    {
+      label: 'Context',
+      value: arch.contextWindow ? formatContextWindow(arch.contextWindow) : '\u2014',
+    },
+    ...(isMoE
+      ? [
+          {
+            label: 'Experts',
+            // Active per token = routed top-k + the always-on shared experts, so
+            // show "6+1/385" (not "6/385"): the shared experts are active too.
+            // Count them from the model (Kimi K3 has 2), never assume one.
+            value: `${arch.activeExperts}${sharedExperts > 0 ? `+${sharedExperts}` : ''}/${arch.numExperts}`,
+          },
+        ]
+      : []),
+  ];
+
+  const specColW = bw / specItems.length;
+
+  g.append('rect')
+    .attr('x', pad.left)
+    .attr('y', specsY)
+    .attr('width', bw)
+    .attr('height', specsH)
+    .attr('rx', 8)
+    .attr('fill', bgSubtle)
+    .attr('stroke', borderColor)
+    .attr('stroke-width', 1);
+
+  specItems.forEach((spec, i) => {
+    const sx = pad.left + i * specColW + specColW / 2;
+
+    g.append('text')
+      .attr('x', sx)
+      .attr('y', specsY + 18)
+      .attr('text-anchor', 'middle')
+      .attr('fill', mutedFg)
+      .attr('font-size', '10px')
+      .attr('font-weight', 500)
+      .attr('font-family', 'inherit')
+      .text(spec.label);
+
+    g.append('text')
+      .attr('x', sx)
+      .attr('y', specsY + 38)
+      .attr('text-anchor', 'middle')
+      .attr('fill', fg)
+      .attr('font-size', '14px')
+      .attr('font-weight', 700)
+      .attr('font-family', 'inherit')
+      .text(spec.value);
+
+    if (i < specItems.length - 1) {
+      g.append('line')
+        .attr('x1', pad.left + (i + 1) * specColW)
+        .attr('y1', specsY + 8)
+        .attr('x2', pad.left + (i + 1) * specColW)
+        .attr('y2', specsY + specsH - 8)
+        .attr('stroke', borderColor)
+        .attr('stroke-width', 1);
+    }
+  });
+}
diff --git a/packages/app/src/components/inference/ui/point-label.ts b/packages/app/src/components/inference/ui/point-label.ts
new file mode 100644
index 000000000..3625b555a
--- /dev/null
+++ b/packages/app/src/components/inference/ui/point-label.ts
@@ -0,0 +1,13 @@
+import type { InferenceData } from '@/components/inference/types';
+import { getPointLabel } from '@/components/inference/utils/tooltipUtils';
+
+export function pointLabelText(
+  point: InferenceData,
+  advanced: boolean,
+  showConcurrency: boolean,
+): string {
+  const base = advanced ? getPointLabel(point) : `${point.tp}`;
+  // Concurrency is opt-in via the advanced "# Concurrent Sessions" toggle so
+  // the default labels stay light-weight.
+  return showConcurrency ? `${base}\nC=${point.conc}` : base;
+}
diff --git a/packages/app/src/components/inference/utils.test.ts b/packages/app/src/components/inference/utils.test.ts
index c69a9cecc..fecdc48c8 100644
--- a/packages/app/src/components/inference/utils.test.ts
+++ b/packages/app/src/components/inference/utils.test.ts
@@ -308,15 +308,48 @@ describe('processOverlayChartData', () => {
     expect(result[0].x).toBe(0.5);
   });
 
-  it('applies cost limit filtering', () => {
+  it('keeps all unofficial-run tokens-per-dollar points without the former cost clamp', () => {
     const data = [
-      pt({ costh: { y: 0.5, roof: false }, median_intvty: 10 } as any),
-      pt({ costh: { y: 100, roof: false }, median_intvty: 20 } as any),
+      pt({ tokensPerDollarH: { y: 500_000, roof: false }, median_intvty: 10 } as any),
+      pt({ tokensPerDollarH: { y: 2_000_000, roof: false }, median_intvty: 20 } as any),
     ];
-    // interactivity chart config has y_cost_limit: 5
+    const result = processOverlayChartData(data, 'interactivity', 'y_tokensPerDollarH', null);
+    expect(result.map((point) => point.y)).toEqual([500_000, 2_000_000]);
+  });
+
+  it('keeps the existing unofficial-run cost-per-million metric independent', () => {
+    const data = [pt({ costh: { y: 0.5, roof: false }, median_intvty: 20 } as any)];
     const result = processOverlayChartData(data, 'interactivity', 'y_costh', null);
+
     expect(result).toHaveLength(1);
     expect(result[0].y).toBe(0.5);
+    expect(result[0].costh.y).toBe(0.5);
+  });
+
+  it('does not classify unofficial-run purchasing-power points as cost overflows', () => {
+    const visible = pt({
+      tokensPerDollarH: { y: 500_000, roof: false },
+      median_intvty: 10,
+      run_url: 'https://github.com/SemiAnalysisAI/InferenceX/actions/runs/123',
+    } as any);
+    const highValue = pt({
+      tokensPerDollarH: { y: 2_000_000, roof: false },
+      median_intvty: 20,
+      run_url: 'https://github.com/SemiAnalysisAI/InferenceX/actions/runs/123',
+    } as any);
+
+    const result = processOverlayChartDataWithClipping(
+      [visible, highValue],
+      'interactivity',
+      'y_tokensPerDollarH',
+      null,
+    );
+
+    expect(result.data).toEqual([
+      expect.objectContaining({ x: 10, y: 500_000, run_url: visible.run_url }),
+      expect.objectContaining({ x: 20, y: 2_000_000, run_url: highValue.run_url }),
+    ]);
+    expect(result.clippedData).toEqual([]);
   });
 
   it('retains clipped unofficial-run points for the overflow continuation path', () => {
diff --git a/packages/app/src/components/inference/utils.ts b/packages/app/src/components/inference/utils.ts
index 718f408c2..8ab4285ce 100644
--- a/packages/app/src/components/inference/utils.ts
+++ b/packages/app/src/components/inference/utils.ts
@@ -4,8 +4,9 @@
  * For Pareto front calculations, see @/lib/chart-utils
  */
 
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
+import chartDefinitions from '@/components/inference/metric-registry';
 import { resolveXAxisField } from '@/components/inference/utils/resolveXAxisField';
+import { remapInferencePoint } from '@/lib/chart-utils';
 
 import type { ChartDefinition, ClippedInferenceData, InferenceData, YAxisMetricKey } from './types';
 
@@ -188,11 +189,7 @@ export function processOverlayChartDataWithClipping(
 
   const processedData = data
     .filter((d) => metricKey in d)
-    .map((d: InferenceData) => {
-      const yValue = (d[metricKey] as { y: number })?.y ?? d.y;
-      const xValue = (d as any)[xAxisField] ?? d.x;
-      return { ...d, x: xValue, y: yValue };
-    });
+    .map((d) => remapInferencePoint(d, metricKey, xAxisField));
 
   // The normalized metric is derived from persisted request traces, which an
   // unofficial overlay does not have. An all-false canonical stamp prevents a
diff --git a/packages/app/src/components/inference/utils/best-series-per-sku.test.ts b/packages/app/src/components/inference/utils/best-series-per-sku.test.ts
index 266b75709..9c3f4421d 100644
--- a/packages/app/src/components/inference/utils/best-series-per-sku.test.ts
+++ b/packages/app/src/components/inference/utils/best-series-per-sku.test.ts
@@ -8,6 +8,16 @@ function point(hw: string, hwKey: string, x: number, y: number): InferenceData {
   return { hw, hwKey, x, y } as InferenceData;
 }
 
+function frameworkPoint(
+  hw: string,
+  hwKey: string,
+  framework: string,
+  x: number,
+  y: number,
+): InferenceData {
+  return { ...point(hw, hwKey, x, y), framework };
+}
+
 describe('bestSeriesPerSku', () => {
   it('selects the highest normalized frontier AUC within each SKU', () => {
     const selected = bestSeriesPerSku(
@@ -57,6 +67,21 @@ describe('bestSeriesPerSku', () => {
     expect(baseSku(point('GB200-NVL72', 'gb200_dynamo-trt', 1, 1))).toBe('GB200');
   });
 
+  it('keeps every eligible TileRT series alongside the SKU AUC winner', () => {
+    const selected = bestSeriesPerSku(
+      [
+        frameworkPoint('B200-8', 'b200_sglang', 'sglang', 20, 1_800),
+        frameworkPoint('B200-8', 'b200_sglang', 'sglang', 60, 1_000),
+        frameworkPoint('B200-8', 'b200_tilert_mtp', 'tilert', 340, 160),
+        frameworkPoint('B200-8', 'b200_tilert_stp', 'tilert', 300, 140),
+        frameworkPoint('B200-8', 'b200_unrelated_endpoint', 'vllm', 400, 120),
+      ],
+      'upper_left',
+    );
+
+    expect(selected).toEqual(new Set(['b200_sglang', 'b200_tilert_mtp', 'b200_tilert_stp']));
+  });
+
   it('ranks unofficial-run overlay series with the same SKU policy', () => {
     const overlayPoint = (hwKey: string, x: number, y: number) =>
       ({
@@ -75,4 +100,22 @@ describe('bestSeriesPerSku', () => {
 
     expect(selected).toEqual(new Set(['b200_overlay_sglang']));
   });
+
+  it('keeps TileRT series from an unofficial-run overlay', () => {
+    const overlayPoint = (hwKey: string, framework: string, x: number, y: number) =>
+      ({
+        ...frameworkPoint('B200-8', hwKey, framework, x, y),
+        run_url: 'https://github.com/example/actions/runs/123',
+      }) as InferenceData;
+    const selected = bestSeriesPerSku(
+      [
+        overlayPoint('b200_overlay_sglang', 'sglang', 20, 1_800),
+        overlayPoint('b200_overlay_sglang', 'sglang', 60, 1_000),
+        overlayPoint('b200_overlay_tilert_mtp', 'tilert', 340, 160),
+      ],
+      'upper_left',
+    );
+
+    expect(selected).toEqual(new Set(['b200_overlay_sglang', 'b200_overlay_tilert_mtp']));
+  });
 });
diff --git a/packages/app/src/components/inference/utils/best-series-per-sku.ts b/packages/app/src/components/inference/utils/best-series-per-sku.ts
index db863e0d2..590a232af 100644
--- a/packages/app/src/components/inference/utils/best-series-per-sku.ts
+++ b/packages/app/src/components/inference/utils/best-series-per-sku.ts
@@ -22,19 +22,25 @@ interface ScoredSeries {
 }
 
 /**
- * Select one configuration line per physical SKU using normalized frontier AUC.
+ * Select the best configuration line per physical SKU using normalized frontier AUC.
  *
  * Every candidate is sampled over the common measured x-domain for that SKU,
  * so a curve cannot win merely because it spans a wider range. The chart's
  * existing Pareto direction and monotone interpolation are reused to keep the
  * ranking aligned with the line users see. Ties are deterministic by hwKey.
+ *
+ * TileRT is the deliberate exception to the one-line policy. Its specialized
+ * operating points can be exceptional even when a broader serving curve wins
+ * the AUC comparison, so every eligible TileRT series remains visible.
  */
 export function bestSeriesPerSku(points: InferenceData[], direction: Direction): Set<string> {
   const bySku = new Map<string, Map<string, InferenceData[]>>();
+  const featured = new Set<string>();
   for (const point of points) {
     if (!isFrontierEligible(point) || !Number.isFinite(point.y)) continue;
     const sku = baseSku(point);
     const key = String(point.hwKey);
+    if (point.framework === 'tilert') featured.add(key);
     let series = bySku.get(sku);
     if (!series) {
       series = new Map();
@@ -45,7 +51,7 @@ export function bestSeriesPerSku(points: InferenceData[], direction: Direction):
     series.set(key, rows);
   }
 
-  const selected = new Set<string>();
+  const selected = new Set(featured);
   const higherYIsBetter = direction.startsWith('upper');
 
   for (const series of bySku.values()) {
diff --git a/packages/app/src/components/inference/utils/knownIssueAnnotations.test.ts b/packages/app/src/components/inference/utils/knownIssueAnnotations.test.ts
index c27b78b36..24025247d 100644
--- a/packages/app/src/components/inference/utils/knownIssueAnnotations.test.ts
+++ b/packages/app/src/components/inference/utils/knownIssueAnnotations.test.ts
@@ -30,6 +30,28 @@ const mi355xAnnotation: KnownIssueAnnotation = {
   points: [{ x: 150, y: 250 }],
 };
 
+const jalapenoPreview: KnownIssueAnnotation = {
+  preview: {
+    id: 'jalapeno-official-preview',
+    summary: 'InferenceX Official Preview',
+    detail: 'Results may change as validation and publication continue.',
+  },
+  label: 'Jalapeño (Teacup)',
+  color: 'rgb(32, 201, 151)',
+  points: [{ x: 180, y: 280 }],
+};
+
+const veraRubinPreview: KnownIssueAnnotation = {
+  preview: {
+    id: 'vera-rubin-official-preview',
+    summary: 'InferenceX Official Preview',
+    detail: 'Results may change as validation and publication continue.',
+  },
+  label: 'Vera Rubin (July)',
+  color: 'rgb(118, 185, 0)',
+  points: [{ x: 210, y: 240 }],
+};
+
 function baseOptions(overrides: Partial<AnnotationRenderOptions> = {}): AnnotationRenderOptions {
   return {
     chartId: 'chart-0',
@@ -71,7 +93,8 @@ describe('renderKnownIssueAnnotations', () => {
     expect(boxes[0].getAttribute('href')).toBe('https://github.com/NVIDIA/srt-slurm/issues/51');
     expect(boxes[0].getAttribute('target')).toBe('_blank');
     expect(boxes[0].textContent).toContain('GB300 NVL72 (Dynamo TRTLLM, MTP)');
-    expect(boxes[0].textContent).toContain('Accuracy issues — filed since Apr 21, 2026');
+    expect(boxes[0].textContent).toContain('Accuracy issues: filed since Apr 21, 2026');
+    expect(boxes[0].textContent).not.toContain('\u2014');
     expect(boxes[0].textContent).toContain('NVIDIA/srt-slurm#51');
 
     expect(boxes[1].getAttribute('href')).toBe(
@@ -80,6 +103,37 @@ describe('renderKnownIssueAnnotations', () => {
     expect(boxes[1].textContent).toContain('MI355X (MoRI SGLang, MTP)');
   });
 
+  it('renders an unlinked preview notice in the same on-chart annotation layer', () => {
+    renderKnownIssueAnnotations(g, defs, baseOptions({ annotations: [jalapenoPreview] }));
+
+    const notice = g.select('[data-testid="jalapeno-official-preview-notice"]');
+    expect(notice.empty()).toBe(false);
+    expect(notice.attr('role')).toBe('note');
+    expect(notice.attr('href')).toBeNull();
+    expect(notice.attr('data-preview-id')).toBe('jalapeno-official-preview');
+    expect(notice.text()).toContain('Jalapeño (Teacup)');
+    expect(notice.text()).toContain('InferenceX Official Preview');
+    expect(notice.text()).toContain('InferenceX Official Preview: Results may change');
+    expect(notice.text()).toContain('Results may change as validation and publication continue.');
+    expect(notice.text()).not.toContain('\u2014');
+    expect(g.selectAll('.known-issue-arrow').nodes()).toHaveLength(1);
+  });
+
+  it('assigns separate selectors to Jalapeño and Vera Rubin preview notices', () => {
+    renderKnownIssueAnnotations(
+      g,
+      defs,
+      baseOptions({ annotations: [jalapenoPreview, veraRubinPreview] }),
+    );
+
+    expect(g.select('[data-testid="jalapeno-official-preview-notice"]').empty()).toBe(false);
+    const rubinNotice = g.select('[data-testid="vera-rubin-official-preview-notice"]');
+    expect(rubinNotice.empty()).toBe(false);
+    expect(rubinNotice.text()).toContain('Vera Rubin (July)');
+    expect(rubinNotice.text()).not.toContain('\u2014');
+    expect(g.selectAll('.known-issue-arrow').nodes()).toHaveLength(2);
+  });
+
   it('stacks boxes without overlap, right-aligned inside the plot', () => {
     renderKnownIssueAnnotations(g, defs, baseOptions());
 
diff --git a/packages/app/src/components/inference/utils/knownIssueAnnotations.ts b/packages/app/src/components/inference/utils/knownIssueAnnotations.ts
index 5fce7fcf7..a77dd8a61 100644
--- a/packages/app/src/components/inference/utils/knownIssueAnnotations.ts
+++ b/packages/app/src/components/inference/utils/knownIssueAnnotations.ts
@@ -6,12 +6,14 @@
  * variables don't survive html-to-image.
  */
 
-import type * as d3 from 'd3';
+import * as d3 from 'd3';
 
 import type { KnownConfigIssue } from '@/lib/known-issues';
+import type { CustomLayerConfig, RenderContext } from '@/lib/d3-chart/D3Chart/types';
+import type { ContinuousScale } from '@/lib/d3-chart/types';
+import { CHART_TYPE } from '@/lib/d3-chart/typography';
 
-export interface KnownIssueAnnotation {
-  issue: KnownConfigIssue;
+interface AnnotationBase {
   /** Display label of the affected series, e.g. "GB300 NVL72 (Dynamo TRTLLM, MTP)" */
   label: string;
   /** Resolved stroke color of the affected series */
@@ -20,6 +22,20 @@ export interface KnownIssueAnnotation {
   points: { x: number; y: number }[];
 }
 
+/** A filed upstream issue or a non-interactive preview notice rendered inside a chart. */
+export type KnownIssueAnnotation = AnnotationBase &
+  (
+    | { issue: KnownConfigIssue; preview?: never }
+    | {
+        issue?: never;
+        preview: {
+          id: string;
+          summary: string;
+          detail: string;
+        };
+      }
+  );
+
 export interface AnnotationRenderOptions {
   chartId: string;
   width: number;
@@ -38,13 +54,64 @@ export interface AnnotationRenderOptions {
   onLinkClick?: (annotation: KnownIssueAnnotation) => void;
 }
 
+export type KnownIssueLayerOptions = Omit<
+  AnnotationRenderOptions,
+  'width' | 'height' | 'xScale' | 'yScale' | 'rightInset'
+>;
+
+/** Build the shared render/zoom lifecycle for chart-level known-issue callouts. */
+export function createKnownIssueLayer(
+  resolveOptions: () => KnownIssueLayerOptions,
+  displayIdentity?: string,
+): CustomLayerConfig {
+  const draw = (ctx: RenderContext, xScale: ContinuousScale, yScale: ContinuousScale): void => {
+    const options = resolveOptions();
+    renderKnownIssueAnnotations(ctx.layout.g, ctx.layout.defs, {
+      ...options,
+      width: ctx.width,
+      height: ctx.height,
+      xScale,
+      yScale,
+      rightInset:
+        options.annotations.length === 0
+          ? 0
+          : measureLegendRightInset(
+              options.chartId,
+              ctx.layout.svg.node(),
+              ctx.layout.margin.left,
+              ctx.width,
+            ),
+    });
+  };
+
+  return {
+    type: 'custom',
+    key: 'known-issues',
+    displayIdentity,
+    render: (_zoomGroup, ctx) =>
+      draw(ctx, ctx.xScale as ContinuousScale, ctx.yScale as ContinuousScale),
+    onDisplayUpdate: (_zoomGroup, ctx) => {
+      const svg = ctx.layout.svg.node();
+      if (!svg) return;
+      const transform = d3.zoomTransform(svg);
+      draw(
+        ctx,
+        transform.rescaleX(ctx.xScale as ContinuousScale),
+        transform.rescaleY(ctx.yScale as ContinuousScale),
+      );
+    },
+    onZoom: (_zoomGroup, ctx) =>
+      draw(ctx, ctx.newXScale as ContinuousScale, ctx.newYScale as ContinuousScale),
+  };
+}
+
 const BOX_TOP = 8;
 const BOX_GAP = 8;
 const BOX_RIGHT_GAP = 10;
 const PAD_X = 10;
 const PAD_Y = 7;
-const LINE1_SIZE = 11;
-const LINE2_SIZE = 10;
+const LINE1_SIZE = CHART_TYPE.annotation;
+const LINE2_SIZE = CHART_TYPE.annotationSub;
 const LINE1_H = 14;
 const LINE2_H = 13;
 const SWATCH_R = 4;
@@ -104,8 +171,7 @@ export function renderKnownIssueAnnotations(
 
   let yCursor = BOX_TOP;
   annotations.forEach((annotation, index) => {
-    const { issue, label, color, points } = annotation;
-
+    const { issue, preview, label, color, points } = annotation;
     const markerId = `known-issue-arrowhead-${chartId}-${index}`;
     defs
       .append('marker')
@@ -121,16 +187,30 @@ export function renderKnownIssueAnnotations(
       .attr('fill', color);
 
     const anchor = layer
-      .append('a')
-      .attr('href', issue.url)
-      .attr('target', '_blank')
-      .attr('rel', 'noopener noreferrer')
+      .append(issue === undefined ? 'g' : 'a')
       .attr('class', 'known-issue-annotation')
-      .attr('data-testid', 'known-issue-annotation')
-      .attr('cursor', 'pointer')
-      .on('click', () => opts.onLinkClick?.(annotation));
+      .attr(
+        'data-testid',
+        preview === undefined ? 'known-issue-annotation' : `${preview.id}-notice`,
+      )
+      .attr('cursor', issue === undefined ? 'default' : 'pointer');
+    if (issue === undefined) {
+      anchor
+        .attr('role', 'note')
+        .attr('aria-label', preview?.summary ?? null)
+        .attr('data-preview-id', preview?.id ?? null);
+    } else {
+      anchor
+        .attr('href', issue.url)
+        .attr('target', '_blank')
+        .attr('rel', 'noopener noreferrer')
+        .on('click', () => opts.onLinkClick?.(annotation));
+    }
 
-    const detail = `${issue.summary} — filed since ${issue.filed} · `;
+    const detail =
+      issue === undefined
+        ? `${preview!.summary}: ${preview!.detail}`
+        : `${issue.summary}: filed since ${issue.filed} · `;
     const text1 = anchor
       .append('text')
       .attr('font-size', LINE1_SIZE)
@@ -139,14 +219,20 @@ export function renderKnownIssueAnnotations(
       .text(label);
     const text2 = anchor.append('text').attr('font-size', LINE2_SIZE);
     text2.append('tspan').attr('fill', opts.mutedForeground).text(detail);
-    text2
-      .append('tspan')
-      .attr('fill', opts.foreground)
-      .attr('text-decoration', 'underline')
-      .text(issue.issueRef);
+    if (issue !== undefined) {
+      text2
+        .append('tspan')
+        .attr('fill', opts.foreground)
+        .attr('text-decoration', 'underline')
+        .text(issue.issueRef);
+    }
 
     const w1 = measureTextWidth(text1.node(), label.length, LINE1_SIZE) + SWATCH_SPACE;
-    const w2 = measureTextWidth(text2.node(), detail.length + issue.issueRef.length, LINE2_SIZE);
+    const w2 = measureTextWidth(
+      text2.node(),
+      detail.length + (issue?.issueRef.length ?? 0),
+      LINE2_SIZE,
+    );
     const boxW = Math.max(w1, w2) + PAD_X * 2;
     const boxH = PAD_Y * 2 + LINE1_H + LINE2_H;
     const boxRight = width - BOX_RIGHT_GAP - (opts.rightInset ?? 0);
diff --git a/packages/app/src/components/inference/utils/label-defaults.test.ts b/packages/app/src/components/inference/utils/label-defaults.test.ts
index e1264b5c0..cee2493b0 100644
--- a/packages/app/src/components/inference/utils/label-defaults.test.ts
+++ b/packages/app/src/components/inference/utils/label-defaults.test.ts
@@ -11,11 +11,13 @@ describe('resolveLabelState', () => {
     });
   });
 
-  it('uses parallelism labels instead of line labels for agentic scenarios', () => {
+  it('turns on parallelism labels for agentic scenarios, and keeps line labels', () => {
+    // Line labels name the curve, point labels name the point — the agentic
+    // view wants both, so it differs from fixed-seq only in the point labels.
     expect(resolveLabelState('agentic', {})).toEqual({
       showPointLabels: true,
       useAdvancedLabels: true,
-      showLineLabels: false,
+      showLineLabels: true,
     });
   });
 
@@ -48,12 +50,12 @@ describe('serializeLabelState', () => {
       serializeLabelState('agentic', {
         showPointLabels: false,
         useAdvancedLabels: false,
-        showLineLabels: true,
+        showLineLabels: false,
       }),
     ).toEqual({
       i_label: '0',
       i_advlabel: '0',
-      i_linelabel: '1',
+      i_linelabel: '0',
     });
   });
 
diff --git a/packages/app/src/components/inference/utils/label-defaults.ts b/packages/app/src/components/inference/utils/label-defaults.ts
index 7ad6c624c..5a7a70100 100644
--- a/packages/app/src/components/inference/utils/label-defaults.ts
+++ b/packages/app/src/components/inference/utils/label-defaults.ts
@@ -22,8 +22,11 @@ export function resolveLabelState(kind: ScenarioKind, params: LabelUrlParams): L
     showPointLabels,
     useAdvancedLabels:
       params.i_advlabel === '1' ? true : params.i_advlabel === '0' ? false : agentic,
-    showLineLabels:
-      params.i_linelabel === '1' ? true : params.i_linelabel === '0' ? false : !agentic,
+    // Line labels are on by default in every scenario, agentic included. They
+    // name the curve rather than the point, so they stay readable where the
+    // per-point labels the agentic view turns on would otherwise be the only
+    // way to tell series apart.
+    showLineLabels: params.i_linelabel === '1' ? true : params.i_linelabel !== '0',
   };
 }
 
diff --git a/packages/app/src/components/inference/utils/legend-points-table.test.ts b/packages/app/src/components/inference/utils/legend-points-table.test.ts
index 86d6f8b37..f7dc70d39 100644
--- a/packages/app/src/components/inference/utils/legend-points-table.test.ts
+++ b/packages/app/src/components/inference/utils/legend-points-table.test.ts
@@ -49,6 +49,16 @@ describe('pointDetailHref', () => {
     });
   });
 
+  it('keeps a Chinese reader on /zh', () => {
+    // The bare path used to hand /zh readers to the English detail page, and
+    // every link out of that page kept them in English.
+    const d = pt({ benchmark_type: 'agentic_traces', id: 206863 });
+    expect(pointDetailHref(d, false, 'zh')).toEqual({
+      href: '/zh/inference/agentic/206863',
+      isExternal: false,
+    });
+  });
+
   it('fixed-seq point links to its GitHub Actions run (repo URL rewritten)', () => {
     const d = pt({
       benchmark_type: 'single_turn',
@@ -119,6 +129,15 @@ describe('buildLegendPointsRows', () => {
     });
   });
 
+  it('threads the locale through to each row href', () => {
+    const rows = buildLegendPointsRows(
+      [pt({ benchmark_type: 'agentic_traces', id: 1 })],
+      false,
+      'zh',
+    );
+    expect(rows[0].href).toBe('/zh/inference/agentic/1');
+  });
+
   it('default-sorts by concurrency ascending', () => {
     const rows = buildLegendPointsRows(
       [pt({ conc: 32 }), pt({ conc: 4 }), pt({ conc: 16 })],
diff --git a/packages/app/src/components/inference/utils/legend-points-table.ts b/packages/app/src/components/inference/utils/legend-points-table.ts
index 87df2fcff..8f7a5f697 100644
--- a/packages/app/src/components/inference/utils/legend-points-table.ts
+++ b/packages/app/src/components/inference/utils/legend-points-table.ts
@@ -1,5 +1,7 @@
 import { updateRepoUrl } from '@/lib/utils';
+import { agenticDetailHref } from '@/lib/agentic-detail-link';
 import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
+import type { Locale } from '@/lib/i18n';
 
 import type { InferenceData } from '@/components/inference/types';
 import { fmt, getPointLabel } from '@/components/inference/utils/tooltipUtils';
@@ -55,10 +57,11 @@ const num = (v: number | undefined | null): number | null =>
 export function pointDetailHref(
   d: InferenceData,
   isOverlay: boolean,
+  locale: Locale = 'en',
 ): { href: string | null; isExternal: boolean } {
   if (isOverlay) return { href: null, isExternal: false };
   if (d.benchmark_type === 'agentic_traces' && isPersistedBenchmarkId(d.id)) {
-    return { href: `/inference/agentic/${d.id}`, isExternal: false };
+    return { href: agenticDetailHref(d.id, locale), isExternal: false };
   }
   if (d.run_url) return { href: updateRepoUrl(d.run_url), isExternal: true };
   return { href: null, isExternal: false };
@@ -72,10 +75,11 @@ export function pointDetailHref(
 export function buildLegendPointsRows(
   points: InferenceData[],
   isOverlay: boolean,
+  locale: Locale = 'en',
 ): LegendPointsTableRow[] {
   return points
     .map((d, i) => {
-      const { href, isExternal } = pointDetailHref(d, isOverlay);
+      const { href, isExternal } = pointDetailHref(d, isOverlay, locale);
       return {
         key: `${d.hwKey}|${d.precision}|${d.conc}|${getPointLabel(d)}|${d.offload_mode ?? ''}|${i}`,
         conc: d.conc,
diff --git a/packages/app/src/components/inference/utils/offload-halo.ts b/packages/app/src/components/inference/utils/offload-halo.ts
new file mode 100644
index 000000000..802bd291d
--- /dev/null
+++ b/packages/app/src/components/inference/utils/offload-halo.ts
@@ -0,0 +1,27 @@
+import type * as d3 from 'd3';
+
+import type { InferenceData } from '@/components/inference/types';
+import {
+  OFFLOAD_HALO_DASHARRAY,
+  OFFLOAD_HALO_RADIUS,
+  OFFLOAD_HALO_STROKE_WIDTH,
+} from '@/components/inference/ui/OffloadHaloLegendKey';
+
+export function renderOffloadHalo(
+  group: d3.Selection<SVGGElement, InferenceData, null, undefined>,
+  point: InferenceData,
+  stroke: string,
+): void {
+  group
+    .selectAll<SVGCircleElement, boolean>('.offload-halo')
+    .data(point.offload_mode === 'on' ? [true] : [])
+    .join('circle')
+    .attr('class', 'offload-halo')
+    .attr('r', OFFLOAD_HALO_RADIUS)
+    .attr('fill', 'none')
+    .attr('stroke', stroke)
+    .attr('stroke-width', OFFLOAD_HALO_STROKE_WIDTH)
+    .attr('stroke-dasharray', OFFLOAD_HALO_DASHARRAY)
+    .attr('opacity', 0.9)
+    .attr('pointer-events', 'none');
+}
diff --git a/packages/app/src/components/inference/utils/parallelism-label.test.ts b/packages/app/src/components/inference/utils/parallelism-label.test.ts
index 643e911a6..a05be01c3 100644
--- a/packages/app/src/components/inference/utils/parallelism-label.test.ts
+++ b/packages/app/src/components/inference/utils/parallelism-label.test.ts
@@ -31,6 +31,17 @@ describe('configSegmentLabel', () => {
     expect(configSegmentLabel(8, 1, false, 0)).toBe('TP8');
     expect(configSegmentLabel(8, 1, false, undefined)).toBe('TP8');
   });
+
+  it('appends non-default DCP and PCP as slash-delimited segments', () => {
+    expect(configSegmentLabel(8, 1, false, 1, 8, 1)).toBe('TP8/DCP8');
+    expect(configSegmentLabel(8, 1, false, 1, 8, 4)).toBe('TP8/DCP8/PCP4');
+    expect(configSegmentLabel(8, 8, false, 2, 8, 4)).toBe('TEP8PP2/DCP8/PCP4');
+  });
+
+  it('omits default DCP and PCP widths', () => {
+    expect(configSegmentLabel(8, 1, false, undefined, 1, 1)).toBe('TP8');
+    expect(configSegmentLabel(8, 1, false, undefined, 0, undefined)).toBe('TP8');
+  });
 });
 
 describe('parallelismLabel', () => {
@@ -94,6 +105,27 @@ describe('parallelismLabel', () => {
     ).toBe('2xTP8PP2+1xTP8');
   });
 
+  it('applies per-role DCP and PCP to multinode-disagg segments', () => {
+    expect(
+      parallelismLabel({
+        tp: 8,
+        ep: 1,
+        disagg: true,
+        isMultinode: true,
+        prefillTp: 8,
+        prefillEp: 1,
+        prefillDcp: 2,
+        prefillPcp: 4,
+        prefillNumWorkers: 1,
+        decodeTp: 8,
+        decodeEp: 1,
+        decodeDcp: 8,
+        decodePcp: 1,
+        decodeNumWorkers: 1,
+      }),
+    ).toBe('1xTP8/DCP2/PCP4+1xTP8/DCP8');
+  });
+
   it('drops the decode segment when the decode pool is absent (0 workers, tp 0)', () => {
     // Prefill-only agentic multinode runs (e.g. Kimi-K3 TP8 PP2 on b200-dgxc)
     // emit decode_num_workers=0 / decode_tp=0 — the label is just the active
diff --git a/packages/app/src/components/inference/utils/parallelism-label.ts b/packages/app/src/components/inference/utils/parallelism-label.ts
index c14a90c37..a4e95efd7 100644
--- a/packages/app/src/components/inference/utils/parallelism-label.ts
+++ b/packages/app/src/components/inference/utils/parallelism-label.ts
@@ -1,6 +1,6 @@
 /**
  * Shared parallelism-config labeling — the single source of truth for the
- * short "TP8 / EP8 / TEP8 / DEP8 / DPAEP8 / 2xEP4+1xDPAEP32" labels.
+ * short "TP8 / TP8/DCP8 / EP8 / TEP8 / DEP8 / 2xEP4+1xDPAEP32" labels.
  *
  * Used by the scatter/GPU chart point labels (via getPointLabel) and the
  * agentic detail page's sibling navigator chips, so both surfaces describe a
@@ -13,22 +13,38 @@
  * - tp == ep and dp-attn true: "DEP{N}"
  * - ep > 1 (tp != ep): "EP{ep}" or "DPAEP{ep}"
  * - ep <= 1 (or no EP): "TP{tp}" or "DPATP{tp}"
- * - pp > 1 appends a "PP{pp}" suffix (e.g. "TP8PP2"); pp <= 1 or absent
- *   renders nothing so pre-PP labels stay byte-identical.
+ * - pp > 1 appends a "PP{pp}" suffix (e.g. "TP8PP2").
+ * - dcp/pcp > 1 append slash-delimited context-parallel segments
+ *   (e.g. "TP8/DCP8" or "TP8/DCP8/PCP4").
+ * - Values <= 1 or absent render nothing, preserving legacy labels.
  */
+export const meaningfulParallelismSize = (
+  ...values: (number | null | undefined)[]
+): number | undefined => {
+  const sizes = values.filter(
+    (value): value is number => value !== null && value !== undefined && value > 1,
+  );
+  return sizes.length > 0 ? Math.max(...sizes) : undefined;
+};
+
 export const configSegmentLabel = (
   tp: number,
   ep: number | undefined,
   dpAttention: boolean | undefined,
   pp?: number,
+  dcp?: number,
+  pcp?: number,
 ): string => {
   const ppSuffix = pp !== null && pp !== undefined && pp > 1 ? `PP${pp}` : '';
+  const dcpSuffix = dcp !== null && dcp !== undefined && dcp > 1 ? `/DCP${dcp}` : '';
+  const pcpSuffix = pcp !== null && pcp !== undefined && pcp > 1 ? `/PCP${pcp}` : '';
+  const suffix = `${ppSuffix}${dcpSuffix}${pcpSuffix}`;
   if (ep !== null && ep !== undefined && ep > 1 && tp === ep) {
-    return `${dpAttention ? 'DEP' : 'TEP'}${tp}${ppSuffix}`;
+    return `${dpAttention ? 'DEP' : 'TEP'}${tp}${suffix}`;
   }
   const dpaPrefix = dpAttention ? 'DPA' : '';
-  if (ep === null || ep === undefined || ep <= 1) return `${dpaPrefix}TP${tp}${ppSuffix}`;
-  return `${dpaPrefix}EP${ep}${ppSuffix}`;
+  if (ep === null || ep === undefined || ep <= 1) return `${dpaPrefix}TP${tp}${suffix}`;
+  return `${dpaPrefix}EP${ep}${suffix}`;
 };
 
 /** Parallelism params for one benchmark config, framework-agnostic. */
@@ -37,17 +53,25 @@ export interface ParallelismFields {
   ep?: number;
   /** Pipeline parallelism. Only rendered when > 1. */
   pp?: number;
+  /** Decode-context parallelism. Only rendered when > 1. */
+  dcp?: number;
+  /** Prefill-context parallelism. Only rendered when > 1. */
+  pcp?: number;
   dpAttention?: boolean;
   disagg?: boolean;
   isMultinode?: boolean;
   prefillTp?: number;
   prefillEp?: number;
   prefillPp?: number;
+  prefillDcp?: number;
+  prefillPcp?: number;
   prefillDpAttention?: boolean;
   prefillNumWorkers?: number;
   decodeTp?: number;
   decodeEp?: number;
   decodePp?: number;
+  decodeDcp?: number;
+  decodePcp?: number;
   decodeDpAttention?: boolean;
   decodeNumWorkers?: number;
 }
@@ -76,12 +100,16 @@ export const parallelismLabel = (f: ParallelismFields): string => {
       f.prefillEp ?? f.ep,
       f.prefillDpAttention ?? f.dpAttention,
       f.prefillPp ?? f.pp,
+      f.prefillDcp ?? f.dcp,
+      f.prefillPcp ?? f.pcp,
     );
     const decodeLabel = configSegmentLabel(
       f.decodeTp ?? f.tp,
       f.decodeEp ?? f.ep,
       f.decodeDpAttention ?? f.dpAttention,
       f.decodePp ?? f.pp,
+      f.decodeDcp ?? f.dcp,
+      f.decodePcp ?? f.pcp,
     );
     const pw = f.prefillNumWorkers ?? 1;
     const dw = f.decodeNumWorkers ?? 1;
@@ -92,5 +120,5 @@ export const parallelismLabel = (f: ParallelismFields): string => {
     return `${pw}x${prefillLabel}+${dw}x${decodeLabel}`;
   }
 
-  return configSegmentLabel(f.tp, f.ep, f.dpAttention, f.pp);
+  return configSegmentLabel(f.tp, f.ep, f.dpAttention, f.pp, f.dcp, f.pcp);
 };
diff --git a/packages/app/src/components/inference/utils/point-identity.test.ts b/packages/app/src/components/inference/utils/point-identity.test.ts
index 5f3ee14ae..26d71f322 100644
--- a/packages/app/src/components/inference/utils/point-identity.test.ts
+++ b/packages/app/src/components/inference/utils/point-identity.test.ts
@@ -2,7 +2,7 @@ import { describe, expect, it } from 'vitest';
 
 import type { InferenceData } from '@/components/inference/types';
 
-import { scatterPointConfigId } from './point-identity';
+import { scatterPointConfigId, scatterPointJoinId } from './point-identity';
 
 const point = (overrides: Partial<InferenceData>): InferenceData =>
   ({
@@ -45,4 +45,43 @@ describe('scatterPointConfigId', () => {
 
     expect(off).not.toBe(on);
   });
+
+  it('keeps recipe variants at the same topology and concurrency distinct', () => {
+    const recipeA = scatterPointConfigId(
+      point({
+        recipe_fingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+      }),
+    );
+    const recipeB = scatterPointConfigId(
+      point({
+        recipe_fingerprint: 'bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb',
+      }),
+    );
+
+    expect(recipeA).not.toBe(recipeB);
+    expect(recipeA).toContain(
+      '|recipe-aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+    );
+  });
+
+  it('preserves the legacy key when no recipe fingerprint exists', () => {
+    expect(scatterPointConfigId(point({}))).not.toContain('|recipe-');
+  });
+
+  it('distinguishes otherwise-identical points when multiple dates render together', () => {
+    const may = point({ date: '2026-05-15' });
+    const june = point({ date: '2026-06-15' });
+    expect(scatterPointConfigId(may)).toBe(scatterPointConfigId(june));
+
+    expect(scatterPointJoinId(may, true)).not.toBe(scatterPointJoinId(june, true));
+    expect(scatterPointJoinId(may, true)).toBe(`${scatterPointConfigId(may)}|date-2026-05-15`);
+  });
+
+  it('preserves current-run and undated identities when dates need no disambiguation', () => {
+    const current = point({ date: '2026-06-15' });
+    const undated = point({ date: '' });
+
+    expect(scatterPointJoinId(current, false)).toBe(scatterPointConfigId(current));
+    expect(scatterPointJoinId(undated, true)).toBe(scatterPointConfigId(undated));
+  });
 });
diff --git a/packages/app/src/components/inference/utils/point-identity.ts b/packages/app/src/components/inference/utils/point-identity.ts
index f578c8e69..d6bc2c2e0 100644
--- a/packages/app/src/components/inference/utils/point-identity.ts
+++ b/packages/app/src/components/inference/utils/point-identity.ts
@@ -17,8 +17,21 @@ export function scatterPointConfigId(point: InferenceData): string {
     key += `|disagg|${point.num_prefill_gpu ?? 0}|${point.num_decode_gpu ?? 0}`;
   }
   if (point.offload_mode) key += `|offload-${point.offload_mode}`;
+  if (point.recipe_fingerprint) key += `|recipe-${point.recipe_fingerprint}`;
   // Agentic series omit spec decoding from hwKey so one curve can mix methods.
   // It remains point identity to avoid collapsing overlapping MTP/STP results.
   key += agenticSpecDecodingKeySuffix(point);
   return key;
 }
+
+/**
+ * Stable D3 join key for an official scatter point.
+ *
+ * Date only participates when a chart is simultaneously rendering multiple
+ * date series. This preserves the long-lived identity of current-run points
+ * while preventing otherwise-identical comparison points from sharing a key.
+ */
+export function scatterPointJoinId(point: InferenceData, distinguishDates: boolean): string {
+  const configId = scatterPointConfigId(point);
+  return distinguishDates && point.date ? `${configId}|date-${point.date}` : configId;
+}
diff --git a/packages/app/src/components/inference/utils/quickFilters.test.ts b/packages/app/src/components/inference/utils/quickFilters.test.ts
index a33347267..106fd2fb4 100644
--- a/packages/app/src/components/inference/utils/quickFilters.test.ts
+++ b/packages/app/src/components/inference/utils/quickFilters.test.ts
@@ -223,4 +223,26 @@ describe('applyQuickFilters', () => {
       [overlay],
     );
   });
+
+  it('filters agentic unofficial overlays by the options exposed in the UI', () => {
+    const official = point({
+      id: 42,
+      benchmark_type: 'agentic_traces',
+      hwKey: 'b200_vllm_mtp',
+      framework: 'vllm',
+      spec_decoding: 'mtp',
+    });
+    const overlay = point({
+      id: undefined,
+      benchmark_type: 'agentic_traces',
+      run_url: 'https://github.com/example/actions/runs/2',
+      hwKey: 'mi300x_sglang_mtp',
+      framework: 'sglang',
+      spec_decoding: 'mtp',
+    });
+
+    expect(
+      applyQuickFilters([official, overlay], filters({ vendors: ['AMD'], frameworks: ['sglang'] })),
+    ).toEqual([overlay]);
+  });
 });
diff --git a/packages/app/src/components/inference/utils/resolveXAxisField.ts b/packages/app/src/components/inference/utils/resolveXAxisField.ts
index 661bc5589..901d052d9 100644
--- a/packages/app/src/components/inference/utils/resolveXAxisField.ts
+++ b/packages/app/src/components/inference/utils/resolveXAxisField.ts
@@ -14,7 +14,7 @@
  */
 import { withPercentile } from '@/lib/benchmark-transform';
 
-import type { ChartDefinition } from '../types';
+import type { AggDataEntry, ChartDefinition } from '../types';
 
 /** Which rung of the branch ladder chose the x field (drives label choice). */
 export type XAxisBranch =
@@ -25,10 +25,9 @@ export type XAxisBranch =
 
 export interface ResolvedXAxis {
   /** The data field the x-axis plots (percentile-adjusted for agentic). */
-  xAxisField: string;
-  /** The chart's natural latency metric, percentile-adjusted — the "no
-   * override" baseline the flip check compares against. */
-  naturalX: string;
+  xAxisField: keyof AggDataEntry;
+  /** The chart's natural latency metric, percentile-adjusted. */
+  naturalX: keyof AggDataEntry;
   isInputMetric: boolean;
   isTtftOverride: boolean;
   branch: XAxisBranch;
@@ -60,7 +59,10 @@ export function resolveXAxisField(
   opts: { isAgentic: boolean; percentile: string },
 ): ResolvedXAxis {
   const { isAgentic, percentile } = opts;
-  const naturalX = withPercentile(chartDef.x, isAgentic ? percentile : 'median');
+  const naturalX = withPercentile(
+    chartDef.x,
+    isAgentic ? percentile : 'median',
+  ) as keyof AggDataEntry;
 
   const metricTitle =
     (chartDef[`${selectedYAxisMetric}_title` as keyof ChartDefinition] as string) || '';
@@ -69,22 +71,22 @@ export function resolveXAxisField(
   // percentile (median/p75/p90/p99) depending on sequence kind.
   const isTtftOverride = typeof effectiveXMetric === 'string' && effectiveXMetric.endsWith('_ttft');
 
-  let xAxisField: string = naturalX;
+  let xAxisField: keyof AggDataEntry = naturalX;
   let branch: XAxisBranch = 'natural';
   if (effectiveXMetric && chartDef.chartType === 'interactivity' && isInputMetric && !isAgentic) {
-    xAxisField = effectiveXMetric;
+    xAxisField = effectiveXMetric as keyof AggDataEntry;
     branch = 'user-input-override';
   } else if (chartDef.chartType === 'interactivity' && isInputMetric) {
     const xOverrideKey = `${selectedYAxisMetric}_x` as keyof ChartDefinition;
-    xAxisField = (chartDef[xOverrideKey] as string) || chartDef.x;
+    xAxisField = ((chartDef[xOverrideKey] as string) || chartDef.x) as keyof AggDataEntry;
     branch = 'config-input-override';
   } else if (chartDef.chartType === 'e2e' && isTtftOverride) {
-    xAxisField = effectiveXMetric!;
+    xAxisField = effectiveXMetric as keyof AggDataEntry;
     branch = 'e2e-ttft-override';
   }
 
   if (isAgentic) {
-    xAxisField = withPercentile(xAxisField, percentile);
+    xAxisField = withPercentile(xAxisField, percentile) as keyof AggDataEntry;
   }
 
   return { xAxisField, naturalX, isInputMetric, isTtftOverride, branch };
diff --git a/packages/app/src/components/inference/utils/tooltip-parallelism.test.ts b/packages/app/src/components/inference/utils/tooltip-parallelism.test.ts
index 1d0ae0b5a..be195c82a 100644
--- a/packages/app/src/components/inference/utils/tooltip-parallelism.test.ts
+++ b/packages/app/src/components/inference/utils/tooltip-parallelism.test.ts
@@ -142,6 +142,41 @@ describe('tooltip parallelism — standard (ep field present)', () => {
     }
   });
 
+  it('shows aggregate DCP and PCP widths when they are greater than one', () => {
+    const html = generateTooltipContent(
+      tooltipConfig({
+        data: pt({
+          tp: 8,
+          ep: 1,
+          prefill_dcp_size: 1,
+          decode_dcp_size: 8,
+          prefill_pcp_size: 4,
+          decode_pcp_size: 1,
+        }),
+      }),
+    );
+
+    expect(html).toContain('<strong>Decode Context Parallelism (DCP):</strong> 8');
+    expect(html).toContain('<strong>Prefill Context Parallelism (PCP):</strong> 4');
+  });
+
+  it('hides default DCP and PCP widths', () => {
+    const html = generateTooltipContent(
+      tooltipConfig({
+        data: pt({
+          tp: 8,
+          ep: 1,
+          prefill_dcp_size: 1,
+          decode_dcp_size: 1,
+          prefill_pcp_size: 1,
+          decode_pcp_size: 1,
+        }),
+      }),
+    );
+
+    expect(html).not.toContain('Context Parallelism');
+  });
+
   it('omits Expert Parallelism line when ep is null', () => {
     const d = pt({ tp: 4 });
     (d as any).ep = null;
@@ -236,6 +271,33 @@ describe('tooltip parallelism — multinode disagg', () => {
     expect(html).not.toContain('PP: 1');
   });
 
+  it('shows per-role DCP and PCP in disaggregated tooltips', () => {
+    const html = generateTooltipContent(
+      tooltipConfig({
+        data: pt({
+          tp: 16,
+          ep: 1,
+          is_multinode: true,
+          disagg: true,
+          prefill_tp: 8,
+          prefill_ep: 1,
+          prefill_dcp_size: 2,
+          prefill_pcp_size: 4,
+          decode_tp: 8,
+          decode_ep: 1,
+          decode_dcp_size: 8,
+          decode_pcp_size: 1,
+          num_prefill_gpu: 8,
+          num_decode_gpu: 8,
+        }),
+      }),
+    );
+
+    expect(html).toMatch(/<strong>Prefill:<\/strong>[^<]*DCP: 2, PCP: 4,/u);
+    expect(html).toMatch(/<strong>Decode:<\/strong>[^<]*DCP: 8,/u);
+    expect(html).not.toContain('PCP: 1');
+  });
+
   it('shows "?" for missing GPU count fields', () => {
     const html = generateTooltipContent(
       tooltipConfig({
@@ -291,10 +353,22 @@ describe('tooltip parallelism — zh locale', () => {
     expect(html).not.toContain('Tensor Parallelism');
   });
 
+  it('localizes DCP and PCP labels', () => {
+    const html = generateTooltipContent(
+      tooltipConfig({
+        data: pt({ tp: 8, ep: 1, decode_dcp_size: 8, prefill_pcp_size: 4 }),
+        locale: 'zh',
+      }),
+    );
+
+    expect(html).toContain('解码上下文并行 (DCP)');
+    expect(html).toContain('预填充上下文并行 (PCP)');
+  });
+
   it('localizes the old-data parallelism strategy line', () => {
     const html = generateTooltipContent(tooltipConfig({ data: pt({ tp: 4 }), locale: 'zh' }));
     expect(html).toContain('并行策略');
-    expect(html).toContain('4 个 Chip');
+    expect(html).toContain('4 个芯片');
   });
 
   it('localizes the multinode disagg role headers', () => {
@@ -311,9 +385,9 @@ describe('tooltip parallelism — zh locale', () => {
         locale: 'zh',
       }),
     );
-    expect(html).toContain('预填充:');
-    expect(html).toContain('解码:');
-    expect(html).toContain('16 个 Chip');
+    expect(html).toContain('预填充：');
+    expect(html).toContain('解码：');
+    expect(html).toContain('16 个芯片');
   });
 });
 
diff --git a/packages/app/src/components/inference/utils/tooltip-utils.test.ts b/packages/app/src/components/inference/utils/tooltip-utils.test.ts
index 7edc1268f..cd469a0c6 100644
--- a/packages/app/src/components/inference/utils/tooltip-utils.test.ts
+++ b/packages/app/src/components/inference/utils/tooltip-utils.test.ts
@@ -99,6 +99,25 @@ describe('getPointLabel', () => {
     expect(getPointLabel(pt({ tp: 4, ep: 1 }))).toBe('TP4');
   });
 
+  it('includes DCP and PCP in the point label when non-default', () => {
+    expect(
+      getPointLabel(
+        pt({
+          tp: 8,
+          decode_tp: 8,
+          ep: 1,
+          prefill_dcp_size: 8,
+          decode_dcp_size: 8,
+          prefill_pcp_size: 1,
+          decode_pcp_size: 1,
+        }),
+      ),
+    ).toBe('TP8/DCP8');
+    expect(
+      getPointLabel(pt({ tp: 8, decode_tp: 8, ep: 1, decode_dcp_size: 8, prefill_pcp_size: 4 })),
+    ).toBe('TP8/DCP8/PCP4');
+  });
+
   it('returns "DPATP4" when ep is 1 and dp_attention is true', () => {
     expect(getPointLabel(pt({ tp: 4, ep: 1, dp_attention: true }))).toBe('DPATP4');
   });
@@ -152,21 +171,89 @@ describe('getPointLabel', () => {
 describe('generateTooltipContent', () => {
   it('renders View charts as a same-tab anchor so browsers offer open-in-new-tab', () => {
     const html = generateTooltipContent(
-      tooltipConfig({ data: pt({ id: 1 }), isPinned: true, hasTrace: true }),
+      tooltipConfig({
+        data: pt({ id: 1, benchmark_type: 'agentic_traces' }),
+        isPinned: true,
+        hasTrace: true,
+      }),
     );
     expect(html).toContain('<a data-action="view-charts"');
     expect(html).toContain('href="/inference/agentic/1"');
     expect(html).not.toContain('data-action="view-charts" target=');
   });
 
+  it('renders View logs only for pinned points with a stored server log', () => {
+    const html = generateTooltipContent(
+      tooltipConfig({
+        data: pt({ id: 7, benchmark_type: 'agentic_traces' }),
+        isPinned: true,
+        hasLog: true,
+      }),
+    );
+    expect(html).toContain('<a data-action="view-logs"');
+    expect(html).toContain('href="/inference/agentic/7?view=logs"');
+    expect(
+      generateTooltipContent(
+        tooltipConfig({
+          data: pt({ id: 7, benchmark_type: 'agentic_traces' }),
+          isPinned: false,
+          hasLog: true,
+        }),
+      ),
+    ).not.toContain('data-action="view-logs"');
+    expect(
+      generateTooltipContent(
+        tooltipConfig({
+          data: pt({ id: 7, benchmark_type: 'agentic_traces' }),
+          isPinned: true,
+          hasLog: false,
+        }),
+      ),
+    ).not.toContain('data-action="view-logs"');
+  });
+
+  it('routes fixed-sequence log actions to the fixed benchmark log viewer', () => {
+    const html = generateTooltipContent(
+      tooltipConfig({
+        data: pt({ id: 96255, benchmark_type: 'single_turn' }),
+        isPinned: true,
+        hasLog: true,
+      }),
+    );
+    expect(html).toContain('<a data-action="view-logs"');
+    expect(html).toContain('href="/inference/logs/96255"');
+    expect(html).not.toContain('/inference/agentic/96255');
+  });
+
+  it('localizes point-detail actions and their /zh routes', () => {
+    const html = generateTooltipContent(
+      tooltipConfig({
+        data: pt({ id: 7, benchmark_type: 'agentic_traces' }),
+        isPinned: true,
+        hasTrace: true,
+        hasLog: true,
+        locale: 'zh',
+      }),
+    );
+    expect(html).toContain('查看图表');
+    expect(html).toContain('href="/zh/inference/agentic/7"');
+    expect(html).toContain('查看日志');
+    expect(html).toContain('href="/zh/inference/agentic/7?view=logs"');
+  });
+
   it('omits View charts when the point id is non-persisted (0 / NaN), even if pinned + hasTrace', () => {
     // Overlay agentic points arrive with id 0 / NaN — the button would otherwise
     // link to /inference/agentic/0, a doomed lookup.
     for (const badId of [0, Number.NaN]) {
       const html = generateTooltipContent(
-        tooltipConfig({ data: pt({ id: badId }), isPinned: true, hasTrace: true }),
+        tooltipConfig({
+          data: pt({ id: badId, benchmark_type: 'agentic_traces' }),
+          isPinned: true,
+          hasTrace: true,
+        }),
       );
       expect(html).not.toContain('data-action="view-charts"');
+      expect(html).not.toContain('data-action="view-logs"');
     }
   });
 
@@ -284,7 +371,7 @@ describe('generateTooltipContent', () => {
     );
 
     expect(enabled).toContain('<strong>Offload Type:</strong> Enabled (legacy data)');
-    expect(disabledZh).toContain('<strong>卸载类型:</strong> 已禁用（旧版数据）');
+    expect(disabledZh).toContain('<strong>offload 类型：</strong> 已禁用（旧版数据）');
   });
 
   it('does not treat the fixed-sequence offload default as legacy metadata', () => {
@@ -371,9 +458,9 @@ describe('generateTooltipContent', () => {
         }),
       }),
     );
-    expect(html).toContain('<strong>卸载类型:</strong> DRAM');
-    expect(html).toContain('<strong>KV 卸载引擎:</strong> LMCache');
-    expect(html).toContain('<strong>路由器:</strong> vLLM Router 0.1.14');
+    expect(html).toContain('<strong>offload 类型：</strong> DRAM');
+    expect(html).toContain('<strong>KV offload 引擎：</strong> LMCache');
+    expect(html).toContain('<strong>路由器：</strong> vLLM Router 0.1.14');
   });
 
   it('omits the offload type row when the canonical tier is none', () => {
@@ -383,7 +470,7 @@ describe('generateTooltipContent', () => {
     );
 
     expect(en).not.toContain('Offload Type');
-    expect(zh).not.toContain('卸载类型');
+    expect(zh).not.toContain('offload 类型');
   });
 
   it('falls back to hwKey when hardware config entry is missing', () => {
@@ -473,6 +560,17 @@ describe('generateOverlayTooltipContent', () => {
     expect(html).toContain('<strong>CPU Cache Hit Rate:</strong> 42.0%');
   });
 
+  it('shows DCP and PCP for unofficial-run points', () => {
+    const html = generateOverlayTooltipContent(
+      overlayConfig({
+        data: pt({ ep: 1, decode_dcp_size: 8, prefill_pcp_size: 4 }),
+      }),
+    );
+
+    expect(html).toContain('<strong>Decode Context Parallelism (DCP):</strong> 8');
+    expect(html).toContain('<strong>Prefill Context Parallelism (PCP):</strong> 4');
+  });
+
   it('shows point-level speculative decoding for mixed agentic overlays', () => {
     const mtp = generateOverlayTooltipContent(
       overlayConfig({
@@ -487,7 +585,22 @@ describe('generateOverlayTooltipContent', () => {
     );
 
     expect(mtp).toContain('<strong>Speculative Decoding:</strong> MTP');
-    expect(standardZh).toContain('<strong>投机解码:</strong> 关闭');
+    expect(standardZh).toContain('<strong>投机解码：</strong> 关闭');
+  });
+
+  it('labels Kimi-K3 speculative decoding "DSpark" rather than the generic MTP', () => {
+    const html = generateOverlayTooltipContent(
+      overlayConfig({
+        data: pt({
+          benchmark_type: 'agentic_traces',
+          model: 'Kimi-K3',
+          spec_decoding: 'mtp',
+        }),
+      }),
+    );
+
+    expect(html).toContain('<strong>Speculative Decoding:</strong> DSpark');
+    expect(html).not.toContain('<strong>Speculative Decoding:</strong> MTP');
   });
 
   it('hides stale CPU cache hits for unofficial overlays without offload', () => {
@@ -556,6 +669,17 @@ describe('generateGPUGraphTooltipContent', () => {
     expect(html).toContain('BF16');
   });
 
+  it('shows DCP and PCP in comparison point tooltips', () => {
+    const html = generateGPUGraphTooltipContent(
+      tooltipConfig({
+        data: pt({ ep: 1, decode_dcp_size: 8, prefill_pcp_size: 4 }),
+      }),
+    );
+
+    expect(html).toContain('<strong>Decode Context Parallelism (DCP):</strong> 8');
+    expect(html).toContain('<strong>Prefill Context Parallelism (PCP):</strong> 4');
+  });
+
   it('splits image and SHA onto separate lines', () => {
     const html = generateGPUGraphTooltipContent(
       tooltipConfig({ data: pt({ image: 'vllm-v0.6.0 abc123' }) }),
@@ -566,22 +690,38 @@ describe('generateGPUGraphTooltipContent', () => {
   it('shows View charts only for pinned points with stored trace data', () => {
     expect(
       generateGPUGraphTooltipContent(
-        tooltipConfig({ data: pt({ id: 1 }), isPinned: true, hasTrace: true }),
+        tooltipConfig({
+          data: pt({ id: 1, benchmark_type: 'agentic_traces' }),
+          isPinned: true,
+          hasTrace: true,
+        }),
       ),
     ).toContain('data-action="view-charts"');
     expect(
       generateGPUGraphTooltipContent(
-        tooltipConfig({ data: pt({ id: 1 }), isPinned: true, hasTrace: true }),
+        tooltipConfig({
+          data: pt({ id: 1, benchmark_type: 'agentic_traces' }),
+          isPinned: true,
+          hasTrace: true,
+        }),
       ),
     ).toContain('href="/inference/agentic/1"');
     expect(
       generateGPUGraphTooltipContent(
-        tooltipConfig({ data: pt({ id: 1 }), isPinned: false, hasTrace: true }),
+        tooltipConfig({
+          data: pt({ id: 1, benchmark_type: 'agentic_traces' }),
+          isPinned: false,
+          hasTrace: true,
+        }),
       ),
     ).not.toContain('data-action="view-charts"');
     expect(
       generateGPUGraphTooltipContent(
-        tooltipConfig({ data: pt({ id: 1 }), isPinned: true, hasTrace: false }),
+        tooltipConfig({
+          data: pt({ id: 1, benchmark_type: 'agentic_traces' }),
+          isPinned: true,
+          hasTrace: false,
+        }),
       ),
     ).not.toContain('data-action="view-charts"');
   });
diff --git a/packages/app/src/components/inference/utils/tooltipUtils.ts b/packages/app/src/components/inference/utils/tooltipUtils.ts
index eff4f0e29..ee740c4f5 100644
--- a/packages/app/src/components/inference/utils/tooltipUtils.ts
+++ b/packages/app/src/components/inference/utils/tooltipUtils.ts
@@ -1,11 +1,15 @@
 import { formatNumber, getDisplayLabel } from '@/lib/utils';
 import { specMethodDisplayLabel } from '@/lib/compare-variant-slug';
+import { agenticDetailHref } from '@/lib/agentic-detail-link';
 import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
 import type { Locale } from '@/lib/i18n';
 import { isKvOffloadEnabled } from '@/lib/kv-offload';
 
 import type { HardwareConfig, InferenceData, OverlayData } from '@/components/inference/types';
-import { parallelismLabel } from '@/components/inference/utils/parallelism-label';
+import {
+  meaningfulParallelismSize,
+  parallelismLabel,
+} from '@/components/inference/utils/parallelism-label';
 import {
   cacheImplementationLabel,
   offloadTypeLabel,
@@ -35,6 +39,8 @@ export interface TooltipConfig {
    * call so we don't ship megabytes of profile JSONL just for this check).
    */
   hasTrace?: boolean;
+  /** Whether this official DB-backed point has a linked `server_logs` row. */
+  hasLog?: boolean;
   /** Page locale for tooltip metadata labels. Defaults to English. */
   locale?: Locale;
 }
@@ -59,28 +65,37 @@ const asBool = (v: boolean | string | undefined): boolean | undefined =>
  * Delegates to the shared {@link parallelismLabel} so the chart points and the
  * agentic sibling navigator describe a config identically.
  */
-export const getPointLabel = (d: InferenceData): string =>
-  parallelismLabel({
+export const getPointLabel = (d: InferenceData): string => {
+  const aggregateDcp = meaningfulParallelismSize(d.prefill_dcp_size, d.decode_dcp_size);
+  const aggregatePcp = meaningfulParallelismSize(d.prefill_pcp_size, d.decode_pcp_size);
+  return parallelismLabel({
     // InferenceData.tp is the TOTAL GPU count (createChartDataPoint folds pp
     // into it for aggregated rows) — the label wants the actual TP width, so
     // prefer the raw decode_tp and keep d.tp only as a legacy fallback.
     tp: d.decode_tp ?? d.tp,
     ep: d.ep,
     pp: d.pp,
+    dcp: d.disagg ? (d.decode_dcp_size ?? d.prefill_dcp_size) : aggregateDcp,
+    pcp: d.disagg ? (d.decode_pcp_size ?? d.prefill_pcp_size) : aggregatePcp,
     dpAttention: asBool(d.dp_attention),
     disagg: d.disagg,
     isMultinode: d.is_multinode,
     prefillTp: d.prefill_tp,
     prefillEp: d.prefill_ep,
     prefillPp: d.prefill_pp,
+    prefillDcp: d.prefill_dcp_size,
+    prefillPcp: d.prefill_pcp_size,
     prefillDpAttention: asBool(d.prefill_dp_attention),
     prefillNumWorkers: d.prefill_num_workers,
     decodeTp: d.decode_tp,
     decodeEp: d.decode_ep,
     decodePp: d.decode_pp,
+    decodeDcp: d.decode_dcp_size,
+    decodePcp: d.decode_pcp_size,
     decodeDpAttention: asBool(d.decode_dp_attention),
     decodeNumWorkers: d.decode_num_workers,
   });
+};
 
 const runLinkHTML = (runUrl?: string) =>
   runUrl
@@ -89,8 +104,10 @@ const runLinkHTML = (runUrl?: string) =>
       </div>`
     : '';
 
+const labelColon = (label: string) => (/[\u4E00-\u9FFF]/u.test(label) ? '：' : ':');
+
 const tooltipLine = (label: string, value: string | number) =>
-  `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${label}:</strong> ${value}</div>`;
+  `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${label}${labelColon(label)}</strong> ${value}</div>`;
 
 const formatPct = (v: number | undefined): string | null =>
   v === undefined || v === null || Number.isNaN(v) ? null : `${(v * 100).toFixed(1)}%`;
@@ -104,6 +121,37 @@ export const fmt = (v: number): string => {
   return String(rounded);
 };
 
+const TOOLTIP_STRINGS = {
+  en: {
+    dismiss: 'Click elsewhere to dismiss',
+    unofficialRun: '✕ UNOFFICIAL RUN',
+    date: 'Date',
+    dataFrom: (d: string) => `(data from ${d})`,
+    image: 'Image',
+    branch: 'Branch',
+    chipConfig: 'Chip Config',
+    totalChips: 'Total Chips',
+    concurrency: 'Concurrency',
+    precision: 'Precision',
+    inputTputPerChip: 'Input Token Throughput per Chip',
+    outputTputPerChip: 'Output Token Throughput per Chip',
+  },
+  zh: {
+    dismiss: '点击其他区域关闭',
+    unofficialRun: '✕ 非官方运行',
+    date: '日期',
+    dataFrom: (d: string) => `（数据来自 ${d}）`,
+    image: '镜像',
+    branch: '分支',
+    chipConfig: '芯片配置',
+    totalChips: '芯片总数',
+    concurrency: '并发数',
+    precision: '精度',
+    inputTputPerChip: '每芯片输入 token 吞吐量',
+    outputTputPerChip: '每芯片输出 token 吞吐量',
+  },
+} as const;
+
 const CACHE_STRINGS = {
   en: {
     offloadType: 'Offload Type',
@@ -117,11 +165,11 @@ const CACHE_STRINGS = {
     legacyDisabled: 'Disabled (legacy data)',
   },
   zh: {
-    offloadType: '卸载类型',
-    offloadBackend: 'KV 卸载引擎',
+    offloadType: 'offload 类型',
+    offloadBackend: 'KV offload 引擎',
     transferEngine: 'KV 传输引擎',
     router: '路由器',
-    gpuHitRate: 'Chip Cache 命中率',
+    gpuHitRate: '芯片 Cache 命中率',
     cpuHitRate: 'CPU Cache 命中率',
     theoreticalHitRate: '理论 Cache 命中率',
     legacyEnabled: '已启用（旧版数据）',
@@ -218,27 +266,49 @@ const generateAgenticHTML = (d: InferenceData, locale: Locale): string => {
   return parts.join('');
 };
 
-/** "View charts" link — only visible when the tooltip is pinned and the
- *  point has stored trace data. Wired up by the scatter/GPU graph click handlers. */
-const viewChartsButtonHTML = (
+const ACTION_STRINGS = {
+  en: { charts: 'View charts', logs: 'View logs' },
+  zh: { charts: '查看图表', logs: '查看日志' },
+} as const;
+
+const pointDetailActionLink = (action: 'view-charts' | 'view-logs', href: string, label: string) =>
+  `<a data-action="${action}" href="${href}" style="
+    display: block; width: 100%; padding: 4px 8px; font-size: 11px; font-weight: 500;
+    border: 1px solid var(--border); border-radius: 6px; cursor: pointer;
+    background: var(--accent); color: var(--accent-foreground); text-align: center; text-decoration: none;
+  ">${label} &rarr;</a>`;
+
+/** Point-detail links rendered only for persisted, pinned official points. */
+const viewActionsHTML = (
   isPinned: boolean,
   hasTraceData: boolean,
+  hasLogData: boolean,
   pointId: number | undefined,
+  benchmarkType: string | undefined,
+  locale: Locale,
 ): string => {
-  if (!isPinned || !hasTraceData || !isPersistedBenchmarkId(pointId)) return '';
-  return `<a data-action="view-charts" href="/inference/agentic/${pointId}" style="
-    display: block; margin-top: 8px; width: 100%; padding: 4px 8px; font-size: 11px; font-weight: 500;
-    border: 1px solid var(--border); border-radius: 6px; cursor: pointer;
-    background: var(--accent); color: var(--accent-foreground); text-align: center; text-decoration: none;
-  ">View charts &rarr;</a>`;
+  const isAgentic = benchmarkType === 'agentic_traces';
+  const showCharts = isAgentic && hasTraceData;
+  if (!isPinned || !isPersistedBenchmarkId(pointId) || (!showCharts && !hasLogData)) return '';
+  const prefix = locale === 'zh' ? '/zh' : '';
+  const agenticHref = agenticDetailHref(pointId, locale);
+  const logHref = isAgentic
+    ? `${agenticHref}${agenticHref.includes('?') ? '&' : '?'}view=logs`
+    : `${prefix}/inference/logs/${pointId}`;
+  const t = ACTION_STRINGS[locale];
+  const actions = [
+    showCharts ? pointDetailActionLink('view-charts', agenticHref, t.charts) : '',
+    hasLogData ? pointDetailActionLink('view-logs', logHref, t.logs) : '',
+  ].filter(Boolean);
+  return `<div style="display: grid; gap: 6px; margin-top: 8px;">${actions.join('')}</div>`;
 };
 
 const shortenSha = (image: string) =>
   image.replaceAll(/(?<shaPrefix>sha256:[a-f0-9]{7})[a-f0-9]+/giu, '$<shaPrefix>…');
 
-const imageTooltipLine = (image: string) =>
+const imageTooltipLine = (image: string, label: string) =>
   `<div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Image:</strong> <span style="display: inline-block; vertical-align: top; overflow-wrap: anywhere;">${shortenSha(image.trim()).replace(/\s+/u, '<br />')}</span>
+        <strong>${label}${labelColon(label)}</strong> <span style="display: inline-block; vertical-align: top; overflow-wrap: anywhere;">${shortenSha(image.trim()).replace(/\s+/u, '<br />')}</span>
       </div>`;
 
 const PARALLELISM_STRINGS = {
@@ -255,6 +325,8 @@ const PARALLELISM_STRINGS = {
     tensorParallelism: 'Tensor Parallelism',
     expertParallelism: 'Expert Parallelism',
     pipelineParallelism: 'Pipeline Parallelism',
+    decodeContextParallelism: 'Decode Context Parallelism (DCP)',
+    prefillContextParallelism: 'Prefill Context Parallelism (PCP)',
     dpAttention: 'DP Attention',
   },
   zh: {
@@ -263,17 +335,26 @@ const PARALLELISM_STRINGS = {
     singleNode: '单节点聚合',
     multiNode: '多节点聚合',
     disaggregated: '分离式',
-    gpuCount: (n: number) => `${n} 个 Chip`,
+    gpuCount: (n: number) => `${n} 个芯片`,
     prefill: '预填充',
     decode: '解码',
-    gpusUnit: '个 Chip',
+    gpusUnit: '个芯片',
     tensorParallelism: '张量并行 (TP)',
     expertParallelism: '专家并行 (EP)',
     pipelineParallelism: '流水线并行 (PP)',
+    decodeContextParallelism: '解码上下文并行 (DCP)',
+    prefillContextParallelism: '预填充上下文并行 (PCP)',
     dpAttention: 'DP Attention',
   },
 } as const;
 
+const contextParallelismParts = (dcp?: number, pcp?: number): string => {
+  const parts: string[] = [];
+  if (dcp !== undefined && dcp > 1) parts.push(`DCP: ${dcp}`);
+  if (pcp !== undefined && pcp > 1) parts.push(`PCP: ${pcp}`);
+  return parts.length > 0 ? `${parts.join(', ')}, ` : '';
+};
+
 /**
  * Generates HTML for the parallelism configuration section of a tooltip.
  * Falls back to GPU count for old data without parallelism fields.
@@ -283,11 +364,18 @@ const PARALLELISM_STRINGS = {
 const generateParallelismHTML = (d: InferenceData, locale: Locale = 'en'): string => {
   const t = PARALLELISM_STRINGS[locale];
   const deployment = d.disagg ? t.disaggregated : d.is_multinode ? t.multiNode : t.singleNode;
+  const aggregateDcp = meaningfulParallelismSize(d.prefill_dcp_size, d.decode_dcp_size);
+  const aggregatePcp = meaningfulParallelismSize(d.prefill_pcp_size, d.decode_pcp_size);
   if (
     (d.ep === null || d.ep === undefined) &&
     (d.prefill_ep === null || d.prefill_ep === undefined)
   ) {
-    return tooltipLine(t.deployment, deployment) + tooltipLine(t.strategy, t.gpuCount(d.tp));
+    return (
+      tooltipLine(t.deployment, deployment) +
+      tooltipLine(t.strategy, t.gpuCount(d.tp)) +
+      (aggregateDcp ? tooltipLine(t.decodeContextParallelism, aggregateDcp) : '') +
+      (aggregatePcp ? tooltipLine(t.prefillContextParallelism, aggregatePcp) : '')
+    );
   }
 
   if (d.is_multinode && d.disagg) {
@@ -301,13 +389,15 @@ const generateParallelismHTML = (d: InferenceData, locale: Locale = 'en'): strin
     const ddpa = d.decode_dp_attention ?? d.dp_attention ?? false;
     const pw = d.prefill_num_workers ?? 1;
     const dw = d.decode_num_workers ?? 1;
+    const prefillContext = contextParallelismParts(d.prefill_dcp_size, d.prefill_pcp_size);
+    const decodeContext = contextParallelismParts(d.decode_dcp_size, d.decode_pcp_size);
     return `
       ${tooltipLine(t.deployment, deployment)}
       <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${t.prefill}:</strong> ${d.num_prefill_gpu ?? '?'} ${t.gpusUnit}, TP: ${ptp}, ${ppp > 1 ? `PP: ${ppp}, ` : ''}EP: ${pep}, DPA: ${pdpa ? 'True' : 'False'}, Workers: ${pw}
+        <strong>${t.prefill}${labelColon(t.prefill)}</strong> ${d.num_prefill_gpu ?? '?'} ${t.gpusUnit}, TP: ${ptp}, ${ppp > 1 ? `PP: ${ppp}, ` : ''}${prefillContext}EP: ${pep}, DPA: ${pdpa ? 'True' : 'False'}, Workers: ${pw}
       </div>
       <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${t.decode}:</strong> ${d.num_decode_gpu ?? '?'} ${t.gpusUnit}, TP: ${dtp}, ${dpp > 1 ? `PP: ${dpp}, ` : ''}EP: ${dep}, DPA: ${ddpa ? 'True' : 'False'}, Workers: ${dw}
+        <strong>${t.decode}${labelColon(t.decode)}</strong> ${d.num_decode_gpu ?? '?'} ${t.gpusUnit}, TP: ${dtp}, ${dpp > 1 ? `PP: ${dpp}, ` : ''}${decodeContext}EP: ${dep}, DPA: ${ddpa ? 'True' : 'False'}, Workers: ${dw}
       </div>`;
   }
 
@@ -315,6 +405,8 @@ const generateParallelismHTML = (d: InferenceData, locale: Locale = 'en'): strin
     ${tooltipLine(t.deployment, deployment)}
     ${tooltipLine(t.tensorParallelism, d.decode_tp ?? d.tp)}
     ${d.pp !== null && d.pp !== undefined && d.pp > 1 ? tooltipLine(t.pipelineParallelism, d.pp) : ''}
+    ${aggregateDcp ? tooltipLine(t.decodeContextParallelism, aggregateDcp) : ''}
+    ${aggregatePcp ? tooltipLine(t.prefillContextParallelism, aggregatePcp) : ''}
     ${d.ep !== null && d.ep !== undefined ? tooltipLine(t.expertParallelism, d.ep) : ''}
     ${tooltipLine(t.dpAttention, d.dp_attention ? 'True' : 'False')}`;
 };
@@ -337,56 +429,43 @@ export const generateTooltipContent = (config: TooltipConfig): string => {
     hasTrace,
   } = config;
   const locale = config.locale ?? 'en';
+  const t = TOOLTIP_STRINGS[locale];
 
   return `
     <div style="background: var(--popover); border: 1px solid var(--border); border-radius: 8px; padding: 12px; box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1); user-select: ${isPinned ? 'text' : 'none'};">
-      ${isPinned ? '<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">Click elsewhere to dismiss</div>' : ''}
+      ${isPinned ? `<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">${t.dismiss}</div>` : ''}
       <div style="color: var(--foreground); font-size: 12px; font-weight: 600; margin-bottom: 8px;">
         ${hardwareConfig[d.hwKey] ? getDisplayLabel(hardwareConfig[d.hwKey]) : d.hwKey}
       </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Date:</strong> ${d.actualDate ?? d.date}
-      </div>
+      ${tooltipLine(t.date, `${d.actualDate ?? d.date}`)}
       ${
         d?.image
           ? `
-      ${imageTooltipLine(d.image)}`
+      ${imageTooltipLine(d.image, t.image)}`
           : ''
       }
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${xLabel}:</strong> ${fmt(d.x)}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${yLabel}:</strong> ${fmt(d.y)}
-      </div>
+      ${tooltipLine(xLabel, fmt(d.x))}
+      ${tooltipLine(yLabel, fmt(d.y))}
       ${
         selectedYAxisMetric === 'y_tpPerGpu' && d['inputTputPerGpu']
           ? `
-          <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-            <strong>Input Token Throughput per Chip:</strong> ${fmt(d['inputTputPerGpu'].y)}
-          </div>`
+          ${tooltipLine(t.inputTputPerChip, `${fmt(d['inputTputPerGpu'].y)}`)}`
           : ''
       }
       ${
         selectedYAxisMetric === 'y_tpPerGpu' && d['outputTputPerGpu']
           ? `
-          <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-            <strong>Output Token Throughput per Chip:</strong> ${fmt(d['outputTputPerGpu'].y)}
-          </div>`
+          ${tooltipLine(t.outputTputPerChip, `${fmt(d['outputTputPerGpu'].y)}`)}`
           : ''
       }
-      ${tooltipLine('Total Chips', d.tp)}
+      ${tooltipLine(t.totalChips, d.tp)}
       ${generateParallelismHTML(d, locale)}
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Concurrency:</strong> ${d.conc}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Precision:</strong> ${d.precision.toUpperCase()}
-      </div>
+      ${tooltipLine(t.concurrency, `${d.conc}`)}
+      ${tooltipLine(t.precision, `${d.precision.toUpperCase()}`)}
       ${generateCacheMetadataHTML(d, locale)}
       ${generateAgenticHTML(d, locale)}
       ${runLinkHTML(runUrl)}
-      ${viewChartsButtonHTML(isPinned, Boolean(hasTrace), d.id)}
+      ${viewActionsHTML(isPinned, Boolean(hasTrace), Boolean(config.hasLog), d.id, d.benchmark_type, locale)}
     </div>
   `;
 };
@@ -401,39 +480,28 @@ export const generateTooltipContent = (config: TooltipConfig): string => {
 export const generateOverlayTooltipContent = (config: OverlayTooltipConfig): string => {
   const { data: d, isPinned, xLabel, yLabel, overlayData } = config;
   const locale = config.locale ?? 'en';
+  const t = TOOLTIP_STRINGS[locale];
   const hwConfig = overlayData.hardwareConfig[d.hwKey];
   const perRow = overlayData.getRunForRow?.(d);
   const branch = perRow?.branch ?? overlayData.label;
 
   return `
     <div style="background: var(--popover); border: 2px solid #dc2626; border-radius: 8px; padding: 12px; box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1); user-select: ${isPinned ? 'text' : 'none'};">
-      ${isPinned ? '<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">Click elsewhere to dismiss</div>' : ''}
+      ${isPinned ? `<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">${t.dismiss}</div>` : ''}
       <div style="color: #dc2626; font-size: 10px; font-weight: 700; margin-bottom: 4px; text-transform: uppercase;">
-        ✕ UNOFFICIAL RUN
+        ${t.unofficialRun}
       </div>
       <div style="color: var(--foreground); font-size: 12px; font-weight: 600; margin-bottom: 8px;">
         ${hwConfig ? getDisplayLabel(hwConfig) : d.hwKey}
       </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Branch:</strong> ${branch}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Date:</strong> ${d.actualDate ?? d.date}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${xLabel}:</strong> ${fmt(d.x)}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${yLabel}:</strong> ${fmt(d.y)}
-      </div>
-      ${tooltipLine('Total Chips', d.tp)}
+      ${tooltipLine(t.branch, `${branch}`)}
+      ${tooltipLine(t.date, `${d.actualDate ?? d.date}`)}
+      ${tooltipLine(xLabel, fmt(d.x))}
+      ${tooltipLine(yLabel, fmt(d.y))}
+      ${tooltipLine(t.totalChips, d.tp)}
       ${generateParallelismHTML(d, locale)}
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Concurrency:</strong> ${d.conc}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Precision:</strong> ${d.precision.toUpperCase()}
-      </div>
+      ${tooltipLine(t.concurrency, `${d.conc}`)}
+      ${tooltipLine(t.precision, `${d.precision.toUpperCase()}`)}
       ${generateCacheMetadataHTML(d, locale)}
       ${generateAgenticHTML(d, locale)}
     </div>
@@ -457,58 +525,44 @@ export const generateGPUGraphTooltipContent = (config: TooltipConfig): string =>
     hardwareConfig,
     runUrl,
     hasTrace,
+    hasLog,
   } = config;
   const locale = config.locale ?? 'en';
+  const t = TOOLTIP_STRINGS[locale];
 
   return `
     <div style="background: var(--popover); border: 1px solid var(--border); border-radius: 8px; padding: 12px; box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1); user-select: ${isPinned ? 'text' : 'none'};">
-      ${isPinned ? '<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">Click elsewhere to dismiss</div>' : ''}
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Date:</strong> ${d.date}${d.actualDate && d.actualDate !== d.date ? ` <span style="opacity: 0.7">(data from ${d.actualDate})</span>` : ''}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Chip Config:</strong> ${hardwareConfig[d.hwKey] ? getDisplayLabel(hardwareConfig[d.hwKey]) : d.hwKey}
-      </div>
+      ${isPinned ? `<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">${t.dismiss}</div>` : ''}
+      ${tooltipLine(t.date, `${d.date}${d.actualDate && d.actualDate !== d.date ? ` <span style="opacity: 0.7">${t.dataFrom(d.actualDate)}</span>` : ''}`)}
+      ${tooltipLine(t.chipConfig, `${hardwareConfig[d.hwKey] ? getDisplayLabel(hardwareConfig[d.hwKey]) : d.hwKey}`)}
       ${
         d?.image
           ? `
-      ${imageTooltipLine(d.image)}`
+      ${imageTooltipLine(d.image, t.image)}`
           : ''
       }
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${xLabel}:</strong> ${fmt(d.x)}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>${yLabel}:</strong> ${fmt(d.y)}
-      </div>
+      ${tooltipLine(xLabel, fmt(d.x))}
+      ${tooltipLine(yLabel, fmt(d.y))}
       ${
         selectedYAxisMetric === 'y_tpPerGpu' && d['inputTputPerGpu']
           ? `
-          <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-            <strong>Input Token Throughput per Chip:</strong> ${fmt(d['inputTputPerGpu'].y)}
-          </div>`
+          ${tooltipLine(t.inputTputPerChip, `${fmt(d['inputTputPerGpu'].y)}`)}`
           : ''
       }
       ${
         selectedYAxisMetric === 'y_tpPerGpu' && d['outputTputPerGpu']
           ? `
-          <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-            <strong>Output Token Throughput per Chip:</strong> ${fmt(d['outputTputPerGpu'].y)}
-          </div>`
+          ${tooltipLine(t.outputTputPerChip, `${fmt(d['outputTputPerGpu'].y)}`)}`
           : ''
       }
-      ${tooltipLine('Total Chips', d.tp)}
+      ${tooltipLine(t.totalChips, d.tp)}
       ${generateParallelismHTML(d, locale)}
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Concurrency:</strong> ${d.conc}
-      </div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;">
-        <strong>Precision:</strong> ${d.precision.toUpperCase()}
-      </div>
+      ${tooltipLine(t.concurrency, `${d.conc}`)}
+      ${tooltipLine(t.precision, `${d.precision.toUpperCase()}`)}
       ${generateCacheMetadataHTML(d, locale)}
       ${generateAgenticHTML(d, locale)}
       ${runLinkHTML(runUrl)}
-      ${viewChartsButtonHTML(isPinned, Boolean(hasTrace), d.id)}
+      ${viewActionsHTML(isPinned, Boolean(hasTrace), Boolean(hasLog), d.id, d.benchmark_type, locale)}
     </div>
   `;
 };
diff --git a/packages/app/src/components/intro-section.tsx b/packages/app/src/components/intro-section.tsx
index e148f3389..6d00b4246 100644
--- a/packages/app/src/components/intro-section.tsx
+++ b/packages/app/src/components/intro-section.tsx
@@ -1,45 +1,52 @@
 import { Quote } from 'lucide-react';
 
 import { Card } from '@/components/ui/card';
-import { MinecraftSplash } from '@/components/minecraft/minecraft-splash';
-import { QuoteCarousel } from '@/components/quote-carousel';
-import { QUOTES, CAROUSEL_ORGS, CAROUSEL_LABELS } from '@/components/quotes/quotes-data';
+import { SupportersStrip } from '@/components/supporters-strip';
+import {
+  QUOTES,
+  CAROUSEL_ORGS,
+  CAROUSEL_LABELS,
+  orgAnchorId,
+} from '@/components/quotes/quotes-data';
 import type { Locale } from '@/lib/i18n';
 
-// Carousel order follows QUOTES order — carousel orgs are listed first there.
-const carouselQuotes = QUOTES.filter((q) => (CAROUSEL_ORGS as readonly string[]).includes(q.org));
+// Strip order follows QUOTES order — supporter orgs are listed first there.
+const stripOrgs = [
+  ...new Set(
+    QUOTES.filter((q) => (CAROUSEL_ORGS as readonly string[]).includes(q.org)).map((q) => q.org),
+  ),
+];
 
-const CAROUSEL_OVERRIDES = {
-  labels: CAROUSEL_LABELS,
-};
+const supporterOrgs = (quotesPath: string) =>
+  stripOrgs.map((org) => ({
+    label: CAROUSEL_LABELS[org] ?? org,
+    href: `${quotesPath}#${orgAnchorId(org)}`,
+  }));
 
 const HEADING = {
-  en: 'Open Source Continuous Inference Benchmark Trusted by GigaWatt Token Factories',
-  zh: '受吉瓦级 token 工厂信赖的开源持续推理基准测试',
+  en: 'Open-Source Continuous Agentic Inference Benchmark Trusted by GigaWatt Token Factories',
+  zh: 'InferenceX 提供持续更新的开源智能体推理基准测试，已获得吉瓦级 token 工厂运营方的信赖。',
 } as const;
 
 export function IntroSection({ locale = 'en' }: { locale?: Locale } = {}) {
   const isZh = locale === 'zh';
-  // Quotes fall back to the English original until a translation lands.
-  const quotes = isZh
-    ? carouselQuotes.map((q) => ({ ...q, text: q.textZh ?? q.text }))
-    : carouselQuotes;
   return (
     <section>
       <Card data-testid="intro-section">
-        <div className="relative flex items-start gap-2 mb-4">
+        {/* The splash moved to the AgentX hero, which now opens the landing
+            page — two copies of the same announcement one scroll apart read
+            as a duplicate rather than a callout. */}
+        <div className="flex items-start gap-2 mb-4">
           <Quote className="size-5 shrink-0 mt-1 text-brand" />
           <h2 className="text-lg font-semibold">{HEADING[locale]}</h2>
-          <MinecraftSplash />
-        </div>
-        <div>
-          <QuoteCarousel
-            quotes={quotes}
-            overrides={CAROUSEL_OVERRIDES}
-            moreHref={isZh ? '/zh/quotes' : '/quotes'}
-            moreLabel={isZh ? '查看更多支持者 →' : undefined}
-          />
         </div>
+        {/* Quote text lives on /quotes now — the band keeps just the org
+            strip and a link out, saving vertical space above the fold. */}
+        <SupportersStrip
+          orgs={supporterOrgs(isZh ? '/zh/quotes' : '/quotes')}
+          moreHref={isZh ? '/zh/quotes' : '/quotes'}
+          moreLabel={isZh ? '查看完整评价与更多支持者 →' : undefined}
+        />
       </Card>
     </section>
   );
diff --git a/packages/app/src/components/landing/curated-view-card.tsx b/packages/app/src/components/landing/curated-view-card.tsx
index 869f3d964..5613643b4 100644
--- a/packages/app/src/components/landing/curated-view-card.tsx
+++ b/packages/app/src/components/landing/curated-view-card.tsx
@@ -38,7 +38,7 @@ export function CuratedViewCard({ preset }: { preset: FavoritePreset }) {
         <h3 className="font-semibold text-sm leading-tight group-hover:text-brand transition-colors duration-200">
           <span className="align-middle">{title}</span>
           {isNew && (
-            <span className="ml-2 inline-flex items-center gap-1.5 align-middle rounded-full bg-brand px-2 py-0.5 text-[10px] font-bold uppercase tracking-wider text-primary-foreground shadow-sm">
+            <span className="ml-2 inline-flex items-center gap-1.5 align-middle rounded-full bg-brand px-2 py-0.5 text-3xs font-bold uppercase tracking-wider text-primary-foreground shadow-sm">
               New
             </span>
           )}
@@ -53,7 +53,7 @@ export function CuratedViewCard({ preset }: { preset: FavoritePreset }) {
           <Badge
             key={tag}
             variant="outline"
-            className="text-[10px] px-2 py-0.5 leading-tight border-border/60 text-muted-foreground group-hover:border-brand/30 group-hover:text-foreground/80 transition-colors duration-200"
+            className="text-3xs px-2 py-0.5 leading-tight border-border/60 text-muted-foreground group-hover:border-brand/30 group-hover:text-foreground/80 transition-colors duration-200"
           >
             {tag}
           </Badge>
diff --git a/packages/app/src/components/landing/landing-page.tsx b/packages/app/src/components/landing/landing-page.tsx
index 968c5f3d8..7a2ac59a5 100644
--- a/packages/app/src/components/landing/landing-page.tsx
+++ b/packages/app/src/components/landing/landing-page.tsx
@@ -1,6 +1,7 @@
-import { ArrowRight, BarChart3, ShieldCheck, Sparkles } from 'lucide-react';
+import { ArrowRight, ShieldCheck, Sparkles } from 'lucide-react';
 
 import { Card } from '@/components/ui/card';
+import { AgentXCompareHero } from '@/components/compare/agentx-compare-hero';
 import { IntroSection } from '@/components/intro-section';
 import { LandingPageAnalytics, LandingTrackedLink } from '@/components/landing/landing-analytics';
 import { CuratedViewCard } from '@/components/landing/curated-view-card';
@@ -11,13 +12,6 @@ import type { Locale } from '@/lib/i18n';
 
 const STRINGS = {
   en: {
-    exploreInferenceX: 'Explore InferenceX',
-    exploreInferenceXLead:
-      'Start with a concise cost overview across active models and key platforms, or open the full dashboard for every model, chip, framework, and metric.',
-    fullDashboard: 'Full Dashboard',
-    platformCoverage:
-      'Compare NVIDIA GB300 NVL72, GB200 NVL72, B300, B200, H200, H100, AMD MI355X, MI325X, MI300X and soon VR200 NVL72, AMD MI455X UALoE72, TPUv7 Ironwood, etc across DeepSeekv4 Pro, Qwen, Kimi, GLM, MiniMax, gpt-oss, Llama and other models.',
-    overview: 'Overview',
     reproTitle: 'Every Result Is Transparently done through Public GitHub Actions Automation',
     reproP1:
       'Every data point on the dashboard is produced by a public GitHub Actions workflow run. The recipe lives in the repo, the run executes on the actual target hardware, and the full logs and artifacts are publicly viewable. Click any point on a chart to jump straight to the run that produced it. All reproducible, auditable, and open source.',
@@ -40,33 +34,33 @@ const STRINGS = {
       'Jump straight into the most popular chip inference benchmark comparisons, curated and ready to explore.',
   },
   zh: {
-    exploreInferenceX: '探索 InferenceX',
-    exploreInferenceXLead:
-      '先查看活跃模型与关键平台的精简成本总览，或打开完整仪表板，探索所有模型、Chip、框架与指标。',
-    fullDashboard: '完整仪表板',
-    platformCoverage:
-      '跨 DeepSeekv4 Pro、Qwen、Kimi、GLM、MiniMax、gpt-oss、Llama 等模型，对比 NVIDIA GB300 NVL72、GB200 NVL72、B300、B200、H200、H100、AMD MI355X、MI325X、MI300X，以及即将上线的 VR200 NVL72、AMD MI455X UALoE72、TPUv7 Ironwood 等硬件。',
-    overview: '总览',
-    reproTitle: '每一条结果都通过公开的 GitHub Actions 自动化流程透明产生',
+    reproTitle: '所有结果均通过公开的 GitHub Actions 流程生成',
     reproP1:
-      '仪表板上的每个数据点都由公开的 GitHub Actions 工作流运行产生。配置方案（recipe）保存在公开仓库中，运行在真实目标硬件上执行，完整日志与产物公开可查。点击图表上的任意数据点即可跳转到生成它的那次运行。一切都可复现、可审计、开源。',
-    reproStat: '平均每周新增 1,000+ 条基准测试数据点。',
-    reproStatTail: '第一时间浏览每个新上线的模型、Chip、框架与配置。',
-    actionsRunsTitle: '公开的 Actions 运行',
-    actionsRunsDesc: '每次基准测试都在 GitHub Actions 上执行，运行过程中即可实时查看完整日志。',
-    openRecipesTitle: '开放的配置方案',
-    openRecipesDesc: '每个模型、框架、精度与并行配置都以 shell 脚本形式提交在公开仓库中。',
+      '仪表板上的每个数据点都来自一次公开的 GitHub Actions 运行。测试配置保存在仓库中，并在对应的真实硬件上执行；完整日志和产物均可公开查看。点击图表中的任意数据点，即可打开生成该结果的运行记录。整个过程可复现、可审计，并完全开源。',
+    reproStat: '平均每周新增 1,000 多个基准测试数据点。',
+    reproStatTail: '新模型、芯片、框架和配置上线后，都可以在这里查看。',
+    actionsRunsTitle: '公开运行记录',
+    actionsRunsDesc: '每次基准测试都通过 GitHub Actions 执行，运行期间即可查看完整日志。',
+    openRecipesTitle: '公开测试配置',
+    openRecipesDesc: '每个模型、框架、精度和并行配置都以 shell 脚本形式保存在公开仓库中。',
     dbSnapshotsTitle: '每周数据库快照',
     dbSnapshotsDesc:
-      '完整基准测试数据库每周以公开 GitHub Release 的形式发布，历史数据集持续可审计。',
+      '完整的基准测试数据库每周都会作为公开的 GitHub Release 发布，方便任何人核查历史数据。',
     browseSubmissions: '浏览提交记录',
-    viewRuns: '在 GitHub Actions 上查看基准测试运行',
-    howItWorks: '工作原理',
+    viewRuns: '查看 GitHub Actions 运行记录',
+    howItWorks: '了解复现流程',
     quickComparisons: '快速对比',
-    quickComparisonsDesc: '一键进入最热门的 Chip 推理基准测试对比，精选视图开箱即用。',
+    quickComparisonsDesc: '一键进入最热门的芯片推理基准测试对比，精选视图开箱即用。',
   },
 } as const;
 
+/**
+ * Quick Comparisons is hidden for now. The card, its `quickComparisons*`
+ * strings, `CuratedViewCard`, and `FAVORITE_PRESETS` are all left in place —
+ * flip this to `true` to bring the section back.
+ */
+const SHOW_QUICK_COMPARISONS = false;
+
 export function LandingPage({ locale = 'en' }: { locale?: Locale } = {}) {
   const t = STRINGS[locale];
   // Internal links stay within the current language tree.
@@ -76,42 +70,16 @@ export function LandingPage({ locale = 'en' }: { locale?: Locale } = {}) {
       <LandingPageAnalytics />
       <NudgeEngine scope="landing" />
       <div className="container mx-auto px-4 lg:px-8 flex flex-col gap-6 lg:gap-4">
+        {/* Same AgentX hero that leads /compare, above the quote carousel. The
+            landing page owns no h1 of its own, but keep this an h2 so the hero
+            stays a section within the page rather than retitling the whole
+            site. */}
+        <AgentXCompareHero locale={locale} headingLevel="h2" surface="landing" />
+
         <IntroSection locale={locale} />
 
         {/* Split: exploration entry points vs presets */}
         <section className="flex flex-col gap-4 pb-8">
-          {/* Primary product entry points */}
-          <Card>
-            <div className="flex items-center gap-2 mb-3">
-              <BarChart3 className="size-5 shrink-0 text-brand" />
-              <h2 className="text-lg font-semibold">{t.exploreInferenceX}</h2>
-            </div>
-            <p className="text-sm text-muted-foreground mb-2">{t.exploreInferenceXLead}</p>
-            <p className="text-sm text-muted-foreground mb-6">{t.platformCoverage}</p>
-            <div className="mt-auto flex flex-col gap-2 sm:flex-row">
-              <LandingTrackedLink
-                href={`${prefix}/overview`}
-                data-testid="landing-overview-link"
-                analyticsEvent="landing_overview_clicked"
-                appNavigation
-                className="inline-flex h-12 w-full items-center justify-center gap-2 rounded-md bg-brand px-8 text-sm font-medium text-primary-foreground transition-colors hover:bg-brand/90 focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-brand focus-visible:ring-offset-2 motion-reduce:transition-none sm:w-auto sm:text-base"
-              >
-                {t.overview}
-                <ArrowRight aria-hidden="true" className="size-4" />
-              </LandingTrackedLink>
-              <LandingTrackedLink
-                href={`${prefix}/inference`}
-                data-testid="landing-full-dashboard-link"
-                analyticsEvent="landing_full_dashboard_clicked"
-                appNavigation
-                className="inline-flex h-12 w-full items-center justify-center gap-2 rounded-md border border-border px-8 text-sm font-medium transition-colors hover:bg-accent focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-brand focus-visible:ring-offset-2 motion-reduce:transition-none sm:w-auto sm:text-base"
-              >
-                {t.fullDashboard}
-                <ArrowRight aria-hidden="true" className="size-4" />
-              </LandingTrackedLink>
-            </div>
-          </Card>
-
           {/* Reproducibility callout */}
           <Card>
             <div className="flex items-center gap-2 mb-3">
@@ -167,19 +135,21 @@ export function LandingPage({ locale = 'en' }: { locale?: Locale } = {}) {
             </div>
           </Card>
 
-          {/* Right - Curated Presets */}
-          <Card>
-            <div className="flex items-center gap-2 mb-3">
-              <Sparkles className="size-5 shrink-0 text-brand" />
-              <h2 className="text-lg font-semibold">{t.quickComparisons}</h2>
-            </div>
-            <p className="text-sm text-muted-foreground mb-4">{t.quickComparisonsDesc}</p>
-            <div className="grid grid-cols-1 sm:grid-cols-2 gap-3">
-              {FAVORITE_PRESETS.filter((preset) => !preset.hidden).map((preset) => (
-                <CuratedViewCard key={preset.id} preset={preset} />
-              ))}
-            </div>
-          </Card>
+          {/* Right - Curated Presets (temporarily hidden, see SHOW_QUICK_COMPARISONS) */}
+          {SHOW_QUICK_COMPARISONS && (
+            <Card data-testid="landing-quick-comparisons">
+              <div className="flex items-center gap-2 mb-3">
+                <Sparkles className="size-5 shrink-0 text-brand" />
+                <h2 className="text-lg font-semibold">{t.quickComparisons}</h2>
+              </div>
+              <p className="text-sm text-muted-foreground mb-4">{t.quickComparisonsDesc}</p>
+              <div className="grid grid-cols-1 sm:grid-cols-2 gap-3">
+                {FAVORITE_PRESETS.filter((preset) => !preset.hidden).map((preset) => (
+                  <CuratedViewCard key={preset.id} preset={preset} />
+                ))}
+              </div>
+            </Card>
+          )}
         </section>
       </div>
     </main>
diff --git a/packages/app/src/components/latest-image/latest-image-content.tsx b/packages/app/src/components/latest-image/latest-image-content.tsx
index 74d8446b8..77a5db1b8 100644
--- a/packages/app/src/components/latest-image/latest-image-content.tsx
+++ b/packages/app/src/components/latest-image/latest-image-content.tsx
@@ -68,7 +68,7 @@ const STRINGS = {
   },
   zh: {
     title: 'InferenceX 当前镜像',
-    description: '各模型与 Chip 配置的 Docker 镜像标签。',
+    description: '各模型与芯片配置的 Docker 镜像标签。',
     loading: '加载中……',
     error: '无法加载镜像数据。',
     noMatch: '没有符合当前筛选条件的镜像数据。',
@@ -80,8 +80,8 @@ const STRINGS = {
     tooltipIslOsl: '输入序列长度 / 输出序列长度（token 数）。',
     labelSpecDecode: '投机解码',
     tooltipSpecDecode: '投机解码方式。MTP = 多 Token 预测。',
-    labelGpuSku: 'Chip SKU',
-    tooltipGpuSku: '按 Chip 型号筛选（如 H200、MI300X、B200）。',
+    labelGpuSku: '芯片型号',
+    tooltipGpuSku: '按芯片型号筛选（如 H200、MI300X、B200）。',
     labelNodeType: '节点类型',
     tooltipNodeType:
       '单节点 = 非分离式服务。分离式 = 使用独立预填充/解码池，包括 Dynamo、Mori 和 llm-d。',
@@ -95,7 +95,7 @@ const STRINGS = {
     allFrameworks: '全部框架',
     thModel: '模型',
     thPrecision: '精度',
-    thGpuSku: 'Chip SKU',
+    thGpuSku: '芯片型号',
     thSpecDecode: '投机解码',
     thCurrentTag: '当前 InferenceX 镜像标签',
     thActualLatest: '实际最新标签',
diff --git a/packages/app/src/components/live-seo/format.ts b/packages/app/src/components/live-seo/format.ts
new file mode 100644
index 000000000..a791b896c
--- /dev/null
+++ b/packages/app/src/components/live-seo/format.ts
@@ -0,0 +1,35 @@
+/**
+ * Number formatting shared by the /rankings and /run server-rendered pages.
+ * Locale-independent by design: both the English and Chinese pages quote
+ * half-width digits with the same precision, per docs/chinese-copy.md.
+ */
+
+/** Tokens/s per GPU: whole numbers above 100, one decimal below. */
+export function fmtThroughput(value: number): string {
+  return value >= 100 ? Math.round(value).toLocaleString('en-US') : value.toFixed(1);
+}
+
+/** $ per million tokens: cents precision, sub-cent gets a third digit. */
+export function fmtCostPerMtok(value: number): string {
+  return `$${value >= 0.1 ? value.toFixed(2) : value.toFixed(3)}`;
+}
+
+/** $/GPU/hr price points from HW_REGISTRY. */
+export function fmtGpuHour(value: number): string {
+  return `$${value.toFixed(2)}`;
+}
+
+/** Percentage delta between a winner and a runner-up, e.g. 23. */
+export function pctFaster(winner: number, runnerUp: number): number {
+  return Math.round((winner / runnerUp - 1) * 100);
+}
+
+/** Percent below the runner-up: a $1 winner against a $2 runner-up is 50. */
+export function pctCheaper(winner: number, runnerUp: number): number {
+  return Math.round((1 - winner / runnerUp) * 100);
+}
+
+/** Milliseconds with sensible precision for TTFT/TPOT quotes. */
+export function fmtMs(value: number): string {
+  return value >= 100 ? `${Math.round(value)} ms` : `${value.toFixed(1)} ms`;
+}
diff --git a/packages/app/src/components/live-seo/rankings-page-sections.tsx b/packages/app/src/components/live-seo/rankings-page-sections.tsx
new file mode 100644
index 000000000..cf8df639a
--- /dev/null
+++ b/packages/app/src/components/live-seo/rankings-page-sections.tsx
@@ -0,0 +1,237 @@
+/**
+ * Shared server-rendered body for `/rankings/<slug>` and `/zh/rankings/<slug>`.
+ * One markup tree, two string bundles — the EN and ZH routes cannot drift.
+ * All numbers come pre-derived from `getRankingPageData`, which reads the
+ * exact cells the /overview leaderboard renders.
+ */
+
+import Link from 'next/link';
+
+import { fmtCostPerMtok, fmtThroughput, pctCheaper, pctFaster } from '@/components/live-seo/format';
+import type { RankingPageData } from '@/lib/run-rankings-data.server';
+import type { RankingPageEntry, RankingRow } from '@/lib/rankings';
+import { getRunPageEntry } from '@/lib/run-pages';
+
+export interface RankingsStrings {
+  backLabel: string;
+  heading: string;
+  scenarioNote: string;
+  tableCaption: string;
+  colRank: string;
+  colGpu: string;
+  colThroughput: string;
+  colCost: string;
+  colPrecision: string;
+  colEngine: string;
+  emptyState: string;
+  siblingLead: string;
+  siblingLabel: string;
+  methodologyHeading: string;
+  methodologyBody: string[];
+  faqHeading: string;
+  faq: { question: string; answer: string }[];
+  exploreHeading: string;
+  exploreLinks: { href: string; label: string }[];
+}
+
+function leadSentence(rows: RankingRow[], entry: RankingPageEntry, zh: boolean): string {
+  if (rows.length < 2) return '';
+  const [first, second] = rows;
+  if (entry.kind === 'fastest-gpu' && first.throughputPerGpu && second.throughputPerGpu) {
+    const pct = pctFaster(first.throughputPerGpu, second.throughputPerGpu);
+    const lead = `${first.hardwareLabel} leads at ${fmtThroughput(first.throughputPerGpu)} tokens/s per GPU, ${pct}% ahead of ${second.hardwareLabel}.`;
+    const leadZh = `${first.hardwareLabel} 以单 GPU 每秒 ${fmtThroughput(first.throughputPerGpu)} token 领先，比第二名 ${second.hardwareLabel} 高 ${pct}%。`;
+    return zh ? leadZh : lead;
+  }
+  if (entry.kind === 'cheapest-gpu' && first.costPerMtok && second.costPerMtok) {
+    const pct = pctCheaper(first.costPerMtok, second.costPerMtok);
+    const lead = `${first.hardwareLabel} is the cheapest at ${fmtCostPerMtok(first.costPerMtok)} per million tokens, ${pct}% below ${second.hardwareLabel}.`;
+    const leadZh = `${first.hardwareLabel} 成本最低，每百万 token 仅 ${fmtCostPerMtok(first.costPerMtok)}，比第二名 ${second.hardwareLabel} 低 ${pct}%。`;
+    return zh ? leadZh : lead;
+  }
+  return '';
+}
+
+export function rankingsLeadSentence(
+  rows: RankingRow[],
+  entry: RankingPageEntry,
+  locale: 'en' | 'zh',
+): string {
+  return leadSentence(rows, entry, locale === 'zh');
+}
+
+export function RankingsDetailContent({
+  entry,
+  data,
+  t,
+  locale,
+}: {
+  entry: RankingPageEntry;
+  data: RankingPageData;
+  t: RankingsStrings;
+  locale: 'en' | 'zh';
+}) {
+  const prefix = locale === 'zh' ? '/zh' : '';
+  const lead = rankingsLeadSentence(data.rows, entry, locale);
+
+  return (
+    <main className="relative">
+      <div className="container mx-auto px-4 lg:px-8">
+        <article className="mx-auto max-w-5xl">
+          <header className="pt-8 md:pt-12">
+            <Link
+              href={`${prefix}/rankings`}
+              className="group inline-flex items-center gap-2 text-sm text-muted-foreground transition-colors hover:text-brand"
+            >
+              <span
+                aria-hidden="true"
+                className="transition-transform group-hover:-translate-x-0.5"
+              >
+                ←
+              </span>
+              {t.backLabel}
+            </Link>
+            <div className="mt-6 flex flex-wrap items-center gap-2">
+              <span className="rounded-full border border-brand/25 bg-brand/8 px-3 py-1 text-xs font-semibold tracking-eyebrow text-brand uppercase">
+                {entry.model.label}
+              </span>
+            </div>
+            <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-5xl">
+              {t.heading}
+            </h1>
+            {lead && <p className="mt-4 max-w-3xl text-lg leading-relaxed">{lead}</p>}
+            <p className="mt-3 max-w-3xl text-sm text-muted-foreground">{t.scenarioNote}</p>
+          </header>
+
+          <section className="mt-10">
+            {data.rows.length === 0 ? (
+              <p className="rounded-xl border border-border/50 p-6 leading-7 text-muted-foreground">
+                {t.emptyState}
+              </p>
+            ) : (
+              <div className="overflow-x-auto rounded-xl border border-border/50">
+                <table className="w-full min-w-[40rem] text-sm">
+                  <caption className="sr-only">{t.tableCaption}</caption>
+                  <thead>
+                    <tr className="border-b border-border/50 text-left text-xs tracking-widest text-muted-foreground uppercase">
+                      <th scope="col" className="px-4 py-3">
+                        {t.colRank}
+                      </th>
+                      <th scope="col" className="px-4 py-3">
+                        {t.colGpu}
+                      </th>
+                      <th scope="col" className="px-4 py-3">
+                        {t.colThroughput}
+                      </th>
+                      <th scope="col" className="px-4 py-3">
+                        {t.colCost}
+                      </th>
+                      <th scope="col" className="px-4 py-3">
+                        {t.colPrecision}
+                      </th>
+                      <th scope="col" className="px-4 py-3">
+                        {t.colEngine}
+                      </th>
+                    </tr>
+                  </thead>
+                  <tbody>
+                    {data.rows.map((row) => {
+                      const runEntry = row.chip
+                        ? getRunPageEntry(`${entry.model.slug}-on-${row.chip.slug}`)
+                        : undefined;
+                      return (
+                        <tr
+                          key={row.hardware}
+                          className="border-b border-border/30 last:border-b-0"
+                        >
+                          <td className="px-4 py-3 font-mono">{row.rank}</td>
+                          <td className="px-4 py-3 font-medium">
+                            {runEntry ? (
+                              <Link
+                                href={`${prefix}/run/${runEntry.slug}`}
+                                className="text-brand hover:underline"
+                              >
+                                {row.hardwareLabel}
+                              </Link>
+                            ) : (
+                              row.hardwareLabel
+                            )}
+                          </td>
+                          <td className="px-4 py-3 font-mono">
+                            {row.throughputPerGpu === null
+                              ? '-'
+                              : fmtThroughput(row.throughputPerGpu)}
+                          </td>
+                          <td className="px-4 py-3 font-mono">
+                            {row.costPerMtok === null ? '-' : fmtCostPerMtok(row.costPerMtok)}
+                          </td>
+                          <td className="px-4 py-3 uppercase">{row.precision ?? '-'}</td>
+                          <td className="px-4 py-3">{row.framework ?? '-'}</td>
+                        </tr>
+                      );
+                    })}
+                  </tbody>
+                </table>
+              </div>
+            )}
+            <p className="mt-4 text-sm text-muted-foreground">
+              {t.siblingLead}{' '}
+              <Link
+                href={`${prefix}/rankings/${entry.kind === 'fastest-gpu' ? 'cheapest-gpu' : 'fastest-gpu'}-for-${entry.model.slug}`}
+                className="font-medium text-brand hover:underline"
+              >
+                {t.siblingLabel}
+              </Link>
+            </p>
+          </section>
+
+          <section
+            aria-labelledby="ranking-methodology"
+            className="mt-10 border-t border-border/50 pt-10"
+          >
+            <h2 id="ranking-methodology" className="text-xl font-semibold tracking-tight">
+              {t.methodologyHeading}
+            </h2>
+            {t.methodologyBody.map((paragraph) => (
+              <p key={paragraph.slice(0, 32)} className="mt-3 leading-7 text-muted-foreground">
+                {paragraph}
+              </p>
+            ))}
+          </section>
+
+          <section aria-labelledby="ranking-faq" className="mt-10 border-t border-border/50 pt-10">
+            <h2 id="ranking-faq" className="text-xl font-semibold tracking-tight">
+              {t.faqHeading}
+            </h2>
+            <dl className="mt-4 space-y-6">
+              {t.faq.map((item) => (
+                <div key={item.question}>
+                  <dt className="font-medium">{item.question}</dt>
+                  <dd className="mt-2 leading-7 text-muted-foreground">{item.answer}</dd>
+                </div>
+              ))}
+            </dl>
+          </section>
+
+          <section
+            aria-labelledby="ranking-explore"
+            className="mt-10 mb-16 rounded-xl border border-brand/20 bg-brand/6 p-5 md:p-6"
+          >
+            <h2 id="ranking-explore" className="text-xl font-semibold tracking-tight">
+              {t.exploreHeading}
+            </h2>
+            <ul className="mt-3 space-y-1">
+              {t.exploreLinks.map((link) => (
+                <li key={link.href}>
+                  <Link href={link.href} className="text-sm font-medium text-brand hover:underline">
+                    {link.label}
+                  </Link>
+                </li>
+              ))}
+            </ul>
+          </section>
+        </article>
+      </div>
+    </main>
+  );
+}
diff --git a/packages/app/src/components/live-seo/run-page-sections.tsx b/packages/app/src/components/live-seo/run-page-sections.tsx
new file mode 100644
index 000000000..1d9d90f5f
--- /dev/null
+++ b/packages/app/src/components/live-seo/run-page-sections.tsx
@@ -0,0 +1,284 @@
+/**
+ * Shared server-rendered body for `/run/<pair>` and `/zh/run/<pair>`.
+ * One markup tree, two string bundles — the EN and ZH routes cannot drift.
+ * All numbers come pre-derived from `getRunPageData`, which reads through the
+ * same cached benchmark query the dashboard renders.
+ */
+
+import Link from 'next/link';
+
+import { fmtCostPerMtok, fmtGpuHour, fmtMs, fmtThroughput } from '@/components/live-seo/format';
+import { getChipHw } from '@/lib/chip-pages';
+import type { RunPageData } from '@/lib/run-rankings-data.server';
+import type { RunPageEntry } from '@/lib/run-pages';
+
+export interface RunStrings {
+  backHref: string;
+  backLabel: string;
+  heading: string;
+  quickAnswerLabel: string;
+  quickAnswer: string;
+  statConfigs: string;
+  statEngines: string;
+  statPrecisions: string;
+  statFreshness: string;
+  ladderHeading: string;
+  ladderIntro: string;
+  colTier: string;
+  colThroughput: string;
+  colCost: string;
+  colEngine: string;
+  colPrecision: string;
+  costHeading: string;
+  costIntro: string;
+  colPriceTier: string;
+  colGpuHour: string;
+  colCostPerMtok: string;
+  priceTierLabels: Record<'hyperscaler' | 'neocloud' | 'retail', string>;
+  emptyState: string;
+  faqHeading: string;
+  faq: { question: string; answer: string }[];
+  exploreHeading: string;
+  exploreLinks: { href: string; label: string }[];
+}
+
+export function RunDetailContent({
+  entry,
+  data,
+  t,
+}: {
+  entry: RunPageEntry;
+  data: RunPageData;
+  t: RunStrings;
+}) {
+  return (
+    <main className="relative">
+      <div className="container mx-auto px-4 lg:px-8">
+        <article className="mx-auto max-w-5xl">
+          <header className="pt-8 md:pt-12">
+            <Link
+              href={t.backHref}
+              className="group inline-flex items-center gap-2 text-sm text-muted-foreground transition-colors hover:text-brand"
+            >
+              <span
+                aria-hidden="true"
+                className="transition-transform group-hover:-translate-x-0.5"
+              >
+                ←
+              </span>
+              {t.backLabel}
+            </Link>
+            <div className="mt-6 flex flex-wrap items-center gap-2">
+              <span className="rounded-full border border-brand/25 bg-brand/8 px-3 py-1 text-xs font-semibold tracking-eyebrow text-brand uppercase">
+                {entry.model.label}
+              </span>
+              <span className="font-mono text-xs tracking-eyebrow text-muted-foreground uppercase">
+                {getChipHw(entry.chip).vendor} {getChipHw(entry.chip).arch}
+              </span>
+            </div>
+            <h1 className="mt-4 max-w-4xl text-4xl font-bold tracking-heading text-balance md:text-5xl">
+              {t.heading}
+            </h1>
+          </header>
+
+          {data.hasData ? (
+            <>
+              <section
+                aria-labelledby="run-quick-answer"
+                className="mt-8 rounded-xl border border-brand/20 bg-brand/6 p-5 md:p-6"
+              >
+                <p
+                  id="run-quick-answer"
+                  className="font-mono text-xs font-semibold tracking-eyebrow text-brand uppercase"
+                >
+                  {t.quickAnswerLabel}
+                </p>
+                <p className="mt-3 text-lg leading-relaxed font-medium text-pretty md:text-xl">
+                  {t.quickAnswer}
+                </p>
+              </section>
+
+              <section
+                aria-label={t.statConfigs}
+                className="mt-8 grid gap-4 sm:grid-cols-2 lg:grid-cols-4"
+              >
+                <div className="rounded-xl border border-border/50 p-4">
+                  <p className="text-xs tracking-widest text-muted-foreground uppercase">
+                    {t.statConfigs}
+                  </p>
+                  <p className="mt-1 font-mono text-2xl font-semibold">{data.configCount}</p>
+                </div>
+                <div className="rounded-xl border border-border/50 p-4">
+                  <p className="text-xs tracking-widest text-muted-foreground uppercase">
+                    {t.statEngines}
+                  </p>
+                  <p className="mt-1 text-sm leading-6 font-medium">{data.frameworks.join(', ')}</p>
+                </div>
+                <div className="rounded-xl border border-border/50 p-4">
+                  <p className="text-xs tracking-widest text-muted-foreground uppercase">
+                    {t.statPrecisions}
+                  </p>
+                  <p className="mt-1 text-sm leading-6 font-medium uppercase">
+                    {data.precisions.join(', ')}
+                  </p>
+                </div>
+                <div className="rounded-xl border border-border/50 p-4">
+                  <p className="text-xs tracking-widest text-muted-foreground uppercase">
+                    {t.statFreshness}
+                  </p>
+                  <p className="mt-1 font-mono text-sm leading-6 font-medium">
+                    {data.oldest} → {data.newest}
+                  </p>
+                </div>
+              </section>
+
+              <section
+                aria-labelledby="run-ladder"
+                className="mt-10 border-t border-border/50 pt-10"
+              >
+                <h2 id="run-ladder" className="text-xl font-semibold tracking-tight">
+                  {t.ladderHeading}
+                </h2>
+                <p className="mt-3 leading-7 text-muted-foreground">{t.ladderIntro}</p>
+                <div className="mt-4 overflow-x-auto rounded-xl border border-border/50">
+                  <table className="w-full min-w-[36rem] text-sm">
+                    <thead>
+                      <tr className="border-b border-border/50 text-left text-xs tracking-widest text-muted-foreground uppercase">
+                        <th scope="col" className="px-4 py-3">
+                          {t.colTier}
+                        </th>
+                        <th scope="col" className="px-4 py-3">
+                          {t.colThroughput}
+                        </th>
+                        <th scope="col" className="px-4 py-3">
+                          {t.colCost}
+                        </th>
+                        <th scope="col" className="px-4 py-3">
+                          {t.colEngine}
+                        </th>
+                        <th scope="col" className="px-4 py-3">
+                          {t.colPrecision}
+                        </th>
+                      </tr>
+                    </thead>
+                    <tbody>
+                      {data.tierLadder.map((read) => (
+                        <tr key={read.tier} className="border-b border-border/30 last:border-b-0">
+                          <td className="px-4 py-3 font-mono">{read.tier} tok/s</td>
+                          <td className="px-4 py-3 font-mono">
+                            {read.throughputPerGpu === null
+                              ? '-'
+                              : fmtThroughput(read.throughputPerGpu)}
+                          </td>
+                          <td className="px-4 py-3 font-mono">
+                            {read.costPerMtok === null ? '-' : fmtCostPerMtok(read.costPerMtok)}
+                          </td>
+                          <td className="px-4 py-3">{read.framework ?? '-'}</td>
+                          <td className="px-4 py-3 uppercase">{read.precision ?? '-'}</td>
+                        </tr>
+                      ))}
+                    </tbody>
+                  </table>
+                </div>
+              </section>
+
+              {data.costTiers.length > 0 && (
+                <section
+                  aria-labelledby="run-cost"
+                  className="mt-10 border-t border-border/50 pt-10"
+                >
+                  <h2 id="run-cost" className="text-xl font-semibold tracking-tight">
+                    {t.costHeading}
+                  </h2>
+                  <p className="mt-3 leading-7 text-muted-foreground">{t.costIntro}</p>
+                  <div className="mt-4 overflow-x-auto rounded-xl border border-border/50">
+                    <table className="w-full min-w-[28rem] text-sm">
+                      <thead>
+                        <tr className="border-b border-border/50 text-left text-xs tracking-widest text-muted-foreground uppercase">
+                          <th scope="col" className="px-4 py-3">
+                            {t.colPriceTier}
+                          </th>
+                          <th scope="col" className="px-4 py-3">
+                            {t.colGpuHour}
+                          </th>
+                          <th scope="col" className="px-4 py-3">
+                            {t.colCostPerMtok}
+                          </th>
+                        </tr>
+                      </thead>
+                      <tbody>
+                        {data.costTiers.map((tier) => (
+                          <tr
+                            key={tier.tierLabel}
+                            className="border-b border-border/30 last:border-b-0"
+                          >
+                            <td className="px-4 py-3 font-medium">
+                              {t.priceTierLabels[tier.tierLabel]}
+                            </td>
+                            <td className="px-4 py-3 font-mono">
+                              {fmtGpuHour(tier.costPerGpuHour)}
+                            </td>
+                            <td className="px-4 py-3 font-mono">
+                              {fmtCostPerMtok(tier.costPerMtok)}
+                            </td>
+                          </tr>
+                        ))}
+                      </tbody>
+                    </table>
+                  </div>
+                </section>
+              )}
+
+              <section aria-labelledby="run-faq" className="mt-10 border-t border-border/50 pt-10">
+                <h2 id="run-faq" className="text-xl font-semibold tracking-tight">
+                  {t.faqHeading}
+                </h2>
+                <dl className="mt-4 space-y-6">
+                  {t.faq.map((item) => (
+                    <div key={item.question}>
+                      <dt className="font-medium">{item.question}</dt>
+                      <dd className="mt-2 leading-7 text-muted-foreground">{item.answer}</dd>
+                    </div>
+                  ))}
+                </dl>
+              </section>
+            </>
+          ) : (
+            <p className="mt-10 mb-16 rounded-xl border border-border/50 p-6 leading-7 text-muted-foreground">
+              {t.emptyState}
+            </p>
+          )}
+
+          <section
+            aria-labelledby="run-explore"
+            className="mt-10 mb-16 rounded-xl border border-brand/20 bg-brand/6 p-5 md:p-6"
+          >
+            <h2 id="run-explore" className="text-xl font-semibold tracking-tight">
+              {t.exploreHeading}
+            </h2>
+            <ul className="mt-3 space-y-1">
+              {t.exploreLinks.map((link) => (
+                <li key={link.href}>
+                  <Link href={link.href} className="text-sm font-medium text-brand hover:underline">
+                    {link.label}
+                  </Link>
+                </li>
+              ))}
+            </ul>
+          </section>
+        </article>
+      </div>
+    </main>
+  );
+}
+
+/**
+ * Formatted best-of-config TTFT and TPOT for the quick answers. Returned as
+ * separate values because each is minimized independently across configs:
+ * the two bests can come from different runs, so composing them as a single
+ * measured pair would misstate the data.
+ */
+export function latencyQuote(data: RunPageData): { ttft: string; tpot: string } | null {
+  if (data.bestMedianTtft === null || data.bestMedianTpot === null) return null;
+  return { ttft: fmtMs(data.bestMedianTtft), tpot: fmtMs(data.bestMedianTpot) };
+}
diff --git a/packages/app/src/components/minecraft/minecraft-decorations.test.tsx b/packages/app/src/components/minecraft/minecraft-decorations.test.tsx
new file mode 100644
index 000000000..7ae4f76cf
--- /dev/null
+++ b/packages/app/src/components/minecraft/minecraft-decorations.test.tsx
@@ -0,0 +1,68 @@
+// @vitest-environment jsdom
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { MinecraftDecorations } from './minecraft-decorations';
+
+interface AudioStub {
+  src: string;
+  volume: number;
+  play: () => Promise<void>;
+  pause: () => void;
+}
+
+let audioInstances: AudioStub[];
+let container: HTMLDivElement;
+let root: Root;
+
+beforeEach(() => {
+  audioInstances = [];
+  document.documentElement.className = 'minecraft';
+  localStorage.removeItem('minecraft-sound');
+
+  vi.stubGlobal(
+    'matchMedia',
+    vi.fn(() => ({ matches: false })),
+  );
+  vi.stubGlobal(
+    'Audio',
+    class implements AudioStub {
+      src: string;
+      volume = 1;
+      play = vi.fn().mockResolvedValue(undefined);
+      pause = vi.fn();
+
+      constructor(src: string) {
+        this.src = src;
+        audioInstances.push(this);
+      }
+    },
+  );
+
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  document.documentElement.className = '';
+  vi.unstubAllGlobals();
+});
+
+describe('MinecraftDecorations', () => {
+  it('plays the dragon growl at quarter volume', async () => {
+    await act(async () => {
+      root.render(<MinecraftDecorations />);
+      await Promise.resolve();
+    });
+
+    expect(audioInstances).toHaveLength(1);
+    expect(audioInstances[0]).toMatchObject({
+      src: '/decorative/minecraft/ender-dragon.mp3',
+      volume: 0.25,
+    });
+  });
+});
diff --git a/packages/app/src/components/minecraft/minecraft-decorations.tsx b/packages/app/src/components/minecraft/minecraft-decorations.tsx
index b3772438d..628dc4890 100644
--- a/packages/app/src/components/minecraft/minecraft-decorations.tsx
+++ b/packages/app/src/components/minecraft/minecraft-decorations.tsx
@@ -123,7 +123,7 @@ export function MinecraftDecorations() {
     if (window.matchMedia('(prefers-reduced-motion: reduce)').matches) return;
 
     const audio = new Audio('/decorative/minecraft/ender-dragon.mp3');
-    audio.volume = 0.5;
+    audio.volume = 0.25;
     const timeout = window.setTimeout(() => {
       audio.play().catch(() => {
         /* browser blocked autoplay — no prior user gesture */
@@ -170,7 +170,7 @@ export function MinecraftDecorations() {
         />
       </div>
 
-      <div className="hidden lg:block fixed bottom-1 right-2 z-0 text-[10px] text-foreground/50 text-right leading-tight">
+      <div className="hidden lg:block fixed bottom-1 right-2 z-0 text-3xs text-foreground/50 text-right leading-tight">
         <div>
           art:{' '}
           <a
diff --git a/packages/app/src/components/minecraft/minecraft-splash.test.tsx b/packages/app/src/components/minecraft/minecraft-splash.test.tsx
new file mode 100644
index 000000000..5ad5a8d18
--- /dev/null
+++ b/packages/app/src/components/minecraft/minecraft-splash.test.tsx
@@ -0,0 +1,85 @@
+// @vitest-environment jsdom
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const pathnameStub = vi.hoisted(() => ({ value: '/' }));
+
+vi.mock('next/navigation', () => ({
+  usePathname: () => pathnameStub.value,
+}));
+
+import { MinecraftSplash, SPLASHES } from './minecraft-splash';
+
+let container: HTMLDivElement;
+let root: Root;
+
+function render() {
+  act(() => root.render(<MinecraftSplash />));
+}
+
+/** MutationObserver callbacks land on the microtask queue. */
+async function flush() {
+  await act(async () => {
+    await Promise.resolve();
+  });
+}
+
+function splashText(): string | null {
+  return container.querySelector('.splash-text')?.textContent ?? null;
+}
+
+beforeEach(() => {
+  pathnameStub.value = '/';
+  document.documentElement.className = '';
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  document.documentElement.className = '';
+});
+
+describe('MinecraftSplash', () => {
+  it('announces AgentX outside the minecraft theme (light + dark)', () => {
+    render();
+    expect(splashText()).toBe('AgentX is here!!');
+  });
+
+  it('announces AgentX in Chinese on /zh pages', () => {
+    pathnameStub.value = '/zh';
+    render();
+    expect(splashText()).toBe('AgentX 来了！！');
+  });
+
+  it('swaps to a random splash from the list when the minecraft theme is on', async () => {
+    document.documentElement.classList.add('minecraft');
+    render();
+    await flush();
+    expect(SPLASHES).toContain(splashText());
+  });
+
+  it('falls back to the announcement when the minecraft theme is turned off', async () => {
+    document.documentElement.classList.add('minecraft');
+    render();
+    await flush();
+
+    await act(async () => {
+      document.documentElement.classList.remove('minecraft');
+      await Promise.resolve();
+    });
+    expect(splashText()).toBe('AgentX is here!!');
+  });
+
+  it('renders the same markup on the server as on the first client render', () => {
+    // The random pick is deferred to an effect precisely so SSR and hydration
+    // agree — a splash chosen during render would mismatch on every load.
+    render();
+    const first = splashText();
+    act(() => root.render(<MinecraftSplash />));
+    expect(splashText()).toBe(first);
+  });
+});
diff --git a/packages/app/src/components/minecraft/minecraft-splash.tsx b/packages/app/src/components/minecraft/minecraft-splash.tsx
index 5f77153ef..0246bef6d 100644
--- a/packages/app/src/components/minecraft/minecraft-splash.tsx
+++ b/packages/app/src/components/minecraft/minecraft-splash.tsx
@@ -2,7 +2,10 @@
 
 import { useState, useEffect } from 'react';
 
-const SPLASHES = [
+import { useLocale } from '@/lib/use-locale';
+
+export const SPLASHES = [
+  'AgentX is here!!',
   'Now with more tokens!',
   'Chip go brrr!',
   'Also try SGLang!',
@@ -35,16 +38,48 @@ const SPLASHES = [
   'Roofline analysis!',
 ];
 
+/**
+ * Splash shown outside the minecraft theme. Light and dark mode get a single
+ * fixed announcement rather than the random rotation — the rotation is a
+ * minecraft-theme easter egg, while this is a launch callout that has to say
+ * the same thing on every load.
+ */
+/** Mirrors the `@media (min-width: 1024px)` rule that reveals `.splash-wrapper`. */
+const SPLASH_MIN_WIDTH = '(min-width: 1024px)';
+
+const ANNOUNCEMENT = {
+  en: 'AgentX is here!!',
+  zh: 'AgentX 来了！！',
+} as const;
+
 /**
  * Splash text — yellow, rotated, bouncing text (Minecraft title screen style)
- * that appears on the landing page only when the minecraft theme is active.
+ * on the landing page. Light and dark mode show the fixed AgentX announcement;
+ * the minecraft theme keeps the random per-load rotation.
  */
 export function MinecraftSplash() {
-  const [splash, setSplash] = useState('');
+  const locale = useLocale();
+  const [randomSplash, setRandomSplash] = useState('');
   const [isMinecraft, setIsMinecraft] = useState(false);
+  const [wideEnough, setWideEnough] = useState(false);
+
+  // Picked client-side only: a server-rendered random pick would mismatch on
+  // hydration. Minecraft mode is the only consumer, and it starts `false`.
+  useEffect(() => {
+    setRandomSplash(SPLASHES[Math.floor(Math.random() * SPLASHES.length)]);
+  }, []);
 
+  // `.splash-wrapper` is `display: none` below 1024px, but hiding it in CSS is
+  // not enough: the browser still fetches the Monocraft webfont the splash
+  // asks for, so a phone pays for a font it never paints. Gate the element out
+  // of the tree entirely below the same breakpoint the stylesheet uses.
+  // Starts `false` so the server render and the first client render agree.
   useEffect(() => {
-    setSplash(SPLASHES[Math.floor(Math.random() * SPLASHES.length)]);
+    const query = window.matchMedia(SPLASH_MIN_WIDTH);
+    const sync = () => setWideEnough(query.matches);
+    sync();
+    query.addEventListener('change', sync);
+    return () => query.removeEventListener('change', sync);
   }, []);
 
   useEffect(() => {
@@ -57,10 +92,11 @@ export function MinecraftSplash() {
     return () => observer.disconnect();
   }, []);
 
-  if (!splash || !isMinecraft) return null;
+  const splash = isMinecraft ? randomSplash : ANNOUNCEMENT[locale];
+  if (!splash || !wideEnough) return null;
 
   return (
-    <div className="splash-wrapper">
+    <div className="splash-wrapper" data-testid="splash-text">
       <span className="splash-text">{splash}</span>
     </div>
   );
diff --git a/packages/app/src/components/model/EmbeddedModelDashboard.tsx b/packages/app/src/components/model/EmbeddedModelDashboard.tsx
new file mode 100644
index 000000000..1ec955a22
--- /dev/null
+++ b/packages/app/src/components/model/EmbeddedModelDashboard.tsx
@@ -0,0 +1,50 @@
+'use client';
+
+import { GlobalFilterProvider } from '@/components/GlobalFilterContext';
+import { InferenceProvider } from '@/components/inference/InferenceContext';
+import InferenceChartDisplay from '@/components/inference/ui/ChartDisplay';
+import { EphemeralUrlStateContext } from '@/hooks/useUrlState';
+import { toModel, toSequence } from '@/lib/compare-enum-coerce';
+import { DEFAULT_Y_AXIS_METRIC } from '@/lib/url-state';
+
+/**
+ * Live InferenceX dashboard embedded on a `/model/[slug]` page, without the
+ * header section (title, description, and model/scenario/metric/chip
+ * selectors). The view is seeded to the page's model, its featured scenario
+ * (AgentX where available, otherwise 8K→1K), and the default
+ * total-tokens-per-$ y-axis metric; every chip config with data is
+ * auto-selected so the embed renders populated charts on first paint.
+ *
+ * The whole embed runs inside an ephemeral URL-state scope: its providers are
+ * the same ones the primary dashboards use, and those persist their state in
+ * the module-scoped store in `url-state.ts` that survives client-side
+ * navigations. Without the scope, the embed's auto-populated selection (all
+ * chip configs, seeded model/scenario) would overwrite that store, so
+ * Back-navigating to a bare `/inference` would rebuild WITH the embed's state
+ * — auto-populated chip configs and an unfolded config changelog the user
+ * never selected.
+ */
+export default function EmbeddedModelDashboard({
+  displayName,
+  sequence,
+}: {
+  displayName: string;
+  sequence: string;
+}) {
+  return (
+    <EphemeralUrlStateContext.Provider value={true}>
+      <GlobalFilterProvider
+        initialModel={toModel(displayName)}
+        initialSequence={toSequence(sequence)}
+      >
+        <InferenceProvider
+          activeTab="inference"
+          initialYAxisMetric={DEFAULT_Y_AXIS_METRIC}
+          autoSelectAllGpus
+        >
+          <InferenceChartDisplay embedded />
+        </InferenceProvider>
+      </GlobalFilterProvider>
+    </EphemeralUrlStateContext.Provider>
+  );
+}
diff --git a/packages/app/src/components/model/ModelArchitectureInline.tsx b/packages/app/src/components/model/ModelArchitectureInline.tsx
new file mode 100644
index 000000000..358e41236
--- /dev/null
+++ b/packages/app/src/components/model/ModelArchitectureInline.tsx
@@ -0,0 +1,25 @@
+'use client';
+
+import dynamic from 'next/dynamic';
+
+import { Skeleton } from '@/components/ui/skeleton';
+import { toModel } from '@/lib/compare-enum-coerce';
+
+// Client-only like the old dashboard drawer: the SVG diagram sizes itself
+// against the rendered container, so it has no meaningful SSR output.
+const ModelArchitectureDiagram = dynamic(
+  () => import('@/components/inference/ui/ModelArchitectureDiagram'),
+  { ssr: false, loading: () => <Skeleton className="h-40 w-full" /> },
+);
+
+/**
+ * Always-expanded architecture diagram for `/model/[slug]` pages. Renders
+ * nothing when the model has no `MODEL_ARCHITECTURES` entry (the diagram
+ * component returns null) — the page's MDX Architecture section still covers
+ * those models in prose.
+ */
+export default function ModelArchitectureInline({ displayName }: { displayName: string }) {
+  const model = toModel(displayName);
+  if (!model) return null;
+  return <ModelArchitectureDiagram model={model} variant="inline" />;
+}
diff --git a/packages/app/src/components/model/ModelDeveloperLogo.tsx b/packages/app/src/components/model/ModelDeveloperLogo.tsx
new file mode 100644
index 000000000..f5bd7d568
--- /dev/null
+++ b/packages/app/src/components/model/ModelDeveloperLogo.tsx
@@ -0,0 +1,28 @@
+import { getModelDeveloperLogo, isMonochromeLogo } from '@/lib/model-logos';
+
+/**
+ * Full-color developer logo for `/model` surfaces. Renders nothing when the
+ * developer has no logo mapping, so pages can treat the logo as optional
+ * decoration. Monochrome (black `currentColor`) marks get a dark-mode invert
+ * so they stay visible on both themes; full-color logos render as-is.
+ */
+export default function ModelDeveloperLogo({
+  developer,
+  className,
+}: {
+  developer: string;
+  className?: string;
+}) {
+  const logo = getModelDeveloperLogo(developer);
+  if (!logo) return null;
+
+  return (
+    <img
+      src={`/logos/${logo}`}
+      alt={`${developer} logo`}
+      width={48}
+      height={48}
+      className={`shrink-0 object-contain ${isMonochromeLogo(logo) ? 'dark:invert' : ''} ${className ?? ''}`}
+    />
+  );
+}
diff --git a/packages/app/src/components/nudge-engine.tsx b/packages/app/src/components/nudge-engine.tsx
index 3a24786e9..76386c963 100644
--- a/packages/app/src/components/nudge-engine.tsx
+++ b/packages/app/src/components/nudge-engine.tsx
@@ -5,7 +5,7 @@ import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
 
 import { track } from '@/lib/analytics';
 import { useLocale } from '@/lib/use-locale';
-import type { Locale } from '@/lib/i18n';
+import { localePath, type Locale } from '@/lib/i18n';
 import {
   isDismissed,
   isPermanentlySuppressed,
@@ -15,10 +15,12 @@ import {
 } from '@/lib/nudges/persistence';
 import { dismissesOnAction } from '@/lib/nudges/policy';
 import { NUDGE_REGISTRY } from '@/lib/nudges/registry';
-import type { NudgeDefinition, NudgeTrigger } from '@/lib/nudges/types';
+import type { NudgeDefinition, NudgeScope, NudgeTrigger } from '@/lib/nudges/types';
 import { LANDING_BANNER_DISMISSED_ATTRIBUTE } from '@/lib/nudges/landing-banner';
 import { BottomToast } from '@/components/ui/bottom-toast';
+import { CoachMark } from '@/components/ui/coach-mark';
 import { Button } from '@/components/ui/button';
+import { NewBadge } from '@/components/ui/new-badge';
 
 // ---------------------------------------------------------------------------
 // Helpers
@@ -29,11 +31,24 @@ function trackNudgeEvent(def: NudgeDefinition, event: 'shown' | 'dismissed' | 'a
   track(name, def.analytics?.properties);
 }
 
-function isEligible(def: NudgeDefinition): boolean {
+/**
+ * `requireAnchor` separates the two questions the engine asks.
+ *
+ * Trigger *setup* asks "could this nudge ever fire?" and must not consult the
+ * anchor: a coach mark's anchor only exists once the chart has painted, and
+ * skipping setup at mount would mean its triggers are never wired up at all.
+ * Showing asks "can it fire right now?" — there a missing anchor is
+ * disqualifying, and the show is simply retried on the next trigger.
+ */
+function isEligible(
+  def: NudgeDefinition,
+  { requireAnchor = true, requireConditions = true } = {},
+): boolean {
+  if (requireAnchor && def.type === 'coach-mark' && !def.content.anchor?.resolve()) return false;
   if (!isWithinSchedule(def.schedule)) return false;
   if (def.permanentSuppressKey && isPermanentlySuppressed(def.permanentSuppressKey)) return false;
   if (isDismissed(def.storageKey, def.dismissal)) return false;
-  if (def.conditions?.some((c) => !c.check())) return false;
+  if (requireConditions && def.conditions?.some((condition) => !condition.check())) return false;
   return true;
 }
 
@@ -42,16 +57,18 @@ function isEligible(def: NudgeDefinition): boolean {
 // ---------------------------------------------------------------------------
 
 interface NudgeEngineProps {
-  scope: 'dashboard' | 'landing' | 'evaluation';
+  scope: NudgeScope;
 }
 
 /**
- * Two independent slots:
+ * Three independent slots:
  *  - **banner** — inline content (one at a time)
  *  - **overlay** — toasts and modals (one at a time)
+ *  - **coach mark** — a callout anchored to a live element (one at a time)
  *
- * A banner and an overlay can be visible simultaneously because they
- * occupy different visual layers.
+ * All three can be visible simultaneously because they occupy different
+ * visual layers: a banner sits in the document flow, an overlay in the
+ * bottom-right corner, and a coach mark next to the element it explains.
  */
 export function NudgeEngine({ scope }: NudgeEngineProps) {
   const scopeNudges = useMemo(() => NUDGE_REGISTRY.filter((n) => n.scope === scope), [scope]);
@@ -68,8 +85,10 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
     return initialBanner?.id ?? null;
   });
   const [activeOverlayId, setActiveOverlayId] = useState<string | null>(null);
+  const [activeCoachMarkId, setActiveCoachMarkId] = useState<string | null>(null);
   const bannerShownRef = useRef(false);
   const overlayShownRef = useRef(false);
+  const coachMarkShownRef = useRef(false);
 
   const triggerCountsRef = useRef<Record<string, number>>({});
   const eventDetailRef = useRef<Record<string, unknown> | null>(null);
@@ -77,6 +96,9 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
 
   const activeBanner = activeBannerId ? scopeNudges.find((n) => n.id === activeBannerId) : null;
   const activeOverlay = activeOverlayId ? scopeNudges.find((n) => n.id === activeOverlayId) : null;
+  const activeCoachMark = activeCoachMarkId
+    ? scopeNudges.find((n) => n.id === activeCoachMarkId)
+    : null;
 
   const showNudge = useCallback((def: NudgeDefinition) => {
     if (sessionDismissedRef.current.has(def.id)) return;
@@ -86,6 +108,10 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
       if (bannerShownRef.current) return;
       bannerShownRef.current = true;
       setActiveBannerId(def.id);
+    } else if (def.type === 'coach-mark') {
+      if (coachMarkShownRef.current) return;
+      coachMarkShownRef.current = true;
+      setActiveCoachMarkId(def.id);
     } else {
       if (overlayShownRef.current) return;
       overlayShownRef.current = true;
@@ -116,6 +142,31 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
     overlayShownRef.current = false;
   }, [activeOverlay]);
 
+  const dismissCoachMark = useCallback(() => {
+    if (!activeCoachMark) return;
+    trackNudgeEvent(activeCoachMark, 'dismissed');
+    markDismissed(activeCoachMark.storageKey, activeCoachMark.dismissal);
+    sessionDismissedRef.current.add(activeCoachMark.id);
+    setActiveCoachMarkId(null);
+    coachMarkShownRef.current = false;
+  }, [activeCoachMark]);
+
+  /**
+   * The user performed the interaction the coach mark was teaching (clicked a
+   * point). That is engagement, not rejection — record it as an action and,
+   * per `dismissesOnAction`, stop showing the tip.
+   */
+  const handleCoachMarkAction = useCallback(() => {
+    if (!activeCoachMark) return;
+    trackNudgeEvent(activeCoachMark, 'action');
+    activeCoachMark.content.action?.onClick(eventDetailRef.current ?? undefined);
+    if (!dismissesOnAction(activeCoachMark)) return;
+    markDismissed(activeCoachMark.storageKey, activeCoachMark.dismissal);
+    sessionDismissedRef.current.add(activeCoachMark.id);
+    setActiveCoachMarkId(null);
+    coachMarkShownRef.current = false;
+  }, [activeCoachMark]);
+
   const handleBannerAction = useCallback(() => {
     if (!activeBanner) return;
     trackNudgeEvent(activeBanner, 'action');
@@ -166,22 +217,43 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
     const sorted = [...scopeNudges].toSorted((a, b) => b.priority - a.priority);
 
     for (const def of sorted) {
-      if (!isEligible(def)) continue;
+      // Already occupying its slot — re-arming would keep re-running the
+      // (potentially expensive) eligibility check for no benefit.
+      if (def.id === activeBannerId || def.id === activeOverlayId) continue;
+      if (def.id === activeCoachMarkId) continue;
+      // Conditions are evaluated when a trigger fires. Wiring triggers only
+      // while a condition is already true strands route-dependent nudges when
+      // the persistent dashboard shell moves between tabs.
+      if (!isEligible(def, { requireAnchor: false, requireConditions: false })) continue;
       if (sessionDismissedRef.current.has(def.id)) continue;
 
       const triggers = Array.isArray(def.trigger) ? def.trigger : [def.trigger];
+      const triggerEvents = new Set(
+        triggers.flatMap((trigger) => (trigger.type === 'event' ? [trigger.event] : [])),
+      );
 
       for (const trigger of triggers) {
         const cleanup = setupTrigger(trigger, def, showNudge, triggerCountsRef, eventDetailRef);
         if (cleanup) cleanups.push(cleanup);
       }
+
+      // A condition may need re-evaluation on an event that is not itself a
+      // display trigger. Avoid a duplicate listener when the trigger already
+      // owns the same event and its configured threshold/delay.
+      for (const condition of def.conditions ?? []) {
+        const event = condition.listenEvent;
+        if (!event || triggerEvents.has(event)) continue;
+        const handler = () => showNudge(def);
+        window.addEventListener(event, handler);
+        cleanups.push(() => window.removeEventListener(event, handler));
+      }
     }
 
     return () => {
       for (const fn of cleanups) fn();
     };
     // eslint-disable-next-line react-hooks/exhaustive-deps
-  }, [activeBannerId, activeOverlayId, scope]);
+  }, [activeBannerId, activeOverlayId, activeCoachMarkId, scope]);
 
   // -------------------------------------------------------------------------
   // Permanent suppress event listener
@@ -197,6 +269,9 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
         if (def.type === 'banner' && activeBannerId === def.id) {
           setActiveBannerId(null);
           bannerShownRef.current = false;
+        } else if (def.type === 'coach-mark' && activeCoachMarkId === def.id) {
+          setActiveCoachMarkId(null);
+          coachMarkShownRef.current = false;
         } else if (activeOverlayId === def.id) {
           setActiveOverlayId(null);
           overlayShownRef.current = false;
@@ -223,7 +298,7 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
         window.removeEventListener(event, handler);
       }
     };
-  }, [activeBannerId, activeOverlayId, scopeNudges]);
+  }, [activeBannerId, activeOverlayId, activeCoachMarkId, scopeNudges]);
 
   // -------------------------------------------------------------------------
   // Render — banner and overlay can coexist
@@ -252,6 +327,13 @@ export function NudgeEngine({ scope }: NudgeEngineProps) {
           onAction={handleOverlayAction}
         />
       )}
+      {activeCoachMark?.content.anchor && (
+        <CoachMarkRenderer
+          def={activeCoachMark}
+          onDismiss={dismissCoachMark}
+          onAction={handleCoachMarkAction}
+        />
+      )}
     </>
   );
 }
@@ -433,9 +515,9 @@ function ModalRenderer({
               <Icon className={`size-5 ${content.iconClassName ?? ''}`} />
               {localized(locale, content.title, content.titleZh)}
               {content.badge && (
-                <span className="ml-1 inline-flex items-center rounded-full bg-brand px-2 py-0.5 text-[10px] font-bold uppercase tracking-wider text-primary-foreground shadow-sm">
+                <NewBadge data-new-badge="launch-modal" className="ml-1">
                   {localized(locale, content.badge, content.badgeZh)}
-                </span>
+                </NewBadge>
               )}
             </h2>
             <p id={`${idPrefix}-description`} className="text-sm text-muted-foreground">
@@ -487,6 +569,34 @@ function ModalRenderer({
   return dialog;
 }
 
+function CoachMarkRenderer({
+  def,
+  onDismiss,
+  onAction,
+}: {
+  def: NudgeDefinition;
+  onDismiss: () => void;
+  onAction: () => void;
+}) {
+  const locale = useLocale();
+  const { content } = def;
+  const Icon = content.icon;
+  // Guarded by the caller — `isEligible` also refuses anchorless coach marks.
+  const anchor = content.anchor!;
+
+  return (
+    <CoachMark
+      anchor={anchor}
+      icon={<Icon className={content.iconClassName} />}
+      title={localized(locale, content.title, content.titleZh)}
+      description={localized(locale, content.description, content.descriptionZh)}
+      onDismiss={onDismiss}
+      onAction={onAction}
+      testId={content.testId}
+    />
+  );
+}
+
 function BannerRenderer({
   def,
   onDismiss,
@@ -500,6 +610,9 @@ function BannerRenderer({
   const rs = RENDERER_STRINGS[locale];
   const { content } = def;
   const Icon = content.icon;
+  const href = content.href?.startsWith('/')
+    ? localePath(content.href, locale)
+    : (content.href ?? '#');
 
   const handleClick = (e: React.MouseEvent<HTMLAnchorElement>) => {
     if (e.metaKey || e.ctrlKey || e.shiftKey || e.altKey || e.button !== 0) return;
@@ -519,7 +632,7 @@ function BannerRenderer({
       data-initial-banner={def.renderOnInitialLoad ? '' : undefined}
     >
       <a
-        href={content.href ?? '#'}
+        href={href}
         onClick={handleClick}
         className="group relative flex items-center gap-3 overflow-hidden rounded-xl border border-brand/40 bg-gradient-to-r from-brand/10 via-brand/5 to-transparent px-4 py-3 transition-all duration-200 hover:border-brand/70 hover:shadow-lg hover:shadow-brand/10"
         data-testid={content.testId}
@@ -535,9 +648,9 @@ function BannerRenderer({
                 {localized(locale, content.title, content.titleZh)}
               </span>
               {content.badge && (
-                <span className="ml-2 inline-flex items-center gap-1.5 align-middle rounded-full bg-brand px-2 py-0.5 text-[10px] font-bold uppercase tracking-wider text-primary-foreground shadow-sm">
+                <NewBadge data-new-badge="launch-banner" className="ml-2 align-middle">
                   {localized(locale, content.badge, content.badgeZh)}
-                </span>
+                </NewBadge>
               )}
             </p>
             <p className="text-xs text-muted-foreground leading-snug truncate">
@@ -545,7 +658,7 @@ function BannerRenderer({
             </p>
           </div>
           <span className="hidden sm:inline-flex items-center gap-1 text-xs font-medium text-brand shrink-0 group-hover:translate-x-0.5 transition-transform duration-200">
-            {rs.explore}
+            {localized(locale, content.linkLabel ?? rs.explore, content.linkLabelZh)}
             <ArrowRight className="size-3.5" />
           </span>
         </div>
diff --git a/packages/app/src/components/official-preview-notice.test.tsx b/packages/app/src/components/official-preview-notice.test.tsx
new file mode 100644
index 000000000..f11dd4d22
--- /dev/null
+++ b/packages/app/src/components/official-preview-notice.test.tsx
@@ -0,0 +1,78 @@
+// @vitest-environment jsdom
+import React, { act, type ReactElement } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import {
+  includesJalapenoResult,
+  includesVeraRubinResult,
+  JalapenoOfficialPreviewNotice,
+  VeraRubinOfficialPreviewNotice,
+} from '@/components/official-preview-notice';
+
+const localeState = vi.hoisted(() => ({ pathname: '/inference' }));
+
+vi.mock('next/navigation', () => ({
+  usePathname: () => localeState.pathname,
+}));
+
+let container: HTMLDivElement;
+let root: Root;
+
+beforeEach(() => {
+  localeState.pathname = '/inference';
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+});
+
+function renderNotice(element: ReactElement, testId: string) {
+  act(() => root.render(element));
+  return container.querySelector(`[data-testid="${testId}"]`);
+}
+
+describe('official preview notices', () => {
+  it('recognizes each preview hardware family without cross-matching', () => {
+    expect(includesJalapenoResult(['h100_vllm', 'jalapeno_teacup'])).toBe(true);
+    expect(includesJalapenoResult(['jalapeno_fp4'])).toBe(true);
+    expect(includesJalapenoResult(['h100_vllm', 'vr200_rubin-july'])).toBe(false);
+
+    expect(includesVeraRubinResult(['h100_vllm', 'vr200_rubin-july'])).toBe(true);
+    expect(includesVeraRubinResult(['vr200_coreweave-vera-rubin'])).toBe(true);
+    expect(includesVeraRubinResult(['h100_vllm', 'jalapeno_teacup'])).toBe(false);
+  });
+
+  it('identifies Jalapeño and Vera Rubin data as official previews in English', () => {
+    const jalapenoNotice = renderNotice(
+      <JalapenoOfficialPreviewNotice />,
+      'jalapeno-official-preview-notice',
+    );
+    expect(jalapenoNotice?.getAttribute('role')).toBe('note');
+    expect(jalapenoNotice?.getAttribute('aria-label')).toBe('InferenceX Official Preview');
+    expect(jalapenoNotice?.textContent).toContain('Jalapeño results are an official preview');
+
+    const rubinNotice = renderNotice(
+      <VeraRubinOfficialPreviewNotice />,
+      'vera-rubin-official-preview-notice',
+    );
+    expect(rubinNotice?.getAttribute('role')).toBe('note');
+    expect(rubinNotice?.textContent).toContain('Vera Rubin (July) results are an official preview');
+  });
+
+  it('renders natural Simplified Chinese copy under /zh', () => {
+    localeState.pathname = '/zh/inference';
+    const notice = renderNotice(
+      <VeraRubinOfficialPreviewNotice />,
+      'vera-rubin-official-preview-notice',
+    );
+
+    expect(notice?.getAttribute('aria-label')).toBe('InferenceX 官方预览');
+    expect(notice?.textContent).toContain('Vera Rubin (July) 结果为官方预览');
+    expect(notice?.textContent).toContain('随着验证和发布工作的推进');
+  });
+});
diff --git a/packages/app/src/components/official-preview-notice.tsx b/packages/app/src/components/official-preview-notice.tsx
new file mode 100644
index 000000000..513432764
--- /dev/null
+++ b/packages/app/src/components/official-preview-notice.tsx
@@ -0,0 +1,107 @@
+'use client';
+
+import { Info } from 'lucide-react';
+
+import { hardwareKeyMatchesAnyBase } from '@/lib/constants';
+import { useLocale } from '@/lib/use-locale';
+
+export const JALAPENO_PREVIEW_STRINGS = {
+  en: {
+    title: 'InferenceX Official Preview',
+    description:
+      'Jalapeño results are an official preview and may change as validation and publication continue.',
+    chartDetail: 'Results may change as validation and publication continue.',
+  },
+  zh: {
+    title: 'InferenceX 官方预览',
+    description: 'Jalapeño 结果为官方预览；随着验证和发布工作的推进，数据可能会调整。',
+    chartDetail: '随着验证和发布工作的推进，结果可能会调整。',
+  },
+} as const;
+
+export const VERA_RUBIN_PREVIEW_STRINGS = {
+  en: {
+    title: 'InferenceX Official Preview',
+    description:
+      'Vera Rubin (July) results are an official preview and may change as validation and publication continue.',
+    chartDetail: 'Results may change as validation and publication continue.',
+  },
+  zh: {
+    title: 'InferenceX 官方预览',
+    description: 'Vera Rubin (July) 结果为官方预览；随着验证和发布工作的推进，数据可能会调整。',
+    chartDetail: '随着验证和发布工作的推进，结果可能会调整。',
+  },
+} as const;
+
+export const OFFICIAL_PREVIEW_SERIES = [
+  {
+    id: 'jalapeno-official-preview',
+    baseGpuKeys: ['jalapeno'],
+    strings: JALAPENO_PREVIEW_STRINGS,
+  },
+  {
+    id: 'vera-rubin-official-preview',
+    baseGpuKeys: ['vr200'],
+    strings: VERA_RUBIN_PREVIEW_STRINGS,
+  },
+] as const;
+
+function includesHardwareResult(
+  hardwareKeys: Iterable<string>,
+  baseGpuKeys: readonly string[],
+): boolean {
+  for (const hardwareKey of hardwareKeys) {
+    if (hardwareKeyMatchesAnyBase(hardwareKey, baseGpuKeys)) return true;
+  }
+  return false;
+}
+
+export function includesJalapenoResult(hardwareKeys: Iterable<string>): boolean {
+  return includesHardwareResult(hardwareKeys, OFFICIAL_PREVIEW_SERIES[0].baseGpuKeys);
+}
+
+export function includesVeraRubinResult(hardwareKeys: Iterable<string>): boolean {
+  return includesHardwareResult(hardwareKeys, OFFICIAL_PREVIEW_SERIES[1].baseGpuKeys);
+}
+
+interface OfficialPreviewNoticeProps {
+  strings: typeof JALAPENO_PREVIEW_STRINGS | typeof VERA_RUBIN_PREVIEW_STRINGS;
+  testId: string;
+}
+
+function OfficialPreviewNotice({ strings, testId }: OfficialPreviewNoticeProps) {
+  const t = strings[useLocale()];
+
+  return (
+    <div
+      role="note"
+      aria-label={t.title}
+      data-testid={testId}
+      className="mb-2 flex items-start gap-3 rounded-lg border border-brand/30 bg-brand/5 px-4 py-3"
+    >
+      <Info aria-hidden="true" className="mt-0.5 size-4 shrink-0 text-brand" />
+      <div className="min-w-0">
+        <p className="text-sm font-semibold text-foreground">{t.title}</p>
+        <p className="mt-0.5 text-xs leading-5 text-muted-foreground">{t.description}</p>
+      </div>
+    </div>
+  );
+}
+
+export function JalapenoOfficialPreviewNotice() {
+  return (
+    <OfficialPreviewNotice
+      strings={JALAPENO_PREVIEW_STRINGS}
+      testId="jalapeno-official-preview-notice"
+    />
+  );
+}
+
+export function VeraRubinOfficialPreviewNotice() {
+  return (
+    <OfficialPreviewNotice
+      strings={VERA_RUBIN_PREVIEW_STRINGS}
+      testId="vera-rubin-official-preview-notice"
+    />
+  );
+}
diff --git a/packages/app/src/components/overview/overview-history-detail-link.tsx b/packages/app/src/components/overview/overview-history-detail-link.tsx
index b357dfd24..76a458845 100644
--- a/packages/app/src/components/overview/overview-history-detail-link.tsx
+++ b/packages/app/src/components/overview/overview-history-detail-link.tsx
@@ -22,7 +22,7 @@ export function OverviewHistoryDetailLink({
       data-testid="overview-history-detail-link"
       href={href}
       aria-label={ariaLabel}
-      className="group mt-0.5 inline-flex min-h-11 items-center gap-1 rounded-sm text-[11px] font-medium text-foreground underline decoration-brand/50 underline-offset-4 transition-colors hover:decoration-brand focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50 motion-reduce:transition-none xl:min-h-8"
+      className="group mt-0.5 inline-flex min-h-11 items-center gap-1 rounded-sm text-2xs font-medium text-foreground underline decoration-brand/50 underline-offset-4 transition-colors hover:decoration-brand focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50 motion-reduce:transition-none xl:min-h-8"
       onClick={() => track('overview_history_detail_clicked', { model, hardware })}
     >
       {children}
diff --git a/packages/app/src/components/overview/overview-history-window-select.tsx b/packages/app/src/components/overview/overview-history-window-select.tsx
index 2765433c0..69d981af6 100644
--- a/packages/app/src/components/overview/overview-history-window-select.tsx
+++ b/packages/app/src/components/overview/overview-history-window-select.tsx
@@ -11,6 +11,7 @@ import { track } from '@/lib/analytics';
 import type { OverviewHistoryWindowKey } from '@/lib/overview-data';
 
 import { useOverviewComparisonMode, useOverviewNavigation } from './overview-navigation';
+import { useIsPresenting } from './overview-presentation';
 
 interface WindowOption {
   href: string;
@@ -34,6 +35,10 @@ export function OverviewHistoryWindowSelect({
   // back to hardware mode carries no window, so keep the committed one.
   const pendingMode = useOverviewComparisonMode();
   const value = pendingMode === 'hardware' ? committedValue : pendingMode;
+  // Same reason as the reference select: `document.body` sits outside both the
+  // fullscreen element and the CSS `zoom`, so a portalled menu would be
+  // invisible or unscaled while presenting.
+  const presenting = useIsPresenting();
 
   return (
     <Select
@@ -53,7 +58,7 @@ export function OverviewHistoryWindowSelect({
       >
         <SelectValue>{options.find((option) => option.value === value)?.label}</SelectValue>
       </SelectTrigger>
-      <SelectContent align="center">
+      <SelectContent align="center" portalled={!presenting}>
         {options.map((option) => (
           <SelectItem key={option.value} value={option.value} data-overview-window={option.value}>
             {option.label}
diff --git a/packages/app/src/components/overview/overview-nav-link.tsx b/packages/app/src/components/overview/overview-nav-link.tsx
index d97434cb6..6b9b5dbc4 100644
--- a/packages/app/src/components/overview/overview-nav-link.tsx
+++ b/packages/app/src/components/overview/overview-nav-link.tsx
@@ -12,7 +12,7 @@ import { useOverviewNavigation } from './overview-navigation';
 const PREFETCH_DWELL_MS = 120;
 
 interface OverviewNavAnalytics {
-  control: 'comparison' | 'engine' | 'models' | 'tier';
+  control: 'comparison' | 'engine' | 'hwrows' | 'models' | 'rows' | 'tier';
   value: string;
 }
 
diff --git a/packages/app/src/components/overview/overview-navigation.test.tsx b/packages/app/src/components/overview/overview-navigation.test.tsx
index 015fcd252..746057542 100644
--- a/packages/app/src/components/overview/overview-navigation.test.tsx
+++ b/packages/app/src/components/overview/overview-navigation.test.tsx
@@ -18,6 +18,7 @@ vi.mock('next/navigation', () => ({
 import {
   OverviewNavigationProvider,
   useOverviewData,
+  useOverviewNavigationError,
   useOverviewNavigation,
   useOverviewReference,
 } from './overview-navigation';
@@ -33,6 +34,8 @@ let root: Root;
 let selectTier: (() => void) | undefined;
 let selectEngine: (() => void) | undefined;
 let selectReference: (() => void) | undefined;
+let selectRowScope: (() => void) | undefined;
+let selectHardwareRowScope: (() => void) | undefined;
 let prefetchTier: (() => void) | undefined;
 
 function pageData(tier: OverviewTier): OverviewPageData {
@@ -43,23 +46,31 @@ function pageData(tier: OverviewTier): OverviewPageData {
     comparisonMode: 'hardware',
     referenceHardware: 'b200',
     modelScope: 'default',
+    rowScope: 'all',
+    hardwareRowScope: 'all',
+    unchangedRowCount: 0,
+    emptyRowCount: 0,
     historicalWindow: null,
   };
 }
 
 function Probe() {
   const navigation = useOverviewNavigation();
+  const navigationError = useOverviewNavigationError();
   const data = useOverviewData();
   const reference = useOverviewReference();
   selectTier = () => navigation.push('/overview?tier=75', ['tier']);
   selectEngine = () => navigation.push('/overview?engine=all', ['engine']);
   selectReference = () => navigation.push('/overview?ref=b300', ['ref']);
+  selectRowScope = () => navigation.push('/overview?compare=30d&rows=changed', ['rows']);
+  selectHardwareRowScope = () => navigation.push('/overview?hwrows=priced', ['hwrows']);
   prefetchTier = () => navigation.prefetch('/overview?tier=75', ['tier']);
   return (
     <>
       <output data-testid="tier">{data.tier}</output>
       <output data-testid="reference">{reference}</output>
       <output data-testid="pending">{navigation.isPending ? 'pending' : 'settled'}</output>
+      <output data-testid="error">{navigationError ? 'error' : 'ok'}</output>
     </>
   );
 }
@@ -105,6 +116,8 @@ afterEach(() => {
   selectTier = undefined;
   selectEngine = undefined;
   selectReference = undefined;
+  selectRowScope = undefined;
+  selectHardwareRowScope = undefined;
   prefetchTier = undefined;
   vi.unstubAllGlobals();
 });
@@ -153,7 +166,7 @@ describe('OverviewNavigationProvider', () => {
     // The shared afterEach unmounts again; react-dom tolerates the repeat.
   });
 
-  it('replaces rather than pushes when the overview request fails', async () => {
+  it('keeps a failed selector entry without falling back to router navigation', async () => {
     const { settlers } = deferredFetch();
     const searchEvents: string[] = [];
     const onSearchChange = (event: Event) => {
@@ -173,12 +186,14 @@ describe('OverviewNavigationProvider', () => {
 
     window.removeEventListener(CLIENT_SEARCH_CHANGE_EVENT, onSearchChange);
 
-    expect(routerStub.replace).toHaveBeenCalledTimes(1);
-    expect(routerStub.replace).toHaveBeenCalledWith('/overview?tier=75', { scroll: false });
+    expect(routerStub.replace).not.toHaveBeenCalled();
     expect(routerStub.push).not.toHaveBeenCalled();
     // The failed selection stays in the address bar so a retry click still works.
     expect(window.location.search).toBe('?tier=75');
     expect(searchEvents).toEqual(['?tier=75']);
+    expect(readProbe('tier')).toBe('50');
+    expect(readProbe('pending')).toBe('settled');
+    expect(readProbe('error')).toBe('error');
   });
 
   it('pushes one history entry when the same selection repeats while pending', () => {
@@ -196,7 +211,7 @@ describe('OverviewNavigationProvider', () => {
     expect(fetch).toHaveBeenCalledTimes(1);
   });
 
-  it('reloads on a failed popstate commit', async () => {
+  it('retains the last data and history when a popstate request fails, then clears the error', async () => {
     renderProvider(pageData(50), '/overview');
 
     const { settlers } = deferredFetch();
@@ -212,6 +227,11 @@ describe('OverviewNavigationProvider', () => {
       origin: 'http://localhost',
       reload,
     });
+    const replaceState = vi.spyOn(History.prototype, 'replaceState');
+    const pushState = vi.spyOn(History.prototype, 'pushState');
+    replaceState.mockClear();
+    pushState.mockClear();
+    const historyLength = window.history.length;
 
     await act(async () => {
       window.dispatchEvent(new PopStateEvent('popstate'));
@@ -220,7 +240,35 @@ describe('OverviewNavigationProvider', () => {
       await Promise.resolve();
     });
 
-    expect(reload).toHaveBeenCalledTimes(1);
+    expect(reload).not.toHaveBeenCalled();
+    expect(routerStub.push).not.toHaveBeenCalled();
+    expect(routerStub.replace).not.toHaveBeenCalled();
+    expect(pushState).not.toHaveBeenCalled();
+    expect(replaceState).not.toHaveBeenCalled();
+    expect(window.history.length).toBe(historyLength);
+    expect(window.location.search).toBe('?tier=75');
+    expect(readProbe('tier')).toBe('50');
+    expect(readProbe('pending')).toBe('settled');
+    expect(readProbe('error')).toBe('error');
+
+    window.history.replaceState({}, '', '/overview?tier=100');
+    await act(async () => {
+      window.dispatchEvent(new PopStateEvent('popstate'));
+      await Promise.resolve();
+    });
+
+    expect(readProbe('pending')).toBe('pending');
+    expect(readProbe('error')).toBe('ok');
+
+    await act(async () => {
+      settlers[1]?.resolve(Response.json(pageData(100)));
+      await Promise.resolve();
+    });
+
+    expect(readProbe('tier')).toBe('100');
+    expect(readProbe('error')).toBe('ok');
+    replaceState.mockRestore();
+    pushState.mockRestore();
   });
 
   it('fetches on a cache-miss popstate and ignores non-overview routes', async () => {
@@ -377,13 +425,37 @@ describe('OverviewNavigationProvider', () => {
     });
 
     expect(settlers).toHaveLength(1);
-    expect(routerStub.replace).toHaveBeenCalledWith('/overview?tier=75&ref=b300', {
-      scroll: false,
-    });
+    expect(routerStub.replace).not.toHaveBeenCalled();
     expect(readProbe('tier')).toBe('50');
     expect(readProbe('reference')).toBe('b300');
   });
 
+  // Both row scopes narrow the rows the server sends, so neither may collapse
+  // into the unfiltered payload's cache key the way `ref` deliberately does.
+  it('requests a narrowed matrix for each row scope instead of reusing the full one', () => {
+    deferredFetch();
+
+    renderProvider(pageData(50), '/overview?compare=30d');
+    act(() => selectRowScope?.());
+
+    expect(fetch).toHaveBeenCalledWith('/api/v1/overview?compare=30d&rows=changed', {
+      headers: { Accept: 'application/json' },
+    });
+    expect(readProbe('pending')).toBe('pending');
+  });
+
+  it('requests a narrowed matrix for the hardware row scope', () => {
+    deferredFetch();
+
+    renderProvider(pageData(50), '/overview');
+    act(() => selectHardwareRowScope?.());
+
+    expect(fetch).toHaveBeenCalledWith('/api/v1/overview?hwrows=priced', {
+      headers: { Accept: 'application/json' },
+    });
+    expect(readProbe('pending')).toBe('pending');
+  });
+
   it('preserves unknown params and the fragment on the first selection', () => {
     deferredFetch();
 
diff --git a/packages/app/src/components/overview/overview-navigation.tsx b/packages/app/src/components/overview/overview-navigation.tsx
index 659b6703d..d3dc933e8 100644
--- a/packages/app/src/components/overview/overview-navigation.tsx
+++ b/packages/app/src/components/overview/overview-navigation.tsx
@@ -1,6 +1,5 @@
 'use client';
 
-import { useRouter } from 'next/navigation';
 import {
   createContext,
   type MutableRefObject,
@@ -22,24 +21,36 @@ import {
   type OverviewReferenceHardware,
   resolveOverviewComparisonMode,
   resolveOverviewEngineScope,
+  resolveOverviewHardwareRowScope,
   resolveOverviewModelScope,
   resolveOverviewReferenceHardware,
+  resolveOverviewRowScope,
   resolveOverviewTier,
 } from '@/lib/overview-data';
 import {
   mergeOverviewControlHref,
   OVERVIEW_CLIENT_ONLY_KEYS,
   OVERVIEW_SEARCH_ORDER,
+  type OverviewClientOnlySearchKey,
   overviewHref,
   type OverviewSearchKey,
 } from '@/lib/overview-links';
 
+const OVERVIEW_SERVER_SEARCH_KEYS = OVERVIEW_SEARCH_ORDER.filter(
+  (key): key is Exclude<OverviewSearchKey, OverviewClientOnlySearchKey> =>
+    !OVERVIEW_CLIENT_ONLY_KEYS.includes(key as OverviewClientOnlySearchKey),
+);
+
 /**
  * Cache and request identity. The payload depends only on the server-resolved
- * params, minus `ref`, which the client derives. Equivalent URLs — explicit
+ * params, minus `ref` and `present`, which the client derives. Equivalent URLs — explicit
  * defaults, reordered params, campaign tags, a fragment — collapse to one key,
  * so a `ref` change is a guaranteed hit and the CDN sees one entry per data
  * state instead of one per link anyone has ever shared.
+ *
+ * Every param the server reads has to appear here. Both row scopes do: each
+ * narrows the rows the response carries, and the dormant one still reaches the
+ * payload so a tab switch can restore the other mode's answer.
  */
 function overviewDataKey(href: string): string {
   const url = new URL(href, 'https://inferencex.local');
@@ -51,10 +62,12 @@ function overviewDataKey(href: string): string {
     resolveOverviewComparisonMode(params.get('compare') ?? undefined),
     OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
     resolveOverviewModelScope(params.get('models') ?? undefined),
+    resolveOverviewRowScope(params.get('rows') ?? undefined),
+    resolveOverviewHardwareRowScope(params.get('hwrows') ?? undefined),
   );
 }
 
-export type OverviewNavControl = 'comparison' | 'engine' | 'models' | 'tier';
+export type OverviewNavControl = 'comparison' | 'engine' | 'hwrows' | 'models' | 'rows' | 'tier';
 
 interface OverviewNavigationValue {
   isPending: boolean;
@@ -62,19 +75,22 @@ interface OverviewNavigationValue {
    *  that replaces it can take the focus its predecessor lost on unmount. */
   focusIntent: MutableRefObject<OverviewNavControl | null>;
   prefetch: (targetHref: string, keys: readonly OverviewSearchKey[]) => void;
+  replaceClientState: (targetHref: string, keys: readonly OverviewClientOnlySearchKey[]) => void;
   resolve: (targetHref: string, keys: readonly OverviewSearchKey[]) => string;
   push: (targetHref: string, keys: readonly OverviewSearchKey[]) => void;
 }
 
 /**
- * Three contexts, not one. A selector click moves the pending href immediately
+ * Split contexts, not one. A selector click moves the pending href immediately
  * and the payload only when the request settles, so a single value would push
  * two full matrix renders per uncached selection. Splitting them means the
- * controls can re-render on their own while the matrix waits for real data.
+ * controls and failure notice can re-render on their own while the matrix waits
+ * for real data.
  */
 const OverviewDataContext = createContext<OverviewPageData | null>(null);
 const OverviewReferenceContext = createContext<OverviewReferenceHardware | null>(null);
 const OverviewComparisonContext = createContext<OverviewComparisonMode | null>(null);
+const OverviewNavigationErrorContext = createContext(false);
 const OverviewNavigationContext = createContext<OverviewNavigationValue | null>(null);
 
 export function OverviewNavigationProvider({
@@ -86,8 +102,8 @@ export function OverviewNavigationProvider({
   initialHref: string;
   children: ReactNode;
 }) {
-  const router = useRouter();
   const [data, setData] = useState(initialData);
+  const [navigationError, setNavigationError] = useState(false);
   const [pendingHref, setPendingHref] = useState(initialHref);
   const [committedHref, setCommittedHref] = useState(initialHref);
   const pendingHrefRef = useRef(initialHref);
@@ -128,6 +144,7 @@ export function OverviewNavigationProvider({
       const navigationId = ++navigationIdRef.current;
       pendingHrefRef.current = href;
       setPendingHref(href);
+      setNavigationError(false);
       if (updateHistory) {
         // Deliberately the pristine prototype method: `window.history.pushState`
         // is patched by Next to dispatch a router action, and that per-click
@@ -149,17 +166,35 @@ export function OverviewNavigationProvider({
       void load(href)
         .then((nextData) => {
           if (navigationId !== navigationIdRef.current) return;
-          committedHrefRef.current = href;
-          setCommittedHref(href);
+          // A client-only control can change while this data request is in
+          // flight. Commit the loaded server state without rolling that newer
+          // local state back out of the address bar or navigation context.
+          const settledHref = mergeOverviewControlHref(
+            href,
+            pendingHrefRef.current,
+            OVERVIEW_CLIENT_ONLY_KEYS,
+          );
+          committedHrefRef.current = settledHref;
+          pendingHrefRef.current = settledHref;
+          setCommittedHref(settledHref);
+          setPendingHref(settledHref);
           if (updateHistory) {
-            History.prototype.replaceState.call(window.history, window.history.state, '', href);
+            History.prototype.replaceState.call(
+              window.history,
+              window.history.state,
+              '',
+              settledHref,
+            );
           }
           setData(nextData);
+          setNavigationError(false);
           // The pushState above is invisible to Next's router, so the app-wide
           // pageview tracker never fires here. Emit once per committed state —
           // in the success branch so Back/Forward is covered too, and so the
-          // failure path's `router.replace` is not double-counted.
-          track('$pageview', { $current_url: new URL(href, window.location.origin).href });
+          // recoverable failure path never records a pageview.
+          track('$pageview', {
+            $current_url: new URL(settledHref, window.location.origin).href,
+          });
         })
         .catch(() => {
           if (navigationId !== navigationIdRef.current) return;
@@ -168,26 +203,18 @@ export function OverviewNavigationProvider({
           // request was in flight needs no payload of its own, and the address
           // bar keeps it either way, so dropping it here would repaint against
           // a column the URL still claims.
+          const clientState = pendingHrefRef.current;
           const rolledBack = mergeOverviewControlHref(
             committedHrefRef.current,
-            href,
+            clientState,
             OVERVIEW_CLIENT_ONLY_KEYS,
           );
           pendingHrefRef.current = rolledBack;
           setPendingHref(rolledBack);
-          if (updateHistory) {
-            // The entry pushed above already carries `href`, so `replace`
-            // rewrites it in place instead of stacking a duplicate the user
-            // has to press Back through twice. Rewinding to the committed
-            // href first would also strand the header language toggle, which
-            // only ever hears about search changes through the event above.
-            router.replace(href, { scroll: false });
-          } else {
-            window.location.reload();
-          }
+          setNavigationError(true);
         });
     },
-    [load, router],
+    [load],
   );
 
   useEffect(() => {
@@ -196,13 +223,14 @@ export function OverviewNavigationProvider({
     // (utm_*, gclid, a fragment) are adopted from the address bar, so a stale
     // location can never key the cache to the wrong payload.
     const actual = `${window.location.pathname}${window.location.search}${window.location.hash}`;
-    const href = mergeOverviewControlHref(actual, initialHref, OVERVIEW_SEARCH_ORDER);
+    const href = mergeOverviewControlHref(actual, initialHref, OVERVIEW_SERVER_SEARCH_KEYS);
     dataCacheRef.current.set(overviewDataKey(href), initialData);
     committedHrefRef.current = href;
     setCommittedHref(href);
     pendingHrefRef.current = href;
     setPendingHref(href);
     setData(initialData);
+    setNavigationError(false);
   }, [initialData, initialHref]);
 
   useEffect(() => {
@@ -236,6 +264,39 @@ export function OverviewNavigationProvider({
     [pendingHref],
   );
 
+  const replaceClientState = useCallback(
+    (targetHref: string, keys: readonly OverviewClientOnlySearchKey[]) => {
+      // A child effect can request client-only cleanup before this provider's
+      // mount effect adopts the loaded URL. Rebase both snapshots on the live
+      // address first so campaign params and the fragment survive that race,
+      // while server-owned keys still come from their pending/committed state.
+      const actualHref = `${window.location.pathname}${window.location.search}${window.location.hash}`;
+      const pendingBase = mergeOverviewControlHref(
+        actualHref,
+        pendingHrefRef.current,
+        OVERVIEW_SERVER_SEARCH_KEYS,
+      );
+      const committedBase = mergeOverviewControlHref(
+        actualHref,
+        committedHrefRef.current,
+        OVERVIEW_SERVER_SEARCH_KEYS,
+      );
+      const pending = mergeOverviewControlHref(pendingBase, targetHref, keys);
+      const committed = mergeOverviewControlHref(committedBase, targetHref, keys);
+      pendingHrefRef.current = pending;
+      committedHrefRef.current = committed;
+      setPendingHref(pending);
+      setCommittedHref(committed);
+
+      const currentHref = `${window.location.pathname}${window.location.search}${window.location.hash}`;
+      if (pending !== currentHref) {
+        History.prototype.replaceState.call(window.history, window.history.state, '', pending);
+        notifyClientSearchChange(pending);
+      }
+    },
+    [],
+  );
+
   /** The URL already shows the new selection while the matrix still shows the
    *  old numbers. Consumers use this to say so. Compared on the data key rather
    *  than the href: a `ref` change repaints from the payload already in hand, so
@@ -277,6 +338,7 @@ export function OverviewNavigationProvider({
       isPending,
       focusIntent: focusIntentRef,
       resolve,
+      replaceClientState,
       prefetch: (targetHref, keys) => {
         const href = mergeOverviewControlHref(pendingHrefRef.current, targetHref, keys);
         void load(href).catch(() => undefined);
@@ -286,16 +348,18 @@ export function OverviewNavigationProvider({
         commit(href, true);
       },
     }),
-    [commit, isPending, load, resolve],
+    [commit, isPending, load, replaceClientState, resolve],
   );
 
   return (
     <OverviewDataContext.Provider value={data}>
       <OverviewReferenceContext.Provider value={referenceHardware}>
         <OverviewComparisonContext.Provider value={comparisonMode}>
-          <OverviewNavigationContext.Provider value={value}>
-            {children}
-          </OverviewNavigationContext.Provider>
+          <OverviewNavigationErrorContext.Provider value={navigationError}>
+            <OverviewNavigationContext.Provider value={value}>
+              {children}
+            </OverviewNavigationContext.Provider>
+          </OverviewNavigationErrorContext.Provider>
         </OverviewComparisonContext.Provider>
       </OverviewReferenceContext.Provider>
     </OverviewDataContext.Provider>
@@ -308,6 +372,10 @@ export function useOverviewNavigation(): OverviewNavigationValue {
   return value;
 }
 
+export function useOverviewNavigationError(): boolean {
+  return useContext(OverviewNavigationErrorContext);
+}
+
 export function useOverviewData(): OverviewPageData {
   const value = useContext(OverviewDataContext);
   if (value === null) throw new Error('Overview controls require OverviewNavigationProvider');
diff --git a/packages/app/src/components/overview/overview-page.test.tsx b/packages/app/src/components/overview/overview-page.test.tsx
new file mode 100644
index 000000000..27a9a871b
--- /dev/null
+++ b/packages/app/src/components/overview/overview-page.test.tsx
@@ -0,0 +1,99 @@
+// @vitest-environment jsdom
+
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import type { OverviewPageData } from '@/lib/overview-data';
+
+import { OverviewPageContent } from './overview-page';
+import type { OverviewLocale } from './overview-strings';
+
+declare global {
+  var IS_REACT_ACT_ENVIRONMENT: boolean;
+}
+
+globalThis.IS_REACT_ACT_ENVIRONMENT = true;
+
+const EMPTY_DATA: OverviewPageData = {
+  models: [],
+  tier: 50,
+  engineScope: 'community',
+  comparisonMode: 'hardware',
+  referenceHardware: 'b200',
+  modelScope: 'default',
+  rowScope: 'all',
+  hardwareRowScope: 'all',
+  unchangedRowCount: 0,
+  emptyRowCount: 0,
+  historicalWindow: null,
+};
+
+let container: HTMLDivElement;
+let root: Root;
+
+function renderPage(locale: OverviewLocale) {
+  act(() => {
+    root.render(<OverviewPageContent data={EMPTY_DATA} locale={locale} />);
+  });
+}
+
+beforeEach(() => {
+  window.history.replaceState({}, '', '/overview');
+  Object.defineProperty(document, 'fullscreenEnabled', { configurable: true, value: false });
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  vi.unstubAllGlobals();
+});
+
+describe('OverviewPageContent failure states', () => {
+  it.each([
+    ['en', 'No overview results match this selection.'],
+    ['zh', '没有符合当前筛选条件的总览结果。'],
+  ] as const)('renders the localized empty state for %s', (locale, message) => {
+    renderPage(locale);
+
+    const empty = container.querySelector('[data-testid="overview-empty-state"]');
+    expect(empty?.textContent).toBe(message);
+    expect(container.querySelector('[data-testid="overview-desktop-matrix"]')).toBeNull();
+    expect(container.querySelector('[data-testid="overview-mobile-list"]')).toBeNull();
+  });
+
+  it.each([
+    ['en', 'Could not load the selected comparison. Showing the last successfully loaded data.'],
+    ['zh', '无法加载所选对比，当前显示的是上次成功加载的数据。'],
+  ] as const)(
+    'renders a localized recoverable navigation alert for %s',
+    async (locale, message) => {
+      vi.stubGlobal(
+        'fetch',
+        vi.fn(() => Promise.reject(new Error('network down'))),
+      );
+      window.history.replaceState({}, '', locale === 'zh' ? '/zh/overview' : '/overview');
+      renderPage(locale);
+
+      // The tier control is a slider, not a link, so the failed navigation is
+      // driven through the engine scope link instead.
+      const scopeLink = container.querySelector<HTMLAnchorElement>(
+        'a[data-overview-engine-scope="all"]',
+      );
+      expect(scopeLink).not.toBeNull();
+
+      await act(async () => {
+        scopeLink?.click();
+        await Promise.resolve();
+        await Promise.resolve();
+      });
+
+      const alert = container.querySelector('[role="alert"]');
+      expect(alert?.textContent).toBe(message);
+      expect(container.querySelector('[aria-busy="true"]')).toBeNull();
+    },
+  );
+});
diff --git a/packages/app/src/components/overview/overview-page.tsx b/packages/app/src/components/overview/overview-page.tsx
index a75830680..1667eae36 100644
--- a/packages/app/src/components/overview/overview-page.tsx
+++ b/packages/app/src/components/overview/overview-page.tsx
@@ -2,9 +2,11 @@
 
 import type { ReactNode } from 'react';
 
+import { TCO_SOURCE_URL } from '@semianalysisai/inferencex-constants';
+
 import { Card } from '@/components/ui/card';
 import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
-import type { OverviewPageData } from '@/lib/overview-data';
+import { OVERVIEW_HISTORY_WINDOW_DAYS, type OverviewPageData } from '@/lib/overview-data';
 import { overviewHref } from '@/lib/overview-links';
 
 import {
@@ -12,24 +14,29 @@ import {
   MobileOverviewList,
   OverviewComparisonSwitcher,
   OverviewEngineScopeSwitcher,
+  OverviewHardwareRowScopeToggle,
   OverviewMethodology,
   OverviewModelScopeToggle,
+  OverviewRowScopeToggle,
   OverviewTierSwitcher,
   overviewFormatters,
-  OVERVIEW_STRINGS,
-  type OverviewLocale,
 } from './overview-scorecard';
+import { OVERVIEW_STRINGS, type OverviewLocale } from './overview-strings';
 import {
   OverviewNavigationProvider,
   useOverviewData,
   useOverviewNavigation,
+  useOverviewNavigationError,
   useOverviewReference,
 } from './overview-navigation';
+import {
+  OverviewPresentationProvider,
+  OverviewPresentationSurface,
+  OverviewPresentToggle,
+  useOverviewPresentation,
+} from './overview-presentation';
 import { useWideViewport } from './use-wide-viewport';
 
-/** The SemiAnalysis AI Cloud TCO model behind `HW_REGISTRY.costh`. */
-const OVERVIEW_SOURCE_HREF = 'https://semianalysis.com/ai-cloud-tco-model/';
-
 interface OverviewPageProps {
   data: OverviewPageData;
   locale: OverviewLocale;
@@ -46,12 +53,17 @@ export function OverviewPageContent({ data, locale }: OverviewPageProps) {
         data.comparisonMode,
         data.referenceHardware,
         data.modelScope,
+        data.rowScope,
+        data.hardwareRowScope,
       )}
     >
       {/* Passed as `children`, never rendered inside the provider's own JSX:
           that keeps this element's identity stable so a pending-state change
-          re-renders the provider without re-rendering the whole matrix. */}
-      <OverviewPageBody locale={locale} />
+          re-renders the provider without re-rendering the whole matrix. The
+          presentation provider is in the same slot for the same reason. */}
+      <OverviewPresentationProvider locale={locale}>
+        <OverviewPageBody locale={locale} />
+      </OverviewPresentationProvider>
     </OverviewNavigationProvider>
   );
 }
@@ -68,6 +80,20 @@ function OverviewPendingStatus({ label }: { label: string }) {
   );
 }
 
+function OverviewNavigationErrorStatus({ label }: { label: string }) {
+  const hasError = useOverviewNavigationError();
+  if (!hasError) return null;
+  return (
+    <p
+      role="alert"
+      data-testid="overview-navigation-error"
+      className="rounded-lg border border-destructive/40 bg-destructive/10 px-4 py-3 text-sm text-foreground"
+    >
+      {label}
+    </p>
+  );
+}
+
 function OverviewMatrixCard({ children }: { children: ReactNode }) {
   const { isPending } = useOverviewNavigation();
   return (
@@ -87,134 +113,332 @@ function OverviewPageBody({ locale }: { locale: OverviewLocale }) {
   // Not `data.referenceHardware`: the reference follows the URL directly, so a
   // cached payload built for another reference still renders the right column.
   const referenceHardware = useOverviewReference();
-  // Both surfaces used to render on every width and hide one with CSS, so every
-  // selection built the matrix twice. The Tailwind classes stay — they carry
-  // SSR and the pre-hydration frame — and this only drops the unused one after.
-  const wide = useWideViewport();
+  const { presenting } = useOverviewPresentation();
   const strings = OVERVIEW_STRINGS[locale];
-  const formatters = overviewFormatters(locale);
 
   return (
     <section data-testid="overview-page" className="flex flex-col gap-4">
       <OverviewPendingStatus label={strings.loadingStatus} />
-      <Card>
-        <header>
-          {/* Two rows at every width: the title, then the metric it is
+      <OverviewNavigationErrorStatus label={strings.navigationError} />
+      {/* Held in a stable child slot: swapping the header out for the surface
+          would remount the surface and drop the browser out of fullscreen. The
+          browser already stops painting it, so this only keeps the hidden
+          duplicates of the SLO and engine controls out of the accessibility
+          tree while the surface renders its own. */}
+      {presenting ? null : (
+        /* Tighter than the default card rhythm: every pixel spent here pushes
+           the matrix further below the fold on a laptop viewport. */
+        <Card className="md:py-5">
+          <header>
+            {/* Two rows at every width: the title, then the metric it is
               measured in and where that measure comes from. */}
-          <div className="flex flex-col gap-y-1">
-            <h1 className="text-lg font-semibold">{strings.title}</h1>
-            {/* Metric, direction and provenance read as one line: the numbers
+            <div className="flex flex-col gap-y-1">
+              <h1 className="text-lg font-semibold">{strings.title}</h1>
+              {/* Metric, direction and provenance read as one line: the numbers
                 and the model they are priced from belong together. */}
-            <p
-              data-testid="overview-scope"
-              aria-label={strings.scopeAria}
-              className="inline-flex flex-wrap items-baseline gap-x-2 leading-snug"
-            >
-              <span
-                data-testid="overview-scope-metric"
-                className="text-base font-semibold text-foreground"
-              >
-                {strings.scopeMetric}
-              </span>{' '}
-              <span aria-hidden="true" className="text-sm text-muted-foreground">
-                ·
-              </span>{' '}
-              <span
-                data-testid="overview-scope-direction"
-                className="text-sm font-normal text-muted-foreground"
+              <p
+                data-testid="overview-scope"
+                aria-label={strings.scopeAria}
+                className="inline-flex flex-wrap items-baseline gap-x-2 leading-snug"
               >
-                {strings.scopeDirection}
-              </span>{' '}
-              <span aria-hidden="true" className="text-sm text-muted-foreground">
-                ·
-              </span>{' '}
-              <span className="text-sm font-normal text-muted-foreground">
-                {strings.sourcePrefix}
-                <a
-                  data-testid="overview-source-link"
-                  href={OVERVIEW_SOURCE_HREF}
-                  target="_blank"
-                  rel="noopener noreferrer"
-                  className="group rounded-sm underline decoration-dotted underline-offset-4 hover:decoration-solid focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50"
+                <span
+                  data-testid="overview-scope-metric"
+                  className="text-base font-semibold text-foreground"
                 >
-                  {strings.sourceLinkText}
-                  <ExternalLinkIcon />
-                </a>
-              </span>
-            </p>
-          </div>
-          <div className="mt-3 flex flex-col gap-2 lg:flex-row lg:flex-wrap lg:items-center lg:gap-x-6">
-            <OverviewTierSwitcher
-              tier={data.tier}
-              engineScope={data.engineScope}
-              comparisonMode={data.comparisonMode}
-              referenceHardware={referenceHardware}
-              modelScope={data.modelScope}
-              locale={locale}
-              strings={strings}
-            />
-            <OverviewEngineScopeSwitcher
-              engineScope={data.engineScope}
-              tier={data.tier}
-              comparisonMode={data.comparisonMode}
-              referenceHardware={referenceHardware}
-              modelScope={data.modelScope}
-              locale={locale}
-              strings={strings}
-            />
-          </div>
-        </header>
-      </Card>
-
-      <OverviewComparisonSwitcher
-        comparisonMode={data.comparisonMode}
-        engineScope={data.engineScope}
-        tier={data.tier}
-        referenceHardware={referenceHardware}
-        modelScope={data.modelScope}
-        locale={locale}
-        strings={strings}
-      />
+                  {strings.scopeMetric}
+                </span>{' '}
+                <span aria-hidden="true" className="text-sm text-muted-foreground">
+                  ·
+                </span>{' '}
+                <span
+                  data-testid="overview-scope-direction"
+                  className="text-sm font-normal text-muted-foreground"
+                >
+                  {strings.scopeDirection}
+                </span>{' '}
+                <span aria-hidden="true" className="text-sm text-muted-foreground">
+                  ·
+                </span>{' '}
+                <span className="text-sm font-normal text-muted-foreground">
+                  {strings.sourcePrefix}
+                  <a
+                    data-testid="overview-source-link"
+                    href={TCO_SOURCE_URL}
+                    target="_blank"
+                    rel="noopener noreferrer"
+                    className="group rounded-sm underline decoration-dotted underline-offset-4 hover:decoration-solid focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50"
+                  >
+                    {strings.sourceLinkText}
+                    <ExternalLinkIcon />
+                  </a>
+                </span>
+              </p>
+            </div>
+            <div className="mt-3 flex flex-col gap-2 lg:flex-row lg:flex-wrap lg:items-center lg:gap-x-6">
+              <OverviewTierSwitcher
+                tier={data.tier}
+                engineScope={data.engineScope}
+                comparisonMode={data.comparisonMode}
+                referenceHardware={referenceHardware}
+                modelScope={data.modelScope}
+                rowScope={data.rowScope}
+                hardwareRowScope={data.hardwareRowScope}
+                locale={locale}
+                strings={strings}
+              />
+              <OverviewEngineScopeSwitcher
+                engineScope={data.engineScope}
+                tier={data.tier}
+                comparisonMode={data.comparisonMode}
+                referenceHardware={referenceHardware}
+                modelScope={data.modelScope}
+                rowScope={data.rowScope}
+                hardwareRowScope={data.hardwareRowScope}
+                locale={locale}
+                strings={strings}
+              />
+            </div>
+          </header>
+        </Card>
+      )}
 
-      {/* Official-only summary; uploaded runs remain in the linked dashboard. */}
-      {/* Clipped on phones for the rounded corners; visible from xl so the
-          desktop matrix header can stick to the page as it scrolls. */}
-      <OverviewMatrixCard>
-        {wide === false ? null : (
-          <DesktopOverviewMatrix
-            models={data.models}
+      <OverviewPresentationSurface>
+        <OverviewMatrixSection locale={locale} />
+      </OverviewPresentationSurface>
+    </section>
+  );
+}
+
+/**
+ * The strip above the matrix. On the page it is just the view tabs plus the
+ * button that starts a presentation. Presenting turns it into the deck's only
+ * toolbar, so it splits into three columns — what is being measured on the
+ * left, which view on the centre, and the way out on the right — rather than
+ * letting a centred row push the tabs off-centre and read Exit as a third tab.
+ */
+function OverviewControlRow({ locale }: { locale: OverviewLocale }) {
+  const data = useOverviewData();
+  const referenceHardware = useOverviewReference();
+  const { presenting } = useOverviewPresentation();
+  const strings = OVERVIEW_STRINGS[locale];
+
+  const views = (
+    <OverviewComparisonSwitcher
+      comparisonMode={data.comparisonMode}
+      engineScope={data.engineScope}
+      tier={data.tier}
+      referenceHardware={referenceHardware}
+      modelScope={data.modelScope}
+      rowScope={data.rowScope}
+      hardwareRowScope={data.hardwareRowScope}
+      locale={locale}
+      strings={strings}
+    />
+  );
+
+  if (!presenting) {
+    // Same three-column skeleton as the presenting toolbar below: the tabs
+    // keep the matrix centre and Present anchors the right edge as an action,
+    // instead of trailing the tabs and reading as a third view.
+    return (
+      <div className="flex flex-wrap items-center justify-center gap-x-3 gap-y-1.5 md:grid md:grid-cols-[1fr_auto_1fr]">
+        <div className="hidden md:block" />
+        <div className="flex flex-wrap items-center justify-center gap-x-3 gap-y-1.5 md:justify-self-center">
+          {views}
+        </div>
+        <div className="md:justify-self-end">
+          <OverviewPresentToggle strings={strings} />
+        </div>
+      </div>
+    );
+  }
+
+  return (
+    <div className="grid grid-cols-[1fr_auto_1fr] items-center gap-x-6">
+      {/* The header card carrying the SLO is gone while presenting, so the
+          control comes along here rather than stranding the audience on
+          whichever tier the deck happened to open on. */}
+      <div className="justify-self-start">
+        <OverviewTierSwitcher
+          tier={data.tier}
+          engineScope={data.engineScope}
+          comparisonMode={data.comparisonMode}
+          referenceHardware={referenceHardware}
+          modelScope={data.modelScope}
+          rowScope={data.rowScope}
+          hardwareRowScope={data.hardwareRowScope}
+          locale={locale}
+          strings={strings}
+        />
+      </div>
+      <div className="justify-self-center">{views}</div>
+      {/* The scope filters live under the matrix on the page, where their full
+          sentences fit. Here they ride next to Exit as chips: the left column
+          is already 355px of SLO, and moving the tabs off the matrix centre to
+          make room there would cost more than it buys. */}
+      <div className="flex items-center gap-x-2 justify-self-end">
+        {data.comparisonMode === 'hardware' ? (
+          <OverviewHardwareRowScopeToggle
+            variant="toolbar"
+            hardwareRowScope={data.hardwareRowScope}
+            emptyRowCount={data.emptyRowCount}
+            tier={data.tier}
+            engineScope={data.engineScope}
+            referenceHardware={referenceHardware}
+            modelScope={data.modelScope}
             locale={locale}
-            formatters={formatters}
             strings={strings}
-            comparisonMode={data.comparisonMode}
-            referenceHardware={referenceHardware}
           />
-        )}
-        {wide === true ? null : (
-          <MobileOverviewList
-            models={data.models}
+        ) : (
+          <OverviewRowScopeToggle
+            variant="toolbar"
+            windowDays={OVERVIEW_HISTORY_WINDOW_DAYS[data.comparisonMode]}
+            rowScope={data.rowScope}
+            unchangedRowCount={data.unchangedRowCount}
+            tier={data.tier}
+            engineScope={data.engineScope}
+            referenceHardware={referenceHardware}
+            modelScope={data.modelScope}
             locale={locale}
-            formatters={formatters}
             strings={strings}
-            comparisonMode={data.comparisonMode}
-            referenceHardware={referenceHardware}
           />
         )}
-        <OverviewMethodology
-          strings={strings}
-          comparisonMode={data.comparisonMode}
-          referenceHardware={referenceHardware}
-        />
         <OverviewModelScopeToggle
+          variant="toolbar"
           modelScope={data.modelScope}
           tier={data.tier}
           engineScope={data.engineScope}
           comparisonMode={data.comparisonMode}
           referenceHardware={referenceHardware}
+          rowScope={data.rowScope}
+          hardwareRowScope={data.hardwareRowScope}
           locale={locale}
           strings={strings}
         />
+        <OverviewPresentToggle strings={strings} />
+      </div>
+    </div>
+  );
+}
+
+/** The half of the page that goes fullscreen: the view tabs and the matrix. */
+function OverviewMatrixSection({ locale }: { locale: OverviewLocale }) {
+  const data = useOverviewData();
+  const referenceHardware = useOverviewReference();
+  const { presenting } = useOverviewPresentation();
+  // Both surfaces used to render on every width and hide one with CSS, so every
+  // selection built the matrix twice. The Tailwind classes stay — they carry
+  // SSR and the pre-hydration frame — and this only drops the unused one after.
+  const wide = useWideViewport();
+  const strings = OVERVIEW_STRINGS[locale];
+  const formatters = overviewFormatters(locale);
+
+  const matrix = (
+    <DesktopOverviewMatrix
+      models={data.models}
+      locale={locale}
+      formatters={formatters}
+      strings={strings}
+      comparisonMode={data.comparisonMode}
+      referenceHardware={referenceHardware}
+      presenting={presenting}
+    />
+  );
+
+  if (data.models.length === 0) {
+    return (
+      <>
+        <OverviewControlRow locale={locale} />
+        <OverviewMatrixCard>
+          <p
+            role="status"
+            data-testid="overview-empty-state"
+            className="px-6 py-12 text-center text-sm text-muted-foreground"
+          >
+            {strings.emptyState}
+          </p>
+        </OverviewMatrixCard>
+      </>
+    );
+  }
+
+  return (
+    <>
+      <OverviewControlRow locale={locale} />
+
+      {/* Official-only summary; uploaded runs remain in the linked dashboard. */}
+      {/* Clipped on phones for the rounded corners; visible from xl so the
+          desktop matrix header can stick to the page as it scrolls. */}
+      <OverviewMatrixCard>
+        {/* A deck lays out at a fixed width and is scaled by `zoom`, so the
+            viewport no longer decides which surface fits: the matrix is the
+            slide at every projector size and the phone list would only ever be
+            dead weight behind it. */}
+        {presenting ? (
+          matrix
+        ) : (
+          <>
+            {wide === false ? null : matrix}
+            {wide === true ? null : (
+              <MobileOverviewList
+                models={data.models}
+                locale={locale}
+                formatters={formatters}
+                strings={strings}
+                comparisonMode={data.comparisonMode}
+                referenceHardware={referenceHardware}
+              />
+            )}
+            {/* One footer bar instead of three stacked link rows: the notes
+                keep the left edge, the scope chips keep the right, and the card
+                ends on a single rule. */}
+            <div className="flex flex-col gap-x-6 gap-y-2 border-t border-border/50 px-4 py-3 lg:flex-row lg:items-center lg:justify-between lg:px-6">
+              <OverviewMethodology
+                strings={strings}
+                comparisonMode={data.comparisonMode}
+                referenceHardware={referenceHardware}
+              />
+              <div className="flex shrink-0 flex-wrap items-center gap-x-2 gap-y-1">
+                {data.comparisonMode === 'hardware' ? (
+                  <OverviewHardwareRowScopeToggle
+                    hardwareRowScope={data.hardwareRowScope}
+                    emptyRowCount={data.emptyRowCount}
+                    tier={data.tier}
+                    engineScope={data.engineScope}
+                    referenceHardware={referenceHardware}
+                    modelScope={data.modelScope}
+                    locale={locale}
+                    strings={strings}
+                  />
+                ) : (
+                  <OverviewRowScopeToggle
+                    windowDays={OVERVIEW_HISTORY_WINDOW_DAYS[data.comparisonMode]}
+                    rowScope={data.rowScope}
+                    unchangedRowCount={data.unchangedRowCount}
+                    tier={data.tier}
+                    engineScope={data.engineScope}
+                    referenceHardware={referenceHardware}
+                    modelScope={data.modelScope}
+                    locale={locale}
+                    strings={strings}
+                  />
+                )}
+                <OverviewModelScopeToggle
+                  modelScope={data.modelScope}
+                  tier={data.tier}
+                  engineScope={data.engineScope}
+                  comparisonMode={data.comparisonMode}
+                  referenceHardware={referenceHardware}
+                  rowScope={data.rowScope}
+                  hardwareRowScope={data.hardwareRowScope}
+                  locale={locale}
+                  strings={strings}
+                />
+              </div>
+            </div>
+          </>
+        )}
       </OverviewMatrixCard>
-    </section>
+    </>
   );
 }
diff --git a/packages/app/src/components/overview/overview-presentation.test.tsx b/packages/app/src/components/overview/overview-presentation.test.tsx
new file mode 100644
index 000000000..0bc104600
--- /dev/null
+++ b/packages/app/src/components/overview/overview-presentation.test.tsx
@@ -0,0 +1,469 @@
+// @vitest-environment jsdom
+
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import { CLIENT_SEARCH_CHANGE_EVENT } from '@/lib/client-navigation';
+import type { OverviewPageData } from '@/lib/overview-data';
+
+const routerStub = vi.hoisted(() => ({ push: vi.fn(), replace: vi.fn() }));
+const trackStub = vi.hoisted(() => vi.fn());
+
+vi.mock('next/navigation', () => ({ useRouter: () => routerStub }));
+vi.mock('@/lib/analytics', () => ({ track: trackStub }));
+
+import { OverviewNavigationProvider, useOverviewNavigation } from './overview-navigation';
+import { DesktopOverviewMatrix, overviewFormatters } from './overview-scorecard';
+import { OVERVIEW_STRINGS, type OverviewStrings } from './overview-strings';
+import {
+  OverviewPresentationProvider,
+  OverviewPresentationSurface,
+  OverviewPresentToggle,
+} from './overview-presentation';
+
+declare global {
+  var IS_REACT_ACT_ENVIRONMENT: boolean;
+}
+
+globalThis.IS_REACT_ACT_ENVIRONMENT = true;
+
+const strings = {
+  presentEnter: 'Present',
+  presentExit: 'Exit',
+  presentEnterAria: 'Show the matrix full screen',
+  presentExitAria: 'Leave full screen',
+  presentShortcutHint: 'Arrow keys switch views · Esc exits',
+} as unknown as OverviewStrings;
+
+describe('OVERVIEW_STRINGS.zh', () => {
+  it('preserves protected identifiers, units, and dynamic values', () => {
+    const zh = OVERVIEW_STRINGS.zh;
+
+    expect(zh.tierUnit).toBe('tok/s/user');
+    expect(zh.scopeAria).toContain('hyperscaler');
+    expect(zh.caption).toContain('token');
+    expect(zh.scenarioLabels.agentx).toContain('AgentX');
+    expect(zh.methodologyNote).toContain('FP4');
+    expect(zh.historyCaption(7)).toContain('7–14');
+    expect(zh.costDeltaAria('20%', true, 'B200')).toEqual(
+      expect.stringMatching(/B200.*20%|20%.*B200/),
+    );
+    expect(zh.historicalDeltaAria('20%', true, '2026年8月1日')).toEqual(
+      expect.stringMatching(/2026年8月1日.*20%|20%.*2026年8月1日/),
+    );
+    expect(zh.rowScopeShow(3, 7)).toEqual(expect.stringMatching(/3.*7|7.*3/));
+    expect(zh.hardwareRowScopeShow(3)).toContain('3');
+  });
+});
+
+const HISTORY_DATA: OverviewPageData = {
+  models: [],
+  tier: 50,
+  engineScope: 'all',
+  comparisonMode: '30d',
+  referenceHardware: 'b200',
+  modelScope: 'default',
+  rowScope: 'all',
+  hardwareRowScope: 'all',
+  unchangedRowCount: 0,
+  emptyRowCount: 0,
+  historicalWindow: null,
+};
+
+let container: HTMLDivElement;
+let root: Root;
+let fullscreenElement: Element | null;
+let requestFullscreen: ReturnType<typeof vi.fn>;
+let exitFullscreen: ReturnType<typeof vi.fn>;
+
+function stubFullscreenApi(enabled: boolean) {
+  fullscreenElement = null;
+  const setFullscreenElement = (element: Element | null) => {
+    fullscreenElement = element;
+    document.dispatchEvent(new Event('fullscreenchange'));
+    return Promise.resolve();
+  };
+  requestFullscreen = vi.fn(function requestFullscreenStub(this: Element) {
+    return setFullscreenElement(this);
+  });
+  exitFullscreen = vi.fn(() => setFullscreenElement(null));
+
+  Object.defineProperty(document, 'fullscreenEnabled', { configurable: true, value: enabled });
+  Object.defineProperty(document, 'fullscreenElement', {
+    configurable: true,
+    get: () => fullscreenElement,
+  });
+  Object.defineProperty(document, 'exitFullscreen', { configurable: true, value: exitFullscreen });
+  Object.defineProperty(Element.prototype, 'requestFullscreen', {
+    configurable: true,
+    value: requestFullscreen,
+  });
+}
+
+function render(initialHref = '/overview?compare=30d') {
+  act(() => {
+    root.render(
+      <OverviewNavigationProvider initialData={HISTORY_DATA} initialHref={initialHref}>
+        <OverviewPresentationProvider locale="en">
+          <OverviewPresentationSurface>
+            <OverviewPresentToggle strings={strings} />
+            <NavigationProbe />
+            <div role="option" tabIndex={-1} data-testid="overview-select-option">
+              7 days
+            </div>
+          </OverviewPresentationSurface>
+        </OverviewPresentationProvider>
+      </OverviewNavigationProvider>,
+    );
+  });
+}
+
+function NavigationProbe() {
+  const navigation = useOverviewNavigation();
+  return (
+    <>
+      <button
+        type="button"
+        data-testid="overview-server-selection"
+        onClick={() => navigation.push('/overview', ['compare'])}
+      >
+        Hardware
+      </button>
+      <output data-testid="overview-next-href">
+        {navigation.resolve('/overview?tier=75', ['tier'])}
+      </output>
+    </>
+  );
+}
+
+const surface = () =>
+  container.querySelector<HTMLElement>('[data-testid="overview-presentation-surface"]');
+const toggle = () =>
+  container.querySelector<HTMLButtonElement>('[data-testid="overview-present-toggle"]');
+const serverSelection = () =>
+  container.querySelector<HTMLButtonElement>('[data-testid="overview-server-selection"]');
+const selectOption = () =>
+  container.querySelector<HTMLElement>('[data-testid="overview-select-option"]');
+
+beforeEach(() => {
+  routerStub.push.mockClear();
+  routerStub.replace.mockClear();
+  trackStub.mockClear();
+  window.history.replaceState({}, '', '/overview?compare=30d');
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+  stubFullscreenApi(true);
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  vi.unstubAllGlobals();
+});
+
+describe('OverviewPresentationSurface', () => {
+  it('hands the surface to the Fullscreen API and mirrors presentation intent in the URL', () => {
+    render();
+    expect(toggle()?.textContent).toBe('Present');
+    expect(surface()?.dataset.presenting).toBe('false');
+
+    act(() => toggle()?.click());
+    expect(requestFullscreen).toHaveBeenCalledTimes(1);
+    expect(window.location.search).toBe('?compare=30d&present=1');
+    expect(fullscreenElement).toBe(surface());
+    expect(surface()?.dataset.presenting).toBe('true');
+    expect(toggle()?.textContent).toBe('Exit');
+    expect(trackStub).toHaveBeenLastCalledWith('overview_presentation_toggled', {
+      action: 'enter',
+    });
+
+    act(() => toggle()?.click());
+    expect(exitFullscreen).toHaveBeenCalledTimes(1);
+    expect(window.location.search).toBe('?compare=30d');
+    expect(surface()?.dataset.presenting).toBe('false');
+    expect(toggle()?.textContent).toBe('Present');
+    expect(trackStub).toHaveBeenLastCalledWith('overview_presentation_toggled', {
+      action: 'exit',
+    });
+  });
+
+  it('follows the browser out of fullscreen when Esc bypasses the button', () => {
+    render();
+    act(() => toggle()?.click());
+    expect(surface()?.dataset.presenting).toBe('true');
+
+    act(() => {
+      fullscreenElement = null;
+      document.dispatchEvent(new Event('fullscreenchange'));
+    });
+    expect(surface()?.dataset.presenting).toBe('false');
+    expect(window.location.search).toBe('?compare=30d');
+  });
+
+  it('reasserts presentation intent when Back or Forward lands on an older URL', async () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(
+        () =>
+          new Promise(() => {
+            // Keep server navigation pending while checking client URL state.
+          }),
+      ),
+    );
+    window.history.replaceState({}, '', '/overview?tier=100&compare=30d');
+    render('/overview?tier=100&compare=30d');
+    act(() => toggle()?.click());
+
+    await act(async () => {
+      window.history.replaceState({}, '', '/overview?tier=75&compare=30d');
+      window.dispatchEvent(new PopStateEvent('popstate'));
+      await Promise.resolve();
+    });
+
+    expect(window.location.search).toBe('?tier=75&compare=30d&present=1');
+    expect(container.querySelector('[data-testid="overview-next-href"]')?.textContent).toBe(
+      '/overview?tier=75&compare=30d&present=1',
+    );
+    expect(surface()?.dataset.presenting).toBe('true');
+    expect(fullscreenElement).toBe(surface());
+    expect(fetch).toHaveBeenCalledWith('/api/v1/overview?tier=75&compare=30d', {
+      headers: { Accept: 'application/json' },
+    });
+  });
+
+  it('does not carry presentation intent onto a route left through browser history', () => {
+    render();
+    act(() => toggle()?.click());
+
+    act(() => {
+      window.history.replaceState({}, '', '/inference');
+      window.dispatchEvent(new PopStateEvent('popstate'));
+    });
+
+    expect(window.location.pathname).toBe('/inference');
+    expect(window.location.search).toBe('');
+  });
+
+  it('preserves every other query parameter and the fragment while notifying persistent chrome', () => {
+    vi.stubGlobal('fetch', vi.fn());
+    window.history.replaceState({}, '', '/overview?tier=75&compare=30d&utm_source=deck#matrix');
+    const searchEvents: string[] = [];
+    const onSearchChange = (event: Event) => {
+      if (event instanceof CustomEvent && typeof event.detail === 'string') {
+        searchEvents.push(event.detail);
+      }
+    };
+    window.addEventListener(CLIENT_SEARCH_CHANGE_EVENT, onSearchChange);
+
+    render('/overview?tier=75&compare=30d');
+    act(() => toggle()?.click());
+
+    expect(window.location.pathname).toBe('/overview');
+    expect(window.location.search).toBe('?tier=75&compare=30d&present=1&utm_source=deck');
+    expect(window.location.hash).toBe('#matrix');
+    expect(searchEvents).toEqual(['?tier=75&compare=30d&present=1&utm_source=deck']);
+    expect(fetch).not.toHaveBeenCalled();
+
+    act(() => toggle()?.click());
+    expect(window.location.search).toBe('?tier=75&compare=30d&utm_source=deck');
+    expect(window.location.hash).toBe('#matrix');
+    expect(searchEvents).toEqual([
+      '?tier=75&compare=30d&present=1&utm_source=deck',
+      '?tier=75&compare=30d&utm_source=deck',
+    ]);
+    window.removeEventListener(CLIENT_SEARCH_CHANGE_EVENT, onSearchChange);
+  });
+
+  it('does not attempt native fullscreen from a shared presentation URL without a user gesture', () => {
+    window.history.replaceState({}, '', '/overview?compare=30d&present=1');
+    render('/overview?compare=30d');
+
+    expect(requestFullscreen).not.toHaveBeenCalled();
+    expect(surface()?.dataset.presenting).toBe('false');
+    expect(window.location.search).toBe('?compare=30d&present=1');
+
+    act(() => toggle()?.click());
+    expect(requestFullscreen).toHaveBeenCalledTimes(1);
+    expect(surface()?.dataset.presenting).toBe('true');
+  });
+
+  it('keeps presentation intent and fullscreen active while paging to another view', async () => {
+    const requested: string[] = [];
+    vi.stubGlobal(
+      'fetch',
+      vi.fn((input: RequestInfo | URL) => {
+        requested.push(String(input));
+        return Promise.resolve(Response.json({ ...HISTORY_DATA, comparisonMode: 'hardware' }));
+      }),
+    );
+    render();
+
+    act(() => toggle()?.click());
+    await act(async () => {
+      window.dispatchEvent(new KeyboardEvent('keydown', { key: 'ArrowRight' }));
+      await Promise.resolve();
+      await Promise.resolve();
+    });
+
+    expect(requested).toHaveLength(1);
+    expect(requested[0]).not.toContain('present=1');
+    expect(window.location.search).toBe('?present=1');
+    expect(surface()?.dataset.presenting).toBe('true');
+    expect(fullscreenElement).toBe(surface());
+  });
+
+  it('does not page when an arrow key belongs to an interactive control', () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() => new Promise<Response>(() => {})),
+    );
+    render();
+    act(() => toggle()?.click());
+
+    act(() =>
+      toggle()?.dispatchEvent(new KeyboardEvent('keydown', { key: 'ArrowRight', bubbles: true })),
+    );
+
+    expect(fetch).not.toHaveBeenCalled();
+    expect(window.location.search).toBe('?compare=30d&present=1');
+  });
+
+  it('does not page when a Radix Select option owns the arrow key', () => {
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(() => new Promise<Response>(() => {})),
+    );
+    render();
+    act(() => toggle()?.click());
+
+    act(() =>
+      selectOption()?.dispatchEvent(
+        new KeyboardEvent('keydown', { key: 'ArrowRight', bubbles: true }),
+      ),
+    );
+
+    expect(fetch).not.toHaveBeenCalled();
+    expect(window.location.search).toBe('?compare=30d&present=1');
+  });
+
+  it('keeps presentation intent when an older server selection fails', async () => {
+    let rejectRequest: ((error: Error) => void) | undefined;
+    vi.stubGlobal(
+      'fetch',
+      vi.fn(
+        () =>
+          new Promise<Response>((_resolve, reject) => {
+            rejectRequest = reject;
+          }),
+      ),
+    );
+    render();
+
+    act(() => serverSelection()?.click());
+    act(() => toggle()?.click());
+    expect(window.location.search).toBe('?present=1');
+
+    await act(async () => {
+      rejectRequest?.(new Error('network down'));
+      await Promise.resolve();
+    });
+
+    expect(window.location.search).toBe('?present=1');
+    expect(surface()?.dataset.presenting).toBe('true');
+    expect(container.querySelector('[data-testid="overview-next-href"]')?.textContent).toContain(
+      'present=1',
+    );
+  });
+
+  it('removes presentation intent if the browser refuses fullscreen', async () => {
+    requestFullscreen.mockRejectedValueOnce(new Error('denied'));
+    render();
+
+    await act(async () => {
+      toggle()?.click();
+      await Promise.resolve();
+    });
+
+    expect(window.location.search).toBe('?compare=30d');
+    expect(surface()?.dataset.presenting).toBe('false');
+  });
+
+  it('keeps presenting without an unhandled rejection if exiting fullscreen is refused', async () => {
+    render();
+    act(() => toggle()?.click());
+    exitFullscreen.mockRejectedValueOnce(new Error('denied'));
+
+    await act(async () => {
+      toggle()?.click();
+      await Promise.resolve();
+    });
+
+    expect(window.location.search).toBe('?compare=30d&present=1');
+    expect(surface()?.dataset.presenting).toBe('true');
+    expect(fullscreenElement).toBe(surface());
+  });
+
+  it('pages between the two views with the arrow keys, but only while presenting', () => {
+    const requested: string[] = [];
+    vi.stubGlobal(
+      'fetch',
+      vi.fn((input: RequestInfo | URL) => {
+        requested.push(String(input));
+        // Never settles: the assertion only cares that the request went out.
+        return new Promise<Response>(() => {});
+      }),
+    );
+    render();
+
+    act(() => window.dispatchEvent(new KeyboardEvent('keydown', { key: 'ArrowRight' })));
+    expect(requested).toEqual([]);
+
+    act(() => toggle()?.click());
+    act(() => window.dispatchEvent(new KeyboardEvent('keydown', { key: 'ArrowRight' })));
+
+    expect(requested).toHaveLength(1);
+    // Leaving the history view is the whole point of the keypress.
+    expect(requested[0]).not.toContain('compare=30d');
+  });
+
+  it('stays out of the way on browsers that refuse fullscreen', () => {
+    stubFullscreenApi(false);
+    window.history.replaceState({}, '', '/overview?compare=30d&present=1&utm_source=deck#matrix');
+    render('/overview?compare=30d');
+    expect(toggle()).toBeNull();
+    expect(surface()).not.toBeNull();
+    expect(window.location.search).toBe('?compare=30d&utm_source=deck');
+    expect(window.location.hash).toBe('#matrix');
+  });
+});
+
+const matrix = (presenting: boolean) => (
+  <DesktopOverviewMatrix
+    models={[]}
+    locale="en"
+    formatters={overviewFormatters('en')}
+    strings={OVERVIEW_STRINGS.en}
+    comparisonMode="hardware"
+    referenceHardware="b200"
+    presenting={presenting}
+  />
+);
+
+describe('DesktopOverviewMatrix', () => {
+  it('drops the viewport gate while presenting so a narrow projector still gets a matrix', () => {
+    const gate = () =>
+      container.querySelector('[data-testid="overview-desktop-matrix"]')?.parentElement?.className;
+
+    act(() => root.render(matrix(false)));
+    expect(gate()).toBe('hidden xl:block');
+
+    // `xl` asks whether the viewport can hold the matrix. That is the right
+    // question on the page and the wrong one on a deck, which lays out at a
+    // fixed width and is scaled by `zoom`: keeping it would blank the slide on
+    // any projector under 1280px, since presenting also drops the phone list.
+    act(() => root.render(matrix(true)));
+    expect(gate()).toBe('block');
+  });
+});
diff --git a/packages/app/src/components/overview/overview-presentation.tsx b/packages/app/src/components/overview/overview-presentation.tsx
new file mode 100644
index 000000000..4c2434a03
--- /dev/null
+++ b/packages/app/src/components/overview/overview-presentation.tsx
@@ -0,0 +1,312 @@
+'use client';
+
+import {
+  createContext,
+  type ReactNode,
+  type RefObject,
+  useCallback,
+  useContext,
+  useEffect,
+  useRef,
+  useState,
+} from 'react';
+
+import { track } from '@/lib/analytics';
+import { notifyClientSearchChange } from '@/lib/client-navigation';
+import { OVERVIEW_DEFAULT_HISTORY_WINDOW, type OverviewComparisonMode } from '@/lib/overview-data';
+import { overviewHref } from '@/lib/overview-links';
+
+import {
+  useOverviewData,
+  useOverviewNavigation,
+  useOverviewReference,
+} from './overview-navigation';
+import type { OverviewLocale, OverviewStrings } from './overview-strings';
+
+/**
+ * Width the matrix is laid out at while presenting, before `zoom` magnifies it.
+ * Fixing it keeps the column proportions identical to the on-page matrix, so a
+ * projected slide reads the same as the page the audience visits afterwards.
+ */
+const PRESENTATION_LAYOUT_WIDTH = 1200;
+
+/** Room left around the matrix so it never runs into the bezel. */
+const PRESENTATION_FILL = { width: 0.96, height: 0.94 };
+
+const PRESENTATION_ZOOM_RANGE = { min: 0.3, max: 3 };
+
+interface OverviewPresentationState {
+  presenting: boolean;
+  /** False until mounted, and on browsers that refuse the Fullscreen API. */
+  supported: boolean;
+  toggle: () => void;
+}
+
+interface OverviewPresentationContextValue extends OverviewPresentationState {
+  surfaceRef: RefObject<HTMLDivElement | null>;
+  scalerRef: RefObject<HTMLDivElement | null>;
+}
+
+const OverviewPresentationContext = createContext<OverviewPresentationContextValue | null>(null);
+
+function usePresentationContext(): OverviewPresentationContextValue {
+  const value = useContext(OverviewPresentationContext);
+  if (value === null) throw new Error('Presentation controls require OverviewPresentationProvider');
+  return value;
+}
+
+export function useOverviewPresentation(): OverviewPresentationState {
+  return usePresentationContext();
+}
+
+/**
+ * For leaves that only need to adapt if a presentation happens to be running.
+ * Unlike the controls, they are reusable outside the overview page, where "not
+ * presenting" is the honest answer rather than a wiring mistake.
+ */
+export function useIsPresenting(): boolean {
+  return useContext(OverviewPresentationContext)?.presenting ?? false;
+}
+
+/**
+ * Owns the fullscreen state for the whole page, not just the surface: the page
+ * body has to know it is presenting so it can drop the chrome that lives
+ * outside the surface from the DOM rather than leaving it for the browser to
+ * merely stop painting.
+ */
+export function OverviewPresentationProvider({
+  locale,
+  children,
+}: {
+  locale: OverviewLocale;
+  children: ReactNode;
+}) {
+  const data = useOverviewData();
+  // Same reason the matrix reads it here: the reference follows the URL, so a
+  // payload still cached from another reference must not rewrite it.
+  const referenceHardware = useOverviewReference();
+  const { push, replaceClientState } = useOverviewNavigation();
+  const surfaceRef = useRef<HTMLDivElement>(null);
+  const scalerRef = useRef<HTMLDivElement>(null);
+  const [presenting, setPresenting] = useState(false);
+  const [supported, setSupported] = useState(false);
+
+  const setPresentationIntent = useCallback(
+    (enabled: boolean) => {
+      const target = new URL(
+        `${window.location.pathname}${window.location.search}${window.location.hash}`,
+        window.location.origin,
+      );
+      if (enabled) target.searchParams.set('present', '1');
+      else target.searchParams.delete('present');
+      replaceClientState(`${target.pathname}${target.search}${target.hash}`, ['present']);
+    },
+    [replaceClientState],
+  );
+
+  useEffect(() => {
+    setSupported(document.fullscreenEnabled);
+    if (!document.fullscreenEnabled) setPresentationIntent(false);
+  }, [setPresentationIntent]);
+
+  useEffect(() => {
+    const syncPresenting = () => {
+      const nextPresenting =
+        surfaceRef.current !== null && document.fullscreenElement === surfaceRef.current;
+      setPresenting(nextPresenting);
+      if (!nextPresenting) setPresentationIntent(false);
+    };
+    document.addEventListener('fullscreenchange', syncPresenting);
+    return () => document.removeEventListener('fullscreenchange', syncPresenting);
+  }, [setPresentationIntent]);
+
+  useEffect(() => {
+    const preserveIntentWhileFullscreen = () => {
+      const onOverview =
+        window.location.pathname === '/overview' || window.location.pathname === '/zh/overview';
+      if (
+        onOverview &&
+        surfaceRef.current !== null &&
+        document.fullscreenElement === surfaceRef.current
+      ) {
+        const target = new URL(
+          `${window.location.pathname}${window.location.search}${window.location.hash}`,
+          window.location.origin,
+        );
+        target.searchParams.set('present', '1');
+        const href = `${target.pathname}${target.search}${target.hash}`;
+        History.prototype.replaceState.call(window.history, window.history.state, '', href);
+        notifyClientSearchChange(href);
+      }
+    };
+    // Capture runs before the navigation provider reads the new location, so
+    // a Back/Forward entry that predates presentation cannot make its client
+    // state disagree with the still-active browser fullscreen surface.
+    window.addEventListener('popstate', preserveIntentWhileFullscreen, true);
+    return () => window.removeEventListener('popstate', preserveIntentWhileFullscreen, true);
+  }, []);
+
+  const toggle = useCallback(() => {
+    const surface = surfaceRef.current;
+    if (surface === null) return;
+    if (document.fullscreenElement === surface) {
+      track('overview_presentation_toggled', { action: 'exit' });
+      void document.exitFullscreen().catch(() => {
+        setPresenting(true);
+        setPresentationIntent(true);
+      });
+    } else {
+      track('overview_presentation_toggled', { action: 'enter' });
+      setPresentationIntent(true);
+      void surface.requestFullscreen().catch(() => {
+        setPresenting(false);
+        setPresentationIntent(false);
+      });
+    }
+  }, [setPresentationIntent]);
+
+  // Magnify rather than restyle: one `zoom` scales type, padding and rules
+  // together, so the projected matrix cannot drift from the page's own layout.
+  const fit = useCallback(() => {
+    const surface = surfaceRef.current;
+    const scaler = scalerRef.current;
+    if (surface === null || scaler === null) return;
+    scaler.style.removeProperty('zoom');
+    if (!presenting) return;
+
+    const { width, height } = scaler.getBoundingClientRect();
+    if (width === 0 || height === 0) return;
+    const factor = Math.min(
+      (surface.clientWidth * PRESENTATION_FILL.width) / width,
+      (surface.clientHeight * PRESENTATION_FILL.height) / height,
+    );
+    scaler.style.zoom = String(
+      Math.min(Math.max(factor, PRESENTATION_ZOOM_RANGE.min), PRESENTATION_ZOOM_RANGE.max),
+    );
+  }, [presenting]);
+
+  // `data` is the dependency that matters beyond resizing: changing the view or
+  // the SLO mid presentation changes the row count, and the matrix has to be
+  // refitted.
+  useEffect(() => {
+    fit();
+    window.addEventListener('resize', fit);
+    return () => window.removeEventListener('resize', fit);
+  }, [fit, data]);
+
+  useEffect(() => {
+    if (!presenting) return undefined;
+    const handleKeyDown = (event: KeyboardEvent) => {
+      if (event.key !== 'ArrowLeft' && event.key !== 'ArrowRight') return;
+      // Key repeat would oscillate between the two views and spam fetches;
+      // one page per physical press.
+      if (event.repeat) return;
+      const eventTarget = event.target;
+      if (
+        eventTarget instanceof Element &&
+        eventTarget.closest(
+          'a, button, input, select, textarea, [contenteditable="true"], [role="button"], [role="combobox"], [role="listbox"], [role="option"], [role="menuitem"]',
+        ) !== null
+      ) {
+        return;
+      }
+      // Two views, so either arrow means "the other one" — the audience reads
+      // it as paging through slides.
+      const target: OverviewComparisonMode =
+        data.comparisonMode === 'hardware' ? OVERVIEW_DEFAULT_HISTORY_WINDOW : 'hardware';
+      event.preventDefault();
+      // Only `compare` is merged out of this href, so both row scopes stay on
+      // the URL as they are and there is nothing to pass for them here.
+      push(
+        overviewHref(
+          locale,
+          data.tier,
+          data.engineScope,
+          target,
+          referenceHardware,
+          data.modelScope,
+        ),
+        ['compare'],
+      );
+    };
+    window.addEventListener('keydown', handleKeyDown);
+    return () => window.removeEventListener('keydown', handleKeyDown);
+  }, [presenting, data, referenceHardware, locale, push]);
+
+  return (
+    <OverviewPresentationContext.Provider
+      value={{ presenting, supported, toggle, surfaceRef, scalerRef }}
+    >
+      {children}
+    </OverviewPresentationContext.Provider>
+  );
+}
+
+/**
+ * The element handed to the Fullscreen API. Everything outside it stops
+ * rendering while it is fullscreen, which is what strips the page down to the
+ * matrix.
+ *
+ * Switching views or SLO while presenting is safe: `OverviewNavigationProvider`
+ * swaps the data client-side without a route change, so this element is never
+ * unmounted and the browser keeps it fullscreen.
+ */
+export function OverviewPresentationSurface({ children }: { children: ReactNode }) {
+  const { presenting, surfaceRef, scalerRef } = usePresentationContext();
+  return (
+    <div
+      ref={surfaceRef}
+      data-testid="overview-presentation-surface"
+      data-presenting={presenting}
+      className={
+        presenting
+          ? 'flex h-full w-full items-center justify-center overflow-auto bg-background'
+          : undefined
+      }
+    >
+      {/* The toolbar needs more air from the matrix once it is the only chrome
+          on a projector than it does as one strip among many on the page. */}
+      <div
+        ref={scalerRef}
+        className={`flex flex-col ${presenting ? 'gap-6' : 'gap-4'}`}
+        style={presenting ? { width: PRESENTATION_LAYOUT_WIDTH } : undefined}
+      >
+        {children}
+      </div>
+    </div>
+  );
+}
+
+export function OverviewPresentToggle({ strings }: { strings: OverviewStrings }) {
+  const { presenting, supported, toggle } = useOverviewPresentation();
+  if (!supported) return null;
+  return (
+    <button
+      type="button"
+      data-testid="overview-present-toggle"
+      onClick={toggle}
+      aria-pressed={presenting}
+      aria-label={presenting ? strings.presentExitAria : strings.presentEnterAria}
+      title={strings.presentShortcutHint}
+      className="inline-flex min-h-11 items-center gap-x-1.5 rounded-md border border-border/60 px-3 text-xs text-muted-foreground transition-colors hover:bg-muted hover:text-foreground focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring/50"
+    >
+      <svg
+        aria-hidden="true"
+        viewBox="0 0 16 16"
+        fill="none"
+        stroke="currentColor"
+        strokeWidth="1.5"
+        strokeLinecap="round"
+        strokeLinejoin="round"
+        className="size-3.5"
+      >
+        {presenting ? (
+          <path d="M6 2v4H2M10 2v4h4M6 14v-4H2M10 14v-4h4" />
+        ) : (
+          <path d="M6 2H2v4M10 2h4v4M6 14H2v-4M10 14h4v-4" />
+        )}
+      </svg>
+      {presenting ? strings.presentExit : strings.presentEnter}
+    </button>
+  );
+}
diff --git a/packages/app/src/components/overview/overview-reference-select.tsx b/packages/app/src/components/overview/overview-reference-select.tsx
index 9acc4b905..0cc2b4c7f 100644
--- a/packages/app/src/components/overview/overview-reference-select.tsx
+++ b/packages/app/src/components/overview/overview-reference-select.tsx
@@ -11,6 +11,7 @@ import { track } from '@/lib/analytics';
 import type { OverviewReferenceHardware } from '@/lib/overview-data';
 
 import { useOverviewNavigation, useOverviewReference } from './overview-navigation';
+import { useIsPresenting } from './overview-presentation';
 
 interface ReferenceOption {
   href: string;
@@ -31,6 +32,11 @@ export function OverviewReferenceSelect({
   // instead of silently discarding it. Nothing to prefetch — a reference change
   // costs no request.
   const value = useOverviewReference();
+  // A portalled menu lands on `document.body`, which is outside the element the
+  // browser is showing fullscreen and outside the `zoom` the matrix is scaled
+  // by. Rendering it in place keeps it both visible and the same size as the
+  // tab that opened it.
+  const presenting = useIsPresenting();
 
   return (
     <Select
@@ -50,7 +56,7 @@ export function OverviewReferenceSelect({
       >
         <SelectValue>{options.find((option) => option.value === value)?.label}</SelectValue>
       </SelectTrigger>
-      <SelectContent align="center">
+      <SelectContent align="center" portalled={!presenting}>
         {options.map((option) => (
           <SelectItem
             key={option.value}
diff --git a/packages/app/src/components/overview/overview-scorecard.tsx b/packages/app/src/components/overview/overview-scorecard.tsx
index 21b8c99c7..38e4a80db 100644
--- a/packages/app/src/components/overview/overview-scorecard.tsx
+++ b/packages/app/src/components/overview/overview-scorecard.tsx
@@ -2,21 +2,24 @@
 
 import { type ComponentPropsWithoutRef, useEffect, useRef } from 'react';
 
+import { ModelLogo } from '@/components/ui/model-logo';
 import {
   OVERVIEW_DEFAULT_HISTORY_WINDOW,
   OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
+  OVERVIEW_DEFAULT_ROW_SCOPE,
   OVERVIEW_HARDWARE,
   OVERVIEW_HISTORY_WINDOW_DAYS,
   OVERVIEW_HISTORY_WINDOWS,
-  OVERVIEW_TIERS,
   overviewHardwareLabel,
   type OverviewComparisonMode,
   type OverviewEngineScope,
+  type OverviewHardwareRowScope,
   type OverviewHistoricalComparison,
   type OverviewModelScope,
   type OverviewModelSummary,
   type OverviewPlatformResult,
   type OverviewReferenceHardware,
+  type OverviewRowScope,
   type OverviewTier,
 } from '@/lib/overview-data';
 import {
@@ -25,7 +28,6 @@ import {
   detailHref,
   overviewEngineScopeHref,
   overviewHref,
-  overviewTierHref,
 } from '@/lib/overview-links';
 
 import { OverviewDetailLink } from './overview-detail-link';
@@ -34,172 +36,8 @@ import { OverviewHistoryWindowSelect } from './overview-history-window-select';
 import { OverviewNavLink } from './overview-nav-link';
 import { type OverviewNavControl, useOverviewNavigation } from './overview-navigation';
 import { OverviewReferenceSelect } from './overview-reference-select';
-
-export type OverviewLocale = 'en' | 'zh';
-
-export const OVERVIEW_STRINGS = {
-  en: {
-    title: 'Inference Cost per Million Tokens',
-    // The active tier is not repeated here — the SLO selector below already
-    // states it.
-    scopeMetric: 'Hyperscaler cost',
-    scopeDirection: '↓ Lower is better',
-    // The unit is dropped from the visible line but kept for screen readers.
-    scopeAria: 'Hyperscaler cost per one million total tokens. Lower is better.',
-    sourcePrefix: 'Source: InferenceX & ',
-    sourceLinkText: 'SemiAnalysis Market July 2026 AI Cloud TCO Model',
-    tierNavLabel: 'SLO',
-    tierUnit: 'tok/s/user',
-    engineScopeNavLabel: 'Engine scope',
-    engineScopeOptions: {
-      all: 'All Platforms',
-      community: 'Open Source Community Engines (vLLM/SGLang)',
-    },
-    comparisonNavLabel: 'Compare',
-    comparisonOptions: {
-      history: 'Change over time',
-    },
-    historyWindowOptions: {
-      '7d': '1 week ago',
-      '30d': '1 month ago',
-      '60d': '2 months ago',
-      '90d': '3 months ago',
-    } as Record<string, string>,
-    historyWindowSelectAria: 'Comparison window',
-    hardwareComparisonLabel: (reference: string) => `vs ${reference}`,
-    referenceSelectorAria: 'Reference hardware',
-    caption:
-      'Cost per million total tokens from each platform’s best observed serving envelope for the scenario shown with each model.',
-    historyCaption: (days: number) =>
-      `Current cost and change versus the latest validated platform result ${days}–${days * 2} days earlier.`,
-    modelHeader: 'Model · Scenario',
-    scenarioLabels: {
-      single_turn_8k1k: '8K/1K',
-      agentx: 'Long Context Multi-Turn Realistic Agentic Scenario (AgentX)',
-    },
-    detailLink: 'View details',
-    detailAria: (modelLabel: string, scenarioLabel: string) =>
-      `View details: ${modelLabel} · ${scenarioLabel}`,
-    compareCurvesLink: 'Compare curves',
-    compareCurvesAria: (modelLabel: string, hardwareLabel: string) =>
-      `Compare current and historical ${hardwareLabel} cost curves for ${modelLabel}`,
-    rawDashboardAria: (evidenceDate: string, modelLabel: string, stack: string) =>
-      `Open raw source dashboard for ${evidenceDate}: ${modelLabel} · ${stack}`,
-    estimatedTooltip: (topologies: readonly string[]) =>
-      topologies.length === 0
-        ? 'Estimated from validated benchmark runs.'
-        : `Estimated from validated ${topologies.join(' and ')} runs.`,
-    estimatedAria: (value: string, explanation: string) => `Approximately ${value}. ${explanation}`,
-    cellStateLegend: (reference: string) => `— = no result. ∞ = ${reference} baseline unavailable.`,
-    missingReasons: (tier: number): Record<string, string> => ({
-      int4_bf16_only: 'INT4/BF16 only',
-      no_scenario_data: 'no data for this scenario',
-      cannot_reach_at_tier: `cannot reach @${tier}`,
-      no_exact_at_tier: `no exact @${tier} result`,
-    }),
-    standardDecodeLabel: 'STP',
-    methodologyNote:
-      'If a chip does not have FP4 spec decoding available, the next best available configuration is used.',
-    costDeltaAria: (pct: string, cheaper: boolean, reference: string) =>
-      `${pct} ${cheaper ? 'cheaper' : 'more expensive'} than ${reference}`,
-    costDeltaEvenAria: (reference: string) => `About the same cost as ${reference}`,
-    noBaselineAria: (reference: string) => `No ${reference} baseline to compare against`,
-    historicalDeltaAria: (pct: string, cheaper: boolean, baselineDate: string) =>
-      `${pct} ${cheaper ? 'cheaper' : 'more expensive'} than this platform’s ${baselineDate} result`,
-    historicalEvenAria: (baselineDate: string) =>
-      `About the same cost as this platform’s ${baselineDate} result`,
-    historyCellStateLegend: (days: number) =>
-      `Platforms without a valid ${days}-day comparison show current cost only.`,
-    referenceHeader: 'Reference',
-    modelScopeNavLabel: 'Inactive models',
-    modelScopeShow: 'Show deprecated & maintenance-mode models',
-    modelScopeHide: 'Hide deprecated & maintenance-mode models',
-    categoryBadges: {
-      maintenance: 'Maintenance',
-      deprecated: 'Deprecated',
-    } as Partial<Record<string, string>>,
-    categoryBadgeTitle: 'Model is no longer actively benchmarked.',
-    loadingStatus: 'Loading the selected comparison…',
-  },
-  zh: {
-    title: '推理每百万 token 成本',
-    scopeMetric: '超大规模云（hyperscaler）成本',
-    scopeDirection: '↓ 越低越好',
-    scopeAria: '超大规模云（hyperscaler）每百万总 token 成本，越低越好。',
-    sourcePrefix: '来源：InferenceX 与 ',
-    sourceLinkText: 'SemiAnalysis Market July 2026 AI Cloud TCO Model',
-    tierNavLabel: 'SLO',
-    tierUnit: 'tok/s/用户',
-    engineScopeNavLabel: '引擎范围',
-    engineScopeOptions: {
-      all: '所有平台',
-      community: '开源社区引擎（vLLM/SGLang）',
-    },
-    comparisonNavLabel: '对比方式',
-    comparisonOptions: {
-      history: '历史变化',
-    },
-    historyWindowOptions: {
-      '7d': '1 周前',
-      '30d': '1 个月前',
-      '60d': '2 个月前',
-      '90d': '3 个月前',
-    } as Record<string, string>,
-    historyWindowSelectAria: '对比时间窗口',
-    hardwareComparisonLabel: (reference: string) => `对比 ${reference}`,
-    referenceSelectorAria: '基准硬件',
-    caption: '按各模型标注的场景，基于各平台最佳观测服务包络线计算每百万总 token 成本。',
-    historyCaption: (days: number) =>
-      `当前成本及其相对 ${days}–${days * 2} 天前最近一次有效平台结果的变化。`,
-    modelHeader: '模型 · 场景',
-    scenarioLabels: {
-      single_turn_8k1k: '8K/1K',
-      agentx: '长上下文多轮真实智能体场景（AgentX）',
-    },
-    detailLink: '查看详情',
-    detailAria: (modelLabel: string, scenarioLabel: string) =>
-      `查看详情：${modelLabel} · ${scenarioLabel}`,
-    compareCurvesLink: '对比曲线',
-    compareCurvesAria: (modelLabel: string, hardwareLabel: string) =>
-      `对比 ${modelLabel} 在 ${hardwareLabel} 上当前与历史成本曲线`,
-    rawDashboardAria: (evidenceDate: string, modelLabel: string, stack: string) =>
-      `打开 ${evidenceDate} 原始数据仪表板：${modelLabel} · ${stack}`,
-    estimatedTooltip: (topologies: readonly string[]) =>
-      topologies.length === 0
-        ? '根据已验证的基准运行结果估算。'
-        : `根据已验证的 ${topologies.join(' 与 ')} 运行结果估算。`,
-    estimatedAria: (value: string, explanation: string) => `约 ${value}。${explanation}`,
-    cellStateLegend: (reference: string) => `— = 无结果。∞ = 缺少 ${reference} 基线。`,
-    missingReasons: (tier: number): Record<string, string> => ({
-      int4_bf16_only: '仅 INT4/BF16',
-      no_scenario_data: '该场景暂无数据',
-      cannot_reach_at_tier: `无法达到 @${tier}`,
-      no_exact_at_tier: `无精确 @${tier} 结果`,
-    }),
-    standardDecodeLabel: 'STP',
-    methodologyNote: '若某款芯片不支持 FP4 推测解码，则采用次优的可用配置。',
-    costDeltaAria: (pct: string, cheaper: boolean, reference: string) =>
-      `比 ${reference} ${cheaper ? '便宜' : '昂贵'} ${pct}`,
-    costDeltaEvenAria: (reference: string) => `与 ${reference} 成本基本持平`,
-    noBaselineAria: (reference: string) => `缺少可比较的 ${reference} 基线`,
-    historicalDeltaAria: (pct: string, cheaper: boolean, baselineDate: string) =>
-      `比该平台 ${baselineDate} 的结果${cheaper ? '便宜' : '昂贵'} ${pct}`,
-    historicalEvenAria: (baselineDate: string) => `与该平台 ${baselineDate} 的结果成本基本持平`,
-    historyCellStateLegend: (days: number) => `缺少有效 ${days} 天对比的平台仅显示当前成本。`,
-    referenceHeader: '基准',
-    modelScopeNavLabel: '非活跃模型',
-    modelScopeShow: '显示已弃用与维护模式模型',
-    modelScopeHide: '隐藏已弃用与维护模式模型',
-    categoryBadges: {
-      maintenance: '维护模式',
-      deprecated: '已弃用',
-    } as Partial<Record<string, string>>,
-    categoryBadgeTitle: '该模型已不再进行活跃基准测试。',
-    loadingStatus: '正在加载所选对比…',
-  },
-} as const;
-
-export type OverviewStrings = (typeof OVERVIEW_STRINGS)[OverviewLocale];
+import { type OverviewLocale, type OverviewStrings } from './overview-strings';
+import { OverviewTierSlider } from './overview-tier-slider';
 
 interface Formatters {
   cost: Intl.NumberFormat;
@@ -277,7 +115,7 @@ function CellMissing({ hardware, reason }: { hardware: string; reason: string })
     >
       <span>{'—'}</span>
       {reason === '' ? null : (
-        <div className="min-w-0 text-[11px] leading-tight font-normal text-muted-foreground/70">
+        <div className="min-w-0 text-2xs leading-tight font-normal text-muted-foreground/70">
           {reason}
         </div>
       )}
@@ -292,8 +130,8 @@ const COST_DELTA_SATURATION = 0.5;
 // Missing comparison evidence is neutral gray, never red/green: availability
 // is not a better/worse judgment.
 const COST_DELTA_CLASS = {
-  cheaper: 'text-emerald-700 dark:text-emerald-400',
-  pricier: 'text-red-700 dark:text-red-400',
+  cheaper: 'text-emerald-900 dark:text-emerald-200',
+  pricier: 'text-red-900 dark:text-red-200',
   even: 'text-muted-foreground',
   'no-baseline': 'text-muted-foreground',
 } as const;
@@ -461,7 +299,7 @@ function CostDeltaBadge({
       title={aria}
       // The cell behind it carries the shade, so the badge itself stays
       // untinted — two washes of the same hue would double up.
-      className={`inline-flex translate-y-px items-center whitespace-nowrap rounded-sm px-1 py-0.5 text-[10px] font-semibold tabular-nums ${
+      className={`inline-flex translate-y-px items-center whitespace-nowrap rounded-sm px-1 py-0.5 text-3xs font-semibold tabular-nums ${
         phoneRow ? 'col-start-2 justify-self-start' : 'xl:col-start-2 xl:justify-self-end'
       } ${COST_DELTA_CLASS[polarity]}`}
     >
@@ -608,7 +446,7 @@ function CellValue({
         )}
       </div>
       {member.precision === null ? null : (
-        <div className="min-w-0 text-[11px] leading-tight font-normal uppercase tracking-wider text-muted-foreground/70">
+        <div className="min-w-0 text-2xs leading-tight font-normal uppercase tracking-wider text-muted-foreground">
           {config === null ? (
             member.precision.toUpperCase()
           ) : phoneRow && stackPrefix !== null && decodeLabel !== null ? (
@@ -667,26 +505,46 @@ function PlatformCell(props: {
 
 function ModelName({ model, strings }: { model: OverviewModelSummary; strings: OverviewStrings }) {
   const badge = strings.categoryBadges[model.category];
+  const label = strings.scenarioLabels[model.scenario];
+  const shortLabel = strings.scenarioLabelsShort[model.scenario];
   return (
     <div>
       <h2 className="text-sm font-semibold leading-snug">
+        {/* Full-color creator mark, same treatment as the /inference chart
+            caption and /model pages. `ModelLogo` renders nothing when the
+            model has no configured logo or the asset fails to load. */}
+        <ModelLogo model={model.model} className="mr-1.5" />
         {model.modelLabel}
         {badge === undefined ? null : (
           <span
             data-testid="overview-model-category-badge"
             data-category={model.category}
             title={strings.categoryBadgeTitle}
-            className="ml-1.5 inline-block rounded-sm border border-border/60 px-1 py-px align-middle text-[10px] font-medium uppercase tracking-wide text-muted-foreground"
+            className="ml-1.5 inline-block rounded-sm border border-border/60 px-1 py-px align-middle text-3xs font-medium uppercase tracking-wide text-muted-foreground"
           >
             {badge}
+            {/* `title` reaches a hovering mouse and nothing else, so a screen
+                reader announced this badge as a bare label with no reason
+                attached. `normal-case` because the badge's `uppercase` inherits
+                and browsers expose the transformed string to the accessibility
+                tree, which turns the sentence into shouted, mis-parsed words. */}
+            <span className="sr-only normal-case">{strings.categoryBadgeTitle}</span>
           </span>
         )}
       </h2>
       <p
         data-testid="overview-model-scenario"
-        className="mt-0.5 text-[11px] font-normal leading-tight text-muted-foreground"
+        title={label}
+        className="mt-0.5 text-2xs font-normal leading-tight text-muted-foreground"
       >
-        {strings.scenarioLabels[model.scenario]}
+        {shortLabel === label ? (
+          label
+        ) : (
+          <>
+            <span className="sr-only">{label}</span>
+            <span aria-hidden="true">{shortLabel}</span>
+          </>
+        )}
       </p>
     </div>
   );
@@ -708,11 +566,16 @@ export function DesktopOverviewMatrix({
   strings,
   comparisonMode,
   referenceHardware,
-}: SurfaceProps) {
+  presenting = false,
+}: SurfaceProps & { presenting?: boolean }) {
   const platforms = models[0]?.platforms ?? [];
   const referenceLabel = overviewHardwareLabel(referenceHardware);
   return (
-    <div className="hidden xl:block">
+    // The `xl` gate asks whether the viewport can hold the matrix, which stops
+    // being the question once presenting: the deck lays out at a fixed width and
+    // is scaled by `zoom`, so on a projector narrower than 1280px this would
+    // hide the matrix on a slide whose phone list has already been dropped.
+    <div className={presenting ? 'block' : 'hidden xl:block'}>
       <table data-testid="overview-desktop-matrix" className="w-full border-collapse text-sm">
         <caption className="sr-only">
           {comparisonMode === 'hardware'
@@ -762,7 +625,7 @@ export function DesktopOverviewMatrix({
                 data-scenario={model.scenario}
                 className="border-b border-border/50 align-top last:border-b-0"
               >
-                <th scope="row" className="px-4 py-4 text-left align-top font-normal lg:px-6">
+                <th scope="row" className="px-4 py-2.5 text-left align-top font-normal lg:px-6">
                   <ModelName model={model} strings={strings} />
                   {/* The link lives with the model it drills into, so the matrix
                     spends no column on a header that is the same every row. */}
@@ -789,7 +652,7 @@ export function DesktopOverviewMatrix({
                       referenceHardware,
                       referenceCost,
                     )}
-                    className={`px-3 py-4 align-top ${comparisonMode === 'hardware' && platform.hardware === referenceHardware ? 'bg-muted/30' : ''}`}
+                    className={`px-3 py-2.5 align-top ${comparisonMode === 'hardware' && platform.hardware === referenceHardware ? 'bg-muted/30' : ''}`}
                   >
                     <PlatformCell
                       locale={locale}
@@ -924,14 +787,15 @@ function ActiveSwitcherOption({
   );
 }
 
-/** Every option remains a copyable server-rendered URL; ordinary clicks use a
- *  soft App Router transition and the displayed tier is never a self-link. */
+/** The six benchmarked service levels are exposed as discrete slider stops. */
 export function OverviewTierSwitcher({
   tier,
   engineScope,
   comparisonMode,
   referenceHardware,
   modelScope,
+  rowScope,
+  hardwareRowScope,
   locale,
   strings,
 }: {
@@ -940,51 +804,24 @@ export function OverviewTierSwitcher({
   comparisonMode: OverviewComparisonMode;
   referenceHardware: OverviewReferenceHardware;
   modelScope: OverviewModelScope;
+  rowScope: OverviewRowScope;
+  hardwareRowScope: OverviewHardwareRowScope;
   locale: OverviewLocale;
   strings: OverviewStrings;
 }) {
-  const optionClass =
-    'inline-flex min-h-11 items-center px-3 tabular-nums focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring';
   return (
-    <nav
-      data-testid="overview-tier-switcher"
-      aria-label={strings.tierNavLabel}
-      className="flex flex-wrap items-center gap-x-2 gap-y-1 text-xs"
-    >
-      <span className="text-muted-foreground">{strings.tierNavLabel}</span>
-      <div className="flex divide-x divide-border/60 overflow-hidden rounded-md border border-border/60">
-        {OVERVIEW_TIERS.map((option) =>
-          option === tier ? (
-            <ActiveSwitcherOption
-              key={option}
-              control="tier"
-              aria-current="page"
-              className={`${optionClass} bg-muted font-semibold text-foreground`}
-            >
-              {option}
-            </ActiveSwitcherOption>
-          ) : (
-            <OverviewNavLink
-              key={option}
-              href={overviewTierHref(
-                locale,
-                option,
-                engineScope,
-                comparisonMode,
-                referenceHardware,
-                modelScope,
-              )}
-              analytics={{ control: 'tier', value: String(option) }}
-              searchKeys={['tier']}
-              className={`${optionClass} text-muted-foreground transition-colors hover:text-foreground`}
-            >
-              {option}
-            </OverviewNavLink>
-          ),
-        )}
-      </div>
-      <span className="text-muted-foreground">{strings.tierUnit}</span>
-    </nav>
+    <OverviewTierSlider
+      tier={tier}
+      engineScope={engineScope}
+      comparisonMode={comparisonMode}
+      referenceHardware={referenceHardware}
+      modelScope={modelScope}
+      rowScope={rowScope}
+      hardwareRowScope={hardwareRowScope}
+      locale={locale}
+      label={strings.tierNavLabel}
+      unit={strings.tierUnit}
+    />
   );
 }
 
@@ -995,6 +832,8 @@ export function OverviewEngineScopeSwitcher({
   comparisonMode,
   referenceHardware,
   modelScope,
+  rowScope,
+  hardwareRowScope,
   locale,
   strings,
 }: {
@@ -1003,6 +842,8 @@ export function OverviewEngineScopeSwitcher({
   comparisonMode: OverviewComparisonMode;
   referenceHardware: OverviewReferenceHardware;
   modelScope: OverviewModelScope;
+  rowScope: OverviewRowScope;
+  hardwareRowScope: OverviewHardwareRowScope;
   locale: OverviewLocale;
   strings: OverviewStrings;
 }) {
@@ -1039,6 +880,8 @@ export function OverviewEngineScopeSwitcher({
                 comparisonMode,
                 referenceHardware,
                 modelScope,
+                rowScope,
+                hardwareRowScope,
               )}
               analytics={{ control: 'engine', value: option }}
               searchKeys={['engine']}
@@ -1059,6 +902,8 @@ export function OverviewComparisonSwitcher({
   tier,
   referenceHardware,
   modelScope,
+  rowScope,
+  hardwareRowScope,
   locale,
   strings,
 }: {
@@ -1067,6 +912,8 @@ export function OverviewComparisonSwitcher({
   tier: OverviewTier;
   referenceHardware: OverviewReferenceHardware;
   modelScope: OverviewModelScope;
+  rowScope: OverviewRowScope;
+  hardwareRowScope: OverviewHardwareRowScope;
   locale: OverviewLocale;
   strings: OverviewStrings;
 }) {
@@ -1074,7 +921,16 @@ export function OverviewComparisonSwitcher({
   const referenceOptions = OVERVIEW_HARDWARE.map((hardware) => ({
     value: hardware,
     label: overviewHardwareLabel(hardware),
-    href: overviewHref(locale, tier, engineScope, 'hardware', hardware, modelScope),
+    href: overviewHref(
+      locale,
+      tier,
+      engineScope,
+      'hardware',
+      hardware,
+      modelScope,
+      rowScope,
+      hardwareRowScope,
+    ),
   }));
   const windowOptions = OVERVIEW_HISTORY_WINDOWS.map((window) => ({
     value: window,
@@ -1082,9 +938,10 @@ export function OverviewComparisonSwitcher({
     href: overviewHref(locale, tier, engineScope, window, referenceHardware, modelScope),
   }));
   // The inactive-only classes live on the inactive branch, not here: Tailwind
-  // emits `border-transparent` after `border-secondary` at equal specificity,
-  // so sharing them left the active underline invisible in light mode. Same
-  // reason for the hover border — it would grey out the active underline.
+  // emits `border-transparent` after the active border at equal specificity,
+  // so sharing them left the active underline invisible in light mode. The
+  // darker light-theme blue also keeps this small active label above AA
+  // contrast on the page background; dark mode retains the brand primary.
   const optionClass =
     'relative inline-flex min-h-11 min-w-[130px] items-center justify-center whitespace-nowrap border-b-2 px-4 py-2 text-sm font-semibold transition-colors duration-200 focus-visible:outline-none focus-visible:ring-[3px] focus-visible:ring-ring sm:min-w-[140px]';
   const inactiveOptionClass =
@@ -1100,7 +957,7 @@ export function OverviewComparisonSwitcher({
           control="comparison"
           data-overview-comparison="hardware"
           aria-current="true"
-          className={`${optionClass} border-secondary text-secondary dark:border-primary dark:text-primary`}
+          className={`${optionClass} border-sky-800 text-sky-800 dark:border-primary dark:text-primary`}
         >
           <span className="inline-flex items-center gap-0.5">
             <span>{locale === 'zh' ? '对比 ' : 'vs '}</span>
@@ -1113,7 +970,16 @@ export function OverviewComparisonSwitcher({
       ) : (
         <OverviewNavLink
           data-overview-comparison="hardware"
-          href={overviewHref(locale, tier, engineScope, 'hardware', referenceHardware, modelScope)}
+          href={overviewHref(
+            locale,
+            tier,
+            engineScope,
+            'hardware',
+            referenceHardware,
+            modelScope,
+            rowScope,
+            hardwareRowScope,
+          )}
           analytics={{ control: 'comparison', value: 'hardware' }}
           searchKeys={['compare']}
           className={`${optionClass} ${inactiveOptionClass}`}
@@ -1131,6 +997,8 @@ export function OverviewComparisonSwitcher({
             OVERVIEW_DEFAULT_HISTORY_WINDOW,
             referenceHardware,
             modelScope,
+            rowScope,
+            hardwareRowScope,
           )}
           analytics={{ control: 'comparison', value: OVERVIEW_DEFAULT_HISTORY_WINDOW }}
           searchKeys={['compare']}
@@ -1143,7 +1011,7 @@ export function OverviewComparisonSwitcher({
           control="comparison"
           data-overview-comparison={comparisonMode}
           aria-current="true"
-          className={`${optionClass} border-secondary text-secondary dark:border-primary dark:text-primary`}
+          className={`${optionClass} border-sky-800 text-sky-800 dark:border-primary dark:text-primary`}
         >
           <span className="inline-flex items-center gap-0.5">
             <span>{locale === 'zh' ? '对比 ' : 'vs '}</span>
@@ -1159,39 +1027,235 @@ export function OverviewComparisonSwitcher({
   );
 }
 
+/**
+ * Where a scope toggle is drawn. `section` is the chip in the footer bar
+ * beneath the matrix; `toolbar` is the same chip riding the deck toolbar while
+ * presenting, where the count badge would be noise at projection size.
+ */
+type OverviewScopeToggleVariant = 'section' | 'toolbar';
+
+/**
+ * Matches Exit, so the right end of the deck toolbar reads as one row of
+ * actions. Deliberately not filled-when-engaged: these labels name the click
+ * rather than the state, and a filled "Show all rows" contradicts itself.
+ */
+const SCOPE_CHIP_CLASS =
+  'inline-flex min-h-11 items-center whitespace-nowrap rounded-md border border-border/60 px-3 text-xs text-muted-foreground transition-colors hover:bg-muted hover:text-foreground';
+
+/** The count the chip's short label elides; the full sentence stays on the
+ *  accessible name, so the badge is presentation only. */
+function ScopeChipCount({ count }: { count: number }) {
+  return (
+    <span
+      aria-hidden="true"
+      className="ml-1.5 inline-flex min-w-5 items-center justify-center rounded-full bg-muted px-1.5 py-0.5 text-3xs leading-none tabular-nums"
+    >
+      {count}
+    </span>
+  );
+}
+
 export function OverviewModelScopeToggle({
   modelScope,
   tier,
   engineScope,
   comparisonMode,
   referenceHardware,
+  rowScope,
+  hardwareRowScope,
   locale,
   strings,
+  variant = 'section',
 }: {
   modelScope: OverviewModelScope;
   tier: OverviewTier;
   engineScope: OverviewEngineScope;
   comparisonMode: OverviewComparisonMode;
   referenceHardware: OverviewReferenceHardware;
+  rowScope: OverviewRowScope;
+  hardwareRowScope: OverviewHardwareRowScope;
   locale: OverviewLocale;
   strings: OverviewStrings;
+  variant?: OverviewScopeToggleVariant;
 }) {
   const target: OverviewModelScope = modelScope === 'all' ? 'default' : 'all';
+  // Inactive models rarely post a 30-day change, so revealing them under the
+  // changed-only scope would filter them straight back out and read as a dead
+  // link. Asking for more models asks for more rows.
+  const targetRowScope: OverviewRowScope = target === 'all' ? 'all' : rowScope;
+  // Same trap in hardware mode, and a deeper one: an inactive model is the most
+  // likely to have no result on any platform, which is exactly what the
+  // priced-only scope removes.
+  const targetHardwareRowScope: OverviewHardwareRowScope =
+    target === 'all' ? 'all' : hardwareRowScope;
+  const sentence = modelScope === 'all' ? strings.modelScopeHide : strings.modelScopeShow;
+  const link = (
+    <OverviewNavLink
+      data-overview-model-scope={target}
+      href={overviewHref(
+        locale,
+        tier,
+        engineScope,
+        comparisonMode,
+        referenceHardware,
+        target,
+        targetRowScope,
+        targetHardwareRowScope,
+      )}
+      analytics={{ control: 'models', value: target }}
+      searchKeys={['models', 'rows', 'hwrows']}
+      aria-label={sentence}
+      title={sentence}
+      className={SCOPE_CHIP_CLASS}
+    >
+      {modelScope === 'all' ? strings.modelScopeChipHide : strings.modelScopeChipShow}
+    </OverviewNavLink>
+  );
+  if (variant === 'toolbar') return link;
   return (
     <nav
       data-testid="overview-model-scope-toggle"
       aria-label={strings.modelScopeNavLabel}
-      className="border-t border-border/50 px-4 text-xs lg:px-6"
+      className="text-xs"
     >
-      <OverviewNavLink
-        data-overview-model-scope={target}
-        href={overviewHref(locale, tier, engineScope, comparisonMode, referenceHardware, target)}
-        analytics={{ control: 'models', value: target }}
-        searchKeys={['models']}
-        className="inline-flex min-h-11 items-center text-muted-foreground underline decoration-dotted underline-offset-4 transition-colors hover:text-foreground hover:decoration-solid"
-      >
-        {modelScope === 'all' ? strings.modelScopeHide : strings.modelScopeShow}
-      </OverviewNavLink>
+      {link}
+    </nav>
+  );
+}
+
+/**
+ * Opt-in narrowing to the rows that moved in the window, and the way back.
+ * Rendered only when the two scopes would differ, so a fully-comparable window
+ * shows no dead control.
+ */
+export function OverviewRowScopeToggle({
+  rowScope,
+  unchangedRowCount,
+  windowDays,
+  tier,
+  engineScope,
+  referenceHardware,
+  modelScope,
+  locale,
+  strings,
+  variant = 'section',
+}: {
+  rowScope: OverviewRowScope;
+  /** Days of the active history window; the copy must name the window the
+   *  filter actually reads. */
+  windowDays: number;
+  unchangedRowCount: number;
+  tier: OverviewTier;
+  engineScope: OverviewEngineScope;
+  referenceHardware: OverviewReferenceHardware;
+  modelScope: OverviewModelScope;
+  locale: OverviewLocale;
+  strings: OverviewStrings;
+  variant?: OverviewScopeToggleVariant;
+}) {
+  if (unchangedRowCount === 0) return null;
+  const target: OverviewRowScope = rowScope === 'all' ? 'changed' : 'all';
+  const sentence =
+    rowScope === 'all'
+      ? strings.rowScopeHide(unchangedRowCount, windowDays)
+      : strings.rowScopeShow(unchangedRowCount, windowDays);
+  const link = (
+    <OverviewNavLink
+      data-overview-row-scope={target}
+      href={overviewHref(
+        locale,
+        tier,
+        engineScope,
+        OVERVIEW_DEFAULT_HISTORY_WINDOW,
+        referenceHardware,
+        modelScope,
+        target,
+      )}
+      analytics={{ control: 'rows', value: target }}
+      searchKeys={['rows']}
+      aria-label={sentence}
+      title={sentence}
+      className={SCOPE_CHIP_CLASS}
+    >
+      {rowScope === 'all' ? strings.rowScopeChipHide : strings.rowScopeChipShow}
+      {variant === 'section' ? <ScopeChipCount count={unchangedRowCount} /> : null}
+    </OverviewNavLink>
+  );
+  if (variant === 'toolbar') return link;
+  return (
+    <nav
+      data-testid="overview-row-scope-toggle"
+      aria-label={strings.rowScopeNavLabel(windowDays)}
+      className="text-xs"
+    >
+      {link}
+    </nav>
+  );
+}
+
+/** The hardware-mode sibling of {@link OverviewRowScopeToggle}. Kept separate
+ *  rather than parameterised: the two filters answer different questions, name
+ *  different counts, and each mode remembers its own answer. */
+export function OverviewHardwareRowScopeToggle({
+  hardwareRowScope,
+  emptyRowCount,
+  tier,
+  engineScope,
+  referenceHardware,
+  modelScope,
+  locale,
+  strings,
+  variant = 'section',
+}: {
+  hardwareRowScope: OverviewHardwareRowScope;
+  emptyRowCount: number;
+  tier: OverviewTier;
+  engineScope: OverviewEngineScope;
+  referenceHardware: OverviewReferenceHardware;
+  modelScope: OverviewModelScope;
+  locale: OverviewLocale;
+  strings: OverviewStrings;
+  variant?: OverviewScopeToggleVariant;
+}) {
+  if (emptyRowCount === 0) return null;
+  const target: OverviewHardwareRowScope = hardwareRowScope === 'all' ? 'priced' : 'all';
+  const sentence =
+    hardwareRowScope === 'all'
+      ? strings.hardwareRowScopeHide(emptyRowCount)
+      : strings.hardwareRowScopeShow(emptyRowCount);
+  const link = (
+    <OverviewNavLink
+      data-overview-hardware-row-scope={target}
+      href={overviewHref(
+        locale,
+        tier,
+        engineScope,
+        'hardware',
+        referenceHardware,
+        modelScope,
+        OVERVIEW_DEFAULT_ROW_SCOPE,
+        target,
+      )}
+      analytics={{ control: 'hwrows', value: target }}
+      searchKeys={['hwrows']}
+      aria-label={sentence}
+      title={sentence}
+      className={SCOPE_CHIP_CLASS}
+    >
+      {hardwareRowScope === 'all'
+        ? strings.hardwareRowScopeChipHide
+        : strings.hardwareRowScopeChipShow}
+      {variant === 'section' ? <ScopeChipCount count={emptyRowCount} /> : null}
+    </OverviewNavLink>
+  );
+  if (variant === 'toolbar') return link;
+  return (
+    <nav
+      data-testid="overview-hardware-row-scope-toggle"
+      aria-label={strings.hardwareRowScopeNavLabel}
+      className="text-xs"
+    >
+      {link}
     </nav>
   );
 }
@@ -1209,7 +1273,7 @@ export function OverviewMethodology({
   return (
     <div
       data-testid="overview-methodology"
-      className="space-y-1 border-t border-border/50 px-4 py-3 text-xs leading-snug text-muted-foreground lg:px-6"
+      className="space-y-1 text-xs leading-snug text-muted-foreground"
     >
       {comparisonMode === 'hardware' ? null : (
         <p>{strings.historyCaption(OVERVIEW_HISTORY_WINDOW_DAYS[comparisonMode])}</p>
diff --git a/packages/app/src/components/overview/overview-strings.ts b/packages/app/src/components/overview/overview-strings.ts
new file mode 100644
index 000000000..478ff8be9
--- /dev/null
+++ b/packages/app/src/components/overview/overview-strings.ts
@@ -0,0 +1,224 @@
+import { TCO_SOURCE_TITLE } from '@semianalysisai/inferencex-constants';
+
+export type OverviewLocale = 'en' | 'zh';
+
+export const OVERVIEW_STRINGS = {
+  en: {
+    title: 'Agentic Inference Costs',
+    // The active tier is not repeated here — the SLO selector below already
+    // states it.
+    scopeMetric: 'Hyperscaler cost',
+    scopeDirection: '↓ Lower is better',
+    // The unit is dropped from the visible line but kept for screen readers.
+    scopeAria: 'Hyperscaler cost per one million total tokens. Lower is better.',
+    sourcePrefix: 'Source: InferenceX & ',
+    sourceLinkText: TCO_SOURCE_TITLE,
+    tierNavLabel: 'SLO',
+    tierUnit: 'tok/s/user',
+    engineScopeNavLabel: 'Engine scope',
+    engineScopeOptions: {
+      all: 'All Platforms',
+      community: 'Open Source Community Engines (vLLM/SGLang)',
+    },
+    comparisonNavLabel: 'Compare',
+    comparisonOptions: {
+      history: 'Change over time',
+    },
+    historyWindowOptions: {
+      '7d': '1 week ago',
+      '30d': '1 month ago',
+      '60d': '2 months ago',
+      '90d': '3 months ago',
+    } as Record<string, string>,
+    historyWindowSelectAria: 'Comparison window',
+    hardwareComparisonLabel: (reference: string) => `vs ${reference}`,
+    referenceSelectorAria: 'Reference hardware',
+    caption:
+      'Cost per million total tokens from each platform’s best observed serving envelope for the scenario shown with each model.',
+    historyCaption: (days: number) =>
+      `Current cost and change versus the latest validated platform result ${days}–${days * 2} days earlier.`,
+    modelHeader: 'Model · Scenario',
+    scenarioLabels: {
+      single_turn_8k1k: '8K/1K',
+      agentx: 'Long Context Multi-Turn Realistic Agentic Scenario (AgentX)',
+    },
+    // Rendered in the row header; the full label above stays the accessible
+    // name and the hover title, so nothing is lost by not wrapping it to three
+    // lines in a 22%-wide column.
+    scenarioLabelsShort: {
+      single_turn_8k1k: '8K/1K',
+      agentx: 'AgentX',
+    },
+    detailLink: 'View details',
+    detailAria: (modelLabel: string, scenarioLabel: string) =>
+      `View details: ${modelLabel} · ${scenarioLabel}`,
+    compareCurvesLink: 'Compare curves',
+    compareCurvesAria: (modelLabel: string, hardwareLabel: string) =>
+      `Compare current and historical ${hardwareLabel} cost curves for ${modelLabel}`,
+    rawDashboardAria: (evidenceDate: string, modelLabel: string, stack: string) =>
+      `Open raw source dashboard for ${evidenceDate}: ${modelLabel} · ${stack}`,
+    estimatedTooltip: (topologies: readonly string[]) =>
+      topologies.length === 0
+        ? 'Estimated from validated benchmark runs.'
+        : `Estimated from validated ${topologies.join(' and ')} runs.`,
+    estimatedAria: (value: string, explanation: string) => `Approximately ${value}. ${explanation}`,
+    cellStateLegend: (reference: string) => `— = no result. ∞ = ${reference} baseline unavailable.`,
+    missingReasons: (tier: number): Record<string, string> => ({
+      int4_bf16_only: 'INT4/BF16 only',
+      no_scenario_data: 'no data for this scenario',
+      cannot_reach_at_tier: `cannot reach @${tier}`,
+      no_exact_at_tier: `no exact @${tier} result`,
+    }),
+    standardDecodeLabel: 'STP',
+    methodologyNote:
+      'If a chip does not have FP4 spec decoding available, the next best available configuration is used.',
+    costDeltaAria: (pct: string, cheaper: boolean, reference: string) =>
+      `${pct} ${cheaper ? 'cheaper' : 'more expensive'} than ${reference}`,
+    costDeltaEvenAria: (reference: string) => `About the same cost as ${reference}`,
+    noBaselineAria: (reference: string) => `No ${reference} baseline to compare against`,
+    historicalDeltaAria: (pct: string, cheaper: boolean, baselineDate: string) =>
+      `${pct} ${cheaper ? 'cheaper' : 'more expensive'} than this platform’s ${baselineDate} result`,
+    historicalEvenAria: (baselineDate: string) =>
+      `About the same cost as this platform’s ${baselineDate} result`,
+    historyCellStateLegend: (days: number) =>
+      `Platforms without a valid ${days}-day comparison show current cost only.`,
+    referenceHeader: 'Reference',
+    modelScopeNavLabel: 'Inactive models',
+    modelScopeShow: 'Show deprecated & maintenance-mode models',
+    modelScopeHide: 'Hide deprecated & maintenance-mode models',
+    rowScopeNavLabel: (days: number) => `Rows without a ${days}-day change`,
+    rowScopeShow: (count: number, days: number) =>
+      `Show ${count} ${count === 1 ? 'row' : 'rows'} with no ${days}-day change`,
+    rowScopeHide: (count: number, days: number) =>
+      `Hide ${count} ${count === 1 ? 'row' : 'rows'} with no ${days}-day change`,
+    hardwareRowScopeNavLabel: 'Rows with no result',
+    hardwareRowScopeShow: (count: number) =>
+      `Show ${count} ${count === 1 ? 'row' : 'rows'} with no result on any platform`,
+    hardwareRowScopeHide: (count: number) =>
+      `Hide ${count} ${count === 1 ? 'row' : 'rows'} with no result on any platform`,
+    // Short forms of the sentences above, for the presentation toolbar. They
+    // carry the same verb, so they flip with the scope the same way; the
+    // sentence with the count stays on the accessible name.
+    rowScopeChipHide: 'Hide unchanged',
+    rowScopeChipShow: 'Show all rows',
+    hardwareRowScopeChipHide: 'Hide blanks',
+    hardwareRowScopeChipShow: 'Show all rows',
+    modelScopeChipShow: 'Show inactive',
+    modelScopeChipHide: 'Hide inactive',
+    presentEnter: 'Present',
+    presentExit: 'Exit',
+    presentEnterAria: 'Show the matrix full screen',
+    presentExitAria: 'Leave full screen',
+    presentShortcutHint: 'Arrow keys switch views · Esc exits',
+    categoryBadges: {
+      maintenance: 'Maintenance',
+      deprecated: 'Deprecated',
+    } as Partial<Record<string, string>>,
+    categoryBadgeTitle: 'Model is no longer actively benchmarked.',
+    loadingStatus: 'Loading the selected comparison…',
+    navigationError:
+      'Could not load the selected comparison. Showing the last successfully loaded data.',
+    emptyState: 'No overview results match this selection.',
+  },
+  zh: {
+    title: '智能体推理成本',
+    scopeMetric: '超大规模云厂商成本',
+    scopeDirection: '↓ 越低越好',
+    scopeAria: '每百万总 token 的超大规模云厂商（hyperscaler）成本，越低越好。',
+    sourcePrefix: '来源：InferenceX 与 ',
+    sourceLinkText: TCO_SOURCE_TITLE,
+    tierNavLabel: 'SLO',
+    tierUnit: 'tok/s/user',
+    engineScopeNavLabel: '引擎范围',
+    engineScopeOptions: {
+      all: '所有平台',
+      community: '开源社区引擎（vLLM/SGLang）',
+    },
+    comparisonNavLabel: '对比方式',
+    comparisonOptions: {
+      history: '历史变化',
+    },
+    historyWindowOptions: {
+      '7d': '1 周前',
+      '30d': '1 个月前',
+      '60d': '2 个月前',
+      '90d': '3 个月前',
+    } as Record<string, string>,
+    historyWindowSelectAria: '对比时间窗口',
+    hardwareComparisonLabel: (reference: string) => `对比 ${reference}`,
+    referenceSelectorAria: '基准硬件',
+    caption: '根据各模型标注的测试场景，采用各平台实测的最优服务包络线，计算每百万总 token 成本。',
+    historyCaption: (days: number) =>
+      `当前成本，以及相较于 ${days}–${days * 2} 天前该平台最近一次有效结果的变化。`,
+    modelHeader: '模型 · 场景',
+    scenarioLabels: {
+      single_turn_8k1k: '8K/1K',
+      agentx: '长上下文、多轮交互的真实智能体场景（AgentX）',
+    },
+    scenarioLabelsShort: {
+      single_turn_8k1k: '8K/1K',
+      agentx: 'AgentX',
+    },
+    detailLink: '查看详情',
+    detailAria: (modelLabel: string, scenarioLabel: string) =>
+      `查看详情：${modelLabel} · ${scenarioLabel}`,
+    compareCurvesLink: '对比曲线',
+    compareCurvesAria: (modelLabel: string, hardwareLabel: string) =>
+      `对比 ${modelLabel} 在 ${hardwareLabel} 上当前与历史成本曲线`,
+    rawDashboardAria: (evidenceDate: string, modelLabel: string, stack: string) =>
+      `打开 ${evidenceDate} 原始数据仪表板：${modelLabel} · ${stack}`,
+    estimatedTooltip: (topologies: readonly string[]) =>
+      topologies.length === 0
+        ? '根据已验证的基准测试结果估算。'
+        : `根据已验证的 ${topologies.join(' 与 ')} 测试结果估算。`,
+    estimatedAria: (value: string, explanation: string) => `约 ${value}。${explanation}`,
+    cellStateLegend: (reference: string) => `— = 无结果。∞ = 缺少 ${reference} 基线。`,
+    missingReasons: (tier: number): Record<string, string> => ({
+      int4_bf16_only: '仅 INT4/BF16',
+      no_scenario_data: '该场景暂无数据',
+      cannot_reach_at_tier: `无法达到 @${tier}`,
+      no_exact_at_tier: `无精确 @${tier} 结果`,
+    }),
+    standardDecodeLabel: 'STP',
+    methodologyNote: '如果某款芯片没有可用的 FP4 投机解码配置，则改用次优配置。',
+    costDeltaAria: (pct: string, cheaper: boolean, reference: string) =>
+      `成本比 ${reference} ${cheaper ? '低' : '高'} ${pct}`,
+    costDeltaEvenAria: (reference: string) => `与 ${reference} 成本基本持平`,
+    noBaselineAria: (reference: string) => `缺少可比较的 ${reference} 基线`,
+    historicalDeltaAria: (pct: string, cheaper: boolean, baselineDate: string) =>
+      `成本比该平台 ${baselineDate} 的结果${cheaper ? '低' : '高'} ${pct}`,
+    historicalEvenAria: (baselineDate: string) => `与该平台 ${baselineDate} 的结果成本基本持平`,
+    historyCellStateLegend: (days: number) => `缺少有效 ${days} 天对比的平台仅显示当前成本。`,
+    referenceHeader: '基准',
+    modelScopeNavLabel: '非活跃模型',
+    modelScopeShow: '显示已弃用与维护模式模型',
+    modelScopeHide: '隐藏已弃用与维护模式模型',
+    rowScopeNavLabel: (days: number) => `${days} 天内无变化的行`,
+    rowScopeShow: (count: number, days: number) => `显示过去 ${days} 天无变化的 ${count} 行数据`,
+    rowScopeHide: (count: number, days: number) => `隐藏过去 ${days} 天无变化的 ${count} 行数据`,
+    hardwareRowScopeNavLabel: '无结果的行',
+    hardwareRowScopeShow: (count: number) => `显示所有平台均无结果的 ${count} 行数据`,
+    hardwareRowScopeHide: (count: number) => `隐藏所有平台均无结果的 ${count} 行数据`,
+    rowScopeChipHide: '隐藏无变化',
+    rowScopeChipShow: '显示全部行',
+    hardwareRowScopeChipHide: '隐藏空行',
+    hardwareRowScopeChipShow: '显示全部行',
+    modelScopeChipShow: '显示停用模型',
+    modelScopeChipHide: '隐藏停用模型',
+    presentEnter: '演示',
+    presentExit: '退出',
+    presentEnterAria: '全屏展示矩阵',
+    presentExitAria: '退出全屏',
+    presentShortcutHint: '方向键切换视图 · Esc 退出',
+    categoryBadges: {
+      maintenance: '维护模式',
+      deprecated: '已弃用',
+    } as Partial<Record<string, string>>,
+    categoryBadgeTitle: '目前已不再对该模型进行持续基准测试。',
+    loadingStatus: '正在加载所选对比…',
+    navigationError: '无法加载所选对比，当前显示的是上次成功加载的数据。',
+    emptyState: '没有符合当前筛选条件的总览结果。',
+  },
+} as const;
+
+export type OverviewStrings = (typeof OVERVIEW_STRINGS)[OverviewLocale];
diff --git a/packages/app/src/components/overview/overview-switcher-focus.test.tsx b/packages/app/src/components/overview/overview-switcher-focus.test.tsx
index e4b399d4a..c7bf6cc8c 100644
--- a/packages/app/src/components/overview/overview-switcher-focus.test.tsx
+++ b/packages/app/src/components/overview/overview-switcher-focus.test.tsx
@@ -4,7 +4,7 @@ import React, { act } from 'react';
 import { createRoot, type Root } from 'react-dom/client';
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
-import type { OverviewPageData, OverviewTier } from '@/lib/overview-data';
+import type { OverviewEngineScope, OverviewPageData } from '@/lib/overview-data';
 
 const routerStub = vi.hoisted(() => ({ push: vi.fn(), replace: vi.fn() }));
 
@@ -13,7 +13,8 @@ vi.mock('next/navigation', () => ({
 }));
 
 import { OverviewNavigationProvider, useOverviewData } from './overview-navigation';
-import { OVERVIEW_STRINGS, OverviewTierSwitcher } from './overview-scorecard';
+import { OverviewEngineScopeSwitcher } from './overview-scorecard';
+import { OVERVIEW_STRINGS } from './overview-strings';
 
 declare global {
   var IS_REACT_ACT_ENVIRONMENT: boolean;
@@ -24,50 +25,54 @@ globalThis.IS_REACT_ACT_ENVIRONMENT = true;
 let container: HTMLDivElement;
 let root: Root;
 
-function pageData(tier: OverviewTier): OverviewPageData {
+function pageData(engineScope: OverviewEngineScope): OverviewPageData {
   return {
     models: [],
-    tier,
-    engineScope: 'community',
+    tier: 50,
+    engineScope,
     comparisonMode: 'hardware',
     referenceHardware: 'b200',
     modelScope: 'default',
+    rowScope: 'all',
+    hardwareRowScope: 'all',
+    unchangedRowCount: 0,
+    emptyRowCount: 0,
     historicalWindow: null,
   };
 }
 
-/** Mirrors the real page: the tier comes from the payload in context, so the
+/** Mirrors the real page: the scope comes from the payload in context, so the
  *  active option moves when the response lands. */
 function Body() {
   const data = useOverviewData();
   return (
-    <OverviewTierSwitcher
-      tier={data.tier}
+    <OverviewEngineScopeSwitcher
       engineScope={data.engineScope}
+      tier={data.tier}
       comparisonMode={data.comparisonMode}
       referenceHardware={data.referenceHardware}
       modelScope={data.modelScope}
+      rowScope={data.rowScope}
+      hardwareRowScope={data.hardwareRowScope}
       locale="en"
       strings={OVERVIEW_STRINGS.en}
     />
   );
 }
 
-function renderSwitcher(tier: OverviewTier, href: string) {
+function renderSwitcher(engineScope: OverviewEngineScope, href: string) {
   act(() => {
     root.render(
-      <OverviewNavigationProvider initialData={pageData(tier)} initialHref={href}>
+      <OverviewNavigationProvider initialData={pageData(engineScope)} initialHref={href}>
         <Body />
       </OverviewNavigationProvider>,
     );
   });
 }
 
-function tierOption(label: string): HTMLElement {
-  const match = [...container.querySelectorAll<HTMLElement>('a, span')].find(
-    (element) => element.textContent === label && element.tagName !== 'NAV',
-  );
-  if (match === undefined) throw new Error(`no tier option labelled ${label}`);
+function scopeOption(scope: OverviewEngineScope): HTMLElement {
+  const match = container.querySelector<HTMLElement>(`[data-overview-engine-scope="${scope}"]`);
+  if (match === null) throw new Error(`no engine scope option ${scope}`);
   return match;
 }
 
@@ -88,11 +93,11 @@ afterEach(() => {
 
 describe('overview switcher focus', () => {
   it('moves focus to the option that replaces the activated link', async () => {
-    vi.stubGlobal('fetch', vi.fn().mockResolvedValue(Response.json(pageData(75))));
+    vi.stubGlobal('fetch', vi.fn().mockResolvedValue(Response.json(pageData('all'))));
 
-    renderSwitcher(50, '/overview');
+    renderSwitcher('community', '/overview');
 
-    const link = tierOption('75');
+    const link = scopeOption('all');
     expect(link.tagName).toBe('A');
     link.focus();
     expect(document.activeElement).toBe(link);
@@ -104,22 +109,22 @@ describe('overview switcher focus', () => {
 
     // The activated anchor is gone — React swapped in the active <span> — and
     // without the focus handoff the browser drops focus to <body>.
-    const active = tierOption('75');
+    const active = scopeOption('all');
     expect(active.tagName).toBe('SPAN');
-    expect(active.getAttribute('aria-current')).toBe('page');
+    expect(active.getAttribute('aria-current')).toBe('true');
     expect(document.activeElement).not.toBe(document.body);
     expect(document.activeElement).toBe(active);
   });
 
   it('leaves focus alone when the click did not come from the focused element', async () => {
-    vi.stubGlobal('fetch', vi.fn().mockResolvedValue(Response.json(pageData(75))));
+    vi.stubGlobal('fetch', vi.fn().mockResolvedValue(Response.json(pageData('all'))));
 
-    renderSwitcher(50, '/overview');
+    renderSwitcher('community', '/overview');
 
     // A pointer click never focuses the anchor first in jsdom, so there is no
     // focus to restore and the active option must not steal it.
     await act(async () => {
-      tierOption('75').dispatchEvent(
+      scopeOption('all').dispatchEvent(
         new MouseEvent('click', { bubbles: true, cancelable: true, button: 0 }),
       );
       await Promise.resolve();
diff --git a/packages/app/src/components/overview/overview-tier-slider.tsx b/packages/app/src/components/overview/overview-tier-slider.tsx
new file mode 100644
index 000000000..061bd9f24
--- /dev/null
+++ b/packages/app/src/components/overview/overview-tier-slider.tsx
@@ -0,0 +1,189 @@
+'use client';
+
+import { type ChangeEvent, useEffect, useRef, useState } from 'react';
+
+import { track } from '@/lib/analytics';
+import {
+  OVERVIEW_TIERS,
+  type OverviewComparisonMode,
+  type OverviewEngineScope,
+  type OverviewHardwareRowScope,
+  type OverviewModelScope,
+  type OverviewReferenceHardware,
+  type OverviewRowScope,
+  type OverviewTier,
+} from '@/lib/overview-data';
+import { overviewTierHref } from '@/lib/overview-links';
+
+import { useOverviewNavigation, useOverviewNavigationError } from './overview-navigation';
+
+interface OverviewTierSliderProps {
+  tier: OverviewTier;
+  engineScope: OverviewEngineScope;
+  comparisonMode: OverviewComparisonMode;
+  referenceHardware: OverviewReferenceHardware;
+  modelScope: OverviewModelScope;
+  rowScope: OverviewRowScope;
+  hardwareRowScope: OverviewHardwareRowScope;
+  locale: 'en' | 'zh';
+  label: string;
+  unit: string;
+}
+
+function tierIndex(tier: OverviewTier): number {
+  return OVERVIEW_TIERS.indexOf(tier);
+}
+
+export function OverviewTierSlider({
+  tier,
+  engineScope,
+  comparisonMode,
+  referenceHardware,
+  modelScope,
+  rowScope,
+  hardwareRowScope,
+  locale,
+  label,
+  unit,
+}: OverviewTierSliderProps) {
+  const navigation = useOverviewNavigation();
+  const navigationError = useOverviewNavigationError();
+  const pointerActive = useRef(false);
+  const committedIndex = useRef(tierIndex(tier));
+  const [selectedIndex, setSelectedIndex] = useState(() => tierIndex(tier));
+
+  useEffect(() => {
+    const index = tierIndex(tier);
+    committedIndex.current = index;
+    // A payload settling mid-drag must not yank the thumb away from the stop
+    // the pointer is choosing; the gesture's own pointerup/cancel settles it.
+    if (!pointerActive.current) setSelectedIndex(index);
+  }, [tier]);
+
+  useEffect(() => {
+    // A failed load leaves the settled tier behind while the thumb and URL
+    // stay on the unsuccessful stop. Re-open the guard so activating that
+    // same stop again retries the navigation instead of returning early.
+    if (navigationError) committedIndex.current = tierIndex(tier);
+  }, [navigationError, tier]);
+
+  const selectedTier = OVERVIEW_TIERS[selectedIndex];
+  const progress = (selectedIndex / (OVERVIEW_TIERS.length - 1)) * 100;
+
+  const commit = (index: number) => {
+    const nextTier = OVERVIEW_TIERS[index];
+    if (index === committedIndex.current) return;
+    committedIndex.current = index;
+
+    track('overview_selector_changed', {
+      control: 'tier',
+      value: String(nextTier),
+    });
+    navigation.push(
+      overviewTierHref(
+        locale,
+        nextTier,
+        engineScope,
+        comparisonMode,
+        referenceHardware,
+        modelScope,
+        rowScope,
+        hardwareRowScope,
+      ),
+      ['tier'],
+    );
+  };
+
+  const handleChange = (event: ChangeEvent<HTMLInputElement>) => {
+    const nextIndex = Number(event.currentTarget.value);
+    setSelectedIndex(nextIndex);
+    if (!pointerActive.current) commit(nextIndex);
+  };
+
+  return (
+    <nav
+      data-testid="overview-tier-switcher"
+      aria-label={label}
+      className="flex min-w-0 items-center gap-3 text-xs"
+    >
+      <span className="shrink-0 text-muted-foreground">{label}</span>
+      <div className="w-[min(18rem,calc(100vw-8.5rem))] min-w-0 pt-1">
+        <div className="relative h-4">
+          <input
+            data-testid="overview-tier-slider"
+            type="range"
+            min={0}
+            max={OVERVIEW_TIERS.length - 1}
+            step={1}
+            value={selectedIndex}
+            aria-label={label}
+            aria-valuetext={`${selectedTier} ${unit}`}
+            data-tier={selectedTier}
+            onChange={handleChange}
+            onPointerDown={() => {
+              pointerActive.current = true;
+            }}
+            onPointerUp={(event) => {
+              pointerActive.current = false;
+              commit(Number(event.currentTarget.value));
+            }}
+            onPointerCancel={() => {
+              pointerActive.current = false;
+              setSelectedIndex(tierIndex(tier));
+            }}
+            className="absolute top-1/2 z-10 h-1.5 w-full -translate-y-1/2 cursor-pointer appearance-none rounded-full bg-border accent-foreground focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring focus-visible:ring-offset-2 focus-visible:ring-offset-background [&::-moz-range-progress]:h-1.5 [&::-moz-range-progress]:rounded-full [&::-moz-range-progress]:bg-foreground [&::-moz-range-thumb]:h-4 [&::-moz-range-thumb]:w-4 [&::-moz-range-thumb]:cursor-grab [&::-moz-range-thumb]:rounded-full [&::-moz-range-thumb]:border-2 [&::-moz-range-thumb]:border-background [&::-moz-range-thumb]:bg-foreground [&::-moz-range-thumb]:shadow-sm [&::-moz-range-thumb]:active:cursor-grabbing [&::-webkit-slider-thumb]:h-4 [&::-webkit-slider-thumb]:w-4 [&::-webkit-slider-thumb]:cursor-grab [&::-webkit-slider-thumb]:appearance-none [&::-webkit-slider-thumb]:rounded-full [&::-webkit-slider-thumb]:border-2 [&::-webkit-slider-thumb]:border-background [&::-webkit-slider-thumb]:bg-foreground [&::-webkit-slider-thumb]:shadow-sm [&::-webkit-slider-thumb]:active:cursor-grabbing"
+            style={{
+              background: `linear-gradient(to right, var(--foreground) 0%, var(--foreground) ${progress}%, var(--border) ${progress}%, var(--border) 100%)`,
+            }}
+          />
+          <div
+            data-testid="overview-tier-ridges"
+            className="pointer-events-none absolute inset-0 z-20"
+            aria-hidden="true"
+          >
+            {OVERVIEW_TIERS.map((option, index) => {
+              const state =
+                index === selectedIndex
+                  ? 'selected'
+                  : index < selectedIndex
+                    ? 'filled'
+                    : 'unfilled';
+              return (
+                <span
+                  key={option}
+                  data-testid="overview-tier-ridge"
+                  data-tier-ridge={option}
+                  data-state={state}
+                  className={`absolute top-1/2 h-2 w-0.5 -translate-x-1/2 -translate-y-1/2 rounded-full ${
+                    state === 'selected'
+                      ? 'opacity-0'
+                      : state === 'filled'
+                        ? 'bg-background/70'
+                        : 'bg-foreground/35'
+                  }`}
+                  style={{ left: `${(index / (OVERVIEW_TIERS.length - 1)) * 100}%` }}
+                />
+              );
+            })}
+          </div>
+        </div>
+        <div className="relative mt-1 h-4 tabular-nums" aria-hidden="true">
+          {OVERVIEW_TIERS.map((option, index) => (
+            <span
+              key={option}
+              data-tier-option={option}
+              data-selected={option === selectedTier ? 'true' : undefined}
+              className={`absolute -translate-x-1/2 ${
+                option === selectedTier ? 'font-semibold text-foreground' : 'text-muted-foreground'
+              }`}
+              style={{ left: `${(index / (OVERVIEW_TIERS.length - 1)) * 100}%` }}
+            >
+              {option}
+            </span>
+          ))}
+        </div>
+      </div>
+      <span className="shrink-0 text-muted-foreground">{unit}</span>
+    </nav>
+  );
+}
diff --git a/packages/app/src/components/overview/use-wide-viewport.ts b/packages/app/src/components/overview/use-wide-viewport.ts
index adcbd622b..194a3f50f 100644
--- a/packages/app/src/components/overview/use-wide-viewport.ts
+++ b/packages/app/src/components/overview/use-wide-viewport.ts
@@ -1,21 +1,33 @@
-import { useEffect, useState } from 'react';
+import { useSyncExternalStore } from 'react';
 
 const QUERY = '(min-width: 80rem)';
 
+const NOOP = () => {};
+
+function subscribe(onStoreChange: () => void): () => void {
+  if (typeof window === 'undefined' || typeof window.matchMedia !== 'function') return NOOP;
+  const mediaQuery = window.matchMedia(QUERY);
+  mediaQuery.addEventListener('change', onStoreChange);
+  return () => mediaQuery.removeEventListener('change', onStoreChange);
+}
+
+function getSnapshot(): boolean {
+  return (
+    typeof window !== 'undefined' &&
+    typeof window.matchMedia === 'function' &&
+    window.matchMedia(QUERY).matches
+  );
+}
+
+function getServerSnapshot(): null {
+  return null;
+}
+
 /**
- * Tri-state on purpose: `null` until the first effect runs, so the server and
- * the hydrating client render both surfaces exactly as the CSS-only version
- * did. Only after hydration is the off-screen one dropped from the tree.
+ * Tri-state on purpose: the server snapshot is `null`, so server rendering and
+ * hydration keep both responsive surfaces mounted. The client snapshot takes
+ * over immediately after hydration and subscribes to viewport changes.
  */
 export function useWideViewport(): boolean | null {
-  const [wide, setWide] = useState<boolean | null>(null);
-  useEffect(() => {
-    if (typeof window === 'undefined' || typeof window.matchMedia !== 'function') return;
-    const mq = window.matchMedia(QUERY);
-    setWide(mq.matches);
-    const onChange = (e: MediaQueryListEvent) => setWide(e.matches);
-    mq.addEventListener('change', onChange);
-    return () => mq.removeEventListener('change', onChange);
-  }, []);
-  return wide;
+  return useSyncExternalStore<boolean | null>(subscribe, getSnapshot, getServerSnapshot);
 }
diff --git a/packages/app/src/components/quote-carousel.tsx b/packages/app/src/components/quote-carousel.tsx
deleted file mode 100644
index b0c46a8b8..000000000
--- a/packages/app/src/components/quote-carousel.tsx
+++ /dev/null
@@ -1,214 +0,0 @@
-'use client';
-
-import Link from 'next/link';
-import { useCallback, useEffect, useMemo, useRef, useState } from 'react';
-
-import { track } from '@/lib/analytics';
-import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
-import { CompanyLogo, highlightBrand } from '@/components/quotes/quote-utils';
-
-export interface CarouselQuote {
-  text: string;
-  name: string;
-  title: string;
-  org: string;
-  logo?: string;
-  link?: string;
-}
-
-export interface QuoteCarouselProps {
-  quotes: CarouselQuote[];
-  overrides?: {
-    /** Override display names in the org strip */
-    labels?: Record<string, string>;
-  };
-  /** Link to a page with all quotes */
-  moreHref?: string;
-  /** Label for the moreHref link (default "See more supporters →") */
-  moreLabel?: string;
-  /** Auto-rotate interval in ms (default 8000) */
-  intervalMs?: number;
-}
-
-interface CompanyEntry {
-  org: string;
-  quote: CarouselQuote;
-}
-
-// One entry per org, first quote wins; entries keep the order of the quotes array.
-function buildCompanyQuotes(quotes: CarouselQuote[]): CompanyEntry[] {
-  const byCompany = new Map<string, CarouselQuote[]>();
-  for (const q of quotes) {
-    const list = byCompany.get(q.org);
-    if (list) list.push(q);
-    else byCompany.set(q.org, [q]);
-  }
-  return [...byCompany.entries()].map(([org, pool]) => ({
-    org,
-    quote: pool[0],
-  }));
-}
-
-// Warm a logo into the browser cache so it paints instantly when its quote
-// becomes active. Called on hover/focus of an org button — i.e. only on user
-// intent, so the initial page load still fetches just the visible logo.
-function prefetchLogo(logo?: string) {
-  if (!logo || typeof window === 'undefined') return;
-  const img = new window.Image();
-  img.src = `/logos/${logo}`;
-}
-
-function QuoteBlock({ quote, renderLogo }: { quote: CarouselQuote; renderLogo: boolean }) {
-  return (
-    <blockquote className="w-full">
-      <p className="text-sm lg:text-base leading-relaxed text-muted-foreground italic">
-        &ldquo;{highlightBrand(quote.text)}&rdquo;
-      </p>
-      <footer className="mt-3 flex items-center gap-3">
-        {renderLogo ? (
-          <CompanyLogo org={quote.org} logo={quote.logo} />
-        ) : (
-          <div aria-hidden="true" className="size-10 shrink-0" />
-        )}
-        <div className="h-12 w-0.5 bg-brand" />
-        <div className="text-sm">
-          {quote.link ? (
-            <a
-              href={quote.link}
-              target="_blank"
-              rel="noopener noreferrer"
-              className="font-semibold text-foreground hover:text-brand transition-colors group"
-            >
-              <span className="group-hover:underline">{quote.name}</span>
-              <ExternalLinkIcon />
-            </a>
-          ) : (
-            <span className="font-semibold text-foreground">{quote.name}</span>
-          )}
-          <span className="block text-muted-foreground text-xs">{quote.title}</span>
-        </div>
-      </footer>
-    </blockquote>
-  );
-}
-
-export function QuoteCarousel({
-  quotes,
-  overrides = {},
-  moreHref,
-  moreLabel,
-  intervalMs = 8_000,
-}: QuoteCarouselProps) {
-  const { labels = {} } = overrides;
-
-  // Keep the first render deterministic so SSR reserves the carousel's full height before hydration.
-  const entries = useMemo(() => buildCompanyQuotes(quotes), [quotes]);
-  const [activeIndex, setActiveIndex] = useState(0);
-  const [fading, setFading] = useState(false);
-  const timerRef = useRef<ReturnType<typeof setInterval> | null>(null);
-  const fadeTimeoutRef = useRef<ReturnType<typeof setTimeout> | null>(null);
-  const hovering = useRef(false);
-
-  const advance = useCallback(() => {
-    if (hovering.current) return;
-    if (fadeTimeoutRef.current) clearTimeout(fadeTimeoutRef.current);
-    setFading(true);
-    fadeTimeoutRef.current = setTimeout(() => {
-      setActiveIndex((prev) => (prev + 1) % (entries.length || 1));
-      setFading(false);
-      fadeTimeoutRef.current = null;
-    }, 300);
-  }, [entries.length]);
-
-  // Auto-rotate
-  useEffect(() => {
-    if (entries.length <= 1) return;
-    timerRef.current = setInterval(advance, intervalMs);
-    return () => {
-      if (timerRef.current) clearInterval(timerRef.current);
-      if (fadeTimeoutRef.current) clearTimeout(fadeTimeoutRef.current);
-    };
-  }, [advance, entries.length, intervalMs]);
-
-  const goTo = useCallback(
-    (index: number) => {
-      if (timerRef.current) clearInterval(timerRef.current);
-      if (fadeTimeoutRef.current) clearTimeout(fadeTimeoutRef.current);
-      setFading(true);
-      fadeTimeoutRef.current = setTimeout(() => {
-        setActiveIndex(index);
-        setFading(false);
-        fadeTimeoutRef.current = null;
-      }, 300);
-      timerRef.current = setInterval(advance, intervalMs);
-      track('quote_carousel_navigated', {
-        toOrg: entries[index]?.org,
-        fromOrg: entries[activeIndex]?.org,
-      });
-    },
-    [advance, intervalMs, entries, activeIndex],
-  );
-
-  return (
-    <div
-      className="flex flex-col gap-4"
-      onMouseEnter={() => {
-        hovering.current = true;
-      }}
-      onMouseLeave={() => {
-        hovering.current = false;
-      }}
-    >
-      {/* Org name strip */}
-      <div className="flex flex-wrap justify-center gap-x-6 md:gap-x-8 gap-y-2 mx-4">
-        {entries.map((e, i) => (
-          <button
-            key={e.org}
-            type="button"
-            onClick={() => goTo(i)}
-            onMouseEnter={() => prefetchLogo(e.quote.logo)}
-            onFocus={() => prefetchLogo(e.quote.logo)}
-            className={`text-xs font-semibold tracking-wide uppercase transition-colors duration-200 ${
-              i === activeIndex ? 'text-foreground' : 'text-[#808488] hover:text-muted-foreground'
-            }`}
-          >
-            {labels[e.org] ?? e.org}
-          </button>
-        ))}
-      </div>
-
-      {/* All quotes stacked in same grid cell — tallest sets height */}
-      <div className="grid items-center">
-        {entries.map((e, i) => {
-          const isActive = i === activeIndex;
-          return (
-            <div
-              key={e.org}
-              className={`col-start-1 row-start-1 ${
-                isActive
-                  ? `transition-opacity duration-300 ease-in-out ${fading ? 'opacity-0' : 'opacity-100'}`
-                  : 'opacity-0 invisible pointer-events-none'
-              }`}
-              aria-hidden={!isActive}
-            >
-              <QuoteBlock quote={e.quote} renderLogo={isActive} />
-            </div>
-          );
-        })}
-      </div>
-
-      {moreHref && (
-        <div className="flex justify-end" data-testid="quote-carousel-more-row">
-          <Link
-            href={moreHref}
-            prefetch={false}
-            className="text-xs font-bold text-brand hover:underline"
-            onClick={() => track('quote_carousel_see_more_clicked')}
-          >
-            {moreLabel ?? 'See more supporters →'}
-          </Link>
-        </div>
-      )}
-    </div>
-  );
-}
diff --git a/packages/app/src/components/quotes/quotes-content.tsx b/packages/app/src/components/quotes/quotes-content.tsx
index d62ccd7c8..20bcdc962 100644
--- a/packages/app/src/components/quotes/quotes-content.tsx
+++ b/packages/app/src/components/quotes/quotes-content.tsx
@@ -10,7 +10,7 @@ import { track } from '@/lib/analytics';
 import type { Locale } from '@/lib/i18n';
 
 import { CompanyLogo, highlightBrand } from './quote-utils';
-import { QUOTES } from './quotes-data';
+import { QUOTES, orgAnchorId } from './quotes-data';
 
 const STRINGS = {
   en: {
@@ -20,22 +20,13 @@ const STRINGS = {
     jumpTo: (org: string) => `Jump to ${org}’s quote`,
   },
   zh: {
-    heading: <>InferenceX&trade; 计划支持者</>,
+    heading: <>业界与社区对 InferenceX&trade; 的评价</>,
     intro:
-      'InferenceX™ 计划获得众多主要算力买家与 ML 社区知名成员的支持，包括来自 MiniMax、Moonshot Kimi、阿里巴巴 Qwen、OpenAI、Microsoft、vLLM、PyTorch 基金会、Oracle 等机构的支持者。',
-    jumpTo: (org: string) => `跳转到 ${org} 的评价`,
+      'InferenceX™ 获得了众多大型算力采购方和 ML 社区知名人士的认可。以下评价来自 MiniMax、Moonshot Kimi、Alibaba Qwen、OpenAI、Microsoft、vLLM、PyTorch Foundation 和 Oracle 等机构的相关人士。',
+    jumpTo: (org: string) => `查看 ${org} 的评价`,
   },
 } as const;
 
-/** Stable anchor id for an org's quote (first occurrence wins). */
-function orgAnchorId(org: string): string {
-  const slug = org
-    .toLowerCase()
-    .replaceAll(/[^a-z0-9]+/gu, '-')
-    .replaceAll(/^-|-$/gu, '');
-  return `quote-${slug}`;
-}
-
 /** Deduplicated logos from all quote orgs. */
 const orgLogos: { org: string; logo: string }[] = [];
 const seenOrgs = new Set<string>();
@@ -149,7 +140,7 @@ export function QuotesContent({ locale = 'en' }: { locale?: Locale } = {}) {
                       }
                       text={locale === 'zh' ? (quote.textZh ?? quote.text) : quote.text}
                       name={quote.name}
-                      title={quote.title}
+                      title={locale === 'zh' ? (quote.titleZh ?? quote.title) : quote.title}
                       org={quote.org}
                       logo={quote.logo}
                       link={quote.link}
diff --git a/packages/app/src/components/quotes/quotes-data.test.ts b/packages/app/src/components/quotes/quotes-data.test.ts
index 29ccaf9b3..6a82abf6f 100644
--- a/packages/app/src/components/quotes/quotes-data.test.ts
+++ b/packages/app/src/components/quotes/quotes-data.test.ts
@@ -8,7 +8,13 @@ import { CAROUSEL_LABELS, CAROUSEL_ORGS, QUOTES } from './quotes-data';
  * supporter come off the carousel without being dropped from the site, so it is
  * asserted here rather than left implicit.
  */
-const OFF_CAROUSEL_ONLY = ['Together AI', 'Nebius', 'White House', 'UC San Diego'] as const;
+const OFF_CAROUSEL_ONLY = [
+  'Together AI',
+  'Nebius',
+  'White House',
+  'UC San Diego',
+  'TensorWave',
+] as const;
 
 describe('quote carousel membership', () => {
   it('keeps every carousel org backed by a real quote', () => {
@@ -22,6 +28,17 @@ describe('quote carousel membership', () => {
     expect(new Set(CAROUSEL_ORGS).size).toBe(CAROUSEL_ORGS.length);
   });
 
+  it('features Mooncake in the landing carousel', () => {
+    expect(CAROUSEL_ORGS).toContain('Mooncake');
+  });
+
+  it('features AMD and NVIDIA in the landing carousel under CEO labels', () => {
+    expect(CAROUSEL_ORGS).toContain('AMD');
+    expect(CAROUSEL_ORGS).toContain('NVIDIA');
+    expect(CAROUSEL_LABELS['AMD']).toBe('AMD Lisa Su');
+    expect(CAROUSEL_LABELS['NVIDIA']).toBe('NVIDIA Jensen');
+  });
+
   it.each(OFF_CAROUSEL_ONLY)('excludes %s from the carousel', (org) => {
     expect((CAROUSEL_ORGS as readonly string[]).includes(org)).toBe(false);
   });
@@ -38,3 +55,15 @@ describe('quote carousel membership', () => {
     }
   });
 });
+
+describe('Chinese quote attribution', () => {
+  it('provides a Chinese role for every quote with an English role', () => {
+    for (const quote of QUOTES) {
+      if (!quote.title) continue;
+      expect(
+        Reflect.get(quote, 'titleZh'),
+        `${quote.name} has an English role but no Chinese role`,
+      ).toBeTruthy();
+    }
+  });
+});
diff --git a/packages/app/src/components/quotes/quotes-data.ts b/packages/app/src/components/quotes/quotes-data.ts
index c79ce1315..a1cb13471 100644
--- a/packages/app/src/components/quotes/quotes-data.ts
+++ b/packages/app/src/components/quotes/quotes-data.ts
@@ -1,9 +1,20 @@
+/** Stable anchor id for an org's quote on the quotes page (first occurrence wins). */
+export function orgAnchorId(org: string): string {
+  const slug = org
+    .toLowerCase()
+    .replaceAll(/[^a-z0-9]+/gu, '-')
+    .replaceAll(/^-|-$/gu, '');
+  return `quote-${slug}`;
+}
+
 export interface Quote {
   text: string;
   /** Simplified Chinese translation of `text`, shown on /zh pages. */
   textZh?: string;
   name: string;
   title: string;
+  /** Simplified Chinese translation of `title`, shown on /zh pages. */
+  titleZh?: string;
   org: string;
   logo?: string;
   link?: string;
@@ -16,6 +27,7 @@ export const QUOTES: Quote[] = [
       '在模型与推理技术栈协同演进的今天，厂商中立、持续更新的基准测试不可或缺。MiniMax M3 在设计之初就兼顾了前沿能力与实际部署效率，而社区第一时间对 vLLM 的支持也体现了我们引以为豪的协作精神。InferenceX 正是生态所需的透明、可复现的数据平台。',
     name: 'Ryan Lee',
     title: 'Head of DevRel, MiniMax',
+    titleZh: 'MiniMax 开发者关系负责人',
     org: 'MiniMax',
     logo: 'minimax.svg',
     link: 'https://x.com/RyanLeeMiniMax',
@@ -56,6 +68,7 @@ export const QUOTES: Quote[] = [
       '在我们以前所未有的规模构建系统之际，机器学习社区拥有开放、透明的基准测试至关重要——它们真实反映了推理在不同硬件和软件上的表现。InferenceMAX™ 的对比基准测试穿透噪音，提供了关于 token 吞吐量、每美元性能和每兆瓦 token 数的动态全景。这种开源努力增强了整个生态，帮助从研究者到前沿数据中心运营者的每一个人做出更明智的决策。',
     name: 'Peter Hoeschele',
     title: 'VP of Infrastructure and Industrial Compute, OpenAI Stargate',
+    titleZh: 'OpenAI Stargate 基础设施与工业计算副总裁',
     org: 'OpenAI',
     logo: 'openai.svg',
     link: 'https://www.linkedin.com/in/peter-hoeschele/',
@@ -66,6 +79,7 @@ export const QUOTES: Quote[] = [
       'Azure 的使命是为客户提供性能最强、效率最高且最具成本效益的 AI 云。SemiAnalysis InferenceMAX™ 通过提供透明、可复现的基准测试来追踪各类 GPU 和软件栈在真实工作负载下的推理性能，有力地支持了这一使命。关于吞吐量、效率和每瓦成本的持续数据增强了我们优化 Azure 推理平台规模化的能力，帮助客户在 Microsoft Cloud 上自信构建。',
     name: 'Scott Guthrie',
     title: 'Executive Vice President, Microsoft Cloud & AI',
+    titleZh: 'Microsoft Cloud & AI 执行副总裁',
     org: 'Microsoft',
     logo: 'microsoft.svg',
     link: 'https://www.linkedin.com/in/guthriescott/',
@@ -76,6 +90,7 @@ export const QUOTES: Quote[] = [
       '在 Microsoft，为客户大规模交付最佳推理性能和经济性，需要深入理解 AI 模型如何与真实硬件和软件交互。像 InferenceMAX™ 这样的开源、可复现基准测试对于产出关于吞吐量、效率和成本的透明洞察至关重要。这些持续信号帮助指导我们的平台战略，使我们能够从芯片到系统再到软件对整个技术栈进行优化，让每一层协同工作，充分释放基础设施的潜力。',
     name: 'Saurabh Dighe',
     title: 'Corporate Vice President, Azure Strategic Planning & Architecture',
+    titleZh: 'Microsoft 公司副总裁、Azure 战略规划与架构负责人',
     org: 'Microsoft',
     logo: 'microsoft.svg',
     link: 'https://www.linkedin.com/in/saurabhdighe/',
@@ -86,6 +101,7 @@ export const QUOTES: Quote[] = [
       'PyTorch 基于一个信念而生：开放工具能加速整个 AI 生态。InferenceX™ 体现了同样的理念——开放、可复现、厂商中立的基准测试，为社区提供真实硬件上的真实数据。随着推理工作负载扩展到服务数十亿用户，在各类加速器上持续更新、透明的性能基线对于做出关键基础设施决策的从业者和平台团队而言不可或缺。',
     name: 'Joseph Spisak',
     title: 'Product Director, Meta Super Intelligence Lab',
+    titleZh: 'Meta Super Intelligence Lab 产品总监',
     org: 'Meta Superintelligence Labs',
     logo: 'meta.svg',
     link: 'https://www.linkedin.com/in/jspisak',
@@ -96,6 +112,7 @@ export const QUOTES: Quote[] = [
       'Oracle Cloud Infrastructure 旨在为前沿实验室和企业提供灵活性与选择，提供多种 GPU SKU 用于大规模 AI。InferenceMAX 通过提供开源、可复现的基准测试来支持这一使命，真实反映最新硬件和软件上的性能、效率与成本。凭借这种透明度，客户可以自信地选择与其 AI 战略最契合的平台。',
     name: 'Jay Jackson',
     title: 'Vice President, Oracle Cloud Infrastructure',
+    titleZh: 'Oracle Cloud Infrastructure 副总裁',
     org: 'Oracle',
     logo: 'oracle.svg',
     link: 'https://www.linkedin.com/in/jayejackson/',
@@ -106,6 +123,7 @@ export const QUOTES: Quote[] = [
       '理论峰值与实际推理吞吐量之间的差距往往取决于系统软件：推理引擎、分布式策略和底层内核。InferenceMAX™ 的价值在于它对最新软件进行基准测试，展示了 FP4、MTP、投机解码和 wide-EP 等优化在不同硬件上的实际效果。这种开放、可复现的结果帮助整个社区更快地前进。',
     name: 'Tri Dao',
     title: 'Chief Scientist of Together AI & Inventor of Flash Attention',
+    titleZh: 'Together AI 首席科学家、FlashAttention 发明者',
     org: 'Together AI',
     logo: 'together-ai.svg',
     link: 'https://tridao.me/',
@@ -116,15 +134,17 @@ export const QUOTES: Quote[] = [
       '行业需要大量公开、可复现的推理性能基准测试。vLLM 团队很高兴与 InferenceMAX™ 合作。更多元化的、人人可信赖和引用的工作负载与场景将推动生态向前发展。公平、透明的测量驱动着技术栈每一层的进步——从模型架构到推理引擎再到硬件。',
     name: 'Simon Mo',
     title: 'vLLM Project Co-Lead',
+    titleZh: 'vLLM 项目联合负责人',
     org: 'vLLM',
     logo: 'vllm.svg',
     link: 'https://www.linkedin.com/in/simon-mo-834217162/',
   },
   {
     text: 'InferenceMAX™ benchmark is pogchamp & W in chat',
-    textZh: 'InferenceMAX™ 基准测试绝绝子，大写的赢',
+    textZh: 'InferenceMAX™ 为推理基准测试树立了新标杆，是社区的一大胜利',
     name: 'Kaichao You',
     title: 'vLLM Project Co-Lead & PhD Student @ Tsinghua University',
+    titleZh: 'vLLM 项目联合负责人、清华大学博士生',
     org: 'vLLM',
     logo: 'vllm.svg',
     link: 'https://www.linkedin.com/in/youkaichao/',
@@ -134,6 +154,7 @@ export const QUOTES: Quote[] = [
     textZh: 'InferenceX 堪称当下最重要的开源基准测试套件',
     name: 'Mark Saroufim',
     title: 'GPU Mode Founder & Meta PyTorch Engineer',
+    titleZh: 'GPU Mode 创始人、Meta PyTorch 工程师',
     org: 'GPU Mode',
     logo: 'gpu-mode.png',
     link: 'https://x.com/marksaroufim',
@@ -144,6 +165,7 @@ export const QUOTES: Quote[] = [
       'InferenceMAX™ 展示了开放生态如何在实践中运作。vLLM、SGLang 和 TensorRT-LLM 等众多领先推理栈均构建于 PyTorch 之上，而这样的基准测试展示了内核、运行时和框架层面的创新如何转化为 NVIDIA 和 AMD GPU 等多种硬件平台上可衡量的性能。凭借开源属性和每夜运行，InferenceMAX™ 提供了一种透明的、社区驱动的方式来追踪进展，并为 PyTorch 用户提供数据驱动的洞察。',
     name: 'Matt White',
     title: 'Executive Director, PyTorch Foundation',
+    titleZh: 'PyTorch Foundation 执行董事',
     org: 'PyTorch Foundation',
     logo: 'pytorch.svg',
     link: 'https://www.linkedin.com/in/mdwdata/',
@@ -154,6 +176,7 @@ export const QUOTES: Quote[] = [
       'InferenceMAX™ 通过提供开放、透明的基准测试来追踪推理在最新 GPU 和软件栈上的实际表现，树立了新标杆。对客户而言，拥有衡量真实每美元 token 数和每瓦 token 数的可复现数据，将抽象的营销数字转化为可操作的洞察。CoreWeave 支持这一努力，因为它为这个快速发展的领域带来了清晰度，帮助整个生态自信构建。',
     name: 'Peter Salanki',
     title: 'CTO, CoreWeave',
+    titleZh: 'CoreWeave 首席技术官',
     org: 'CoreWeave',
     logo: 'coreweave.svg',
     link: 'https://www.linkedin.com/in/salanki/',
@@ -164,6 +187,7 @@ export const QUOTES: Quote[] = [
       'InferenceMAX™ 通过提供开放、透明的基准测试，揭示了推理在当今领先 GPU 和软件栈上的表现，树立了新标准。凭借衡量真实每美元 token 数和每瓦 token 数的可复现数据，客户可以超越营销宣传，获得可操作的洞察。对于作为全栈 AI 云服务商的 Nebius 而言，这一计划帮助我们自信地构建推理平台，并确保与生态保持一致。',
     name: 'Roman Chernin',
     title: 'Co-Founder & Chief Business Officer, Nebius',
+    titleZh: 'Nebius 联合创始人兼首席商务官',
     org: 'Nebius',
     logo: 'nebius.svg',
     link: 'https://www.linkedin.com/in/roman-chernin-1b4b8758/',
@@ -174,6 +198,7 @@ export const QUOTES: Quote[] = [
       '在 TensorWave，我们基于 AMD GPU 构建下一代云，因为我们相信当客户拥有强有力的替代方案时，创新才能蓬勃发展。InferenceMAX™ 通过提供开源、可复现的基准测试来追踪最新硬件和软件的吞吐量、效率与成本，强化了这一愿景。它穿透合成数据，突出真实推理性能，帮助客户看到 AMD 平台在大规模 AI 中的全部潜力。',
     name: 'Darrick Horton',
     title: 'CEO, TensorWave',
+    titleZh: 'TensorWave 首席执行官',
     org: 'TensorWave',
     logo: 'tensorwave.svg',
     link: 'https://www.linkedin.com/in/darrick-horton/',
@@ -184,6 +209,7 @@ export const QUOTES: Quote[] = [
       'SGLang 是 xAI Grok 等众多生产级推理工厂背后的推理引擎，被誉为推理之王。在大规模场景中，我们深刻体会到性能在不同硬件、模型和配置间的巨大差异。InferenceX™ 每夜在所有主流 GPU 平台上对 SGLang 进行基准测试，以其他基准测试无法做到的方式——持续且可复现地——捕捉这种差异。',
     name: 'Mingyi Lu',
     title: 'SGLang Product Lead',
+    titleZh: 'SGLang 产品负责人',
     org: 'SGLang',
     logo: 'sglang.webp',
     link: 'https://www.linkedin.com/in/mingyi-lu/',
@@ -194,6 +220,7 @@ export const QUOTES: Quote[] = [
       'InferenceX™ 恰好体现了这一点——开放、可复现的基准测试，随着 xPU 加速器（GPU/TPU/LPU）、内存、存储和软件栈的演进而持续更新。我很高兴看到 InferenceX 基准测试路线图纳入了对 CPU KV Cache 卸载乃至即将到来的 NVMe KV Cache 卸载施压的智能体编程工作负载。WEKA 通过构建为这些 xPU 供给的 KV Cache 基础设施来帮助突破内存墙，拥有这种对推理性能的深度可见性有助于整个生态做出更明智的投资决策。',
     name: 'Val Bercovici',
     title: 'Chief AI Officer, WEKA',
+    titleZh: 'WEKA 首席 AI 官',
     org: 'WEKA',
     logo: 'weka.svg',
     link: 'https://www.linkedin.com/in/valentinbercovici/',
@@ -204,6 +231,7 @@ export const QUOTES: Quote[] = [
       '对于从事推理优化的研究者而言，理解新技术如何在软硬件栈中交互至关重要，却极难衡量。InferenceX™ 提供了亟需的洞察，展示了推理性能在各主要硬件平台上的演进轨迹，以开放、可复现的数据让差距与进展清晰可见，推动了该领域的发展。',
     name: 'Simon Guo',
     title: 'PhD Student, Stanford CS',
+    titleZh: 'Stanford CS 博士生',
     org: 'Stanford',
     logo: 'stanford.svg',
     link: 'https://simonguo.tech/',
@@ -214,6 +242,7 @@ export const QUOTES: Quote[] = [
       'Hugging Face 的存在是为了让 AI 对每个人都开放且可及。InferenceX™ 将这一使命延伸到 AI 芯片性能领域，直接从 Hub 拉取模型，在所有主流加速器上持续、透明地进行基准测试。当社区能够实时看到前沿开源模型在真实硬件上的确切表现时，整个生态的标准都将被提升。',
     name: 'Clement Delangue',
     title: 'CEO, Hugging Face',
+    titleZh: 'Hugging Face 首席执行官',
     org: 'Hugging Face',
     logo: 'huggingface.svg',
     link: 'https://www.linkedin.com/in/cdelangue/',
@@ -224,6 +253,7 @@ export const QUOTES: Quote[] = [
       'Lambda 致力于让 GPU 算力对 AI 团队——从个人研究者到大型实验室——都简单易用。InferenceX™ 通过为社区提供开放、可复现的基准测试来衡量真正重要的指标：真实吞吐量、成本效率以及最新硬件和软件栈上的每瓦性能，与这一使命高度契合。团队可以基于透明、持续更新的数据做出明智的算力选择。',
     name: 'Stephen Balaban',
     title: 'Co-founder and CEO, Lambda',
+    titleZh: 'Lambda 联合创始人兼首席执行官',
     org: 'Lambda',
     logo: 'lambda.svg',
     link: 'https://www.linkedin.com/in/sbalaban/',
@@ -234,6 +264,7 @@ export const QUOTES: Quote[] = [
       '当我们推出 DistServe 时，核心论点很简单：将预填充和解码分离，分别优化。十八个月后，解聚已成为行业默认架构。InferenceX™ 是在整条帕累托曲线上对比解聚与聚合服务的基准测试。InferenceX 精确展示了 P/D 分离在 TTFT、TPOT、吞吐量和成本方面何时何地带来收益。',
     name: 'Hao Zhang',
     title: 'Assistant Professor, UC San Diego & Co-Creator of DistServe, vLLM, and FastVideo',
+    titleZh: 'UC San Diego 助理教授、DistServe、vLLM 和 FastVideo 联合创建者',
     org: 'UC San Diego',
     logo: 'uc-san-diego.svg',
     link: 'https://haozhang.ai/',
@@ -243,6 +274,7 @@ export const QUOTES: Quote[] = [
     textZh: '这基准测试真不错',
     name: 'Michael Goin',
     title: 'vLLM Core Maintainer & Senior Principal Engineer at Red Hat',
+    titleZh: 'vLLM 核心维护者、Red Hat 资深首席工程师',
     org: 'Red Hat',
     logo: 'redhat.svg',
     link: 'https://www.linkedin.com/in/michael-goin/',
@@ -252,6 +284,7 @@ export const QUOTES: Quote[] = [
     textZh: '现在经常听到"我们想要X领域的 SemiAnalysis"。这是对 @dylan522p 所构建之物的最好证明。',
     name: 'Sriram Krishnan',
     title: 'White House Senior AI Advisor',
+    titleZh: '白宫高级 AI 顾问',
     org: 'White House',
     logo: 'white-house.svg',
     link: 'https://x.com/sriramk/status/2048824255702262135',
@@ -262,6 +295,7 @@ export const QUOTES: Quote[] = [
       '高端推理（Premium Inference）正在成为一个全新品类，其驱动力来自智能体：它们需要快速、可交互的 token，并要求在前沿智能模型上实现最佳的单芯片吞吐量。InferenceX™ 是少数在所有主流加速器上公开衡量这类芯片性能的基准测试之一。我们期待以 SN50 参与官方 InferenceX，在透明追踪之下展示我们在前沿开源模型上的高端解码性能。',
     name: 'Rodrigo Liang',
     title: 'Co-Founder & CEO, SambaNova',
+    titleZh: 'SambaNova 联合创始人兼首席执行官',
     org: 'SambaNova',
     logo: 'sambanova.webp',
     link: 'https://www.linkedin.com/in/rodrigo-liang/',
@@ -282,6 +316,7 @@ export const QUOTES: Quote[] = [
       '开放协作正在推动 AI 创新的下一个时代。开源的 InferenceMAX 基准测试为社区提供透明的每夜结果，激发信任并加速进步。它突出了我们 AMD Instinct MI300、MI325X 和 MI355X GPU 在多样化工作负载中极具竞争力的 TCO 表现，彰显了我们平台的实力以及我们致力于让开发者实时了解软件进展的承诺。',
     name: 'Dr. Lisa Su',
     title: 'Chair and CEO, AMD',
+    titleZh: 'AMD 董事长兼首席执行官',
     org: 'AMD',
     logo: 'amd.svg',
     link: 'https://www.linkedin.com/in/lisasu-amd/',
@@ -292,6 +327,7 @@ export const QUOTES: Quote[] = [
       '推理需求在长上下文推理的驱动下呈指数级增长。NVIDIA Grace Blackwell NVL72 正是为这个思考型 AI 的新时代而生。NVIDIA 通过持续的硬件和软件创新来满足这一需求，推动 AI 的下一步发展。通过高频基准测试，InferenceMAX™ 为行业提供了 LLM 推理在真实工作负载上性能的透明视角。结果一目了然：Grace Blackwell NVL72 搭配 TRT-LLM 和 Dynamo 提供了无与伦比的每美元性能和每兆瓦性能——驱动着全球最高效、最具成本效益的 AI 工厂。',
     name: 'Jensen Huang',
     title: 'Founder & CEO, NVIDIA',
+    titleZh: 'NVIDIA 创始人兼首席执行官',
     org: 'NVIDIA',
     logo: 'nvidia.svg',
     link: 'https://www.linkedin.com/in/jenhsunhuang/',
@@ -302,6 +338,7 @@ export const QUOTES: Quote[] = [
       '速度就是护城河。InferenceMAX™ 的每夜基准测试与 AMD 软件栈的改进速度同步。看到 AMD MI300、MI325 和 MI355 GPU 在多样化工作负载和交互级别上表现如此出色，令人振奋。',
     name: 'Anush Elangovan',
     title: 'VP GPU Software, AMD',
+    titleZh: 'AMD GPU 软件副总裁',
     org: 'AMD',
     logo: 'amd.svg',
     link: 'https://www.linkedin.com/in/anushelangovan/',
@@ -312,6 +349,7 @@ export const QUOTES: Quote[] = [
       'InferenceMAX™ 聚焦机器学习社区关注的工作负载。在 NVIDIA，我们欢迎这些对比，因为它们凸显了我们全栈方案的优势——从 GPU 硬件到 NVLink 网络，到 NVL72 机架级系统，再到 Dynamo 解聚服务，持续提供业界领先的推理性能和大规模投资回报率。',
     name: 'Ian Buck',
     title: 'VP & GM, Hyperscale, NVIDIA & Inventor of CUDA',
+    titleZh: 'NVIDIA 超大规模计算业务副总裁兼总经理、CUDA 发明者',
     org: 'NVIDIA',
     logo: 'nvidia.svg',
     link: 'https://www.linkedin.com/in/ian-buck-19201315/',
@@ -322,6 +360,7 @@ export const QUOTES: Quote[] = [
       'InferenceMAX™ 的每夜结果突出展示了 AMD 软件栈的快速进步。能够见证一个开源项目的诞生令人兴奋——它在 AMD 软件团队的工作与其对我们 MI300、MI325 和 MI355 GPU 上特定机器学习用例的影响之间建立了紧密的反馈闭环。我期待看到 InferenceMAX 的下一步发展，并展示 AMD 平台的能力。AMD GPU 将持续每周变得更快。',
     name: 'Quentin Colombet',
     title: 'Senior Director, AMD, Ex-Brium CEO',
+    titleZh: 'AMD 高级总监、Brium 前首席执行官',
     org: 'AMD',
     logo: 'amd.svg',
     link: 'https://www.linkedin.com/in/quentincolombet/',
@@ -332,6 +371,7 @@ export const QUOTES: Quote[] = [
       '在 Crusoe，我们相信成为优秀合作伙伴意味着赋予客户选择权和清晰度。这就是我们自豪地支持 InferenceMAX™ 的原因——它为整个 AI 社区提供最新硬件上开源、可复现的基准测试。通过提供关于吞吐量、效率和成本的透明真实数据，InferenceMAX™ 穿透炒作，帮助客户自信地为其独特工作负载选择最佳平台。',
     name: 'Chase Lochmiller',
     title: 'Co-Founder & CEO, Crusoe',
+    titleZh: 'Crusoe 联合创始人兼首席执行官',
     org: 'Crusoe',
     logo: 'crusoe.svg',
     link: 'https://www.linkedin.com/in/chase-lochmiller-604483341/',
@@ -342,6 +382,7 @@ export const QUOTES: Quote[] = [
       'Supermicro 对 InferenceMAX™ 的发布感到振奋——这是 SemiAnalysis 的基准测试系统，衡量真实吞吐量、每美元性能和能效。这一开源工具在最新硬件和软件上提供可复现的基准测试，帮助 AI 实验室和企业在大规模场景中选择最佳平台。',
     name: 'Charles Liang',
     title: 'Founder & CEO, Supermicro',
+    titleZh: 'Supermicro 创始人兼首席执行官',
     org: 'Supermicro',
     logo: 'supermicro.svg',
     link: 'https://en.wikipedia.org/wiki/Charles_Liang',
@@ -352,6 +393,7 @@ export const QUOTES: Quote[] = [
       'Vultr 致力于提供一个开放生态，让开发者自由选择如何构建和扩展 AI——无论是在 NVIDIA 还是 AMD GPU 上。借助 InferenceMAX™，客户获得开放、可复现的基准测试，对前沿硬件和软件的吞吐量、效率与成本提供清晰洞察。通过展示真实性能，我们赋能团队自信地为其 AI 工作负载选择合适的平台。',
     name: 'Nathan Goulding',
     title: 'SVP of Engineering, Vultr',
+    titleZh: 'Vultr 工程高级副总裁',
     org: 'Vultr',
     logo: 'vultr.svg',
     link: 'https://www.linkedin.com/in/nathangoulding/',
@@ -362,6 +404,7 @@ export const QUOTES: Quote[] = [
       '在 Prime Intellect，我们正在推动 AI 后训练和开放研究的前沿。InferenceX™ 通过提供开放、可复现的基准测试来追踪推理性能在不断演进的硬件和软件栈上的真实表现，与我们的工作形成互补。对于像我们这样的研究者，拥有关于吞吐量和效率的透明、持续更新的数据意味着我们可以专注于构建更好的模型，而不必为基础设施选择纠结。这正是加速每个人进步的社区驱动力量。',
     name: 'Jack Min Ong',
     title: 'Researcher, Prime Intellect',
+    titleZh: 'Prime Intellect 研究员',
     org: 'Prime Intellect',
     logo: 'prime-intellect.svg',
     link: 'https://www.linkedin.com/in/jackminong/',
@@ -372,6 +415,7 @@ export const QUOTES: Quote[] = [
       '在 Firmus，我们正在建造全球最节能的 AI 工厂——而效率只有在可衡量时才有意义。InferenceX™ 为行业提供开放、可复现的基准测试，追踪最新 GPU 平台和软件栈上的真实吞吐量、成本和每瓦性能。随着我们在亚太和澳洲扩展吉瓦级可再生能源驱动的 AI 基础设施，这种透明、持续更新的数据帮助整个生态了解这些系统的实际交付能力。',
     name: 'Tim Rosenfield',
     title: 'Co-Founder & Co-CEO, Firmus',
+    titleZh: 'Firmus 联合创始人兼联席首席执行官',
     org: 'Firmus',
     logo: 'firmus.svg',
     link: 'https://www.linkedin.com/in/tim-rosenfield-a735a4112',
@@ -381,6 +425,7 @@ export const QUOTES: Quote[] = [
     textZh: 'InferenceMAX 对我们很有用，即使 Dylan Patel 是个有感情的可爱小伙子',
     name: 'Matthew Leavitt',
     title: 'Chief Science Officer, DatologyAI',
+    titleZh: 'DatologyAI 首席科学官',
     org: 'DatologyAI',
     logo: 'datologyai.svg',
     link: 'https://www.linkedin.com/in/matthew-leavitt-6797703b/',
@@ -391,6 +436,7 @@ export const QUOTES: Quote[] = [
       'InferenceX™ 提供了社区所需的开源测量——真实工作负载、真实硬件和真实软件栈上的每夜结果。作为一位大量撰写过理论性能与实际系统性能差距的人，我很高兴看到一个让这种差距对每个人都清晰可见、可追踪的项目。',
     name: 'Stas Bekman',
     title: 'Developer & Author of Machine Learning Engineering Open Book (17.5K+ ⭐)',
+    titleZh: '开发者、《Machine Learning Engineering Open Book》作者（17.5K+ ⭐）',
     org: 'Stas Bekman',
     link: 'https://github.com/stas00/ml-engineering',
   },
@@ -400,6 +446,7 @@ export const QUOTES: Quote[] = [
       '我们在 Adaptive ML 自己也使用 InferenceX 基准测试作为帮助我们做出基础设施决策的关键数据点之一。推理性能对于大规模强化学习工作负载至关重要，快速生成直接影响客户的上市时间和收入。InferenceX™ 持续对全栈进行基准测试——引擎、模型、软件和硬件，覆盖 GB300 NVL72 等机架级系统。这正是生态一直缺少的那种开放、透明、可复现的信号。',
     name: 'Julien Launay',
     title: 'Co-Founder & CEO, Adaptive ML',
+    titleZh: 'Adaptive ML 联合创始人兼首席执行官',
     org: 'Adaptive ML',
     logo: 'adaptive-ml.svg',
     link: 'https://www.linkedin.com/in/julienlaunay/',
@@ -410,6 +457,7 @@ export const QUOTES: Quote[] = [
       '我们的客户使用前沿开源模型将 AI 投入生产——在大规模场景中，每秒每个 token 和每百万 token 的每一美元都至关重要。InferenceX™ 为生态提供了我们一直需要的东西：一个客观、开放的基准测试，持续追踪 GB300 NVL72、GB200 NVL72、H100 以及即将到来的 Rubin、TPU 和 Trainium 等硬件上的真实推理性能。这对帮助更广泛的社区理解行业格局并建立清晰的性能分类体系非常有价值。',
     name: 'Alex Ker',
     title: 'Engineer, Baseten',
+    titleZh: 'Baseten 工程师',
     org: 'Baseten',
     logo: 'baseten.svg',
     link: 'https://www.linkedin.com/in/alex-ker/',
@@ -420,6 +468,7 @@ export const QUOTES: Quote[] = [
       '我们创立 Verda 是为了让 AI 工程师无障碍地使用前沿算力，没有门槛。InferenceX 通过为 AI 构建者提供开放、可复现的基准测试来支持这一使命，展示 GPU 在真实推理工作负载下的实际交付能力。我们希望客户看到透明、持续更新的性能数据，没有营销虚辞。InferenceX 恰好提供了这一切。',
     name: 'Ruben Bryon',
     title: 'Founder & CEO, Verda',
+    titleZh: 'Verda 创始人兼首席执行官',
     org: 'Verda',
     logo: 'verda.svg',
     link: 'https://www.linkedin.com/in/ruben-bryon/',
@@ -430,6 +479,7 @@ export const QUOTES: Quote[] = [
       'Voltage Park 旨在为 AI 团队提供快速、经济的大规模 GPU 算力。InferenceX™ 通过提供开放、可复现的基准测试来展示推理在最新硬件和软件栈上的实际表现，有力支持了这一目标。凭借关于吞吐量、效率和成本的透明、持续更新数据，团队可以自信地做出算力决策而非凭空猜测。我们很高兴支持一项为生态带来如此清晰度的工作。',
     name: 'Saurabh Giri',
     title: 'CTO, Voltage Park',
+    titleZh: 'Voltage Park 首席技术官',
     org: 'Voltage Park',
     logo: 'voltage-park.svg',
     link: 'https://www.linkedin.com/in/saurabh-giri/',
@@ -440,6 +490,7 @@ export const QUOTES: Quote[] = [
       '在 Periodic Labs，我们正在构建将算力转化为真实科学发现的 AI 科学家。这意味着我们非常关注每块 GPU 的实际交付能力。InferenceX™ 提供开放、可复现的基准测试，穿透规格表，展示最新硬件和软件栈上的真实吞吐量、效率与成本。在数千块 GPU 上做过推理后，我可以说这种透明、持续更新的数据正是从业者做出明智基础设施决策所需要的。',
     name: 'Xander Dunn',
     title: 'Founding Team, Periodic Labs',
+    titleZh: 'Periodic Labs 创始团队成员',
     org: 'Periodic Labs',
     logo: 'periodic-labs.png',
     link: 'https://www.linkedin.com/in/xanderdunn/',
@@ -450,6 +501,7 @@ export const QUOTES: Quote[] = [
       '随着 AI 基础设施在全球范围内扩展，没有任何单一厂商或地区能够定义适用于所有人的基准测试。InferenceX 是朝着共享、透明的推理性能和 TCO 视角迈出的重要一步，为主权 AI 云运营商带来更理性的投资决策、更健康的竞争，并最终在全球范围内提供更可及的 AI 算力。',
     name: 'Talal M. Al Kaissi',
     title: 'CEO',
+    titleZh: '首席执行官',
     org: 'Core42',
     logo: 'core42.webp',
   },
@@ -459,6 +511,7 @@ export const QUOTES: Quote[] = [
       '拥有一个开放且持续更新的平台来对推理引擎在真实工作负载和多样化硬件上进行基准测试非常重要。InferenceX 提供了这种透明、实用的评估，帮助社区更好地理解真实系统瓶颈和权衡。这样的基准测试对于构建更高效、更可扩展的 AI 系统至关重要。此外，随着 LLM 智能体在改进系统方面日益强大，这样的平台可以提供闭合自动优化循环所需的可靠反馈，进一步推动该领域的进步。',
     name: 'Cao Shiyi',
     title: 'Researcher, Sky Computing Lab',
+    titleZh: 'Sky Computing Lab 研究员',
     org: 'UC Berkeley',
     logo: 'sky-berkeley.webp',
   },
@@ -468,6 +521,7 @@ export const QUOTES: Quote[] = [
       '在 GMI Cloud，我们认为推理已成为 AI 价值创造的核心。SemiAnalysis 通过 InferenceX 做了行业期盼已久的事——将推理从一个黑箱变成了一个被持续衡量的真实系统。InferenceX 不仅对硬件进行基准测试，还覆盖完整技术栈——模型、运行时和分布式系统，反映的是 AI 在生产中的实际运行方式，而非营销宣传。',
     name: 'Alex Yeh',
     title: 'Founder & CEO, GMI Cloud',
+    titleZh: 'GMI Cloud 创始人兼首席执行官',
     org: 'GMI Cloud',
     logo: 'gmi-cloud.svg',
     link: 'https://www.linkedin.com/in/gmi-yeh',
@@ -478,6 +532,7 @@ export const QUOTES: Quote[] = [
       '在 EmbeddedLLM，我们的团队深耕于生产推理栈，包括 vLLM 的核心维护和贡献工作，因此我们每天都能看到真实 AI 性能在多大程度上取决于完整系统：模型、运行时、内核、调度和硬件。InferenceX™ 之所以重要，是因为它持续且公开地对完整系统进行基准测试。它将推理从营销话题转变为工程学科，为 AI 实验室、新型云服务商和企业提供在生产规模上做出吞吐量、成本和效率决策所需的数据。',
     name: 'Pin Siang Tan',
     title: 'Co-founder & CTO, EmbeddedLLM',
+    titleZh: 'EmbeddedLLM 联合创始人兼首席技术官',
     org: 'EmbeddedLLM',
     logo: 'embeddedllm.webp',
     link: 'https://www.linkedin.com/in/tanpinsiang',
@@ -488,6 +543,7 @@ export const QUOTES: Quote[] = [
       '智能体工作负载让上下文缓存的价值以非常具体的方式显现出来。相同的工具、指令、文档和对话状态会在许多步骤中被反复使用，而每一次重新计算的前缀都会增加延迟和 GPU 成本。Tensormesh 的目标就是让这种复用在生产环境中可见、可用——无论团队是通过 OpenAI 兼容 API 还是在预留算力上运行。InferenceX 的价值在于它按照这些工作负载的真实行为进行测试：长会话、重复上下文，以及足够的缓存压力，足以检验服务系统是否做对了。',
     name: 'Kuntai Du',
     title: 'Chief Scientist at Tensormesh & Co-Creator of LMCache',
+    titleZh: 'Tensormesh 首席科学家、LMCache 联合创建者',
     org: 'Tensormesh',
     logo: 'tensormesh.svg',
     link: 'https://www.linkedin.com/in/kuntai-du/',
@@ -498,10 +554,21 @@ export const QUOTES: Quote[] = [
       'LMCache 源于一个简单的观察：KV cache 应该由服务系统统一管理，而不是被当作单个引擎内部用完即弃的状态。智能体轨迹让这一点尤为清晰：它们会产生重复前缀、共享上下文、长对话，以及在 GPU 显存、主机内存、存储和其他 worker 之间迁移缓存的机会。InferenceX 让社区能够直接测量这些行为——包括 KV 卸载、缓存共享和预填充效率——而所用的工作负载远比固定提示词基准测试更接近真实的智能体使用场景。',
     name: 'Yihua Cheng',
     title: 'Co-Creator of LMCache & CTO at Tensormesh',
+    titleZh: 'LMCache 联合创建者、Tensormesh 首席技术官',
     org: 'LMCache',
     logo: 'lmcache.webp',
     link: 'https://www.linkedin.com/in/yihuacheng-215133327/',
   },
+  {
+    text: 'Mooncake makes KV cache a first-class system resource for disaggregated LLM serving. Real-world performance depends on the full configuration model, hardware, prefill/decode topology, parallelism strategy, and workload. InferenceX™ tests Mooncake across a broad and growing range of configurations, openly and reproducibly, giving the community a clear view of the tradeoffs and helping us improve the system faster.',
+    textZh:
+      'Mooncake 将 KV cache 作为分离式 LLM 推理中的一等系统资源。真实性能取决于完整的配置组合——模型、硬件、预填充/解码拓扑、并行策略和工作负载。InferenceX™ 以开放、可复现的方式，在持续扩展的丰富配置上测试 Mooncake，让社区清晰了解其中的权衡，也帮助我们更快地改进系统。',
+    name: 'Mooncake Team',
+    title: '',
+    org: 'Mooncake',
+    logo: 'kvcache-ai.svg',
+    link: 'https://kvcache.ai/',
+  },
 ];
 
 /**
@@ -515,6 +582,8 @@ export const CAROUSEL_ORGS = [
   'Alibaba Qwen',
   'Zhipu GLM',
   'OpenAI',
+  'AMD',
+  'NVIDIA',
   'Microsoft',
   'Meta Superintelligence Labs',
   'Oracle',
@@ -522,7 +591,6 @@ export const CAROUSEL_ORGS = [
   'GPU Mode',
   'PyTorch Foundation',
   'CoreWeave',
-  'TensorWave',
   'SGLang',
   'WEKA',
   'Stanford',
@@ -531,6 +599,7 @@ export const CAROUSEL_ORGS = [
   'Red Hat',
   'SambaNova',
   'TileRT',
+  'Mooncake',
 ] as const;
 
 /**
@@ -545,4 +614,6 @@ export const CAROUSEL_LABELS: Record<string, string> = {
   'PyTorch Foundation': 'PyTorch',
   'Meta Superintelligence Labs': 'Meta',
   'Moonshot AI': 'Moonshot Kimi',
+  AMD: 'AMD Lisa Su',
+  NVIDIA: 'NVIDIA Jensen',
 };
diff --git a/packages/app/src/components/reliability/ReliabilityContext.test.ts b/packages/app/src/components/reliability/ReliabilityContext.test.ts
new file mode 100644
index 000000000..f97365b3c
--- /dev/null
+++ b/packages/app/src/components/reliability/ReliabilityContext.test.ts
@@ -0,0 +1,110 @@
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+import type { ReliabilityRow } from '@/lib/api';
+
+import { aggregateByDateRange } from './ReliabilityContext';
+
+const NOW = new Date('2026-08-20T12:00:00.000Z');
+
+function row(hardware: string, date: string, n_success = 1, total = 2): ReliabilityRow {
+  return { hardware, date, n_success, total };
+}
+
+beforeEach(() => {
+  vi.useFakeTimers();
+  vi.setSystemTime(NOW);
+});
+
+afterEach(() => {
+  vi.useRealTimers();
+});
+
+describe('aggregateByDateRange', () => {
+  it('preserves the existing fixture totals and rounded rates in every window', () => {
+    const result = aggregateByDateRange([
+      row('a', '2026-08-20T12:00:00.000Z', 3, 4),
+      row('a', '2026-08-17T12:00:00.000Z', 1, 2),
+      row('a', '2026-08-17T11:59:59.999Z', 2, 2),
+      row('a', '2026-08-13T12:00:00.000Z', 0, 2),
+      row('a', '2026-07-21T12:00:00.000Z', 4, 4),
+      row('a', '2026-05-22T12:00:00.000Z', 1, 4),
+      row('a', '2026-05-22T11:59:59.999Z', 2, 4),
+      row('b', '2026-08-19T12:00:00.000Z', 5, 10),
+      row('zero-total', '2026-08-20T12:00:00.000Z', 0, 0),
+    ]);
+
+    expect(result).toEqual({
+      'last-3-days': {
+        a: { rate: 66.67, total: 6, n_success: 4 },
+        b: { rate: 50, total: 10, n_success: 5 },
+      },
+      'last-7-days': {
+        a: { rate: 60, total: 10, n_success: 6 },
+        b: { rate: 50, total: 10, n_success: 5 },
+      },
+      'last-month': {
+        a: { rate: 71.43, total: 14, n_success: 10 },
+        b: { rate: 50, total: 10, n_success: 5 },
+      },
+      'last-3-months': {
+        a: { rate: 61.11, total: 18, n_success: 11 },
+        b: { rate: 50, total: 10, n_success: 5 },
+      },
+      'all-time': {
+        a: { rate: 59.09, total: 22, n_success: 13 },
+        b: { rate: 50, total: 10, n_success: 5 },
+      },
+    });
+  });
+
+  it('includes exact cutoffs and excludes dates one millisecond before them', () => {
+    const result = aggregateByDateRange([
+      row('exact-3', '2026-08-17T12:00:00.000Z'),
+      row('before-3', '2026-08-17T11:59:59.999Z'),
+      row('date-only-3', '2026-08-17'),
+      row('exact-7', '2026-08-13T12:00:00.000Z'),
+      row('before-7', '2026-08-13T11:59:59.999Z'),
+      row('exact-30', '2026-07-21T12:00:00.000Z'),
+      row('before-30', '2026-07-21T11:59:59.999Z'),
+      row('exact-90', '2026-05-22T12:00:00.000Z'),
+      row('before-90', '2026-05-22T11:59:59.999Z'),
+    ]);
+
+    expect(Object.keys(result['last-3-days'])).toEqual(['exact-3']);
+    expect(Object.keys(result['last-7-days'])).toEqual([
+      'exact-3',
+      'before-3',
+      'date-only-3',
+      'exact-7',
+    ]);
+    expect(Object.keys(result['last-month'])).toEqual([
+      'exact-3',
+      'before-3',
+      'date-only-3',
+      'exact-7',
+      'before-7',
+      'exact-30',
+    ]);
+    expect(Object.keys(result['last-3-months'])).toEqual([
+      'exact-3',
+      'before-3',
+      'date-only-3',
+      'exact-7',
+      'before-7',
+      'exact-30',
+      'before-30',
+      'exact-90',
+    ]);
+    expect(Object.keys(result['all-time'])).toEqual([
+      'exact-3',
+      'before-3',
+      'date-only-3',
+      'exact-7',
+      'before-7',
+      'exact-30',
+      'before-30',
+      'exact-90',
+      'before-90',
+    ]);
+  });
+});
diff --git a/packages/app/src/components/reliability/ReliabilityContext.tsx b/packages/app/src/components/reliability/ReliabilityContext.tsx
index a157c38e0..cfa18751c 100644
--- a/packages/app/src/components/reliability/ReliabilityContext.tsx
+++ b/packages/app/src/components/reliability/ReliabilityContext.tsx
@@ -7,13 +7,14 @@ import {
   useContext,
   useEffect,
   useMemo,
+  useRef,
   useState,
 } from 'react';
 
 import {
+  resolveAvailableSelection,
   useChartUIState,
   useChartToggleSet,
-  useAutoInitializeToggleSet,
   useUrlStateSync,
 } from '@/hooks/useChartContext';
 import { useReliability } from '@/hooks/api/use-reliability';
@@ -30,31 +31,36 @@ import type {
 /** @internal Exported for test provider wrapping only. */
 export const ReliabilityContext = createContext<ReliabilityChartContextType | undefined>(undefined);
 
-/** Aggregate raw reliability rows into date-range buckets. */
-function aggregateByDateRange(rows: ReliabilityRow[]): DateRangeSuccessRateData {
-  const now = new Date();
-  const cutoffs: Record<string, Date | null> = {
-    'last-3-days': new Date(now.getTime() - 3 * 86400000),
-    'last-7-days': new Date(now.getTime() - 7 * 86400000),
-    'last-month': new Date(now.getTime() - 30 * 86400000),
-    'last-3-months': new Date(now.getTime() - 90 * 86400000),
-    'all-time': null,
-  };
+/** @internal Aggregate raw reliability rows into date-range buckets. */
+export function aggregateByDateRange(rows: ReliabilityRow[]): DateRangeSuccessRateData {
+  const now = Date.now();
+  const ranges = [
+    ['last-3-days', now - 3 * 86400000],
+    ['last-7-days', now - 7 * 86400000],
+    ['last-month', now - 30 * 86400000],
+    ['last-3-months', now - 90 * 86400000],
+    ['all-time', null],
+  ] as const;
+  const aggregates = Object.fromEntries(
+    ranges.map(([range]) => [range, {} as Record<string, { n_success: number; total: number }>]),
+  ) as Record<(typeof ranges)[number][0], Record<string, { n_success: number; total: number }>>;
 
-  const result: DateRangeSuccessRateData = {};
-
-  for (const [range, cutoff] of Object.entries(cutoffs)) {
-    const agg: Record<string, { n_success: number; total: number }> = {};
-    for (const row of rows) {
-      if (cutoff && new Date(row.date) < cutoff) continue;
-      if (!agg[row.hardware]) agg[row.hardware] = { n_success: 0, total: 0 };
-      agg[row.hardware].n_success += row.n_success;
-      agg[row.hardware].total += row.total;
+  for (const row of rows) {
+    const rowTime = new Date(row.date).getTime();
+    for (const [range, cutoff] of ranges) {
+      if (cutoff !== null && rowTime < cutoff) continue;
+      const stats = (aggregates[range][row.hardware] ??= { n_success: 0, total: 0 });
+      stats.n_success += row.n_success;
+      stats.total += row.total;
     }
+  }
+
+  const result: DateRangeSuccessRateData = {};
+  for (const [range] of ranges) {
     result[range] = {};
-    for (const [hw, stats] of Object.entries(agg)) {
+    for (const [hardware, stats] of Object.entries(aggregates[range])) {
       if (stats.total === 0) continue;
-      result[range][hw] = {
+      result[range][hardware] = {
         rate: Math.round((stats.n_success / stats.total) * 10000) / 100,
         total: stats.total,
         n_success: stats.n_success,
@@ -67,7 +73,12 @@ function aggregateByDateRange(rows: ReliabilityRow[]): DateRangeSuccessRateData
 
 export function ReliabilityProvider({ children }: { children: ReactNode }) {
   const { getUrlParam } = useUrlState();
-  const { data: rawRows, isLoading: loading, error: queryError } = useReliability();
+  const {
+    data: rawRows,
+    isLoading: loading,
+    isSuccess: reliabilitySettled,
+    error: queryError,
+  } = useReliability();
 
   const error = queryError ? queryError.message : null;
 
@@ -110,19 +121,15 @@ export function ReliabilityProvider({ children }: { children: ReactNode }) {
     return rangeData ? Object.keys(rangeData) : [];
   }, [dateRangeSuccessRateData, dateRange]);
 
-  useAutoInitializeToggleSet(availableModels, enabledModels, setEnabledModels);
-
   const filteredReliabilityData = useMemo(() => {
     const selectedRangeData = dateRangeSuccessRateData[dateRange];
     if (!selectedRangeData) return [];
-    return Object.entries(selectedRangeData).map(
-      ([model, stats]): ModelSuccessRateData => ({
-        model,
-        successRate: stats.rate,
-        total: stats.total,
-        n_success: stats.n_success,
-      }),
-    );
+    return Object.entries(selectedRangeData).map(([model, stats]): ModelSuccessRateData => ({
+      model,
+      successRate: stats.rate,
+      total: stats.total,
+      n_success: stats.n_success,
+    }));
   }, [dateRangeSuccessRateData, dateRange]);
 
   const chartData = useMemo(
@@ -152,16 +159,28 @@ export function ReliabilityProvider({ children }: { children: ReactNode }) {
   );
   const removeModel = useCallback((model: string) => removeModelRaw(model), [removeModelRaw]);
 
+  const lastModelScopeRef = useRef('');
   useEffect(() => {
-    if (modelsWithData.size === 0) return;
-    if (pendingActiveModels) {
-      const restored = new Set([...pendingActiveModels].filter((k) => modelsWithData.has(k)));
-      setEnabledModels(restored.size > 0 ? restored : modelsWithData);
-      setPendingActiveModels(null);
-      return;
-    }
-    setEnabledModels(modelsWithData);
-  }, [dateRange, modelsWithData]);
+    const scopeChanged = lastModelScopeRef.current !== dateRange;
+    const resolution = resolveAvailableSelection({
+      active: enabledModels,
+      available: modelsWithData,
+      pending: pendingActiveModels,
+      scopeChanged,
+      settled: reliabilitySettled,
+    });
+    if (!reliabilitySettled) return;
+    lastModelScopeRef.current = dateRange;
+    if (resolution.selection !== enabledModels) setEnabledModels(resolution.selection);
+    if (resolution.consumedPending) setPendingActiveModels(null);
+  }, [
+    dateRange,
+    enabledModels,
+    modelsWithData,
+    pendingActiveModels,
+    reliabilitySettled,
+    setEnabledModels,
+  ]);
 
   const selectAllModels = useCallback(
     () => selectAllModelsRaw(modelsWithData),
diff --git a/packages/app/src/components/reliability/ui/BarChartD3.tsx b/packages/app/src/components/reliability/ui/BarChartD3.tsx
index a0d003ac7..e443351b6 100644
--- a/packages/app/src/components/reliability/ui/BarChartD3.tsx
+++ b/packages/app/src/components/reliability/ui/BarChartD3.tsx
@@ -14,6 +14,7 @@ import { computeLeftMargin, measureTextWidth } from '@/lib/d3-chart/dynamic-marg
 import { useReliabilityContext } from '@/components/reliability/ReliabilityContext';
 import type { ModelSuccessRateData } from '@/components/reliability/types';
 import { useThemeColors } from '@/hooks/useThemeColors';
+import { type Locale } from '@/lib/i18n';
 import { useLocale } from '@/lib/use-locale';
 import ChartLegend from '@/components/ui/chart-legend';
 
@@ -21,15 +22,35 @@ type ChartItem = ModelSuccessRateData & { modelLabel: string };
 
 const BASE_MARGIN = { top: 24, right: 24, bottom: 40 };
 
-const generateReliabilityTooltipContent = (data: ChartItem, isPinned: boolean): string => {
+const TOOLTIP_STRINGS = {
+  en: {
+    dismiss: 'Click elsewhere to dismiss',
+    successRate: 'Success Rate:',
+    successful: 'Successful:',
+    totalRuns: 'Total Runs:',
+  },
+  zh: {
+    dismiss: '点击其他区域关闭',
+    successRate: '成功率：',
+    successful: '成功次数：',
+    totalRuns: '总运行次数：',
+  },
+} as const;
+
+const generateReliabilityTooltipContent = (
+  data: ChartItem,
+  isPinned: boolean,
+  locale: Locale = 'en',
+): string => {
+  const t = TOOLTIP_STRINGS[locale];
   const modelLabel = getHardwareConfig(data.model).label;
   return `
     <div style="background: var(--popover); border: 1px solid var(--border); border-radius: 8px; padding: 12px; box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1); user-select: ${isPinned ? 'text' : 'none'};">
-      ${isPinned ? '<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">Click elsewhere to dismiss</div>' : ''}
+      ${isPinned ? `<div style="color: var(--muted-foreground); font-size: 10px; margin-bottom: 6px; font-style: italic;">${t.dismiss}</div>` : ''}
       <div style="color: var(--foreground); font-size: 12px; font-weight: 600; margin-bottom: 8px;">${modelLabel}</div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>Success Rate:</strong> ${data.successRate.toFixed(2)}%</div>
-      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>Successful:</strong> ${data.n_success}</div>
-      <div style="color: var(--muted-foreground); font-size: 11px;"><strong>Total Runs:</strong> ${data.total}</div>
+      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${t.successRate}</strong> ${data.successRate.toFixed(2)}%</div>
+      <div style="color: var(--muted-foreground); font-size: 11px; margin-bottom: 4px;"><strong>${t.successful}</strong> ${data.n_success}</div>
+      <div style="color: var(--muted-foreground); font-size: 11px;"><strong>${t.totalRuns}</strong> ${data.total}</div>
     </div>
   `;
 };
@@ -157,6 +178,88 @@ export default function ReliabilityBarChartD3({ caption }: { caption?: ReactNode
     [chartData],
   );
 
+  const dataIdentity = useMemo(
+    () =>
+      JSON.stringify(
+        sortedChartData.map((datum) => JSON.stringify([datum.model, datum.modelLabel])).toSorted(),
+      ),
+    [sortedChartData],
+  );
+  const metricIdentity = useMemo(
+    () =>
+      JSON.stringify(
+        sortedChartData
+          .map((datum) =>
+            JSON.stringify([
+              datum.model,
+              datum.modelLabel,
+              datum.successRate,
+              datum.n_success,
+              datum.total,
+            ]),
+          )
+          .toSorted(),
+      ),
+    [sortedChartData],
+  );
+  const paletteIdentity = useMemo(
+    () =>
+      JSON.stringify(
+        sortedChartData
+          .map((datum) => JSON.stringify([datum.model, getCssColor(resolveColor(datum.model))]))
+          .toSorted(),
+      ),
+    [getCssColor, resolveColor, sortedChartData],
+  );
+  // Reverse so first in sort order appears at top (band scale range is [height, 0]).
+  const yDomain = useMemo(
+    () => [...sortedChartData].toReversed().map((datum) => datum.modelLabel),
+    [sortedChartData],
+  );
+  const xScaleConfig = useMemo(
+    () => ({ type: 'linear' as const, domain: [0, 100] as [number, number] }),
+    [],
+  );
+  const yScaleConfig = useMemo(
+    () => ({ type: 'band' as const, domain: yDomain, padding: 0.15 }),
+    [yDomain],
+  );
+  const zoomConfig = useMemo(
+    () => ({
+      enabled: true,
+      axes: 'x' as const,
+      scaleExtent: [0.1, 1] as [number, number],
+      rescaleX: (xScale: ContinuousScale, transform: d3.ZoomTransform) =>
+        xScale.copy().domain([0, 100 / transform.k]) as ContinuousScale,
+      customTransformStorage: (transform: d3.ZoomTransform) => d3.zoomIdentity.scale(transform.k),
+    }),
+    [],
+  );
+  const tooltipConfig = useMemo(
+    () => ({
+      rulerType: 'vertical' as const,
+      content: (datum: ChartItem, isPinned: boolean) =>
+        generateReliabilityTooltipContent(datum, isPinned, locale),
+      getRulerX: () => hoveredBarXRef.current,
+      getRulerY: (
+        datum: ChartItem,
+        scale: d3.ScaleLinear<number, number> | d3.ScaleLogarithmic<number, number>,
+      ) => {
+        const bandScale = scale as unknown as d3.ScaleBand<string>;
+        return (bandScale(datum.modelLabel) ?? 0) + bandScale.bandwidth() / 2;
+      },
+      onHoverStart: (selection: d3.Selection<SVGRectElement, ChartItem, SVGGElement, unknown>) => {
+        hoveredBarXRef.current = Number.parseFloat(selection.attr('width') || '0');
+        selection.attr('stroke', 'var(--foreground)').attr('stroke-width', 1.5);
+      },
+      onHoverEnd: (selection: d3.Selection<SVGRectElement, ChartItem, SVGGElement, unknown>) => {
+        selection.attr('stroke', 'none');
+      },
+      attachToLayer: 0,
+    }),
+    [locale],
+  );
+
   const dynamicHeight = useMemo(() => {
     const barCount = sortedChartData.length || 1;
     return Math.max(600, barCount * 45 + 80);
@@ -166,6 +269,7 @@ export default function ReliabilityBarChartD3({ caption }: { caption?: ReactNode
     (): LayerConfig<ChartItem>[] => [
       {
         type: 'horizontalBar',
+        key: 'bars',
         data: sortedChartData,
         config: {
           getY: (d) => d.modelLabel,
@@ -179,6 +283,7 @@ export default function ReliabilityBarChartD3({ caption }: { caption?: ReactNode
       {
         type: 'custom',
         key: 'bar-labels',
+        displayIdentity: paletteIdentity,
         render: (group, ctx) => {
           const yScale = ctx.yScale as d3.ScaleBand<string>;
 
@@ -212,19 +317,22 @@ export default function ReliabilityBarChartD3({ caption }: { caption?: ReactNode
             getCssColor(resolveColor(d.model)),
           );
         },
+        onDisplayUpdate: (group, ctx) => {
+          const baseXScale = ctx.xScale as d3.ScaleLinear<number, number>;
+          const svgNode = ctx.layout.svg.node();
+          const transform = svgNode ? d3.zoomTransform(svgNode) : d3.zoomIdentity;
+          const currentXScale = baseXScale.copy().domain([0, 100 / transform.k]);
+          positionLabelPairs(group, currentXScale, (datum) =>
+            getCssColor(resolveColor(datum.model)),
+          );
+        },
         onZoom: (group, ctx) => {
           const newXScale = ctx.newXScale as d3.ScaleLinear<number, number>;
           positionLabelPairs(group, newXScale, (d) => getCssColor(resolveColor(d.model)));
         },
       },
     ],
-    [sortedChartData, getCssColor, resolveColor],
-  );
-
-  // Reverse so first in sort order appears at top (band scale range is [height, 0])
-  const yDomain = useMemo(
-    () => [...sortedChartData].toReversed().map((d) => d.modelLabel),
-    [sortedChartData],
+    [sortedChartData, getCssColor, paletteIdentity, resolveColor],
   );
 
   const yAxisConfig = useMemo(() => ({ customize: twoRowYAxisLabels() }), []);
@@ -260,6 +368,9 @@ export default function ReliabilityBarChartD3({ caption }: { caption?: ReactNode
       <D3Chart<ChartItem>
         chartId="reliability-chart"
         data={sortedChartData}
+        dataIdentity={dataIdentity}
+        metricIdentity={metricIdentity}
+        displayIdentity={paletteIdentity}
         height={dynamicHeight}
         margin={chartMargin}
         watermark="logo"
@@ -268,36 +379,13 @@ export default function ReliabilityBarChartD3({ caption }: { caption?: ReactNode
         caption={caption}
         noDataOverlay={emptyOverlay}
         instructions="Shift+Scroll to zoom horizontally · Drag to pan · Double-click to reset · Hover for details"
-        xScale={{ type: 'linear', domain: [0, 100] }}
-        yScale={{ type: 'band', domain: yDomain, padding: 0.15 }}
+        xScale={xScaleConfig}
+        yScale={yScaleConfig}
         xAxis={xAxisConfig}
         yAxis={yAxisConfig}
         layers={layers}
-        zoom={{
-          enabled: true,
-          axes: 'x',
-          scaleExtent: [0.1, 1],
-          rescaleX: (xScale, transform) =>
-            xScale.copy().domain([0, 100 / transform.k]) as ContinuousScale,
-          customTransformStorage: (transform) => d3.zoomIdentity.scale(transform.k),
-        }}
-        tooltip={{
-          rulerType: 'vertical',
-          content: generateReliabilityTooltipContent,
-          getRulerX: () => hoveredBarXRef.current,
-          getRulerY: (d, ys) => {
-            const bandScale = ys as unknown as d3.ScaleBand<string>;
-            return (bandScale(d.modelLabel) ?? 0) + bandScale.bandwidth() / 2;
-          },
-          onHoverStart: (sel) => {
-            hoveredBarXRef.current = parseFloat(sel.attr('width') || '0');
-            sel.attr('stroke', 'var(--foreground)').attr('stroke-width', 1.5);
-          },
-          onHoverEnd: (sel) => {
-            sel.attr('stroke', 'none');
-          },
-          attachToLayer: 0,
-        }}
+        zoom={zoomConfig}
+        tooltip={tooltipConfig}
         legendElement={
           <ChartLegend
             variant="sidebar"
diff --git a/packages/app/src/components/reliability/ui/ChartControls.tsx b/packages/app/src/components/reliability/ui/ChartControls.tsx
index 161ab3508..56e8a9c42 100644
--- a/packages/app/src/components/reliability/ui/ChartControls.tsx
+++ b/packages/app/src/components/reliability/ui/ChartControls.tsx
@@ -29,7 +29,7 @@ const STRINGS = {
   zh: {
     dateRangeLabel: '时间范围',
     dateRangeTooltip:
-      '计算 Chip 可靠性指标的时间窗口。更长的范围可提供更稳定的统计数据，但可能无法反映近期的硬件性能变化。',
+      '计算芯片可靠性指标的时间窗口。更长的范围可提供更稳定的统计数据，但可能无法反映近期的硬件性能变化。',
     dateRangePlaceholder: '选择时间范围',
     last3Days: '最近 3 天',
     last7Days: '最近 7 天',
diff --git a/packages/app/src/components/reliability/ui/ChartDisplay.tsx b/packages/app/src/components/reliability/ui/ChartDisplay.tsx
index 4d7c196a1..0b8240b5e 100644
--- a/packages/app/src/components/reliability/ui/ChartDisplay.tsx
+++ b/packages/app/src/components/reliability/ui/ChartDisplay.tsx
@@ -23,9 +23,9 @@ const STRINGS = {
     captionSource: 'Source: SemiAnalysis InferenceX™',
   },
   zh: {
-    heading: 'Chip 可靠性',
-    description: '各 Chip 型号推理运行的成功率百分比，展示硬件在一段时间内的推理运行可靠性。',
-    captionHeading: '各 Chip 型号成功率',
+    heading: '芯片可靠性',
+    description: '各芯片型号推理运行的成功率百分比，展示硬件在一段时间内的推理运行可靠性。',
+    captionHeading: '各芯片型号成功率',
     captionSource: '数据来源：SemiAnalysis InferenceX™',
   },
 } as const;
diff --git a/packages/app/src/components/share-buttons.tsx b/packages/app/src/components/share-buttons.tsx
index b61dc3d76..a7cd53c7c 100644
--- a/packages/app/src/components/share-buttons.tsx
+++ b/packages/app/src/components/share-buttons.tsx
@@ -9,13 +9,13 @@ const SITE_URL = 'https://inferencex.semianalysis.com';
 const STRINGS = {
   en: {
     shareText:
-      'Check out InferenceX — open-source ML inference benchmarks comparing chips across real-world workloads. Transparent, up-to-date data for the ML community.',
+      'Explore InferenceX: open-source agentic and fixed-sequence AI inference benchmarks, including AgentX results from public chip runs.',
     twitter: 'Share on X (Twitter)',
     linkedin: 'Share on LinkedIn',
   },
   zh: {
     shareText:
-      '来看 InferenceX——开源 ML 推理基准测试，跨真实工作负载对比 Chip 性能。为 ML 社区提供透明、最新的数据。',
+      '探索 InferenceX：开源测试智能体推理与固定序列 AI 推理，并提供来自公开芯片运行的 AgentX 结果。',
     twitter: '分享到 X（推特）',
     linkedin: '分享到 LinkedIn',
   },
diff --git a/packages/app/src/components/submissions/SubmissionsTable.tsx b/packages/app/src/components/submissions/SubmissionsTable.tsx
index 4b1243e85..54550d4fb 100644
--- a/packages/app/src/components/submissions/SubmissionsTable.tsx
+++ b/packages/app/src/components/submissions/SubmissionsTable.tsx
@@ -1,7 +1,7 @@
 'use client';
 
 import { ChevronDown, ChevronRight, GitCompare, Info } from 'lucide-react';
-import { useCallback, useEffect, useMemo, useState } from 'react';
+import { useCallback, useMemo, useState } from 'react';
 
 import {
   DB_MODEL_TO_DISPLAY,
@@ -91,7 +91,7 @@ const STRINGS = {
   },
   zh: {
     searchPlaceholder: '搜索配置...',
-    colGpu: 'Chip',
+    colGpu: '芯片',
     colModel: '模型',
     colPrecision: '精度',
     colSpecMethod: '推测解码',
@@ -103,25 +103,25 @@ const STRINGS = {
     noData: '暂无提交数据。',
     vsPrev: '对比',
     vendorLabel: '厂商：',
-    vendorTip: 'Chip 制造商',
+    vendorTip: '芯片制造商',
     specMethodLabel: '推测解码方法：',
     specMethodTip: '推测解码方法（如 MTP、Eagle）',
     disaggLabel: '分离式部署：',
-    disaggTip: 'Prefill 和 Decode 在不同 Chip 池上运行',
+    disaggTip: 'Prefill 和 Decode 在不同芯片池上运行',
     multinodeLabel: '多节点：',
     multinodeTip: '配置跨多个物理节点',
-    totalGpusLabel: '总 Chip 数：',
-    totalGpusTip: '物理 Chip 总数。分离式部署时，Prefill 和 Decode 使用不同的 Chip 池',
-    prefillGpusLabel: 'Prefill Chip 数：',
-    prefillGpusTip: '用于 Prefill（提示处理）阶段的 Chip',
-    decodeGpusLabel: 'Decode Chip 数：',
-    decodeGpusTip: '用于 Decode（Token 生成）阶段的 Chip',
+    totalGpusLabel: '总芯片数：',
+    totalGpusTip: '物理芯片总数。分离式部署时，Prefill 和 Decode 使用不同的芯片池',
+    prefillGpusLabel: 'Prefill 芯片数：',
+    prefillGpusTip: '用于 Prefill（提示处理）阶段的芯片',
+    decodeGpusLabel: 'Decode 芯片数：',
+    decodeGpusTip: '用于 Decode（Token 生成）阶段的芯片',
     prefillTpEpLabel: 'Prefill TP/EP：',
     prefillTpEpTip: 'Prefill 的张量并行 / 专家并行',
     decodeTpEpLabel: 'Decode TP/EP：',
     decodeTpEpTip: 'Decode 的张量并行 / 专家并行',
-    aggregateGpusLabel: '聚合推理 Chip 数：',
-    aggregateGpusTip: '单个聚合推理引擎使用的 Chip 数',
+    aggregateGpusLabel: '聚合推理芯片数：',
+    aggregateGpusTip: '单个聚合推理引擎使用的芯片数',
     aggregateTpEpLabel: '聚合推理 TP/EP：',
     aggregateTpEpTip: '聚合推理引擎的张量并行 / 专家并行',
     sequencesLabel: '序列组合：',
@@ -237,6 +237,7 @@ export default function SubmissionsTable({ data }: SubmissionsTableProps) {
 
   const handleSort = useCallback(
     (key: SortKey) => {
+      setVisibleCount(ROW_PAGE_SIZE);
       if (sortKey === key) {
         setSortDir((d) => (d === 'asc' ? 'desc' : 'asc'));
       } else {
@@ -273,11 +274,10 @@ export default function SubmissionsTable({ data }: SubmissionsTableProps) {
     });
   }, [filtered, sortKey, sortDir]);
 
-  // Reset visible count when the filtered/sorted view changes so the user
-  // always lands at the top of the new result set instead of mid-list.
-  useEffect(() => {
+  const handleSearchChange = (nextSearch: string) => {
+    setSearch(nextSearch);
     setVisibleCount(ROW_PAGE_SIZE);
-  }, [search, sortKey, sortDir]);
+  };
 
   const visibleRows = useMemo(() => sorted.slice(0, visibleCount), [sorted, visibleCount]);
   const hiddenCount = Math.max(0, sorted.length - visibleRows.length);
@@ -306,7 +306,7 @@ export default function SubmissionsTable({ data }: SubmissionsTableProps) {
       <input
         type="text"
         value={search}
-        onChange={(e) => setSearch(e.target.value)}
+        onChange={(e) => handleSearchChange(e.target.value)}
         onBlur={() => {
           if (search.trim()) track('submissions_table_searched', { query: search.trim() });
         }}
diff --git a/packages/app/src/components/supporters-strip.tsx b/packages/app/src/components/supporters-strip.tsx
new file mode 100644
index 000000000..9c4762cc4
--- /dev/null
+++ b/packages/app/src/components/supporters-strip.tsx
@@ -0,0 +1,59 @@
+'use client';
+
+import Link from 'next/link';
+
+import { track } from '@/lib/analytics';
+
+export interface SupporterOrg {
+  /** Display name shown in the org strip */
+  label: string;
+  /** Link to the org's quote on the supporters page */
+  href: string;
+}
+
+export interface SupportersStripProps {
+  /** Supporter orgs shown in the strip, in display order */
+  orgs: SupporterOrg[];
+  /** Link to the page with all quotes and supporters */
+  moreHref: string;
+  /** Label for the moreHref link (default "See full quotes & more supporters →") */
+  moreLabel?: string;
+}
+
+/**
+ * Compact replacement for the landing quote carousel: renders the supporter
+ * org strip plus a single link out to the full quotes page instead of a
+ * rotating quote block, saving vertical space above the fold. Each org links
+ * to its quote anchor on the supporters page.
+ */
+export function SupportersStrip({ orgs, moreHref, moreLabel }: SupportersStripProps) {
+  return (
+    <div className="flex flex-col gap-4">
+      {/* Org name strip */}
+      <div className="flex flex-wrap justify-center gap-x-6 md:gap-x-8 gap-y-2 mx-4">
+        {orgs.map((org) => (
+          <Link
+            key={org.label}
+            href={org.href}
+            prefetch={false}
+            className="text-xs font-semibold tracking-wide uppercase text-[#808488] transition-colors duration-200 hover:text-foreground hover:underline"
+            onClick={() => track('supporters_strip_org_clicked', { org: org.label })}
+          >
+            {org.label}
+          </Link>
+        ))}
+      </div>
+
+      <div className="flex justify-end" data-testid="supporters-more-row">
+        <Link
+          href={moreHref}
+          prefetch={false}
+          className="text-xs font-bold text-brand hover:underline"
+          onClick={() => track('quote_carousel_see_more_clicked')}
+        >
+          {moreLabel ?? 'See full quotes & more supporters →'}
+        </Link>
+      </div>
+    </div>
+  );
+}
diff --git a/packages/app/src/components/tab-nav.tsx b/packages/app/src/components/tab-nav.tsx
index 9df8f66a3..e6464e206 100644
--- a/packages/app/src/components/tab-nav.tsx
+++ b/packages/app/src/components/tab-nav.tsx
@@ -3,9 +3,19 @@
 import { ChevronDown } from 'lucide-react';
 import Link from 'next/link';
 import { usePathname, useRouter } from 'next/navigation';
-import { useContext, useEffect, useState } from 'react';
+import { useMemo, useState } from 'react';
 
 import { track } from '@/lib/analytics';
+import {
+  DASHBOARD_ROUTES,
+  dashboardRouteForPathname,
+  getDashboardRoute,
+  isDashboardRouteKey,
+  type DashboardRoute,
+  type DashboardRouteKey,
+} from '@/lib/dashboard-routes';
+import { localePath } from '@/lib/i18n';
+import { TAB_LABELS_ZH } from '@/lib/tab-meta-zh';
 import { useFeatureGate } from '@/lib/use-feature-gate';
 import { Card } from '@/components/ui/card';
 import { Label } from '@/components/ui/label';
@@ -20,34 +30,27 @@ import {
   SelectTrigger,
   SelectValue,
 } from '@/components/ui/select';
-import { UnofficialRunContext } from '@/components/unofficial-run-provider';
-import { hasZhSibling, isZhPathname, ZH_PREFIX } from '@/lib/i18n';
-import { TAB_LABELS_ZH } from '@/lib/tab-meta-zh';
+import { useClientSearchParams } from '@/hooks/useClientSearch';
 import { cn } from '@/lib/utils';
 
-const VISIBLE_TABS = [
-  { href: '/inference', label: 'Inference Performance', testId: 'tab-trigger-inference' },
-  { href: '/evaluation', label: 'Accuracy Evals', testId: 'tab-trigger-evaluation' },
-  { href: '/historical', label: 'Historical Trends', testId: 'tab-trigger-historical' },
-  { href: '/calculator', label: 'TCO Calculator', testId: 'tab-trigger-calculator' },
-  { href: '/gpu-specs', label: 'Chip Specs', testId: 'tab-trigger-gpu-specs' },
-  { href: '/submissions', label: 'Submissions', testId: 'tab-trigger-submissions' },
-] as const;
-
-const GATED_TABS = [
-  { href: '/collectivex', label: 'CollectiveX', testId: 'tab-trigger-collectivex' },
-  { href: '/ai-chart', label: 'AI Chart', testId: 'tab-trigger-ai-chart' },
-  { href: '/gpu-metrics', label: 'PowerX', testId: 'tab-trigger-gpu-metrics' },
-  {
-    href: '/current-inferencex-image',
-    label: 'Images',
-    testId: 'tab-trigger-current-inferencex-image',
-  },
-  { href: '/feedback', label: 'Feedback', testId: 'tab-trigger-feedback' },
-] as const;
-
-const TAB_VALUES = new Set([...VISIBLE_TABS, ...GATED_TABS].map((t) => t.href.slice(1)));
-const GATED_VALUES = new Set(GATED_TABS.map((t) => t.href.slice(1)));
+const TAB_LABELS_EN: Record<DashboardRouteKey, string> = {
+  inference: 'Inference Performance',
+  evaluation: 'Accuracy Evals',
+  historical: 'Historical Trends',
+  calculator: 'TCO Calculator',
+  fleet: 'Fleet Lifecycle',
+  reliability: 'Reliability',
+  'gpu-specs': 'Chip Specs',
+  submissions: 'Submissions',
+  collectivex: 'CollectiveX',
+  'ai-chart': 'AI Chart',
+  'gpu-metrics': 'PowerX',
+  'current-inferencex-image': 'Images',
+  feedback: 'Feedback',
+};
+
+const PRIMARY_TABS = DASHBOARD_ROUTES.filter((route) => route.navGroup === 'primary');
+const GATED_TABS = DASHBOARD_ROUTES.filter((route) => route.navGroup === 'feature-gated');
 
 const tabLinkClass = cn(
   'relative inline-flex items-center justify-center',
@@ -63,13 +66,7 @@ const currentTabClass = (active: boolean) =>
     ? 'border-secondary dark:border-primary text-secondary dark:text-primary'
     : 'hover:border-muted-foreground/30';
 
-function activeTab(pathname: string): string {
-  const segments = pathname.split('/').filter(Boolean);
-  if (segments[0] === ZH_PREFIX.slice(1)) segments.shift();
-  return segments[0] || 'inference';
-}
-
-function handleDesktopClick(tab: string) {
+function handleDesktopClick(tab: DashboardRouteKey) {
   window.dispatchEvent(new CustomEvent('inferencex:tab-change'));
   track('tab_changed', { tab });
 }
@@ -78,49 +75,27 @@ export function TabNav() {
   const pathname = usePathname();
   const router = useRouter();
   const featureGateUnlocked = useFeatureGate();
-  const isZh = isZhPathname(pathname);
-  const current = activeTab(pathname);
-  const selectedTab = TAB_VALUES.has(current) ? current : '';
-  // On /zh pages, tabs with a Chinese sibling navigate within the Chinese
-  // tree and show Chinese labels; the rest (most gated tabs) keep English
-  // targets.
-  const tabLabel = (tab: { href: string; label: string }) =>
-    isZh ? (TAB_LABELS_ZH[tab.href.slice(1)] ?? tab.label) : tab.label;
-  const localizedPath = (path: string) =>
-    isZh && hasZhSibling(path) ? `${ZH_PREFIX}${path}` : path;
-
-  // Preserve the `unofficialrun(s)` URL param across tab navigation so an
-  // overlay loaded on /inference doesn't get dropped when switching to
-  // /evaluation, etc. The URL is the source of truth (it's still set during
-  // the in-flight fetch and even when the fetch fails), so we read it from
-  // window.location and re-sync on pathname change, context update
-  // (dismiss/clear writes via history.pushState), and popstate.
-  const unofficialCtx = useContext(UnofficialRunContext);
-  const ctxRunInfos = unofficialCtx?.unofficialRunInfos;
-  const [unofficialIds, setUnofficialIds] = useState('');
-  useEffect(() => {
-    if (typeof window === 'undefined') return undefined;
-    const sync = () => {
-      const sp = new URLSearchParams(window.location.search);
-      for (const [k, v] of sp) {
-        if (/^unofficialruns?$/iu.test(k) && v) {
-          setUnofficialIds(v);
-          return;
-        }
-      }
-      setUnofficialIds('');
-    };
-    sync();
-    window.addEventListener('popstate', sync);
-    return () => window.removeEventListener('popstate', sync);
-  }, [pathname, ctxRunInfos]);
+  const locale = pathname === '/zh' || pathname.startsWith('/zh/') ? 'zh' : 'en';
+  const current = dashboardRouteForPathname(pathname)?.key ?? 'inference';
+  const selectedTab = getDashboardRoute(current).navGroup === 'footer-only' ? '' : current;
+  const tabLabel = (route: DashboardRoute) =>
+    locale === 'zh' ? TAB_LABELS_ZH[route.key] : TAB_LABELS_EN[route.key];
+
+  const searchParams = useClientSearchParams();
+  const unofficialIds = useMemo(() => {
+    for (const [key, value] of searchParams) {
+      if (/^unofficialruns?$/iu.test(key) && value) return value;
+    }
+    return '';
+  }, [searchParams]);
   const tabHref = (path: string) =>
     unofficialIds ? `${path}?unofficialruns=${unofficialIds}` : path;
 
   const handleMobileChange = (value: string) => {
+    if (!isDashboardRouteKey(value)) return;
     window.dispatchEvent(new CustomEvent('inferencex:tab-change'));
     track('tab_changed', { tab: value });
-    router.push(tabHref(localizedPath(`/${value}`)));
+    router.push(tabHref(localePath(getDashboardRoute(value).path, locale)));
   };
 
   return (
@@ -130,37 +105,35 @@ export function TabNav() {
         <div className="w-full pb-6" />
         <Card>
           <div className="space-y-2">
-            <Label htmlFor="chart-select">{isZh ? '选择图表' : 'Select Chart'}</Label>
+            <Label htmlFor="chart-select">{locale === 'zh' ? '选择图表' : 'Select Chart'}</Label>
             <Select value={selectedTab} onValueChange={handleMobileChange}>
               <SelectTrigger id="chart-select" data-testid="mobile-chart-select" className="w-full">
-                <SelectValue placeholder={isZh ? '选择图表' : 'Select Chart'} />
+                <SelectValue placeholder={locale === 'zh' ? '选择图表' : 'Select Chart'} />
               </SelectTrigger>
               <SelectContent>
-                {VISIBLE_TABS.map((tab) => {
-                  const value = tab.href.slice(1);
-                  return (
-                    <SelectItem key={value} value={value} data-ph-capture-attribute-tab={value}>
-                      {tabLabel(tab)}
-                    </SelectItem>
-                  );
-                })}
+                {PRIMARY_TABS.map((route) => (
+                  <SelectItem
+                    key={route.key}
+                    value={route.key}
+                    data-ph-capture-attribute-tab={route.key}
+                  >
+                    {tabLabel(route)}
+                  </SelectItem>
+                ))}
                 {featureGateUnlocked && (
                   <>
                     <SelectSeparator />
                     <SelectGroup>
-                      <SelectLabel>{isZh ? '隐藏' : 'Hidden'}</SelectLabel>
-                      {GATED_TABS.map((tab) => {
-                        const value = tab.href.slice(1);
-                        return (
-                          <SelectItem
-                            key={value}
-                            value={value}
-                            data-ph-capture-attribute-tab={value}
-                          >
-                            {tabLabel(tab)}
-                          </SelectItem>
-                        );
-                      })}
+                      <SelectLabel>{locale === 'zh' ? '隐藏' : 'Hidden'}</SelectLabel>
+                      {GATED_TABS.map((route) => (
+                        <SelectItem
+                          key={route.key}
+                          value={route.key}
+                          data-ph-capture-attribute-tab={route.key}
+                        >
+                          {tabLabel(route)}
+                        </SelectItem>
+                      ))}
                     </SelectGroup>
                   </>
                 )}
@@ -177,25 +150,25 @@ export function TabNav() {
             data-testid="chart-section-tabs"
             className="relative flex items-center justify-evenly min-w-0"
           >
-            {VISIBLE_TABS.map((tab) => (
+            {PRIMARY_TABS.map((route) => (
               <Link
-                key={tab.href}
-                href={tabHref(localizedPath(tab.href))}
-                data-testid={tab.testId}
-                data-ph-capture-attribute-tab={tab.href.slice(1)}
-                onClick={() => handleDesktopClick(tab.href.slice(1))}
-                className={cn(tabLinkClass, currentTabClass(current === tab.href.slice(1)))}
+                key={route.key}
+                href={tabHref(localePath(route.path, locale))}
+                data-testid={`tab-trigger-${route.key}`}
+                data-ph-capture-attribute-tab={route.key}
+                onClick={() => handleDesktopClick(route.key)}
+                className={cn(tabLinkClass, currentTabClass(current === route.key))}
               >
-                {tabLabel(tab)}
+                {tabLabel(route)}
               </Link>
             ))}
             {featureGateUnlocked && (
               <HiddenTabsPopover
                 current={current}
-                tabHref={(path) => tabHref(localizedPath(path))}
+                tabHref={(path) => tabHref(localePath(path, locale))}
                 onSelect={handleDesktopClick}
                 tabLabel={tabLabel}
-                isZh={isZh}
+                locale={locale}
               />
             )}
           </nav>
@@ -210,16 +183,16 @@ function HiddenTabsPopover({
   tabHref,
   onSelect,
   tabLabel,
-  isZh,
+  locale,
 }: {
-  current: string;
+  current: DashboardRouteKey;
   tabHref: (path: string) => string;
-  onSelect: (tab: string) => void;
-  tabLabel: (tab: { href: string; label: string }) => string;
-  isZh: boolean;
+  onSelect: (tab: DashboardRouteKey) => void;
+  tabLabel: (route: DashboardRoute) => string;
+  locale: 'en' | 'zh';
 }) {
   const [open, setOpen] = useState(false);
-  const active = GATED_VALUES.has(current);
+  const active = getDashboardRoute(current).navGroup === 'feature-gated';
 
   return (
     <Popover open={open} onOpenChange={setOpen}>
@@ -228,7 +201,7 @@ function HiddenTabsPopover({
         data-ph-capture-attribute-tab="hidden"
         className={cn(tabLinkClass, currentTabClass(active), 'gap-1 cursor-pointer')}
       >
-        {isZh ? '隐藏' : 'Hidden'}
+        {locale === 'zh' ? '隐藏' : 'Hidden'}
         <ChevronDown
           className={cn('size-4 transition-transform', open && 'rotate-180')}
           aria-hidden
@@ -236,18 +209,17 @@ function HiddenTabsPopover({
       </PopoverTrigger>
       <PopoverContent align="center" className="w-44 p-1" data-testid="tab-hidden-popover">
         <ul className="flex flex-col">
-          {GATED_TABS.map((tab) => {
-            const value = tab.href.slice(1);
-            const isActive = current === value;
+          {GATED_TABS.map((route) => {
+            const isActive = current === route.key;
             return (
-              <li key={tab.href}>
+              <li key={route.key}>
                 <Link
-                  href={tabHref(tab.href)}
-                  data-testid={tab.testId}
-                  data-ph-capture-attribute-tab={value}
+                  href={tabHref(route.path)}
+                  data-testid={`tab-trigger-${route.key}`}
+                  data-ph-capture-attribute-tab={route.key}
                   onClick={() => {
                     setOpen(false);
-                    onSelect(value);
+                    onSelect(route.key);
                   }}
                   className={cn(
                     'block rounded-sm px-2 py-1.5 text-sm',
@@ -257,7 +229,7 @@ function HiddenTabsPopover({
                       : 'text-muted-foreground hover:bg-accent hover:text-foreground',
                   )}
                 >
-                  {tabLabel(tab)}
+                  {tabLabel(route)}
                 </Link>
               </li>
             );
diff --git a/packages/app/src/components/trends/HistoricalTrendsDisplay.tsx b/packages/app/src/components/trends/HistoricalTrendsDisplay.tsx
index c35943973..a227718e0 100644
--- a/packages/app/src/components/trends/HistoricalTrendsDisplay.tsx
+++ b/packages/app/src/components/trends/HistoricalTrendsDisplay.tsx
@@ -4,7 +4,12 @@ import { track } from '@/lib/analytics';
 import { useLocale } from '@/lib/use-locale';
 import React, { useCallback, useMemo, useState } from 'react';
 
-import { useInference } from '@/components/inference/InferenceContext';
+import {
+  useInferenceActions,
+  useInferenceData,
+  useInferenceDisplay,
+  useInferenceFilters,
+} from '@/components/inference/InferenceContext';
 import { useInterpolatedTrendData } from '@/components/inference/hooks/useInterpolatedTrendData';
 import type { TrendLineConfig } from '@/components/inference/types';
 import ChartControls from '@/components/inference/ui/ChartControls';
@@ -31,6 +36,12 @@ import {
 } from '@/lib/data-mappings';
 import { getDisplayLabel } from '@/lib/utils';
 import { useThemeColors } from '@/hooks/useThemeColors';
+import {
+  includesJalapenoResult,
+  includesVeraRubinResult,
+  JalapenoOfficialPreviewNotice,
+  VeraRubinOfficialPreviewNotice,
+} from '@/components/official-preview-notice';
 
 const STRINGS = {
   en: {
@@ -53,7 +64,7 @@ const STRINGS = {
     heading: '历史趋势',
     description: '在固定交互性操作点下，各性能指标随时间的插值变化。',
     targetLabel: '目标交互性 (tok/s/user)',
-    targetTooltip: '用于插值的交互性操作点。移动滑块可查看各 Chip 在不同交互性水平下的性能变化。',
+    targetTooltip: '用于插值的交互性操作点。移动滑块可查看各芯片在不同交互性水平下的性能变化。',
     captionTitle: (yTitle: string, target: number) =>
       `${yTitle} 随时间变化（交互性 ${target} tok/s/user）`,
     source: '来源：SemiAnalysis InferenceX™',
@@ -67,28 +78,19 @@ const STRINGS = {
 
 export default function HistoricalTrendsDisplay() {
   const t = STRINGS[useLocale()];
+  const { graphs, loading, hardwareConfig, hwTypesWithData, availableDates } = useInferenceData();
+  const { selectedModel, selectedSequence, selectedPrecisions, activeHwTypes, selectedRunDate } =
+    useInferenceFilters();
+  const { selectedYAxisMetric, tokenRevenuePricing, logScale, isLegendExpanded, highContrast } =
+    useInferenceDisplay();
   const {
-    graphs,
-    loading,
-    selectedModel,
-    selectedSequence,
-    selectedPrecisions,
-    selectedYAxisMetric,
-    hardwareConfig,
-    activeHwTypes,
-    hwTypesWithData,
     toggleHwType,
     removeHwType,
     selectAllHwTypes,
-    availableDates,
-    logScale,
     setLogScale,
-    isLegendExpanded,
     setIsLegendExpanded,
-    workflowInfo,
-    highContrast,
     setHighContrast,
-  } = useInference();
+  } = useInferenceActions();
 
   // Check if interactivity chart data exists
   const hasInteractivityChart = graphs.some((g) => g.chartDefinition.chartType === 'interactivity');
@@ -153,6 +155,7 @@ export default function HistoricalTrendsDisplay() {
     selectedYAxisMetric,
     targetInteractivity,
     availableDates,
+    tokenRevenuePricing,
     enabled: hasInteractivityChart,
   });
 
@@ -190,6 +193,8 @@ export default function HistoricalTrendsDisplay() {
         }),
     [trendLines, activeHwTypes, hardwareConfig, selectedPrecisions, resolveColor],
   );
+  const showsJalapenoPreview = includesJalapenoResult(lineConfigs.map((config) => config.hwKey));
+  const showsVeraRubinPreview = includesVeraRubinResult(lineConfigs.map((config) => config.hwKey));
 
   if (loading || graphs.length === 0 || trendLoading) {
     return (
@@ -326,15 +331,18 @@ export default function HistoricalTrendsDisplay() {
                         .map((prec: string) => getPrecisionLabel(prec as Precision))
                         .join(', ')}{' '}
                       • {getSequenceLabel(selectedSequence as Sequence)} • {t.source}
-                      {workflowInfo && workflowInfo.length > 0 && workflowInfo[0]?.run_date && (
+                      {selectedRunDate && (
                         <>
                           {' '}
-                          • {t.updated} {workflowInfo[0].run_date.split(',')[0]}
+                          • {t.updated} {selectedRunDate}
                         </>
                       )}
                     </p>
+                    {showsJalapenoPreview && <JalapenoOfficialPreviewNotice />}
+                    {showsVeraRubinPreview && <VeraRubinOfficialPreviewNotice />}
                     <MetricAssumptionNotes
                       selectedYAxisMetric={selectedYAxisMetric}
+                      activeHwKeys={activeHwTypes}
                       includeAllPowerThroughputMetrics={false}
                       includePowerThroughputCaveat={false}
                     />
diff --git a/packages/app/src/components/ui/accordion.tsx b/packages/app/src/components/ui/accordion.tsx
deleted file mode 100644
index 45edcfde8..000000000
--- a/packages/app/src/components/ui/accordion.tsx
+++ /dev/null
@@ -1,66 +0,0 @@
-'use client';
-
-import * as AccordionPrimitive from '@radix-ui/react-accordion';
-import { ChevronDownIcon } from 'lucide-react';
-import * as React from 'react';
-
-import { cn } from '@/lib/utils';
-
-function Accordion({ ...props }: React.ComponentProps<typeof AccordionPrimitive.Root>) {
-  return (
-    <AccordionPrimitive.Root data-slot="accordion" className="flex flex-col gap-8" {...props} />
-  );
-}
-
-function AccordionItem({
-  className,
-  ...props
-}: React.ComponentProps<typeof AccordionPrimitive.Item>) {
-  return (
-    <AccordionPrimitive.Item
-      data-slot="accordion-item"
-      className={cn('border rounded-md', className)}
-      {...props}
-    />
-  );
-}
-
-function AccordionTrigger({
-  className,
-  children,
-  ...props
-}: React.ComponentProps<typeof AccordionPrimitive.Trigger>) {
-  return (
-    <AccordionPrimitive.Header className="flex">
-      <AccordionPrimitive.Trigger
-        data-slot="accordion-trigger"
-        className={cn(
-          'focus-visible:border-ring focus-visible:ring-ring flex flex-1 bg-card/80 rounded-t-md data-[state=closed]:rounded-md items-start justify-between gap-4 p-4 text-left text-lg font-medium transition-all outline-none hover:underline focus-visible:ring-[3px] disabled:pointer-events-none disabled:opacity-50 [&[data-state=open]>svg]:rotate-180',
-          className,
-        )}
-        {...props}
-      >
-        {children}
-        <ChevronDownIcon className="text-muted-foreground pointer-events-none size-6 shrink-0 translate-y-0.5 transition-transform duration-200" />
-      </AccordionPrimitive.Trigger>
-    </AccordionPrimitive.Header>
-  );
-}
-
-function AccordionContent({
-  className,
-  children,
-  ...props
-}: React.ComponentProps<typeof AccordionPrimitive.Content>) {
-  return (
-    <AccordionPrimitive.Content
-      data-slot="accordion-content"
-      className="data-[state=closed]:animate-accordion-up bg-card/80 rounded-md p-4 data-[state=open]:animate-accordion-down overflow-hidden"
-      {...props}
-    >
-      <div className={cn('pt-0', className)}>{children}</div>
-    </AccordionPrimitive.Content>
-  );
-}
-
-export { Accordion, AccordionContent, AccordionItem, AccordionTrigger };
diff --git a/packages/app/src/components/ui/atom-engine-footnote.tsx b/packages/app/src/components/ui/atom-engine-footnote.tsx
new file mode 100644
index 000000000..717115cb2
--- /dev/null
+++ b/packages/app/src/components/ui/atom-engine-footnote.tsx
@@ -0,0 +1,38 @@
+'use client';
+
+import { useLocale } from '@/lib/use-locale';
+import { cn } from '@/lib/utils';
+
+/**
+ * Superscript marker appended to ATOM-family framework labels (e.g. "ATOM¹",
+ * "Mooncake ATOMesh¹"). Consumers detect it in a display label to decide
+ * whether the footnote below applies to the chart.
+ */
+export const ATOM_FOOTNOTE_MARKER = '¹';
+
+const STRINGS = {
+  en: {
+    atomFootnote:
+      'The ATOM engine is promising, however it has yet to serve production tokens. It is still in its infant stage.',
+  },
+  zh: {
+    atomFootnote: 'ATOM 引擎前景可期，但尚未用于生产环境 token 服务，仍处于早期阶段。',
+  },
+} as const;
+
+/**
+ * Footnote explaining the ¹ marker on ATOM-family series labels. Rendered in
+ * the evaluation chart legend and in the inference chart's axis-metric info
+ * footer, so the copy lives in one place.
+ */
+export function AtomEngineFootnote({ className }: { className?: string }) {
+  const locale = useLocale();
+  return (
+    <p
+      data-testid="atom-engine-footnote"
+      className={cn('text-3xs text-muted-foreground/70 leading-tight', className)}
+    >
+      <sup>1</sup> {STRINGS[locale].atomFootnote}
+    </p>
+  );
+}
diff --git a/packages/app/src/components/ui/calendar-picker-utils.test.ts b/packages/app/src/components/ui/calendar-picker-utils.test.ts
index cea929206..d323bf929 100644
--- a/packages/app/src/components/ui/calendar-picker-utils.test.ts
+++ b/packages/app/src/components/ui/calendar-picker-utils.test.ts
@@ -52,14 +52,12 @@ describe('calendar-picker-utils', () => {
     expect(formatCalendarDate(bounds.latestMonth)).toBe('2026-04-20');
   });
 
-  it('supports inclusive and exclusive min/max boundary checks', () => {
+  it('treats min/max boundaries as inclusive', () => {
     const min = parseCalendarDate('2026-01-10');
     const max = parseCalendarDate('2026-01-20');
 
     expect(isCalendarDateOutOfRange(parseCalendarDate('2026-01-10'), min, max)).toBe(false);
     expect(isCalendarDateOutOfRange(parseCalendarDate('2026-01-20'), min, max)).toBe(false);
-    expect(isCalendarDateOutOfRange(parseCalendarDate('2026-01-10'), min, max, true)).toBe(true);
-    expect(isCalendarDateOutOfRange(parseCalendarDate('2026-01-20'), min, max, true)).toBe(true);
     expect(isCalendarDateOutOfRange(parseCalendarDate('2026-01-09'), min, max)).toBe(true);
     expect(isCalendarDateOutOfRange(parseCalendarDate('2026-01-21'), min, max)).toBe(true);
   });
diff --git a/packages/app/src/components/ui/calendar-picker-utils.tsx b/packages/app/src/components/ui/calendar-picker-utils.tsx
index 7381b5bab..289bb2f04 100644
--- a/packages/app/src/components/ui/calendar-picker-utils.tsx
+++ b/packages/app/src/components/ui/calendar-picker-utils.tsx
@@ -1,15 +1,24 @@
 'use client';
 
-import { useEffect, useState } from 'react';
+import { useReducer, type Dispatch, type SetStateAction } from 'react';
 
 import { Button } from '@/components/ui/button';
+import type { Locale } from '@/lib/i18n';
+import { useLocale } from '@/lib/use-locale';
 import { cn } from '@/lib/utils';
 
-const DISPLAY_DATE_FORMATTER = new Intl.DateTimeFormat('en-US', {
-  month: 'short',
-  day: 'numeric',
-  year: 'numeric',
-});
+const DISPLAY_DATE_FORMATTERS: Record<Locale, Intl.DateTimeFormat> = {
+  en: new Intl.DateTimeFormat('en-US', {
+    month: 'short',
+    day: 'numeric',
+    year: 'numeric',
+  }),
+  zh: new Intl.DateTimeFormat('zh-CN', {
+    month: 'short',
+    day: 'numeric',
+    year: 'numeric',
+  }),
+};
 
 export interface CalendarDateBounds {
   minAllowedDate: Date;
@@ -40,6 +49,31 @@ export interface CalendarMonthPanelProps {
 
 type CalendarMonthResetDep = string | number | boolean | null | undefined;
 
+interface CalendarMonthState {
+  resetKey: string;
+  currentMonth: Date;
+}
+
+type CalendarMonthAction =
+  | { type: 'reset'; resetKey: string; currentMonth: Date }
+  | { type: 'set'; currentMonth: SetStateAction<Date> };
+
+function calendarMonthReducer(
+  state: CalendarMonthState,
+  action: CalendarMonthAction,
+): CalendarMonthState {
+  if (action.type === 'reset') {
+    return { resetKey: action.resetKey, currentMonth: action.currentMonth };
+  }
+  return {
+    ...state,
+    currentMonth:
+      typeof action.currentMonth === 'function'
+        ? action.currentMonth(state.currentMonth)
+        : action.currentMonth,
+  };
+}
+
 export function parseCalendarDate(dateStr: string): Date {
   if (dateStr.includes('-') && !dateStr.includes(',')) {
     const [year, month, day] = dateStr.split('-');
@@ -58,8 +92,8 @@ export function formatCalendarDate(date: Date): string {
   return `${year}-${month}-${day}`;
 }
 
-export function formatDisplayDate(dateStr: string): string {
-  return DISPLAY_DATE_FORMATTER.format(parseCalendarDate(dateStr));
+export function formatDisplayDate(dateStr: string, locale: Locale = 'en'): string {
+  return DISPLAY_DATE_FORMATTERS[locale].format(parseCalendarDate(dateStr));
 }
 
 export function getLatestSelectableDate(availableDates?: string[], maxDate?: string): string {
@@ -103,8 +137,6 @@ export function isCalendarDateOutOfRange(
   date: Date,
   minAllowedDate: Date,
   maxAllowedDate: Date,
-  // MultiDatePicker keeps legacy exclusive min/max boundaries; the other pickers use inclusive edges.
-  excludeBoundaryDates = false,
 ): boolean {
   const dateOnly = new Date(date.getFullYear(), date.getMonth(), date.getDate());
   const minDateOnly = new Date(
@@ -117,10 +149,6 @@ export function isCalendarDateOutOfRange(
     maxAllowedDate.getMonth(),
     maxAllowedDate.getDate(),
   );
-  if (excludeBoundaryDates) {
-    return dateOnly <= minDateOnly || dateOnly >= maxDateOnly;
-  }
-
   return dateOnly < minDateOnly || dateOnly > maxDateOnly;
 }
 
@@ -158,18 +186,28 @@ export function useCalendarMonth(
   maxAllowedDate: Date,
   deps: readonly CalendarMonthResetDep[],
 ) {
-  const resetMonthKey = formatCalendarDate(
-    getInitialCalendarMonth(selectedDate, availableDates, maxAllowedDate),
-  );
-  const availableDatesKey = availableDates?.join(',') ?? '';
-  const maxAllowedDateKey = formatCalendarDate(maxAllowedDate);
-  const selectionResetKey = deps.map((dep) => String(dep ?? '')).join('\u001F');
-  const [currentMonth, setCurrentMonth] = useState(() => parseCalendarDate(resetMonthKey));
-
-  useEffect(() => {
-    setCurrentMonth(parseCalendarDate(resetMonthKey));
-  }, [availableDatesKey, maxAllowedDateKey, resetMonthKey, selectionResetKey]);
+  const initialMonth = getInitialCalendarMonth(selectedDate, availableDates, maxAllowedDate);
+  const resetMonthKey = formatCalendarDate(initialMonth);
+  const resetKey = [
+    availableDates?.join(',') ?? '',
+    formatCalendarDate(maxAllowedDate),
+    resetMonthKey,
+    deps.map((dep) => String(dep ?? '')).join('\u001F'),
+  ].join('\u001E');
+  const [state, dispatch] = useReducer(calendarMonthReducer, {
+    resetKey,
+    currentMonth: initialMonth,
+  });
+
+  let currentMonth = state.currentMonth;
+  if (state.resetKey !== resetKey) {
+    currentMonth = initialMonth;
+    dispatch({ type: 'reset', resetKey, currentMonth });
+  }
 
+  const setCurrentMonth: Dispatch<SetStateAction<Date>> = (nextMonth) => {
+    dispatch({ type: 'set', currentMonth: nextMonth });
+  };
   return [currentMonth, setCurrentMonth] as const;
 }
 
@@ -228,8 +266,18 @@ export function CalendarMonthPanel({
   onDateClick,
   onDateHover,
 }: CalendarMonthPanelProps) {
-  const monthName = month.toLocaleDateString('en-US', { month: 'long', year: 'numeric' });
+  const locale = useLocale();
+  const monthName = month.toLocaleDateString(locale === 'zh' ? 'zh-CN' : 'en-US', {
+    month: 'long',
+    year: 'numeric',
+  });
+  const weekdays =
+    locale === 'zh'
+      ? ['日', '一', '二', '三', '四', '五', '六']
+      : ['Su', 'Mo', 'Tu', 'We', 'Th', 'Fr', 'Sa'];
   const days = getCalendarMonthDays(month);
+  const previousMonthLabel = locale === 'zh' ? '上个月' : 'Previous month';
+  const nextMonthLabel = locale === 'zh' ? '下个月' : 'Next month';
 
   return (
     <div className="space-y-4">
@@ -241,6 +289,7 @@ export function CalendarMonthPanel({
             onClick={onPreviousMonth}
             disabled={isDisabled || !canGoPrevious}
             className={cn(!canGoPrevious && 'opacity-30')}
+            aria-label={previousMonthLabel}
           >
             ‹
           </Button>
@@ -255,6 +304,7 @@ export function CalendarMonthPanel({
             onClick={onNextMonth}
             disabled={isDisabled || !canGoNext}
             className={cn(!canGoNext && 'opacity-30')}
+            aria-label={nextMonthLabel}
           >
             ›
           </Button>
@@ -264,7 +314,7 @@ export function CalendarMonthPanel({
       </div>
 
       <div className="grid grid-cols-7 gap-2">
-        {['Su', 'Mo', 'Tu', 'We', 'Th', 'Fr', 'Sa'].map((weekday) => (
+        {weekdays.map((weekday) => (
           <div key={weekday} className="text-center text-xs font-medium text-muted-foreground py-2">
             {weekday}
           </div>
diff --git a/packages/app/src/components/ui/chart-display-helpers.test.tsx b/packages/app/src/components/ui/chart-display-helpers.test.tsx
index 0aa96af48..9f77a4477 100644
--- a/packages/app/src/components/ui/chart-display-helpers.test.tsx
+++ b/packages/app/src/components/ui/chart-display-helpers.test.tsx
@@ -3,6 +3,8 @@ import React, { act } from 'react';
 import { createRoot, type Root } from 'react-dom/client';
 import { afterEach, beforeEach, describe, expect, it } from 'vitest';
 
+import { TCO_SOURCE_TITLE, TCO_SOURCE_URL } from '@semianalysisai/inferencex-constants';
+
 import { ChartShareActions, MetricAssumptionNotes } from '@/components/ui/chart-display-helpers';
 
 let container: HTMLDivElement;
@@ -52,6 +54,30 @@ describe('MetricAssumptionNotes', () => {
     expect(getVisibleCaveatText()).toContain('calculate power per decode chip or per prefill chip');
   });
 
+  // Total tok/s/MW divides throughput per chip overall by per-chip power — the
+  // same denominator an aggregated config uses — so, like the total-token cost
+  // metrics, it keeps the power badges but must not carry the disagg caveat.
+  it('hides the disaggregation caveat for the total per-MW metric', () => {
+    renderUi(<MetricAssumptionNotes selectedYAxisMetric="y_tpPerMw" />);
+
+    expect(getVisibleText()).toContain('All in Power/Chip:');
+    expect(getVisibleText()).toContain('SemiAnalysis Datacenter Industry Model');
+    expect(getVisibleCaveatText()).not.toContain(
+      'calculate power per decode chip or per prefill chip',
+    );
+  });
+
+  it.each(['y_inputTputPerMw', 'y_outputTputPerMw'])(
+    'shows the disaggregation caveat for per-token-type per-MW metric %s',
+    (metric) => {
+      renderUi(<MetricAssumptionNotes selectedYAxisMetric={metric} />);
+
+      expect(getVisibleCaveatText()).toContain(
+        'calculate power per decode chip or per prefill chip',
+      );
+    },
+  );
+
   it('preserves historical-trends semantics when both compatibility flags are disabled', () => {
     renderUi(
       <MetricAssumptionNotes
@@ -80,45 +106,116 @@ describe('MetricAssumptionNotes', () => {
     );
   });
 
-  it('renders TCO notes, source attribution, and the cost disaggregation caveat', () => {
-    renderUi(<MetricAssumptionNotes selectedYAxisMetric="y_costhOutput" />);
+  it('renders TCO notes, source attribution, and the purchasing-power caveat', () => {
+    renderUi(<MetricAssumptionNotes selectedYAxisMetric="y_outputTokensPerDollarH" />);
 
     expect(getVisibleText()).toContain('TCO $/chip/hr:');
-    expect(getVisibleText()).toContain(
-      'SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model',
+    expect(getVisibleText()).toContain(TCO_SOURCE_TITLE);
+    expect(container.querySelector(`a[href="${TCO_SOURCE_URL}"]`)).not.toBeNull();
+    expect(getVisibleCaveatText()).toContain(
+      'calculate tokens per $1 USD per decode chip or per prefill chip',
     );
-    expect(getVisibleCaveatText()).toContain('calculate cost per decode chip or per prefill chip');
   });
 
-  // The prefill/decode split only skews the per-token-type costs; the
-  // total-token cost divides by the whole chip count, exactly as an aggregated
+  it('describes the existing cost-per-million metric', () => {
+    renderUi(<MetricAssumptionNotes selectedYAxisMetric="y_costhOutput" />);
+
+    expect(getVisibleCaveatText()).toContain(
+      'calculate cost per million tokens per decode chip or per prefill chip',
+    );
+    expect(getVisibleCaveatText()).toContain('token cost comparison');
+  });
+
+  // The prefill/decode split only skews per-token-type metrics; the
+  // total-token metric divides by the whole chip count, exactly as an aggregated
   // config does, so it must not carry the caveat.
+  it.each([
+    'y_outputTokensPerDollarH',
+    'y_outputTokensPerDollarN',
+    'y_outputTokensPerDollarR',
+    'y_inputTokensPerDollarH',
+    'y_inputTokensPerDollarN',
+    'y_inputTokensPerDollarR',
+  ])('shows the purchasing-power caveat for per-token-type metric %s', (metric) => {
+    renderUi(<MetricAssumptionNotes selectedYAxisMetric={metric} />);
+
+    expect(getVisibleCaveatText()).toContain(
+      'calculate tokens per $1 USD per decode chip or per prefill chip',
+    );
+  });
+
   it.each(['y_costhOutput', 'y_costnOutput', 'y_costrOutput', 'y_costhi', 'y_costni', 'y_costri'])(
-    'shows the cost disaggregation caveat for per-token-type cost metric %s',
+    'shows the token-cost caveat for per-token-type metric %s',
     (metric) => {
       renderUi(<MetricAssumptionNotes selectedYAxisMetric={metric} />);
 
       expect(getVisibleCaveatText()).toContain(
-        'calculate cost per decode chip or per prefill chip',
+        'calculate cost per million tokens per decode chip or per prefill chip',
       );
     },
   );
 
-  it.each(['y_costh', 'y_costn', 'y_costr'])(
-    'hides the cost disaggregation caveat for total-token cost metric %s',
-    (metric) => {
-      renderUi(<MetricAssumptionNotes selectedYAxisMetric={metric} />);
+  it.each([
+    'y_costh',
+    'y_costn',
+    'y_costr',
+    'y_tokensPerDollarH',
+    'y_tokensPerDollarN',
+    'y_tokensPerDollarR',
+  ])('hides the purchasing-power caveat for total-token metric %s', (metric) => {
+    renderUi(<MetricAssumptionNotes selectedYAxisMetric={metric} />);
+
+    // The TCO badges and source attribution still explain the hourly-price input.
+    expect(getVisibleText()).toContain('TCO $/chip/hr:');
+    expect(getVisibleText()).toContain(TCO_SOURCE_TITLE);
+    expect(container.querySelector(`a[href="${TCO_SOURCE_URL}"]`)).not.toBeNull();
+    expect(getVisibleCaveatText()).not.toContain(
+      'calculate tokens per $1 USD per decode chip or per prefill chip',
+    );
+  });
 
-      // The TCO badges and source attribution still belong on a cost chart.
-      expect(getVisibleText()).toContain('TCO $/chip/hr:');
-      expect(getVisibleText()).toContain(
-        'SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model',
-      );
-      expect(getVisibleCaveatText()).not.toContain(
-        'calculate cost per decode chip or per prefill chip',
-      );
-    },
-  );
+  it('narrows the TCO badges to the base GPUs of the active legend selection', () => {
+    renderUi(
+      <MetricAssumptionNotes
+        selectedYAxisMetric="y_tokensPerDollarH"
+        activeHwKeys={['h200_dynamo-sglang', 'gb300_dynamo-sglang']}
+      />,
+    );
+
+    expect(getVisibleText()).toContain('TCO $/chip/hr:');
+    expect(getVisibleText()).toContain('H200:');
+    expect(getVisibleText()).toContain('GB300:');
+    expect(getVisibleText()).not.toContain('H100:');
+    expect(getVisibleText()).not.toContain('MI300X:');
+  });
+
+  it('narrows the power badges to the active legend selection', () => {
+    renderUi(
+      <MetricAssumptionNotes selectedYAxisMetric="y_tpPerMw" activeHwKeys={new Set(['mi300x'])} />,
+    );
+
+    expect(getVisibleText()).toContain('All in Power/Chip:');
+    expect(getVisibleText()).toContain('MI300X:');
+    expect(getVisibleText()).not.toContain('H100:');
+    expect(getVisibleText()).not.toContain('H200:');
+  });
+
+  it('falls back to every registry GPU when the selection is empty or unrecognized', () => {
+    renderUi(<MetricAssumptionNotes selectedYAxisMetric="y_tokensPerDollarH" activeHwKeys={[]} />);
+
+    expect(getVisibleText()).toContain('H100:');
+    expect(getVisibleText()).toContain('MI300X:');
+
+    renderUi(
+      <MetricAssumptionNotes
+        selectedYAxisMetric="y_tokensPerDollarH"
+        activeHwKeys={['not-a-gpu_dynamo-sglang']}
+      />,
+    );
+
+    expect(getVisibleText()).toContain('H100:');
+    expect(getVisibleText()).toContain('MI300X:');
+  });
 
   it('renders metric-specific throughput caveats and preserves Joules wording semantics', () => {
     renderUi(<MetricAssumptionNotes selectedYAxisMetric="y_inputTputPerGpu" />);
diff --git a/packages/app/src/components/ui/chart-display-helpers.tsx b/packages/app/src/components/ui/chart-display-helpers.tsx
index c2e98d068..9be67d1ae 100644
--- a/packages/app/src/components/ui/chart-display-helpers.tsx
+++ b/packages/app/src/components/ui/chart-display-helpers.tsx
@@ -1,21 +1,62 @@
 'use client';
 
 import Link from 'next/link';
-import type { ReactNode } from 'react';
+import { useMemo, type ReactNode } from 'react';
+
+import {
+  HW_REGISTRY,
+  TCO_SOURCE_TITLE,
+  TCO_SOURCE_URL,
+} from '@semianalysisai/inferencex-constants';
 
 import { Badge } from '@/components/ui/badge';
 import { ExternalLinkIcon } from '@/components/ui/external-link-icon';
 import { ShareButton } from '@/components/ui/share-button';
-import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
 import { useLocale } from '@/lib/use-locale';
 import type { Locale } from '@/lib/i18n';
 
 // Keep these metric-key groups in sync with chart-utils/chart configs when new source-backed
 // metrics are added; this helper owns which caption notes and caveats appear for each family.
 const POWER_SOURCE_METRICS = new Set(['y_tpPerMw', 'y_inputTputPerMw', 'y_outputTputPerMw']);
-const TOTAL_COST_METRICS = new Set(['y_costh', 'y_costn', 'y_costr']);
-const OUTPUT_COST_METRICS = new Set(['y_costhOutput', 'y_costnOutput', 'y_costrOutput']);
-const INPUT_COST_METRICS = new Set(['y_costhi', 'y_costni', 'y_costri']);
+// The disaggregation caveat only applies to the per-token-type per-MW metrics: a
+// disaggregated run reports input/output throughput per prefill or per decode chip,
+// so dividing by per-chip power inherits that skew. Total tok/s/MW divides
+// throughput per chip overall by the same per-chip power an aggregated config
+// uses, so it needs no caveat — the same split the cost caveats below make.
+const PER_TOKEN_TYPE_POWER_METRICS = new Set(['y_inputTputPerMw', 'y_outputTputPerMw']);
+const TOTAL_COST_METRICS = new Set([
+  'y_costh',
+  'y_costn',
+  'y_costr',
+  'y_tokensPerDollarH',
+  'y_tokensPerDollarN',
+  'y_tokensPerDollarR',
+  'y_tokensPerRmbH',
+  'y_tokensPerRmbN',
+  'y_tokensPerRmbR',
+]);
+const OUTPUT_COST_METRICS = new Set([
+  'y_costhOutput',
+  'y_costnOutput',
+  'y_costrOutput',
+  'y_outputTokensPerDollarH',
+  'y_outputTokensPerDollarN',
+  'y_outputTokensPerDollarR',
+  'y_outputTokensPerRmbH',
+  'y_outputTokensPerRmbN',
+  'y_outputTokensPerRmbR',
+]);
+const INPUT_COST_METRICS = new Set([
+  'y_costhi',
+  'y_costni',
+  'y_costri',
+  'y_inputTokensPerDollarH',
+  'y_inputTokensPerDollarN',
+  'y_inputTokensPerDollarR',
+  'y_inputTokensPerRmbH',
+  'y_inputTokensPerRmbN',
+  'y_inputTokensPerRmbR',
+]);
 const POWER_VALUES = Object.fromEntries(
   Object.entries(HW_REGISTRY).map(([base, specs]) => [base, `${specs.power}kW`]),
 );
@@ -32,7 +73,7 @@ function MetricBadges({
       {label}{' '}
       {Object.entries(values).map(([base, value]) => (
         <Badge key={base} variant="outline">
-          {base.toUpperCase()}: {value}
+          {HW_REGISTRY[base]?.badgeLabel ?? base.toUpperCase()}: {value}
         </Badge>
       ))}
     </p>
@@ -63,6 +104,11 @@ function SourceLink({
 
 const NOUN_ZH: Record<string, string> = {
   cost: '成本',
+  'cost per million tokens': '每百万 token 成本',
+  'token cost': 'token 成本',
+  'tokens per $1 USD': '每 1 美元可购买的 token 数',
+  'tokens per ¥1 RMB': '每 1 元人民币可购买的 token 数',
+  'purchasing power': '购买力',
   'input throughput': '输入吞吐量',
   'output throughput': '输出吞吐量',
   power: '功耗',
@@ -84,10 +130,10 @@ function DisaggCaveat({
   const content =
     locale === 'zh' ? (
       <>
-        <strong>注意：</strong>分离式推理配置（如 MoRI SGLang、Dynamo TRTLLM）按解码 Chip 或预填充
-        Chip 计算
+        <strong>注意：</strong>分离式推理配置（如 MoRI SGLang、Dynamo TRTLLM）按解码芯片或预填充
+        芯片计算
         {NOUN_ZH[calculationNoun] ?? calculationNoun}
-        ，而非按 Chip 总数计算。因此，与聚合配置进行
+        ，而非按芯片总数计算。因此，与聚合配置进行
         {NOUN_ZH[comparisonNoun] ?? comparisonNoun}
         的直接对比并不完全等价。
       </>
@@ -119,11 +165,23 @@ function getCostValues(selectedYAxisMetric: string) {
       base,
       selectedYAxisMetric === 'y_costh' ||
       selectedYAxisMetric === 'y_costhOutput' ||
-      selectedYAxisMetric === 'y_costhi'
+      selectedYAxisMetric === 'y_costhi' ||
+      selectedYAxisMetric === 'y_tokensPerRmbH' ||
+      selectedYAxisMetric === 'y_outputTokensPerRmbH' ||
+      selectedYAxisMetric === 'y_inputTokensPerRmbH' ||
+      selectedYAxisMetric === 'y_tokensPerDollarH' ||
+      selectedYAxisMetric === 'y_outputTokensPerDollarH' ||
+      selectedYAxisMetric === 'y_inputTokensPerDollarH'
         ? specs.costh
         : selectedYAxisMetric === 'y_costn' ||
             selectedYAxisMetric === 'y_costnOutput' ||
-            selectedYAxisMetric === 'y_costni'
+            selectedYAxisMetric === 'y_costni' ||
+            selectedYAxisMetric === 'y_tokensPerRmbN' ||
+            selectedYAxisMetric === 'y_outputTokensPerRmbN' ||
+            selectedYAxisMetric === 'y_inputTokensPerRmbN' ||
+            selectedYAxisMetric === 'y_tokensPerDollarN' ||
+            selectedYAxisMetric === 'y_outputTokensPerDollarN' ||
+            selectedYAxisMetric === 'y_inputTokensPerDollarN'
           ? specs.costn
           : specs.costr,
     ]),
@@ -136,16 +194,43 @@ export function ChartShareActions() {
 
 export function MetricAssumptionNotes({
   selectedYAxisMetric,
+  activeHwKeys,
   includeAllPowerThroughputMetrics = true,
   includePowerThroughputCaveat = true,
 }: {
   selectedYAxisMetric: string;
+  /**
+   * Active legend hardware keys (e.g. `gb300_dynamo-sglang`). When provided,
+   * the TCO $/chip/hr and Power/Chip badges are narrowed to the base GPUs the
+   * selection covers, so the caption only quotes chips that can appear on the
+   * plot. Omitted (or when the selection maps to no registry GPU) every
+   * registry GPU is shown, preserving the historical behavior.
+   */
+  activeHwKeys?: ReadonlySet<string> | readonly string[];
   // Historical trends only annotates y_tpPerMw and intentionally omits per-MW caveats to preserve
   // the tab's existing caption contract while sharing the same helper as inference.
   includeAllPowerThroughputMetrics?: boolean;
   includePowerThroughputCaveat?: boolean;
 }) {
   const locale = useLocale();
+  // Legend keys are `{base}` or `{base}_{framework/variant}`; badge maps are
+  // keyed by registry base, so reduce the selection to its base GPUs.
+  const activeBases = useMemo(() => {
+    const bases = new Set<string>();
+    for (const key of activeHwKeys ?? []) {
+      const base = key.split('_')[0];
+      if (base in HW_REGISTRY) bases.add(base);
+    }
+    return bases;
+  }, [activeHwKeys]);
+  const filterToActive = (values: Record<string, string | number>) => {
+    if (activeBases.size === 0) return values;
+    const filtered = Object.fromEntries(
+      Object.entries(values).filter(([base]) => activeBases.has(base)),
+    );
+    // Defensive: never render a badge row with an empty value list.
+    return Object.keys(filtered).length > 0 ? filtered : values;
+  };
   const showPowerSource = includeAllPowerThroughputMetrics
     ? POWER_SOURCE_METRICS.has(selectedYAxisMetric)
     : selectedYAxisMetric === 'y_tpPerMw';
@@ -154,13 +239,16 @@ export function MetricAssumptionNotes({
   const showInputCostSource = INPUT_COST_METRICS.has(selectedYAxisMetric);
   const showInputThroughputCaveat = selectedYAxisMetric === 'y_inputTputPerGpu';
   const showOutputThroughputCaveat = selectedYAxisMetric === 'y_outputTputPerGpu';
-  // Per-token-type cost only. A disagg config's prefill and decode chips are
-  // counted separately, so the input- and output-token costs are attributed to
-  // one side of the split and can't be lined up against an aggregated config.
-  // The total-token cost divides by the whole chip count, which is the same
+  // Per-token-type cost and purchasing power only. A disagg config's prefill and decode
+  // chips are counted separately, so input/output economics are attributed to one side
+  // of the split and can't be lined up against an aggregated config.
+  // The total-token metric uses the whole chip count, which is the same
   // denominator an aggregated config uses, so it needs no caveat — the same
   // split the throughput caveats above already make (input/output, not total).
   const showCostCaveat = showOutputCostSource || showInputCostSource;
+  const isTokensPerDollar = selectedYAxisMetric.includes('TokensPerDollar');
+  const isTokensPerRmb = selectedYAxisMetric.includes('TokensPerRmb');
+  const isTokensPerCurrency = isTokensPerDollar || isTokensPerRmb;
   const showJouleSource = selectedYAxisMetric.startsWith('y_j');
 
   const costValues =
@@ -168,15 +256,15 @@ export function MetricAssumptionNotes({
       ? getCostValues(selectedYAxisMetric)
       : null;
 
-  const powerLabel = locale === 'zh' ? '全包功耗/Chip：' : 'All in Power/Chip:';
-  const costLabel = locale === 'zh' ? 'TCO $/chip/小时：' : 'TCO $/chip/hr:';
+  const powerLabel = locale === 'zh' ? '全含功率/芯片：' : 'All in Power/Chip:';
+  const costLabel = locale === 'zh' ? 'TCO $/chip/hr：' : 'TCO $/chip/hr:';
   const sourceLabel = locale === 'zh' ? '来源：' : 'Source:';
 
   return (
     <>
       {showPowerSource && (
         <>
-          <MetricBadges label={powerLabel} values={POWER_VALUES} />
+          <MetricBadges label={powerLabel} values={filterToActive(POWER_VALUES)} />
           <SourceLink
             href="https://semianalysis.com/datacenter-industry-model/"
             sourceLabel={sourceLabel}
@@ -187,13 +275,24 @@ export function MetricAssumptionNotes({
       )}
       {costValues && (
         <>
-          <MetricBadges label={costLabel} values={costValues} />
-          <SourceLink href="https://semianalysis.com/ai-cloud-tco-model/" sourceLabel={sourceLabel}>
-            SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model
+          <MetricBadges label={costLabel} values={filterToActive(costValues)} />
+          <SourceLink href={TCO_SOURCE_URL} sourceLabel={sourceLabel}>
+            {TCO_SOURCE_TITLE}
           </SourceLink>
         </>
       )}
-      <DisaggCaveat visible={showCostCaveat} calculationNoun="cost" locale={locale} />
+      <DisaggCaveat
+        visible={showCostCaveat}
+        calculationNoun={
+          isTokensPerRmb
+            ? 'tokens per ¥1 RMB'
+            : isTokensPerDollar
+              ? 'tokens per $1 USD'
+              : 'cost per million tokens'
+        }
+        comparisonNoun={isTokensPerCurrency ? 'purchasing power' : 'token cost'}
+        locale={locale}
+      />
       <DisaggCaveat
         visible={showInputThroughputCaveat}
         calculationNoun="input throughput"
@@ -206,14 +305,14 @@ export function MetricAssumptionNotes({
       />
       {includePowerThroughputCaveat && (
         <DisaggCaveat
-          visible={POWER_SOURCE_METRICS.has(selectedYAxisMetric)}
+          visible={PER_TOKEN_TYPE_POWER_METRICS.has(selectedYAxisMetric)}
           calculationNoun="power"
           locale={locale}
         />
       )}
       {showJouleSource && (
         <>
-          <MetricBadges label={powerLabel} values={POWER_VALUES} />
+          <MetricBadges label={powerLabel} values={filterToActive(POWER_VALUES)} />
           <SourceLink
             href="https://semianalysis.com/datacenter-industry-model/"
             sourceLabel={sourceLabel}
diff --git a/packages/app/src/components/ui/chart-legend-item.tsx b/packages/app/src/components/ui/chart-legend-item.tsx
index 69b1a0d6a..deae0f677 100644
--- a/packages/app/src/components/ui/chart-legend-item.tsx
+++ b/packages/app/src/components/ui/chart-legend-item.tsx
@@ -97,7 +97,7 @@ const ChartLegendItem: React.FC<CommonLegendItemProps> = ({
         htmlFor={id}
         className={cn(
           'group/item flex items-center cursor-pointer hover:underline peer-focus-visible:ring-1 peer-focus-visible:ring-offset-1 peer-focus-visible:outline-none rounded-sm',
-          isLegendExpanded ? 'w-fit whitespace-nowrap' : '',
+          isLegendExpanded ? 'w-fit whitespace-nowrap' : 'max-w-full',
         )}
         title={!isLegendExpanded && isLongText ? label : title}
         onMouseEnter={onHover && isActive ? () => onHover(hw || name) : undefined}
@@ -170,7 +170,7 @@ const ChartLegendItem: React.FC<CommonLegendItemProps> = ({
           className={cn(
             'pr-2',
             isLongText && !isLegendExpanded
-              ? 'truncate'
+              ? 'truncate min-w-0'
               : isLegendExpanded
                 ? 'whitespace-nowrap'
                 : 'whitespace-normal',
diff --git a/packages/app/src/components/ui/chart-legend.tsx b/packages/app/src/components/ui/chart-legend.tsx
index 400132cad..86a22f6d7 100644
--- a/packages/app/src/components/ui/chart-legend.tsx
+++ b/packages/app/src/components/ui/chart-legend.tsx
@@ -1,19 +1,17 @@
 'use client';
 
-import { track } from '@/lib/analytics';
 import {
-  ArrowLeftToLine,
-  ArrowRightToLine,
   ChevronDown,
   ChevronRight,
   Circle,
   Diamond,
   Info,
+  PanelRight,
   Square,
   Triangle,
   X,
 } from 'lucide-react';
-import React, { useCallback, useId, useLayoutEffect, useMemo, useRef, useState } from 'react';
+import React, { useId, useLayoutEffect, useMemo, useRef, useState } from 'react';
 
 import { SHAPE_ORDER, type ShapeKey, getShapeKeyForPrecision } from '@/lib/chart-rendering';
 import { type Precision, getPrecisionLabel } from '@/lib/data-mappings';
@@ -28,6 +26,7 @@ const SHAPE_ICON: Record<ShapeKey, React.ComponentType<{ size?: number; classNam
   diamond: Diamond,
 };
 
+import { ATOM_FOOTNOTE_MARKER, AtomEngineFootnote } from './atom-engine-footnote';
 import ChartLegendItem, { type CommonLegendItemProps } from './chart-legend-item';
 import { Label } from './label';
 import { Switch } from './switch';
@@ -38,32 +37,19 @@ export type { CommonLegendItemProps } from './chart-legend-item';
 const STRINGS = {
   en: {
     advanced: 'Advanced',
-    collapse: 'Collapse',
-    expand: 'Expand',
-    searchPlaceholder: 'Search...',
-    searchAria: 'Search legend',
-    clearSearch: 'Clear search',
     moreInfo: (label: string) => `More info about ${label}`,
-    collapseLegend: 'Collapse legend',
-    expandLegend: 'Expand legend',
+    hideLegend: 'Hide legend',
+    showLegend: 'Show legend',
     hide: (label: string) => `Hide ${label}`,
     showPoints: (label: string) => `Show all ${label} data points`,
-    atomFootnote:
-      'The ATOM engine is promising, however it has yet to serve production tokens. It is still in its infant stage.',
   },
   zh: {
     advanced: '高级',
-    collapse: '收起',
-    expand: '展开',
-    searchPlaceholder: '搜索…',
-    searchAria: '搜索图例',
-    clearSearch: '清除搜索',
     moreInfo: (label: string) => `查看${label}的更多信息`,
-    collapseLegend: '收起图例',
-    expandLegend: '展开图例',
+    hideLegend: '隐藏图例',
+    showLegend: '显示图例',
     hide: (label: string) => `隐藏${label}`,
     showPoints: (label: string) => `显示${label}的全部数据点`,
-    atomFootnote: 'ATOM 引擎前景可期，但尚未用于生产环境 token 服务，仍处于早期阶段。',
   },
 } as const;
 
@@ -109,6 +95,12 @@ export interface ChartLegendProps {
   onItemHoverEnd?: () => void;
   onItemRemove?: (name: string) => void;
   onAdvancedExpandedChange?: (expanded: boolean) => void;
+  /**
+   * Suppress the ATOM engine footnote even when a legend label carries the ¹
+   * marker. Inference charts pass this because they render the footnote in
+   * the axis-metric info footer below the chart instead of in the legend.
+   */
+  hideAtomFootnote?: boolean;
 }
 
 export default function ChartLegend({
@@ -129,18 +121,20 @@ export default function ChartLegend({
   onItemHoverEnd,
   onItemRemove,
   onAdvancedExpandedChange,
+  hideAtomFootnote = false,
 }: ChartLegendProps) {
   const locale = useLocale();
   const t = STRINGS[locale];
   const isSidebar = variant === 'sidebar';
-  const [hasLongText, setHasLongText] = useState(false);
   const scrollRef = useRef<HTMLDivElement>(null);
   const [isOverflowing, setIsOverflowing] = useState(false);
-  const [searchQuery, setSearchQuery] = useState('');
   const [isAdvancedExpanded, setIsAdvancedExpanded] = useState(false);
   const advancedControlsId = useId();
 
-  const effectiveExpanded = isLegendExpanded;
+  // Sidebar items always render in the compact (truncate + title tooltip)
+  // layout: the panel is a fixed-width in-flow column and must never grow
+  // over the plot area.
+  const itemsExpanded = isSidebar ? false : isLegendExpanded;
   // Counts only removable series: the guard below exists to stop the user
   // emptying the chart, and label-only entries (unofficial runs) are not
   // something removing leaves you without.
@@ -152,10 +146,6 @@ export default function ChartLegend({
   const removeFor = (item: CommonLegendItemProps) =>
     item.isRemovable === false ? undefined : effectiveRemove;
 
-  useLayoutEffect(() => {
-    setHasLongText(legendItems.some((item) => item.label && item.label.length > 8));
-  }, [legendItems]);
-
   useLayoutEffect(() => {
     const el = scrollRef.current;
     if (!el) return;
@@ -172,22 +162,6 @@ export default function ChartLegend({
     return filterAndSortLegendItems(legendItems, '', !disableActiveSort);
   }, [legendItems, isSidebar, disableActiveSort]);
 
-  // Compute which items match the search query (used to hide non-matching via CSS)
-  const hiddenNames = useMemo(() => {
-    if (!isSidebar) return new Set<string>();
-    const query = searchQuery.trim().toLowerCase();
-    if (!query) return new Set<string>();
-    return new Set(
-      legendItems
-        .filter(
-          (item) =>
-            !item.label.toLowerCase().includes(query) &&
-            !(item.title && item.title.toLowerCase().includes(query)),
-        )
-        .map((item) => item.name),
-    );
-  }, [legendItems, searchQuery, isSidebar]);
-
   const rows = useMemo(() => {
     if (!grouped) return null;
     const items = isSidebar ? sortedItems : legendItems;
@@ -207,17 +181,15 @@ export default function ChartLegend({
     return result.filter((row) => row.length > 0);
   }, [grouped, legendItems, sortedItems, isSidebar]);
 
-  const handleLegendExpand = () => {
+  const toggleLegendOpen = () => {
     onExpandedChange(!isLegendExpanded);
   };
 
+  // Sidebar: a fixed in-flow panel that takes layout space next to the plot
+  // (Epoch-style). It never overlays the chart; closing it (X) removes it
+  // entirely and the chart reclaims the width.
   const outerClasses = isSidebar
-    ? cn(
-        'p-2 rounded-sm text-sm flex flex-col h-full legend-container sidebar-legend transition-all',
-        isLegendExpanded
-          ? 'absolute right-0 top-0 z-10 w-auto min-w-fit border bg-accent'
-          : 'w-full',
-      )
+    ? 'p-3 rounded-md border border-border/60 bg-background text-sm flex flex-col h-full legend-container sidebar-legend w-full'
     : grouped
       ? cn(
           'py-1 px-2 md:py-1 rounded-sm border text-sm top-0 right-0 bg-accent transition-all md:flex md:flex-col legend-container',
@@ -239,8 +211,9 @@ export default function ChartLegend({
 
   // Show ATOM footnote when any legend item label contains the ¹ marker
   const hasAtomFootnote = useMemo(
-    () => legendItems.some((item) => item.label.includes('¹')),
-    [legendItems],
+    () =>
+      !hideAtomFootnote && legendItems.some((item) => item.label.includes(ATOM_FOOTNOTE_MARKER)),
+    [legendItems, hideAtomFootnote],
   );
 
   const shapeIndicators = useMemo(() => {
@@ -256,8 +229,7 @@ export default function ChartLegend({
     (shapeIndicators !== null ||
       keyIndicators ||
       (switches && switches.length > 0) ||
-      (actions && actions.length > 0) ||
-      hasLongText);
+      (actions && actions.length > 0));
   const scrollClasses = isSidebar
     ? cn(
         'overflow-y-auto flex-1 min-h-0 space-y-0.5',
@@ -267,41 +239,40 @@ export default function ChartLegend({
       ? 'flex gap-x-4 flex-wrap flex-row md:block md:overflow-y-auto md:flex-1 md:min-h-0'
       : 'flex flex-row flex-wrap gap-x-4 gap-y-2 md:block md:overflow-y-auto md:flex-1 md:min-h-0';
 
-  const trackSearchOnBlur = useCallback(() => {
-    if (searchQuery.trim()) {
-      track('inference_legend_searched', { query: searchQuery.trim() });
-    }
-  }, [searchQuery]);
-
-  const searchInput =
-    isSidebar && isLegendExpanded ? (
-      <div className="pb-1.5 no-export">
-        <div className="relative">
-          <input
-            type="text"
-            value={searchQuery}
-            onChange={(e) => setSearchQuery(e.target.value)}
-            onBlur={trackSearchOnBlur}
-            placeholder={t.searchPlaceholder}
-            aria-label={t.searchAria}
-            className="w-full px-2 py-1 pr-6 rounded-md border border-border bg-background text-foreground text-xs placeholder:text-muted-foreground focus:outline-none focus:ring-1 focus:ring-sky-500/50 focus:border-sky-500/50"
-          />
-          {searchQuery && (
-            <button
-              type="button"
-              onClick={() => {
-                track('inference_legend_search_cleared');
-                setSearchQuery('');
-              }}
-              className="absolute right-1.5 top-1/2 -translate-y-1/2 p-0.5 rounded-sm text-muted-foreground hover:text-foreground transition-colors"
-              aria-label={t.clearSearch}
-            >
-              <X size={12} />
-            </button>
-          )}
-        </div>
+  // Fully-closed sidebar: render only a reopen affordance at the top-right of
+  // the chart area (all hooks above have already run unconditionally).
+  if (isSidebar && !isLegendExpanded) {
+    return (
+      <div className="flex justify-end no-export">
+        <button
+          type="button"
+          data-testid="legend-open-button"
+          onClick={toggleLegendOpen}
+          aria-label={t.showLegend}
+          title={t.showLegend}
+          className="p-1.5 rounded-md border border-border bg-background text-muted-foreground hover:text-foreground transition-colors cursor-pointer"
+        >
+          <PanelRight size={16} />
+        </button>
       </div>
-    ) : null;
+    );
+  }
+
+  // Slim header: just the close (X) affordance, right-aligned (Epoch-style).
+  const panelHeader = isSidebar ? (
+    <div className="pb-1 no-export flex justify-end">
+      <button
+        type="button"
+        data-testid="legend-close-button"
+        onClick={toggleLegendOpen}
+        aria-label={t.hideLegend}
+        title={t.hideLegend}
+        className="p-1 rounded-sm text-muted-foreground hover:text-foreground transition-colors cursor-pointer"
+      >
+        <X size={14} />
+      </button>
+    </div>
+  ) : null;
 
   const standardSwitches = switches?.filter((sw) => !sw.advanced) ?? [];
   const advancedSwitches = switches?.filter((sw) => sw.advanced) ?? [];
@@ -407,7 +378,7 @@ export default function ChartLegend({
     <div
       className={cn(
         'w-full md:w-auto mt-2 px-1 pr-2 gap-x-4 gap-y-1',
-        effectiveExpanded ? 'flex flex-wrap' : 'grid grid-cols-2',
+        itemsExpanded ? 'flex flex-wrap' : 'grid grid-cols-2',
       )}
     >
       {shapeIndicators.map(({ precision, shapeKey }) => {
@@ -424,22 +395,8 @@ export default function ChartLegend({
     </div>
   ) : null;
 
-  const expandButton = hasLongText ? (
-    <div className="hidden lg:block mt-2 no-export">
-      <button
-        type="button"
-        onClick={handleLegendExpand}
-        className="text-xs text-accent-foreground hover:text-foreground flex items-center gap-1"
-        aria-label={isLegendExpanded ? t.collapseLegend : t.expandLegend}
-      >
-        {isLegendExpanded ? <ArrowRightToLine size={16} /> : <ArrowLeftToLine size={16} />}
-        {isLegendExpanded ? t.collapse : t.expand}
-      </button>
-    </div>
-  ) : null;
-
   // Compute li className for a legend item (shared by tooltip and non-tooltip paths)
-  const itemClassName = (item: CommonLegendItemProps, isHidden: boolean) =>
+  const itemClassName = (item: CommonLegendItemProps) =>
     cn(
       'transition-opacity duration-300',
       isSidebar
@@ -449,12 +406,11 @@ export default function ChartLegend({
         : item.isActive
           ? 'opacity-100'
           : 'opacity-50 no-export',
-      effectiveExpanded && 'md:w-full md:block',
-      isHidden && 'h-0 m-0! p-0! overflow-hidden',
+      itemsExpanded && 'md:w-full md:block',
     );
 
   // Render a single legend item, optionally wrapped with a tooltip
-  const renderItem = (item: CommonLegendItemProps, isHidden: boolean) => {
+  const renderItem = (item: CommonLegendItemProps) => {
     const legendItem = (
       <ChartLegendItem
         name={item.name}
@@ -473,19 +429,19 @@ export default function ChartLegend({
         hideAriaLabel={t.hide(item.label)}
         showPointsAriaLabel={t.showPoints(item.label)}
         asFragment
-        isLegendExpanded={effectiveExpanded}
+        isLegendExpanded={itemsExpanded}
         sidebarMode={isSidebar}
       />
     );
 
     return (
-      <li key={item.name} className={itemClassName(item, isHidden)}>
+      <li key={item.name} className={itemClassName(item)}>
         {enableTooltips ? (
           <TooltipRoot>
             <TooltipTrigger asChild>
               {/* Full width when the row carries a points-table icon so the
                   ml-auto icon pins to a consistent right-edge column. */}
-              <div className={item.onShowPoints ? 'w-full' : 'w-fit'}>{legendItem}</div>
+              <div className={item.onShowPoints ? 'w-full' : 'w-fit max-w-full'}>{legendItem}</div>
             </TooltipTrigger>
             {item.isHighlighted && item.tooltip && (
               <TooltipContent side="bottom" collisionPadding={10}>
@@ -500,26 +456,16 @@ export default function ChartLegend({
     );
   };
 
-  // Bottom controls (switches, FP indicators, expand button, actions)
+  // Bottom controls (switches, FP indicators, actions)
   const hasBottomControls =
-    switchElements ||
-    actionElements ||
-    fpIndicators ||
-    keyIndicators ||
-    expandButton ||
-    hasAtomFootnote;
+    switchElements || actionElements || fpIndicators || keyIndicators || hasAtomFootnote;
   const bottomControls = hasBottomControls ? (
     <div className="shrink-0 grow-0">
       {actionElements}
       {switchElements}
       {fpIndicators}
       {keyIndicators}
-      {expandButton}
-      {hasAtomFootnote && (
-        <p className="mt-2 text-[10px] text-muted-foreground/70 leading-tight no-export">
-          <sup>1</sup> {t.atomFootnote}
-        </p>
-      )}
+      {hasAtomFootnote && <AtomEngineFootnote className="mt-2 no-export" />}
     </div>
   ) : null;
 
@@ -531,56 +477,43 @@ export default function ChartLegend({
         style={isSidebar || isOverflowing ? { scrollbarGutter: 'stable' } : undefined}
         className={cn(scrollClasses, 'custom-scrollbar')}
       >
-        {rows.map((row, i) => {
-          const allHidden =
-            isSidebar && row.every((item: CommonLegendItemProps) => hiddenNames.has(item.name));
-          return (
-            <div
-              key={i}
+        {rows.map((row, i) => (
+          <div key={i} className={cn('p-1 rounded-sm shrink-0', i > 0 && 'mt-2')}>
+            <div className="text-sm font-medium text-muted-foreground gpu-legend-title whitespace-nowrap overflow-ellipsis overflow-hidden">
+              {row[0].title}
+            </div>
+            <ul
               className={cn(
-                'p-1 rounded-sm shrink-0',
-                i > 0 && 'mt-2',
-                allHidden && 'h-0 m-0! p-0! overflow-hidden',
+                'flex flex-wrap gap-x-2 gap-y-1',
+                itemsExpanded && 'md:block md:space-y-1',
               )}
             >
-              <div className="text-sm font-medium text-muted-foreground gpu-legend-title whitespace-nowrap overflow-ellipsis overflow-hidden">
-                {row[0].title}
-              </div>
-              <ul
-                className={cn(
-                  'flex flex-wrap gap-x-2 gap-y-1',
-                  effectiveExpanded && 'md:block md:space-y-1',
-                )}
-              >
-                {row.map((item: CommonLegendItemProps) => {
-                  const isHidden = isSidebar && hiddenNames.has(item.name);
-                  return (
-                    <li key={item.name} className={cn(isHidden && 'h-0 m-0! p-0! overflow-hidden')}>
-                      <ChartLegendItem
-                        name={item.name}
-                        hw={item.hw}
-                        label={item.label}
-                        color={item.color}
-                        lineDasharray={item.lineDasharray}
-                        title={item.title}
-                        isActive={item.isActive}
-                        onClick={item.onClick}
-                        onHover={onItemHover}
-                        onHoverEnd={onItemHoverEnd}
-                        onRemove={removeFor(item)}
-                        onShowPoints={item.onShowPoints}
-                        hideAriaLabel={t.hide(item.label)}
-                        showPointsAriaLabel={t.showPoints(item.label)}
-                        sidebarMode={isSidebar}
-                        asFragment
-                      />
-                    </li>
-                  );
-                })}
-              </ul>
-            </div>
-          );
-        })}
+              {row.map((item: CommonLegendItemProps) => (
+                <li key={item.name}>
+                  <ChartLegendItem
+                    name={item.name}
+                    hw={item.hw}
+                    label={item.label}
+                    color={item.color}
+                    lineDasharray={item.lineDasharray}
+                    title={item.title}
+                    isActive={item.isActive}
+                    onClick={item.onClick}
+                    onHover={onItemHover}
+                    onHoverEnd={onItemHoverEnd}
+                    onRemove={removeFor(item)}
+                    onShowPoints={item.onShowPoints}
+                    hideAriaLabel={t.hide(item.label)}
+                    showPointsAriaLabel={t.showPoints(item.label)}
+                    isLegendExpanded={itemsExpanded}
+                    sidebarMode={isSidebar}
+                    asFragment
+                  />
+                </li>
+              ))}
+            </ul>
+          </div>
+        ))}
       </div>
     ) : (
       <ul
@@ -588,16 +521,14 @@ export default function ChartLegend({
         style={isSidebar || isOverflowing ? { scrollbarGutter: 'stable' } : undefined}
         className={cn(scrollClasses, 'custom-scrollbar')}
       >
-        {(isSidebar ? sortedItems : legendItems).map((item) =>
-          renderItem(item, isSidebar && hiddenNames.has(item.name)),
-        )}
+        {(isSidebar ? sortedItems : legendItems).map((item) => renderItem(item))}
       </ul>
     );
 
   const content = (
     <div className={isSidebar ? 'h-full' : 'relative'}>
       <div data-testid="chart-legend" className={outerClasses} style={outerStyle}>
-        {searchInput}
+        {panelHeader}
         {scrollContent}
         {bottomControls}
       </div>
diff --git a/packages/app/src/components/ui/chart-selectors.tsx b/packages/app/src/components/ui/chart-selectors.tsx
index 2a7815132..88205e2fd 100644
--- a/packages/app/src/components/ui/chart-selectors.tsx
+++ b/packages/app/src/components/ui/chart-selectors.tsx
@@ -1,10 +1,13 @@
 'use client';
 
 import { Info } from 'lucide-react';
+import type { ReactNode } from 'react';
 
 import { LabelWithTooltip } from '@/components/ui/label-with-tooltip';
 import { track } from '@/lib/analytics';
+import { ModelLogo } from '@/components/ui/model-logo';
 import { MultiSelect } from '@/components/ui/multi-select';
+import { NewBadge } from '@/components/ui/new-badge';
 import {
   Select,
   SelectContent,
@@ -36,46 +39,66 @@ const STRINGS = {
   en: {
     model: 'Model',
     modelTooltip: 'The language model being benchmarked.',
+    newModel: 'NEW',
     islOsl: 'ISL / OSL',
     islOslTooltip:
       'Input Sequence Length / Output Sequence Length. Defines the number of input and output tokens for the benchmark (e.g., 1K/8K means 1,024 input tokens and 8,192 output tokens).',
     scenario: 'Scenario',
     scenarioTooltip:
-      'Benchmark scenario. Fixed Sequence Length runs use a defined input/output token count (ISL/OSL). Agentic Traces replay real agentic workloads with variable inputs/outputs.',
+      'Benchmark scenario. Fixed Sequence Length runs use a defined input/output token count (ISL/OSL). The Agentic scenario replays real agentic workloads with variable inputs/outputs.',
+    agenticScenarioTooltip: 'Realistic Long Context Multi Turn Agentic Workload with Sub Agents.',
+    agenticScenarioLearnMore: 'Learn More Here',
     latencyPercentile: 'Latency Percentile',
     latencyPercentileTooltip:
       'Percentile of the latency distribution used for the chart x-axis on agentic runs.',
     precision: 'Precision',
     precisionTooltip:
       "Numerical precision used for model weights. Lower precision like 'FP4' uses less memory and increases throughput but may slightly reduce accuracy compared to higher precisions like 'FP8'.",
-    agenticGroup: 'Agentic',
     fixedSequenceLength: 'Fixed Sequence Length',
+    experimentalSupport: 'Experimental Support (WIP)',
+    maintenanceMode: 'Maintenance Mode',
+    maintenanceReason: 'Updated at a lower priority because these models are irrelevant.',
     deprecated: 'Deprecated',
+    deprecatedModelReason: 'Model is no longer actively benchmarked.',
     deprecatedSequenceReason:
       'CI capacity was reallocated to agentic coding and multi-turn chat scenarios.',
   },
   zh: {
     model: '模型',
     modelTooltip: '正在进行基准测试的语言模型。',
+    newModel: '新',
     islOsl: 'ISL / OSL',
     islOslTooltip:
       '输入序列长度 / 输出序列长度（Input Sequence Length / Output Sequence Length）。定义基准测试的输入和输出 token 数量（如 1K/8K 表示 1,024 个输入 token 和 8,192 个输出 token）。',
     scenario: '场景',
     scenarioTooltip:
-      '基准测试场景。Fixed Sequence Length 使用预设的输入/输出 token 数（ISL/OSL）。Agentic Traces 回放具有可变输入/输出的真实智能体工作负载。',
+      '基准测试场景。Fixed Sequence Length 使用预设的输入/输出 token 数（ISL/OSL）。Agentic 场景回放具有可变输入/输出的真实智能体工作负载。',
+    agenticScenarioTooltip: '真实的长上下文、多轮、带子智能体（sub-agent）的智能体工作负载。',
+    agenticScenarioLearnMore: '点此了解更多',
     latencyPercentile: '延迟分位数',
     latencyPercentileTooltip: '用于智能体运行图表 X 轴的延迟分布分位数。',
     precision: '精度',
     precisionTooltip:
       '模型权重的数值精度。FP4 等低精度占用更少显存并提高吞吐量，但与 FP8 等高精度相比可能略微降低准确度。',
-    agenticGroup: '智能体',
     fixedSequenceLength: '固定序列长度',
+    experimentalSupport: '实验性支持（开发中）',
+    maintenanceMode: '维护模式',
+    maintenanceReason: '这些模型的相关性较低，因此以较低优先级更新。',
     deprecated: '已弃用',
+    deprecatedModelReason: '该模型已不再进行活跃基准测试。',
     deprecatedSequenceReason: 'CI 容量已重新分配给智能体编程和多轮对话场景。',
   },
 } as const;
 
-function CategorySectionTitle({ label, reason }: { label: string; reason: string }) {
+function CategorySectionTitle({
+  id,
+  label,
+  reason,
+}: {
+  id: string;
+  label: string;
+  reason: string;
+}) {
   return (
     <span className="flex items-center gap-1">
       {label}
@@ -83,7 +106,7 @@ function CategorySectionTitle({ label, reason }: { label: string; reason: string
         <TooltipTrigger asChild>
           <Info
             className="size-3 text-muted-foreground cursor-help"
-            data-testid={`selector-category-${label.toLowerCase().replaceAll(' ', '-')}-info`}
+            data-testid={`selector-category-${id}-info`}
           />
         </TooltipTrigger>
         <TooltipContent side="top" collisionPadding={10} className="z-[130]">
@@ -94,6 +117,57 @@ function CategorySectionTitle({ label, reason }: { label: string; reason: string
   );
 }
 
+/**
+ * Info affordance shown next to the scenario selector while an agentic scenario
+ * is selected. The agentic workload isn't self-describing from its name alone,
+ * and it's now the opening scenario for the AgentX models — so the explainer
+ * sits beside the closed trigger rather than inside the dropdown, where the
+ * "learn more" link would be swallowed by the select's outside-click handling.
+ */
+function AgenticScenarioInfo({
+  tooltip,
+  learnMore,
+  href,
+}: {
+  tooltip: string;
+  learnMore: string;
+  href: string;
+}) {
+  return (
+    <TooltipRoot>
+      <TooltipTrigger asChild>
+        <Info
+          className="size-3.5 shrink-0 text-muted-foreground cursor-help"
+          data-testid="scenario-agentic-info"
+        />
+      </TooltipTrigger>
+      <TooltipContent side="top" collisionPadding={10} className="z-[130]">
+        <span>
+          {tooltip}{' '}
+          <a
+            href={href}
+            className="underline underline-offset-2"
+            data-testid="scenario-agentic-info-link"
+            onClick={() => track('selector_scenario_agentx_link')}
+          >
+            {learnMore}
+          </a>
+        </span>
+      </TooltipContent>
+    </TooltipRoot>
+  );
+}
+
+// Full-color creator logo beside each model name, in the dropdown rows and
+// the closed trigger alike. `ModelLogo` renders nothing for models without
+// a configured logo, so rows degrade to plain text instead of breaking.
+const toOption = (model: string, badge?: ReactNode) => ({
+  value: model,
+  label: getModelLabel(model as Model),
+  icon: <ModelLogo model={model as Model} />,
+  badge,
+});
+
 interface ModelSelectorProps {
   id?: string;
   value: string;
@@ -102,6 +176,18 @@ interface ModelSelectorProps {
   onOpenChange?: (open: boolean) => void;
   availableModels: string[];
   'data-testid'?: string;
+  /**
+   * Optional affordance rendered beside the closed trigger (outside the
+   * dropdown, mirroring the scenario selector's info icon) — e.g. the
+   * model-architecture deep-dive link on the inference dashboard.
+   */
+  trailing?: ReactNode;
+  /**
+   * Models whose dropdown rows carry a NEW pill (localized). The inference
+   * dashboard passes the featured AgentX set; surfaces that don't highlight
+   * new models simply omit the prop.
+   */
+  newModels?: ReadonlySet<string>;
 }
 
 export function ModelSelector({
@@ -112,26 +198,29 @@ export function ModelSelector({
   onOpenChange,
   availableModels,
   'data-testid': testId,
+  trailing,
+  newModels,
 }: ModelSelectorProps) {
   const t = STRINGS[useLocale()];
   const groups = groupByCategory(availableModels, (m) => getModelCategory(m as Model));
+  const optionFor = (model: string) =>
+    toOption(
+      model,
+      newModels?.has(model) ? (
+        <NewBadge data-new-badge="model-option">{t.newModel}</NewBadge>
+      ) : undefined,
+    );
   const sections = [
     {
       id: 'default',
-      options: groups.default.map((model) => ({
-        value: model,
-        label: getModelLabel(model as Model),
-      })),
+      options: groups.default.map(optionFor),
     },
     ...(groups.experimental.length > 0
       ? [
           {
             id: 'experimental',
-            header: 'Experimental Support (WIP)',
-            options: groups.experimental.map((model) => ({
-              value: model,
-              label: getModelLabel(model as Model),
-            })),
+            header: t.experimentalSupport,
+            options: groups.experimental.map(optionFor),
           },
         ]
       : []),
@@ -141,14 +230,12 @@ export function ModelSelector({
             id: 'maintenance',
             header: (
               <CategorySectionTitle
-                label="Maintenance Mode"
-                reason="Updated at a lower priority because these models are irrelevant."
+                id="maintenance-mode"
+                label={t.maintenanceMode}
+                reason={t.maintenanceReason}
               />
             ),
-            options: groups.maintenance.map((model) => ({
-              value: model,
-              label: getModelLabel(model as Model),
-            })),
+            options: groups.maintenance.map(optionFor),
           },
         ]
       : []),
@@ -158,14 +245,12 @@ export function ModelSelector({
             id: 'deprecated',
             header: (
               <CategorySectionTitle
-                label="Deprecated"
-                reason="Model is no longer actively benchmarked."
+                id="deprecated"
+                label={t.deprecated}
+                reason={t.deprecatedModelReason}
               />
             ),
-            options: groups.deprecated.map((model) => ({
-              value: model,
-              label: getModelLabel(model as Model),
-            })),
+            options: groups.deprecated.map(optionFor),
           },
         ]
       : []),
@@ -174,28 +259,31 @@ export function ModelSelector({
   return (
     <div className="flex flex-col space-y-1.5 lg:col-span-2">
       <LabelWithTooltip htmlFor={id} label={t.model} tooltip={t.modelTooltip} />
-      <div>
-        <MultiSelect
-          sections={sections}
-          value={[value]}
-          onChange={(values) => {
-            const next = values[0];
-            if (!next) return;
-            track('selector_model_changed', { model: next });
-            onChange(next as Model);
-          }}
-          open={open}
-          onOpenChange={onOpenChange}
-          triggerId={id}
-          triggerTestId={testId}
-          placeholder="Model"
-          minSelections={1}
-          maxSelections={1}
-          showClearAll={false}
-          searchable={false}
-          plainSelectedText
-          showSelectionSummary={false}
-        />
+      <div className="flex items-center gap-1.5">
+        <div className="min-w-0 flex-1">
+          <MultiSelect
+            sections={sections}
+            value={[value]}
+            onChange={(values) => {
+              const next = values[0];
+              if (!next) return;
+              track('selector_model_changed', { model: next });
+              onChange(next as Model);
+            }}
+            open={open}
+            onOpenChange={onOpenChange}
+            triggerId={id}
+            triggerTestId={testId}
+            placeholder={t.model}
+            minSelections={1}
+            maxSelections={1}
+            showClearAll={false}
+            searchable={false}
+            plainSelectedText
+            showSelectionSummary={false}
+          />
+        </div>
+        {trailing}
       </div>
     </div>
   );
@@ -237,8 +325,9 @@ export function SequenceSelector({
             id: 'deprecated',
             header: (
               <CategorySectionTitle
-                label="Deprecated"
-                reason="CI capacity was reallocated to agentic coding and multi-turn chat scenarios."
+                id="deprecated"
+                label={t.deprecated}
+                reason={t.deprecatedSequenceReason}
               />
             ),
             options: groups.deprecated.map((seq) => ({
@@ -267,7 +356,7 @@ export function SequenceSelector({
           onOpenChange={onOpenChange}
           triggerId={id}
           triggerTestId={testId}
-          placeholder="ISL / OSL"
+          placeholder={t.islOsl}
           minSelections={1}
           maxSelections={1}
           showClearAll={false}
@@ -294,6 +383,11 @@ interface ScenarioSelectorProps {
  * Scenario selector — fixed-seq-len rows grouped under "Fixed Sequence Length",
  * agentic-trace rows rendered flat below. Label is "Scenario" (the ISL/OSL
  * framing only applies to the fixed-seq subset).
+ *
+ * Renders nothing when fewer than two scenarios are available: a dropdown
+ * with a single choice is dead UI (e.g. agentic-only models like Kimi K3),
+ * and a static "Scenario: Agentic" readout is just as redundant when there is
+ * nothing to choose — so the whole control disappears.
  */
 export function ScenarioSelector({
   id = 'scenario-select',
@@ -309,62 +403,78 @@ export function ScenarioSelector({
   const fixedSeq = availableSequences.filter((s) => sequenceKind(s as Sequence) === 'fixed-seq');
   const agentic = availableSequences.filter((s) => sequenceKind(s as Sequence) === 'agentic');
   const fixedGroups = groupByCategory(fixedSeq, (s) => getSequenceCategory(s as Sequence));
+  const isAgenticSelected = sequenceKind(value as Sequence) === 'agentic';
+
+  if (availableSequences.length < 2) return null;
 
   return (
     <div className="flex flex-col space-y-1.5 lg:col-span-1">
       <LabelWithTooltip htmlFor={id} label={t.scenario} tooltip={t.scenarioTooltip} />
-      <Select
-        value={value}
-        onValueChange={(v) => {
-          track('selector_scenario_changed', { scenario: v });
-          onChange(v as Sequence);
-        }}
-        open={open}
-        onOpenChange={onOpenChange}
-      >
-        <SelectTrigger id={id} data-testid={testId} className="w-full">
-          <SelectValue />
-        </SelectTrigger>
-        <SelectContent>
-          {/* Agentic group listed first when available (display order only —
-              the app default scenario is 8K/1K). */}
-          {agentic.length > 0 && (
-            <SelectGroup>
-              <SelectLabel>{t.agenticGroup}</SelectLabel>
-              {agentic.map((seq) => (
-                <SelectItem key={seq} value={seq}>
-                  {getSequenceLabel(seq as Sequence, locale)}
-                </SelectItem>
-              ))}
-            </SelectGroup>
-          )}
-          {fixedSeq.length > 0 && (
-            <SelectGroup>
-              <SelectLabel>{t.fixedSequenceLength}</SelectLabel>
-              {fixedGroups.default.map((seq) => (
-                <SelectItem key={seq} value={seq}>
-                  {getSequenceLabel(seq as Sequence, locale)}
-                </SelectItem>
-              ))}
-              {fixedGroups.deprecated.length > 0 && (
-                <>
-                  <SelectLabel>
-                    <CategorySectionTitle
-                      label={t.deprecated}
-                      reason={t.deprecatedSequenceReason}
-                    />
-                  </SelectLabel>
-                  {fixedGroups.deprecated.map((seq) => (
-                    <SelectItem key={seq} value={seq}>
-                      {getSequenceLabel(seq as Sequence, locale)}
-                    </SelectItem>
-                  ))}
-                </>
-              )}
-            </SelectGroup>
-          )}
-        </SelectContent>
-      </Select>
+      <div className="flex items-center gap-1.5">
+        <Select
+          value={value}
+          onValueChange={(v) => {
+            track('selector_scenario_changed', { scenario: v });
+            onChange(v as Sequence);
+          }}
+          open={open}
+          onOpenChange={onOpenChange}
+        >
+          <SelectTrigger id={id} data-testid={testId} className="w-full min-w-0">
+            <SelectValue />
+          </SelectTrigger>
+          <SelectContent>
+            {/* Agentic entries listed first when available (display order only
+                — availability decides which scenario opens by default). They
+                carry no group header: they are named "Agentic" themselves, so a
+                heading above them would just repeat the word. They stay in their
+                own SelectGroup so the "Fixed Sequence Length" heading below
+                still reads as a separate section. */}
+            {agentic.length > 0 && (
+              <SelectGroup>
+                {agentic.map((seq) => (
+                  <SelectItem key={seq} value={seq}>
+                    {getSequenceLabel(seq as Sequence, locale)}
+                  </SelectItem>
+                ))}
+              </SelectGroup>
+            )}
+            {fixedSeq.length > 0 && (
+              <SelectGroup>
+                <SelectLabel>{t.fixedSequenceLength}</SelectLabel>
+                {fixedGroups.default.map((seq) => (
+                  <SelectItem key={seq} value={seq}>
+                    {getSequenceLabel(seq as Sequence, locale)}
+                  </SelectItem>
+                ))}
+                {fixedGroups.deprecated.length > 0 && (
+                  <>
+                    <SelectLabel>
+                      <CategorySectionTitle
+                        id="deprecated"
+                        label={t.deprecated}
+                        reason={t.deprecatedSequenceReason}
+                      />
+                    </SelectLabel>
+                    {fixedGroups.deprecated.map((seq) => (
+                      <SelectItem key={seq} value={seq}>
+                        {getSequenceLabel(seq as Sequence, locale)}
+                      </SelectItem>
+                    ))}
+                  </>
+                )}
+              </SelectGroup>
+            )}
+          </SelectContent>
+        </Select>
+        {isAgenticSelected && (
+          <AgenticScenarioInfo
+            tooltip={t.agenticScenarioTooltip}
+            learnMore={t.agenticScenarioLearnMore}
+            href={locale === 'zh' ? '/zh/agentx' : '/agentx'}
+          />
+        )}
+      </div>
     </div>
   );
 }
@@ -427,6 +537,11 @@ interface PrecisionSelectorProps {
   'data-testid'?: string;
 }
 
+/**
+ * Precision multi-select. Renders nothing when fewer than two precisions are
+ * available — a single-precision model (e.g. Kimi K3) has nothing to toggle,
+ * so the control disappears instead of offering a no-op menu.
+ */
 export function PrecisionSelector({
   id = 'precision-select',
   value,
@@ -437,6 +552,9 @@ export function PrecisionSelector({
   'data-testid': testId,
 }: PrecisionSelectorProps) {
   const t = STRINGS[useLocale()];
+
+  if (availablePrecisions.length < 2) return null;
+
   return (
     <div className="flex flex-col space-y-1.5 lg:col-span-1">
       <LabelWithTooltip htmlFor={id} label={t.precision} tooltip={t.precisionTooltip} />
diff --git a/packages/app/src/components/ui/coach-mark.test.tsx b/packages/app/src/components/ui/coach-mark.test.tsx
new file mode 100644
index 000000000..fd745e232
--- /dev/null
+++ b/packages/app/src/components/ui/coach-mark.test.tsx
@@ -0,0 +1,251 @@
+// @vitest-environment jsdom
+import React, { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi, type Mock } from 'vitest';
+
+vi.mock('next/navigation', () => ({
+  usePathname: () => '/inference',
+}));
+
+import { CoachMark } from '@/components/ui/coach-mark';
+import type { NudgeAnchor } from '@/lib/nudges/types';
+
+const ACTION_SELECTOR = '.dot-group';
+
+function stubRect(element: Element, left: number, top: number, size = 12): void {
+  element.getBoundingClientRect = () =>
+    ({
+      left,
+      top,
+      width: size,
+      height: size,
+      right: left + size,
+      bottom: top + size,
+      x: left,
+      y: top,
+      toJSON: () => ({}),
+    }) as DOMRect;
+}
+
+/** A point the resolver can find, mid-viewport so it passes the on-screen check. */
+function addPoint(parent: ParentNode = document.body): HTMLElement {
+  const point = document.createElement('div');
+  point.className = 'dot-group';
+  parent.append(point);
+  stubRect(point, 500, 300);
+  return point;
+}
+
+let container: HTMLDivElement;
+let root: Root;
+let onDismiss: Mock<() => void>;
+let onAction: Mock<() => void>;
+
+function renderCoachMark(anchor: NudgeAnchor) {
+  act(() =>
+    root.render(
+      <CoachMark
+        anchor={anchor}
+        icon={<span />}
+        title="Every point has a story"
+        description="Click any point to view server metrics and logs."
+        onDismiss={onDismiss}
+        onAction={onAction}
+        testId="coach-mark"
+      />,
+    ),
+  );
+}
+
+function pressEscape() {
+  act(() => {
+    window.dispatchEvent(new KeyboardEvent('keydown', { key: 'Escape' }));
+  });
+}
+
+const isVisible = () => container.querySelector('[data-testid="coach-mark-target"]') !== null;
+
+/**
+ * Let the component notice a DOM change on its own: the MutationObserver
+ * callback is a microtask, and the reposition it schedules runs on the next
+ * animation frame.
+ */
+async function flushReposition() {
+  await act(async () => {
+    await Promise.resolve();
+    await new Promise((resolve) => {
+      requestAnimationFrame(() => resolve(null));
+    });
+  });
+}
+
+beforeEach(() => {
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+  onDismiss = vi.fn<() => void>();
+  onAction = vi.fn<() => void>();
+});
+
+afterEach(() => {
+  act(() => root.unmount());
+  container.remove();
+  document.body.innerHTML = '';
+  vi.restoreAllMocks();
+});
+
+describe('CoachMark input handling', () => {
+  describe('when an anchor is on screen', () => {
+    let point: HTMLElement;
+
+    beforeEach(() => {
+      point = addPoint();
+      renderCoachMark({ resolve: () => point, actionSelector: ACTION_SELECTOR });
+    });
+
+    it('renders the callout', () => {
+      expect(isVisible()).toBe(true);
+      expect(container.querySelector('[data-testid="coach-mark"]')?.textContent).toContain(
+        'Every point has a story',
+      );
+    });
+
+    it('dismisses on Escape', () => {
+      pressEscape();
+      expect(onDismiss).toHaveBeenCalledTimes(1);
+    });
+
+    it('treats a click on the anchored element as taking the action', () => {
+      act(() => {
+        point.dispatchEvent(new MouseEvent('click', { bubbles: true }));
+      });
+      expect(onAction).toHaveBeenCalledTimes(1);
+    });
+  });
+
+  describe('when no anchor resolves (scrolled away, or every point filtered out)', () => {
+    beforeEach(() => {
+      // The card stays mounted so it can be measured and can re-appear, but
+      // nothing is drawn — the user has never seen this tip.
+      renderCoachMark({ resolve: () => null, actionSelector: ACTION_SELECTOR });
+    });
+
+    it('renders nothing visible', () => {
+      expect(isVisible()).toBe(false);
+      expect(
+        container.querySelector<HTMLElement>('[data-testid="coach-mark"]')?.style.visibility,
+      ).toBe('hidden');
+    });
+
+    it('ignores Escape rather than burning the permanent dismissal', () => {
+      // Escape is a global key: the user is closing some other UI entirely.
+      // Acting on it here would permanently dismiss a tip never shown, and it
+      // would not come back on scroll.
+      pressEscape();
+      expect(onDismiss).not.toHaveBeenCalled();
+    });
+
+    it('ignores a click on a point rather than recording phantom engagement', () => {
+      const point = addPoint();
+      act(() => {
+        point.dispatchEvent(new MouseEvent('click', { bubbles: true }));
+      });
+      expect(onAction).not.toHaveBeenCalled();
+    });
+  });
+
+  it('starts responding to Escape once its point comes into view', async () => {
+    let point: HTMLElement | null = null;
+    renderCoachMark({ resolve: () => point, actionSelector: ACTION_SELECTOR });
+    expect(isVisible()).toBe(false);
+
+    pressEscape();
+    expect(onDismiss).not.toHaveBeenCalled();
+
+    // The point appears — as it does when the chart is scrolled into view.
+    // No re-render: the component's own MutationObserver has to notice.
+    point = addPoint();
+    await flushReposition();
+    expect(isVisible()).toBe(true);
+
+    pressEscape();
+    expect(onDismiss).toHaveBeenCalledTimes(1);
+  });
+});
+
+describe('CoachMark anchor reuse', () => {
+  it('reuses one resolved anchor across scoped transform mutations', async () => {
+    const chart = document.createElement('div');
+    document.body.append(chart);
+    const point = addPoint(chart);
+    const resolve = vi.fn(() => point);
+    const getRect = vi.fn((element: Element) => element.getBoundingClientRect());
+
+    renderCoachMark({
+      resolve,
+      getRect,
+      getMutationRoot: () => chart,
+      actionSelector: ACTION_SELECTOR,
+    });
+    expect(resolve).toHaveBeenCalledTimes(1);
+    expect(getRect).toHaveBeenCalledTimes(1);
+
+    const outside = document.createElement('div');
+    document.body.append(outside);
+    outside.setAttribute('transform', 'translate(1,1)');
+    await flushReposition();
+    expect(getRect, 'outside mutation ignored').toHaveBeenCalledTimes(1);
+
+    for (let index = 0; index < 20; index++) {
+      point.setAttribute('transform', `translate(${index},${index})`);
+    }
+    await flushReposition();
+
+    expect(resolve, 'full resolver stays cached').toHaveBeenCalledTimes(1);
+    expect(getRect, 'one cached measurement for the frame').toHaveBeenCalledTimes(2);
+  });
+
+  it('resolves a replacement when the cached anchor becomes unusable', async () => {
+    const chart = document.createElement('div');
+    document.body.append(chart);
+    const first = addPoint(chart);
+    const second = addPoint(chart);
+    stubRect(second, 700, 400);
+    let current = first;
+    let firstUsable = true;
+    const resolve = vi.fn(() => current);
+    const getRect = vi.fn((element: Element) =>
+      element === first && !firstUsable ? null : element.getBoundingClientRect(),
+    );
+
+    renderCoachMark({
+      resolve,
+      getRect,
+      getMutationRoot: () => chart,
+      actionSelector: ACTION_SELECTOR,
+    });
+    expect(resolve).toHaveBeenCalledTimes(1);
+
+    current = second;
+    firstUsable = false;
+    first.setAttribute('transform', 'translate(1,1)');
+    await flushReposition();
+
+    expect(resolve).toHaveBeenCalledTimes(2);
+    expect(container.querySelector('[data-testid="coach-mark-target"]')?.getAttribute('cx')).toBe(
+      '706',
+    );
+  });
+
+  it('stays hidden when a resolved anchor fails validation', () => {
+    const point = addPoint();
+
+    renderCoachMark({
+      resolve: () => point,
+      getRect: () => null,
+      actionSelector: ACTION_SELECTOR,
+    });
+
+    expect(isVisible()).toBe(false);
+  });
+});
diff --git a/packages/app/src/components/ui/coach-mark.tsx b/packages/app/src/components/ui/coach-mark.tsx
new file mode 100644
index 000000000..be70a5094
--- /dev/null
+++ b/packages/app/src/components/ui/coach-mark.tsx
@@ -0,0 +1,306 @@
+'use client';
+
+import { X } from 'lucide-react';
+import { useCallback, useEffect, useLayoutEffect, useRef, useState } from 'react';
+
+import {
+  computeCoachMarkPlacement,
+  isAnchorOnScreen,
+  viewportSize,
+  type CoachMarkPlacement,
+  type Size,
+} from '@/lib/nudges/anchor';
+import type { NudgeAnchor } from '@/lib/nudges/types';
+import { useLocale } from '@/lib/use-locale';
+
+const STRINGS = {
+  en: { close: 'Dismiss tip' },
+  zh: { close: '关闭提示' },
+} as const;
+
+/** Card width in px. Mirrors the `w-72` utility so placement can use it before paint. */
+const CARD_WIDTH = 288;
+/** Stop the pointer short of the anchor so the arrowhead doesn't cover the point. */
+const ARROW_STANDOFF = 13;
+
+interface CoachMarkProps {
+  anchor: NudgeAnchor;
+  icon: React.ReactNode;
+  title: string;
+  description: string;
+  /** Fired by the X button and by Escape. */
+  onDismiss: () => void;
+  /** Fired when the user clicks something matching `anchor.actionSelector`. */
+  onAction?: () => void;
+  testId?: string;
+}
+
+function samePlacement(a: CoachMarkPlacement | null, b: CoachMarkPlacement | null): boolean {
+  if (a === null || b === null) return a === b;
+  return (
+    Math.abs(a.left - b.left) < 0.5 &&
+    Math.abs(a.top - b.top) < 0.5 &&
+    Math.abs(a.arrowX1 - b.arrowX1) < 0.5 &&
+    Math.abs(a.arrowY1 - b.arrowY1) < 0.5 &&
+    Math.abs(a.arrowX2 - b.arrowX2) < 0.5 &&
+    Math.abs(a.arrowY2 - b.arrowY2) < 0.5 &&
+    a.side === b.side
+  );
+}
+
+/**
+ * A non-modal callout pinned to a live element: a small card plus a pointer
+ * drawn to the anchor's centre.
+ *
+ * Deliberately *not* a focus trap and not backed by a scrim — the whole point
+ * is to teach an interaction the user then performs on the element underneath,
+ * so everything except the card itself stays `pointer-events: none`. The card
+ * hides (without dismissing) whenever the anchor is missing or scrolled out of
+ * view, and re-appears when it comes back.
+ */
+export function CoachMark({
+  anchor,
+  icon,
+  title,
+  description,
+  onDismiss,
+  onAction,
+  testId,
+}: CoachMarkProps) {
+  const locale = useLocale();
+  const strings = STRINGS[locale];
+  const cardRef = useRef<HTMLDivElement>(null);
+  const anchorRef = useRef<Element | null>(null);
+  const [placement, setPlacement] = useState<CoachMarkPlacement | null>(null);
+
+  const resolve = anchor.resolve;
+  const getRect = anchor.getRect;
+  const getMutationRoot = anchor.getMutationRoot;
+  const repositionEvents = anchor.repositionEvents;
+  const actionSelector = anchor.actionSelector;
+
+  // ── Placement, kept in sync with the live chart ──────────────────────────
+  useLayoutEffect(() => {
+    let frame = 0;
+
+    const measure = () => {
+      frame = 0;
+      const card = cardRef.current;
+      if (!card) return;
+
+      let element = getRect && anchorRef.current?.isConnected ? anchorRef.current : null;
+      let rect = element && getRect ? getRect(element) : null;
+      if (!element || !rect) {
+        element = resolve();
+        rect = element ? (getRect ? getRect(element) : element.getBoundingClientRect()) : null;
+      }
+      anchorRef.current = element && rect ? element : null;
+
+      if (!element || !rect) {
+        setPlacement((prev) => (prev === null ? prev : null));
+        return;
+      }
+      const viewport = viewportSize();
+      if (!isAnchorOnScreen(rect, viewport)) {
+        setPlacement((prev) => (prev === null ? prev : null));
+        return;
+      }
+      const size: Size = {
+        width: card.offsetWidth || CARD_WIDTH,
+        height: card.offsetHeight,
+      };
+      const next = computeCoachMarkPlacement(rect, size, viewport);
+      setPlacement((prev) => (samePlacement(prev, next) ? prev : next));
+    };
+
+    const schedule = () => {
+      if (frame !== 0) return;
+      frame = requestAnimationFrame(measure);
+    };
+
+    measure();
+
+    window.addEventListener('resize', schedule);
+    // Capture phase so scrolling inside any container (not just the document)
+    // repositions the pointer.
+    window.addEventListener('scroll', schedule, { capture: true, passive: true });
+    for (const event of repositionEvents ?? []) window.addEventListener(event, schedule);
+
+    // D3 coordinate transitions rewrite `transform` on every point each
+    // frame. Keep the selected anchor for those mutations and validate just
+    // that point. Structural, visibility, and trace-availability changes can
+    // change which point is eligible, so they invalidate the cached choice.
+    const observer = new MutationObserver((records) => {
+      if (
+        records.some(
+          (record) => record.type === 'childList' || record.attributeName !== 'transform',
+        )
+      ) {
+        anchorRef.current = null;
+      }
+      schedule();
+    });
+    observer.observe(getMutationRoot?.(anchorRef.current) ?? document.body, {
+      attributes: true,
+      attributeFilter: [
+        'transform',
+        'style',
+        'data-has-trace',
+        'data-trace-availability',
+        'data-benchmark-type',
+      ],
+      childList: true,
+      subtree: true,
+    });
+
+    return () => {
+      if (frame !== 0) cancelAnimationFrame(frame);
+      observer.disconnect();
+      window.removeEventListener('resize', schedule);
+      window.removeEventListener('scroll', schedule, { capture: true });
+      for (const event of repositionEvents ?? []) window.removeEventListener(event, schedule);
+    };
+  }, [resolve, getRect, getMutationRoot, repositionEvents]);
+
+  /**
+   * Whether the callout is actually on screen. The card stays mounted while
+   * hidden — it has to be in the DOM to be measured, and it re-appears when
+   * its point comes back — so every input handler below is gated on this.
+   * Dismissal is permanent: reacting to a keypress or a click while nothing
+   * is visible would burn the first-visit tip the user never saw.
+   */
+  const visible = placement !== null;
+
+  // ── Escape to dismiss ────────────────────────────────────────────────────
+  const onDismissRef = useRef(onDismiss);
+  onDismissRef.current = onDismiss;
+  useEffect(() => {
+    if (!visible) return;
+    const handler = (event: KeyboardEvent) => {
+      if (event.key === 'Escape') onDismissRef.current();
+    };
+    window.addEventListener('keydown', handler);
+    return () => window.removeEventListener('keydown', handler);
+  }, [visible]);
+
+  // ── Clicking the thing the tip is about counts as taking the action ──────
+  const onActionRef = useRef(onAction);
+  onActionRef.current = onAction;
+  useEffect(() => {
+    if (!actionSelector || !visible) return;
+    const handler = (event: MouseEvent) => {
+      const target = event.target as Element | null;
+      if (target?.closest?.(actionSelector)) onActionRef.current?.();
+    };
+    // Capture phase: d3's own point handler calls stopPropagation().
+    document.addEventListener('click', handler, true);
+    return () => document.removeEventListener('click', handler, true);
+  }, [actionSelector, visible]);
+
+  const handleDismiss = useCallback(() => onDismissRef.current(), []);
+
+  const titleId = testId ? `${testId}-title` : undefined;
+  const descriptionId = testId ? `${testId}-description` : undefined;
+
+  return (
+    <div
+      className="pointer-events-none fixed inset-0 z-50"
+      // The card is measured before it is ever placed, so it must be in the
+      // DOM from the first paint — `aria-hidden` keeps that measuring pass out
+      // of the accessibility tree.
+      aria-hidden={visible ? undefined : 'true'}
+      data-testid={testId ? `${testId}-layer` : undefined}
+    >
+      {visible && (
+        <svg
+          className="pointer-events-none absolute inset-0 size-full overflow-visible"
+          aria-hidden="true"
+        >
+          <defs>
+            <marker
+              id="coach-mark-arrowhead"
+              markerWidth="7"
+              markerHeight="7"
+              refX="5.5"
+              refY="3"
+              orient="auto"
+            >
+              <path d="M0,0 L6,3 L0,6 Z" fill="var(--brand)" />
+            </marker>
+          </defs>
+          {(() => {
+            const dx = placement.arrowX2 - placement.arrowX1;
+            const dy = placement.arrowY2 - placement.arrowY1;
+            const length = Math.hypot(dx, dy) || 1;
+            const scale = Math.max(0, length - ARROW_STANDOFF) / length;
+            return (
+              <line
+                data-testid={testId ? `${testId}-pointer` : undefined}
+                x1={placement.arrowX1}
+                y1={placement.arrowY1}
+                x2={placement.arrowX1 + dx * scale}
+                y2={placement.arrowY1 + dy * scale}
+                stroke="var(--brand)"
+                strokeWidth={2}
+                strokeLinecap="round"
+                markerEnd="url(#coach-mark-arrowhead)"
+              />
+            );
+          })()}
+          {/* Pulsing ring on the anchor itself. Unlike the pointer line —
+              which stops short so its arrowhead doesn't cover the point — this
+              is centred exactly on what the tip is about. */}
+          <circle
+            data-testid={testId ? `${testId}-target` : undefined}
+            cx={placement.arrowX2}
+            cy={placement.arrowY2}
+            r={11}
+            fill="none"
+            stroke="var(--brand)"
+            strokeWidth={2}
+            className="animate-ping"
+            style={{ transformBox: 'fill-box', transformOrigin: 'center' }}
+          />
+        </svg>
+      )}
+
+      <div
+        ref={cardRef}
+        data-testid={testId}
+        role="dialog"
+        aria-modal="false"
+        aria-labelledby={titleId}
+        aria-describedby={descriptionId}
+        className={`pointer-events-auto absolute w-72 rounded-lg border border-brand/50 bg-card p-4 shadow-lg transition-opacity duration-200 ${
+          visible ? 'opacity-100' : 'opacity-0'
+        }`}
+        style={{
+          left: placement?.left ?? 0,
+          top: placement?.top ?? 0,
+          visibility: visible ? undefined : 'hidden',
+        }}
+      >
+        <button
+          type="button"
+          onClick={handleDismiss}
+          className="absolute right-2 top-2 rounded-sm text-muted-foreground transition-colors hover:text-foreground focus:outline-none focus:ring-2 focus:ring-ring"
+          aria-label={strings.close}
+          data-testid={testId ? `${testId}-dismiss` : undefined}
+        >
+          <X className="size-3.5" />
+        </button>
+        <div className="flex items-start gap-3 pr-4">
+          <span className="mt-0.5 shrink-0 [&_svg]:size-4">{icon}</span>
+          <div className="flex flex-col gap-1">
+            <p id={titleId} className="text-sm font-semibold text-foreground">
+              {title}
+            </p>
+            <p id={descriptionId} className="text-xs text-muted-foreground">
+              {description}
+            </p>
+          </div>
+        </div>
+      </div>
+    </div>
+  );
+}
diff --git a/packages/app/src/components/ui/collapsible-section.tsx b/packages/app/src/components/ui/collapsible-section.tsx
new file mode 100644
index 000000000..06fab8fa6
--- /dev/null
+++ b/packages/app/src/components/ui/collapsible-section.tsx
@@ -0,0 +1,75 @@
+'use client';
+
+import { ChevronDownIcon } from 'lucide-react';
+import React, { useId, useState } from 'react';
+
+import { cn } from '@/lib/utils';
+
+export interface CollapsibleSectionProps {
+  /** Section heading. */
+  title: string;
+  children: React.ReactNode;
+  /** Whether the section starts expanded. Uncontrolled from then on. */
+  defaultOpen?: boolean;
+  /**
+   * Set false when the body already carries its own heading — a chart caption,
+   * say. The header then shows the title only while the section is folded, so an
+   * expanded section never shows the same heading twice.
+   */
+  titleWhenOpen?: boolean;
+  /** Called with the new state on every toggle. Callers use it for analytics. */
+  onToggle?: (open: boolean) => void;
+  /** Accessible name for the toggle. Localised by the caller. */
+  toggleLabel: string;
+  testId?: string;
+  className?: string;
+}
+
+/**
+ * A section that folds away behind a chevron.
+ *
+ * The body is unmounted rather than hidden: these sections hold D3 charts and
+ * tables, and leaving them mounted under `display: none` keeps them measuring and
+ * re-rendering off-screen for a reader who has said they are not interested.
+ * Chart zoom and pinned tooltips do not survive a fold, which is the trade.
+ */
+export function CollapsibleSection({
+  title,
+  children,
+  defaultOpen = true,
+  titleWhenOpen = true,
+  onToggle,
+  toggleLabel,
+  testId,
+  className,
+}: CollapsibleSectionProps) {
+  const [open, setOpen] = useState(defaultOpen);
+  const bodyId = useId();
+
+  return (
+    <div className={className}>
+      <div className="flex items-center gap-2 mb-2">
+        <button
+          type="button"
+          aria-expanded={open}
+          aria-controls={bodyId}
+          aria-label={toggleLabel}
+          title={toggleLabel}
+          data-testid={testId}
+          onClick={() => {
+            const next = !open;
+            setOpen(next);
+            onToggle?.(next);
+          }}
+          className="text-muted-foreground hover:text-foreground focus-visible:ring-ring rounded-sm outline-none focus-visible:ring-[3px] cursor-pointer"
+        >
+          <ChevronDownIcon
+            className={cn('size-5 transition-transform duration-200', open && 'rotate-180')}
+          />
+        </button>
+        {(open ? titleWhenOpen : true) && <h2 className="text-lg font-semibold">{title}</h2>}
+      </div>
+      {open && <div id={bodyId}>{children}</div>}
+    </div>
+  );
+}
diff --git a/packages/app/src/components/ui/d3-chart-wrapper.tsx b/packages/app/src/components/ui/d3-chart-wrapper.tsx
index 44013b1bf..04f5db6da 100644
--- a/packages/app/src/components/ui/d3-chart-wrapper.tsx
+++ b/packages/app/src/components/ui/d3-chart-wrapper.tsx
@@ -10,9 +10,11 @@ import { createPortal } from 'react-dom';
  * coordinates by the d3 layer.
  */
 function PortalTooltip({
+  chartId,
   tooltipRef,
   pinned,
 }: {
+  chartId: string;
   tooltipRef: React.RefObject<HTMLDivElement | null>;
   pinned: boolean;
 }) {
@@ -21,7 +23,7 @@ function PortalTooltip({
   const node = (
     <div
       ref={tooltipRef}
-      data-chart-tooltip
+      data-chart-tooltip={chartId}
       style={{
         position: 'fixed',
         left: 0,
@@ -81,8 +83,13 @@ export function D3ChartWrapper({
       <div className="flex flex-col lg:flex-row w-full">
         <div ref={setContainerRef} className="relative flex-1 min-w-0">
           <div className="relative">
+            {/* Stable hook for tests. `[data-testid="scatter-graph"] svg` also
+                matches every Lucide icon inside the card — dozens of them —
+                so picking "the first svg" silently grabs an icon whenever the
+                selected metric renders one above the chart. */}
             <svg
               ref={svgRef}
+              data-testid="d3-chart-svg"
               width="100%"
               height={dimensions.height}
               style={{ cursor: grabCursor ? 'grab' : undefined }}
@@ -111,7 +118,11 @@ export function D3ChartWrapper({
                 rise above sibling chart cards' stacking contexts. The d3 layer
                 writes viewport-coords into style.left/top — see
                 computeTooltipPosition. */}
-            <PortalTooltip tooltipRef={tooltipRef} pinned={Boolean(pinnedPoint)} />
+            <PortalTooltip
+              chartId={chartId}
+              tooltipRef={tooltipRef}
+              pinned={Boolean(pinnedPoint)}
+            />
             {noDataOverlay}
           </div>
           <p className="no-export text-xs text-muted-foreground text-center mt-2">{instructions}</p>
@@ -120,7 +131,13 @@ export function D3ChartWrapper({
           </div>
         </div>
         {legendElement && (
-          <div className="w-full h-96 lg:h-[575px] lg:w-48 lg:shrink-0 relative mt-3 lg:mt-0">
+          /* Sizes to the legend content: when the sidebar legend panel is open
+             (.sidebar-legend present) the column grows to fit the widest
+             legend label (capped) so full names display without truncation,
+             while still sitting next to the plot without overlapping it; when
+             closed the legend renders only a small reopen button and the
+             chart reclaims the width. */
+          <div className="w-full lg:w-auto lg:shrink-0 relative mt-3 lg:mt-0 has-[.sidebar-legend]:h-96 lg:has-[.sidebar-legend]:h-[575px] lg:has-[.sidebar-legend]:w-fit lg:has-[.sidebar-legend]:min-w-48 lg:has-[.sidebar-legend]:max-w-96">
             {legendElement}
           </div>
         )}
diff --git a/packages/app/src/components/ui/data-table.tsx b/packages/app/src/components/ui/data-table.tsx
index 88ffa33d0..0f60c5429 100644
--- a/packages/app/src/components/ui/data-table.tsx
+++ b/packages/app/src/components/ui/data-table.tsx
@@ -53,6 +53,12 @@ interface DataTableProps<T> {
   analyticsPrefix?: string;
   /** Show watermark (default: true). */
   watermark?: boolean;
+  /**
+   * Show the search box (default: true). Turn it off for a table whose rows are
+   * few and already named elsewhere — one line per chip, say — where a search
+   * field is a control the reader has no use for.
+   */
+  searchable?: boolean;
 }
 
 const PAGE_SIZE_OPTIONS = [10, 25, 50, 100, 250, 500] as const;
@@ -106,6 +112,7 @@ export function DataTable<T>({
   testId = 'data-table',
   analyticsPrefix = 'table',
   watermark = true,
+  searchable = true,
 }: DataTableProps<T>) {
   const locale = useLocale();
   const t = STRINGS[locale];
@@ -138,7 +145,7 @@ export function DataTable<T>({
 
   // Search: match against all columns with sortValue
   const filtered = useMemo(() => {
-    if (!search.trim()) return data;
+    if (!searchable || !search.trim()) return data;
     const q = search.trim().toLowerCase();
     return data.filter((row) =>
       columns.some((col) => {
@@ -146,7 +153,7 @@ export function DataTable<T>({
         return String(col.sortValue(row)).toLowerCase().includes(q);
       }),
     );
-  }, [data, search, columns]);
+  }, [data, search, columns, searchable]);
 
   const sorted = useMemo(() => {
     if (sort.dir === null || sort.columnIndex < 0) return filtered;
@@ -176,35 +183,37 @@ export function DataTable<T>({
   return (
     <div data-testid={testId} className="mt-3 min-w-0 w-full max-w-full">
       {/* Search */}
-      <div className="mb-3 max-w-xs relative">
-        <Search className="absolute left-2.5 top-1/2 -translate-y-1/2 size-3.5 text-muted-foreground" />
-        <input
-          ref={searchRef}
-          type="text"
-          value={search}
-          onChange={(e) => {
-            setSearch(e.target.value);
-            setPage(0);
-          }}
-          placeholder={t.search}
-          className="w-full h-7 pl-8 pr-7 text-xs bg-transparent border border-border rounded-md focus:outline-none focus:ring-1 focus:ring-ring"
-          aria-label={t.searchAria}
-        />
-        {search && (
-          <button
-            type="button"
-            onClick={() => {
-              setSearch('');
+      {searchable && (
+        <div className="mb-3 max-w-xs relative">
+          <Search className="absolute left-2.5 top-1/2 -translate-y-1/2 size-3.5 text-muted-foreground" />
+          <input
+            ref={searchRef}
+            type="text"
+            value={search}
+            onChange={(e) => {
+              setSearch(e.target.value);
               setPage(0);
-              searchRef.current?.focus();
             }}
-            className="absolute right-2 top-1/2 -translate-y-1/2 text-muted-foreground hover:text-foreground"
-            aria-label={t.clearSearch}
-          >
-            <X className="size-3" />
-          </button>
-        )}
-      </div>
+            placeholder={t.search}
+            className="w-full h-7 pl-8 pr-7 text-xs bg-transparent border border-border rounded-md focus:outline-none focus:ring-1 focus:ring-ring"
+            aria-label={t.searchAria}
+          />
+          {search && (
+            <button
+              type="button"
+              onClick={() => {
+                setSearch('');
+                setPage(0);
+                searchRef.current?.focus();
+              }}
+              className="absolute right-2 top-1/2 -translate-y-1/2 text-muted-foreground hover:text-foreground"
+              aria-label={t.clearSearch}
+            >
+              <X className="size-3" />
+            </button>
+          )}
+        </div>
+      )}
 
       <div className="overflow-x-auto relative">
         {watermark && isUnofficialDomain === false && (
diff --git a/packages/app/src/components/ui/date-picker.tsx b/packages/app/src/components/ui/date-picker.tsx
index 41bd4c4ec..985058f02 100644
--- a/packages/app/src/components/ui/date-picker.tsx
+++ b/packages/app/src/components/ui/date-picker.tsx
@@ -1,7 +1,7 @@
 'use client';
 
 import { Calendar, ChevronLeft, ChevronRight, Loader2 } from 'lucide-react';
-import { useEffect, useState } from 'react';
+import { useState } from 'react';
 
 import { track } from '@/lib/analytics';
 
@@ -26,6 +26,42 @@ import {
   DialogTitle,
   DialogTrigger,
 } from '@/components/ui/dialog';
+import { useLocale } from '@/lib/use-locale';
+
+const STRINGS = {
+  en: {
+    placeholder: 'Select date',
+    selectDateError: 'Please select a date',
+    unavailableDatePrefix: 'This date does not exist: ',
+    latest: 'Latest',
+    previousDate: 'Previous available date',
+    nextDate: 'Next available date',
+    runDate: 'Run Date:',
+    title: 'Select a Run Date',
+    description: 'Select a run date to view the performance data for that run.',
+    checking: 'Checking available dates...',
+    goToLatest: 'Go to Latest',
+    cancel: 'Cancel',
+    applying: 'Applying...',
+    apply: 'Apply',
+  },
+  zh: {
+    placeholder: '选择日期',
+    selectDateError: '请选择一个日期',
+    unavailableDatePrefix: '此日期不存在：',
+    latest: '最新',
+    previousDate: '上一个可用日期',
+    nextDate: '下一个可用日期',
+    runDate: '运行日期：',
+    title: '选择运行日期',
+    description: '选择运行日期以查看该次运行的性能数据。',
+    checking: '正在检查可用日期...',
+    goToLatest: '前往最新日期',
+    cancel: '取消',
+    applying: '正在应用...',
+    apply: '应用',
+  },
+} as const;
 
 export interface DatePickerProps {
   date?: string;
@@ -46,10 +82,13 @@ export function DatePicker({
   onChange,
   minDate,
   maxDate,
-  placeholder = 'Select date',
+  placeholder,
   availableDates,
   isCheckingAvailableDates,
 }: DatePickerProps) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const resolvedPlaceholder = placeholder ?? t.placeholder;
   const [open, setOpen] = useState(false);
   const [tempDate, setTempDate] = useState<string | undefined>(date);
   const [isApplying, _setIsApplying] = useState(false);
@@ -58,9 +97,9 @@ export function DatePicker({
   // Get display text for the input
   const getDisplayText = () => {
     if (!date) {
-      return placeholder;
+      return resolvedPlaceholder;
     }
-    return formatDisplayDate(date);
+    return formatDisplayDate(date, locale);
   };
 
   // Handle date selection in calendar
@@ -82,12 +121,12 @@ export function DatePicker({
   // Apply selection
   const handleApply = () => {
     if (!tempDate) {
-      setError('Please select a date');
+      setError(t.selectDateError);
       return;
     }
 
     if (availableDates && !availableDates.includes(tempDate)) {
-      setError(`This date does not exist: ${formatDisplayDate(tempDate)}`);
+      setError(`${t.unavailableDatePrefix}${formatDisplayDate(tempDate, locale)}`);
       return;
     }
 
@@ -176,16 +215,13 @@ export function DatePicker({
   // Reset when opening
   const handleOpenChange = (isOpen: boolean) => {
     track(isOpen ? 'date_picker_opened' : 'date_picker_closed');
+    setError('');
     if (isOpen) {
       setTempDate(date);
     }
     setOpen(isOpen);
   };
 
-  useEffect(() => {
-    setError('');
-  }, [open]);
-
   return (
     <div className="space-y-2">
       <div className="flex items-center gap-0.5">
@@ -196,7 +232,7 @@ export function DatePicker({
           disabled={isCurrentDateLatest() || Boolean(isCheckingAvailableDates)}
           className="text-xs px-2"
         >
-          Latest
+          {t.latest}
         </Button>
         <Button
           variant="ghost"
@@ -205,6 +241,7 @@ export function DatePicker({
           disabled={!canGoPrevious() || Boolean(isCheckingAvailableDates)}
           className="size-8"
           suppressHydrationWarning
+          aria-label={t.previousDate}
         >
           <ChevronLeft className="size-4" />
         </Button>
@@ -215,7 +252,7 @@ export function DatePicker({
               className="!px-5 min-w-[200px] dark:bg-input/90 dark:hover:bg-input/50"
             >
               <Calendar className="mr-0 size-4" />
-              <strong>Run Date:</strong>
+              <strong>{t.runDate}</strong>
               <span className="tabular-nums inline-block w-[6.5em] text-left">
                 {getDisplayText()}
               </span>
@@ -223,10 +260,8 @@ export function DatePicker({
           </DialogTrigger>
           <DialogContent className="sm:max-w-[600px]">
             <DialogHeader>
-              <DialogTitle>Select a Run Date</DialogTitle>
-              <DialogDescription>
-                Select a run date to view the performance data for that run.
-              </DialogDescription>
+              <DialogTitle>{t.title}</DialogTitle>
+              <DialogDescription>{t.description}</DialogDescription>
             </DialogHeader>
             <div className="py-4 relative">
               <CalendarGrid
@@ -241,7 +276,7 @@ export function DatePicker({
                 <div className="absolute inset-0 bg-background/80 backdrop-blur-sm flex items-center justify-center z-10 rounded-md">
                   <div className="flex flex-col items-center gap-2">
                     <Loader2 className="size-6 animate-spin text-primary" />
-                    <p className="text-sm text-muted-foreground">Checking available dates...</p>
+                    <p className="text-sm text-muted-foreground">{t.checking}</p>
                   </div>
                 </div>
               )}
@@ -253,16 +288,16 @@ export function DatePicker({
                 onClick={handleGoToLatest}
                 disabled={isLatestDateSelected() || Boolean(isCheckingAvailableDates)}
               >
-                Go to Latest
+                {t.goToLatest}
               </Button>
               <div className="flex gap-2">
                 <DialogClose asChild>
                   <Button variant="outline" onClick={handleCancel}>
-                    Cancel
+                    {t.cancel}
                   </Button>
                 </DialogClose>
                 <Button onClick={handleApply} disabled={isApplying}>
-                  {isApplying ? 'Applying...' : 'Apply'}
+                  {isApplying ? t.applying : t.apply}
                 </Button>
               </div>
             </div>
@@ -274,6 +309,7 @@ export function DatePicker({
           onClick={handleGoNext}
           disabled={!canGoNext() || Boolean(isCheckingAvailableDates)}
           className="size-8"
+          aria-label={t.nextDate}
         >
           <ChevronRight className="size-4" />
         </Button>
diff --git a/packages/app/src/components/ui/date-range-picker.tsx b/packages/app/src/components/ui/date-range-picker.tsx
index aaa941ab3..8b1e0be63 100644
--- a/packages/app/src/components/ui/date-range-picker.tsx
+++ b/packages/app/src/components/ui/date-range-picker.tsx
@@ -1,7 +1,7 @@
 'use client';
 
 import { Calendar, Loader2 } from 'lucide-react';
-import { useEffect, useState } from 'react';
+import { useState } from 'react';
 
 import { track } from '@/lib/analytics';
 
@@ -26,6 +26,53 @@ import {
   DialogTrigger,
 } from '@/components/ui/dialog';
 import { cn } from '@/lib/utils';
+import { useLocale } from '@/lib/use-locale';
+
+const STRINGS = {
+  en: {
+    placeholder: 'Select date range',
+    unavailableDatesPrefix: 'These dates do not exist: ',
+    title: 'Select Date Range',
+    selected: 'Selected:',
+    startDate: 'Start date:',
+    chooseEndDate: 'Choose an end date',
+    chooseRange: 'Choose a start and end date to define your date range.',
+    checking: 'Checking available dates...',
+    noAvailableDates: 'No available dates',
+    noAvailableDatesDescription: 'Please change Model, ISL/OSL, or Chip to see available dates.',
+    oneDateAvailable: 'Only 1 date available',
+    oneDateDescription:
+      'Historical comparison requires at least 2 dates. Please change Model, ISL/OSL, or chip selection.',
+    viewAnyway: 'View anyway',
+    maxRange: 'Max Range',
+    last90Days: 'Last 90 Days',
+    last30Days: 'Last 30 Days',
+    cancel: 'Cancel',
+    applying: 'Applying...',
+    apply: 'Apply',
+  },
+  zh: {
+    placeholder: '选择日期范围',
+    unavailableDatesPrefix: '以下日期不存在：',
+    title: '选择日期范围',
+    selected: '已选择：',
+    startDate: '开始日期：',
+    chooseEndDate: '请选择结束日期',
+    chooseRange: '请选择开始和结束日期以定义日期范围。',
+    checking: '正在检查可用日期...',
+    noAvailableDates: '没有可用日期',
+    noAvailableDatesDescription: '请更改模型、ISL/OSL 或芯片以查看可用日期。',
+    oneDateAvailable: '仅有 1 个可用日期',
+    oneDateDescription: '历史对比至少需要 2 个日期。请更改模型、ISL/OSL 或芯片选择。',
+    viewAnyway: '仍然查看',
+    maxRange: '最大范围',
+    last90Days: '最近 90 天',
+    last30Days: '最近 30 天',
+    cancel: '取消',
+    applying: '正在应用...',
+    apply: '应用',
+  },
+} as const;
 
 export interface DateRange {
   startDate: string;
@@ -51,12 +98,15 @@ export function DateRangePicker({
   dateRange,
   onChange,
   className,
-  placeholder = 'Select date range',
+  placeholder,
   minDate,
   maxDate,
   availableDates,
   isCheckingAvailableDates,
 }: DateRangePickerProps) {
+  const locale = useLocale();
+  const t = STRINGS[locale];
+  const resolvedPlaceholder = placeholder ?? t.placeholder;
   const [open, setOpen] = useState(false);
   const [tempRange, setTempRange] = useState<DateRange>(dateRange);
   const [selectingStart, setSelectingStart] = useState(true);
@@ -67,18 +117,18 @@ export function DateRangePicker({
   // Get display text for the input
   const getDisplayText = () => {
     if (!dateRange.startDate && !dateRange.endDate) {
-      return placeholder;
+      return resolvedPlaceholder;
     }
     if (dateRange.startDate && dateRange.endDate) {
       if (dateRange.startDate === dateRange.endDate) {
-        return formatDisplayDate(dateRange.startDate);
+        return formatDisplayDate(dateRange.startDate, locale);
       }
-      return `${formatDisplayDate(dateRange.startDate)} - ${formatDisplayDate(dateRange.endDate)}`;
+      return `${formatDisplayDate(dateRange.startDate, locale)} - ${formatDisplayDate(dateRange.endDate, locale)}`;
     }
     if (dateRange.startDate) {
-      return `${formatDisplayDate(dateRange.startDate)} - ...`;
+      return `${formatDisplayDate(dateRange.startDate, locale)} - ...`;
     }
-    return placeholder;
+    return resolvedPlaceholder;
   };
 
   // Handle date selection in calendar
@@ -108,7 +158,7 @@ export function DateRangePicker({
         const dates = [tempRange.startDate, tempRange.endDate];
         const failedDates = dates.filter((date) => !availableDates.includes(date));
         if (failedDates.length > 0) {
-          setError(`These dates do not exist: ${failedDates.join(', ')}`);
+          setError(`${t.unavailableDatesPrefix}${failedDates.join(', ')}`);
           return;
         }
       }
@@ -125,9 +175,9 @@ export function DateRangePicker({
     setOpen(false);
   };
 
-  // Reset when opening
   const handleOpenChange = (isOpen: boolean) => {
     track(isOpen ? 'date_range_picker_opened' : 'date_range_picker_closed');
+    setError('');
     if (isOpen) {
       setTempRange(dateRange);
       setSelectingStart(!dateRange.startDate || Boolean(dateRange.endDate));
@@ -135,10 +185,6 @@ export function DateRangePicker({
     setOpen(isOpen);
   };
 
-  useEffect(() => {
-    setError('');
-  }, [open]);
-
   return (
     <div className="space-y-2">
       <Dialog open={open} onOpenChange={handleOpenChange}>
@@ -157,26 +203,26 @@ export function DateRangePicker({
         </DialogTrigger>
         <DialogContent className="sm:max-w-[900px]">
           <DialogHeader>
-            <DialogTitle>Select Date Range</DialogTitle>
+            <DialogTitle>{t.title}</DialogTitle>
             <DialogDescription>
               {tempRange.startDate && tempRange.endDate ? (
                 <span>
-                  Selected:{' '}
+                  {t.selected}{' '}
                   <span className="font-semibold text-foreground">
-                    {formatDisplayDate(tempRange.startDate)} -{' '}
-                    {formatDisplayDate(tempRange.endDate)}
+                    {formatDisplayDate(tempRange.startDate, locale)} -{' '}
+                    {formatDisplayDate(tempRange.endDate, locale)}
                   </span>
                 </span>
               ) : tempRange.startDate ? (
                 <span>
-                  Start date:{' '}
+                  {t.startDate}{' '}
                   <span className="font-semibold text-foreground">
-                    {formatDisplayDate(tempRange.startDate)}
+                    {formatDisplayDate(tempRange.startDate, locale)}
                   </span>{' '}
-                  - Choose an end date
+                  - {t.chooseEndDate}
                 </span>
               ) : (
-                'Choose a start and end date to define your date range.'
+                t.chooseRange
               )}
             </DialogDescription>
           </DialogHeader>
@@ -196,7 +242,7 @@ export function DateRangePicker({
               <div className="absolute inset-0 bg-background/80 backdrop-blur-sm flex items-center justify-center z-10 rounded-md">
                 <div className="flex flex-col items-center gap-2">
                   <Loader2 className="size-6 animate-spin text-primary" />
-                  <p className="text-sm text-muted-foreground">Checking available dates...</p>
+                  <p className="text-sm text-muted-foreground">{t.checking}</p>
                 </div>
               </div>
             )}
@@ -205,10 +251,8 @@ export function DateRangePicker({
               availableDates.length === 0 && (
                 <div className="absolute inset-0 bg-background/80 backdrop-blur-sm flex items-center justify-center z-10 rounded-md">
                   <div className="flex flex-col items-center gap-2 text-center px-4">
-                    <p className="text-sm font-medium text-foreground">No available dates</p>
-                    <p className="text-xs text-muted-foreground">
-                      Please change Model, ISL/OSL, or Chip to see available dates.
-                    </p>
+                    <p className="text-sm font-medium text-foreground">{t.noAvailableDates}</p>
+                    <p className="text-xs text-muted-foreground">{t.noAvailableDatesDescription}</p>
                   </div>
                 </div>
               )}
@@ -217,11 +261,8 @@ export function DateRangePicker({
               availableDates.length === 1 && (
                 <div className="absolute inset-0 bg-background/80 backdrop-blur-sm flex items-center justify-center z-10 rounded-md">
                   <div className="flex flex-col items-center gap-2 text-center px-4">
-                    <p className="text-sm font-medium text-foreground">Only 1 date available</p>
-                    <p className="text-xs text-muted-foreground">
-                      Historical comparison requires at least 2 dates. Please change Model, ISL/OSL,
-                      or chip selection.
-                    </p>
+                    <p className="text-sm font-medium text-foreground">{t.oneDateAvailable}</p>
+                    <p className="text-xs text-muted-foreground">{t.oneDateDescription}</p>
                     <Button
                       variant="outline"
                       size="sm"
@@ -233,7 +274,7 @@ export function DateRangePicker({
                         setOpen(false);
                       }}
                     >
-                      View anyway
+                      {t.viewAnyway}
                     </Button>
                   </div>
                 </div>
@@ -245,14 +286,16 @@ export function DateRangePicker({
               <div className="flex flex-wrap gap-1.5">
                 {[
                   {
-                    label: 'Max Range',
+                    label: t.maxRange,
+                    analyticsLabel: 'Max Range',
                     getRange: () => ({
                       startDate: availableDates[0],
                       endDate: availableDates.at(-1)!,
                     }),
                   },
                   {
-                    label: 'Last 90 Days',
+                    label: t.last90Days,
+                    analyticsLabel: 'Last 90 Days',
                     getRange: () => {
                       const cutoff = new Date();
                       cutoff.setDate(cutoff.getDate() - 90);
@@ -263,7 +306,8 @@ export function DateRangePicker({
                     },
                   },
                   {
-                    label: 'Last 30 Days',
+                    label: t.last30Days,
+                    analyticsLabel: 'Last 30 Days',
                     getRange: () => {
                       const cutoff = new Date();
                       cutoff.setDate(cutoff.getDate() - 30);
@@ -273,7 +317,7 @@ export function DateRangePicker({
                       return { startDate: filtered[0], endDate: filtered.at(-1)! };
                     },
                   },
-                ].map(({ label, getRange }) => {
+                ].map(({ label, analyticsLabel, getRange }) => {
                   const range = getRange();
                   if (!range) return null;
                   return (
@@ -282,7 +326,7 @@ export function DateRangePicker({
                       variant="outline"
                       onClick={() => {
                         setTempRange(range);
-                        track('date_range_picker_quick_select', { label });
+                        track('date_range_picker_quick_select', { label: analyticsLabel });
                       }}
                     >
                       {label}
@@ -296,7 +340,7 @@ export function DateRangePicker({
             <div className="flex gap-2">
               <DialogClose asChild>
                 <Button variant="outline" onClick={handleCancel}>
-                  Cancel
+                  {t.cancel}
                 </Button>
               </DialogClose>
               <Button
@@ -308,7 +352,7 @@ export function DateRangePicker({
                   (availableDates !== undefined && availableDates.length < 2)
                 }
               >
-                {isApplying ? 'Applying...' : 'Apply'}
+                {isApplying ? t.applying : t.apply}
               </Button>
             </div>
           </DialogFooter>
diff --git a/packages/app/src/components/ui/dialog.tsx b/packages/app/src/components/ui/dialog.tsx
index b0f6d3cab..8d71d1941 100644
--- a/packages/app/src/components/ui/dialog.tsx
+++ b/packages/app/src/components/ui/dialog.tsx
@@ -5,6 +5,7 @@ import { X } from 'lucide-react';
 import * as React from 'react';
 
 import { cn } from '@/lib/utils';
+import { useLocale } from '@/lib/use-locale';
 
 const Dialog = DialogPrimitive.Root;
 
@@ -32,25 +33,28 @@ DialogOverlay.displayName = DialogPrimitive.Overlay.displayName;
 const DialogContent = React.forwardRef<
   React.ElementRef<typeof DialogPrimitive.Content>,
   React.ComponentPropsWithoutRef<typeof DialogPrimitive.Content>
->(({ className, children, ...props }, ref) => (
-  <DialogPortal>
-    <DialogOverlay />
-    <DialogPrimitive.Content
-      ref={ref}
-      className={cn(
-        'fixed left-[50%] top-[50%] z-50 grid w-full max-w-lg translate-x-[-50%] translate-y-[-50%] gap-4 border bg-background p-6 shadow-lg duration-200 data-[state=open]:animate-in data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=open]:fade-in-0 data-[state=closed]:zoom-out-95 data-[state=open]:zoom-in-95 data-[state=closed]:slide-out-to-left-1/2 data-[state=closed]:slide-out-to-top-[48%] data-[state=open]:slide-in-from-left-1/2 data-[state=open]:slide-in-from-top-[48%] rounded-lg',
-        className,
-      )}
-      {...props}
-    >
-      {children}
-      <DialogPrimitive.Close className="absolute right-4 top-4 rounded-sm opacity-70 ring-offset-background transition-opacity hover:opacity-100 focus:outline-none focus:ring-2 focus:ring-ring focus:ring-offset-2 disabled:pointer-events-none data-[state=open]:bg-accent data-[state=open]:text-muted-foreground">
-        <X className="size-4" />
-        <span className="sr-only">Close</span>
-      </DialogPrimitive.Close>
-    </DialogPrimitive.Content>
-  </DialogPortal>
-));
+>(({ className, children, ...props }, ref) => {
+  const closeLabel = useLocale() === 'zh' ? '关闭' : 'Close';
+  return (
+    <DialogPortal>
+      <DialogOverlay />
+      <DialogPrimitive.Content
+        ref={ref}
+        className={cn(
+          'fixed left-[50%] top-[50%] z-50 grid w-full max-w-lg translate-x-[-50%] translate-y-[-50%] gap-4 border bg-background p-6 shadow-lg duration-200 data-[state=open]:animate-in data-[state=closed]:animate-out data-[state=closed]:fade-out-0 data-[state=open]:fade-in-0 data-[state=closed]:zoom-out-95 data-[state=open]:zoom-in-95 data-[state=closed]:slide-out-to-left-1/2 data-[state=closed]:slide-out-to-top-[48%] data-[state=open]:slide-in-from-left-1/2 data-[state=open]:slide-in-from-top-[48%] rounded-lg',
+          className,
+        )}
+        {...props}
+      >
+        {children}
+        <DialogPrimitive.Close className="absolute right-4 top-4 rounded-sm opacity-70 ring-offset-background transition-opacity hover:opacity-100 focus:outline-none focus:ring-2 focus:ring-ring focus:ring-offset-2 disabled:pointer-events-none data-[state=open]:bg-accent data-[state=open]:text-muted-foreground">
+          <X className="size-4" />
+          <span className="sr-only">{closeLabel}</span>
+        </DialogPrimitive.Close>
+      </DialogPrimitive.Content>
+    </DialogPortal>
+  );
+});
 DialogContent.displayName = DialogPrimitive.Content.displayName;
 
 const DialogHeader = ({ className, ...props }: React.HTMLAttributes<HTMLDivElement>) => (
diff --git a/packages/app/src/components/ui/eyebrow.tsx b/packages/app/src/components/ui/eyebrow.tsx
new file mode 100644
index 000000000..127d845bd
--- /dev/null
+++ b/packages/app/src/components/ui/eyebrow.tsx
@@ -0,0 +1,47 @@
+import { cva, type VariantProps } from 'class-variance-authority';
+import * as React from 'react';
+
+import { cn } from '@/lib/utils';
+
+/**
+ * Uppercase mono eyebrow label — the small kicker above headings. Wraps the
+ * single most repeated typography recipe in the app so its size, weight, and
+ * letter-spacing stay uniform. See docs/typography.md.
+ */
+const eyebrowVariants = cva('font-mono text-xs font-semibold uppercase', {
+  variants: {
+    tone: {
+      brand: 'text-brand',
+      muted: 'text-muted-foreground',
+    },
+    wide: {
+      true: 'tracking-eyebrow-wide',
+      false: 'tracking-eyebrow',
+    },
+  },
+  defaultVariants: {
+    tone: 'brand',
+    wide: false,
+  },
+});
+
+function Eyebrow({
+  className,
+  tone,
+  wide,
+  as: Tag = 'span',
+  ...props
+}: React.HTMLAttributes<HTMLElement> &
+  VariantProps<typeof eyebrowVariants> & {
+    as?: 'span' | 'p' | 'div';
+  }) {
+  return (
+    <Tag
+      data-slot="eyebrow"
+      className={cn(eyebrowVariants({ tone, wide, className }))}
+      {...props}
+    />
+  );
+}
+
+export { Eyebrow, eyebrowVariants };
diff --git a/packages/app/src/components/ui/heading.test.ts b/packages/app/src/components/ui/heading.test.ts
new file mode 100644
index 000000000..56566359e
--- /dev/null
+++ b/packages/app/src/components/ui/heading.test.ts
@@ -0,0 +1,51 @@
+import { describe, expect, it } from 'vitest';
+
+import { eyebrowVariants } from './eyebrow';
+import { headingVariants } from './heading';
+import { cn } from '@/lib/utils';
+
+describe('headingVariants', () => {
+  it('renders the page-title recipe used across route h1s', () => {
+    expect(headingVariants({ level: 'page' })).toBe(
+      'text-foreground text-2xl lg:text-4xl font-bold tracking-tight',
+    );
+  });
+
+  it('defaults to the section level', () => {
+    expect(headingVariants({})).toContain('text-xl');
+  });
+});
+
+describe('eyebrowVariants', () => {
+  it('matches the dominant eyebrow recipe by default', () => {
+    const classes = eyebrowVariants({});
+    for (const cls of [
+      'font-mono',
+      'text-xs',
+      'font-semibold',
+      'uppercase',
+      'text-brand',
+      'tracking-eyebrow',
+    ]) {
+      expect(classes).toContain(cls);
+    }
+  });
+
+  it('switches to wide tracking', () => {
+    expect(eyebrowVariants({ wide: true })).toContain('tracking-eyebrow-wide');
+    expect(eyebrowVariants({ wide: true })).not.toContain('tracking-eyebrow ');
+  });
+});
+
+describe('cn merging of custom typography tokens', () => {
+  it('resolves conflicts between named tracking tokens and built-ins (last wins)', () => {
+    expect(cn('tracking-eyebrow', 'tracking-tight')).toBe('tracking-tight');
+    expect(cn('tracking-tight', 'tracking-eyebrow')).toBe('tracking-eyebrow');
+    expect(cn('tracking-eyebrow', 'tracking-eyebrow-wide')).toBe('tracking-eyebrow-wide');
+  });
+
+  it('resolves conflicts between micro size tokens and the standard scale', () => {
+    expect(cn('text-sm', 'text-2xs')).toBe('text-2xs');
+    expect(cn('text-3xs', 'text-lg')).toBe('text-lg');
+  });
+});
diff --git a/packages/app/src/components/ui/heading.tsx b/packages/app/src/components/ui/heading.tsx
new file mode 100644
index 000000000..f478e3937
--- /dev/null
+++ b/packages/app/src/components/ui/heading.tsx
@@ -0,0 +1,49 @@
+import { cva, type VariantProps } from 'class-variance-authority';
+import * as React from 'react';
+
+import { cn } from '@/lib/utils';
+
+/**
+ * Visual heading levels, derived from the recipes already in use across the
+ * app. The rendered tag is chosen independently via `as`, so document outline
+ * and visual size never fight each other. New pages should reach for this
+ * instead of retyping a size/weight/tracking combo; existing headings migrate
+ * as their files get touched. See docs/typography.md.
+ */
+const headingVariants = cva('text-foreground', {
+  variants: {
+    level: {
+      /** Landing/marketing hero. */
+      display: 'text-4xl font-bold tracking-heading text-balance md:text-5xl',
+      /** Route-level page title. */
+      page: 'text-2xl lg:text-4xl font-bold tracking-tight',
+      /** Section head within a page. */
+      section: 'text-xl font-semibold',
+      /** Card or panel head. */
+      card: 'text-base font-semibold',
+      /** Dense sub-head above tables or control groups. */
+      label: 'text-sm font-medium',
+    },
+  },
+  defaultVariants: {
+    level: 'section',
+  },
+});
+
+type HeadingTag = 'h1' | 'h2' | 'h3' | 'h4' | 'h5' | 'h6';
+
+function Heading({
+  className,
+  level,
+  as: Tag = 'h2',
+  ...props
+}: React.ComponentProps<'h2'> &
+  VariantProps<typeof headingVariants> & {
+    as?: HeadingTag;
+  }) {
+  return (
+    <Tag data-slot="heading" className={cn(headingVariants({ level, className }))} {...props} />
+  );
+}
+
+export { Heading, headingVariants };
diff --git a/packages/app/src/components/ui/hw-vendor-logo.tsx b/packages/app/src/components/ui/hw-vendor-logo.tsx
new file mode 100644
index 000000000..ee66ce10c
--- /dev/null
+++ b/packages/app/src/components/ui/hw-vendor-logo.tsx
@@ -0,0 +1,45 @@
+'use client';
+
+import { useState } from 'react';
+
+import { isMonochromeLogo } from '@/lib/model-logos';
+import { getHwVendorLogo } from '@/lib/vendor-logos';
+import { cn } from '@/lib/utils';
+
+/**
+ * Small inline hardware-vendor mark, rendered beside a GPU label (e.g. in the
+ * `/compare` pair cards). Resolves the asset via `getHwVendorLogo` and renders
+ * nothing when the vendor has no configured logo or the asset fails to load,
+ * so surrounding text never gains a broken-image box.
+ *
+ * Full-color marks (NVIDIA green) render as-is in both themes; monochrome
+ * marks (AMD — black / reversed-white only per brand guidelines) get the
+ * shared dark-mode invert. The image is decorative — the hardware label is
+ * the adjacent text — so it is hidden from assistive technology.
+ */
+export function HwVendorLogo({ vendor, className }: { vendor?: string; className?: string }) {
+  // Track WHICH asset failed rather than a boolean, mirroring ModelLogo: a
+  // reused instance should not suppress a different vendor's logo after one
+  // load error.
+  const [failedSrc, setFailedSrc] = useState<string | null>(null);
+  const logo = getHwVendorLogo(vendor);
+
+  if (!logo || failedSrc === logo) return null;
+
+  return (
+    <img
+      src={`/logos/${logo}`}
+      alt=""
+      aria-hidden="true"
+      className={cn(
+        // Height-bounded with natural aspect ratio: the NVIDIA eye mark is
+        // square-ish while the AMD lockup is a wide wordmark; max-w keeps the
+        // wordmark from crowding the card title.
+        'inline-block h-4 w-auto max-w-12 shrink-0 object-contain align-text-bottom',
+        isMonochromeLogo(logo) && 'dark:invert',
+        className,
+      )}
+      onError={() => setFailedSrc(logo)}
+    />
+  );
+}
diff --git a/packages/app/src/components/ui/label-with-tooltip.tsx b/packages/app/src/components/ui/label-with-tooltip.tsx
index 11b0bb96c..c0e1214d0 100644
--- a/packages/app/src/components/ui/label-with-tooltip.tsx
+++ b/packages/app/src/components/ui/label-with-tooltip.tsx
@@ -7,7 +7,12 @@ import { Label } from '@/components/ui/label';
 import { TooltipContent, TooltipRoot, TooltipTrigger } from '@/components/ui/tooltip';
 
 interface LabelWithTooltipProps {
-  htmlFor: string;
+  /**
+   * Omit for controls that are not labelable elements — a segmented toggle is a
+   * `role="tablist"`, so `for` would dangle and its `ariaLabel` is the accessible
+   * name instead.
+   */
+  htmlFor?: string;
   label: string;
   tooltip: ReactNode;
 }
diff --git a/packages/app/src/components/ui/model-logo.tsx b/packages/app/src/components/ui/model-logo.tsx
new file mode 100644
index 000000000..ac5dfc408
--- /dev/null
+++ b/packages/app/src/components/ui/model-logo.tsx
@@ -0,0 +1,46 @@
+'use client';
+
+import { useState } from 'react';
+
+import { getModelLogo, type Model } from '@/lib/data-mappings';
+import { isMonochromeLogo } from '@/lib/model-logos';
+import { cn } from '@/lib/utils';
+
+/**
+ * Small inline logo for the model creator, rendered beside the model name
+ * (e.g. in the inference chart caption). Resolves the asset via
+ * `getModelLogo` and renders nothing when the model has no configured logo or
+ * the asset fails to load, so surrounding text never gains a broken-image box.
+ *
+ * Monochrome marks (black / `currentColor`, e.g. kimi, openai) are inverted
+ * in dark mode so they stay visible; full-color logos (e.g. deepseek-color,
+ * qwen-color) are shown as-is in both themes. The image is decorative — the
+ * model name is the adjacent text — so it is hidden from assistive
+ * technology.
+ */
+export function ModelLogo({ model, className }: { model: Model; className?: string }) {
+  // Track WHICH asset failed rather than a boolean: chart sections are keyed by
+  // index, so one ModelLogo instance is reused across model switches and a
+  // stale boolean would keep suppressing every later model's logo after a
+  // single load error.
+  const [failedSrc, setFailedSrc] = useState<string | null>(null);
+  const logo = getModelLogo(model);
+
+  if (!logo || failedSrc === logo) return null;
+
+  return (
+    <img
+      src={`/logos/${logo}`}
+      alt=""
+      aria-hidden="true"
+      width={16}
+      height={16}
+      className={cn(
+        'inline-block size-4 shrink-0 object-contain align-text-bottom',
+        isMonochromeLogo(logo) && 'dark:invert',
+        className,
+      )}
+      onError={() => setFailedSrc(logo)}
+    />
+  );
+}
diff --git a/packages/app/src/components/ui/multi-date-picker.tsx b/packages/app/src/components/ui/multi-date-picker.tsx
deleted file mode 100644
index 31cf69c29..000000000
--- a/packages/app/src/components/ui/multi-date-picker.tsx
+++ /dev/null
@@ -1,301 +0,0 @@
-'use client';
-
-import { Calendar, X } from 'lucide-react';
-import { useEffect, useState } from 'react';
-
-import { track } from '@/lib/analytics';
-
-import { Button } from '@/components/ui/button';
-import {
-  CalendarMonthPanel,
-  formatCalendarDate,
-  formatDisplayDate,
-  getCalendarMonthNavState,
-  isCalendarDateOutOfRange,
-  resolveCalendarDateBounds,
-  useCalendarMonth,
-} from '@/components/ui/calendar-picker-utils';
-import {
-  Dialog,
-  DialogClose,
-  DialogContent,
-  DialogDescription,
-  DialogFooter,
-  DialogHeader,
-  DialogTitle,
-  DialogTrigger,
-} from '@/components/ui/dialog';
-import { cn } from '@/lib/utils';
-export interface MultiDatePickerProps {
-  dates: string[];
-  onChange: (dates: string[]) => void;
-  maxDates?: number;
-  minDate?: string;
-  maxDate?: string;
-  className?: string;
-  placeholder?: string;
-  availableDates?: string[];
-}
-
-/**
- * Multi-date picker component that allows selecting multiple dates via a modal calendar.
- * Displays individual date for 1 date, "x vs y" for 2 dates, or "N dates selected" for 3+ dates.
- */
-export function MultiDatePicker({
-  dates,
-  onChange,
-  maxDates = 2,
-  minDate,
-  maxDate,
-  className,
-  placeholder = 'Select dates',
-  availableDates,
-}: MultiDatePickerProps) {
-  const [open, setOpen] = useState(false);
-  const [tempDates, setTempDates] = useState<string[]>(dates);
-  const [isApplying, _setIsApplying] = useState(false);
-  const [error, setError] = useState('');
-
-  // Get display text for the input
-  const getDisplayText = () => {
-    if (dates.length === 0) {
-      return placeholder;
-    }
-    if (dates.length === 1) {
-      return formatDisplayDate(dates[0]);
-    }
-    if (dates.length === 2) {
-      return `${formatDisplayDate(dates[0])} vs ${formatDisplayDate(dates[1])}`;
-    }
-    return `${dates.length} dates selected`;
-  };
-
-  // Handle date selection in calendar
-  const handleDateClick = (date: Date) => {
-    const dateStr = formatCalendarDate(date);
-    const isSelected = tempDates.includes(dateStr);
-
-    if (isSelected) {
-      setTempDates(tempDates.filter((d) => d !== dateStr));
-    } else if (tempDates.length < maxDates) {
-      setTempDates([...tempDates, dateStr].toSorted());
-    }
-    track('multi_date_picker_date_clicked', { date: dateStr, selected: !isSelected });
-  };
-
-  // Remove a specific date from temp selection
-  const handleRemoveTempDate = (dateStr: string) => {
-    track('multi_date_picker_date_removed', { date: dateStr });
-    setTempDates(tempDates.filter((d) => d !== dateStr));
-  };
-
-  // Apply selection
-  const handleApply = () => {
-    if (availableDates) {
-      const failedDates = tempDates.filter((date) => !availableDates.includes(date));
-      if (failedDates.length > 0) {
-        setError(`These dates do not exist: ${failedDates.join(', ')}`);
-        return;
-      }
-    }
-
-    track('multi_date_picker_applied', { dates: tempDates });
-    onChange(tempDates);
-    setOpen(false);
-  };
-
-  // Cancel selection
-  const handleCancel = () => {
-    setTempDates(dates);
-    setOpen(false);
-  };
-
-  // Reset when opening
-  const handleOpenChange = (isOpen: boolean) => {
-    track(isOpen ? 'multi_date_picker_opened' : 'multi_date_picker_closed');
-    if (isOpen) {
-      setTempDates(dates);
-    }
-    setOpen(isOpen);
-  };
-
-  useEffect(() => {
-    setError('');
-  }, [open]);
-
-  return (
-    <div className="space-y-2">
-      <Dialog open={open} onOpenChange={handleOpenChange}>
-        <DialogTrigger asChild>
-          <Button
-            variant="outline"
-            className={cn(
-              'w-full justify-start text-left font-normal',
-              dates.length === 0 && 'text-muted-foreground',
-              className,
-            )}
-          >
-            <Calendar className="mr-2 size-4" />
-            {getDisplayText()}
-          </Button>
-        </DialogTrigger>
-        <DialogContent className="sm:max-w-[600px]">
-          <DialogHeader>
-            <DialogTitle>Select Comparison Dates</DialogTitle>
-            <DialogDescription>
-              Choose up to {maxDates} dates to compare chip performance over time.
-            </DialogDescription>
-          </DialogHeader>
-          <div className="py-4">
-            <CalendarGrid
-              selectedDates={tempDates}
-              onDateClick={handleDateClick}
-              maxDates={maxDates}
-              minDate={minDate}
-              maxDate={maxDate}
-              availableDates={availableDates}
-            />
-            {tempDates.length > 0 && maxDates > 1 && (
-              <div className="mt-4 p-3 bg-muted/30 rounded-md">
-                <div className="flex items-center justify-between mb-2">
-                  <p className="text-sm font-medium">Selected Dates:</p>
-                  <Button
-                    variant="ghost"
-                    size="sm"
-                    onClick={() => {
-                      track('multi_date_picker_cleared', { dateCount: tempDates.length });
-                      setTempDates([]);
-                    }}
-                    className="h-6 px-2 text-xs"
-                  >
-                    Clear All
-                  </Button>
-                </div>
-                <div className="flex flex-wrap gap-2">
-                  {tempDates.map((dateStr, index) => (
-                    <div
-                      key={index}
-                      className="px-2 py-1 bg-primary text-primary-foreground rounded-md text-xs flex items-center gap-1 group"
-                    >
-                      {formatDisplayDate(dateStr)}
-                      <button
-                        type="button"
-                        onClick={(e) => {
-                          e.stopPropagation();
-                          handleRemoveTempDate(dateStr);
-                        }}
-                        className="ml-1 hover:bg-primary-foreground/20 rounded-sm p-0.5 transition-colors"
-                        aria-label={`Remove ${formatDisplayDate(dateStr)}`}
-                      >
-                        <X className="size-3" />
-                      </button>
-                    </div>
-                  ))}
-                </div>
-              </div>
-            )}
-          </div>
-          {error && <p className="text-md text-center text-red-500">{error}</p>}
-          <DialogFooter>
-            <DialogClose asChild>
-              <Button variant="outline" onClick={handleCancel}>
-                Cancel
-              </Button>
-            </DialogClose>
-            <Button onClick={handleApply} disabled={isApplying}>
-              {isApplying ? 'Applying...' : 'Apply'}
-            </Button>
-          </DialogFooter>
-        </DialogContent>
-      </Dialog>
-    </div>
-  );
-}
-
-interface CalendarGridProps {
-  selectedDates: string[];
-  onDateClick: (date: Date) => void;
-  maxDates: number;
-  minDate?: string;
-  maxDate?: string;
-  availableDates?: string[];
-}
-
-function CalendarGrid({
-  selectedDates,
-  onDateClick,
-  maxDates,
-  minDate,
-  maxDate,
-  availableDates,
-}: CalendarGridProps) {
-  const { minAllowedDate, maxAllowedDate, earliestMonth, latestMonth } = resolveCalendarDateBounds(
-    minDate,
-    maxDate,
-    availableDates,
-    '2025-10-10',
-  );
-  const [currentMonth, setCurrentMonth] = useCalendarMonth(
-    selectedDates[0],
-    availableDates,
-    maxAllowedDate,
-    [selectedDates.join(',')],
-  );
-
-  const isDateSelected = (date: Date) => selectedDates.includes(formatCalendarDate(date));
-
-  const getDayState = (date: Date) => {
-    const outOfRange = isCalendarDateOutOfRange(date, minAllowedDate, maxAllowedDate, true);
-    const selected = isDateSelected(date);
-    const dateStr = formatCalendarDate(date);
-
-    return {
-      selected,
-      disabled:
-        outOfRange ||
-        (availableDates !== undefined && !availableDates.includes(dateStr)) ||
-        (selectedDates.length >= maxDates && !selected),
-      outOfRange,
-    };
-  };
-
-  const { canGoPrevious, canGoNext } = getCalendarMonthNavState(
-    currentMonth,
-    earliestMonth,
-    latestMonth,
-  );
-
-  const goToPreviousMonth = () => {
-    if (canGoPrevious) {
-      const newMonth = new Date(currentMonth.getFullYear(), currentMonth.getMonth() - 1);
-      track('multi_date_picker_month_navigated', {
-        direction: 'previous',
-        month: newMonth.toISOString().slice(0, 7),
-      });
-      setCurrentMonth(newMonth);
-    }
-  };
-
-  const goToNextMonth = () => {
-    if (canGoNext) {
-      const newMonth = new Date(currentMonth.getFullYear(), currentMonth.getMonth() + 1);
-      track('multi_date_picker_month_navigated', {
-        direction: 'next',
-        month: newMonth.toISOString().slice(0, 7),
-      });
-      setCurrentMonth(newMonth);
-    }
-  };
-
-  return (
-    <CalendarMonthPanel
-      month={currentMonth}
-      onPreviousMonth={goToPreviousMonth}
-      onNextMonth={goToNextMonth}
-      canGoPrevious={canGoPrevious}
-      canGoNext={canGoNext}
-      getDayState={getDayState}
-      onDateClick={onDateClick}
-    />
-  );
-}
diff --git a/packages/app/src/components/ui/multi-select.tsx b/packages/app/src/components/ui/multi-select.tsx
index 767edcab1..5f5b38e33 100644
--- a/packages/app/src/components/ui/multi-select.tsx
+++ b/packages/app/src/components/ui/multi-select.tsx
@@ -5,10 +5,40 @@ import * as React from 'react';
 
 import { track } from '@/lib/analytics';
 import { cn } from '@/lib/utils';
+import { useLocale } from '@/lib/use-locale';
+
+const STRINGS = {
+  en: {
+    placeholder: 'Select items...',
+    searchPlaceholder: 'Search...',
+    searchAriaLabel: 'Search options',
+    noResultsLabel: 'No results',
+    clearSearchLabel: 'Clear search',
+    selectedSuffix: ' selected',
+    minimumPrefix: 'Minimum: ',
+    removePrefix: 'Remove ',
+    clearAllSelections: 'Clear all selections',
+  },
+  zh: {
+    placeholder: '选择项目...',
+    searchPlaceholder: '搜索...',
+    searchAriaLabel: '搜索选项',
+    noResultsLabel: '没有结果',
+    clearSearchLabel: '清除搜索',
+    selectedSuffix: ' 项已选择',
+    minimumPrefix: '最少：',
+    removePrefix: '移除 ',
+    clearAllSelections: '清除所有选择',
+  },
+} as const;
 
 interface MultiSelectOption {
   value: string;
   label: string;
+  /** Optional leading visual (e.g. a brand logo) rendered before the label. */
+  icon?: React.ReactNode;
+  /** Optional trailing visual (e.g. a NEW pill) rendered after the label. */
+  badge?: React.ReactNode;
 }
 
 export interface MultiSelectSection {
@@ -39,6 +69,7 @@ interface MultiSelectProps {
   plainSelectedText?: boolean;
   showSelectionSummary?: boolean;
   searchPlaceholder?: string;
+  searchAriaLabel?: string;
   noResultsLabel?: string;
   clearSearchLabel?: string;
   selectedSuffix?: string;
@@ -54,7 +85,7 @@ function MultiSelect({
   triggerTestId,
   open,
   onOpenChange,
-  placeholder = 'Select items...',
+  placeholder,
   size = 'default',
   className,
   disabled = false,
@@ -64,12 +95,21 @@ function MultiSelect({
   searchable = true,
   plainSelectedText = false,
   showSelectionSummary = true,
-  searchPlaceholder = 'Search...',
-  noResultsLabel = 'No results',
-  clearSearchLabel = 'Clear search',
-  selectedSuffix = ' selected',
-  minimumPrefix = 'Minimum: ',
+  searchPlaceholder,
+  searchAriaLabel,
+  noResultsLabel,
+  clearSearchLabel,
+  selectedSuffix,
+  minimumPrefix,
 }: MultiSelectProps) {
+  const t = STRINGS[useLocale()];
+  const resolvedPlaceholder = placeholder ?? t.placeholder;
+  const resolvedSearchPlaceholder = searchPlaceholder ?? t.searchPlaceholder;
+  const resolvedSearchAriaLabel = searchAriaLabel ?? t.searchAriaLabel;
+  const resolvedNoResultsLabel = noResultsLabel ?? t.noResultsLabel;
+  const resolvedClearSearchLabel = clearSearchLabel ?? t.clearSearchLabel;
+  const resolvedSelectedSuffix = selectedSuffix ?? t.selectedSuffix;
+  const resolvedMinimumPrefix = minimumPrefix ?? t.minimumPrefix;
   const [internalIsOpen, setInternalIsOpen] = React.useState(false);
   const [search, setSearch] = React.useState('');
   const listboxId = React.useId();
@@ -85,19 +125,24 @@ function MultiSelect({
   const isOpen = isControlledOpen ? open : internalIsOpen;
   const setIsOpen = React.useCallback(
     (nextOpen: boolean) => {
+      if (!nextOpen && isOpen) {
+        if (searchUsedRef.current) {
+          track('multi_select_searched', { query: searchStateRef.current });
+          searchUsedRef.current = false;
+        }
+        setSearch('');
+      }
       if (!isControlledOpen) {
         setInternalIsOpen(nextOpen);
       }
       onOpenChange?.(nextOpen);
     },
-    [isControlledOpen, onOpenChange],
+    [isControlledOpen, isOpen, onOpenChange],
   );
 
   const isMaxReached = maxSelections !== undefined && value.length >= maxSelections;
   const isMinReached = minSelections !== undefined && value.length <= minSelections;
 
-  const prevIsOpenRef = React.useRef(isOpen);
-
   React.useEffect(() => {
     const handlePointerDownOutside = (event: PointerEvent) => {
       if (containerRef.current && !containerRef.current.contains(event.target as Node)) {
@@ -134,20 +179,7 @@ function MultiSelect({
       document.removeEventListener('focusin', handleFocusOutside);
       document.removeEventListener('keydown', handleKeyDown);
     };
-  }, [isOpen]);
-
-  React.useEffect(() => {
-    const wasOpen = prevIsOpenRef.current;
-    prevIsOpenRef.current = isOpen;
-
-    if (wasOpen && !isOpen) {
-      if (searchUsedRef.current) {
-        track('multi_select_searched', { query: searchStateRef.current });
-        searchUsedRef.current = false;
-      }
-      setSearch('');
-    }
-  }, [isOpen]);
+  }, [isOpen, setIsOpen]);
 
   const flatOptions = React.useMemo(() => {
     if (sections?.length) {
@@ -242,10 +274,10 @@ function MultiSelect({
   };
 
   // Preserve the order of selected values, not the order of options
-  const selectedLabels = value.map((val) => {
-    const option = flatOptions.find((opt) => opt.value === val);
-    return option ? option.label : val;
-  });
+  const selectedOptions = value.map(
+    (val) => flatOptions.find((opt) => opt.value === val) ?? { value: val, label: val },
+  );
+  const selectedLabels = selectedOptions.map((opt) => opt.label);
 
   return (
     <div ref={containerRef} className="relative">
@@ -270,8 +302,9 @@ function MultiSelect({
         <div className="flex gap-1 flex-1 min-w-0 items-center min-h-5 flex-wrap">
           {value.length > 0 ? (
             plainSelectedText ? (
-              <span className="text-foreground block min-w-0 truncate">
-                {selectedLabels.join(', ')}
+              <span className="text-foreground flex min-w-0 items-center gap-1.5">
+                {selectedOptions.length === 1 && selectedOptions[0].icon}
+                <span className="block min-w-0 truncate">{selectedLabels.join(', ')}</span>
               </span>
             ) : (
               selectedLabels.map((label, index) => (
@@ -279,6 +312,7 @@ function MultiSelect({
                   key={value[index]}
                   className="bg-transparent text-foreground border border-border dark:bg-[#0a6ca8] dark:border-border inline-flex items-center gap-1 rounded px-1.5 py-0.5 text-xs font-medium transition-colors shrink-0"
                 >
+                  {selectedOptions[index]?.icon}
                   {label}
                   <span
                     role="button"
@@ -294,7 +328,7 @@ function MultiSelect({
                       'hover:bg-primary/20 rounded-sm cursor-pointer transition-colors',
                       (disabled || isMinReached) && 'hidden',
                     )}
-                    aria-label={`Remove ${label}`}
+                    aria-label={`${t.removePrefix}${label}`}
                     aria-disabled={disabled || isMinReached}
                   >
                     <XIcon className="size-4 text-foreground" />
@@ -303,7 +337,7 @@ function MultiSelect({
               ))
             )
           ) : (
-            <span className="text-muted-foreground">{placeholder}</span>
+            <span className="text-muted-foreground">{resolvedPlaceholder}</span>
           )}
         </div>
         {value.length > 0 && showClearAll && (
@@ -322,7 +356,7 @@ function MultiSelect({
               (disabled || (minSelections !== undefined && minSelections > 0)) &&
                 'cursor-not-allowed opacity-50 pointer-events-none',
             )}
-            aria-label="Clear all selections"
+            aria-label={t.clearAllSelections}
             aria-disabled={disabled || (minSelections !== undefined && minSelections > 0)}
           >
             <XIcon className="size-4" />
@@ -358,7 +392,8 @@ function MultiSelect({
                     setSearch(e.target.value);
                     if (e.target.value) searchUsedRef.current = true;
                   }}
-                  placeholder={searchPlaceholder}
+                  placeholder={resolvedSearchPlaceholder}
+                  aria-label={resolvedSearchAriaLabel}
                   className="w-full bg-transparent py-1.5 text-sm outline-none placeholder:text-muted-foreground"
                 />
                 {search && (
@@ -369,7 +404,7 @@ function MultiSelect({
                       searchRef.current?.focus();
                     }}
                     className="shrink-0 text-muted-foreground hover:text-foreground transition-colors"
-                    aria-label={clearSearchLabel}
+                    aria-label={resolvedClearSearchLabel}
                   >
                     <XIcon className="size-3.5" />
                   </button>
@@ -381,10 +416,10 @@ function MultiSelect({
                 <div className="text-muted-foreground px-2 py-1.5 text-xs border-b mb-1">
                   {value.length}
                   {maxSelections !== undefined && ` / ${maxSelections}`}
-                  {selectedSuffix}
+                  {resolvedSelectedSuffix}
                   {minSelections !== undefined && minSelections > 0 && (
                     <span className="block text-xs mt-0.5">
-                      {minimumPrefix}
+                      {resolvedMinimumPrefix}
                       {minSelections}
                     </span>
                   )}
@@ -392,7 +427,7 @@ function MultiSelect({
               )}
             {filteredOptions.length === 0 && (
               <div className="text-muted-foreground px-2 py-1.5 text-sm text-center">
-                {noResultsLabel}
+                {resolvedNoResultsLabel}
               </div>
             )}
             {filteredSections
@@ -427,7 +462,11 @@ function MultiSelect({
                             <span className="absolute right-2 flex size-3.5 items-center justify-center">
                               {isSelected && <CheckIcon className="size-4 text-primary" />}
                             </span>
-                            <span className="flex items-center gap-2">{option.label}</span>
+                            <span className="flex min-w-0 items-center gap-2">
+                              {option.icon}
+                              {option.label}
+                              {option.badge}
+                            </span>
                           </div>
                         );
                       })}
@@ -456,7 +495,11 @@ function MultiSelect({
                       <span className="absolute right-2 flex size-3.5 items-center justify-center">
                         {isSelected && <CheckIcon className="size-4 text-primary" />}
                       </span>
-                      <span className="flex items-center gap-2">{option.label}</span>
+                      <span className="flex min-w-0 items-center gap-2">
+                        {option.icon}
+                        {option.label}
+                        {option.badge}
+                      </span>
                     </div>
                   );
                 })}
diff --git a/packages/app/src/components/ui/new-badge.tsx b/packages/app/src/components/ui/new-badge.tsx
new file mode 100644
index 000000000..2e1799ed3
--- /dev/null
+++ b/packages/app/src/components/ui/new-badge.tsx
@@ -0,0 +1,23 @@
+import type { ComponentPropsWithoutRef } from 'react';
+
+import { cn } from '@/lib/utils';
+
+export function NewBadge({ children, className, ...props }: ComponentPropsWithoutRef<'span'>) {
+  return (
+    <span
+      className={cn(
+        // 32x16 pill. `NEW` inks ~23px at 10px bold, so the horizontal room is
+        // slack rather than padding to spend: `min-w-8` holds the shared size
+        // for short labels (zh `新`) while letting an unusually wide label
+        // widen the pill instead of spilling its glyphs past the edge.
+        'inline-flex h-4 min-w-8 shrink-0 items-center justify-center rounded-full bg-brand px-0.5 text-3xs font-bold leading-none uppercase tracking-normal text-primary-foreground shadow-sm',
+        className,
+      )}
+      {...props}
+    >
+      <span data-new-badge-label className="flex h-full items-center justify-center text-center">
+        {children}
+      </span>
+    </span>
+  );
+}
diff --git a/packages/app/src/components/ui/searchable-select.tsx b/packages/app/src/components/ui/searchable-select.tsx
index 1c7d80ad5..3485208e6 100644
--- a/packages/app/src/components/ui/searchable-select.tsx
+++ b/packages/app/src/components/ui/searchable-select.tsx
@@ -67,15 +67,17 @@ export function SearchableSelect({
     setMounted(true);
   }, []);
 
-  React.useEffect(() => {
-    if (!isOpen) {
+  const handleOpenChange = (nextOpen: boolean) => {
+    if (disabled) return;
+    if (!nextOpen && isOpen) {
       if (searchUsedRef.current && trackPrefix) {
         track(`${trackPrefix}_searched`, { query: search });
         searchUsedRef.current = false;
       }
       setSearch('');
     }
-  }, [isOpen, search, trackPrefix]);
+    setIsOpen(nextOpen);
+  };
 
   const filteredGroups = React.useMemo(() => {
     if (!search) return groups;
@@ -101,7 +103,7 @@ export function SearchableSelect({
   const handleSelect = (optionValue: string) => {
     if (disabled) return;
     onValueChange(optionValue);
-    setIsOpen(false);
+    handleOpenChange(false);
   };
   const focusOption = (index: number) => {
     const options = listboxRef.current?.querySelectorAll<HTMLElement>('[role="option"]');
@@ -132,7 +134,7 @@ export function SearchableSelect({
   };
 
   return (
-    <Popover open={isOpen} onOpenChange={(open) => !disabled && setIsOpen(open)}>
+    <Popover open={isOpen} onOpenChange={handleOpenChange}>
       <div className="relative">
         <PopoverTrigger asChild>
           <button
diff --git a/packages/app/src/components/ui/share-button.tsx b/packages/app/src/components/ui/share-button.tsx
index 867c383c1..5ca4b1560 100644
--- a/packages/app/src/components/ui/share-button.tsx
+++ b/packages/app/src/components/ui/share-button.tsx
@@ -6,11 +6,12 @@ import { useCallback, useEffect, useRef, useState } from 'react';
 import { ShareLinkedInButton, ShareTwitterButton } from '@/components/share-buttons';
 import { track } from '@/lib/analytics';
 import { buildShareUrl } from '@/lib/url-state';
+import { cn } from '@/lib/utils';
 
 import { Button } from './button';
 import { Popover, PopoverContent, PopoverTrigger } from './popover';
 
-export function ShareButton() {
+export function ShareButton({ className }: { className?: string } = {}) {
   const [open, setOpen] = useState(false);
   const [copied, setCopied] = useState(false);
   const [url, setUrl] = useState('');
@@ -55,7 +56,10 @@ export function ShareButton() {
         <Button
           data-testid="share-button"
           size="sm"
-          className="h-8 gap-1.5 bg-brand text-primary-foreground hover:bg-brand/90 text-xs font-medium"
+          className={cn(
+            'h-8 gap-1.5 bg-brand text-primary-foreground hover:bg-brand/90 text-xs font-medium',
+            className,
+          )}
           title="Share this view"
         >
           <Share2 className="size-3.5" />
diff --git a/packages/app/src/components/unofficial-run-provider.query.test.tsx b/packages/app/src/components/unofficial-run-provider.query.test.tsx
new file mode 100644
index 000000000..b00e89fd9
--- /dev/null
+++ b/packages/app/src/components/unofficial-run-provider.query.test.tsx
@@ -0,0 +1,185 @@
+// @vitest-environment jsdom
+import { QueryClient, QueryClientProvider } from '@tanstack/react-query';
+import { act, createElement, useState } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+import { CLIENT_SEARCH_CHANGE_EVENT } from '@/lib/client-navigation';
+
+import {
+  UnofficialRunProvider,
+  useUnofficialRun,
+  type UnofficialRunContextType,
+} from './unofficial-run-provider';
+
+vi.mock('@/components/ui/unofficial-banner', () => ({ UnofficialBanner: () => null }));
+
+interface PendingRequest {
+  url: string;
+  signal: AbortSignal;
+  resolve: (response: Response) => void;
+}
+
+const requests: PendingRequest[] = [];
+
+function runInfo(id: number) {
+  return {
+    id,
+    name: `run ${id}`,
+    branch: `branch-${id}`,
+    sha: String(id),
+    createdAt: '2026-08-20T00:00:00Z',
+    url: `https://github.com/o/r/actions/runs/${id}`,
+    conclusion: 'success',
+    status: 'completed',
+    isNonMainBranch: true,
+  };
+}
+
+let latestRunIds: number[] = [];
+let dismissLatest: ((runId: string) => void) | null = null;
+let clearLatest: (() => void) | null = null;
+let latestContextValue: UnofficialRunContextType | null = null;
+let rerenderProvider: (() => void) | null = null;
+
+function Probe() {
+  const value = useUnofficialRun();
+  latestContextValue = value;
+  latestRunIds = value.unofficialRunInfos.map((run) => run.id);
+  dismissLatest = value.dismissRun;
+  clearLatest = value.clearUnofficialRun;
+  return null;
+}
+
+function ProviderHarness() {
+  const [, setRevision] = useState(0);
+  rerenderProvider = () => setRevision((revision) => revision + 1);
+  return createElement(UnofficialRunProvider, null, createElement(Probe));
+}
+
+function installPendingFetch(): void {
+  vi.spyOn(globalThis, 'fetch').mockImplementation((input, init) => {
+    const { promise, resolve } = Promise.withResolvers<Response>();
+    requests.push({
+      url: String(input),
+      signal: init?.signal as AbortSignal,
+      resolve,
+    });
+    return promise;
+  });
+}
+
+function mountProvider(): Root {
+  const client = new QueryClient({
+    defaultOptions: { queries: { retry: false } },
+  });
+  const container = document.createElement('div');
+  const root = createRoot(container);
+  act(() => {
+    root.render(createElement(QueryClientProvider, { client }, createElement(ProviderHarness)));
+  });
+  return root;
+}
+
+async function flush(): Promise<void> {
+  await act(async () => {
+    const { promise, resolve } = Promise.withResolvers<void>();
+    window.setTimeout(resolve, 0);
+    await promise;
+  });
+}
+
+function hasRunIds(expected: number[]): boolean {
+  return (
+    latestRunIds.length === expected.length &&
+    expected.every((runId, index) => latestRunIds[index] === runId)
+  );
+}
+
+async function waitForRunIds(expected: number[]): Promise<void> {
+  for (let attempt = 0; attempt < 10; attempt++) {
+    if (hasRunIds(expected)) return;
+    await flush();
+  }
+  expect(latestRunIds).toEqual(expected);
+}
+
+afterEach(() => {
+  requests.length = 0;
+  latestRunIds = [];
+  clearLatest = null;
+  dismissLatest = null;
+  latestContextValue = null;
+  rerenderProvider = null;
+  window.history.replaceState({}, '', '/inference');
+  vi.restoreAllMocks();
+});
+
+describe('UnofficialRunProvider query lifecycle', () => {
+  it('preserves the context value identity when provider fields are unchanged', () => {
+    const root = mountProvider();
+    const initialValue = latestContextValue;
+
+    act(() => rerenderProvider?.());
+
+    expect(latestContextValue).toBe(initialValue);
+    act(() => root.unmount());
+  });
+
+  it('aborts the old key and ignores its response after a URL race', async () => {
+    installPendingFetch();
+    window.history.replaceState({}, '', '/inference?unofficialrun=1');
+    const root = mountProvider();
+    await flush();
+    expect(requests[0].url).toContain('runId=1');
+
+    act(() => {
+      window.history.pushState({}, '', '/inference?unofficialruns=2');
+      window.dispatchEvent(new CustomEvent(CLIENT_SEARCH_CHANGE_EVENT));
+    });
+    await flush();
+    expect(requests[0].signal.aborted).toBe(true);
+    expect(requests[1].url).toContain('runId=2');
+
+    requests[1].resolve({
+      ok: true,
+      json: () => Promise.resolve({ runInfos: [runInfo(2)], benchmarks: [], evaluations: [] }),
+    } as Response);
+    await flush();
+    requests[0].resolve({
+      ok: true,
+      json: () => Promise.resolve({ runInfos: [runInfo(1)], benchmarks: [], evaluations: [] }),
+    } as Response);
+    await flush();
+    expect(latestRunIds).toEqual([2]);
+    act(() => root.unmount());
+  });
+
+  it('dismisses one run with pushState while retaining the other run data', async () => {
+    installPendingFetch();
+    window.history.replaceState({}, '', '/inference?unofficialruns=1,2');
+    const root = mountProvider();
+    await flush();
+    requests[0].resolve({
+      ok: true,
+      json: () =>
+        Promise.resolve({
+          runInfos: [runInfo(1), runInfo(2)],
+          benchmarks: [],
+          evaluations: [],
+        }),
+    } as Response);
+    await waitForRunIds([1, 2]);
+
+    act(() => dismissLatest?.('1'));
+    await flush();
+    expect(window.location.search).toBe('?unofficialruns=2');
+    expect(latestRunIds).toEqual([2]);
+
+    act(() => clearLatest?.());
+    await flush();
+    expect(window.location.search).toBe('');
+    expect(latestRunIds).toEqual([]);
+    act(() => root.unmount());
+  });
+});
diff --git a/packages/app/src/components/unofficial-run-provider.test.ts b/packages/app/src/components/unofficial-run-provider.test.ts
index 3c24d32bd..30465fcfe 100644
--- a/packages/app/src/components/unofficial-run-provider.test.ts
+++ b/packages/app/src/components/unofficial-run-provider.test.ts
@@ -2,8 +2,15 @@ import { describe, expect, it } from 'vitest';
 
 import type { BenchmarkRow } from '@/lib/api';
 import { Model, Sequence } from '@/lib/data-mappings';
+import { normalizeArtifactRows } from '@/app/api/unofficial-run/route';
 
-import { buildChartData, parseAvailableModelsAndSequences } from './unofficial-run-provider';
+import {
+  buildChartData,
+  overlaySelectionReducer,
+  parseAvailableModelsAndSequences,
+  parseUnofficialRunIds,
+  type OverlaySelectionState,
+} from './unofficial-run-provider';
 
 // ---------------------------------------------------------------------------
 // Helpers
@@ -43,6 +50,31 @@ function stubRow(overrides: Partial<BenchmarkRow> = {}): BenchmarkRow {
   };
 }
 
+function rawPowerArtifact(overrides: Record<string, unknown> = {}): Record<string, unknown> {
+  return {
+    infmax_model_prefix: 'dsr1',
+    hw: 'h200-nv',
+    framework: 'sglang',
+    precision: 'fp8',
+    isl: 1024,
+    osl: 1024,
+    conc: 128,
+    disagg: true,
+    prefill_tp: 8,
+    prefill_ep: 1,
+    prefill_num_workers: 1,
+    decode_tp: 8,
+    decode_ep: 1,
+    decode_num_workers: 1,
+    num_prefill_gpu: 8,
+    num_decode_gpu: 8,
+    median_e2el: 1.4,
+    median_intvty: 48,
+    tput_per_gpu: 100.5,
+    ...overrides,
+  };
+}
+
 // ---------------------------------------------------------------------------
 // parseAvailableModelsAndSequences
 // ---------------------------------------------------------------------------
@@ -276,3 +308,157 @@ describe('buildChartData', () => {
     });
   });
 });
+
+describe('schema-v2 measured-power overlay data flow', () => {
+  it('normalizes an overlay artifact and exposes J/query, Wh/query, and percent TDP', () => {
+    const rows = normalizeArtifactRows(
+      [
+        rawPowerArtifact({
+          power_valid: '1',
+          power_metric_schema_version: '2',
+          avg_power_w: 560,
+          joules_per_successful_query: 1800,
+        }),
+      ],
+      '2026-08-12',
+    );
+    const point = buildChartData(rows)['DeepSeek-R1-0528_1k/1k'].interactivity.data[0];
+
+    expect(rows[0].metrics.power_valid).toBe(1);
+    expect(rows[0].metrics.power_metric_schema_version).toBe(2);
+    expect(point.measuredJPerSuccessfulQuery?.y).toBe(1800);
+    expect(point.measuredWhPerSuccessfulQuery?.y).toBe(0.5);
+    expect(point.measuredPowerPercentTdp?.y).toBe(80);
+  });
+
+  it.each([
+    {
+      name: 'malformed boolean verdict and junk schema',
+      contract: { power_valid: true, power_metric_schema_version: '2garbage' },
+    },
+    {
+      name: 'malformed string verdict',
+      contract: { power_valid: 'garbage', power_metric_schema_version: 2 },
+    },
+    {
+      name: 'explicit invalid verdict',
+      contract: { power_valid: 0, power_metric_schema_version: 2 },
+    },
+  ])('withholds measured values for $name', ({ contract }) => {
+    const rows = normalizeArtifactRows(
+      [
+        rawPowerArtifact({
+          ...contract,
+          avg_power_w: 560,
+          joules_per_successful_query: 1800,
+          avg_temp_c: 68.4,
+          workers: [{ role: 'agg', worker_idx: 0, num_gpus: 8, avg_power_w: 560 }],
+        }),
+      ],
+      '2026-08-12',
+    );
+    const point = buildChartData(rows)['DeepSeek-R1-0528_1k/1k'].interactivity.data[0];
+
+    expect(rows[0].metrics.power_valid).toBe(0);
+    expect(point.avg_power_w).toBeUndefined();
+    expect(point.joules_per_successful_query).toBeUndefined();
+    expect(point.avg_temp_c).toBeUndefined();
+    expect(point.workers).toBeUndefined();
+    expect(point.measuredJPerSuccessfulQuery).toBeUndefined();
+    expect(point.measuredWhPerSuccessfulQuery).toBeUndefined();
+    expect(point.measuredPowerPercentTdp).toBeUndefined();
+  });
+});
+
+const initialSelection = (): OverlaySelectionState => ({
+  availabilityKey: '1,2',
+  activeOverlayHwTypes: new Set(['b200', 'h100']),
+  availableOverlayHwTypes: new Set(['b200', 'h100']),
+  localOfficialOverride: null,
+  scopeKey: null,
+  scopeOverlayHwTypes: new Set(),
+  scopeReady: false,
+  bestSelectionKey: '',
+  bestPerSku: false,
+});
+
+describe('unofficial URL and overlay selection state', () => {
+  it('canonicalizes whitespace, leading zeroes, and duplicate run IDs', () => {
+    expect(parseUnofficialRunIds('?UnofficialRuns=0022,%2011,22')).toEqual(['22', '11']);
+  });
+
+  it('seeds a mixed official/overlay scope only when official readiness is known', () => {
+    const pending = overlaySelectionReducer(initialSelection(), {
+      type: 'scope',
+      input: {
+        scopeKey: 'model|sequence|1,2',
+        officialHwTypes: new Set(['h200']),
+        overlayHwTypes: new Set(['b200']),
+        bestOfficialHwTypes: new Set(['h200']),
+        bestOverlayHwTypes: new Set(['b200']),
+        bestPerSku: false,
+        ready: false,
+      },
+    });
+    expect(pending.localOfficialOverride).toBeNull();
+    expect(pending.activeOverlayHwTypes).toEqual(new Set(['b200', 'h100']));
+
+    const ready = overlaySelectionReducer(pending, {
+      type: 'scope',
+      input: {
+        scopeKey: 'model|sequence|1,2',
+        officialHwTypes: new Set(['h200']),
+        overlayHwTypes: new Set(['b200']),
+        bestOfficialHwTypes: new Set(['h200']),
+        bestOverlayHwTypes: new Set(['b200']),
+        bestPerSku: false,
+        ready: true,
+      },
+    });
+    expect(ready.localOfficialOverride).toEqual(new Set(['h200']));
+  });
+
+  it('reconciles best-per-SKU in the reducer and restores the full scope when disabled', () => {
+    const scope = {
+      scopeKey: 'model|sequence|1,2',
+      officialHwTypes: new Set(['h100', 'h200']),
+      overlayHwTypes: new Set(['b100', 'b200']),
+      bestOfficialHwTypes: new Set(['h200']),
+      bestOverlayHwTypes: new Set(['b200']),
+      bestPerSku: true,
+      ready: true,
+    };
+    const best = overlaySelectionReducer(initialSelection(), { type: 'scope', input: scope });
+    expect(best.localOfficialOverride).toEqual(new Set(['h200']));
+    expect(best.activeOverlayHwTypes).toEqual(new Set(['h100', 'b200']));
+
+    const restored = overlaySelectionReducer(best, {
+      type: 'scope',
+      input: { ...scope, bestPerSku: false },
+    });
+    expect(restored.localOfficialOverride).toEqual(new Set(['h100', 'h200']));
+    expect(restored.activeOverlayHwTypes).toEqual(new Set(['h100', 'b100', 'b200']));
+  });
+
+  it('resets dismissed-run selection exactly once when the canonical query key changes', () => {
+    const selected = overlaySelectionReducer(initialSelection(), {
+      type: 'selection',
+      official: new Set(['h200']),
+      overlay: new Set(['b200']),
+    });
+    const dismissed = overlaySelectionReducer(selected, {
+      type: 'availability',
+      key: '1',
+      allOverlayHwTypes: new Set(['b100']),
+    });
+    expect(dismissed.activeOverlayHwTypes).toEqual(new Set(['b100']));
+    expect(dismissed.localOfficialOverride).toBeNull();
+    expect(
+      overlaySelectionReducer(dismissed, {
+        type: 'availability',
+        key: '1',
+        allOverlayHwTypes: new Set(['b100']),
+      }),
+    ).toBe(dismissed);
+  });
+});
diff --git a/packages/app/src/components/unofficial-run-provider.tsx b/packages/app/src/components/unofficial-run-provider.tsx
index 2ee7138c8..d699ba457 100644
--- a/packages/app/src/components/unofficial-run-provider.tsx
+++ b/packages/app/src/components/unofficial-run-provider.tsx
@@ -1,5 +1,6 @@
 'use client';
 
+import { useQuery, useQueryClient } from '@tanstack/react-query';
 import {
   type ReactNode,
   createContext,
@@ -7,20 +8,21 @@ import {
   useContext,
   useEffect,
   useMemo,
-  useState,
+  useReducer,
 } from 'react';
 
 import type { ChartDefinition, HardwareConfig, InferenceData } from '@/components/inference/types';
 import { UnofficialBanner } from '@/components/ui/unofficial-banner';
 import { DB_MODEL_TO_DISPLAY, rowToSequence } from '@semianalysisai/inferencex-constants';
-import { computeToggle } from '@/hooks/useTogglableSet';
 import type { BenchmarkRow, EvalRow } from '@/lib/api';
-import { normalizeEvalHardwareKey } from '@/lib/chart-utils';
-
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
 import { transformBenchmarkRows } from '@/lib/benchmark-transform';
+import { normalizeEvalHardwareKey } from '@/lib/chart-utils';
+import { notifyClientSearchChange } from '@/lib/client-navigation';
+import { useClientSearch } from '@/hooks/useClientSearch';
 import { Model, Sequence } from '@/lib/data-mappings';
 
+import chartDefinitions from '@/components/inference/metric-registry';
+
 interface UnofficialRunInfo {
   id: number;
   name: string;
@@ -41,7 +43,22 @@ type UnofficialChartData = Record<
   }
 >;
 
+interface UnofficialRunResponse {
+  runInfos: UnofficialRunInfo[];
+  benchmarks: BenchmarkRow[];
+  evaluations: EvalRow[];
+}
+
+const EMPTY_UNOFFICIAL_RUN_INFOS: UnofficialRunInfo[] = [];
+
 const UNOFFICIAL_RUN_PARAM_RE = /^unofficialruns?$/iu;
+function deleteUnofficialRunParams(searchParams: URLSearchParams): void {
+  const keysToDelete: string[] = [];
+  for (const key of searchParams.keys()) {
+    if (UNOFFICIAL_RUN_PARAM_RE.test(key)) keysToDelete.push(key);
+  }
+  for (const key of keysToDelete) searchParams.delete(key);
+}
 
 export interface AvailableModelSequence {
   model: Model;
@@ -49,37 +66,27 @@ export interface AvailableModelSequence {
   precisions: string[];
 }
 
+export interface OverlayScopeInput {
+  scopeKey: string;
+  officialHwTypes: Set<string>;
+  overlayHwTypes: Set<string>;
+  bestOfficialHwTypes: Set<string>;
+  bestOverlayHwTypes: Set<string>;
+  bestPerSku: boolean;
+  ready: boolean;
+}
+
 export interface UnofficialRunContextType {
   isUnofficialRun: boolean;
-  /** First run in the loaded set — kept as a convenience alias for overlay labels. */
   unofficialRunInfo: UnofficialRunInfo | null;
-  /** All runs loaded from the `unofficialrun(s)` URL param (comma-separated). */
   unofficialRunInfos: UnofficialRunInfo[];
-  /**
-   * Position of each run in the loaded set, keyed by both `run.url` and the
-   * numeric id as a string. Used to derive a distinct hue shift per run for
-   * overlay points so multiple runs are visually separable.
-   */
   runIndexByUrl: Record<string, number>;
   unofficialChartData: UnofficialChartData | null;
-  /**
-   * Raw benchmark rows as returned by the unofficial-run API, before any chart
-   * transform. Kept alongside `unofficialChartData` so consumers that build
-   * their own derived shapes from raw rows — today the TCO calculator's
-   * `useThroughputData` — can run overlay rows through the exact same mapping
-   * as official DB rows instead of re-deriving them from chart data.
-   */
   unofficialBenchmarkRows: BenchmarkRow[] | null;
   unofficialEvalRows: EvalRow[] | null;
   loading: boolean;
   error: string | null;
-  /** Clear every unofficial run. Wipes state + URL. */
   clearUnofficialRun: () => void;
-  /**
-   * Drop a single run ID. Rewrites the URL to the remaining IDs and filters
-   * local state (chart data + eval rows + run infos) by `run_url` without
-   * refetching the others.
-   */
   dismissRun: (runId: string) => void;
   availableModelsAndSequences: AvailableModelSequence[];
   getOverlayData: (
@@ -90,17 +97,14 @@ export interface UnofficialRunContextType {
     data: InferenceData[];
     hardwareConfig: HardwareConfig;
   } | null;
-  // Shared overlay toggle state — both charts read/write the same sets
   activeOverlayHwTypes: Set<string>;
-  setActiveOverlayHwTypes: (v: Set<string>) => void;
   allOverlayHwTypes: Set<string>;
-  toggleOverlayHwType: (key: string) => void;
-  resetOverlayHwTypes: () => void;
   localOfficialOverride: Set<string> | null;
-  setLocalOfficialOverride: (v: Set<string> | null) => void;
+  reconcileOverlayScope: (scope: OverlayScopeInput) => void;
+  setUnifiedOverlaySelection: (official: Set<string>, overlay: Set<string>) => void;
+  resetOverlaySelection: () => void;
 }
 
-/** @internal Exported for test provider wrapping only. */
 export const UnofficialRunContext = createContext<UnofficialRunContextType | undefined>(undefined);
 
 export function useUnofficialRun() {
@@ -110,11 +114,15 @@ export function useUnofficialRun() {
   }
   return context;
 }
+/** Registers derived chart availability; the provider reducer owns reconciliation. */
+export function useOverlayScopeReconciliation(input: OverlayScopeInput | null): void {
+  const { reconcileOverlayScope } = useUnofficialRun();
+  useEffect(() => {
+    if (input) reconcileOverlayScope(input);
+  }, [input, reconcileOverlayScope]);
+}
 
-/** Build chart data from raw benchmark rows returned by the unofficial-run API. */
 export function buildChartData(benchmarks: BenchmarkRow[]): UnofficialChartData {
-  // Group benchmarks by display model name + Sequence enum value
-  // (keys must match getOverlayData which looks up `${Model}_${Sequence}`)
   const groups = new Map<string, BenchmarkRow[]>();
   for (const row of benchmarks) {
     const displayModel = DB_MODEL_TO_DISPLAY[row.model] ?? row.model;
@@ -128,7 +136,6 @@ export function buildChartData(benchmarks: BenchmarkRow[]): UnofficialChartData
   const result: UnofficialChartData = {};
   for (const [key, rows] of groups) {
     const { chartData, hardwareConfig } = transformBenchmarkRows(rows);
-    // chartData indices match chartDefinitions order — look up by chartType
     const e2eIdx = (chartDefinitions as ChartDefinition[]).findIndex((d) => d.chartType === 'e2e');
     const interactivityIdx = (chartDefinitions as ChartDefinition[]).findIndex(
       (d) => d.chartType === 'interactivity',
@@ -138,7 +145,6 @@ export function buildChartData(benchmarks: BenchmarkRow[]): UnofficialChartData
       interactivity: { data: chartData[interactivityIdx] ?? [], gpus: hardwareConfig },
     };
   }
-
   return result;
 }
 
@@ -150,210 +156,305 @@ export function parseAvailableModelsAndSequences(
   const result: AvailableModelSequence[] = [];
   const allModels = Object.values(Model);
   const allSequences = Object.values(Sequence);
-
   for (const key of Object.keys(chartData)) {
     const lastUnderscoreIndex = key.lastIndexOf('_');
     if (lastUnderscoreIndex === -1) continue;
-    const modelPart = key.slice(0, lastUnderscoreIndex);
-    const sequencePart = key.slice(lastUnderscoreIndex + 1);
-    const model = allModels.find((m) => m === modelPart);
-    const sequence = allSequences.find((s) => s === sequencePart);
+    const model = allModels.find((candidate) => candidate === key.slice(0, lastUnderscoreIndex));
+    const sequence = allSequences.find(
+      (candidate) => candidate === key.slice(lastUnderscoreIndex + 1),
+    );
     if (!model || !sequence) continue;
     const group = chartData[key];
     const precisions = [
       ...new Set(
-        [...(group?.e2e.data ?? []), ...(group?.interactivity.data ?? [])].map((d) => d.precision),
+        [...(group?.e2e.data ?? []), ...(group?.interactivity.data ?? [])].map(
+          (point) => point.precision,
+        ),
       ),
     ];
-    if (!result.some((r) => r.model === model && r.sequence === sequence)) {
+    if (!result.some((entry) => entry.model === model && entry.sequence === sequence)) {
       result.push({ model, sequence, precisions });
     }
   }
-
   return result;
 }
 
+/** Canonicalize the first unofficial-run URL value without changing server validation policy. */
+export function parseUnofficialRunIds(search: string): string[] {
+  const params = new URLSearchParams(search);
+  for (const [key, value] of params) {
+    if (!UNOFFICIAL_RUN_PARAM_RE.test(key) || !value) continue;
+    return [
+      ...new Set(
+        value
+          .split(',')
+          .map((id) => id.trim())
+          .filter(Boolean)
+          .map((id) => (/^\d+$/u.test(id) ? id.replace(/^0+(?=\d)/u, '') : id)),
+      ),
+    ];
+  }
+  return [];
+}
+
+export interface OverlaySelectionState {
+  availabilityKey: string;
+  activeOverlayHwTypes: Set<string>;
+  availableOverlayHwTypes: Set<string>;
+  localOfficialOverride: Set<string> | null;
+  scopeKey: string | null;
+  scopeOverlayHwTypes: Set<string>;
+  scopeReady: boolean;
+  bestSelectionKey: string;
+  bestPerSku: boolean;
+}
+
+export type OverlaySelectionAction =
+  | { type: 'availability'; key: string; allOverlayHwTypes: Set<string> }
+  | { type: 'scope'; input: OverlayScopeInput }
+  | { type: 'selection'; official: Set<string>; overlay: Set<string> }
+  | { type: 'reset' };
+
+const EMPTY_SELECTION_STATE: OverlaySelectionState = {
+  availabilityKey: '',
+  activeOverlayHwTypes: new Set(),
+  availableOverlayHwTypes: new Set(),
+  localOfficialOverride: null,
+  scopeKey: null,
+  scopeOverlayHwTypes: new Set(),
+  scopeReady: false,
+  bestSelectionKey: '',
+  bestPerSku: false,
+};
+
+function sortedSetKey(values: Set<string>): string {
+  return [...values].toSorted().join(',');
+}
+
+export function overlaySelectionReducer(
+  state: OverlaySelectionState,
+  action: OverlaySelectionAction,
+): OverlaySelectionState {
+  if (action.type === 'availability') {
+    if (action.key !== state.availabilityKey) {
+      return {
+        ...EMPTY_SELECTION_STATE,
+        availabilityKey: action.key,
+        availableOverlayHwTypes: new Set(action.allOverlayHwTypes),
+        activeOverlayHwTypes: new Set(action.allOverlayHwTypes),
+      };
+    }
+    if (sortedSetKey(action.allOverlayHwTypes) === sortedSetKey(state.availableOverlayHwTypes)) {
+      return state;
+    }
+    const activeOverlayHwTypes = new Set(state.activeOverlayHwTypes);
+    for (const key of state.availableOverlayHwTypes) {
+      if (!action.allOverlayHwTypes.has(key)) activeOverlayHwTypes.delete(key);
+    }
+    for (const key of action.allOverlayHwTypes) {
+      if (!state.availableOverlayHwTypes.has(key)) activeOverlayHwTypes.add(key);
+    }
+    return {
+      ...state,
+      availableOverlayHwTypes: new Set(action.allOverlayHwTypes),
+      activeOverlayHwTypes,
+    };
+  }
+  if (action.type === 'selection') {
+    return {
+      ...state,
+      bestPerSku: false,
+      bestSelectionKey: '',
+      localOfficialOverride: new Set(action.official),
+      activeOverlayHwTypes: new Set(action.overlay),
+    };
+  }
+  if (action.type === 'reset') {
+    return {
+      ...state,
+      activeOverlayHwTypes: new Set(state.availableOverlayHwTypes),
+      localOfficialOverride: null,
+      bestPerSku: false,
+      bestSelectionKey: '',
+    };
+  }
+
+  const { input } = action;
+  const desiredOfficial = input.bestPerSku
+    ? input.bestOfficialHwTypes.size > 0
+      ? input.bestOfficialHwTypes
+      : input.officialHwTypes
+    : input.officialHwTypes;
+  const desiredOverlay = input.bestPerSku
+    ? input.bestOverlayHwTypes.size > 0
+      ? input.bestOverlayHwTypes
+      : input.overlayHwTypes
+    : input.overlayHwTypes;
+  const bestSelectionKey = input.bestPerSku
+    ? `${sortedSetKey(desiredOfficial)}|${sortedSetKey(desiredOverlay)}`
+    : '';
+  const scopeChanged = state.scopeKey !== input.scopeKey;
+  const bestSelectionChanged =
+    state.bestPerSku !== input.bestPerSku ||
+    (input.bestPerSku && state.bestSelectionKey !== bestSelectionKey);
+  const becameReady = input.ready && !state.scopeReady;
+  if (!scopeChanged && !bestSelectionChanged && !becameReady) return state;
+
+  let activeOverlayHwTypes = state.activeOverlayHwTypes;
+  if (scopeChanged || bestSelectionChanged) {
+    activeOverlayHwTypes = new Set(state.activeOverlayHwTypes);
+    state.scopeOverlayHwTypes.forEach((key) => activeOverlayHwTypes.delete(key));
+    input.overlayHwTypes.forEach((key) => activeOverlayHwTypes.delete(key));
+    desiredOverlay.forEach((key) => activeOverlayHwTypes.add(key));
+  }
+  return {
+    ...state,
+    activeOverlayHwTypes,
+    localOfficialOverride:
+      input.ready && (scopeChanged || bestSelectionChanged || becameReady)
+        ? new Set(desiredOfficial)
+        : scopeChanged
+          ? null
+          : state.localOfficialOverride,
+    scopeKey: input.scopeKey,
+    scopeOverlayHwTypes: new Set(input.overlayHwTypes),
+    scopeReady: input.ready,
+    bestSelectionKey,
+    bestPerSku: input.bestPerSku,
+  };
+}
+
+function selectUnofficialRunResponse(
+  response: UnofficialRunResponse,
+  runIds: readonly string[],
+): UnofficialRunResponse {
+  const selectedIds = new Set(runIds);
+  const runInfos = response.runInfos.filter((info) => selectedIds.has(String(info.id)));
+  const selectedUrls = new Set(runInfos.map((info) => info.url).filter(Boolean));
+  const belongsToSelectedRun = (runUrl?: string | null) => {
+    if (!runUrl) return runIds.length === 1;
+    if (selectedUrls.has(runUrl)) return true;
+    const runId = runUrl.match(/\/runs\/(?<runId>\d+)/u)?.groups?.runId;
+    return Boolean(runId && selectedIds.has(runId));
+  };
+  return {
+    runInfos,
+    benchmarks: response.benchmarks.filter((row) => belongsToSelectedRun(row.run_url)),
+    evaluations: response.evaluations.filter((row) => belongsToSelectedRun(row.run_url)),
+  };
+}
+
+async function fetchUnofficialRuns(
+  ids: string[],
+  signal: AbortSignal,
+): Promise<UnofficialRunResponse> {
+  const response = await fetch(`/api/unofficial-run?runId=${encodeURIComponent(ids.join(','))}`, {
+    cache: 'no-store',
+    signal,
+  });
+  const data = (await response.json()) as Partial<UnofficialRunResponse> & { error?: string };
+  if (!response.ok) throw new Error(data.error || 'Failed to fetch unofficial run');
+  return selectUnofficialRunResponse(
+    {
+      runInfos: Array.isArray(data.runInfos) ? data.runInfos : [],
+      benchmarks: Array.isArray(data.benchmarks) ? data.benchmarks : [],
+      evaluations: Array.isArray(data.evaluations) ? data.evaluations : [],
+    },
+    ids,
+  );
+}
+
 export function UnofficialRunProvider({ children }: { children: ReactNode }) {
-  const [unofficialRunInfos, setUnofficialRunInfos] = useState<UnofficialRunInfo[]>([]);
+  const queryClient = useQueryClient();
+  const search = useClientSearch();
+  const runIds = useMemo(() => parseUnofficialRunIds(search), [search]);
+  const runIdsKey = runIds.join(',');
+  const query = useQuery({
+    queryKey: ['unofficial-runs', runIdsKey] as const,
+    queryFn: ({ signal }) => fetchUnofficialRuns(runIds, signal),
+    enabled: runIds.length > 0,
+    staleTime: 0,
+    gcTime: 0,
+    retry: false,
+    refetchOnMount: 'always',
+  });
+
+  const unofficialRunInfos = query.data?.runInfos ?? EMPTY_UNOFFICIAL_RUN_INFOS;
   const unofficialRunInfo = unofficialRunInfos[0] ?? null;
-  const [unofficialChartData, setUnofficialChartData] = useState<UnofficialChartData | null>(null);
-  const [unofficialBenchmarkRows, setUnofficialBenchmarkRows] = useState<BenchmarkRow[] | null>(
-    null,
+  const unofficialBenchmarkRows = query.data?.benchmarks ?? null;
+  const unofficialEvalRows = query.data?.evaluations ?? null;
+  const unofficialChartData = useMemo(
+    () => (unofficialBenchmarkRows ? buildChartData(unofficialBenchmarkRows) : null),
+    [unofficialBenchmarkRows],
+  );
+  const availableModelsAndSequences = useMemo(
+    () => parseAvailableModelsAndSequences(unofficialChartData),
+    [unofficialChartData],
   );
-  const [unofficialEvalRows, setUnofficialEvalRows] = useState<EvalRow[] | null>(null);
-  const [loading, setLoading] = useState(false);
-  const [error, setError] = useState<string | null>(null);
-  const [availableModelsAndSequences, setAvailableModelsAndSequences] = useState<
-    AvailableModelSequence[]
-  >([]);
-
-  // --- Shared overlay toggle state (unified across both charts) ---
-  const [activeOverlayHwTypes, setActiveOverlayHwTypes] = useState<Set<string>>(new Set());
-  const [localOfficialOverride, setLocalOfficialOverrideRaw] = useState<Set<string> | null>(null);
-
-  // Derive all overlay hw types from chart data
   const allOverlayHwTypes = useMemo(() => {
     const hwTypes = new Set<string>();
-    if (unofficialChartData) {
-      for (const group of Object.values(unofficialChartData)) {
-        for (const chartType of [group.e2e, group.interactivity]) {
-          chartType.data.forEach((p) => {
-            if (p.hwKey) hwTypes.add(p.hwKey as string);
-          });
-        }
+    for (const group of Object.values(unofficialChartData ?? {})) {
+      for (const chartType of [group.e2e, group.interactivity]) {
+        chartType.data.forEach((point) => {
+          if (point.hwKey) hwTypes.add(String(point.hwKey));
+        });
       }
     }
-    if (unofficialEvalRows) {
-      unofficialEvalRows.forEach((row) => {
-        const hwKey = normalizeEvalHardwareKey(row.hardware, row.framework, row.spec_method);
-        if (hwKey !== 'unknown') hwTypes.add(hwKey);
-      });
+    for (const row of unofficialEvalRows ?? []) {
+      const hwKey = normalizeEvalHardwareKey(row.hardware, row.framework, row.spec_method);
+      if (hwKey !== 'unknown') hwTypes.add(hwKey);
     }
     return hwTypes;
   }, [unofficialChartData, unofficialEvalRows]);
 
-  // Reset overlay state when chart data changes
+  const [selection, dispatchSelection] = useReducer(overlaySelectionReducer, EMPTY_SELECTION_STATE);
   useEffect(() => {
-    setActiveOverlayHwTypes(allOverlayHwTypes);
-    setLocalOfficialOverrideRaw(null);
-  }, [allOverlayHwTypes]);
-
-  const toggleOverlayHwType = useCallback(
-    (key: string) => {
-      setActiveOverlayHwTypes((prev) => computeToggle(prev, key, allOverlayHwTypes));
-    },
-    [allOverlayHwTypes],
-  );
-
-  const resetOverlayHwTypes = useCallback(() => {
-    setActiveOverlayHwTypes(allOverlayHwTypes);
-  }, [allOverlayHwTypes]);
-
-  const setLocalOfficialOverride = useCallback(
-    (v: Set<string> | null) => setLocalOfficialOverrideRaw(v),
-    [],
-  );
-
-  const setActiveOverlayHwTypesStable = useCallback(
-    (v: Set<string>) => setActiveOverlayHwTypes(v),
-    [],
-  );
+    dispatchSelection({
+      type: 'availability',
+      key: runIdsKey,
+      allOverlayHwTypes,
+    });
+  }, [runIdsKey, allOverlayHwTypes]);
 
   const clearUnofficialRun = useCallback(() => {
-    setUnofficialRunInfos([]);
-    setUnofficialChartData(null);
-    setUnofficialBenchmarkRows(null);
-    setUnofficialEvalRows(null);
-    setError(null);
-    setAvailableModelsAndSequences([]);
     const url = new URL(window.location.href);
-    for (const key of url.searchParams.keys()) {
-      if (UNOFFICIAL_RUN_PARAM_RE.test(key)) url.searchParams.delete(key);
-    }
+    deleteUnofficialRunParams(url.searchParams);
     window.history.pushState({}, '', url);
+    notifyClientSearchChange(url.href);
   }, []);
 
-  /**
-   * Drop a single run from the URL + state. Since benchmark rows are tagged
-   * with `run_url` and eval rows have their own `run_url`, we can filter local
-   * state by the dismissed run's URL/id without refetching the remaining runs.
-   */
   const dismissRun = useCallback(
     (runId: string) => {
-      const target = unofficialRunInfos.find((r) => String(r.id) === runId);
+      const target = unofficialRunInfos.find((run) => String(run.id) === runId);
       if (!target) return;
-
-      const remaining = unofficialRunInfos.filter((r) => String(r.id) !== runId);
-
-      // Rewrite URL to the remaining IDs (or drop param if none left).
-      const url = new URL(window.location.href);
-      const existingKeys: string[] = [];
-      for (const key of url.searchParams.keys()) {
-        if (UNOFFICIAL_RUN_PARAM_RE.test(key)) existingKeys.push(key);
-      }
-      for (const key of existingKeys) url.searchParams.delete(key);
-      if (remaining.length > 0) {
-        url.searchParams.set('unofficialruns', remaining.map((r) => r.id).join(','));
+      const remainingIds = runIds.filter((id) => id !== runId);
+      if (remainingIds.length > 0 && query.data) {
+        queryClient.setQueryData<UnofficialRunResponse>(
+          ['unofficial-runs', remainingIds.join(',')],
+          selectUnofficialRunResponse(query.data, remainingIds),
+        );
       }
+      const url = new URL(window.location.href);
+      deleteUnofficialRunParams(url.searchParams);
+      if (remainingIds.length > 0) url.searchParams.set('unofficialruns', remainingIds.join(','));
       window.history.pushState({}, '', url);
-
-      if (remaining.length === 0) {
-        setUnofficialRunInfos([]);
-        setUnofficialChartData(null);
-        setUnofficialBenchmarkRows(null);
-        setUnofficialEvalRows(null);
-        setError(null);
-        setAvailableModelsAndSequences([]);
-        return;
-      }
-
-      setUnofficialRunInfos(remaining);
-
-      // Filter chart data by stamped `run_url`. A row belongs to the dismissed
-      // run if its URL matches exactly OR the numeric id parses to the same.
-      const belongsToDismissed = (rowUrl?: string | null) => {
-        if (!rowUrl) return false;
-        if (rowUrl === target.url) return true;
-        const m = rowUrl.match(/\/runs\/(?<runId>\d+)/u);
-        return m?.groups?.runId === runId;
-      };
-
-      // Compute the filtered chart data BEFORE any setState so we can pass the
-      // same value to setUnofficialChartData and parseAvailableModelsAndSequences.
-      // Writing to an outer variable from inside a setState updater and then
-      // reading it synchronously is unsafe: React 18 invokes updaters during
-      // render, not at the call site, so the read would see the initial null.
-      const nextChartData: UnofficialChartData | null = unofficialChartData
-        ? (() => {
-            const next: UnofficialChartData = {};
-            for (const [key, group] of Object.entries(unofficialChartData)) {
-              const e2eData = group.e2e.data.filter((d) => !belongsToDismissed(d.run_url));
-              const intvData = group.interactivity.data.filter(
-                (d) => !belongsToDismissed(d.run_url),
-              );
-              if (e2eData.length === 0 && intvData.length === 0) continue;
-              next[key] = {
-                e2e: { data: e2eData, gpus: group.e2e.gpus },
-                interactivity: { data: intvData, gpus: group.interactivity.gpus },
-              };
-            }
-            return next;
-          })()
-        : null;
-      setUnofficialChartData(nextChartData);
-      // Re-derive available (model, sequence) pairs from surviving runs so the
-      // model/sequence picker doesn't still offer combos that only existed in
-      // the dismissed run.
-      setAvailableModelsAndSequences(parseAvailableModelsAndSequences(nextChartData));
-
-      setUnofficialBenchmarkRows((prev) =>
-        prev ? prev.filter((row) => !belongsToDismissed(row.run_url)) : prev,
-      );
-
-      setUnofficialEvalRows((prev) =>
-        prev ? prev.filter((row) => !belongsToDismissed(row.run_url)) : prev,
-      );
+      notifyClientSearchChange(url.href);
     },
-    [unofficialRunInfos, unofficialChartData],
+    [runIds, unofficialRunInfos, query.data, queryClient],
   );
 
-  // Build a url → index lookup. Keyed by the full run.url AND by the numeric id
-  // as a string, since `updateRepoUrl` can rewrite hosts/orgs between the
-  // overlay rendering path and the run metadata.
   const runIndexByUrl = useMemo(() => {
     const map: Record<string, number> = {};
-    unofficialRunInfos.forEach((info, idx) => {
-      if (info.url) map[info.url] = idx;
-      if (info.id !== undefined && info.id !== null) map[String(info.id)] = idx;
+    unofficialRunInfos.forEach((info, index) => {
+      if (info.url) map[info.url] = index;
+      map[String(info.id)] = index;
     });
     return map;
   }, [unofficialRunInfos]);
 
   const getOverlayData = useCallback(
     (model: Model, sequence: Sequence, chartType: 'e2e' | 'interactivity') => {
-      if (!unofficialChartData) return null;
-      const dataKey = `${model}_${sequence}`;
-      const chartGroup = unofficialChartData[dataKey];
+      const chartGroup = unofficialChartData?.[`${model}_${sequence}`];
       if (!chartGroup) return null;
       const dataForChart = chartType === 'e2e' ? chartGroup.e2e : chartGroup.interactivity;
       return { data: dataForChart.data, hardwareConfig: dataForChart.gpus };
@@ -361,85 +462,65 @@ export function UnofficialRunProvider({ children }: { children: ReactNode }) {
     [unofficialChartData],
   );
 
-  useEffect(() => {
-    const load = () => {
-      const params = new URLSearchParams(window.location.search);
-      let unofficialRunIdParam: string | undefined;
-      for (const [key, value] of params) {
-        if (UNOFFICIAL_RUN_PARAM_RE.test(key) && value) {
-          unofficialRunIdParam = value;
-          break;
-        }
-      }
-      if (!unofficialRunIdParam) {
-        setUnofficialRunInfos([]);
-        setUnofficialChartData(null);
-        setUnofficialBenchmarkRows(null);
-        setUnofficialEvalRows(null);
-        setError(null);
-        setAvailableModelsAndSequences([]);
-        return;
-      }
-
-      setLoading(true);
-      setError(null);
-
-      // Pass the raw param value through — it may be a single id or a comma-separated list.
-      // encodeURIComponent preserves commas while escaping any accidental whitespace/symbols.
-      fetch(`/api/unofficial-run?runId=${encodeURIComponent(unofficialRunIdParam)}`)
-        .then(async (response) => {
-          const data = await response.json();
-          if (!response.ok) throw new Error(data.error || 'Failed to fetch unofficial run');
-
-          setUnofficialRunInfos(Array.isArray(data.runInfos) ? data.runInfos : []);
-          const benchmarkRows: BenchmarkRow[] = data.benchmarks ?? [];
-          const chartData = buildChartData(benchmarkRows);
-          setUnofficialChartData(chartData);
-          setUnofficialBenchmarkRows(benchmarkRows);
-          setUnofficialEvalRows(data.evaluations ?? []);
-          setAvailableModelsAndSequences(parseAvailableModelsAndSequences(chartData));
-        })
-        .catch((caughtError) => {
-          setError(caughtError instanceof Error ? caughtError.message : 'Unknown error');
-          setUnofficialRunInfos([]);
-          setUnofficialChartData(null);
-          setUnofficialBenchmarkRows(null);
-          setUnofficialEvalRows(null);
-          setAvailableModelsAndSequences([]);
-        })
-        .finally(() => setLoading(false));
-    };
-
-    load();
-    window.addEventListener('popstate', load);
-    return () => window.removeEventListener('popstate', load);
+  const resetOverlaySelection = useCallback(() => {
+    dispatchSelection({ type: 'reset' });
+  }, []);
+  const reconcileOverlayScope = useCallback((input: OverlayScopeInput) => {
+    dispatchSelection({ type: 'scope', input });
+  }, []);
+  const setUnifiedOverlaySelection = useCallback((official: Set<string>, overlay: Set<string>) => {
+    dispatchSelection({ type: 'selection', official, overlay });
   }, []);
 
+  const loading = runIds.length > 0 && query.isFetching;
+  const error = query.error instanceof Error ? query.error.message : null;
+
+  const contextValue = useMemo<UnofficialRunContextType>(
+    () => ({
+      isUnofficialRun: unofficialRunInfos.length > 0,
+      unofficialRunInfo,
+      unofficialRunInfos,
+      runIndexByUrl,
+      unofficialChartData,
+      unofficialBenchmarkRows,
+      unofficialEvalRows,
+      loading,
+      error,
+      clearUnofficialRun,
+      dismissRun,
+      availableModelsAndSequences,
+      getOverlayData,
+      activeOverlayHwTypes: selection.activeOverlayHwTypes,
+      allOverlayHwTypes,
+      localOfficialOverride: selection.localOfficialOverride,
+      reconcileOverlayScope,
+      setUnifiedOverlaySelection,
+      resetOverlaySelection,
+    }),
+    [
+      unofficialRunInfos,
+      unofficialRunInfo,
+      runIndexByUrl,
+      unofficialChartData,
+      unofficialBenchmarkRows,
+      unofficialEvalRows,
+      loading,
+      error,
+      clearUnofficialRun,
+      dismissRun,
+      availableModelsAndSequences,
+      getOverlayData,
+      selection.activeOverlayHwTypes,
+      allOverlayHwTypes,
+      selection.localOfficialOverride,
+      reconcileOverlayScope,
+      setUnifiedOverlaySelection,
+      resetOverlaySelection,
+    ],
+  );
+
   return (
-    <UnofficialRunContext.Provider
-      value={{
-        isUnofficialRun: unofficialRunInfos.length > 0,
-        unofficialRunInfo,
-        unofficialRunInfos,
-        runIndexByUrl,
-        unofficialChartData,
-        unofficialBenchmarkRows,
-        unofficialEvalRows,
-        loading,
-        error,
-        clearUnofficialRun,
-        dismissRun,
-        availableModelsAndSequences,
-        getOverlayData,
-        activeOverlayHwTypes,
-        setActiveOverlayHwTypes: setActiveOverlayHwTypesStable,
-        allOverlayHwTypes,
-        toggleOverlayHwType,
-        resetOverlayHwTypes,
-        localOfficialOverride,
-        setLocalOfficialOverride,
-      }}
-    >
+    <UnofficialRunContext.Provider value={contextValue}>
       {unofficialRunInfos.length > 0 && (
         <UnofficialBanner
           runs={unofficialRunInfos}
diff --git a/packages/app/src/components/zh/zh-tab-intro.tsx b/packages/app/src/components/zh/zh-tab-intro.tsx
index 751e1d7e1..a2483bbac 100644
--- a/packages/app/src/components/zh/zh-tab-intro.tsx
+++ b/packages/app/src/components/zh/zh-tab-intro.tsx
@@ -1,18 +1,33 @@
 import { Card } from '@/components/ui/card';
-import { TAB_INTRO_ZH, TAB_META_ZH, type ZhTabKey } from '@/lib/tab-meta-zh';
+import type { DashboardRouteKey } from '@/lib/dashboard-routes';
+import { TAB_INTRO_ZH, TAB_META_ZH } from '@/lib/tab-meta-zh';
 
 /**
  * Server-rendered Chinese intro above the interactive dashboard on /zh tab
  * pages. The charts below render in English; this block gives crawlers and
  * readers genuine Chinese content describing what the page shows.
  */
-export function ZhTabIntro({ tab }: { tab: ZhTabKey }) {
+export function ZhTabIntro({
+  tab,
+  title,
+  intro,
+}: {
+  tab: DashboardRouteKey;
+  /** Model-specific overrides used by the per-model routes
+   *  (/zh/historical/<slug>); the tab defaults apply when omitted. */
+  title?: string;
+  intro?: string;
+}) {
   return (
     <Card data-testid="zh-tab-intro">
-      <h1 className="text-xl lg:text-2xl font-bold tracking-tight">{TAB_META_ZH[tab].title}</h1>
-      <p className="mt-2 text-sm lg:text-base text-muted-foreground">{TAB_INTRO_ZH[tab]}</p>
+      <h1 className="text-xl lg:text-2xl font-bold tracking-tight">
+        {title ?? TAB_META_ZH[tab].title}
+      </h1>
+      <p className="mt-2 text-sm lg:text-base text-muted-foreground">
+        {intro ?? TAB_INTRO_ZH[tab]}
+      </p>
       <p className="mt-2 text-xs text-muted-foreground">
-        图表中的模型、Chip、框架与指标名称均沿用业界通用英文名称。
+        图表中的模型、芯片、框架与指标名称均沿用业界通用英文名称。
       </p>
     </Card>
   );
diff --git a/packages/app/src/hooks/api/ai-chart-data.test.ts b/packages/app/src/hooks/api/ai-chart-data.test.ts
new file mode 100644
index 000000000..b7350fd99
--- /dev/null
+++ b/packages/app/src/hooks/api/ai-chart-data.test.ts
@@ -0,0 +1,262 @@
+import { describe, expect, it } from 'vitest';
+
+import chartDefinitions from '@/components/inference/metric-registry';
+
+import {
+  buildAiLineData,
+  getAiMetricDirection,
+  normalizeAiRadarRows,
+  rankAiHardwareKeys,
+  readAiMetric,
+} from './ai-chart-data';
+
+const chartDefinition = chartDefinitions[0] as Record<string, unknown>;
+
+describe('readAiMetric', () => {
+  it('preserves an explicitly measured zero while rejecting a missing telemetry property', () => {
+    expect(
+      readAiMetric({ measuredAvgPower: { y: 0 } }, 'measuredAvgPower.y', 'y_measuredAvgPower'),
+    ).toBe(0);
+    expect(readAiMetric({}, 'measuredAvgPower.y', 'y_measuredAvgPower')).toBeNull();
+    expect(
+      readAiMetric(
+        { measuredAvgPower: { y: Number.NaN } },
+        'measuredAvgPower.y',
+        'y_measuredAvgPower',
+      ),
+    ).toBeNull();
+  });
+
+  it.each([
+    ['costh.y', 'y_costh'],
+    ['tpPerGpu.y', 'y_tpPerGpu'],
+    ['jTotal.y', 'y_jTotal'],
+  ])('rejects a synthetic zero sentinel for %s', (path, metric) => {
+    const [field] = path.split('.');
+    expect(readAiMetric({ [field]: { y: 0 } }, path, metric)).toBeNull();
+  });
+});
+
+describe('getAiMetricDirection', () => {
+  const lowerIsBetterMetrics = Object.entries(chartDefinition)
+    .filter(([key, value]) => key.endsWith('_roofline') && String(value).startsWith('lower_'))
+    .map(([key]) => key.slice(0, -'_roofline'.length));
+
+  it.each(lowerIsBetterMetrics)('derives lower-is-better for %s from its roofline', (metric) => {
+    expect(getAiMetricDirection(metric, chartDefinition)).toBe('lower');
+  });
+
+  it('derives higher-is-better from an upper roofline', () => {
+    expect(getAiMetricDirection('y_tpPerGpu', chartDefinition)).toBe('higher');
+  });
+
+  it('defaults a directionless metric to higher-is-better', () => {
+    expect(getAiMetricDirection('y_measuredPowerPercentTdp', chartDefinition)).toBe('higher');
+  });
+});
+
+describe('buildAiLineData', () => {
+  it('keeps missing X positions as line gaps while preserving a real zero', () => {
+    const points = [
+      { hwKey: 'b200_vllm', x: 10, metric: { y: 5 } },
+      { hwKey: 'b200_vllm', x: 20 },
+      { hwKey: 'b200_vllm', x: 30, metric: { y: 0 } },
+    ];
+
+    const lines = buildAiLineData(points, 'y_measuredAvgPower', 'metric.y', new Set(['b200_vllm']));
+
+    expect(lines.b200_vllm).toHaveLength(3);
+    expect(lines.b200_vllm?.map((point) => point.x)).toEqual([10, 20, 30]);
+    expect(lines.b200_vllm?.[0]?.y).toBe(5);
+    expect(lines.b200_vllm?.[1]?.y).toBeNaN();
+    expect(lines.b200_vllm?.[2]?.y).toBe(0);
+  });
+
+  it('excludes a line whose selected metric is missing at every point', () => {
+    const lines = buildAiLineData(
+      [
+        { hwKey: 'b200_vllm', x: 10 },
+        { hwKey: 'b200_vllm', x: 20 },
+      ],
+      'y_measuredAvgPower',
+      'metric.y',
+      new Set(['b200_vllm']),
+    );
+
+    expect(lines).toEqual({});
+  });
+});
+
+describe('normalizeAiRadarRows', () => {
+  it('places lower measured power farther out on the radar', () => {
+    const normalized = normalizeAiRadarRows(
+      new Map([
+        ['low-power', [300]],
+        ['high-power', [600]],
+      ]),
+      ['y_measuredAvgPower'],
+      chartDefinition,
+    );
+
+    expect(normalized.get('low-power')?.[0]).toBe(1);
+    expect(normalized.get('high-power')?.[0]).toBe(0);
+  });
+
+  it('preserves missing radar metrics without treating them as zero', () => {
+    const normalized = normalizeAiRadarRows(
+      new Map([
+        ['missing', [null]],
+        ['low', [5]],
+        ['positive', [10]],
+      ]),
+      ['y_tpPerGpu'],
+      chartDefinition,
+    );
+
+    expect(normalized.get('missing')?.[0]).toBeNull();
+    expect(normalized.get('low')?.[0]).toBe(0);
+    expect(normalized.get('positive')?.[0]).toBe(1);
+  });
+
+  it('rejects synthetic zero sentinels while preserving measured zero telemetry', () => {
+    const unavailableCost = normalizeAiRadarRows(
+      new Map([
+        ['unavailable', [0]],
+        ['valid-low', [1]],
+        ['valid-high', [2]],
+      ]),
+      ['y_costh'],
+      chartDefinition,
+    );
+    const measuredPower = normalizeAiRadarRows(
+      new Map([
+        ['zero', [0]],
+        ['positive', [100]],
+      ]),
+      ['y_measuredAvgPower'],
+      chartDefinition,
+    );
+
+    expect(unavailableCost.get('unavailable')?.[0]).toBeNull();
+    expect(measuredPower.get('zero')?.[0]).toBe(1);
+  });
+});
+
+describe('synthetic metric availability', () => {
+  it('turns an unavailable cost into no bar value, a radar gap, and a line gap', () => {
+    const unavailable = { hwKey: 'b200_vllm', x: 20, costh: { y: 0 } };
+    const lines = buildAiLineData(
+      [
+        { hwKey: 'b200_vllm', x: 10, costh: { y: 2 } },
+        unavailable,
+        { hwKey: 'b200_vllm', x: 30, costh: { y: 1 } },
+      ],
+      'y_costh',
+      'costh.y',
+      new Set(['b200_vllm']),
+    );
+    const radar = normalizeAiRadarRows(
+      new Map([
+        ['unavailable', [0]],
+        ['valid-low', [1]],
+        ['valid-high', [2]],
+      ]),
+      ['y_costh'],
+      chartDefinition,
+    );
+
+    expect(readAiMetric(unavailable, 'costh.y', 'y_costh')).toBeNull();
+    expect(lines.b200_vllm?.[1]?.y).toBeNaN();
+    expect(radar.get('unavailable')?.[0]).toBeNull();
+  });
+});
+
+describe('rankAiHardwareKeys', () => {
+  const baseOptions = {
+    chartDefinition,
+    topN: 1,
+    distinctGpus: false,
+  };
+
+  it('selects minima for lower-is-better cost metrics', () => {
+    const points = [
+      { hwKey: 'b200_vllm', x: 10, costh: { y: 2 } },
+      { hwKey: 'b200_vllm', x: 20, costh: { y: 1.8 } },
+      { hwKey: 'mi355x_sglang', x: 10, costh: { y: 1.2 } },
+    ];
+
+    expect(
+      rankAiHardwareKeys(points, {
+        ...baseOptions,
+        metric: 'y_costh',
+        metricPath: 'costh.y',
+      }),
+    ).toEqual(['mi355x_sglang']);
+  });
+
+  it('selects maxima for higher-is-better throughput metrics', () => {
+    const points = [
+      { hwKey: 'b200_vllm', x: 10, tpPerGpu: { y: 80 } },
+      { hwKey: 'b200_vllm', x: 20, tpPerGpu: { y: 100 } },
+      { hwKey: 'mi355x_sglang', x: 10, tpPerGpu: { y: 90 } },
+    ];
+
+    expect(
+      rankAiHardwareKeys(points, {
+        ...baseOptions,
+        metric: 'y_tpPerGpu',
+        metricPath: 'tpPerGpu.y',
+      }),
+    ).toEqual(['b200_vllm']);
+  });
+
+  it('picks one best config per GPU before ranking distinct GPU families', () => {
+    const points = [
+      { hwKey: 'b200_vllm', x: 10, costh: { y: 2 } },
+      { hwKey: 'b200_sglang', x: 10, costh: { y: 1 } },
+      { hwKey: 'mi355x_vllm', x: 10, costh: { y: 1.5 } },
+      { hwKey: 'h200_vllm', x: 10, costh: { y: 3 } },
+    ];
+
+    expect(
+      rankAiHardwareKeys(points, {
+        chartDefinition,
+        metric: 'y_costh',
+        metricPath: 'costh.y',
+        topN: 2,
+        distinctGpus: true,
+      }),
+    ).toEqual(['b200_sglang', 'mi355x_vllm']);
+  });
+
+  it('excludes missing and synthetic zero values from lower-is-better ranking', () => {
+    const points = [
+      { hwKey: 'missing', x: 10 },
+      { hwKey: 'zero', x: 10, costh: { y: 0 } },
+      { hwKey: 'positive', x: 10, costh: { y: 1 } },
+    ];
+
+    expect(
+      rankAiHardwareKeys(points, {
+        ...baseOptions,
+        metric: 'y_costh',
+        metricPath: 'costh.y',
+      }),
+    ).toEqual(['positive']);
+  });
+
+  it('preserves measured zero telemetry in lower-is-better ranking', () => {
+    const points = [
+      { hwKey: 'zero', x: 10, measuredAvgPower: { y: 0 } },
+      { hwKey: 'positive', x: 10, measuredAvgPower: { y: 1 } },
+    ];
+
+    expect(
+      rankAiHardwareKeys(points, {
+        ...baseOptions,
+        metric: 'y_measuredAvgPower',
+        metricPath: 'measuredAvgPower.y',
+      }),
+    ).toEqual(['zero']);
+  });
+});
diff --git a/packages/app/src/hooks/api/ai-chart-data.ts b/packages/app/src/hooks/api/ai-chart-data.ts
new file mode 100644
index 000000000..99787b2cc
--- /dev/null
+++ b/packages/app/src/hooks/api/ai-chart-data.ts
@@ -0,0 +1,158 @@
+export type AiMetricDirection = 'higher' | 'lower';
+
+export interface AiMetricPoint {
+  hwKey?: string | null;
+  x: number;
+}
+
+export interface RankAiHardwareOptions {
+  metric: string;
+  metricPath: string;
+  chartDefinition: Record<string, unknown>;
+  topN: number;
+  distinctGpus: boolean;
+}
+
+function isMeasuredTelemetryMetric(metric: string): boolean {
+  return metric.startsWith('y_measured');
+}
+
+export function isAiMetricValueValid(metric: string, value: unknown): value is number {
+  return (
+    typeof value === 'number' &&
+    Number.isFinite(value) &&
+    (value > 0 || (value === 0 && isMeasuredTelemetryMetric(metric)))
+  );
+}
+
+export function readAiMetric(point: unknown, path: string, metric: string): number | null {
+  let value: unknown = point;
+
+  for (const key of path.split('.')) {
+    if (typeof value !== 'object' || value === null || !(key in value)) return null;
+    value = (value as Record<string, unknown>)[key];
+  }
+
+  return isAiMetricValueValid(metric, value) ? value : null;
+}
+
+export function getAiMetricDirection(
+  metric: string,
+  chartDefinition: Record<string, unknown>,
+): AiMetricDirection {
+  const roofline = chartDefinition[`${metric}_roofline`];
+  return typeof roofline === 'string' && roofline.startsWith('lower_') ? 'lower' : 'higher';
+}
+
+export function buildAiLineData<T extends AiMetricPoint>(
+  points: readonly T[],
+  metric: string,
+  metricPath: string,
+  visibleHardwareKeys: ReadonlySet<string>,
+): Record<string, { x: number; y: number }[]> {
+  const lines: Record<string, { x: number; y: number }[]> = {};
+
+  for (const point of points) {
+    const hwKey = point.hwKey ?? '';
+    if (!hwKey || !visibleHardwareKeys.has(hwKey)) continue;
+
+    lines[hwKey] ??= [];
+    lines[hwKey].push({
+      x: point.x,
+      y: readAiMetric(point, metricPath, metric) ?? Number.NaN,
+    });
+  }
+
+  for (const [hwKey, line] of Object.entries(lines)) {
+    line.sort((a, b) => a.x - b.x);
+    if (!line.some(({ y }) => Number.isFinite(y))) delete lines[hwKey];
+  }
+  return lines;
+}
+
+export function normalizeAiRadarRows(
+  rawRows: ReadonlyMap<string, readonly (number | null)[]>,
+  metrics: readonly string[],
+  chartDefinition: Record<string, unknown>,
+): Map<string, (number | null)[]> {
+  const mins = metrics.map(() => Infinity);
+  const maxs = metrics.map(() => -Infinity);
+
+  for (const values of rawRows.values()) {
+    for (let index = 0; index < metrics.length; index++) {
+      const value = values[index];
+      if (!isAiMetricValueValid(metrics[index], value)) continue;
+      mins[index] = Math.min(mins[index], value);
+      maxs[index] = Math.max(maxs[index], value);
+    }
+  }
+
+  const normalizedRows = new Map<string, (number | null)[]>();
+  for (const [hwKey, values] of rawRows) {
+    normalizedRows.set(
+      hwKey,
+      metrics.map((metric, index) => {
+        const value = values[index];
+        if (
+          !isAiMetricValueValid(metric, value) ||
+          !Number.isFinite(mins[index]) ||
+          maxs[index] === mins[index]
+        ) {
+          return null;
+        }
+
+        const normalized = (value - mins[index]) / (maxs[index] - mins[index]);
+        return getAiMetricDirection(metric, chartDefinition) === 'lower'
+          ? 1 - normalized
+          : normalized;
+      }),
+    );
+  }
+
+  return normalizedRows;
+}
+
+export function rankAiHardwareKeys<T extends AiMetricPoint>(
+  points: readonly T[],
+  options: RankAiHardwareOptions,
+): string[] {
+  const direction = getAiMetricDirection(options.metric, options.chartDefinition);
+  const bestByHardware = new Map<string, number>();
+
+  for (const point of points) {
+    const hwKey = point.hwKey ?? '';
+    if (!hwKey) continue;
+
+    const value = readAiMetric(point, options.metricPath, options.metric);
+    if (value === null) continue;
+
+    const existing = bestByHardware.get(hwKey);
+    if (existing === undefined || (direction === 'lower' ? value < existing : value > existing)) {
+      bestByHardware.set(hwKey, value);
+    }
+  }
+
+  const compare = (left: number, right: number) =>
+    direction === 'lower' ? left - right : right - left;
+
+  if (!options.distinctGpus) {
+    return [...bestByHardware.entries()]
+      .toSorted(([, left], [, right]) => compare(left, right))
+      .slice(0, options.topN)
+      .map(([hwKey]) => hwKey);
+  }
+
+  const bestPerGpu = new Map<string, { hwKey: string; value: number }>();
+  for (const [hwKey, value] of bestByHardware) {
+    const baseGpu = hwKey.split('_')[0];
+    const existing = bestPerGpu.get(baseGpu);
+    if (!existing || compare(value, existing.value) < 0) {
+      bestPerGpu.set(baseGpu, { hwKey, value });
+    }
+  }
+
+  return [...bestPerGpu.values()]
+    .toSorted((left, right) => compare(left.value, right.value))
+    .slice(0, options.topN)
+    .map(({ hwKey }) => hwKey);
+}
diff --git a/packages/app/src/hooks/api/benchmark-id-query.test.tsx b/packages/app/src/hooks/api/benchmark-id-query.test.tsx
new file mode 100644
index 000000000..5f847322c
--- /dev/null
+++ b/packages/app/src/hooks/api/benchmark-id-query.test.tsx
@@ -0,0 +1,38 @@
+// @vitest-environment jsdom
+import { act, createElement } from 'react';
+import { createRoot } from 'react-dom/client';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+const useQueryMock = vi.hoisted(() => vi.fn((options: unknown) => options));
+
+vi.mock('@tanstack/react-query', () => ({ useQuery: useQueryMock }));
+
+import { useByIdQuery } from './benchmark-id-query';
+
+const decode = (value: { encoded: number }) => ({ decoded: value.encoded });
+
+describe('useByIdQuery', () => {
+  beforeEach(() => useQueryMock.mockClear());
+
+  it('keeps the nullable select wrapper stable across parent renders', () => {
+    function Harness({ caption: _caption }: { caption: string }) {
+      useByIdQuery('request-chart-data', 439331, true, decode);
+      return null;
+    }
+
+    const container = document.createElement('div');
+    const root = createRoot(container);
+
+    act(() => root.render(createElement(Harness, { caption: 'first' })));
+    const firstSelect = useQueryMock.mock.calls.at(-1)?.[0].select;
+
+    act(() => root.render(createElement(Harness, { caption: 'second' })));
+    const secondSelect = useQueryMock.mock.calls.at(-1)?.[0].select;
+
+    expect(secondSelect).toBe(firstSelect);
+    expect(secondSelect({ encoded: 42 })).toEqual({ decoded: 42 });
+    expect(secondSelect(null)).toBeNull();
+
+    act(() => root.unmount());
+  });
+});
diff --git a/packages/app/src/hooks/api/benchmark-id-query.ts b/packages/app/src/hooks/api/benchmark-id-query.ts
index 0aa506876..d1540f1d8 100644
--- a/packages/app/src/hooks/api/benchmark-id-query.ts
+++ b/packages/app/src/hooks/api/benchmark-id-query.ts
@@ -1,3 +1,4 @@
+import { useCallback } from 'react';
 import { useQuery } from '@tanstack/react-query';
 
 /**
@@ -43,7 +44,18 @@ export function useBulkIdsQuery<T>(
 }
 
 /** Single-payload query for one benchmark_results id; 404 resolves to null. */
-export function useByIdQuery<T>(endpoint: string, id: number | null, enabled: boolean) {
+export function useByIdQuery<T, TSelected = T>(
+  endpoint: string,
+  id: number | null,
+  enabled: boolean,
+  select?: (data: T) => TSelected,
+) {
+  const selectNullable = useCallback(
+    (data: T | null): TSelected | null =>
+      data === null ? null : (select?.(data) ?? (data as unknown as TSelected)),
+    [select],
+  );
+
   return useQuery({
     queryKey: [endpoint, id] as const,
     queryFn: async ({ signal }): Promise<T | null> => {
@@ -55,5 +67,9 @@ export function useByIdQuery<T>(endpoint: string, id: number | null, enabled: bo
     },
     enabled,
     staleTime: STALE_TIME_MS,
+    // TanStack Query re-runs select when its function identity changes. Keep
+    // this wrapper stable so decoding a large compact payload happens only
+    // when the cached wire data or the caller's selector actually changes.
+    select: selectNullable,
   });
 }
diff --git a/packages/app/src/hooks/api/use-agentic-aggregates.ts b/packages/app/src/hooks/api/use-agentic-aggregates.ts
index 7ca029cff..0979618f9 100644
--- a/packages/app/src/hooks/api/use-agentic-aggregates.ts
+++ b/packages/app/src/hooks/api/use-agentic-aggregates.ts
@@ -5,6 +5,7 @@ export interface MetricPercentiles {
   p50: number;
   p75: number;
   p90: number;
+  p95: number;
   p99: number;
   n: number;
 }
@@ -22,7 +23,7 @@ export type AgenticAggregateMap = Record<number, AgenticAggregate>;
 const fetchAgenticAggregates = bulkIdsFetcher<AgenticAggregate>('agentic-aggregates');
 
 /**
- * Fetch per-id aggregate stats (mean/p50/p75/p90/p99) for ISL, OSL, KV
+ * Fetch per-id aggregate stats (mean/p50/p75/p90/p95/p99) for ISL, OSL, KV
  * cache utilization, and prefix cache hit rate. Used by the "Aggregates
  * across configs" view on the agentic detail page.
  */
diff --git a/packages/app/src/hooks/api/use-ai-chart.ts b/packages/app/src/hooks/api/use-ai-chart.ts
index 92783463a..adddb90d5 100644
--- a/packages/app/src/hooks/api/use-ai-chart.ts
+++ b/packages/app/src/hooks/api/use-ai-chart.ts
@@ -26,14 +26,17 @@ import {
   dedupeAgenticHistoryRuns,
   dedupeRowsToLatestPerConfig,
 } from '@/lib/benchmark-run-selection';
-import {
-  getNestedYValue,
-  normalizeEvalHardwareKey,
-  generateHighContrastColors,
-} from '@/lib/chart-utils';
+import { normalizeEvalHardwareKey, generateHighContrastColors } from '@/lib/chart-utils';
 import { getHardwareConfig, getModelSortIndex } from '@/lib/constants';
 
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
+import {
+  buildAiLineData,
+  normalizeAiRadarRows,
+  rankAiHardwareKeys,
+  readAiMetric,
+} from './ai-chart-data';
+
+import chartDefinitions from '@/components/inference/metric-registry';
 
 // ---------------------------------------------------------------------------
 // Result types
@@ -123,8 +126,8 @@ function buildBenchmarkBarData(
       }
     }
 
-    const value = getNestedYValue(closest, yFieldPath);
-    if (value <= 0) continue;
+    const value = readAiMetric(closest, yFieldPath, spec.yAxisMetric);
+    if (value === null) continue;
 
     const config = getHardwareConfig(hwKey, closest.model);
     bars.push({
@@ -271,19 +274,7 @@ function buildLineData(
 ): Record<string, { x: number; y: number }[]> {
   const chartDef = (chartDefinitions as any[])[0];
   const yFieldPath: string = chartDef[spec.yAxisMetric] ?? 'tpPerGpu.y';
-
-  const lines: Record<string, { x: number; y: number }[]> = {};
-  for (const p of points) {
-    const hwKey = p.hwKey ?? '';
-    if (!hwKey || !colorMap[hwKey]) continue;
-    if (!lines[hwKey]) lines[hwKey] = [];
-    lines[hwKey].push({ x: p.x, y: getNestedYValue(p, yFieldPath) });
-  }
-  // Sort each line by x
-  for (const pts of Object.values(lines)) {
-    pts.sort((a, b) => a.x - b.x);
-  }
-  return lines;
+  return buildAiLineData(points, spec.yAxisMetric, yFieldPath, new Set(Object.keys(colorMap)));
 }
 
 function buildRadarData(
@@ -307,51 +298,25 @@ function buildRadarData(
   }
 
   // Extract raw values per metric per GPU
-  const rawMatrix = new Map<string, number[]>();
+  const rawMatrix = new Map<string, (number | null)[]>();
   for (const [hwKey, point] of groups) {
     const vals = metrics.map((m) => {
       const path: string = chartDef[m] ?? m;
-      return getNestedYValue(point, path);
+      return readAiMetric(point, path, m);
     });
     rawMatrix.set(hwKey, vals);
   }
-
-  // Find min/max per metric for normalization
-  const mins = metrics.map(() => Infinity);
-  const maxs = metrics.map(() => -Infinity);
-  for (const vals of rawMatrix.values()) {
-    for (let i = 0; i < vals.length; i++) {
-      if (vals[i] > 0) {
-        mins[i] = Math.min(mins[i], vals[i]);
-        maxs[i] = Math.max(maxs[i], vals[i]);
-      }
-    }
-  }
-
-  // For cost/energy metrics, invert normalization (lower is better)
-  const invertMetric = new Set([
-    'y_costh',
-    'y_costn',
-    'y_costr',
-    'y_jTotal',
-    'y_jOutput',
-    'y_jInput',
-  ]);
+  const normalizedMatrix = normalizeAiRadarRows(rawMatrix, metrics, chartDef);
 
   const items: AiRadarItem[] = [];
   for (const [hwKey, rawVals] of rawMatrix) {
     const config = getHardwareConfig(hwKey, groups.get(hwKey)?.model);
-    const normalized = rawVals.map((v, i) => {
-      if (v <= 0 || !isFinite(mins[i]) || maxs[i] === mins[i]) return null;
-      const norm = (v - mins[i]) / (maxs[i] - mins[i]);
-      return invertMetric.has(metrics[i]) ? 1 - norm : norm;
-    });
     items.push({
       hwKey,
       label: config ? `${config.label}${config.suffix ? ` ${config.suffix}` : ''}` : hwKey,
       color: colorMap[hwKey] ?? '#888',
-      values: normalized,
-      rawValues: rawVals.map((v) => (v > 0 ? v : null)),
+      values: normalizedMatrix.get(hwKey) ?? metrics.map(() => null),
+      rawValues: rawVals,
     });
   }
 
@@ -459,43 +424,19 @@ async function resolveSpec(spec: AiChartSpec): Promise<AiSingleChartResult> {
     });
   }
 
-  // topN: rank configs by peak metric value and keep only the best N
+  // topN: rank configs by their best value in the metric's canonical direction.
   if (spec.topN) {
     const chartDef = (chartDefinitions as any[])[0];
     const yFieldPath: string = chartDef[spec.yAxisMetric] ?? 'tpPerGpu.y';
-    const peakByHw = new Map<string, number>();
-    for (const p of points) {
-      const hw = p.hwKey ?? '';
-      if (!hw) continue;
-      const val = getNestedYValue(p, yFieldPath);
-      peakByHw.set(hw, Math.max(peakByHw.get(hw) ?? 0, val));
-    }
-
-    let topHwKeys: Set<string>;
-    if (spec.topNDistinctGpus === false) {
-      // Rank individual configs regardless of GPU family
-      topHwKeys = new Set(
-        [...peakByHw.entries()]
-          .toSorted(([, a], [, b]) => b - a)
-          .slice(0, spec.topN)
-          .map(([k]) => k),
-      );
-    } else {
-      // Group by base GPU, pick the best config per GPU, then take top N GPUs
-      const bestPerGpu = new Map<string, { hwKey: string; peak: number }>();
-      for (const [hwKey, peak] of peakByHw) {
-        const base = hwKey.split('_')[0];
-        const existing = bestPerGpu.get(base);
-        if (!existing || peak > existing.peak) {
-          bestPerGpu.set(base, { hwKey, peak });
-        }
-      }
-      const topBases = [...bestPerGpu.entries()]
-        .toSorted(([, a], [, b]) => b.peak - a.peak)
-        .slice(0, spec.topN);
-      // Include only the single best config per winning GPU
-      topHwKeys = new Set(topBases.map(([, v]) => v.hwKey));
-    }
+    const topHwKeys = new Set(
+      rankAiHardwareKeys(points, {
+        metric: spec.yAxisMetric,
+        metricPath: yFieldPath,
+        chartDefinition: chartDef,
+        topN: spec.topN,
+        distinctGpus: spec.topNDistinctGpus !== false,
+      }),
+    );
     points = points.filter((p) => topHwKeys.has(p.hwKey ?? ''));
   }
 
diff --git a/packages/app/src/hooks/api/use-availability.ts b/packages/app/src/hooks/api/use-availability.ts
index cb4d283ad..1f138df0b 100644
--- a/packages/app/src/hooks/api/use-availability.ts
+++ b/packages/app/src/hooks/api/use-availability.ts
@@ -1,10 +1,11 @@
 import { useQuery } from '@tanstack/react-query';
 
 import { fetchAvailability } from '@/lib/api';
+import { withSupplementalAvailability } from '@/lib/supplemental-benchmarks';
 
 export function useAvailability() {
   return useQuery({
     queryKey: ['availability'],
-    queryFn: ({ signal }) => fetchAvailability(signal),
+    queryFn: async ({ signal }) => withSupplementalAvailability(await fetchAvailability(signal)),
   });
 }
diff --git a/packages/app/src/hooks/api/use-benchmark-history.ts b/packages/app/src/hooks/api/use-benchmark-history.ts
index 971919d19..ad300c935 100644
--- a/packages/app/src/hooks/api/use-benchmark-history.ts
+++ b/packages/app/src/hooks/api/use-benchmark-history.ts
@@ -1,18 +1,38 @@
 import { useQuery } from '@tanstack/react-query';
 
 import { fetchBenchmarkHistory } from '@/lib/api';
+import { withSupplementalBenchmarkHistory } from '@/lib/supplemental-benchmarks';
 
+/**
+ * Full benchmark history for one model + sequence — every run date, not just
+ * the latest. This is a large response (single-digit MB for a well-swept
+ * model), so `enabled` matters: pass `false` until a consumer actually needs it.
+ */
 export function useBenchmarkHistory(
   model: string,
   isl: number,
   osl: number,
-  benchmarkType?: 'agentic_traces',
+  options?: {
+    /** Agentic traces have no fixed sequence, so they ignore isl/osl entirely. */
+    benchmarkType?: 'agentic_traces';
+    /** Trim to the calculator's metric allowlist. Not safe for measured-power charts. */
+    view?: 'calculator';
+    /** Gate the fetch. Defaults to true; combined with the model/isl/osl check. */
+    enabled?: boolean;
+  },
 ) {
+  const { benchmarkType, view } = options ?? {};
   return useQuery({
-    queryKey: benchmarkType
-      ? ['benchmark-history', model, isl, osl, benchmarkType]
-      : ['benchmark-history', model, isl, osl],
-    queryFn: ({ signal }) => fetchBenchmarkHistory(model, isl, osl, signal, benchmarkType),
-    enabled: Boolean(model && (benchmarkType === 'agentic_traces' || (isl && osl))),
+    // Both discriminators are part of the key: these responses have different
+    // shapes and must never share a cache entry.
+    queryKey: ['benchmark-history', model, isl, osl, benchmarkType ?? 'fixed', view ?? 'full'],
+    queryFn: async ({ signal }) =>
+      withSupplementalBenchmarkHistory(
+        await fetchBenchmarkHistory(model, isl, osl, signal, benchmarkType, view),
+        { model, isl, osl, benchmarkType },
+      ),
+    enabled:
+      Boolean(model && (benchmarkType === 'agentic_traces' || (isl && osl))) &&
+      (options?.enabled ?? true),
   });
 }
diff --git a/packages/app/src/hooks/api/use-benchmark-siblings.ts b/packages/app/src/hooks/api/use-benchmark-siblings.ts
index 7cdb990f0..7471dd45b 100644
--- a/packages/app/src/hooks/api/use-benchmark-siblings.ts
+++ b/packages/app/src/hooks/api/use-benchmark-siblings.ts
@@ -7,11 +7,15 @@ export interface BenchmarkSibling {
   decode_tp: number;
   decode_ep: number;
   decode_pp: number | null;
+  decode_dcp_size: number | null;
+  decode_pcp_size: number | null;
   decode_dp_attention: boolean;
   decode_num_workers: number;
   prefill_tp: number;
   prefill_ep: number;
   prefill_pp: number | null;
+  prefill_dcp_size: number | null;
+  prefill_pcp_size: number | null;
   prefill_dp_attention: boolean;
   prefill_num_workers: number;
   num_prefill_gpu: number;
diff --git a/packages/app/src/hooks/api/use-benchmarks.initial-data.test.tsx b/packages/app/src/hooks/api/use-benchmarks.initial-data.test.tsx
new file mode 100644
index 000000000..4c272ee3e
--- /dev/null
+++ b/packages/app/src/hooks/api/use-benchmarks.initial-data.test.tsx
@@ -0,0 +1,182 @@
+// @vitest-environment jsdom
+import { QueryClient, QueryClientProvider } from '@tanstack/react-query';
+import { act, createElement } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+import type { BenchmarkRow } from '@/lib/api';
+
+const { mockFetchBenchmarks } = vi.hoisted(() => ({ mockFetchBenchmarks: vi.fn() }));
+vi.mock('@/lib/api', () => ({ fetchBenchmarks: mockFetchBenchmarks }));
+
+import { useBenchmarks } from './use-benchmarks';
+
+const INITIAL_ROWS = [{ date: '2026-08-01', precision: 'fp8' }] as BenchmarkRow[];
+let observed: BenchmarkRow[] | undefined;
+
+function Probe({ sequence, initialData }: { sequence: string; initialData?: BenchmarkRow[] }) {
+  const query = useBenchmarks(
+    'DeepSeek-R1-0528',
+    '',
+    true,
+    undefined,
+    undefined,
+    {
+      type: 'calculator',
+      sequence,
+      ...(initialData ? { cacheScope: 'compare-initial:h100,h200' } : {}),
+    },
+    initialData,
+  );
+  observed = query.data;
+  return null;
+}
+
+function FullProbe({
+  model,
+  date,
+  initialData,
+}: {
+  model: string;
+  date: string;
+  initialData?: BenchmarkRow[];
+}) {
+  const query = useBenchmarks(
+    model,
+    date,
+    true,
+    undefined,
+    undefined,
+    undefined,
+    initialData,
+    'compare-pair:h100:h200',
+  );
+  observed = query.data;
+  return null;
+}
+
+describe('useBenchmarks initial calculator data', () => {
+  let root: Root | undefined;
+
+  afterEach(() => {
+    if (root) act(() => root?.unmount());
+    root = undefined;
+    observed = undefined;
+    vi.clearAllMocks();
+  });
+
+  it('skips the hydrated initial request but fetches a later selector key', async () => {
+    mockFetchBenchmarks.mockResolvedValue([{ date: '2026-08-02', precision: 'bf16' }]);
+    const client = new QueryClient({
+      defaultOptions: { queries: { staleTime: Infinity, retry: false } },
+    });
+    const container = document.createElement('div');
+    root = createRoot(container);
+
+    await act(() => {
+      root?.render(
+        createElement(
+          QueryClientProvider,
+          { client },
+          createElement(Probe, { sequence: '1k/1k', initialData: INITIAL_ROWS }),
+        ),
+      );
+    });
+
+    expect(observed).toBe(INITIAL_ROWS);
+    expect(mockFetchBenchmarks).not.toHaveBeenCalled();
+
+    await act(() => {
+      root?.render(
+        createElement(QueryClientProvider, { client }, createElement(Probe, { sequence: '8k/1k' })),
+      );
+    });
+
+    expect(mockFetchBenchmarks).toHaveBeenCalledTimes(1);
+    expect(mockFetchBenchmarks).toHaveBeenCalledWith(
+      'DeepSeek-R1-0528',
+      '',
+      undefined,
+      expect.any(AbortSignal),
+      undefined,
+      undefined,
+      { type: 'calculator', sequence: '8k/1k' },
+    );
+  });
+
+  it('isolates a hydrated pair-scoped full query and fetches later dates normally', async () => {
+    mockFetchBenchmarks.mockResolvedValue([{ date: '2026-07-01', precision: 'fp8' }]);
+    const client = new QueryClient({
+      defaultOptions: { queries: { staleTime: Infinity, retry: false } },
+    });
+    root = createRoot(document.createElement('div'));
+
+    await act(() => {
+      root?.render(
+        createElement(
+          QueryClientProvider,
+          { client },
+          createElement(FullProbe, {
+            model: 'DeepSeek-R1-0528',
+            date: '',
+            initialData: INITIAL_ROWS,
+          }),
+        ),
+      );
+    });
+
+    expect(observed).toBe(INITIAL_ROWS);
+    expect(mockFetchBenchmarks).not.toHaveBeenCalled();
+    expect(
+      client.getQueryData(['benchmarks', 'DeepSeek-R1-0528', '', 'latest', 'all', 'asof']),
+    ).toBeUndefined();
+
+    await act(() => {
+      root?.render(
+        createElement(
+          QueryClientProvider,
+          { client },
+          createElement(FullProbe, {
+            model: 'DeepSeek-R1-0528',
+            date: '2026-07-01',
+          }),
+        ),
+      );
+    });
+
+    expect(mockFetchBenchmarks).toHaveBeenCalledTimes(1);
+  });
+
+  it('fetches normally when a model override has no matching initial rows', async () => {
+    mockFetchBenchmarks.mockResolvedValue([]);
+    const client = new QueryClient({
+      defaultOptions: { queries: { staleTime: Infinity, retry: false } },
+    });
+    root = createRoot(document.createElement('div'));
+
+    await act(() => {
+      root?.render(
+        createElement(
+          QueryClientProvider,
+          { client },
+          createElement(FullProbe, {
+            model: 'Overridden-Model',
+            date: '',
+            initialData: undefined,
+          }),
+        ),
+      );
+    });
+
+    expect(mockFetchBenchmarks).toHaveBeenCalledTimes(1);
+    expect(mockFetchBenchmarks).toHaveBeenCalledWith(
+      'Overridden-Model',
+      '',
+      undefined,
+      expect.any(AbortSignal),
+      undefined,
+      undefined,
+      undefined,
+    );
+  });
+});
diff --git a/packages/app/src/hooks/api/use-benchmarks.test.ts b/packages/app/src/hooks/api/use-benchmarks.test.ts
index 7da761852..6d2922e43 100644
--- a/packages/app/src/hooks/api/use-benchmarks.test.ts
+++ b/packages/app/src/hooks/api/use-benchmarks.test.ts
@@ -83,6 +83,40 @@ describe('benchmarkQueryOptions', () => {
     expect(calculator.queryKey).not.toEqual(otherSequence.queryKey);
   });
 
+  it('seeds only the matching initial calculator query', () => {
+    const rows = [{ id: 1 }] as never[];
+    const initial = benchmarkQueryOptions(
+      'DeepSeek-R1-0528',
+      '',
+      true,
+      undefined,
+      undefined,
+      undefined,
+      { type: 'calculator', sequence: '1k/1k' },
+      rows,
+    );
+    const changed = benchmarkQueryOptions(
+      'DeepSeek-R1-0528',
+      '',
+      true,
+      undefined,
+      undefined,
+      undefined,
+      { type: 'calculator', sequence: '8k/1k' },
+    );
+
+    expect(initial.initialData).toBe(rows);
+    expect('initialData' in changed).toBe(false);
+    expect(initial.queryKey).not.toEqual(changed.queryKey);
+  });
+
+  it('canonicalizes every no-date request to the materialized-view key', () => {
+    const omittedDate = benchmarkQueryOptions('m', 'latest');
+    const emptyDate = benchmarkQueryOptions('m', '');
+    expect(omittedDate.queryKey).toEqual(emptyDate.queryKey);
+    expect(emptyDate.queryKey[2]).toBe('');
+  });
+
   it('is enabled when model is non-empty', () => {
     const opts = benchmarkQueryOptions('DeepSeek-R1-0528', '2026-03-01');
     expect(opts.enabled).toBe(true);
diff --git a/packages/app/src/hooks/api/use-benchmarks.ts b/packages/app/src/hooks/api/use-benchmarks.ts
index b5e4546a7..659514368 100644
--- a/packages/app/src/hooks/api/use-benchmarks.ts
+++ b/packages/app/src/hooks/api/use-benchmarks.ts
@@ -1,6 +1,7 @@
 import { useQuery } from '@tanstack/react-query';
 
-import { fetchBenchmarks } from '@/lib/api';
+import { fetchBenchmarks, type BenchmarkRow } from '@/lib/api';
+import { withSupplementalBenchmarks } from '@/lib/supplemental-benchmarks';
 
 /** Shared query options — reused by useQueries for comparison dates. */
 export function benchmarkQueryOptions(
@@ -12,21 +13,33 @@ export function benchmarkQueryOptions(
   runId?: string,
   /** When true with a runId, fetch exactly that run's results (GPU comparison). */
   exactRun?: boolean,
-  view?: { type: 'calculator'; sequence: string },
+  view?: { type: 'calculator'; sequence: string; cacheScope?: string },
+  initialData?: BenchmarkRow[],
+  scope?: string,
 ) {
+  const canonicalDate = date === 'latest' ? '' : date;
   return {
     queryKey: [
       'benchmarks',
       model,
-      date,
+      canonicalDate,
       exact ? 'exact' : 'latest',
       runId ?? 'all',
       exactRun ? 'run' : 'asof',
-      ...(view ? ([view.type, view.sequence] as const) : []),
+      ...(view
+        ? ([view.type, view.sequence, ...(view.cacheScope ? [view.cacheScope] : [])] as const)
+        : []),
+      ...(scope ? (['scope', scope] as const) : []),
     ] as const,
-    queryFn: ({ signal }: { signal: AbortSignal }) =>
-      fetchBenchmarks(model, date, exact, signal, runId, exactRun, view),
+    queryFn: async ({ signal }: { signal: AbortSignal }) =>
+      withSupplementalBenchmarks(
+        await fetchBenchmarks(model, canonicalDate, exact, signal, runId, exactRun, view),
+        { model, date: canonicalDate, exact, runId, view },
+      ),
     enabled: enabled && Boolean(model),
+    // Pair-scoped SSR payloads must preserve their exact identity and contents;
+    // supplemental rows are merged on every ordinary network read above.
+    ...(initialData ? { initialData } : {}),
   };
 }
 
@@ -36,9 +49,21 @@ export function useBenchmarks(
   enabled = true,
   runId?: string,
   exactRun?: boolean,
-  view?: { type: 'calculator'; sequence: string },
+  view?: { type: 'calculator'; sequence: string; cacheScope?: string },
+  initialData?: BenchmarkRow[],
+  scope?: string,
 ) {
   return useQuery(
-    benchmarkQueryOptions(model, date ?? 'latest', enabled, undefined, runId, exactRun, view),
+    benchmarkQueryOptions(
+      model,
+      date ?? '',
+      enabled,
+      undefined,
+      runId,
+      exactRun,
+      view,
+      initialData,
+      scope,
+    ),
   );
 }
diff --git a/packages/app/src/hooks/api/use-comparison-changelogs.ts b/packages/app/src/hooks/api/use-comparison-changelogs.ts
index b56976140..27752fac5 100644
--- a/packages/app/src/hooks/api/use-comparison-changelogs.ts
+++ b/packages/app/src/hooks/api/use-comparison-changelogs.ts
@@ -1,12 +1,8 @@
 import { useQueries } from '@tanstack/react-query';
 import { useMemo } from 'react';
 
-import {
-  fetchWorkflowInfo,
-  type ChangelogRow,
-  type RunConfigRow,
-  type WorkflowInfoResponse,
-} from '@/lib/api';
+import { type ChangelogRow, type RunConfigRow, type WorkflowInfoResponse } from '@/lib/api';
+import { workflowInfoQueryOptions } from '@/hooks/api/use-workflow-info';
 
 export interface ComparisonChangelogEntry {
   config_keys: string[];
@@ -64,12 +60,7 @@ export function useComparisonChangelogs(
   ]);
 
   const queries = useQueries({
-    queries: datesToQuery.map((date) => ({
-      queryKey: benchmarkType ? ['workflow-info', date, benchmarkType] : ['workflow-info', date],
-      queryFn: ({ signal }: { signal: AbortSignal }) =>
-        fetchWorkflowInfo(date, signal, benchmarkType),
-      enabled: hasGPUs,
-    })),
+    queries: datesToQuery.map((date) => workflowInfoQueryOptions(date, benchmarkType, hasGPUs)),
   });
 
   const changelogs = useMemo(() => {
diff --git a/packages/app/src/hooks/api/use-datasets.test.tsx b/packages/app/src/hooks/api/use-datasets.test.tsx
new file mode 100644
index 000000000..a93c831ff
--- /dev/null
+++ b/packages/app/src/hooks/api/use-datasets.test.tsx
@@ -0,0 +1,71 @@
+// @vitest-environment jsdom
+import { QueryClient, QueryClientProvider } from '@tanstack/react-query';
+import { act, createElement } from 'react';
+import { createRoot } from 'react-dom/client';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+import { useDatasetConversations } from './use-datasets';
+
+function ConversationProbe({ search }: { search: string }) {
+  useDatasetConversations({ slug: 'trace', search });
+  return null;
+}
+
+async function flushQueries(): Promise<void> {
+  await act(async () => {
+    await Promise.resolve();
+    await Promise.resolve();
+  });
+}
+
+afterEach(() => {
+  vi.useRealTimers();
+  vi.restoreAllMocks();
+});
+
+describe('useDatasetConversations cache lifecycle', () => {
+  it('collects an inactive search key after the bounded conversation gc window', async () => {
+    vi.useFakeTimers();
+    vi.spyOn(globalThis, 'fetch').mockImplementation(() =>
+      Promise.resolve(Response.json({ total: 0, items: [] }, { status: 200 })),
+    );
+
+    const client = new QueryClient({
+      defaultOptions: { queries: { retry: false, gcTime: Infinity } },
+    });
+    const container = document.createElement('div');
+    const root = createRoot(container);
+    act(() =>
+      root.render(
+        createElement(
+          QueryClientProvider,
+          { client },
+          createElement(ConversationProbe, { search: 'alpha' }),
+        ),
+      ),
+    );
+    await flushQueries();
+    act(() =>
+      root.render(
+        createElement(
+          QueryClientProvider,
+          { client },
+          createElement(ConversationProbe, { search: 'beta' }),
+        ),
+      ),
+    );
+    await flushQueries();
+
+    const oldKey = ['dataset-conversations', 'trace', 'alpha', 50, 0, 'tokens'] as const;
+    expect(client.getQueryData(oldKey)).toEqual({ total: 0, items: [] });
+
+    act(() => vi.advanceTimersByTime(5 * 60 * 1000 - 1));
+    expect(client.getQueryData(oldKey)).toEqual({ total: 0, items: [] });
+
+    act(() => vi.advanceTimersByTime(1));
+    expect(client.getQueryData(oldKey)).toBeUndefined();
+
+    act(() => root.unmount());
+    client.clear();
+  });
+});
diff --git a/packages/app/src/hooks/api/use-datasets.ts b/packages/app/src/hooks/api/use-datasets.ts
index ea1b17cfd..4c79709f8 100644
--- a/packages/app/src/hooks/api/use-datasets.ts
+++ b/packages/app/src/hooks/api/use-datasets.ts
@@ -1,42 +1,26 @@
 import { useQuery, keepPreviousData } from '@tanstack/react-query';
 
 import type {
+  ConversationDetail,
+  ConversationList,
+  ConversationListItem,
   ConversationStructure,
+  DatasetDetail as RawDatasetDetail,
+  DatasetRecord,
+  DatasetSummary,
+  ListConversationsOpts,
   StructureNode,
-} from '@semianalysisai/inferencex-db/etl/weka-structure';
+} from '@semianalysisai/inferencex-db/queries/datasets';
 
-export type { ConversationStructure, StructureNode };
-
-export interface DatasetSummary {
-  blockSize?: number;
-  hashIdScope?: string | null;
-  totalIn?: number;
-  totalOut?: number;
-  totalCached?: number;
-  cachedPct?: number;
-  mainTurns?: number;
-  subagentGroups?: number;
-  subagentTurns?: number;
-  meanRequestsPerConversation?: number;
-  medianRequestsPerConversation?: number;
-  meanSubagentsPerTrace?: number;
-  medianSubagentsPerTrace?: number;
-  modelMix?: Record<string, number>;
-  [k: string]: unknown;
-}
-
-export interface DatasetRecord {
-  id: string;
-  slug: string;
-  label: string;
-  variant: string;
-  description: string | null;
-  hf_url: string | null;
-  license: string | null;
-  conversation_count: number;
-  summary: DatasetSummary;
-  ingested_at: string;
-}
+export type {
+  ConversationDetail,
+  ConversationList,
+  ConversationListItem,
+  ConversationStructure,
+  DatasetRecord,
+  DatasetSummary,
+  StructureNode,
+};
 
 export interface HistogramBin {
   x0: number;
@@ -75,40 +59,15 @@ export interface DatasetChartData {
   [k: string]: unknown;
 }
 
-export interface DatasetDetail extends DatasetRecord {
+export type DatasetDetail = RawDatasetDetail & {
   chart_data: DatasetChartData;
-}
-
-export interface ConversationListItem {
-  conv_id: string;
-  models: string[];
-  num_turns: number;
-  num_subagent_groups: number;
-  total_in: number;
-  total_out: number;
-  total_cached: number;
-}
-
-export interface ConversationList {
-  total: number;
-  items: ConversationListItem[];
-}
-
-export interface ConversationDetail {
-  conv_id: string;
-  models: string[];
-  num_turns: number;
-  num_subagent_groups: number;
-  total_in: number;
-  total_out: number;
-  total_cached: number;
-  structure: ConversationStructure;
-}
+};
 
-export type ConversationSort = 'tokens' | 'turns' | 'subagents' | 'id';
+export type ConversationSort = NonNullable<ListConversationsOpts['sort']>;
 
 // Dataset contents only change on (rare) re-ingest, so cache aggressively.
 const DAY = 24 * 60 * 60 * 1000;
+const CONVERSATION_QUERY_GC_TIME = 5 * 60 * 1000;
 
 /** Shared fetch for the per-dataset endpoints: 404 → null, other errors throw. */
 async function fetchJsonOr404<T>(
@@ -180,6 +139,7 @@ export function useDatasetConversations({
     enabled: Boolean(slug),
     placeholderData: keepPreviousData,
     staleTime: DAY,
+    gcTime: CONVERSATION_QUERY_GC_TIME,
   });
 }
 
diff --git a/packages/app/src/hooks/api/use-framework-releases.ts b/packages/app/src/hooks/api/use-framework-releases.ts
index cbc5c55ff..4200e0d86 100644
--- a/packages/app/src/hooks/api/use-framework-releases.ts
+++ b/packages/app/src/hooks/api/use-framework-releases.ts
@@ -5,7 +5,7 @@ import { fetchFrameworkReleases } from '@/lib/api';
 export function useFrameworkReleases() {
   return useQuery({
     queryKey: ['framework-releases'],
-    queryFn: fetchFrameworkReleases,
+    queryFn: ({ signal }) => fetchFrameworkReleases(signal),
     staleTime: 60 * 60 * 1000, // 1 hour
   });
 }
diff --git a/packages/app/src/hooks/api/use-latest-images.ts b/packages/app/src/hooks/api/use-latest-images.ts
index 5bc896a2b..0194e5bbc 100644
--- a/packages/app/src/hooks/api/use-latest-images.ts
+++ b/packages/app/src/hooks/api/use-latest-images.ts
@@ -5,6 +5,6 @@ import { fetchLatestImages } from '@/lib/api';
 export function useLatestImages() {
   return useQuery({
     queryKey: ['latest-images'],
-    queryFn: fetchLatestImages,
+    queryFn: ({ signal }) => fetchLatestImages(signal),
   });
 }
diff --git a/packages/app/src/hooks/api/use-log-availability.ts b/packages/app/src/hooks/api/use-log-availability.ts
new file mode 100644
index 000000000..8bb8bbb24
--- /dev/null
+++ b/packages/app/src/hooks/api/use-log-availability.ts
@@ -0,0 +1,10 @@
+import { bulkIdsFetcher, useBulkIdsQuery } from './benchmark-id-query';
+
+export type LogAvailabilityMap = Record<number, true>;
+
+const fetchLogAvailability = bulkIdsFetcher<true>('log-availability');
+
+/** Bulk presence lookup for DB-backed benchmark points with a stored server log. */
+export function useLogAvailability(ids: number[], enabled = true) {
+  return useBulkIdsQuery('log-availability', ids, enabled, fetchLogAvailability);
+}
diff --git a/packages/app/src/hooks/api/use-openrouter-pricing.test.ts b/packages/app/src/hooks/api/use-openrouter-pricing.test.ts
new file mode 100644
index 000000000..0161d8688
--- /dev/null
+++ b/packages/app/src/hooks/api/use-openrouter-pricing.test.ts
@@ -0,0 +1,39 @@
+import { describe, expect, it } from 'vitest';
+
+import { openRouterPricingForModel } from './use-openrouter-pricing';
+
+describe('openRouterPricingForModel', () => {
+  it('converts OpenRouter per-token strings to $/M token prices', () => {
+    const pricing = openRouterPricingForModel(
+      {
+        data: [
+          {
+            id: 'deepseek/deepseek-v4-pro-0813',
+            pricing: {
+              prompt: '0.000001122',
+              completion: '0.000003366',
+            },
+          },
+        ],
+      },
+      'deepseek/deepseek-v4-pro-0813',
+    );
+
+    expect(pricing).toEqual({
+      source: 'openrouter',
+      inputPerMillion: 1.122,
+      outputPerMillion: 3.366,
+      openRouterModelId: 'deepseek/deepseek-v4-pro-0813',
+    });
+  });
+
+  it('returns null for a missing model or incomplete prices', () => {
+    expect(openRouterPricingForModel({ data: [] }, 'missing/model')).toBeNull();
+    expect(
+      openRouterPricingForModel(
+        { data: [{ id: 'incomplete/model', pricing: { prompt: '0.1' } }] },
+        'incomplete/model',
+      ),
+    ).toBeNull();
+  });
+});
diff --git a/packages/app/src/hooks/api/use-openrouter-pricing.ts b/packages/app/src/hooks/api/use-openrouter-pricing.ts
new file mode 100644
index 000000000..ee152d7d2
--- /dev/null
+++ b/packages/app/src/hooks/api/use-openrouter-pricing.ts
@@ -0,0 +1,67 @@
+import { useQuery } from '@tanstack/react-query';
+
+import type { TokenRevenuePricing } from '@/components/inference/types';
+
+export const OPENROUTER_MODELS_URL = 'https://openrouter.ai/api/v1/models';
+
+interface OpenRouterModel {
+  id?: unknown;
+  pricing?: {
+    prompt?: unknown;
+    completion?: unknown;
+  };
+}
+
+interface OpenRouterModelsResponse {
+  data?: OpenRouterModel[];
+}
+
+function dollarsPerMillion(value: unknown): number | null {
+  if (typeof value !== 'string' && typeof value !== 'number') return null;
+  const parsed = Number(value);
+  return Number.isFinite(parsed) && parsed >= 0 ? parsed * 1_000_000 : null;
+}
+
+/** Parse OpenRouter's per-token catalog prices into dashboard $/M tok units. */
+export function openRouterPricingForModel(
+  payload: OpenRouterModelsResponse,
+  modelId: string,
+): TokenRevenuePricing | null {
+  const model = payload.data?.find((candidate) => candidate.id === modelId);
+  const inputPerMillion = dollarsPerMillion(model?.pricing?.prompt);
+  const outputPerMillion = dollarsPerMillion(model?.pricing?.completion);
+  if (inputPerMillion === null || outputPerMillion === null) return null;
+
+  return {
+    source: 'openrouter',
+    inputPerMillion,
+    outputPerMillion,
+    openRouterModelId: modelId,
+  };
+}
+
+export async function fetchOpenRouterPricing(
+  modelId: string,
+  signal?: AbortSignal,
+): Promise<TokenRevenuePricing> {
+  const response = await fetch(OPENROUTER_MODELS_URL, { signal });
+  if (!response.ok) throw new Error(`OpenRouter pricing request failed (${response.status})`);
+  const pricing = openRouterPricingForModel(
+    (await response.json()) as OpenRouterModelsResponse,
+    modelId,
+  );
+  if (!pricing) throw new Error(`OpenRouter pricing is unavailable for ${modelId}`);
+  return pricing;
+}
+
+/** Fetch the public OpenRouter model catalog only when its price source is active. */
+export function useOpenRouterPricing(modelId: string | null, enabled: boolean) {
+  return useQuery({
+    queryKey: ['openrouter-pricing', modelId],
+    queryFn: ({ signal }) => fetchOpenRouterPricing(modelId!, signal),
+    enabled: enabled && Boolean(modelId),
+    staleTime: 5 * 60 * 1_000,
+    gcTime: 60 * 60 * 1_000,
+    retry: 1,
+  });
+}
diff --git a/packages/app/src/hooks/api/use-request-chart-data.ts b/packages/app/src/hooks/api/use-request-chart-data.ts
new file mode 100644
index 000000000..96a90331d
--- /dev/null
+++ b/packages/app/src/hooks/api/use-request-chart-data.ts
@@ -0,0 +1,76 @@
+import { useByIdQuery } from './benchmark-id-query';
+
+type RequestChartTuple = [
+  cid: number,
+  phase: number,
+  startUs: number,
+  endUs: number,
+  ttftMs: number | null,
+  tpotMs: number | null,
+  isl: number | null,
+  osl: number | null,
+  cancelled: 0 | 1,
+];
+
+interface RequestChartDataWire {
+  version: number;
+  timelineVersion: number;
+  startNs: number;
+  endNs: number;
+  durationS: number;
+  cids: string[];
+  phases: string[];
+  requests: RequestChartTuple[];
+}
+
+export interface RequestChartRecord {
+  cid: string;
+  phase: string;
+  start: number;
+  end: number;
+  ttftMs: number | null;
+  tpotMs: number | null;
+  isl: number | null;
+  osl: number | null;
+  cancelled: boolean;
+}
+
+export interface RequestChartData {
+  version: number;
+  timelineVersion: number;
+  startNs: number;
+  endNs: number;
+  durationS: number;
+  requests: RequestChartRecord[];
+}
+
+function decodeRequestChartData(wire: RequestChartDataWire): RequestChartData {
+  return {
+    version: wire.version,
+    timelineVersion: wire.timelineVersion,
+    startNs: wire.startNs,
+    endNs: wire.endNs,
+    durationS: wire.durationS,
+    requests: wire.requests.map((request) => ({
+      cid: wire.cids[request[0]] ?? '',
+      phase: wire.phases[request[1]] ?? '',
+      start: request[2] * 1_000,
+      end: request[3] * 1_000,
+      ttftMs: request[4],
+      tpotMs: request[5],
+      isl: request[6],
+      osl: request[7],
+      cancelled: request[8] === 1,
+    })),
+  };
+}
+
+/** Fetches the compact request dataset used by the default point charts. */
+export function useRequestChartData(id: number | null, enabled = false) {
+  return useByIdQuery<RequestChartDataWire, RequestChartData>(
+    'request-chart-data',
+    id,
+    enabled && Boolean(id),
+    decodeRequestChartData,
+  );
+}
diff --git a/packages/app/src/hooks/api/use-request-timeline.ts b/packages/app/src/hooks/api/use-request-timeline.ts
index 6f43de250..048a8a494 100644
--- a/packages/app/src/hooks/api/use-request-timeline.ts
+++ b/packages/app/src/hooks/api/use-request-timeline.ts
@@ -3,6 +3,8 @@ import { useByIdQuery } from './benchmark-id-query';
 export interface RequestRecord {
   /** Conversation id (groups turns of one agent session). */
   cid: string;
+  /** Compact replay-lane id; separates concurrent reuses of one conversation. */
+  ri?: number;
   /** Zero-based turn index within the conversation. */
   ti: number;
   /** Source trace id from the original raw dataset, when provided by AIPerf. */
diff --git a/packages/app/src/hooks/api/use-resident-sequence-lengths.test.ts b/packages/app/src/hooks/api/use-resident-sequence-lengths.test.ts
new file mode 100644
index 000000000..87584671e
--- /dev/null
+++ b/packages/app/src/hooks/api/use-resident-sequence-lengths.test.ts
@@ -0,0 +1,10 @@
+import { describe, expect, it } from 'vitest';
+
+import { chunkResidentPointIds } from './use-resident-sequence-lengths';
+
+describe('chunkResidentPointIds', () => {
+  it('keeps resident-point requests within the API limit', () => {
+    const chunks = chunkResidentPointIds(Array.from({ length: 451 }, (_, index) => index + 1));
+    expect(chunks.map((chunk) => chunk.length)).toEqual([200, 200, 51]);
+  });
+});
diff --git a/packages/app/src/hooks/api/use-resident-sequence-lengths.ts b/packages/app/src/hooks/api/use-resident-sequence-lengths.ts
new file mode 100644
index 000000000..5fa469b53
--- /dev/null
+++ b/packages/app/src/hooks/api/use-resident-sequence-lengths.ts
@@ -0,0 +1,68 @@
+import { useQuery } from '@tanstack/react-query';
+
+import {
+  mergeTokenLengthSketches,
+  tokenLengthPercentiles,
+  type TokenLengthSketch,
+  type TokenLengthPercentiles,
+} from '@semianalysisai/inferencex-constants';
+
+interface ResidentSequenceLengthSketches {
+  isl: TokenLengthSketch | null;
+  osl: TokenLengthSketch | null;
+  coveredPoints: number;
+  requestedPoints: number;
+}
+
+export interface ResidentSequenceLengthPercentiles {
+  isl: TokenLengthPercentiles;
+  osl: TokenLengthPercentiles;
+  coveredPoints: number;
+  requestedPoints: number;
+}
+
+const MAX_IDS_PER_REQUEST = 200;
+const STALE_TIME_MS = 5 * 60 * 1000;
+
+export function chunkResidentPointIds(ids: readonly number[]): number[][] {
+  const chunks: number[][] = [];
+  for (let index = 0; index < ids.length; index += MAX_IDS_PER_REQUEST) {
+    chunks.push(ids.slice(index, index + MAX_IDS_PER_REQUEST));
+  }
+  return chunks;
+}
+
+async function fetchResidentSequenceLengths(
+  ids: number[],
+  signal?: AbortSignal,
+): Promise<ResidentSequenceLengthPercentiles | null> {
+  const chunks = await Promise.all(
+    chunkResidentPointIds(ids).map(async (chunk) => {
+      const response = await fetch(`/api/v1/resident-sequence-lengths?ids=${chunk.join(',')}`, {
+        signal,
+      });
+      if (!response.ok) throw new Error(`resident-sequence-lengths ${response.status}`);
+      return (await response.json()) as ResidentSequenceLengthSketches;
+    }),
+  );
+
+  const isl = tokenLengthPercentiles(mergeTokenLengthSketches(chunks.map((chunk) => chunk.isl)));
+  const osl = tokenLengthPercentiles(mergeTokenLengthSketches(chunks.map((chunk) => chunk.osl)));
+  if (!isl || !osl) return null;
+  return {
+    isl,
+    osl,
+    coveredPoints: chunks.reduce((sum, chunk) => sum + chunk.coveredPoints, 0),
+    requestedPoints: chunks.reduce((sum, chunk) => sum + chunk.requestedPoints, 0),
+  };
+}
+
+export function useResidentSequenceLengths(ids: number[], enabled = true) {
+  const sortedIds = [...new Set(ids)].toSorted((a, b) => a - b);
+  return useQuery({
+    queryKey: ['resident-sequence-lengths', sortedIds.join(',')] as const,
+    queryFn: ({ signal }) => fetchResidentSequenceLengths(sortedIds, signal),
+    enabled: enabled && sortedIds.length > 0,
+    staleTime: STALE_TIME_MS,
+  });
+}
diff --git a/packages/app/src/hooks/api/use-server-log-files.ts b/packages/app/src/hooks/api/use-server-log-files.ts
new file mode 100644
index 000000000..4f1659df3
--- /dev/null
+++ b/packages/app/src/hooks/api/use-server-log-files.ts
@@ -0,0 +1,6 @@
+import { useByIdQuery } from './benchmark-id-query';
+
+/** Artifact-relative .log/.out filenames available for one benchmark point. */
+export function useServerLogFiles(id: number | null, enabled = false) {
+  return useByIdQuery<string[]>('server-log-files', id, enabled);
+}
diff --git a/packages/app/src/hooks/api/use-server-log-search.ts b/packages/app/src/hooks/api/use-server-log-search.ts
new file mode 100644
index 000000000..1737d176e
--- /dev/null
+++ b/packages/app/src/hooks/api/use-server-log-search.ts
@@ -0,0 +1,37 @@
+import { useQuery } from '@tanstack/react-query';
+
+export interface ServerLogSearchMatch {
+  fileName: string;
+  offset: number;
+  before: string;
+  match: string;
+  after: string;
+}
+
+export interface ServerLogSearchResponse {
+  id: number;
+  query: string;
+  matches: ServerLogSearchMatch[];
+  truncated: boolean;
+}
+
+async function fetchServerLogSearch(
+  id: number,
+  query: string,
+  signal?: AbortSignal,
+): Promise<ServerLogSearchResponse> {
+  const params = new URLSearchParams({ id: String(id), q: query });
+  const response = await fetch(`/api/v1/server-log-search?${params}`, { signal });
+  if (!response.ok) throw new Error(`server-log-search ${response.status}`);
+  return (await response.json()) as ServerLogSearchResponse;
+}
+
+/** Search complete stored log files; the browser does not need to load them first. */
+export function useServerLogSearch(id: number, query: string, enabled = false) {
+  return useQuery({
+    queryKey: ['server-log-search', id, query] as const,
+    queryFn: ({ signal }) => fetchServerLogSearch(id, query, signal),
+    enabled: enabled && query.length > 0,
+    staleTime: 5 * 60 * 1000,
+  });
+}
diff --git a/packages/app/src/hooks/api/use-server-log.ts b/packages/app/src/hooks/api/use-server-log.ts
new file mode 100644
index 000000000..e4e457134
--- /dev/null
+++ b/packages/app/src/hooks/api/use-server-log.ts
@@ -0,0 +1,50 @@
+import { useInfiniteQuery } from '@tanstack/react-query';
+
+export const SERVER_LOG_CHUNK_SIZE = 64 * 1024;
+const STALE_TIME_MS = 5 * 60 * 1000;
+
+export interface ServerLogChunk {
+  id: number;
+  fileName: string;
+  serverLog: string;
+  offset: number;
+  nextOffset: number | null;
+}
+
+async function fetchServerLogChunk(
+  id: number,
+  fileName: string,
+  offset: number,
+  signal?: AbortSignal,
+): Promise<ServerLogChunk | null> {
+  const params = new URLSearchParams({
+    id: String(id),
+    file: fileName,
+    offset: String(offset),
+    limit: String(SERVER_LOG_CHUNK_SIZE),
+  });
+  const response = await fetch(`/api/v1/server-log?${params}`, { signal });
+  if (response.status === 404) return null;
+  if (!response.ok) throw new Error(`server-log ${response.status}`);
+  return (await response.json()) as ServerLogChunk;
+}
+
+/** Incrementally fetch immutable server-log chunks for one benchmark point. */
+export function useServerLog(
+  id: number | null,
+  fileName: string | null,
+  enabled = false,
+  initialOffset = 0,
+) {
+  return useInfiniteQuery({
+    queryKey: ['server-log', id, fileName, 'chunks', initialOffset] as const,
+    queryFn: ({ pageParam, signal }) => {
+      if (!id || !fileName) return Promise.resolve(null);
+      return fetchServerLogChunk(id, fileName, pageParam, signal);
+    },
+    initialPageParam: initialOffset,
+    getNextPageParam: (lastPage) => lastPage?.nextOffset ?? undefined,
+    enabled: enabled && Boolean(id) && Boolean(fileName),
+    staleTime: STALE_TIME_MS,
+  });
+}
diff --git a/packages/app/src/hooks/api/use-trace-histograms.ts b/packages/app/src/hooks/api/use-trace-histograms.ts
deleted file mode 100644
index 8197147a6..000000000
--- a/packages/app/src/hooks/api/use-trace-histograms.ts
+++ /dev/null
@@ -1,25 +0,0 @@
-import { bulkIdsFetcher, useBulkIdsQuery } from './benchmark-id-query';
-
-export interface TraceHistogramPoint {
-  id: number;
-  /** Input sequence length (tokens) per completed request. */
-  isl: number[];
-  /** Output sequence length (tokens) per completed request. */
-  osl: number[];
-}
-
-export type TraceHistogramMap = Record<number, TraceHistogramPoint>;
-
-const fetchTraceHistograms = bulkIdsFetcher<TraceHistogramPoint>('trace-histograms');
-
-/**
- * Fetch per-request ISL/OSL arrays for a set of benchmark_results.id values.
- * Ids without a stored trace_replay blob are silently omitted from the response.
- *
- * Caller passes the agentic id set currently on screen; React Query handles
- * dedup + stale-while-revalidate. Cache key is sorted-ids-comma-joined so
- * any permutation of the same set hits the same cache entry.
- */
-export function useTraceHistograms(ids: number[], enabled = true) {
-  return useBulkIdsQuery('trace-histograms', ids, enabled, fetchTraceHistograms);
-}
diff --git a/packages/app/src/hooks/api/use-trace-server-metrics.ts b/packages/app/src/hooks/api/use-trace-server-metrics.ts
index c802ac388..68ad6d441 100644
--- a/packages/app/src/hooks/api/use-trace-server-metrics.ts
+++ b/packages/app/src/hooks/api/use-trace-server-metrics.ts
@@ -1,3 +1,5 @@
+import { useQuery } from '@tanstack/react-query';
+
 import { useByIdQuery } from './benchmark-id-query';
 
 export interface TimeSeriesPoint {
@@ -63,6 +65,10 @@ export interface MetricSourceSeries {
   kvCacheUsageByEngine: { engineLabel: string; points: TimeSeriesPoint[] }[];
 }
 
+export interface MetricSourceDescriptor {
+  source: MetricSource;
+}
+
 export interface TraceServerMetrics {
   meta: PointMeta;
   startNs: number;
@@ -90,7 +96,7 @@ export interface TraceServerMetrics {
    */
   kvCachePoolTokens: number | null;
   /** Orchestrator-normalized metrics grouped by endpoint/worker. */
-  metricSources: MetricSourceSeries[];
+  metricSources: MetricSourceDescriptor[];
 }
 
 /**
@@ -101,3 +107,20 @@ export interface TraceServerMetrics {
 export function useTraceServerMetrics(id: number | null, enabled = false) {
   return useByIdQuery<TraceServerMetrics>('trace-server-metrics', id, enabled && Boolean(id));
 }
+
+/** Fetch one endpoint/worker's full arrays only after it is selected. */
+export function useTraceServerMetricSource(id: number | null, source: string, enabled = false) {
+  return useQuery({
+    queryKey: ['trace-server-metric-source', id, source] as const,
+    queryFn: async ({ signal }): Promise<MetricSourceSeries | null> => {
+      if (!id || !source || source === 'all') return null;
+      const params = new URLSearchParams({ id: String(id), source });
+      const response = await fetch(`/api/v1/trace-server-metric-source?${params}`, { signal });
+      if (response.status === 404) return null;
+      if (!response.ok) throw new Error(`trace-server-metric-source ${response.status}`);
+      return (await response.json()) as MetricSourceSeries;
+    },
+    enabled: enabled && Boolean(id) && source !== 'all',
+    staleTime: 5 * 60 * 1000,
+  });
+}
diff --git a/packages/app/src/hooks/api/use-workflow-info.test.ts b/packages/app/src/hooks/api/use-workflow-info.test.ts
new file mode 100644
index 000000000..d3ff7f5b9
--- /dev/null
+++ b/packages/app/src/hooks/api/use-workflow-info.test.ts
@@ -0,0 +1,25 @@
+import { describe, expect, it } from 'vitest';
+
+import { workflowInfoQueryOptions } from '@/hooks/api/use-workflow-info';
+
+describe('workflowInfoQueryOptions', () => {
+  it('shares the canonical single-date key', () => {
+    expect(workflowInfoQueryOptions('2026-08-20').queryKey).toEqual([
+      'workflow-info',
+      '2026-08-20',
+    ]);
+  });
+
+  it('adds benchmark type only when the endpoint request is scoped by it', () => {
+    expect(workflowInfoQueryOptions('2026-08-20', 'agentic_traces').queryKey).toEqual([
+      'workflow-info',
+      '2026-08-20',
+      'agentic_traces',
+    ]);
+  });
+
+  it('does not enable an empty date or override an explicit disabled query', () => {
+    expect(workflowInfoQueryOptions('').enabled).toBe(false);
+    expect(workflowInfoQueryOptions('2026-08-20', undefined, false).enabled).toBe(false);
+  });
+});
diff --git a/packages/app/src/hooks/api/use-workflow-info.ts b/packages/app/src/hooks/api/use-workflow-info.ts
index 3fcd003f4..b781edee5 100644
--- a/packages/app/src/hooks/api/use-workflow-info.ts
+++ b/packages/app/src/hooks/api/use-workflow-info.ts
@@ -2,10 +2,21 @@ import { useQuery } from '@tanstack/react-query';
 
 import { fetchWorkflowInfo } from '@/lib/api';
 
+export function workflowInfoQueryOptions(
+  date: string,
+  benchmarkType?: 'agentic_traces',
+  enabled = true,
+) {
+  return {
+    queryKey: benchmarkType
+      ? (['workflow-info', date, benchmarkType] as const)
+      : (['workflow-info', date] as const),
+    queryFn: ({ signal }: { signal: AbortSignal }) =>
+      fetchWorkflowInfo(date, signal, benchmarkType),
+    enabled: enabled && Boolean(date),
+  };
+}
+
 export function useWorkflowInfo(date: string) {
-  return useQuery({
-    queryKey: ['workflow-info', date],
-    queryFn: ({ signal }) => fetchWorkflowInfo(date, signal),
-    enabled: Boolean(date),
-  });
+  return useQuery(workflowInfoQueryOptions(date));
 }
diff --git a/packages/app/src/hooks/toggle-logic.test.ts b/packages/app/src/hooks/toggle-logic.test.ts
index 634417db4..e6786de58 100644
--- a/packages/app/src/hooks/toggle-logic.test.ts
+++ b/packages/app/src/hooks/toggle-logic.test.ts
@@ -42,17 +42,29 @@ describe('computeToggle', () => {
     });
   });
 
+  // Clicking an active entry always isolates it, even from a partial
+  // selection. Solo used to require EVERY item to be active, which made it
+  // unreachable wherever something starts deselected — on the agentic chart
+  // the engine guard switches SGLang off on load, so the first legend click
+  // removed the entry instead of isolating it. Removing one entry has its own
+  // control (the X on a legend row).
   describe('when some (but not all/one) items are active', () => {
-    it('removes an active item', () => {
+    it('solos the clicked active item', () => {
       const prev = new Set(['a', 'b']);
       const result = computeToggle(prev, 'a', ALL);
-      expect(result).toEqual(new Set(['b']));
+      expect(result).toEqual(new Set(['a']));
     });
 
-    it('removes the other active item', () => {
+    it('solos the other active item', () => {
       const prev = new Set(['a', 'b']);
       const result = computeToggle(prev, 'b', ALL);
-      expect(result).toEqual(new Set(['a']));
+      expect(result).toEqual(new Set(['b']));
+    });
+
+    it('does not mutate the input set', () => {
+      const prev = new Set(['a', 'b']);
+      computeToggle(prev, 'a', ALL);
+      expect(prev).toEqual(new Set(['a', 'b']));
     });
   });
 
@@ -113,14 +125,14 @@ describe('computeToggle', () => {
       expect(step2).toEqual(ALL);
     });
 
-    it('full cycle: all → solo → add → remove → single → restore', () => {
+    it('full cycle: all → solo → add → re-solo → restore', () => {
       const s1 = computeToggle(new Set(['a', 'b', 'c']), 'a', ALL); // solo a
       expect(s1).toEqual(new Set(['a']));
       const s2 = computeToggle(s1, 'b', ALL); // add b
       expect(s2).toEqual(new Set(['a', 'b']));
-      const s3 = computeToggle(s2, 'a', ALL); // remove a
-      expect(s3).toEqual(new Set(['b']));
-      const s4 = computeToggle(s3, 'b', ALL); // restore all (only b active)
+      const s3 = computeToggle(s2, 'a', ALL); // clicking a active entry re-solos
+      expect(s3).toEqual(new Set(['a']));
+      const s4 = computeToggle(s3, 'a', ALL); // restore all (only a active)
       expect(s4).toEqual(ALL);
     });
   });
diff --git a/packages/app/src/hooks/useChartContext.test.ts b/packages/app/src/hooks/useChartContext.test.ts
index a9289b883..f2f497e2b 100644
--- a/packages/app/src/hooks/useChartContext.test.ts
+++ b/packages/app/src/hooks/useChartContext.test.ts
@@ -1,6 +1,6 @@
 import { describe, it, expect } from 'vitest';
 
-import { reconcileActiveSet } from '@/hooks/useChartContext';
+import { reconcileActiveSet, resolveAvailableSelection } from '@/hooks/useChartContext';
 
 describe('reconcileActiveSet', () => {
   it('initializes with all available items when no previous selection', () => {
@@ -47,3 +47,61 @@ describe('reconcileActiveSet', () => {
     expect(reconcileActiveSet(pruned, full, true)).toBe(pruned);
   });
 });
+
+describe('resolveAvailableSelection', () => {
+  it('waits through transient empty data without consuming URL intent', () => {
+    const active = new Set(['h100']);
+    const result = resolveAvailableSelection({
+      active,
+      available: new Set(),
+      pending: new Set(['b200']),
+      scopeChanged: true,
+      settled: false,
+    });
+    expect(result).toEqual({ selection: active, consumedPending: false });
+  });
+
+  it('restores the valid part of pending URL intent once data settles', () => {
+    const result = resolveAvailableSelection({
+      active: new Set(),
+      available: new Set(['h100', 'b200']),
+      pending: new Set(['b200', 'removed']),
+      scopeChanged: true,
+      settled: true,
+    });
+    expect(result.selection).toEqual(new Set(['b200']));
+    expect(result.consumedPending).toBe(true);
+  });
+
+  it('defaults a new scope to all available items', () => {
+    const available = new Set(['h100', 'b200']);
+    const result = resolveAvailableSelection({
+      active: new Set(['old']),
+      available,
+      scopeChanged: true,
+      settled: true,
+    });
+    expect(result.selection).toBe(available);
+  });
+
+  it('preserves a deliberate empty selection inside the same scope', () => {
+    const active = new Set<string>();
+    const result = resolveAvailableSelection({
+      active,
+      available: new Set(['h100', 'b200']),
+      scopeChanged: false,
+      settled: true,
+    });
+    expect(result.selection).toBe(active);
+  });
+
+  it('clears stale items for a settled empty scope', () => {
+    const result = resolveAvailableSelection({
+      active: new Set(['h100']),
+      available: new Set(),
+      scopeChanged: false,
+      settled: true,
+    });
+    expect(result.selection).toEqual(new Set());
+  });
+});
diff --git a/packages/app/src/hooks/useChartContext.ts b/packages/app/src/hooks/useChartContext.ts
index be095430e..b40a37548 100644
--- a/packages/app/src/hooks/useChartContext.ts
+++ b/packages/app/src/hooks/useChartContext.ts
@@ -28,6 +28,60 @@ export function reconcileActiveSet<T>(
   return filtered;
 }
 
+export interface AvailableSelectionInput<T> {
+  active: Set<T>;
+  available: Set<T>;
+  pending?: Set<T> | null;
+  /** A user-visible scope changed (model, date range, benchmark, etc.). */
+  scopeChanged: boolean;
+  /** False while an empty available set may still be a loading transition. */
+  settled: boolean;
+}
+
+export interface AvailableSelectionResult<T> {
+  selection: Set<T>;
+  consumedPending: boolean;
+}
+
+/**
+ * Shared policy for legend selections backed by asynchronously available data.
+ *
+ * Pending URL intent wins once data settles. A real scope transition starts
+ * fully enabled, while updates inside the same scope only prune stale keys.
+ * An explicitly empty selection is preserved inside a scope rather than being
+ * mistaken for an uninitialized set.
+ */
+export function resolveAvailableSelection<T>({
+  active,
+  available,
+  pending,
+  scopeChanged,
+  settled,
+}: AvailableSelectionInput<T>): AvailableSelectionResult<T> {
+  if (!settled) return { selection: active, consumedPending: false };
+  if (available.size === 0) {
+    return {
+      selection: active.size === 0 ? active : new Set(),
+      consumedPending: pending !== null && pending !== undefined,
+    };
+  }
+
+  if (pending) {
+    const restored = new Set([...pending].filter((item) => available.has(item)));
+    return {
+      selection: restored.size > 0 ? restored : available,
+      consumedPending: true,
+    };
+  }
+
+  if (scopeChanged) return { selection: available, consumedPending: false };
+  if (active.size === 0) return { selection: active, consumedPending: false };
+  return {
+    selection: reconcileActiveSet(active, available, false),
+    consumedPending: false,
+  };
+}
+
 /**
  * Common chart context state and logic shared across all chart types.
  * Extracts duplicated patterns from InferenceChartContext, ReliabilityChartContext,
@@ -116,22 +170,6 @@ export function useChartToggleSet<T extends string = string>() {
   };
 }
 
-/**
- * Automatically initializes a togglable set when available items change.
- * Prevents unnecessary reinitialization when the set is already populated.
- */
-export function useAutoInitializeToggleSet<T>(
-  availableItems: T[],
-  activeSet: Set<T>,
-  setActiveSet: (set: Set<T>) => void,
-) {
-  useEffect(() => {
-    if (availableItems.length > 0 && activeSet.size === 0) {
-      setActiveSet(new Set(availableItems));
-    }
-  }, [availableItems, activeSet.size, setActiveSet]);
-}
-
 /**
  * Syncs state to URL parameters, skipping the initial render to avoid
  * overwriting URL-provided values.
@@ -200,17 +238,3 @@ export function useChartDataFilter<TData, TKey extends string>(
 
   return itemsWithData;
 }
-
-/**
- * Creates a memoized filtered dataset based on active items.
- */
-export function useFilteredData<TData, TKey extends string>(
-  data: TData[],
-  activeSet: Set<TKey>,
-  extractKey: (item: TData) => TKey,
-) {
-  return useMemo(
-    () => data.filter((item) => activeSet.has(extractKey(item))),
-    [data, activeSet, extractKey],
-  );
-}
diff --git a/packages/app/src/hooks/useChartExport.test.ts b/packages/app/src/hooks/useChartExport.test.ts
index 458b222aa..0d8ac88cd 100644
--- a/packages/app/src/hooks/useChartExport.test.ts
+++ b/packages/app/src/hooks/useChartExport.test.ts
@@ -1,7 +1,11 @@
 // @vitest-environment jsdom
 import { describe, expect, it } from 'vitest';
 
-import { getExportFontFamily } from './useChartExport';
+import {
+  getExportCaptureDimensions,
+  getExportFontFamily,
+  normalizeChartSvgWidthsForExport,
+} from './useChartExport';
 
 describe('getExportFontFamily', () => {
   it('uses Minecraft font stack when minecraft theme is active', () => {
@@ -21,3 +25,54 @@ describe('getExportFontFamily', () => {
     expect(getExportFontFamily()).toContain('"Segoe UI"');
   });
 });
+
+describe('getExportCaptureDimensions', () => {
+  it('includes content that overflows the export host to the right', () => {
+    const element = document.createElement('div');
+    Object.defineProperties(element, {
+      clientWidth: { value: 980 },
+      clientHeight: { value: 604 },
+      scrollWidth: { value: 1168 },
+      scrollHeight: { value: 604 },
+    });
+    element.getBoundingClientRect = () => ({ width: 980, height: 604 }) as DOMRect;
+
+    expect(getExportCaptureDimensions(element)).toEqual({ width: 1168, height: 604 });
+  });
+
+  it('rounds fractional layout bounds up so the edge pixel is not cropped', () => {
+    const element = document.createElement('div');
+    Object.defineProperties(element, {
+      clientWidth: { value: 1168 },
+      clientHeight: { value: 604 },
+      scrollWidth: { value: 1168 },
+      scrollHeight: { value: 604 },
+    });
+    element.getBoundingClientRect = () => ({ width: 1168.25, height: 604.5 }) as DOMRect;
+
+    expect(getExportCaptureDimensions(element)).toEqual({ width: 1169, height: 605 });
+  });
+});
+
+describe('normalizeChartSvgWidthsForExport', () => {
+  it('resizes only the chart SVG and leaves UI icons unchanged', () => {
+    const root = document.createElement('div');
+    root.innerHTML = `
+      <svg data-testid="d3-chart-svg" style="width: 800px"></svg>
+      <a href="#"><svg data-testid="external-link-icon" style="width: 12px"></svg></a>
+      <button><svg data-testid="legend-info-icon" style="width: 14px"></svg></button>
+    `;
+
+    normalizeChartSvgWidthsForExport(root);
+
+    expect(root.querySelector<SVGElement>('svg[data-testid="d3-chart-svg"]')?.style.width).toBe(
+      '100%',
+    );
+    expect(
+      root.querySelector<SVGElement>('svg[data-testid="external-link-icon"]')?.style.width,
+    ).toBe('12px');
+    expect(root.querySelector<SVGElement>('svg[data-testid="legend-info-icon"]')?.style.width).toBe(
+      '14px',
+    );
+  });
+});
diff --git a/packages/app/src/hooks/useChartExport.ts b/packages/app/src/hooks/useChartExport.ts
index 6c6158323..6d9496956 100644
--- a/packages/app/src/hooks/useChartExport.ts
+++ b/packages/app/src/hooks/useChartExport.ts
@@ -1,5 +1,7 @@
 import { useCallback, useState } from 'react';
 
+import { CHART_FONT_MINECRAFT, CHART_FONT_SANS } from '@/lib/d3-chart/typography';
+
 interface UseChartExportOptions {
   chartId: string;
   setIsLegendExpanded?: (expanded: boolean) => void;
@@ -24,10 +26,10 @@ export function getExportFontFamily(): string {
       document.body.classList.contains('minecraft'));
 
   if (isMinecraftTheme) {
-    return 'var(--font-minecraft), "Monocraft", monospace';
+    return CHART_FONT_MINECRAFT;
   }
 
-  return 'var(--font-dm-sans), -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif';
+  return CHART_FONT_SANS;
 }
 
 function getResolvedExportFontFamily(): string {
@@ -169,12 +171,39 @@ function waitForRender(): Promise<void> {
   });
 }
 
+/**
+ * html-to-image measures `clientWidth`/`clientHeight`, which exclude content
+ * overflowing the export host. The sidebar legend deliberately grows beyond
+ * that host for long labels, so capture the larger scroll dimensions instead.
+ */
+export function getExportCaptureDimensions(element: HTMLElement): {
+  width: number;
+  height: number;
+} {
+  const bounds = element.getBoundingClientRect();
+  return {
+    width: Math.ceil(Math.max(bounds.width, element.clientWidth, element.scrollWidth)),
+    height: Math.ceil(Math.max(bounds.height, element.clientHeight, element.scrollHeight)),
+  };
+}
+
+/** Keep the plot responsive without stretching small UI icons in the clone. */
+export function normalizeChartSvgWidthsForExport(root: HTMLElement): void {
+  for (const svg of root.querySelectorAll<SVGElement>('svg[data-testid="d3-chart-svg"]')) {
+    svg.style.width = '100%';
+  }
+}
+
+function watermarkFont(size: number): string {
+  return `bold ${size}px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif`;
+}
+
 /** Add a subtle watermark bar at the bottom of the exported image */
 function addWatermark(dataUrl: string, bgColor: string): Promise<string> {
   return new Promise((resolve) => {
     const img = new Image();
     img.addEventListener('load', () => {
-      const WATERMARK_HEIGHT = 48;
+      const WATERMARK_HEIGHT = 240;
       const canvas = document.createElement('canvas');
       canvas.width = img.width;
       canvas.height = img.height + WATERMARK_HEIGHT;
@@ -195,16 +224,20 @@ function addWatermark(dataUrl: string, bgColor: string): Promise<string> {
       ctx.fillStyle = isDark ? '#1a1a2e' : '#f5f5f5';
       ctx.fillRect(0, img.height, canvas.width, WATERMARK_HEIGHT);
 
-      // Draw watermark text
+      // Draw watermark text (shrink to fit on narrow exports)
+      const WATERMARK_TEXT = 'InferenceX — github.com/SemiAnalysisAI/InferenceX';
+      let fontSize = 80;
+      ctx.font = watermarkFont(fontSize);
+      const maxTextWidth = canvas.width - 48;
+      const textWidth = ctx.measureText(WATERMARK_TEXT).width;
+      if (textWidth > maxTextWidth) {
+        fontSize = Math.max(16, Math.floor((fontSize * maxTextWidth) / textWidth));
+        ctx.font = watermarkFont(fontSize);
+      }
       ctx.fillStyle = isDark ? '#aaa' : '#555';
-      ctx.font = 'bold 16px -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif';
       ctx.textAlign = 'center';
       ctx.textBaseline = 'middle';
-      ctx.fillText(
-        'InferenceX — github.com/SemiAnalysisAI/InferenceX',
-        canvas.width / 2,
-        img.height + WATERMARK_HEIGHT / 2,
-      );
+      ctx.fillText(WATERMARK_TEXT, canvas.width / 2, img.height + WATERMARK_HEIGHT / 2);
 
       resolve(canvas.toDataURL('image/png'));
     });
@@ -226,13 +259,14 @@ export function useChartExport({
   const exportToImage = useCallback(async () => {
     setIsExporting(true);
 
-    // Temporarily expand the legend so the clone captures expanded state
-    let wasCollapsed = false;
+    // A fully-closed sidebar legend renders only its reopen button (no
+    // .legend-container), so temporarily open it for the clone and restore
+    // the closed state right after.
+    let wasClosed = false;
     if (setIsLegendExpanded) {
       const el = document.querySelector(`#${chartId}`);
-      const legend = el?.getElementsByClassName('legend-container')[0];
-      wasCollapsed = Boolean(legend) && !legend!.classList.contains('bg-accent');
-      if (wasCollapsed) {
+      wasClosed = Boolean(el?.querySelector('[data-testid="legend-open-button"]'));
+      if (wasClosed) {
         setIsLegendExpanded(true);
         await waitForRender();
       }
@@ -274,8 +308,8 @@ export function useChartExport({
 
       exportElement.append(clone);
 
-      // Restore collapsed state immediately after cloning
-      if (wasCollapsed && setIsLegendExpanded) {
+      // Restore closed state immediately after cloning
+      if (wasClosed && setIsLegendExpanded) {
         setIsLegendExpanded(false);
       }
 
@@ -312,7 +346,7 @@ export function useChartExport({
       // Force legend into inline flow
       if (legendContainer) {
         legendContainer.style.cssText +=
-          '; position: relative !important; right: auto !important; top: auto !important; left: auto !important; bottom: auto !important; width: auto !important; min-width: fit-content !important; z-index: auto !important; overflow: visible !important; padding: 8px !important;';
+          '; position: relative !important; right: auto !important; top: auto !important; left: auto !important; bottom: auto !important; width: auto !important; min-width: fit-content !important; height: auto !important; min-height: 0 !important; max-height: none !important; z-index: auto !important; overflow: visible !important; padding: 8px !important;';
 
         const scrollContainer = legendContainer.querySelector(
           'ul, [class*="overflow"]',
@@ -408,9 +442,7 @@ export function useChartExport({
       for (const span of clone.querySelectorAll('span')) {
         span.style.fontSize = '14px';
       }
-      for (const svg of clone.querySelectorAll('svg')) {
-        svg.style.width = '100%';
-      }
+      normalizeChartSvgWidthsForExport(clone);
 
       // Wait for fonts before capture
       try {
@@ -432,7 +464,9 @@ export function useChartExport({
           // Fallback to @font-face extraction from loaded stylesheets.
         }
       }
+      const captureDimensions = getExportCaptureDimensions(exportElement);
       const chartDataUrl = await toPng(exportElement, {
+        ...captureDimensions,
         quality: 1,
         pixelRatio: 2,
         backgroundColor: bgColor,
@@ -460,7 +494,7 @@ export function useChartExport({
     } catch (error) {
       console.error('Error exporting image:', error);
       alert('Failed to export image. Please try again.');
-      if (wasCollapsed && setIsLegendExpanded) setIsLegendExpanded(false);
+      if (wasClosed && setIsLegendExpanded) setIsLegendExpanded(false);
     } finally {
       setIsExporting(false);
       const exportElement = document.querySelector<HTMLElement>(`#${chartId}-export`);
diff --git a/packages/app/src/hooks/useChartTooltipHandlers.ts b/packages/app/src/hooks/useChartTooltipHandlers.ts
index 3b74872ef..da7a75129 100644
--- a/packages/app/src/hooks/useChartTooltipHandlers.ts
+++ b/packages/app/src/hooks/useChartTooltipHandlers.ts
@@ -1,8 +1,13 @@
 import * as d3 from 'd3';
 import { useCallback } from 'react';
-import { computeTooltipPosition } from '@/lib/d3-chart/layers/scatter-points';
+import {
+  computeTooltipPosition,
+  getTooltipContainerGeometry,
+  invalidateTooltipGeometry,
+} from '@/lib/d3-chart/layers/scatter-points';
 import { useStickyTooltip } from './useStickyTooltip';
 
+import { overlayMarkerPosition } from '@/lib/d3-chart/overlay-x-marker';
 export type RulerType = 'vertical' | 'horizontal' | 'crosshair' | 'none';
 
 export interface ChartTooltipConfig<TData> {
@@ -257,30 +262,42 @@ export function useChartTooltipHandlers<TData>(): ChartTooltipHandlers<TData> {
             .style('display', 'block')
             .style('pointer-events', 'none')
             .html(config.generateTooltipContent(d, false));
+          invalidateTooltipGeometry(tooltipElement.node());
 
-          // Position rulers
+          // Position rulers. During a data transition the bound scales already
+          // describe the destination while the marker is still interpolating.
+          // Prefer the live translated marker position so rulers pass through
+          // what is actually painted.
+          const markerPosition = this instanceof Element ? overlayMarkerPosition(this) : null;
           const { curX: currentXScale, curY: currentYScale } = getZoomedScales();
 
           if (verticalRuler || horizontalRuler) {
             rulerGroup.style('display', 'block');
           }
           if (verticalRuler && config.getRulerX) {
-            const x = config.getRulerX(d, currentXScale);
+            const x = markerPosition?.x ?? config.getRulerX(d, currentXScale);
             verticalRuler.attr('x1', x).attr('x2', x);
           }
 
           if (horizontalRuler && currentYScale && config.getRulerY) {
-            const y = config.getRulerY(d, currentYScale);
+            const y = markerPosition?.y ?? config.getRulerY(d, currentYScale);
             horizontalRuler.attr('y1', y).attr('y2', y);
           }
         })
         .on('mousemove', (event) => {
           if (isPinned()) return;
 
-          const rect = containerElement.getBoundingClientRect();
-          const mx = event.clientX - rect.left;
-          const my = event.clientY - rect.top;
-          const pos = computeTooltipPosition(mx, my, tooltipElement, containerElement);
+          const geometry = getTooltipContainerGeometry(containerElement);
+          const mx = event.clientX - geometry.bounds.left;
+          const my = event.clientY - geometry.bounds.top;
+          const pos = computeTooltipPosition(
+            mx,
+            my,
+            tooltipElement,
+            containerElement,
+            10,
+            geometry,
+          );
           tooltipElement.style('left', `${pos.left}px`).style('top', `${pos.top}px`);
         })
         .on('mouseleave', function (_event, d) {
@@ -297,15 +314,23 @@ export function useChartTooltipHandlers<TData>(): ChartTooltipHandlers<TData> {
           tooltipElement.style('opacity', 0).style('display', 'none');
           rulerGroup.style('display', 'none');
         })
-        .on('click', (event, d) => {
+        .on('click', function (event, d) {
           event.stopPropagation();
 
           // Set content first so dimensions are available for position calc
-          const rect = containerElement.getBoundingClientRect();
-          const mx = event.clientX - rect.left;
-          const my = event.clientY - rect.top;
+          const geometry = getTooltipContainerGeometry(containerElement);
+          const mx = event.clientX - geometry.bounds.left;
+          const my = event.clientY - geometry.bounds.top;
           tooltipElement.html(config.generateTooltipContent(d, true));
-          const pos = computeTooltipPosition(mx, my, tooltipElement, containerElement);
+          invalidateTooltipGeometry(tooltipElement.node());
+          const pos = computeTooltipPosition(
+            mx,
+            my,
+            tooltipElement,
+            containerElement,
+            10,
+            geometry,
+          );
           tooltipElement
             .style('left', `${pos.left}px`)
             .style('top', `${pos.top}px`)
@@ -313,27 +338,22 @@ export function useChartTooltipHandlers<TData>(): ChartTooltipHandlers<TData> {
             .style('display', 'block')
             .style('pointer-events', 'auto');
 
-          // Position rulers at the clicked point
+          const markerPosition = this instanceof Element ? overlayMarkerPosition(this) : null;
           const { curX: currentXScale, curY: currentYScale } = getZoomedScales();
           if (verticalRuler || horizontalRuler) {
             rulerGroup.style('display', 'block');
           }
           if (verticalRuler && config.getRulerX) {
-            const x = config.getRulerX(d, currentXScale);
+            const x = markerPosition?.x ?? config.getRulerX(d, currentXScale);
             verticalRuler.attr('x1', x).attr('x2', x);
           }
           if (horizontalRuler && currentYScale && config.getRulerY) {
-            const y = config.getRulerY(d, currentYScale);
+            const y = markerPosition?.y ?? config.getRulerY(d, currentYScale);
             horizontalRuler.attr('y1', y).attr('y2', y);
           }
 
-          // Pin the tooltip
           pinTooltip(d);
-
-          // Call optional analytics tracking
-          if (config.onPointClick) {
-            config.onPointClick(d);
-          }
+          config.onPointClick?.(d);
         });
     },
     [isPinned, pinTooltip],
diff --git a/packages/app/src/hooks/useClientPathname.ts b/packages/app/src/hooks/useClientPathname.ts
new file mode 100644
index 000000000..8016eab21
--- /dev/null
+++ b/packages/app/src/hooks/useClientPathname.ts
@@ -0,0 +1,56 @@
+'use client';
+
+import { useSyncExternalStore } from 'react';
+
+import { CLIENT_PATHNAME_CHANGE_EVENT, getClientPathnameOverride } from '@/lib/client-navigation';
+
+const subscribers = new Set<() => void>();
+let listening = false;
+
+function emitPathnameChange(): void {
+  for (const subscriber of subscribers) subscriber();
+}
+
+function startListening(): void {
+  if (listening || typeof window === 'undefined') return;
+  window.addEventListener('popstate', emitPathnameChange);
+  window.addEventListener(CLIENT_PATHNAME_CHANGE_EVENT, emitPathnameChange);
+  listening = true;
+}
+
+function stopListening(): void {
+  if (!listening || subscribers.size > 0 || typeof window === 'undefined') return;
+  window.removeEventListener('popstate', emitPathnameChange);
+  window.removeEventListener(CLIENT_PATHNAME_CHANGE_EVENT, emitPathnameChange);
+  listening = false;
+}
+
+function subscribeClientPathname(subscriber: () => void): () => void {
+  subscribers.add(subscriber);
+  startListening();
+  return () => {
+    subscribers.delete(subscriber);
+    stopListening();
+  };
+}
+
+/**
+ * The router pathname, corrected for in-place `replaceClientPathname`
+ * rewrites. `usePathname` deliberately keeps the server-rendered value after
+ * per-model dashboard routes rewrite the URL on a model switch; controls that
+ * map the CURRENT address to a destination (the header language toggle) read
+ * this instead. The override is honored only while the address bar still
+ * matches it, so App Router navigations (which update `routerPathname`) and
+ * popstate transitions retire it naturally — and environments where the
+ * browser URL is not the route (component tests) keep the router value.
+ */
+export function useClientPathname(routerPathname: string): string {
+  return useSyncExternalStore(
+    subscribeClientPathname,
+    () => {
+      const override = getClientPathnameOverride();
+      return override !== null && window.location.pathname === override ? override : routerPathname;
+    },
+    () => routerPathname,
+  );
+}
diff --git a/packages/app/src/hooks/useClientSearch.test.tsx b/packages/app/src/hooks/useClientSearch.test.tsx
new file mode 100644
index 000000000..44e272cb3
--- /dev/null
+++ b/packages/app/src/hooks/useClientSearch.test.tsx
@@ -0,0 +1,80 @@
+// @vitest-environment jsdom
+import { act, createElement } from 'react';
+import { createRoot } from 'react-dom/client';
+import { renderToString } from 'react-dom/server';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+import { CLIENT_SEARCH_CHANGE_EVENT } from '@/lib/client-navigation';
+
+import { useClientSearch } from './useClientSearch';
+
+afterEach(() => {
+  window.history.replaceState({}, '', '/');
+  vi.restoreAllMocks();
+});
+
+function SearchProbe({ id }: { id: string }) {
+  const search = useClientSearch();
+  return createElement('output', { id }, search);
+}
+
+describe('useClientSearch', () => {
+  it('uses the empty server snapshot during hydration rendering', () => {
+    window.history.replaceState({}, '', '/inference?unofficialrun=123');
+    expect(renderToString(createElement(SearchProbe, { id: 'server' }))).toContain(
+      '<output id="server"></output>',
+    );
+  });
+
+  it('shares one pair of browser listeners across subscribers', () => {
+    const addEventListener = vi.spyOn(window, 'addEventListener');
+    const container = document.createElement('div');
+    const root = createRoot(container);
+    act(() => {
+      root.render(
+        createElement(
+          'div',
+          null,
+          createElement(SearchProbe, { id: 'first' }),
+          createElement(SearchProbe, { id: 'second' }),
+        ),
+      );
+    });
+
+    expect(addEventListener.mock.calls.filter(([event]) => event === 'popstate')).toHaveLength(1);
+    expect(
+      addEventListener.mock.calls.filter(([event]) => event === CLIENT_SEARCH_CHANGE_EVENT),
+    ).toHaveLength(1);
+    act(() => root.unmount());
+  });
+
+  it('updates every subscriber for explicit search writes and back/forward events', () => {
+    const container = document.createElement('div');
+    const root = createRoot(container);
+    act(() => {
+      root.render(
+        createElement(
+          'div',
+          null,
+          createElement(SearchProbe, { id: 'first' }),
+          createElement(SearchProbe, { id: 'second' }),
+        ),
+      );
+    });
+
+    act(() => {
+      window.history.pushState({}, '', '/inference?unofficialruns=1,2');
+      window.dispatchEvent(new CustomEvent(CLIENT_SEARCH_CHANGE_EVENT));
+    });
+    expect(container.querySelector('#first')?.textContent).toBe('?unofficialruns=1,2');
+    expect(container.querySelector('#second')?.textContent).toBe('?unofficialruns=1,2');
+
+    act(() => {
+      window.history.replaceState({}, '', '/inference?unofficialruns=1');
+      window.dispatchEvent(new PopStateEvent('popstate'));
+    });
+    expect(container.querySelector('#first')?.textContent).toBe('?unofficialruns=1');
+    expect(container.querySelector('#second')?.textContent).toBe('?unofficialruns=1');
+    act(() => root.unmount());
+  });
+});
diff --git a/packages/app/src/hooks/useClientSearch.ts b/packages/app/src/hooks/useClientSearch.ts
new file mode 100644
index 000000000..51fd0c849
--- /dev/null
+++ b/packages/app/src/hooks/useClientSearch.ts
@@ -0,0 +1,61 @@
+'use client';
+
+import { useMemo, useSyncExternalStore } from 'react';
+
+import { CLIENT_SEARCH_CHANGE_EVENT } from '@/lib/client-navigation';
+
+const subscribers = new Set<() => void>();
+let listening = false;
+
+function emitSearchChange(): void {
+  for (const subscriber of subscribers) subscriber();
+}
+
+function startListening(): void {
+  if (listening || typeof window === 'undefined') return;
+  window.addEventListener('popstate', emitSearchChange);
+  window.addEventListener(CLIENT_SEARCH_CHANGE_EVENT, emitSearchChange);
+  listening = true;
+}
+
+function stopListening(): void {
+  if (!listening || subscribers.size > 0 || typeof window === 'undefined') return;
+  window.removeEventListener('popstate', emitSearchChange);
+  window.removeEventListener(CLIENT_SEARCH_CHANGE_EVENT, emitSearchChange);
+  listening = false;
+}
+
+/** One shared external store for the browser's current query string. */
+export function subscribeClientSearch(subscriber: () => void): () => void {
+  subscribers.add(subscriber);
+  startListening();
+  return () => {
+    subscribers.delete(subscriber);
+    stopListening();
+  };
+}
+
+export function getClientSearchSnapshot(): string {
+  return typeof window === 'undefined' ? '' : window.location.search;
+}
+
+export function getServerSearchSnapshot(): string {
+  return '';
+}
+
+/**
+ * Hydration-safe live query-string snapshot. History writes remain explicit;
+ * callers that use pushState/replaceState must dispatch CLIENT_SEARCH_CHANGE_EVENT.
+ */
+export function useClientSearch(): string {
+  return useSyncExternalStore(
+    subscribeClientSearch,
+    getClientSearchSnapshot,
+    getServerSearchSnapshot,
+  );
+}
+
+export function useClientSearchParams(): URLSearchParams {
+  const search = useClientSearch();
+  return useMemo(() => new URLSearchParams(search), [search]);
+}
diff --git a/packages/app/src/hooks/useOpenDropdown.ts b/packages/app/src/hooks/useOpenDropdown.ts
new file mode 100644
index 000000000..b6b78bf8a
--- /dev/null
+++ b/packages/app/src/hooks/useOpenDropdown.ts
@@ -0,0 +1,20 @@
+'use client';
+
+import { useCallback, useState } from 'react';
+
+/** Coordinates a group of controlled dropdowns so opening one closes the previous dropdown. */
+export function useOpenDropdown<DropdownKey extends string>() {
+  const [openDropdown, setOpenDropdown] = useState<DropdownKey | null>(null);
+  const handleDropdownOpenChange = useCallback(
+    (dropdownKey: DropdownKey) => (isOpen: boolean) => {
+      if (isOpen) {
+        setOpenDropdown(dropdownKey);
+        return;
+      }
+      setOpenDropdown((current) => (current === dropdownKey ? null : current));
+    },
+    [],
+  );
+
+  return { openDropdown, handleDropdownOpenChange } as const;
+}
diff --git a/packages/app/src/hooks/useUrlState.ephemeral.test.tsx b/packages/app/src/hooks/useUrlState.ephemeral.test.tsx
new file mode 100644
index 000000000..1995a3f86
--- /dev/null
+++ b/packages/app/src/hooks/useUrlState.ephemeral.test.tsx
@@ -0,0 +1,103 @@
+// @vitest-environment jsdom
+import { act } from 'react';
+import { createRoot, type Root } from 'react-dom/client';
+import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
+
+const mocks = vi.hoisted(() => ({
+  writeUrlParams: vi.fn(),
+  refreshUrlParamsOnNavigation: vi.fn(() => true),
+  readUrlParams: vi.fn(() => ({}) as Record<string, string>),
+}));
+
+vi.mock('next/navigation', () => ({
+  usePathname: () => '/model/test-model',
+}));
+
+vi.mock('@/lib/url-state', () => ({
+  readUrlParams: mocks.readUrlParams,
+  refreshUrlParamsOnNavigation: mocks.refreshUrlParamsOnNavigation,
+  writeUrlParams: mocks.writeUrlParams,
+}));
+
+import { EphemeralUrlStateContext, useEphemeralUrlState, useUrlState } from './useUrlState';
+
+let root: Root | undefined;
+let container: HTMLDivElement | undefined;
+let hookValue: ReturnType<typeof useUrlState> | undefined;
+let ephemeralValue: boolean | undefined;
+
+function Probe() {
+  hookValue = useUrlState();
+  ephemeralValue = useEphemeralUrlState();
+  return null;
+}
+
+function mount(ui: React.ReactElement): void {
+  container = document.createElement('div');
+  document.body.append(container);
+  root = createRoot(container);
+  act(() => root?.render(ui));
+}
+
+beforeEach(() => {
+  (globalThis as Record<string, unknown>).IS_REACT_ACT_ENVIRONMENT = true;
+  mocks.writeUrlParams.mockClear();
+  mocks.readUrlParams.mockReturnValue({});
+});
+
+afterEach(() => {
+  act(() => root?.unmount());
+  container?.remove();
+  root = undefined;
+  container = undefined;
+  hookValue = undefined;
+  ephemeralValue = undefined;
+});
+
+describe('useUrlState outside an ephemeral scope', () => {
+  it('writes params to the shared store', () => {
+    mount(<Probe />);
+    expect(ephemeralValue).toBe(false);
+
+    act(() => hookValue?.setUrlParam('i_gpus', 'B200_vllm'));
+    expect(mocks.writeUrlParams).toHaveBeenCalledWith({ i_gpus: 'B200_vllm' });
+
+    act(() => hookValue?.setUrlParams({ g_model: 'Kimi-K3', i_seq: 'agentic-traces' }));
+    expect(mocks.writeUrlParams).toHaveBeenCalledWith({
+      g_model: 'Kimi-K3',
+      i_seq: 'agentic-traces',
+    });
+  });
+});
+
+describe('useUrlState inside an ephemeral scope', () => {
+  // The /model/[slug] embed reuses the dashboard providers, whose URL-sync
+  // effects write every state change (auto-selected chip configs included)
+  // into the module-scoped share-link store. Inside the ephemeral scope those
+  // writes must be dropped, or Back-navigating to a bare /inference rebuilds
+  // with the embed's auto-populated state instead of the user's own.
+  it('drops writes so embed state cannot leak into the dashboard', () => {
+    mount(
+      <EphemeralUrlStateContext.Provider value={true}>
+        <Probe />
+      </EphemeralUrlStateContext.Provider>,
+    );
+    expect(ephemeralValue).toBe(true);
+
+    act(() => hookValue?.setUrlParam('i_gpus', 'B200_vllm,GB200_dynamo'));
+    act(() => hookValue?.setUrlParams({ g_model: 'Kimi-K3' }));
+    expect(mocks.writeUrlParams).not.toHaveBeenCalled();
+  });
+
+  it('still reads params (share links into the embed keep working)', () => {
+    mocks.readUrlParams.mockReturnValue({ i_gpus: 'MI355X_atom' });
+    mount(
+      <EphemeralUrlStateContext.Provider value={true}>
+        <Probe />
+      </EphemeralUrlStateContext.Provider>,
+    );
+
+    expect(hookValue?.getUrlParam('i_gpus')).toBe('MI355X_atom');
+    expect(hookValue?.hasUrlParam('i_gpus')).toBe(true);
+  });
+});
diff --git a/packages/app/src/hooks/useUrlState.ts b/packages/app/src/hooks/useUrlState.ts
index b440350ff..180ae1d46 100644
--- a/packages/app/src/hooks/useUrlState.ts
+++ b/packages/app/src/hooks/useUrlState.ts
@@ -1,22 +1,65 @@
 'use client';
 
-import { useCallback, useRef } from 'react';
+import { usePathname } from 'next/navigation';
+import { createContext, useCallback, useContext, useRef } from 'react';
 
 import {
   type UrlStateKey,
   type UrlStateParams,
   readUrlParams,
+  refreshUrlParamsOnNavigation,
   writeUrlParams,
 } from '@/lib/url-state';
 
+/**
+ * Marks a subtree whose chart providers must treat the shared URL-state store
+ * as READ-ONLY.
+ *
+ * The store in `url-state.ts` is module-scoped and deliberately outlives
+ * client-side navigations, so the dashboard's filters survive a round-trip
+ * through other pages. Embedded chart instances reuse the same providers
+ * (`GlobalFilterProvider` → `InferenceProvider`), so without this scope their
+ * seeded state — e.g. the `/model/[slug]` embed's auto-selected "every chip
+ * config with data" — is written into the store and becomes the state a bare
+ * `/inference` rebuilds from on Back-navigation, silently replacing whatever
+ * the user actually had selected.
+ *
+ * Reads stay live inside the scope: explicit share-link params on the embed's
+ * own URL keep working, and writes made by the primary dashboards are still
+ * visible. Only writes (including `rememberChartStateInUrl` stamping — see the
+ * chart components) are suppressed.
+ */
+export const EphemeralUrlStateContext = createContext(false);
+
+/** Whether the calling component is inside an ephemeral URL-state scope. */
+export function useEphemeralUrlState(): boolean {
+  return useContext(EphemeralUrlStateContext);
+}
+
 /**
  * React hook for URL state synchronization.
  * Reads URL params once on mount (cached in ref), and provides
  * functions to write params back to the URL.
+ *
+ * Inside an `EphemeralUrlStateContext` scope the write functions are no-ops
+ * (see the context doc above); reads are unaffected.
  */
 export function useUrlState() {
   const initialParams = useRef<UrlStateParams | null>(null);
 
+  // The load-time snapshot in `url-state.ts` is captured once per document, so
+  // on a client-side navigation every provider below would initialise from the
+  // params of the page the user came FROM. Refresh during render — before the
+  // `useState` initialisers and mount effects of the providers underneath —
+  // and only once per navigation, so a component mounting later on the same
+  // path cannot re-import the URL over filter changes made since.
+  //
+  // `usePathname`, not `useSearchParams`: the latter forces every statically
+  // prerendered page that mounts a filter provider behind a Suspense boundary,
+  // which fails the build on /ai-chart.
+  const pathname = usePathname();
+  refreshUrlParamsOnNavigation(pathname ?? '');
+
   // read URL params only once (synchronous, before first render)
   if (initialParams.current === null) {
     initialParams.current = readUrlParams();
@@ -32,13 +75,23 @@ export function useUrlState() {
     [],
   );
 
-  const setUrlParam = useCallback((key: UrlStateKey, value: string) => {
-    writeUrlParams({ [key]: value });
-  }, []);
+  const ephemeral = useContext(EphemeralUrlStateContext);
 
-  const setUrlParams = useCallback((params: UrlStateParams) => {
-    writeUrlParams(params);
-  }, []);
+  const setUrlParam = useCallback(
+    (key: UrlStateKey, value: string) => {
+      if (ephemeral) return;
+      writeUrlParams({ [key]: value });
+    },
+    [ephemeral],
+  );
+
+  const setUrlParams = useCallback(
+    (params: UrlStateParams) => {
+      if (ephemeral) return;
+      writeUrlParams(params);
+    },
+    [ephemeral],
+  );
 
   return {
     initialParams: initialParams.current,
diff --git a/packages/app/src/lib/agentic-catalog.test.ts b/packages/app/src/lib/agentic-catalog.test.ts
new file mode 100644
index 000000000..899ff6ba9
--- /dev/null
+++ b/packages/app/src/lib/agentic-catalog.test.ts
@@ -0,0 +1,88 @@
+import type { AgenticCatalogEntry } from '@semianalysisai/inferencex-db/queries/agentic-catalog';
+import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
+import { describe, expect, it } from 'vitest';
+
+import { groupAgenticCatalog } from './agentic-catalog';
+
+function entry(overrides: Partial<AgenticCatalogEntry> = {}): AgenticCatalogEntry {
+  return {
+    id: 1,
+    model: 'dsv4',
+    hardware: 'b200',
+    framework: 'vllm',
+    precision: 'fp4',
+    points: 10,
+    minConc: 1,
+    maxConc: 64,
+    latestDate: '2026-08-19',
+    ...overrides,
+  };
+}
+
+describe('groupAgenticCatalog', () => {
+  it('groups configs under their model and resolves registry labels', () => {
+    const groups = groupAgenticCatalog([
+      entry({ id: 1, hardware: 'b200', framework: 'vllm' }),
+      entry({ id: 2, hardware: 'h200', framework: 'sglang', points: 5 }),
+    ]);
+
+    expect(groups).toHaveLength(1);
+    const [group] = groups;
+    expect(group.key).toBe('dsv4');
+    // Same label the AgentX ledger uses, so the two catalogs agree.
+    expect(group.label).toBe('DeepSeekv4 Pro 0813 1.6T');
+    expect(group.totalPoints).toBe(15);
+    expect(group.cards.map((card) => card.label)).toEqual(['B200 · vLLM', 'H200 · SGLang']);
+    expect(group.cards[0]).toMatchObject({
+      id: 1,
+      hardwareKey: 'b200',
+      precisionLabel: 'FP4',
+      vendor: 'NVIDIA',
+      arch: 'Blackwell',
+    });
+  });
+
+  it('orders cards by the chart hardware sort, not insertion order', () => {
+    const groups = groupAgenticCatalog([
+      entry({ id: 1, hardware: 'h100' }),
+      entry({ id: 2, hardware: 'b200' }),
+      entry({ id: 3, hardware: 'h200' }),
+    ]);
+
+    const keys = groups[0].cards.map((card) => card.hardwareKey);
+    const sorts = keys.map((key) => HW_REGISTRY[key].sort);
+    expect(sorts).toEqual([...sorts].sort((a, b) => a - b));
+  });
+
+  it('leads with the model carrying the most telemetry points', () => {
+    const groups = groupAgenticCatalog([
+      entry({ id: 1, model: 'dsv4', points: 3 }),
+      entry({ id: 2, model: 'kimik3', points: 40 }),
+      entry({ id: 3, model: 'glm5.2', points: 12 }),
+    ]);
+
+    expect(groups.map((group) => group.key)).toEqual(['kimik3', 'glm5.2', 'dsv4']);
+  });
+
+  it('drops rows whose model or hardware is unknown to the registries', () => {
+    // A card we cannot label is a card the reader cannot act on. Dropping beats
+    // rendering "undefined · vLLM" or a raw DB key.
+    const groups = groupAgenticCatalog([
+      entry({ id: 1, model: 'not-a-model' }),
+      entry({ id: 2, hardware: 'not-a-gpu' }),
+      entry({ id: 3 }),
+    ]);
+
+    expect(groups).toHaveLength(1);
+    expect(groups[0].cards.map((card) => card.id)).toEqual([3]);
+  });
+
+  it('keeps a model/SKU/engine pair with two precisions as two cards', () => {
+    const groups = groupAgenticCatalog([
+      entry({ id: 1, model: 'qwen3.5', precision: 'fp4' }),
+      entry({ id: 2, model: 'qwen3.5', precision: 'fp8' }),
+    ]);
+
+    expect(groups[0].cards.map((card) => card.precisionLabel)).toEqual(['FP4', 'FP8']);
+  });
+});
diff --git a/packages/app/src/lib/agentic-catalog.ts b/packages/app/src/lib/agentic-catalog.ts
new file mode 100644
index 000000000..6433b97c9
--- /dev/null
+++ b/packages/app/src/lib/agentic-catalog.ts
@@ -0,0 +1,132 @@
+/**
+ * Server-side helper for the `/inference/agentic` telemetry catalog.
+ *
+ * Turns the raw per-config rows from `getAgenticCatalog` into the
+ * model-grouped, label-resolved shape the page renders. Mirrors how
+ * `compare-availability.ts` feeds the `/compare` index: a cached DB read on
+ * the server, a fixture in e2e mode, and all presentation mapping done here
+ * rather than in the query.
+ *
+ * Rows whose model or hardware key is unknown to the registries are dropped —
+ * a config we cannot label is a card the reader cannot act on, and the
+ * catalog is navigation, not a data export.
+ */
+
+import { FIXTURES_MODE, getDb } from '@semianalysisai/inferencex-db/connection';
+
+import {
+  type AgenticCatalogEntry,
+  getAgenticCatalog,
+} from '@semianalysisai/inferencex-db/queries/agentic-catalog';
+
+import { FRAMEWORK_LABELS, HW_REGISTRY } from '@semianalysisai/inferencex-constants';
+
+import { cachedQuery } from '@/lib/api-cache';
+import { COMPARE_MODEL_SLUGS } from '@/lib/compare-slug';
+import { loadFixture } from '@/lib/test-fixtures';
+
+/**
+ * The fixture is regenerated by hand rather than by `capture:fixtures`: the
+ * catalog has no public API route to fetch, so there is nothing for the
+ * capture script to hit. 47 rows of pure metadata, so it stays cheap to refresh.
+ */
+export const getCachedAgenticCatalog = cachedQuery(() => {
+  if (FIXTURES_MODE) return Promise.resolve(loadFixture<AgenticCatalogEntry[]>('agentic-catalog'));
+
+  return getAgenticCatalog(getDb());
+}, 'agentic-catalog');
+
+export interface AgenticCatalogCard {
+  /** `/inference/agentic/<id>` — the representative point for this config. */
+  id: number;
+  /** "B200 · vLLM" — what the card leads with. */
+  label: string;
+  hardwareKey: string;
+  hardwareLabel: string;
+  frameworkLabel: string;
+  precisionLabel: string;
+  vendor: string;
+  arch: string;
+  points: number;
+  minConc: number;
+  maxConc: number;
+  latestDate: string;
+}
+
+export interface AgenticCatalogModelGroup {
+  /** DB model key, used as the section anchor. */
+  key: string;
+  /** "Kimi K3 2.8T" — the same label the AgentX ledger shows. */
+  label: string;
+  cards: AgenticCatalogCard[];
+  /** Trace-backed points across every config in this model. */
+  totalPoints: number;
+}
+
+/** DB model key → the compare catalog's display label, so the telemetry
+ *  catalog and the AgentX ledger name the same model the same way. */
+function buildModelLabels(): Map<string, string> {
+  const labels = new Map<string, string>();
+  for (const model of COMPARE_MODEL_SLUGS) {
+    for (const dbKey of model.dbKeys) labels.set(dbKey, model.label);
+  }
+  return labels;
+}
+
+/** Precision keys are short and uppercase in the UI ("fp4" → "FP4"). */
+function precisionLabel(precision: string): string {
+  return precision.toUpperCase();
+}
+
+export function groupAgenticCatalog(rows: AgenticCatalogEntry[]): AgenticCatalogModelGroup[] {
+  const modelLabels = buildModelLabels();
+  const groups = new Map<string, AgenticCatalogModelGroup>();
+
+  for (const row of rows) {
+    const modelLabel = modelLabels.get(row.model);
+    const hw = HW_REGISTRY[row.hardware];
+    if (!modelLabel || !hw) continue;
+
+    const frameworkLabel = FRAMEWORK_LABELS[row.framework] ?? row.framework;
+    const group = groups.get(row.model) ?? {
+      key: row.model,
+      label: modelLabel,
+      cards: [],
+      totalPoints: 0,
+    };
+    group.cards.push({
+      id: row.id,
+      label: `${hw.label} · ${frameworkLabel}`,
+      hardwareKey: row.hardware,
+      hardwareLabel: hw.label,
+      frameworkLabel,
+      precisionLabel: precisionLabel(row.precision),
+      vendor: hw.vendor,
+      arch: hw.arch,
+      points: row.points,
+      minConc: row.minConc,
+      maxConc: row.maxConc,
+      latestDate: row.latestDate,
+    });
+    group.totalPoints += row.points;
+    groups.set(row.model, group);
+  }
+
+  const ordered = [...groups.values()];
+  // Cards follow the chart's hardware sort order so a reader scanning the
+  // catalog meets SKUs in the same sequence the legend uses.
+  for (const group of ordered) {
+    group.cards.sort(
+      (a, b) =>
+        (HW_REGISTRY[a.hardwareKey]?.sort ?? 0) - (HW_REGISTRY[b.hardwareKey]?.sort ?? 0) ||
+        a.label.localeCompare(b.label),
+    );
+  }
+  // Models with the most telemetry lead, so the richest sections are on top.
+  ordered.sort((a, b) => b.totalPoints - a.totalPoints || a.label.localeCompare(b.label));
+  return ordered;
+}
+
+export async function getAgenticCatalogGroups(): Promise<AgenticCatalogModelGroup[]> {
+  return groupAgenticCatalog(await getCachedAgenticCatalog());
+}
diff --git a/packages/app/src/lib/agentic-detail-link.test.ts b/packages/app/src/lib/agentic-detail-link.test.ts
new file mode 100644
index 000000000..9b877c81a
--- /dev/null
+++ b/packages/app/src/lib/agentic-detail-link.test.ts
@@ -0,0 +1,74 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+
+function setupWindow(search = '', pathname = '/inference') {
+  vi.stubGlobal('window', {
+    location: { search, pathname, hash: '', origin: 'https://example.com' },
+    history: { replaceState: vi.fn(), state: null },
+  });
+}
+
+describe('agenticDetailHref', () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+    vi.resetModules();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+  });
+
+  it('keeps a Chinese reader on /zh', async () => {
+    setupWindow('', '/zh/inference');
+    const { agenticDetailHref } = await import('@/lib/agentic-detail-link');
+
+    // Without the prefix the /zh chart handed readers to the English detail
+    // page, and every link out of that page kept them in English.
+    expect(agenticDetailHref(440106, 'zh')).toBe('/zh/inference/agentic/440106');
+    expect(agenticDetailHref(440106, 'en')).toBe('/inference/agentic/440106');
+  });
+
+  it('defaults to the English path when no locale is given', async () => {
+    setupWindow('', '/inference');
+    const { agenticDetailHref } = await import('@/lib/agentic-detail-link');
+    expect(agenticDetailHref(440106)).toBe('/inference/agentic/440106');
+  });
+
+  it('carries the chart state so the detail page can link back to it', async () => {
+    setupWindow('', '/inference');
+    const { agenticDetailHref } = await import('@/lib/agentic-detail-link');
+    const { writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3', i_active: 'gb200_dynamo-vllm' });
+    const href = agenticDetailHref(440106, 'en');
+    const params = new URLSearchParams(href.slice(href.indexOf('?')));
+
+    expect(href.startsWith('/inference/agentic/440106?')).toBe(true);
+    expect(params.get('g_model')).toBe('Kimi-K3');
+    expect(params.get('i_active')).toBe('gb200_dynamo-vllm');
+  });
+
+  it('carries an unofficial-run overlay so the round trip keeps the overlay', async () => {
+    setupWindow('?unofficialruns=987654321', '/inference');
+    const { agenticDetailHref } = await import('@/lib/agentic-detail-link');
+    const { writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3' });
+    const href = agenticDetailHref(440106, 'zh');
+    const params = new URLSearchParams(href.slice(href.indexOf('?')));
+
+    expect(href.startsWith('/zh/inference/agentic/440106?')).toBe(true);
+    expect(params.get('unofficialruns')).toBe('987654321');
+  });
+
+  it('stays a bare path when the chart is on defaults', async () => {
+    setupWindow('', '/inference');
+    const { agenticDetailHref } = await import('@/lib/agentic-detail-link');
+    const { writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'DeepSeek-V4-Pro' });
+    await vi.advanceTimersByTimeAsync(200);
+
+    expect(agenticDetailHref(440106, 'en')).toBe('/inference/agentic/440106');
+  });
+});
diff --git a/packages/app/src/lib/agentic-detail-link.ts b/packages/app/src/lib/agentic-detail-link.ts
new file mode 100644
index 000000000..a497f904d
--- /dev/null
+++ b/packages/app/src/lib/agentic-detail-link.ts
@@ -0,0 +1,24 @@
+import { ZH_PREFIX, type Locale } from '@/lib/i18n';
+import { withChartState } from '@/lib/url-state';
+
+/**
+ * In-app href for an agentic point's detail page.
+ *
+ * Two things the bare `/inference/agentic/<id>` path got wrong:
+ *
+ *  - It dropped the `/zh` prefix, so a Chinese reader clicking a point landed
+ *    on the English detail page and every link out of it kept them there.
+ *  - It carried no chart state. These links are plain `<a href>` (they must
+ *    support open-in-new-tab), so following one is a full-document navigation
+ *    that destroys the in-memory filter store. Without the params in the URL
+ *    the detail page has no way to send the reader back to the chart they
+ *    left, and it opens on the default model with the default legend.
+ *
+ * The params are stripped from the address bar again on load (see
+ * `url-state.ts`), so the detail page still shows a clean URL — they survive
+ * in the in-memory store, which is what `withChartState` reads back when the
+ * detail page builds its "Inference chart" link.
+ */
+export function agenticDetailHref(pointId: number, locale: Locale = 'en'): string {
+  return withChartState(`${locale === 'zh' ? ZH_PREFIX : ''}/inference/agentic/${pointId}`);
+}
diff --git a/packages/app/src/lib/agentx-optimizations-zh.ts b/packages/app/src/lib/agentx-optimizations-zh.ts
new file mode 100644
index 000000000..b580cdff9
--- /dev/null
+++ b/packages/app/src/lib/agentx-optimizations-zh.ts
@@ -0,0 +1,611 @@
+/**
+ * Simplified Chinese port of `agentx-optimizations.ts`.
+ *
+ * Translations are keyed by framework slug and section id, and carry only the
+ * prose: PR references, figure assets, and link targets come from the English
+ * registry, so a PR list can never drift between the two languages.
+ * `agentx-optimizations.test.ts` fails if a section, highlight, paragraph, or
+ * link here does not line up with its English counterpart.
+ */
+
+import {
+  AGENTX_OPTIMIZATION_FRAMEWORKS,
+  OPTIMIZATIONS_OVERVIEW,
+  type OptimizationFigureCopy,
+  type OptimizationFramework,
+  type OptimizationLayer,
+  type OptimizationSection,
+  type OptimizationsOverview,
+} from './agentx-optimizations';
+import type { Locale } from './i18n';
+
+interface SectionTranslation {
+  heading: string;
+  paragraphs: readonly string[];
+  figure?: OptimizationFigureCopy;
+  /** Link labels, in the same order as the English section's `links`. */
+  links?: readonly string[];
+}
+
+interface FrameworkTranslation {
+  summary: string;
+  lead: string;
+  /** Highlight labels, in order. The values stay in English (units, numbers). */
+  highlights: readonly string[];
+  sections: Readonly<Record<string, SectionTranslation>>;
+}
+
+const OVERVIEW_ZH = {
+  eyebrow: 'AgentX 的行业影响',
+  title: '面向智能体负载的优化',
+  lead: 'AgentX 推出后的最初几个月，最有价值的成果并非开源数据集，而是 AgentX 合作伙伴提交的 50 多个上游 PR。这些改动以 AgentX 为北极星，面向真实智能体负载进行优化。',
+  intro: [
+    'AgentX 回放的是真实智能体流量，因此它衡量的远不止 prefill 与 decode 内核本身，而是端到端的完整链路：KV cache 生命周期、混合注意力（hybrid attention）cache 正确性、CPU KV offload、传输进度、路由亲和性、增量 tokenization、请求序列化以及调度器记账。这些开销每一项都是生产环境智能体部署已经在付的成本，而单轮 8k/1k 场景一项也测不出来。',
+    'SemiAnalysis 与 AMD 软件团队多年协作、推动其软件开发方式现代化，下面的许多工作正是这一协作的成果，它们让 AMD 开源栈在智能体负载上更接近一流水平。',
+  ],
+  highlights: ['上游 PR', '上游项目', '涉及的栈层级', '作为北极星的 trace'],
+  frameworksTitle: '按项目查看优化',
+  frameworksIntro:
+    '下面每个页面汇总一个项目中由 AgentX 驱动的工作，并按其所处的栈层级归类。文中标注为 open 的改动只是已提出、尚未合入，不代表上游当前行为。',
+  sections: {
+    ecosystem: {
+      heading: '分布式推理生态简介',
+      paragraphs: [
+        '智能体推理是系统级问题，而不是单纯的芯片或内核问题。当分布式系统需要处理数十万个智能体请求时，请求调度与 KV cache 管理就不再只是记账工作，而会实实在在地影响性能。例如 subagent 会产生突发式的 KV cache 访问模式，若不加优化，就会把主 agent 的 cache 挤掉。',
+        '在栈的最上层，router（有时也称 frontend）把请求分发给各个 worker。当服务端启用数据并行注意力（data-parallel attention）时，每个 DP rank 都有独立的 KV cache，因此请求会按 consistent hashing 等策略路由：同一会话或同一 subagent 的请求按其唯一 ID 落到同一个 rank，而不是把所有 cache 都搅乱。',
+        '大多数路由策略在不同实现之间差别不大。有些 router 是独立组件，例如 vLLM router 和 llm-d router；有些则集成在引擎内部，例如 SGLang model gateway 和 ATOM Mesh。',
+        '完成路由后，请求进入 vLLM、SGLang 等推理引擎的调度器。引擎负责执行推理并通过 API 返回结果，同时提供接口把引擎内部的 KV cache 接到外部 KV cache 管理器上——这正是该生态可插拔的原因：一个 cache 管理器可以对接多种引擎。',
+        '当前 AgentX 结果所用的简单部署方式，是让 Mooncake 与 vLLM 运行在同一节点上。每个 vLLM worker 内嵌一个 Mooncake Store 客户端，并贡献一部分主机 DRAM 给外部 KV cache 池。vLLM 通过 MooncakeStoreConnector 接口访问这个池：把可复用的 KV block 载入 GPU 显存，并把新计算出的 block 写回主机内存。Mooncake Store 负责放置与淘汰策略，Mooncake Transfer Engine 负责在 GPU 与 CPU 内存之间实际搬运数据。',
+        '不同的 KV cache 管理器会使用不同的传输引擎，在内存层级之间或机器之间（例如 prefill 与 decode worker 之间）搬运字节。一套部署可以用 Mooncake Store 把可复用的 KV block offload 到主机 DRAM，同时用 NIXL 把请求专属的 KV 直接从 prefill GPU 传给 decode GPU；在支持的环境下 NIXL 会走 UCX 与 GPUDirect RDMA。多条 KV 管理与传输路径可以在同一个推理引擎中共存。',
+        '整个生态由许多独立组件构成：推理引擎、router、KV cache 管理器、数据传输库以及集群控制器。NVIDIA Dynamo、llm-d、AMD Infera 等平台把其中若干组合打包成完整发行版，提供彼此兼容的容器镜像、connector、部署清单和编排逻辑。最终交付的通常是一组协同工作的容器，而不是单一的巨型服务。',
+      ],
+      figure: {
+        alt: '分布式推理栈的五层示意图：路由/前端层、推理引擎、KV cache 管理层、数据搬运层，以及加速器与分层存储。',
+        caption:
+          '分布式推理栈。每一层都有多个可互换的实现，Dynamo、llm-d、AMD Infera 等平台会把其中若干组合打包成一套发行版。',
+      },
+      links: [
+        'Consistent hash 路由策略',
+        'vLLM router',
+        'llm-d router',
+        'SGLang model gateway',
+        'ATOM Mesh',
+        'Mooncake Transfer Engine',
+      ],
+    },
+    other: {
+      heading: '其他优化：day-zero 支持与正确性',
+      paragraphs: [
+        '各项目页面汇总的工作主要针对长上下文开销：prefix 必须活下来、hybrid cache 必须保持正确、传输必须跟得上。这一节的改动性质不同，它们属于 day-zero 支持与正确性缺陷——它们让一个请求彻底失败的程度，和百万 token 会话遇到的问题一样严重。',
+        'MiniMax-M3 检验了这些 ROCm 工作能否累积成 day-zero 就绪能力，Advancing AI 一文直接给出了对比：AMD 首个公开的 disaggregated 配方 MI355X FP4 在 1 月才登陆 InferenceX，比 NVIDIA 晚了数月；而 M3 FP4 disaggregation 在 day zero 就已就位。相比 DeepSeek-R1 时期需要数月才追平，这是明显进步。day-zero 路径上有三个 vLLM 修复，且都属于正确性问题而非性能问题。',
+        '首先被卡住的是 disaggregation。NixlConnector 的握手逻辑假设 SPLIT 区域的 block_len 会随 prefill 与 decode 的 TP 比例缩放，但 block_len 实际取决于每个 rank 的 KV head 数量。M3 只有 4 个 KV head，因此 TP4 prefill 搭配 TP8 decode 时，两侧都被 GQA 限制为每 rank 一个 head，两个长度相等，而断言却要求相差两倍。握手因此被拒绝，KV 完全没有传输，decode 只能从零重新生成，gsm8k 得分为 0。改为按真实 head 比例校验后问题解决。',
+        '另外两个问题源于平台差异。M3 的稀疏注意力后端在所有 E4M3 配置下都把按字节存储的 FP8 cache 读作 float8_e4m3fn，但 gfx942 的平台 dtype 是 e4m3fnuz；两种编码并不相同，因此 K 和 V 在进入内核前就已被改变，而且 prefill 与 decode 的 wrapper 也漏掉了 FP8 检查中的 FNUZ 类型。改为按平台 dtype 解释 cache 视图后，两处一并修复。此外，M3 的 NVIDIA 与 AMD 模型实现位于不同文件，只有 NVIDIA 版实现了 EAGLE3 接口，导致 ROCm 上的 speculative decoding 在引擎初始化时就以 model-does-not-support 报错退出。把 AMD 版补齐到同等能力后恢复正常，MI355X 的 gsm8k 与未启用 EAGLE3 的 MI355X 运行以及 B200 结果一致。',
+        'TensorRT-LLM 页面覆盖了 M3 中与长上下文直接相关的工作：disaggregated KV 传输中的 descriptor 爆炸、context graph 捕获、稀疏 block stride、autotuner 候选集，以及必须从候选池中移除的错误 split-K MoE tactic。',
+      ],
+      links: ['Can AMD break the CUDA moat? — Advancing AI'],
+    },
+    'what-activates-this': {
+      heading: 'AgentX 测试矩阵激活了什么',
+      paragraphs: [
+        '本地 AgentX 矩阵同时包含会话感知或 KV 感知路由、长度多变的长对话历史、MTP、hybrid attention、聚合式与 disaggregated 部署，以及跨越 HBM 容量悬崖的并发扫描；既有全部驻留 GPU 的对比，也有通过 vLLM SimpleCPU、Mooncake、LMCache 和 SGLang HiCache 进行的 CPU DRAM offload。正是这种组合激活了上面这些上游工作。',
+        '旧的固定序列矩阵通常只创建一个 prompt、执行一次 prefill、解码固定长度的续写，然后丢弃请求。因此它无法衡量 cache 跨轮次的存活、重复 tokenization、会话亲和性、cache 事件流量、offload 抖动、调度器停顿期间的传输进度，以及长生命周期的归属记账。',
+        '允许的优化策略把 CPU KV offload 视为可选项。厂商可以使用 vLLM connector、LMCache、SGLang HiCache、Mooncake、Dynamo KVBM 或其他 CPU DRAM connector；如果关闭 offload 能得到更好的延迟与吞吐量组合，也可以关闭。NVMe offload 暂缓。CPU DRAM 必须按所用 GPU 比例缩放，其中非标准化 DRAM 系统上限为 3 TB；标准化 DRAM 系统没有硬性上限，但同样遵守比例规则。本地生成器目前对所有 runner 一律套用 3 TB 上限，因此尚未实现标准化 DRAM 的例外条款。',
+        '新增的优化面并不只是"更长的注意力"，而是不断增长的会话状态在保存、搬运、路由、重建和反复处理过程中的全部开销。AgentX 把这些成本放大到足以推动 vLLM、SGLang、TensorRT-LLM、ATOM、AITER、Dynamo 和 LMCache 做出通用的上游改动。对 NIXL 与 Mooncake 的直接检索没有发现此处之外带 AgentX 标记的运行时 PR，因此它们的影响仍通过上述引擎 connector 改动体现。',
+      ],
+      links: ['所有提及 AgentX 的上游 PR'],
+    },
+  } as Readonly<Record<string, SectionTranslation>>,
+  layerLabels: {
+    engine: '推理引擎',
+    router: 'Router 与编排',
+    'kv-cache': 'KV cache 层',
+    kernels: '内核',
+    transfer: '传输引擎',
+  } as Readonly<Record<OptimizationLayer, string>>,
+  ui: {
+    backToAgentX: '← 返回 AgentX',
+    backToOverview: '← 全部 AgentX 优化',
+    prsLabel: '上游 PR',
+    referencesLabel: '参考链接',
+    readMore: '查看优化详情',
+    onThisPage: '本页目录',
+    allProjects: '其他项目',
+    figureCta: '查看原始分辨率图片',
+    prSearch: '检索全部 AgentX 相关 PR',
+  },
+};
+
+const FRAMEWORKS_ZH: Readonly<Record<string, FrameworkTranslation>> = {
+  vllm: {
+    summary:
+      '混合注意力 prefix 保留、面向 hybrid 模型的 CPU KV offload，以及收窄后的 store 与 load 路径。',
+    lead: '我们与来自 Inferact、Red Hat、NVIDIA 和 AMD 的 vLLM maintainer 一起，以 AgentX 的真实回放器为北极星开展工作，相关修复均已合入上游，其中大部分可以直接迁移到生产环境。',
+    highlights: [
+      '1M 上下文下的 prefix cache 命中率',
+      'hybrid CPU offload 带来的输出吞吐量提升',
+      '平均端到端延迟下降',
+    ],
+    sections: {
+      'hybrid-prefix-caching': {
+        heading: '混合注意力 prefix caching',
+        paragraphs: [
+          'vLLM 改进了 hybrid attention 的 prefix caching，使短生命周期的 sliding-window 分配不再挤掉有价值的长上下文 checkpoint。选择性保留（selective retention）会保住稀疏的回放边界，在 14 个并发请求、上下文最长 100 万 token 的条件下，报告 prefix cache 命中率超过 95%。',
+          '同样的可达性策略也应用到了 Mooncake，并移除了不可达的 sliding-window 查找。更早的后续工作还停止 offload 那些永远不会被复用的 sliding-window block，并把 speculative lookahead block 保留在被保住的 prefix 内。',
+        ],
+        figure: {
+          alt: '前后对比示意图：未做保留时每个 sliding-window 尾部都会被释放，整个 prefix 必须重算；采用选择性保留后少量 checkpoint 得以存活，prefix 仍可复用。',
+          caption:
+            '改动前没有任何 window 尾部存活，因此没有位置可恢复，full-attention KV 虽然驻留但无法使用；采用选择性保留后少量尾部得以保留，prefix cache 命中率超过 95%。',
+        },
+      },
+      'cpu-offload': {
+        heading: '面向 hybrid 模型的 CPU KV offload',
+        paragraphs: [
+          '高并发智能体负载必须依赖 offload，而 AgentX 推动的工作让 CPU KV offload 从只支持统一 full-attention 模型，扩展到可用于 hybrid 模型。这一区别很关键：统一模型每个 token 只有一种 KV 布局，connector 用单一 block 几何形状就能描述要保存什么；hybrid 模型同时携带多个 cache group，形状与生命周期各不相同，而假设单一布局的 connector 无法表达某个 block 属于哪一组。结果就是，最需要 offload 的长会话模型反而用不了 offload。',
+          '通用的 SimpleCPU connector 最先落地，随后在 ROCm 上启用，并进一步扩展到 DeepSeek-V4 的 hybrid attention：与 prefix 超出 HBM 后直接重算相比，报告输出吞吐量提高 81.7%、平均端到端延迟降低 46.6%。Mooncake 也获得了等价的 hybrid 内存分配支持。',
+        ],
+      },
+      'store-path': {
+        heading: '收窄 store 路径',
+        paragraphs: [
+          '在真实负载上做 profiling 后发现，offload 一旦能跑通，开销就转移到了 store 路径上——写得太多、也太频繁。三条规则收窄了它。',
+          '现在，若已有相同内容的传输在途，就跳过这次 store，于是共享同一 prefix 的并发会话只需付一次代价，而不是各付一次；store 只覆盖新生成的 KV 区间，因此延长历史的会话只写增量，而不是每轮重写整个 prefix；store 也不再取决于相同 block 是否仍在 HBM 中，因此已经排定的工作不会因为底下发生淘汰而被丢弃。',
+        ],
+        figure: {
+          alt: 'vLLM connector 层位于 GPU HBM 与 CPU DRAM 池之间的示意图，标注了相关 connector PR 以及收窄写入的三条 store 侧规则。',
+          caption:
+            '位于 GPU HBM 与 CPU DRAM 池之间的 connector 层。store 侧规则会跳过在途的重复传输、只写新生成的 KV，并与 block 是否仍在 HBM 解耦；调度器侧的查找则改为异步。',
+        },
+      },
+      'load-path': {
+        heading: '把查找移出关键路径',
+        paragraphs: [
+          'load 路径需要单独调优，因为查找发生在每一次调度决策上，而不只是在真正搬运数据时。把调度器路径中的查找改为异步，可以让 connector 离开单步执行的关键路径，一个 step 不必等待 CPU 侧 cache 查询就能开始接纳新工作。',
+          '此后，紧凑的零拷贝查找 key、接收侧并行加载，以及预先构建的 Mooncake key 字符串，进一步消除了残留的 CPU 与传输开销。',
+        ],
+      },
+      correctness: {
+        heading: '长生命周期 hybrid 状态下的正确性与记账',
+        paragraphs: [
+          '长生命周期的 hybrid 状态还带来了一批固定形状请求很少触及的正确性与记账修复。vLLM 现在按 hybrid cache group 分别发出 cache 事件，正确处理分布式 context 下 store 的 stride，并在分布式 context 与 prefill 下正确计算查找 prefix；相关的 context-parallel 记账改动则让 cache 归属与分片后的 token 区间对齐。',
+          'Speculative 状态现在会在合并后的 Mooncake group 之间以及 SimpleCPU coordinator 中正确传递，避免多轮复用的 cache 悄悄丢失 EAGLE 状态。',
+        ],
+      },
+      'rocm-decode': {
+        heading: 'ROCm：cache 层之下',
+        paragraphs: [
+          '在 ROCm 侧，近期工作继续深入到 cache 层以下，那里的开销是按层而非按请求计的。当 prefix 能够存活并及时到达之后，剩下的就是 decode 步骤本身；而一次会话中要执行数千次的 decode，会为每一次可以避免的拷贝、每一个不匹配的内核付出代价。三个仍为 open 的改动针对的正是这一层。',
+          '一个 Kimi-K3 改动把 KDA decode 结果直接写入该层的输出 buffer，为每个 KDA 层省去一次设备内拷贝；单看收益很小，但它会在每个解码 token 的每一层上重复发生。第二个改动用 AITER 的 sparse-MLA decode 内核替代通用路径，报告 AgentX 输出吞吐量提升 5.22%，token 间延迟大幅下降。',
+          '第三个改动很好地说明了"测量形状"有多重要：配套改动把 full-graph 注意力投影切换到调优过的 AITER GEMM，在低并发固定序列上取得 2.3% 的提升。内核级改动可以在统一形状上给出干净的收益，却在智能体 trace 上被 cache 与调度波动淹没。',
+        ],
+      },
+    },
+  },
+  sglang: {
+    summary:
+      'Sliding-window 分配、HiCache 混合 offload、以 runtime scalar 传入上下文长度，以及 cache 感知的 DP 路由。',
+    lead: '我们与来自 RadixArk、Meta、NVIDIA 和 AMD 的 SGLang maintainer 合作，在 AgentX 计划推动下完成了一批面向智能体负载的优化，这些改动同样显著提升了生产环境的推理服务表现。',
+    highlights: [
+      '并发 384 下的输出吞吐量提升',
+      '同一改动带来的平均 TTFT 下降',
+      'staging 修复后正确命中的 needle 数（此前为 2/128）',
+    ],
+    sections: {
+      'sliding-window': {
+        heading: 'Sliding-window 分配',
+        paragraphs: [
+          'SGLang 的 sliding-window 工作处理的冲突与 vLLM 的保留策略相同，只是从分配器一侧入手。window 页与 prefix 页取自同一个池，而 window 是更贪婪的消费者：它不断翻新，prefix 却长期不动，于是在压力下这种临时分配会挤掉本应长期保留的部分。',
+          '三种设计从不同角度解决该问题。其一，在页面离开 window 时就主动释放，而不是等淘汰压力来找它们，让失效的 window 状态不再争抢它已经用不上的页面。其二，把 compute lock 限制在单个 window 内，从而限定一个在途请求最多能钉住多少池空间。其三，清除已失去价值的陈旧 full-KV 条目。',
+          '与之并列的 ROCm ring-cache 修复属于正确性而非容量问题：ring buffer 的设计本身就会复用槽位，而复用一个旧内容仍被引用的槽位，得到的是错误输出而不是变慢的输出。这些问题在单个 8k prompt 上完全看不出来——window 永远不会绕过 prefix，池也不会有竞争；但在多轮 hybrid 会话中，它们决定了昂贵的 full-attention 历史在下一轮是否还在。',
+        ],
+      },
+      hicache: {
+        heading: 'HiCache offload',
+        paragraphs: [
+          'HiCache 是 SGLang 内置的一流 offload 机制。它面对的 hybrid 问题与 vLLM connector 相同，但解法是一种不对称设计：offload full-attention cache，而在回载时重建较短的 sliding-window 尾部。只有昂贵的那一半值得跨总线搬运，便宜的一半重建比取回更快。在 AMD 上，分阶段写回让这种搬运不会阻塞引擎。',
+          '剩下的缺口是 recurrent 状态——它无法像 window 尾部那样由相邻 token 重建。FlashInfer 的 GDN checkpoint 让它得以参与 prefix 复用，在 92.4% cache 命中率下把吞吐量从 47,771 提升到 53,004 tok/s/GPU。',
+        ],
+        figure: {
+          alt: '示意图展示三类状态从 GPU HBM 到 CPU DRAM 的处理方式：full-attention KV 按字节 offload，sliding-window 尾部不搬运、在加载时重建，recurrent 状态则做 checkpoint。',
+          caption:
+            'HiCache 的不对称设计。full-attention KV 体积大且无法重建，因此跨总线搬运；window 尾部重建成本低，所以留在原地；recurrent 状态虽小却无法重建，因此以 checkpoint 形式保存。',
+        },
+      },
+      'variable-length': {
+        heading: '长度多变的流量对内核流水线意味着什么',
+        paragraphs: [
+          '与生产流量一样，AgentX 会话的上下文长度连续变化。如果运行时天真地按长度做特化，几乎每来一个请求就要编译一个新内核。SGLang maintainer 的解决办法是把上下文长度作为 runtime scalar 传入，从而收敛为一次编译：AgentX 并发 384 下输出吞吐量提升 26.75%，平均 TTFT 降低 36.25%——收益来自消除编译，而不是把计算做得更快。',
+          '同理，移除每步一次的 device-to-host 序列长度同步，也消除了一个 decode 气泡；这个气泡的唯一成因，就是 host 想知道一个设备端早已掌握的长度。',
+        ],
+      },
+      routing: {
+        heading: 'Cache 感知的数据并行路由',
+        paragraphs: [
+          '当请求不带任何可复用历史（例如某个 subagent 刚启动）时，发给哪个 worker 都行，唯一值得考虑的就是负载均衡；但当它带着上兆字节的已缓存 prefix 时，把它发给一个并不持有该 prefix 的空闲 worker 就是昂贵选择，router 必须知道状态究竟在哪里。',
+          'SGLang 加入了 DP cache affinity，让会话粘在持有其 cache 的 rank 上。同一批工作还实现了 DP 感知的 prefill 与 decode 路由，使 disaggregated 部署的两侧做出一致决策，并把 cache 均衡作为路由信号，避免亲和性退化成单个热点 worker。router 只能依据它被告知的信息行动，因此 hybrid cache 事件也变得 radix cache 感知与 sliding-window 感知。',
+        ],
+      },
+      speculative: {
+        heading: 'Speculative decoding',
+        paragraphs: [
+          'Speculative decoding 需要特别关注，因为 MTP 会引入第二份、体量更小的每请求状态，而它必须与主 cache 一样在各种情况下存活。SGLang 修复了 disaggregated 部署中的 draft window 传输，使该状态完整跨越 prefill 到 decode 的边界；为高并发在线 decode 增加了 overlap 调度；移除了一处无实际作用的 EAGLE 重归一化；并避免了 EAGLE prefill 期间的 host 同步。',
+          '仍为 open 的 resource-lease 调度工作与数据并行 graph metadata 修复延续同一方向：当请求可能被撤回并恢复、而不是一路跑完时，让 overlap 依然安全。',
+        ],
+      },
+      staging: {
+        heading: '异构 disaggregation 下的 prefix 感知 staging',
+        paragraphs: [
+          '异构的 prefill 与 decode 拓扑需要 prefix 感知的 staging，而这正是 prefix caching 与 disaggregation 相互冲突之处。当两侧分片方式不同时，KV 无法作为一条连续流拷贝过去，必须按传输网格切分，并在 decode 侧期望的偏移处重新拼装。prefix 命中会让这件事更难而不是更容易，因为 prefill worker 此时只发送未命中的剩余部分，decode 侧却仍然期望得到一份完整、位置正确的 cache。staging buffer 中的 radix cache 支持会按该网格切分已缓存的发送内容，并散布到正确的 decode 偏移上。',
+          '它修复的是一个正确性问题。一项 127,500 token 的共享 prefix 测试，从 128 根 needle 只答对 2 根变为 128 根全对，说明此前 cache 一直悄悄落在错误位置——而吞吐量基准测试只会把它记成一次又快又自信的错误回答。相应的 AgentX 对比还把单用户输出吞吐量中位数提高了 9.61%，而每 GPU 总吞吐量几乎不变。',
+          '同一条线上仍有 open 的工作：UMBP 中的多池 DeepSeek-V4 支持、经由 MoRI 承载的统一 KV HiSparse 状态，以及在 decode 未产生可见内容就终止时保留 prefill 侧持有的 token。HiSparse 相关工作应被理解为长上下文的容量与正确性使能，而不是高并发下的吞吐量收益——它目前还不是。',
+        ],
+        figure: {
+          alt: 'staging buffer 索引的前后对比：改动前未命中的剩余部分覆盖到 prefix 所在的错误行；改动后它被放到 decode 侧已持有 prefix 之后的正确偏移上。',
+          caption:
+            '改动前 staging 索引从 0 开始计数，忽略 decode 侧已经持有的内容，于是剩余部分覆盖了 prefix；采用 radix 感知 staging 后，已缓存的发送按传输网格切分，并散布到正确的 decode 偏移上。',
+        },
+      },
+    },
+  },
+  'tensorrt-llm': {
+    summary:
+      '边界感知的增量 tokenization、disaggregated KV 的 descriptor 合并，以及调度器生命周期修复。',
+    lead: 'TensorRT-LLM 的 AgentX 工作从前端开始——那里有一项只因负载是多轮才存在的开销——随后沿着请求路径深入到传输粒度、内核选择与调度器生命周期。',
+    highlights: [
+      '每轮 tokenization 平均耗时',
+      '并发 5 时请求关键路径 KV 的 p99',
+      'context graph 带来的单用户输出吞吐量提升',
+    ],
+    sections: {
+      'incremental-tokenization': {
+        heading: '边界感知的增量 tokenization',
+        paragraphs: [
+          '对话每一轮都会把完整历史再加一点新内容重新发送一遍，而朴素实现会把这些内容整体重新 tokenize。按千字节计算 tokenization 很便宜，但当同样的 100,000 个 token 每轮都被重新处理一次时，代价就变得难以承受。',
+          '看上去显而易见的修法——只 tokenize 新增的后缀——却错在一个容易被忽视的地方：BPE 并非与位置无关。token 可能跨接缝合并，因此在边界处切分文本再把两段 token 序列拼接，可能得到与整体 tokenize 不同的序列，从而悄悄偏离 prefix cache 所依据的序列。',
+          'TensorRT-LLM 实现了边界感知的增量 tokenization：先找到渲染文本的公共前缀，回退一个完整 token 以便重新计算任何跨接缝的合并，再从该处只对变化的后缀做 tokenize。在 Qwen3.5 的 AgentX trace 上，它在全部 1,087 次转换中都与完整 tokenize 结果一致——正确性是被验证的，而不是被假设的——平均处理时间从 185.1 ms 降到 11.3 ms。固定的 8k/1k 请求没有上一轮渲染结果可复用，因此完全体现不出这一点。',
+          '相关地，chat template 渲染被移入输入处理线程池，长模板不再让主请求循环排在它后面串行等待。',
+        ],
+      },
+      'disaggregated-kv': {
+        heading: 'Disaggregated KV 搬运与 descriptor 粒度',
+        paragraphs: [
+          'MiniMax-M3 的工作聚焦于 disaggregated KV 搬运，那里的问题出在粒度上。当 prefill 与 decode 对 head 布局理解不一致时，一个逻辑请求的 KV 就不再是少数几块大的连续区域，而变成数千个带 stride 的小片段，每一片都会生成自己的传输 descriptor。搬运的字节数没变，爆炸的是每个 descriptor 的开销，而且恰恰在最需要关注的长 prompt 上最严重。',
+          '修正后的多池映射与分块 NIXL bounce 路径把这些碎片通过一块有界可复用的 arena 合并起来，以一次额外的 staging 拷贝换取数量级更少的 descriptor。AgentX 诊断显示：并发 5 时请求关键路径 KV 的 p99 从 26.74 秒降到 125 ms，并发 40 时从 10.15 秒降到 288 ms。',
+          '非阻塞的 context 传输轮询保护同一条路径：即使调度停顿，也能及时回收已完成的传输，打破"已完成的 KV block 仍被钉住、导致无法接纳新请求"的反馈回路。另有一个 draft cache 传输提案已关闭且未合入，不应算作 TensorRT-LLM 已交付的行为。',
+        ],
+      },
+      'execution-paths': {
+        heading: '面向不规则长上下文的执行路径',
+        paragraphs: [
+          'TensorRT-LLM 还把不规则的长上下文工作迁移到更高效的执行路径上。面向 MiniMax-M3 的 context graph producer 捕获稳定的稀疏 producer，同时让依赖请求的注意力保持 eager 执行，在其 AgentX 测试中单用户输出吞吐量提升 12.58%。另有一个仍为 open 的原生 KV 事件生成改动，减少了 KV 感知路由路径上的内存分配与格式转换开销。',
+        ],
+      },
+      'kernel-selection': {
+        heading: '内核选择与调度器生命周期',
+        paragraphs: [
+          'AgentX 还暴露出只有在规模和持续时间上去之后才会出现的内核选择与调度器生命周期问题。其中两个关乎"选中了哪个内核"。MiniMax-M3 为 MXFP8 autotuning 增加了 CuTeDSL 候选，扩大候选集后，在低并发聚合点上每 GPU 输出吞吐量提升约 7% 到 10%。反方向的例子是：TensorRT-LLM 在错误的 split-K MoE tactic 导致七次 AgentX 运行中有五次崩溃后禁用了它们，之后七次对照运行无一崩溃。一个又快又错的 tactic 比单纯慢的更糟，而 autotuner 只要不把它从候选池中剔除，就会兴高采烈地选中它。',
+          '另外两个属于生命周期缺陷，这类问题是长时间运行（而非大规模运行）的典型失败模式。序列槽位余量与一致的按槽位索引的 buffer 尺寸，处理的是"一个请求即将完成、另一个刚被接纳、两者同时需要槽位"的短暂重叠——持续的到达与离开会不断撞上这个窗口，而固定批次永远撞不到。后续的 attention 数据并行 dummy request 修复，让九个 Qwen3.5 disaggregated 单元保持存活，而此前多数单元几分钟内就会失败：这正是"能跑完基准测试的配置"与"能撑过一次会话的配置"之间的差别。',
+        ],
+      },
+      'pipelined-transfer': {
+        heading: '面向超长 prompt 的流水线 KV 传输',
+        paragraphs: [
+          '两个仍为 open 的传输改动针对超长的 disaggregated prompt，它们放在一起还说明了一个修复如何制造出下一个瓶颈。在默认安排下，decode worker 必须等整段 prompt 完成 prefill 并传输完毕才能开始，于是两个昂贵阶段前后串行，尽管第一个阶段本可以增量产出。流水线 KV 传输会在每个 prefill 分块完成时就开始发送，让传输与 prefill 计算重叠，只有最后一个分块留在关键路径上。',
+          '该改动使分块处理变得频繁，从而暴露出原本只做一次的工作。其后续改动改为只取当前分块所属的 block ID，而不是每次都取整段 prompt 的 block 列表。对于切成 1,024 token 分块的 128,000 token prompt，这就是"构建一次 4,096 条目的列表"与"为每个层组重建 128 次"的差别——一项随 prompt 总长增长的按分块开销，正好会吃掉流水线刚刚换来的收益。',
+        ],
+        figure: {
+          alt: '时间线示意图：改动前四个 prefill 分块全部完成后才整体传输 prompt，decode 随后才开始；改动后每个完成的分块边传输边计算下一块，decode 更早开始。',
+          caption:
+            '改动前 prefill 与传输前后串行，decode 要等两者都完成；采用流水线传输后，每个完成的分块在下一块计算期间发送，只有最后一个分块位于关键路径上。',
+        },
+      },
+    },
+  },
+  atom: {
+    summary:
+      '稀疏 checkpoint 保留、recurrent 状态 checkpoint、CPU offload 归属管理，以及长 prefill 并行。',
+    lead: 'AMD 的 ATOM 引擎当初是为单轮负载设计的，而不是面向真实世界的多轮智能体生产流量，因此支持长上下文多轮负载需要对核心引擎及其内核做大量改动。相比 vLLM 与 SGLang，ATOM 在智能体负载上仍有很长的路要走，而 AgentX 正是其重构的真实北极星目标。',
+    highlights: [
+      '并发 48 下的 prefix 命中率',
+      'sliding-window 关卡处的丢弃率',
+      '采用分块 PP prefill 后的 TTFT 中位数',
+    ],
+    sections: {
+      'sparse-checkpoints': {
+        heading: '稀疏 checkpoint 保留',
+        paragraphs: [
+          'ATOM 为 DeepSeek-V4 的分页 sliding-window attention 实现了稀疏 checkpoint 保留，这说明该问题属于负载本身的性质，而不是某个代码库特有的。已合入的实现会保住选定的 window 尾部，使分支与回放请求能够在有意义的边界处恢复。',
+          '它的测量把两种效应清楚地区分开：在同一条 AgentX trace、并发 48 的条件下，实际 prefix 命中率从 5.6% 升到 96.45%，sliding-window 关卡处的丢弃率从 91.35% 降到 0.16%。后一个数字正是前一个数字背后的机制——十次 prefix 匹配里有九次被找到后又因缺少 window 尾部而被丢弃，也就是说 cache 并不是没命中，而是命中后被否决了。',
+        ],
+      },
+      'cache-manager': {
+        heading: '必须先落地的 cache 管理器修复',
+        paragraphs: [
+          '在上述效果能够被测量之前，还有两个更早的 cache 管理器修复必须先合入；它们都是"cache 自报健康却什么也没做"的典型例子。其一，避免 free pool 命中破坏共享的 cache 条目；其二是一处 deferred-output 修复，它在默认调度模式下恢复了 prefix hash，使重复的长 prompt 从零缓存 token 变为可复用每一个完整的 prefix block。',
+          '另有一个改动让命中 prefix 的 prefill 继续走优化过的 sink attention 内核，而不是回退到通用路径，使得一次 cache 命中不会悄悄抵消掉它节省的一部分开销。',
+        ],
+      },
+      'recurrent-state': {
+        heading: 'Recurrent 状态 checkpoint',
+        paragraphs: [
+          'Hybrid 模型还携带 recurrent 或 compressor 状态，它与普通 KV 有一个决定性差异：无法由周围的 token 重建。window 尾部可以由相邻上下文重算，但 recurrent 状态是此前全部内容累积的结果，一旦丢弃，唯一的办法就是重放整个序列。',
+          'ATOM 为这份每请求状态引入了内容寻址的 checkpoint 生命周期，让已生成的轮次留下可复用的恢复点，而无需为它们单独预留一块受保护的 cache。在一次测试中，某个请求复用了 512 个已生成 token，只需计算两个 token 的后缀。',
+          '调优细节与特性本身同样重要。无条件发布 checkpoint 会在零命中流量上损失 17.5% 吞吐量——这是所有不会再回来的会话为那些会回来的会话付出的代价。按 token 间隔发布 checkpoint 避免了这项惩罚，且固定 1k/1k 的吞吐量变化仍在测量噪声范围内；这正是关键的安全性质：一个面向智能体复用的特性，不应向永远用不到它的负载收税。',
+        ],
+      },
+      'cpu-path': {
+        heading: 'CPU offload 路径：归属与索引位置',
+        paragraphs: [
+          'ATOM 与 AgentX 相关的 CPU 路径，首先要看 offload 本身是否划算。独立运行的 LMCache offload 从 CPU 重新载入 32,000 token 的 prefix 约需 0.32 秒，而重算约需 2.5 秒，相差八倍。这说明在这种上下文长度下跨总线搬运是值得的，而在短 prompt 上并不成立。',
+          '路径的其余部分关乎归属与索引位置，而不是带宽。ATOM 移植了 vLLM 的多 connector 设计，使 prefill worker 可以同时把 KV 发给远端 decode worker 并把同一 prefix 保存到 CPU，且在两个消费方都完成之前不释放这些 block。同一批 block 有两个独立读者，这是单轮流量永远不会出现的情形。',
+          '把恢复回来的 block 重新提升进 GPU prefix 索引，修复的是一种更隐蔽的浪费：如果不这样做，从 CPU 载入的 prefix 用过之后并不会被登记为常驻，于是下一轮又要跨总线取一次同样的热点 prefix，为一份其实已在 HBM 中的 cache 反复付出传输代价。后续工作一并修复了异步保存顺序、packed KV 几何布局、非对齐交接以及远端请求记账，在两轮共 2,638 个请求的验证中消除了重载损坏。这个缺陷只有在同一批 block 被反复保存、淘汰、恢复许多次时才会显现。',
+        ],
+        figure: {
+          alt: '流程示意图：从 CPU DRAM 恢复到 GPU block 的 prefix，要么被移出 GPU prefix 索引（导致下一轮再次取回），要么被提升进索引（下一轮直接在 HBM 命中）。',
+          caption:
+            '重新载入 32,000 token 的 prefix 约需 0.32 秒，而重算约需 2.5 秒。把恢复回来的 block 提升进 GPU prefix 索引，才能让下一轮不必再付一次传输开销。',
+        },
+      },
+      routing: {
+        heading: 'ATOM Mesh 中的 cache 感知路由',
+        paragraphs: [
+          '分布式路径在另一个代码库里重演了 SGLang 与 Dynamo 已经展示过的规律：路由必须知道状态在哪里。ATOM 的 router ATOM Mesh 是 SGLang router 的一个精简 fork，砍掉了大部分特性——其中就包括后来发现必不可少的 cache 感知路由。',
+          'ATOM 因此补上了面向 cache 感知 router 的 KV 生命周期事件（router 首先要能知道状态在哪）、多节点的 prefill 与 decode 路由，以及会话粘性的数据并行路由。这套粘性策略是一个值得明说的双向折中：会话会回到持有其状态且健康的 worker，但空闲的绑定关系会过期，从而避免为早已离开的会话长期牺牲集群均衡。',
+        ],
+      },
+      disaggregation: {
+        heading: 'Disaggregation 必须搬运模型真正保留的东西',
+        paragraphs: [
+          'Disaggregation 必须搬运模型真正保留的状态，而它并不总是一份统一的 cache。DeepSeek-V4 会传输其 FP8 与 BF16 混合布局的两个 buffer；EAGLE disaggregation 则在目标 cache 之外一并搬运 draft 模型独立的 KV cache——这正是 TensorRT-LLM 与 SGLang 各自也必须解决的"第二份 cache"问题。',
+          '远端 KV 的准入与背压机制补上了最后一环：阻止 decode 侧接受超出其安全恢复能力的挂起传输，这相当于 disaggregated 形态下的"接下自己完不成的活"。',
+        ],
+      },
+      'long-prefill': {
+        heading: '面向长 prefill 的并行',
+        paragraphs: [
+          '长 prefill 会用到固定 8k prompt 难以充分锻炼的并行能力，因为 8k 本身可切分的余地有限，TTFT 也已经很短。Prefill context parallelism（PCP）把 DeepSeek-V4 的 query token 切分到多张 GPU 上，报告平均首 token 时间降低 35% 到 43%，在 64,000 token 输入下总吞吐量提升最高约 49%——这种收益随输入长度增长，而不是随 batch size 增长。',
+          '要在实践中可用，它还必须与会话依赖的其他机制共存，因此 decode context parallelism 被改造为兼容 prefix caching、chunked prefill 与 FP8 KV，随后又扩展到 MTP。无法与 prefix cache 共存的并行策略，只会用一种长上下文收益换掉另一种。',
+          '分块流水线并行 prefill 从内存一侧解决同一问题，用流式的层级交接取代反复的张量并行集合通信。它在高负载下的 GLM-5.2 结果是本节中最完整的一组：输出吞吐量翻倍，首 token 时间中位数从 28.6 秒降到 8.7 秒，每张 prefill GPU 可容纳的 KV block 数量提升到 3.68 倍。最后这个数字应当最先看，因为每张 prefill GPU 的容量决定了在部署撞上 HBM 悬崖之前，可以同时承载多少个长会话。',
+        ],
+      },
+    },
+  },
+  aiter: {
+    summary: 'Context-parallel 进程组、面向大 cache 池的 64 位寻址，以及常驻式 MLA decode 内核。',
+    lead: 'ATOM、AMD vLLM 与 AMD SGLang 的长上下文执行都依赖与之匹配的底层 AITER 内核，因为引擎层的并行策略只有在内核能够表达时才是真实可用的。',
+    highlights: [
+      '加宽行索引后可覆盖的 token 数',
+      '超过 4 GB 的池所需的寻址位宽',
+      '可避免静默访问错误行的行数量级',
+    ],
+    sections: {
+      'context-parallel': {
+        heading: 'Context-parallel 进程组',
+        paragraphs: [
+          'Prefill 的 context-parallel 进程组提供了 prefill context parallelism 所需的额外 query 分片维度，同时把融合内核的行索引加宽到可支持 131,000 token 以上的 prompt。Decode context parallelism（DCP）则把 KV 切分到已有的张量并行 GPU 上，使更长的序列或更大的 batch 无需在每个 rank 上复制整份 cache 就能容纳。',
+        ],
+      },
+      'address-width': {
+        heading: '地址位宽：短请求几乎永远碰不到的失败',
+        paragraphs: [
+          '大 cache 暴露出一类短固定请求基本碰不到的失败：地址位宽。在单个 cache 池跨过边界之前，32 位偏移完全够用；一旦越界，运算就会回绕，内核会访问到错误的行，而且不会抛出任何错误。',
+          'AITER 为超过 4 GB 的 batch prefill 增加了运行时 64 位分发，为超过 2 GB 的场景提供 64 位 MLA 偏移，并在 DeepSeek-V4 的统一 cache 路径中全面采用 64 位寻址——最后一项避免了在约 1.5 亿行的池中静默读写到错误的行。',
+        ],
+        figure: {
+          alt: '约 1.5 亿行的统一 KV 池的两幅对比图：使用 32 位偏移时，超过 4 GB 的访问会回绕到另一行；使用 64 位偏移时，访问到的正是目标行。',
+          caption:
+            '在单个 cache 池跨过 4 GB 之前，32 位偏移都够用；一旦越界，运算回绕，内核就会读写到另一行，并且不会报错。',
+        },
+      },
+      'persistent-mla': {
+        heading: '面向常见 head 打包方式的常驻 MLA decode',
+        paragraphs: [
+          'DeepSeek-V4 的 decode 还获得了面向 64 head 与 128 head MTP 打包方式的常驻 MLA 内核。这两种 head 数量正是普通解码与 speculative 验证实际产生的形态，因此引擎在常见形状上获得了专门的长上下文路径，而不是把它们当作某个为短上下文编写的内核的附带变体。这与前面 vLLM 选择 AITER sparse-MLA 内核的论点相同：在长上下文下，通用路径不是一种小幅妥协，而根本就是错误的内核。',
+        ],
+      },
+    },
+  },
+  dynamo: {
+    summary:
+      '批量化 KV 匹配、以 request lease 表示归属、更廉价的 router 状态，以及更精简的请求平面。',
+    lead: 'NVIDIA 提交的相当一部分结果使用 Dynamo 推理编排与 router 系统。它的 AgentX 系列工作说明：当引擎内核变快之后，分布式服务层本身可能成为瓶颈。router 的工作量正比于活跃 prefix 的数量与长度，而不是生成的 token 数，因此大量长期存在、彼此重叠的长会话对它造成的压力，是固定形状流量永远制造不出来的。',
+    highlights: [
+      '并发 512 下输出吞吐量中位数提升',
+      '采用 request lease 后 AgentX 回放耗时下降（vLLM 后端）',
+      '前端每秒请求数',
+    ],
+    sections: {
+      'routing-cost': {
+        heading: '每次路由决策的成本',
+        paragraphs: [
+          '第一批 PR 降低了单次路由决策的成本：减少查找热路径上的工作、去掉冗余的后缀失效操作，最后把 KV 匹配、注册、归属登记与终止解引用批量化，报告在并发 512 下输出吞吐量中位数提升 22.2%。批量化在这里奏效的原因与它在引擎中奏效的原因一样：单项开销已经压过了单项本身。',
+        ],
+      },
+      ownership: {
+        heading: '归属关系如何表示',
+        paragraphs: [
+          '第二批 PR 改变了归属关系的表示方式，这才是底层更难的问题。每个缓存 block 都需要归属到依赖它的请求上，才能做到"仍在使用时不释放、所有人用完后不再钉住"。当成千上万个并发会话共享彼此重叠的 prefix 时，这套记账本身的开销就变得可观。',
+          'Dynamo 从共享 block 链，转向 arena 级别的归属计数，最终采用按后端划分的 request lease，每一步都把被跟踪的单位变粗。lease 设计让 vLLM 后端的 AgentX 回放耗时降低 23.71%、SGLang 降低 22.02%，同时降低了峰值内存——这说明问题出在此前的表示方式上，而不是流量本身。',
+        ],
+        figure: {
+          alt: 'block 归属关系的前后对比：改动前每个 block 各自持有引用计数、请求指向它用到的每一个 block；改动后计数集中在一张带行号、时间戳与父行的表中，每个请求只保留一个书签。',
+          caption:
+            '归属表示的前后对比。计数从各个 block 中移出、集中到一张表里，请求只需保留一个书签，而不必引用它触及的每一个 block。',
+        },
+      },
+      'router-state': {
+        heading: '曾经很小的 router 状态',
+        paragraphs: [
+          '后续的 router profiling 又消除了一批形态相同的开销：某个周期性扫描或全量重算之所以此前可以接受，只是因为活跃状态一直很小。分桶式过期清理取代了正比于全部被跟踪对象的扫描，让高churn 的 AgentX 吞吐量提升 13.7%。仅处理增量的后缀清理只处理发生变化的部分，在同一时间窗口内多吸收约 28 倍的写入与删除事件。压缩 prompt 路径把前端 CPU 占用降低 35.3%，并显著改善尾部首 token 时间——这很重要，因为该负载的 prompt 既长又高度重复。过载状态现在也改为增量跟踪而非重新计算。',
+          '有一项路由改动是有意的取舍，而非纯粹的收益：Dynamo 现在可以把进行中的 decode 请求计入路由评分，于是一个已经承担长时间 decode 的 worker，会显得比其队列深度所暗示的更"贵"。在其报告的调参点上，这改善了 AgentX 延迟中位数，代价是少量吞吐量——只有当请求会长时间占用 worker 时，这种权衡才会显现出来。',
+        ],
+      },
+      'request-plane': {
+        heading: '请求平面',
+        paragraphs: [
+          '接下来优化的是请求平面，因为智能体 trace 并不是"一个请求、一个响应"。它会发出大量内容高度雷同的相关请求，并把每个 token 作为独立帧流式返回，因此序列化与拷贝是按轮次、按 token 付出的，而不是只付一次。改用 MessagePack 请求负载在其 AgentX 测试中把吞吐量提高 8.1%、平均首 token 时间降低 9.7%；直接的 Python 转码则彻底去掉了该路径上的中间值树。',
+          '随后是一系列"去掉一次拷贝"而非"把某步做快"的改动：不再拷贝 MessagePack 事件负载、不再拷贝收到的 ZeroMQ 帧、不再为每个 token 承担完整的 token 间延迟指标开销；chat 流式热路径也因同样理由被缩短。单看这些都平淡无奇，但乘上每个并发会话的每个流式 token，它们决定了前端每秒能承载多少请求。',
+        ],
+      },
+      profiling: {
+        heading: '与数据搬运毫无关系的开销',
+        paragraphs: [
+          '高并发 profiling 随后发现了一批与搬运数据毫无关系的开销。静态日志过滤器移除了一处共享的 span matcher 锁——这是竞争点问题而非流量问题——使报告中的前端吞吐量从每秒 932 个请求提升到 1,133 个。更简单的位置式 radix 分桶在 32 worker 的运行中把 mocker 的峰值内存降低了 5.51 GiB。',
+          '还有一个仍为 open 的改动，把 detokenization 指标改为每个响应汇总一次，而不是在每个流式分块上更新累计计数器，在其对照诊断 profile 中把前端 CPU 时间大约减半。这是该类问题最清楚的例子：单次调用的埋点很便宜，但按每个 token 调用一次就变得难以承受。',
+        ],
+      },
+    },
+  },
+  lmcache: {
+    summary:
+      '分块的外部 cache 加载、hybrid group 存储优化、AMD Instinct 支持，以及 DCP 感知的 offload。',
+    lead: 'LMCache 是一个开源 KV cache 层，位于 vLLM 等推理引擎之下，按 prefix hash 为 key，把可复用的 KV chunk 存放在 CPU DRAM、本地 NVMe 以及 Mooncake、Redis、S3 等远端后端上。它可以作为 vLLM 原生 offload connector 的替代方案。',
+    highlights: [
+      '旧路径死锁前后完成的请求数对比',
+      'hybrid group 每 token 存储量的下降幅度',
+      'MI350X 上通过的 KV 传输内核测试',
+    ],
+    sections: {
+      'chunked-loading': {
+        heading: '分块的外部 cache 加载',
+        paragraphs: [
+          'LMCache 的多进程路径针对智能体场景下 cache 搬运的规模与形态做了修改，起点是一种不是"变慢"而是"卡死"的失败：当大量上下文超过 100,000 token 的请求各自在开始前就为整次加载预留 block 时，池会被一群全都在等待、没有一个在推进的请求耗尽。',
+          '分块的外部 cache 加载改为按分块预留，于是各次加载可以交错进行并逐步排空。在并发 32 的验证中，旧路径在第 28 个请求处死锁，新路径完成了 120 个请求；并发 48 时即使 KV 池已占用 98.5%，运行依然继续。',
+        ],
+        figure: {
+          alt: '示意图：多个 vLLM 实例之上共享一个 LMCache 层，包含 CPU DRAM、经 GDS 访问的本地 NVMe 以及远端存储层，后到的请求可以在另一个实例上读到先前写入的内容。',
+          caption:
+            'LMCache 作为按 hash 寻址的 KV 层位于所有实例之下，横跨 CPU DRAM、本地 NVMe 与远端存储，因此后到的请求即使落在另一个实例上，也能读到先前写入的内容。',
+        },
+      },
+      'moving-less': {
+        heading: '搬得更少，也更少挡路',
+        paragraphs: [
+          '其余改动的目标是减少搬运量以及运行时的干扰。只存储 DeepSeek-V4 hybrid group 中真正有用的部分，把每 token 的存储量降低了近二十倍；sliding-window 预取现在只加载仍然有效的 window，而不是那些永远不会被读取的 window 状态——这与 vLLM 在 offload 上采用的可达性论证相同，只是从存储一侧切入。',
+          '随后，按对象组只做一次原生传输调用，消除了 staging 拷贝与内核启动之间反复的 Python 锁交接；这类开销正比于片段数量，而不是其中的字节数。',
+        ],
+      },
+      'lock-accounting': {
+        heading: 'Open：hybrid 锁记账',
+        paragraphs: [
+          '当前有两个 LMCache 改动与 AgentX 特别相关，但都仍为 open。hybrid 锁记账修复防止一个请求释放掉另一个请求在共享 sliding-window 或 recurrent 状态 chunk 上的读锁。要复现它需要三个条件同时成立：多个请求共享同一批 chunk、记账按持有者而非按 chunk 进行、并且淘汰确实开始发生。',
+          '开启 DRAM offload 的长时间 Kimi-K3 运行同时满足了这三点，产生了数万条告警、损坏的生成结果，并在淘汰开始后最终导致 GPU 崩溃。只要不是长时间、共享且内存吃紧的运行，这个缺陷就一直潜伏。',
+        ],
+      },
+      'amd-enablement': {
+        heading: 'AMD Instinct 支持',
+        paragraphs: [
+          '另一条并行的工作线让上述能力在 AMD Instinct 硬件上真正可用。CacheBlend 的非 prefix 复用依赖仅支持 CUDA 的 FlashInfer，因此一个 Triton block-sparse 注意力后端重新实现了它需要的三个内核：带 CSR 索引并输出 log-sum-exp 的 block-sparse attention、causal prefill，以及基于 log-sum-exp 的输出融合；在检测到 ROCm 或缺少 FlashInfer 时会自动切换到这些实现。ROCm Dockerfile 则对齐了 CUDA 的构建与轻量镜像。一个 AMD hipFile 后端扩展了此前只能通过 NVIDIA cuFile 访问存储的 GDS L1 slab 文件层：它通过 ctypes 绑定 ROCm 的 hipFile，并按 torch.version.hip 分发，cuFile 路径保持不变。',
+          '最后的缺口是分发方式。CUDA 用户安装预编译 wheel，AMD 用户却要从源码构建。我们与 AMD 一起发布了预编译的 gfx942 与 gfx950 wheel 来补上这一环：它可以装入上游镜像，并在 MI350X 上通过全部 56 项 KV 传输内核测试；它发布到 GitHub release 而非 PyPI，因此直接执行 pip install lmcache 仍会安装 CUDA 版本。一个单行的后续修复把 bind mount 进来的仓库标记为 git safe directory——这个问题只在 CI 中出现，因为容器以 root 身份运行在 runner 所有的检出目录上，setup.py 中的版本探测会拒绝读取它。',
+        ],
+      },
+      'dcp-offload': {
+        heading: 'DCP 感知的 CPU offload',
+        paragraphs: [
+          'DCP 感知的 CPU offload 解决了两个在长上下文下必须同时启用的特性之间的直接冲突。启用 decode context parallelism 后，每个 rank 只持有 KV 的一个 stride，因此单个 rank 能保存的内容并不构成可用的 prefix。修复方案在保存前汇集这些带 stride 的分片，并在加载后重新分发。如果不这样做，启用 context parallelism 就会悄悄让 CPU cache 命中失效，而失效的恰恰是催生这两个特性的长 prefix。其验证记录了超过 30,000 次 CPU 命中事件，单个请求的加载量可达数十万 token。',
+        ],
+      },
+    },
+  },
+  mooncake: {
+    summary:
+      '通过 HIP dmabuf 在 ROCm 上实现 GPU-direct RDMA，并提供已发布的 ROCm wheel 与发版流程。',
+    lead: 'Mooncake 承载着 Moonshot 的 Kimi 生产流量以及多家实验室的生产流量，同时也是 disaggregated vLLM 与 SGLang 配置底层的传输引擎。直到最近，它在 AMD 上的支持既缺少 RDMA 注册，也缺少可直接安装的软件包。',
+    highlights: [
+      'ROCm 上的 GPU-direct RDMA 路径',
+      '同一个架构无关 wheel 覆盖的目标',
+      '按 tag 触发的发布矩阵',
+    ],
+    sections: {
+      'rdma-registration': {
+        heading: 'ROCm 上的 GPU-direct RDMA 注册',
+        paragraphs: [
+          '在 NVIDIA 上，为 RDMA 注册 GPU 显存要么使用 nvidia-peermem 内核模块，要么导出 dmabuf 文件描述符。AMD 没有 nvidia-peermem 的对应物，因此 GPU-direct RDMA 完全没有可走的路径，部署只能退回到通过主机 DRAM 中转 KV。',
+          '一个 HIP dmabuf 注册分支补上了与现有 CUDA dmabuf 路径对应的实现：改用 ROCm 导出，而不是调用 CUDA 的 handle 接口，并且会先解析出真实的分配基址，因为带缓存的分配器会把张量打包在一块更大分配内部的某个偏移处。主机内存仍然直接注册。',
+        ],
+      },
+      packaging: {
+        heading: '装不上的支持不算支持',
+        paragraphs: [
+          'Mooncake 发布了 CUDA 与 MUSA 的 wheel，却没有 ROCm 包，因此 AMD 用户只能在每个镜像里从源码构建引擎。ROCm wheel、CI 与发版流程把 mooncake-transfer-engine-rocm 一并发布到 PyPI。这条工作线来自 AMD 工程师 Andy Luo，他在用 AgentX 自测（dogfooding）智能体负载时注意到：在 ROCm 上从源码构建 Mooncake 并不是一流的使用体验。',
+          '该传输引擎没有设备端内核，也不依赖 torch，因此一个架构无关的 wheel 即可覆盖 gfx942 与 gfx950；ROCm 运行时在加载时绑定而非随包分发，这意味着同一个 wheel 在上游 vLLM ROCm 镜像与 SGLang ROCm 镜像中都可直接使用。验证覆盖了完整的交叉组合：MI300X 与 MI355X，各自在 vllm/vllm-openai-rocm 与 lmsysorg/sglang 下运行 master 二进制及带数据校验的 HIP buffer 传输测试。该 PR 还增加了按 tag 触发、覆盖 Python 3.10 到 3.13 的发布流程。',
+          '一个仍为 open 的后续改动增加了自托管的双节点 MI350X 外部 prefill 与 decode 环境，让 ROCm 的 disaggregated 路径在真实硬件上被实际执行，而不只是通过编译。综合这些工作，AMD 上的 AgentX 运行现在可以直接从已发布产物把传输引擎与 KV cache 层装进标准上游镜像，并在 GPU 显存与网络之间直接搬运 KV。',
+        ],
+        links: ['PyPI 上的 mooncake-transfer-engine'],
+      },
+    },
+  },
+};
+
+function localizeSection(
+  section: OptimizationSection,
+  zh: SectionTranslation,
+): OptimizationSection {
+  return {
+    ...section,
+    heading: zh.heading,
+    paragraphs: zh.paragraphs,
+    figure: section.figure && zh.figure ? { ...section.figure, ...zh.figure } : section.figure,
+    links: section.links?.map((link, index) => ({
+      ...link,
+      label: zh.links?.[index] ?? link.label,
+    })),
+  };
+}
+
+/** The overview content for a locale, with English structure and localized prose. */
+export function getOptimizationsOverview(locale: Locale): OptimizationsOverview {
+  if (locale === 'en') return OPTIMIZATIONS_OVERVIEW;
+  return {
+    ...OPTIMIZATIONS_OVERVIEW,
+    eyebrow: OVERVIEW_ZH.eyebrow,
+    title: OVERVIEW_ZH.title,
+    lead: OVERVIEW_ZH.lead,
+    intro: OVERVIEW_ZH.intro,
+    highlights: OPTIMIZATIONS_OVERVIEW.highlights.map((highlight, index) => ({
+      ...highlight,
+      label: OVERVIEW_ZH.highlights[index] ?? highlight.label,
+    })),
+    frameworksTitle: OVERVIEW_ZH.frameworksTitle,
+    frameworksIntro: OVERVIEW_ZH.frameworksIntro,
+    sections: OPTIMIZATIONS_OVERVIEW.sections.map((section) => {
+      const zh = OVERVIEW_ZH.sections[section.id];
+      return zh ? localizeSection(section, zh) : section;
+    }),
+    layerLabels: OVERVIEW_ZH.layerLabels,
+    ui: OVERVIEW_ZH.ui,
+  };
+}
+
+/** One framework's content for a locale. */
+export function getLocalizedFramework(
+  framework: OptimizationFramework,
+  locale: Locale,
+): OptimizationFramework {
+  if (locale === 'en') return framework;
+  const zh = FRAMEWORKS_ZH[framework.slug];
+  if (!zh) return framework;
+  return {
+    ...framework,
+    summary: zh.summary,
+    lead: zh.lead,
+    highlights: framework.highlights.map((highlight, index) => ({
+      ...highlight,
+      label: zh.highlights[index] ?? highlight.label,
+    })),
+    sections: framework.sections.map((section) => {
+      const sectionZh = zh.sections[section.id];
+      return sectionZh ? localizeSection(section, sectionZh) : section;
+    }),
+  };
+}
+
+/** Every framework for a locale, in registry order. */
+export function getLocalizedFrameworks(locale: Locale): readonly OptimizationFramework[] {
+  return AGENTX_OPTIMIZATION_FRAMEWORKS.map((framework) =>
+    getLocalizedFramework(framework, locale),
+  );
+}
+
+/** Exported for the parity test, which walks the same keys the resolvers read. */
+export const OPTIMIZATIONS_ZH_INTERNALS = { OVERVIEW_ZH, FRAMEWORKS_ZH };
diff --git a/packages/app/src/lib/agentx-optimizations.test.ts b/packages/app/src/lib/agentx-optimizations.test.ts
new file mode 100644
index 000000000..67add110c
--- /dev/null
+++ b/packages/app/src/lib/agentx-optimizations.test.ts
@@ -0,0 +1,191 @@
+import { existsSync } from 'node:fs';
+import path from 'node:path';
+
+import { describe, expect, it } from 'vitest';
+
+import {
+  AGENTX_OPTIMIZATION_FRAMEWORKS,
+  AGENTX_OPTIMIZATION_SLUGS,
+  OPTIMIZATION_FIGURES,
+  OPTIMIZATION_LAYERS,
+  OPTIMIZATIONS_OVERVIEW,
+  PR_REPOS,
+  countReferencedPrs,
+  getOptimizationFramework,
+  prLabel,
+  prUrl,
+  type OptimizationSection,
+} from './agentx-optimizations';
+import {
+  getLocalizedFramework,
+  getLocalizedFrameworks,
+  getOptimizationsOverview,
+} from './agentx-optimizations-zh';
+
+const CJK = /[一-鿿]/u;
+
+function allSections(): OptimizationSection[] {
+  return [
+    ...OPTIMIZATIONS_OVERVIEW.sections,
+    ...AGENTX_OPTIMIZATION_FRAMEWORKS.flatMap((framework) => framework.sections),
+  ];
+}
+
+describe('AgentX optimizations registry', () => {
+  it('has unique slugs and section ids', () => {
+    expect(new Set(AGENTX_OPTIMIZATION_SLUGS).size).toBe(AGENTX_OPTIMIZATION_SLUGS.length);
+    for (const framework of AGENTX_OPTIMIZATION_FRAMEWORKS) {
+      const ids = framework.sections.map((section) => section.id);
+      expect(new Set(ids).size, framework.slug).toBe(ids.length);
+    }
+  });
+
+  it('uses URL-safe slugs and known layers', () => {
+    for (const framework of AGENTX_OPTIMIZATION_FRAMEWORKS) {
+      expect(framework.slug).toMatch(/^[a-z0-9-]+$/u);
+      expect(OPTIMIZATION_LAYERS).toContain(framework.layer);
+      expect(framework.name.length).toBeGreaterThan(0);
+      expect(framework.sections.length).toBeGreaterThan(0);
+    }
+  });
+
+  it('resolves a framework by slug and rejects unknown ones', () => {
+    expect(getOptimizationFramework('vllm')?.name).toBe('vLLM');
+    expect(getOptimizationFramework('not-a-framework')).toBeUndefined();
+  });
+
+  it('references only known repositories, and builds their PR URLs', () => {
+    for (const section of allSections()) {
+      for (const pr of section.prs ?? []) {
+        expect(Object.keys(PR_REPOS)).toContain(pr.repo);
+        expect(Number.isInteger(pr.number) && pr.number > 0).toBe(true);
+        expect(prUrl(pr)).toBe(`https://github.com/${pr.repo}/pull/${pr.number}`);
+        expect(prLabel(pr)).toContain(`#${pr.number}`);
+      }
+    }
+  });
+
+  it('never lists the same PR twice inside one section', () => {
+    for (const section of allSections()) {
+      const urls = (section.prs ?? []).map(prUrl);
+      expect(new Set(urls).size, section.id).toBe(urls.length);
+    }
+  });
+
+  it('backs the "50+ upstream PRs" claim in the lead copy', () => {
+    expect(OPTIMIZATIONS_OVERVIEW.lead).toContain('50+');
+    expect(countReferencedPrs()).toBeGreaterThanOrEqual(50);
+  });
+
+  it('links only to https references', () => {
+    for (const section of allSections()) {
+      for (const link of section.links ?? []) {
+        expect(link.href.startsWith('https://'), link.href).toBe(true);
+        expect(link.label.length).toBeGreaterThan(0);
+      }
+    }
+  });
+
+  it('ships every figure it references', () => {
+    const publicDir = path.resolve(import.meta.dirname, '../../public');
+    for (const section of allSections()) {
+      if (!section.figure) continue;
+      const asset = OPTIMIZATION_FIGURES[section.figure.key];
+      expect(asset, section.figure.key).toBeDefined();
+      expect(existsSync(path.join(publicDir, asset.src)), asset.src).toBe(true);
+      // Dimensions feed next/image, so a wrong aspect ratio would shift layout.
+      expect(asset.width).toBeGreaterThan(0);
+      expect(asset.height).toBeGreaterThan(0);
+      expect(section.figure.alt.length).toBeGreaterThan(20);
+      expect(section.figure.caption.length).toBeGreaterThan(20);
+    }
+  });
+
+  it('uses every declared figure exactly once', () => {
+    const used = allSections()
+      .map((section) => section.figure?.key)
+      .filter((key): key is NonNullable<typeof key> => Boolean(key));
+    expect(new Set(used).size).toBe(used.length);
+    expect(new Set(used)).toEqual(new Set(Object.keys(OPTIMIZATION_FIGURES)));
+  });
+
+  it('carries no unresolved editorial notes from the source document', () => {
+    const prose = [
+      OPTIMIZATIONS_OVERVIEW.lead,
+      ...OPTIMIZATIONS_OVERVIEW.intro,
+      ...allSections().flatMap((section) => [section.heading, ...section.paragraphs]),
+      ...AGENTX_OPTIMIZATION_FRAMEWORKS.flatMap((framework) => [framework.lead, framework.summary]),
+    ].join('\n');
+    for (const marker of ['TO BE UPDATED', 'TODO', 'Bryan Shan', 'Source: SemiAnalysis']) {
+      expect(prose, marker).not.toContain(marker);
+    }
+  });
+});
+
+describe('AgentX optimizations Chinese port', () => {
+  it('translates the overview without dropping structure', () => {
+    const en = OPTIMIZATIONS_OVERVIEW;
+    const zh = getOptimizationsOverview('zh');
+
+    expect(zh.title).not.toBe(en.title);
+    expect(CJK.test(zh.title)).toBe(true);
+    expect(CJK.test(zh.lead)).toBe(true);
+    expect(zh.intro).toHaveLength(en.intro.length);
+    expect(zh.highlights).toHaveLength(en.highlights.length);
+    expect(zh.sections).toHaveLength(en.sections.length);
+
+    for (const [index, section] of zh.sections.entries()) {
+      const source = en.sections[index];
+      expect(section.id).toBe(source.id);
+      expect(section.paragraphs).toHaveLength(source.paragraphs.length);
+      expect(CJK.test(section.heading), section.id).toBe(true);
+      expect(section.prs).toEqual(source.prs);
+      expect(section.links?.map((link) => link.href)).toEqual(
+        source.links?.map((link) => link.href),
+      );
+      expect(section.figure?.key).toBe(source.figure?.key);
+    }
+
+    for (const value of Object.values(zh.ui)) expect(CJK.test(value)).toBe(true);
+    for (const label of Object.values(zh.layerLabels)) expect(label.length).toBeGreaterThan(0);
+  });
+
+  it('returns the English registry untouched for the en locale', () => {
+    expect(getOptimizationsOverview('en')).toBe(OPTIMIZATIONS_OVERVIEW);
+    expect(getLocalizedFrameworks('en')).toEqual(AGENTX_OPTIMIZATION_FRAMEWORKS);
+  });
+
+  it('translates every framework, section, figure, and highlight', () => {
+    for (const framework of AGENTX_OPTIMIZATION_FRAMEWORKS) {
+      const zh = getLocalizedFramework(framework, 'zh');
+
+      expect(zh.slug).toBe(framework.slug);
+      // Product names, PR references, and figure assets stay language-neutral.
+      expect(zh.name).toBe(framework.name);
+      expect(CJK.test(zh.summary), framework.slug).toBe(true);
+      expect(CJK.test(zh.lead), framework.slug).toBe(true);
+      expect(zh.highlights).toHaveLength(framework.highlights.length);
+
+      for (const [index, highlight] of zh.highlights.entries()) {
+        expect(highlight.value).toBe(framework.highlights[index].value);
+        expect(CJK.test(highlight.label), `${framework.slug} highlight ${index}`).toBe(true);
+      }
+
+      expect(zh.sections).toHaveLength(framework.sections.length);
+      for (const [index, section] of zh.sections.entries()) {
+        const source = framework.sections[index];
+        const where = `${framework.slug}/${source.id}`;
+        expect(section.id).toBe(source.id);
+        expect(section.paragraphs, where).toHaveLength(source.paragraphs.length);
+        expect(CJK.test(section.heading) || section.heading === source.heading, where).toBe(true);
+        for (const paragraph of section.paragraphs) expect(CJK.test(paragraph), where).toBe(true);
+        expect(section.prs).toEqual(source.prs);
+        expect(section.figure?.key).toBe(source.figure?.key);
+        if (source.figure) {
+          expect(CJK.test(section.figure?.alt ?? ''), where).toBe(true);
+          expect(CJK.test(section.figure?.caption ?? ''), where).toBe(true);
+        }
+      }
+    }
+  });
+});
diff --git a/packages/app/src/lib/agentx-optimizations.ts b/packages/app/src/lib/agentx-optimizations.ts
new file mode 100644
index 000000000..b18e189dd
--- /dev/null
+++ b/packages/app/src/lib/agentx-optimizations.ts
@@ -0,0 +1,985 @@
+/**
+ * AgentX industry impact — the upstream optimization work that AgentX traces
+ * drove across inference engines, routers, KV-cache layers, and kernels.
+ *
+ * English content lives here; the Simplified Chinese port is a 1:1 mirror in
+ * `agentx-optimizations-zh.ts`, keyed by the same slugs and section ids.
+ * `agentx-optimizations.test.ts` enforces that the two stay structurally
+ * identical, so a section added here without a translation fails CI.
+ *
+ * Every claim links to the upstream pull request it came from. Sections list
+ * PRs as `{ repo, number }` rather than prose links so the reference survives
+ * translation unchanged — a PR number is an identifier, not copy.
+ */
+
+/** Where a project sits in the serving stack; drives the index-page grouping. */
+export const OPTIMIZATION_LAYERS = ['engine', 'router', 'kv-cache', 'kernels', 'transfer'] as const;
+
+export type OptimizationLayer = (typeof OPTIMIZATION_LAYERS)[number];
+
+/** Upstream repositories referenced by this page, with their display names. */
+export const PR_REPOS = {
+  'vllm-project/vllm': 'vLLM',
+  'sgl-project/sglang': 'SGLang',
+  'NVIDIA/TensorRT-LLM': 'TensorRT-LLM',
+  'ROCm/ATOM': 'ATOM',
+  'ROCm/aiter': 'AITER',
+  'ai-dynamo/dynamo': 'Dynamo',
+  'LMCache/LMCache': 'LMCache',
+  'kvcache-ai/Mooncake': 'Mooncake',
+} as const;
+
+export type PrRepo = keyof typeof PR_REPOS;
+
+export interface UpstreamPr {
+  repo: PrRepo;
+  number: number;
+}
+
+export function prUrl(pr: UpstreamPr): string {
+  return `https://github.com/${pr.repo}/pull/${pr.number}`;
+}
+
+export function prLabel(pr: UpstreamPr): string {
+  return `${PR_REPOS[pr.repo]} #${pr.number}`;
+}
+
+/** Figures exported from the source deck, served from /public. */
+export const OPTIMIZATION_FIGURES = {
+  servingStack: {
+    src: '/images/agentx-optimizations/serving-stack.png',
+    width: 1814,
+    height: 2048,
+  },
+  vllmSelectiveRetention: {
+    src: '/images/agentx-optimizations/vllm-selective-retention.png',
+    width: 2048,
+    height: 632,
+  },
+  vllmConnectorLayer: {
+    src: '/images/agentx-optimizations/vllm-connector-layer.png',
+    width: 2048,
+    height: 837,
+  },
+  sglangHiCacheAsymmetry: {
+    src: '/images/agentx-optimizations/sglang-hicache-asymmetry.png',
+    width: 2048,
+    height: 720,
+  },
+  sglangRadixStaging: {
+    src: '/images/agentx-optimizations/sglang-radix-staging.png',
+    width: 2048,
+    height: 544,
+  },
+  trtllmPipelinedTransfer: {
+    src: '/images/agentx-optimizations/trtllm-pipelined-transfer.png',
+    width: 2048,
+    height: 769,
+  },
+  atomPrefixPromotion: {
+    src: '/images/agentx-optimizations/atom-prefix-promotion.png',
+    width: 2048,
+    height: 420,
+  },
+  aiter64BitOffsets: {
+    src: '/images/agentx-optimizations/aiter-64bit-offsets.png',
+    width: 1408,
+    height: 1046,
+  },
+  dynamoOwnershipTable: {
+    src: '/images/agentx-optimizations/dynamo-ownership-table.png',
+    width: 2048,
+    height: 576,
+  },
+  lmcacheTiers: {
+    src: '/images/agentx-optimizations/lmcache-tiers.png',
+    width: 2048,
+    height: 638,
+  },
+} as const;
+
+export type OptimizationFigureKey = keyof typeof OPTIMIZATION_FIGURES;
+
+export interface OptimizationFigureCopy {
+  alt: string;
+  caption: string;
+}
+
+export interface ReferenceLink {
+  label: string;
+  href: string;
+}
+
+export interface OptimizationSection {
+  /** Stable anchor id, used for in-page navigation and as the translation key. */
+  id: string;
+  heading: string;
+  paragraphs: readonly string[];
+  figure?: { key: OptimizationFigureKey } & OptimizationFigureCopy;
+  prs?: readonly UpstreamPr[];
+  links?: readonly ReferenceLink[];
+}
+
+export interface OptimizationHighlight {
+  value: string;
+  label: string;
+}
+
+export interface OptimizationFramework {
+  /** URL segment: /agentx/optimizations/<slug>. */
+  slug: string;
+  /** Product name — never translated. */
+  name: string;
+  layer: OptimizationLayer;
+  /** One-line description used on the index card and in page metadata. */
+  summary: string;
+  /** Intro rendered above the first section on the detail page. */
+  lead: string;
+  highlights: readonly OptimizationHighlight[];
+  sections: readonly OptimizationSection[];
+}
+
+export interface OptimizationsOverview {
+  eyebrow: string;
+  title: string;
+  lead: string;
+  intro: readonly string[];
+  highlights: readonly OptimizationHighlight[];
+  frameworksTitle: string;
+  frameworksIntro: string;
+  sections: readonly OptimizationSection[];
+  layerLabels: Readonly<Record<OptimizationLayer, string>>;
+  /** Shared UI strings for both the index and the detail pages. */
+  ui: {
+    backToAgentX: string;
+    backToOverview: string;
+    prsLabel: string;
+    referencesLabel: string;
+    readMore: string;
+    onThisPage: string;
+    allProjects: string;
+    figureCta: string;
+    prSearch: string;
+  };
+}
+
+export const OPTIMIZATIONS_OVERVIEW: OptimizationsOverview = {
+  eyebrow: 'AgentX industry impact',
+  title: 'Optimizations for Agentic Workloads',
+  lead: 'The most useful thing AgentX produced in its first months was not the open-source datasets. It was 50+ upstream pull requests from AgentX partners, optimizing real-world agentic workloads with AgentX as the north star.',
+  intro: [
+    'AgentX replays real agentic traffic, so it benchmarks more than raw prefill and decode kernels. It exercises the whole path end to end: KV-cache lifecycle, hybrid-attention cache correctness, CPU KV offload, transfer progress, routing affinity, incremental tokenization, request serialization, and scheduler bookkeeping. Every one of those costs is already paid by every production agentic deployment, and none of them is visible in a single-turn 8k/1k scenario.',
+    'SemiAnalysis has also worked with AMD software development over several years to modernize their development principles. Much of the work below is the result, and it moves AMD open source closer to first-class support for agentic workloads.',
+  ],
+  highlights: [
+    { value: '50+', label: 'upstream PRs' },
+    { value: '8', label: 'upstream projects' },
+    { value: '5', label: 'layers of the stack' },
+    { value: 'AgentX', label: 'the north star trace' },
+  ],
+  frameworksTitle: 'Optimizations by project',
+  frameworksIntro:
+    'Each page collects the AgentX-driven work in one project, grouped by the part of the stack it touches. Changes described as open are proposed but not merged, and are not shipped behavior.',
+  sections: [
+    {
+      id: 'ecosystem',
+      heading: 'A brief introduction to the distributed inference ecosystem',
+      paragraphs: [
+        'Agentic inference is a system-wide problem rather than a chip or kernel problem. Once a distributed system handles hundreds of thousands of agentic requests, request scheduling and KV-cache management stop being bookkeeping and start having legitimate performance implications. Subagents, for example, produce bursty KV-cache patterns that will evict the main agent’s cache if nothing prevents it.',
+        'At the top of the stack, routers — sometimes called frontends — send requests to workers. When a server runs data-parallel attention there is a separate KV cache per DP rank, so requests are routed under policies such as consistent hashing, where every request in the same session or subagent follows its unique ID to the same rank rather than thrashing all of them.',
+        'Most routing policies do not differ much between implementations. Some routers are separate components, such as the vLLM router and the llm-d router; others are integrated into the engine, such as the SGLang model gateway and ATOM Mesh.',
+        'After routing, a request reaches the scheduler of an inference engine such as vLLM or SGLang. The engine performs the inference and returns the result over an API. Each engine also exposes an interface connecting its internal KV cache to external KV-cache managers, which is what makes the ecosystem pluggable: one cache manager can integrate with many engines.',
+        'The simple deployment used in the current AgentX results runs Mooncake alongside vLLM on the same node. Each vLLM worker embeds a Mooncake Store client and contributes part of host DRAM to the external KV-cache pool. vLLM reaches that pool through the MooncakeStoreConnector interface, which loads reusable KV blocks into GPU memory and saves newly computed blocks back to host memory. Mooncake Store handles placement and eviction; Mooncake Transfer Engine moves the bytes between GPU and CPU memory.',
+        'Different KV-cache managers use different transfer engines to move bytes between memory tiers or machines, such as between prefill and decode workers. A deployment can use Mooncake Store to offload reusable KV blocks to host DRAM while simultaneously using NIXL to move request-specific KV directly from prefill GPUs to decode GPUs, with NIXL using UCX and GPUDirect RDMA where supported. Multiple KV-management and transfer paths can coexist inside one inference engine.',
+        'The ecosystem is made of many independent components: inference engines, routers, KV-cache managers, data-transfer libraries, and cluster controllers. Platforms such as NVIDIA Dynamo, llm-d, and AMD Infera package selected combinations into complete distributions, publishing compatible container images, connectors, deployment manifests, and orchestration logic. The result is usually a set of coordinated containers rather than one monolithic service.',
+      ],
+      figure: {
+        key: 'servingStack',
+        alt: 'Five-layer diagram of the distributed inference stack: routing and frontend layer, inference engines, KV-cache management layer, data movement layer, and accelerators with tiered storage.',
+        caption:
+          'The distributed inference stack. Each layer has several interchangeable implementations, and platforms such as Dynamo, llm-d, and AMD Infera package selected combinations into one distribution.',
+      },
+      links: [
+        {
+          label: 'Consistent-hash routing policy',
+          href: 'https://github.com/vllm-project/router/blob/main/src/policies/consistent_hash.rs',
+        },
+        { label: 'vLLM router', href: 'https://github.com/vllm-project/router' },
+        { label: 'llm-d router', href: 'https://github.com/llm-d/llm-d-router' },
+        {
+          label: 'SGLang model gateway',
+          href: 'https://github.com/sgl-project/sglang/tree/main/sgl-model-gateway',
+        },
+        { label: 'ATOM Mesh', href: 'https://github.com/ROCm/ATOM/tree/main/atom/mesh' },
+        {
+          label: 'Mooncake Transfer Engine',
+          href: 'https://github.com/kvcache-ai/Mooncake/tree/main/mooncake-transfer-engine',
+        },
+      ],
+    },
+    {
+      id: 'other',
+      heading: 'Other optimizations: day-zero enablement and correctness',
+      paragraphs: [
+        'The work collected on the project pages addresses long-context cost: a prefix that has to survive, a hybrid cache that has to stay correct, a transfer that has to keep up. The changes here are different in kind. They are day-zero enablement and correctness bugs, and they break a request exactly as badly as a million-token session does.',
+        'MiniMax-M3 tested whether the ROCm work compounds into day-zero readiness, and the Advancing AI writeup draws the comparison directly: AMD’s first public disaggregated recipe, MI355X FP4, reached InferenceX in January, months behind NVIDIA, while M3 FP4 disaggregation landed on day zero. That is a marked improvement on the DeepSeek-R1 period, when parity took months. Three vLLM fixes sat on the day-zero path, and each was a correctness failure rather than a performance one.',
+        'Disaggregation was blocked first. NixlConnector’s handshake asserted that the SPLIT-region block_len scales with the prefill-to-decode TP ratio, but block_len follows per-rank KV heads. M3 has 4 KV heads, so a TP4 prefill paired with a TP8 decode is GQA-capped to one head per rank on both sides, and the two lengths are equal where the assertion demanded a factor of two. The handshake was rejected, no KV moved, decode regenerated everything from scratch, and gsm8k scored 0. Validating against the actual head ratio fixed it.',
+        'The other two were platform splits. M3’s sparse-attention backend read the byte-backed FP8 cache as float8_e4m3fn for every E4M3 configuration, but gfx942’s platform dtype is e4m3fnuz. The two encodings differ, so K and V were altered before the kernels consumed them, and the prefill and decode wrappers had also omitted the FNUZ types from their FP8 checks. Using the platform dtype for the cache view fixed both halves. Separately, M3 ships as separate NVIDIA and AMD model files and only the NVIDIA one implemented the EAGLE3 interface, so speculative decoding aborted at engine init on ROCm with a model-does-not-support error. Bringing the AMD model to parity restored it, with MI355X gsm8k matching both the non-EAGLE3 MI355X run and B200.',
+        'The TensorRT-LLM page covers the M3 work that is long-context specific: descriptor explosion in disaggregated KV transfer, context graph capture, sparse block strides, autotuner candidates, and the corrupt split-K MoE tactics that had to be removed from the pool.',
+      ],
+      prs: [
+        { repo: 'vllm-project/vllm', number: 45879 },
+        { repo: 'vllm-project/vllm', number: 45720 },
+        { repo: 'vllm-project/vllm', number: 45546 },
+      ],
+      links: [
+        {
+          label: 'Can AMD break the CUDA moat? — Advancing AI',
+          href: 'https://newsletter.semianalysis.com/p/can-amd-break-the-cuda-moat-amd-advancing',
+        },
+      ],
+    },
+    {
+      id: 'what-activates-this',
+      heading: 'What the AgentX matrix activates',
+      paragraphs: [
+        'The local AgentX matrix combines session-aware or KV-aware routing, long and variable conversation histories, MTP, hybrid attention, aggregated and disaggregated serving, and concurrency sweeps that cross the HBM capacity cliff. It includes GPU-resident comparisons as well as CPU DRAM offload through vLLM SimpleCPU, Mooncake, LMCache, and SGLang HiCache. That combination is what activates the upstream work above.',
+        'The old fixed-sequence matrix usually creates one prompt, performs one prefill, decodes one fixed continuation, and discards the request. It therefore does not measure cache survival across turns, repeated tokenization, session affinity, cache-event traffic, offload churn, transfer progress during scheduler stalls, or long-lived ownership bookkeeping.',
+        'The allowed optimization policy treats CPU KV offload as optional. A vendor may use vLLM connectors, LMCache, SGLang HiCache, Mooncake, Dynamo KVBM, or another CPU DRAM connector — or disable offload when the resulting latency and throughput point is better. NVMe offload is deferred. CPU DRAM must scale with the fraction of GPUs used, including the 3 TB cap for systems without standardized DRAM. Systems with standardized DRAM have no hard cap but keep the same proportionality rule; the local generator currently applies the 3 TB cap to every runner, so it does not yet implement the standardized-DRAM exception.',
+        'The net new optimization surface is not simply longer attention. It is the preservation, movement, routing, reconstruction, and repeated processing of a growing session state. AgentX made those costs large enough to drive generic upstream changes across vLLM, SGLang, TensorRT-LLM, ATOM, AITER, Dynamo, and LMCache. Direct searches of NIXL and Mooncake did not identify additional AgentX-tagged runtime PRs beyond those listed here, so their effects remain represented through the engine connector changes.',
+      ],
+      links: [
+        {
+          label: 'Every upstream PR mentioning AgentX',
+          href: 'https://github.com/search?q=agentx&type=pullrequests&s=created&o=desc',
+        },
+      ],
+    },
+  ],
+  layerLabels: {
+    engine: 'Inference engine',
+    router: 'Router and orchestration',
+    'kv-cache': 'KV-cache layer',
+    kernels: 'Kernels',
+    transfer: 'Transfer engine',
+  },
+  ui: {
+    backToAgentX: '← Back to AgentX',
+    backToOverview: '← All AgentX optimizations',
+    prsLabel: 'Upstream pull requests',
+    referencesLabel: 'References',
+    readMore: 'Read the optimizations',
+    onThisPage: 'On this page',
+    allProjects: 'Other projects',
+    figureCta: 'View full-resolution image',
+    prSearch: 'Search every AgentX pull request',
+  },
+};
+
+const FRAMEWORKS: readonly OptimizationFramework[] = [
+  {
+    slug: 'vllm',
+    name: 'vLLM',
+    layer: 'engine',
+    summary:
+      'Hybrid-attention prefix retention, CPU KV offload for hybrid models, and a narrowed store and load path.',
+    lead: 'Working alongside vLLM maintainers from Inferact, Red Hat, NVIDIA, and AMD, we used AgentX’s realistic replayer as a north star. The resulting fixes landed upstream, and most of them transfer directly to production.',
+    highlights: [
+      { value: '>95%', label: 'prefix-cache hit rate at 1M context' },
+      { value: '+81.7%', label: 'output throughput with hybrid CPU offload' },
+      { value: '−46.6%', label: 'mean end-to-end latency' },
+    ],
+    sections: [
+      {
+        id: 'hybrid-prefix-caching',
+        heading: 'Hybrid-attention prefix caching',
+        paragraphs: [
+          'vLLM improved hybrid-attention prefix caching so short-lived sliding-window allocations do not evict useful long-context checkpoints. Selective retention preserves sparse replay boundaries and reported a prefix-cache hit rate above 95% with fourteen concurrent requests and contexts up to one million tokens.',
+          'The same reachability policy was applied to Mooncake, and unreachable sliding-window lookups were removed. Earlier follow-up work also stopped offloading sliding-window blocks that could never be reused, and kept the speculative lookahead block inside the retained prefix.',
+        ],
+        figure: {
+          key: 'vllmSelectiveRetention',
+          alt: 'Before and after diagram: without retention every sliding-window tail is freed and the whole prefix must be recomputed; with selective retention a few checkpoints survive so the prefix stays reusable.',
+          caption:
+            'Before, no window tail survives, so no position is resumable and the full-attention KV is resident but unusable. After selective retention, a few tails survive and the prefix-cache hit rate exceeds 95%.',
+        },
+        prs: [
+          { repo: 'vllm-project/vllm', number: 43447 },
+          { repo: 'vllm-project/vllm', number: 44774 },
+          { repo: 'vllm-project/vllm', number: 45444 },
+          { repo: 'vllm-project/vllm', number: 42258 },
+          { repo: 'vllm-project/vllm', number: 44082 },
+        ],
+      },
+      {
+        id: 'cpu-offload',
+        heading: 'CPU KV offload for hybrid models',
+        paragraphs: [
+          'Highly concurrent agentic workloads require offload, and AgentX drove the work that made CPU KV offload usable for hybrid models rather than only for uniform full-attention models. The distinction matters: a uniform model has one KV layout per token, so a connector can describe what to save with a single block geometry. A hybrid model carries several cache groups at once, each with a different shape and a different lifetime, and a connector that assumes one uniform layout cannot express which group a given block belongs to. Offload was therefore unavailable for exactly the models whose long sessions needed it most.',
+          'The general SimpleCPU connector came first, was enabled on ROCm, and was then extended to DeepSeek-V4 hybrid attention, which reported 81.7% higher output throughput and 46.6% lower mean end-to-end latency against recomputing the prefix once it no longer fits in HBM. Mooncake gained the equivalent hybrid-memory allocation support.',
+        ],
+        prs: [
+          { repo: 'vllm-project/vllm', number: 37160 },
+          { repo: 'vllm-project/vllm', number: 40549 },
+          { repo: 'vllm-project/vllm', number: 42296 },
+          { repo: 'vllm-project/vllm', number: 42828 },
+        ],
+      },
+      {
+        id: 'store-path',
+        heading: 'Narrowing the store path',
+        paragraphs: [
+          'Profiling realistic workloads showed that once offload worked at all, the cost moved to the store path, which was writing too much and too often. Three rules narrowed it.',
+          'A store is now skipped while an identical transfer is already in flight, so concurrent sessions sharing a prefix pay for it once rather than once each. A store covers only newly generated KV ranges, so a session that extends its history writes the delta instead of rewriting the whole prefix every turn. And a store no longer depends on whether the same blocks still sit in HBM, so work already scheduled is not discarded when an eviction lands underneath it.',
+        ],
+        figure: {
+          key: 'vllmConnectorLayer',
+          alt: 'Diagram of the vLLM connector layer between GPU HBM and a CPU DRAM pool, listing the connector PRs and the three store-side rules that narrow writes.',
+          caption:
+            'The connector layer between GPU HBM and the CPU DRAM pool. Store-side rules skip in-flight duplicates, write only newly generated KV, and decouple the store from whether the blocks still sit in HBM; the scheduler-side lookup is asynchronous.',
+        },
+        prs: [
+          { repo: 'vllm-project/vllm', number: 41289 },
+          { repo: 'vllm-project/vllm', number: 46412 },
+          { repo: 'vllm-project/vllm', number: 46906 },
+        ],
+      },
+      {
+        id: 'load-path',
+        heading: 'Keeping lookups off the critical path',
+        paragraphs: [
+          'The load path was tuned separately, because lookups happen on every scheduling decision rather than only when data actually moves. Making lookups asynchronous in the scheduler path keeps the connector off the step’s critical path, so a step no longer waits on CPU-side cache queries before it can admit work.',
+          'Compact zero-copy lookup keys, parallel receive-side loading, and prebuilt Mooncake key strings then removed the CPU and transport overhead that remained.',
+        ],
+        prs: [
+          { repo: 'vllm-project/vllm', number: 45659 },
+          { repo: 'vllm-project/vllm', number: 45969 },
+          { repo: 'vllm-project/vllm', number: 45971 },
+          { repo: 'vllm-project/vllm', number: 46188 },
+        ],
+      },
+      {
+        id: 'correctness',
+        heading: 'Correctness and accounting under long-lived hybrid state',
+        paragraphs: [
+          'Long-lived hybrid state forced correctness and accounting fixes that fixed-shape requests rarely reach. vLLM now emits cache events per hybrid cache group, strides distributed-context stores correctly, and computes lookup prefixes correctly under distributed context and prefill. A related context-parallel accounting change aligns cache ownership with sharded token ranges.',
+          'Speculative state is now propagated across merged Mooncake groups and through the SimpleCPU coordinator, which prevents the repeated-turn cache from silently losing EAGLE state.',
+        ],
+        prs: [
+          { repo: 'vllm-project/vllm', number: 44103 },
+          { repo: 'vllm-project/vllm', number: 45371 },
+          { repo: 'vllm-project/vllm', number: 46855 },
+          { repo: 'vllm-project/vllm', number: 45340 },
+          { repo: 'vllm-project/vllm', number: 49069 },
+          { repo: 'vllm-project/vllm', number: 49071 },
+        ],
+      },
+      {
+        id: 'rocm-decode',
+        heading: 'ROCm: below the cache layer',
+        paragraphs: [
+          'On the ROCm side the recent work continues below the cache layer, where the remaining cost is per-layer rather than per-request. Once the prefix survives and arrives on time, what is left is the decode step itself, and a decode step that runs thousands of times per session pays for every avoidable copy and every mismatched kernel. Three open changes attack that layer.',
+          'A Kimi-K3 change writes KDA decode results directly into the layer output buffer, removing one device copy per KDA layer; the saving is small in isolation and repeated for every layer of every decoded token. A second change selects an AITER sparse-MLA decode kernel in place of the generic path, and reported 5.22% higher AgentX output throughput with substantially lower inter-token latency.',
+          'The third is a useful illustration of how much the measurement shape matters. A companion change routes full-graph attention projections through tuned AITER GEMMs and reached a 2.3% gain on fixed sequences at low concurrency. A kernel-level change can show a clean gain on uniform shapes and then be swamped, on an agentic trace, by the cache and scheduling variance the trace introduces.',
+        ],
+        prs: [
+          { repo: 'vllm-project/vllm', number: 51183 },
+          { repo: 'vllm-project/vllm', number: 51714 },
+          { repo: 'vllm-project/vllm', number: 51713 },
+        ],
+      },
+    ],
+  },
+  {
+    slug: 'sglang',
+    name: 'SGLang',
+    layer: 'engine',
+    summary:
+      'Sliding-window allocation, HiCache hybrid offload, runtime-scalar context length, and cache-aware DP routing.',
+    lead: 'Working alongside SGLang maintainers from RadixArk, Meta, NVIDIA, and AMD, the AgentX initiative drove optimizations for agentic workloads that also improve production serving substantially.',
+    highlights: [
+      { value: '+26.75%', label: 'output throughput at concurrency 384' },
+      { value: '−36.25%', label: 'mean TTFT from the same change' },
+      { value: '128/128', label: 'needles correct after the staging fix, from 2/128' },
+    ],
+    sections: [
+      {
+        id: 'sliding-window',
+        heading: 'Sliding-window allocation',
+        paragraphs: [
+          'SGLang’s sliding-window work addresses the same conflict as vLLM’s retention policy, but from the allocator side. Window pages and prefix pages are drawn from one pool, and the window is the greedier consumer: it turns over constantly while the prefix sits still, so under pressure the transient allocation displaces the durable one.',
+          'Three designs attack that from different angles. One proactively frees pages as they leave the window rather than waiting for eviction pressure to find them, so dead window state stops competing for pages it can no longer use. Another caps compute locks to a single window, bounding how much of the pool an in-flight request can hold pinned at once. A third removes stale full-KV entries that outlive their usefulness.',
+          'Alongside those, the ROCm ring-cache fix is a correctness change rather than a capacity change: a ring buffer reuses slots by construction, and reusing one whose old contents are still referenced yields wrong output rather than slow output. None of this is visible on a single 8k prompt, where the window never laps the prefix and the pool is never contended. On a multi-turn hybrid session, these changes decide whether the expensive full-attention history is still there on the next turn.',
+        ],
+        prs: [
+          { repo: 'sgl-project/sglang', number: 26907 },
+          { repo: 'sgl-project/sglang', number: 27210 },
+          { repo: 'sgl-project/sglang', number: 29369 },
+          { repo: 'sgl-project/sglang', number: 30339 },
+        ],
+      },
+      {
+        id: 'hicache',
+        heading: 'HiCache offload',
+        paragraphs: [
+          'HiCache is SGLang’s first-class in-tree offloading mechanism. It faced the same hybrid problem vLLM’s connectors did, and solved it with an asymmetry: offload the full-attention cache and reconstruct the short sliding-window tail on the way back. Only the expensive half is worth moving across the bus, and the cheap half can be rebuilt faster than it can be fetched. On AMD, staged write-back keeps that movement from blocking the engine while it happens.',
+          'Recurrent state was the remaining gap, because it cannot be rebuilt from neighbouring tokens the way a window tail can. FlashInfer GDN checkpoints let it participate in prefix reuse at all, and raised throughput from 47,771 to 53,004 tok/s/GPU at a 92.4% cache-hit rate.',
+        ],
+        figure: {
+          key: 'sglangHiCacheAsymmetry',
+          alt: 'Diagram showing three kinds of state crossing from GPU HBM to CPU DRAM: full-attention KV is offloaded byte for byte, the sliding-window tail is not moved and is rebuilt on load, and recurrent state is checkpointed.',
+          caption:
+            'HiCache’s asymmetry. Full-attention KV is large and not reconstructible, so it crosses the bus; the window tail is cheap to rebuild and stays behind; recurrent state is small but cannot be reconstructed, so it is checkpointed.',
+        },
+        prs: [
+          { repo: 'sgl-project/sglang', number: 29417 },
+          { repo: 'sgl-project/sglang', number: 28534 },
+          { repo: 'sgl-project/sglang', number: 29735 },
+        ],
+      },
+      {
+        id: 'variable-length',
+        heading: 'What variable-length traffic does to a kernel pipeline',
+        paragraphs: [
+          'Like production traffic, AgentX sessions arrive at continuously varying context lengths. A naive runtime that specializes on length will compile a fresh kernel for nearly every request it sees. SGLang maintainers solved this by passing context length as a runtime scalar, collapsing that into one compilation, and improved AgentX output throughput at concurrency 384 by 26.75% and mean TTFT by 36.25% — by removing compilation, not by computing anything faster.',
+          'In the same spirit, removing a per-step device-to-host sequence-length synchronization eliminates a decode bubble that exists only because the host wanted to know a length the device already had.',
+        ],
+        prs: [
+          { repo: 'sgl-project/sglang', number: 30255 },
+          { repo: 'sgl-project/sglang', number: 30365 },
+        ],
+      },
+      {
+        id: 'routing',
+        heading: 'Cache-aware data-parallel routing',
+        paragraphs: [
+          'When a request carries no reusable history, such as the start of a subagent, any worker will do and load balancing is the only question worth asking. When it carries a megabyte of cached prefix, sending it to an idle worker that does not hold that prefix is the expensive choice, and the router needs to know where the state already lives.',
+          'SGLang added DP cache affinity so a session is sticky to the rank holding its cache. The same work implements DP-aware prefill and decode routing so both halves of a disaggregated deployment make that decision consistently, and adds cache balance as a routing signal so affinity does not degenerate into one hot worker. A router can only act on what it is told, so hybrid cache events also became radix-cache aware and sliding-window aware.',
+        ],
+        prs: [
+          { repo: 'sgl-project/sglang', number: 26091 },
+          { repo: 'sgl-project/sglang', number: 26245 },
+          { repo: 'sgl-project/sglang', number: 26293 },
+          { repo: 'sgl-project/sglang', number: 26387 },
+          { repo: 'sgl-project/sglang', number: 26579 },
+        ],
+      },
+      {
+        id: 'speculative',
+        heading: 'Speculative decoding',
+        paragraphs: [
+          'Speculative decoding receives special attention, because MTP adds a second, smaller piece of per-request state that has to survive everything the main cache survives. SGLang fixed draft-window transfer in disaggregated serving so that state crosses the prefill-to-decode boundary intact, added overlap scheduling for high-concurrency online decoding, removed a no-op EAGLE renormalization, and avoided host synchronizations during EAGLE prefill.',
+          'The open resource-lease scheduling work and the data-parallel graph-metadata fix continue the same effort: making overlap safe when requests can be retracted and resumed rather than simply run to completion.',
+        ],
+        prs: [
+          { repo: 'sgl-project/sglang', number: 30461 },
+          { repo: 'sgl-project/sglang', number: 30497 },
+          { repo: 'sgl-project/sglang', number: 31294 },
+          { repo: 'sgl-project/sglang', number: 33662 },
+          { repo: 'sgl-project/sglang', number: 32042 },
+          { repo: 'sgl-project/sglang', number: 32196 },
+        ],
+      },
+      {
+        id: 'staging',
+        heading: 'Prefix-aware staging in heterogeneous disaggregation',
+        paragraphs: [
+          'Heterogeneous prefill and decode topologies need prefix-aware staging, and this is where prefix caching and disaggregation interact badly. When the two sides are not sharded identically, KV cannot be copied across as one contiguous stream; it has to be split on a transfer grid and reassembled at the offsets the decode side expects. A prefix hit makes that harder, not easier, because the prefill worker now sends only the uncached remainder while the decode side still expects a complete, correctly positioned cache. Radix-cache support in the staging buffer splits cached sends on that grid and scatters them at the correct decode offsets.',
+          'The failure this fixes is a correctness failure. A 127,500-token shared-prefix test went from 2 correct needles out of 128 to 128 out of 128, meaning the cache had been silently landing in the wrong places — which a throughput benchmark would have scored as a fast, confident, wrong answer. The AgentX comparison additionally raised median per-user output throughput by 9.61% at nearly unchanged total throughput per GPU.',
+          'Open work continues along the same seam: multi-pool DeepSeek-V4 support in UMBP, unified-KV HiSparse state carried over MoRI, and preserving the prefill-owned token when decode terminates without visible content. The HiSparse work should be read as a capacity and correctness enabler for long contexts rather than as a throughput win at high concurrency, which it is not yet.',
+        ],
+        figure: {
+          key: 'sglangRadixStaging',
+          alt: 'Before and after diagram of staging-buffer indexing: before, the uncached remainder lands on top of the prefix at the wrong rows; after, it is placed at the offsets after the prefix the decode side already holds.',
+          caption:
+            'Before, the staging index counts from zero and ignores what the decode side already holds, so the remainder overwrites the prefix. After radix-aware staging, cached sends are split on the transfer grid and scattered at the correct decode offsets.',
+        },
+        prs: [
+          { repo: 'sgl-project/sglang', number: 30545 },
+          { repo: 'sgl-project/sglang', number: 30762 },
+          { repo: 'sgl-project/sglang', number: 32368 },
+          { repo: 'sgl-project/sglang', number: 34216 },
+        ],
+      },
+    ],
+  },
+  {
+    slug: 'tensorrt-llm',
+    name: 'TensorRT-LLM',
+    layer: 'engine',
+    summary:
+      'Boundary-aware incremental tokenization, disaggregated KV descriptor coalescing, and scheduler-lifetime fixes.',
+    lead: 'TensorRT-LLM’s AgentX work starts at the frontend, with a cost that exists only because the workload is multi-turn, and then follows the request down into transfer granularity, kernel selection, and scheduler lifetime.',
+    highlights: [
+      { value: '185.1 ms → 11.3 ms', label: 'mean tokenization time per turn' },
+      { value: '26.74 s → 125 ms', label: 'request-critical KV p99 at concurrency 5' },
+      { value: '+12.58%', label: 'per-user output throughput from context graphs' },
+    ],
+    sections: [
+      {
+        id: 'incremental-tokenization',
+        heading: 'Boundary-aware incremental tokenization',
+        paragraphs: [
+          'Every turn of a conversation re-sends the entire history plus a little more, and the naive implementation re-tokenizes all of it. Tokenization is cheap per kilobyte and ruinous when the same 100,000 tokens are tokenized again on every turn.',
+          'The obvious fix — tokenize only the new suffix — is wrong in a way that is easy to miss, because byte-pair encoding is not position-independent. Tokens can merge across the join, so splitting the text at the boundary and concatenating the two token sequences can produce a different sequence than tokenizing the whole string, which quietly diverges from the sequence the prefix cache was built against.',
+          'TensorRT-LLM implements boundary-aware incremental tokenization, which finds the rendered-text common prefix, rolls back one complete token so any merge that spans the join is recomputed, and tokenizes only the changed suffix from there. On the Qwen3.5 AgentX trace it matched full tokenization on all 1,087 transitions — the correctness claim tested rather than assumed — and reduced mean processing time from 185.1 ms to 11.3 ms. A fixed 8k/1k request has no prior rendered turn to reuse, so none of this appears there.',
+          'Relatedly, chat-template rendering moved into the input-processing pool, so a long template no longer serializes the main request loop behind it.',
+        ],
+        prs: [
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17462 },
+          { repo: 'NVIDIA/TensorRT-LLM', number: 16231 },
+        ],
+      },
+      {
+        id: 'disaggregated-kv',
+        heading: 'Disaggregated KV movement and descriptor granularity',
+        paragraphs: [
+          'The MiniMax-M3 work focuses on disaggregated KV movement, where the failure is one of granularity. When prefill and decode do not agree on head layout, the KV for one logical request stops being a few large contiguous regions and becomes thousands of small strided pieces, each of which turns into its own transfer descriptor. The bytes moved are unchanged; the per-descriptor overhead is what explodes, and it explodes worst on exactly the long prompts that matter.',
+          'Corrected multi-pool mapping and a chunked NIXL bounce path coalesce those pieces through a bounded reusable arena, trading an extra staging copy for orders of magnitude fewer descriptors. The AgentX diagnostic reduced request-critical KV p99 from 26.74 seconds to 125 ms at concurrency five, and from 10.15 seconds to 288 ms at concurrency forty.',
+          'Nonblocking context-transfer polling protects the same path by reaping completed transfers even when scheduling stalls, breaking a feedback loop in which finished KV blocks stay pinned and prevent new admissions. A separate draft-cache transfer proposal was closed without merge and should not be counted as shipped TensorRT-LLM behavior.',
+        ],
+        prs: [
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17518 },
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17428 },
+        ],
+      },
+      {
+        id: 'execution-paths',
+        heading: 'Execution paths for irregular long context',
+        paragraphs: [
+          'TensorRT-LLM also moved irregular long-context work onto more efficient execution paths. Context graph producers for MiniMax-M3 capture stable sparse producers while leaving request-dependent attention eager, and per-user output throughput improved by 12.58% in its AgentX test. An open native KV-event production change reduces allocation and conversion work on the KV-aware routing path.',
+        ],
+        prs: [
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17473 },
+          { repo: 'NVIDIA/TensorRT-LLM', number: 16876 },
+        ],
+      },
+      {
+        id: 'kernel-selection',
+        heading: 'Kernel selection and scheduler lifetime',
+        paragraphs: [
+          'AgentX also exposed kernel-selection and scheduler-lifetime failures that only appear at scale and duration. Two are about which kernel gets picked. MiniMax-M3 added CuTeDSL choices to MXFP8 autotuning, widening the candidate set and improving output throughput per GPU by roughly 7 to 10% at low-concurrency aggregated points. In the opposite direction, TensorRT-LLM disabled corrupt split-K MoE tactics after they crashed five of seven AgentX runs, with no crashes in seven matched runs afterwards. A tactic that is fast and wrong is worse than one that is merely slow, and an autotuner will select it enthusiastically unless it is removed from the pool.',
+          'The other two are lifetime bugs, which are the characteristic failure of long runs rather than large ones. Sequence-slot headroom and consistent slot-indexed buffer sizing handle the transient overlap where a completing request and a newly admitted one both need a slot — a window that a steady stream of arrivals and departures hits constantly and a fixed batch never hits at all. A later attention-data-parallel dummy-request fix kept nine Qwen3.5 disaggregated cells alive where most earlier cells had failed within minutes: the difference between a configuration that benchmarks and one that survives a session.',
+        ],
+        prs: [
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17316 },
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17105 },
+          { repo: 'NVIDIA/TensorRT-LLM', number: 16279 },
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17278 },
+        ],
+      },
+      {
+        id: 'pipelined-transfer',
+        heading: 'Pipelined KV transfer for very long prompts',
+        paragraphs: [
+          'Two open transfer changes target very long disaggregated prompts, and together they show how a fix can create the next bottleneck. In the default arrangement a decode worker cannot start until the entire prompt has been prefilled and then transferred, so two expensive phases run back to back even though the first produces its output incrementally. Pipelined KV transfer begins sending each completed prefill chunk as it lands, so transfer overlaps prefill compute and only the final chunk is on the critical path.',
+          'That change makes chunk handling frequent, which exposes work that used to happen once. Its follow-up retrieves only the block IDs belonging to the current chunk rather than the whole prompt’s block list each time. For a 128,000-token prompt split into 1,024-token chunks, that is the difference between building a 4,096-entry list once and rebuilding it 128 times for every layer group — a per-chunk cost that scales with total prompt length, which is exactly the shape that eats the gain the pipelining just bought.',
+        ],
+        figure: {
+          key: 'trtllmPipelinedTransfer',
+          alt: 'Timeline diagram: before, four prefill chunks complete and then the entire prompt is transferred before decode starts; after, each finished chunk is transferred while the next is computed, so decode starts earlier.',
+          caption:
+            'Before, prefill and transfer run back to back and decode waits for both. With pipelined transfer each finished chunk is sent while the next is computed, so only the final chunk sits on the critical path.',
+        },
+        prs: [
+          { repo: 'NVIDIA/TensorRT-LLM', number: 15727 },
+          { repo: 'NVIDIA/TensorRT-LLM', number: 17526 },
+        ],
+      },
+    ],
+  },
+  {
+    slug: 'atom',
+    name: 'AMD ATOM',
+    layer: 'engine',
+    summary:
+      'Sparse checkpoint retention, recurrent-state checkpoints, CPU offload ownership, and long-prefill parallelism.',
+    lead: 'AMD’s ATOM engine was designed for single-turn workloads rather than real-world multi-turn agentic production traffic, so supporting long-context multi-turn work required changes to the core engine and its kernels. ATOM still has a long way to go relative to where vLLM and SGLang are, and AgentX is the realistic north star for its refactor.',
+    highlights: [
+      { value: '5.6% → 96.45%', label: 'prefix hit rate at concurrency 48' },
+      { value: '91.35% → 0.16%', label: 'losses at the sliding-window gate' },
+      { value: '28.6 s → 8.7 s', label: 'median TTFT with chunked PP prefill' },
+    ],
+    sections: [
+      {
+        id: 'sparse-checkpoints',
+        heading: 'Sparse checkpoint retention',
+        paragraphs: [
+          'ATOM implemented sparse checkpoint retention for DeepSeek-V4 paged sliding-window attention, which shows the problem is a property of the workload rather than of any one codebase. The merged implementation keeps selected window tails alive so branch and replay requests can resume at useful boundaries.',
+          'Its measurements separate the two effects cleanly: on the same AgentX trace at concurrency 48, the actual prefix hit rate rose from 5.6% to 96.45%, and losses at the sliding-window gate fell from 91.35% to 0.16%. The second number is the mechanism behind the first. Nine out of ten prefix matches were being found and then discarded for want of a window tail, so the cache was not missing — it was being overruled.',
+        ],
+        prs: [{ repo: 'ROCm/ATOM', number: 1640 }],
+      },
+      {
+        id: 'cache-manager',
+        heading: 'Cache-manager fixes that had to land first',
+        paragraphs: [
+          'Two earlier cache-manager fixes had to land before any of that could be measured, and both are worth noting as examples of a cache that reports itself healthy while doing nothing. One stopped free-pool hits from destroying shared cache entries. The other, a deferred-output fix, restored prefix hashing in the default scheduler mode and moved repeated long prompts from zero cached tokens to reuse of every complete prefix block.',
+          'A separate change lets prefix-hit prefill stay on the optimized sink attention kernel rather than falling back to the generic path, so a cache hit does not quietly cost part of what it saves.',
+        ],
+        prs: [
+          { repo: 'ROCm/ATOM', number: 902 },
+          { repo: 'ROCm/ATOM', number: 939 },
+          { repo: 'ROCm/ATOM', number: 1345 },
+        ],
+      },
+      {
+        id: 'recurrent-state',
+        heading: 'Recurrent-state checkpoints',
+        paragraphs: [
+          'Hybrid models also carry a recurrent or compressor state, which differs from ordinary KV in one decisive way: it cannot be reconstructed from the tokens around it. A window tail can be recomputed from neighbouring context, but recurrent state is the accumulated result of everything that came before, so if it is dropped the only way back is to replay the sequence.',
+          'ATOM gave this per-request state a content-addressed checkpoint lifecycle, letting generated turns leave reusable resume points without reserving a separate protected cache for them. In one test a request reused 512 generated tokens and computed only a two-token suffix.',
+          'The tuning detail matters as much as the feature. Publishing a checkpoint unconditionally cost 17.5% throughput on zero-hit traffic — the price paid by every session that never comes back, in order to help the ones that do. Spacing checkpoints by token interval avoided that penalty, and fixed 1k/1k throughput stayed within measurement noise, which is the relevant safety property: a feature aimed at agentic reuse should not tax workloads that will never use it.',
+        ],
+        prs: [{ repo: 'ROCm/ATOM', number: 1771 }],
+      },
+      {
+        id: 'cpu-path',
+        heading: 'The CPU offload path: ownership and index placement',
+        paragraphs: [
+          'ATOM’s AgentX-relevant CPU path starts from the arithmetic that justifies offload at all. Standalone LMCache offload reloads a 32,000-token prefix from CPU in about 0.32 seconds against roughly 2.5 seconds to recompute it, an eight-fold margin. That makes crossing the bus worth doing at these context lengths, and it would not hold for a short prompt.',
+          'The rest of the path is about ownership and index placement rather than bandwidth. ATOM copied vLLM’s multi-connector design, which lets a prefill worker send KV to a remote decode worker and save the same prefix to CPU at once, without freeing the blocks until both consumers are finished. Two independent readers of the same blocks is a situation single-turn traffic never produces.',
+          'Promoting restored blocks back into the GPU prefix index fixes a subtler waste: without it, a prefix loaded from CPU is used and then not registered as resident, so the next turn fetches the same hot prefix across the bus again, paying the transfer repeatedly for a cache that was already in HBM. Follow-up work fixed asynchronous save ordering, packed-KV geometry, unaligned handoffs, and remote request accounting together, eliminating reload corruption across a two-round, 2,638-request validation. That bug surfaces only when the same blocks are saved, evicted, and restored many times over.',
+        ],
+        figure: {
+          key: 'atomPrefixPromotion',
+          alt: 'Flow diagram: a prefix restored from CPU DRAM into GPU blocks is either dropped from the GPU prefix index, so the next turn fetches it again, or promoted into the index so the next turn hits in HBM.',
+          caption:
+            'Reloading a 32,000-token prefix costs about 0.32 s against roughly 2.5 s to recompute it. Promoting restored blocks into the GPU prefix index is what stops the next turn paying that transfer again.',
+        },
+        prs: [
+          { repo: 'ROCm/ATOM', number: 1318 },
+          { repo: 'ROCm/ATOM', number: 1406 },
+          { repo: 'ROCm/ATOM', number: 1725 },
+          { repo: 'ROCm/ATOM', number: 1807 },
+        ],
+      },
+      {
+        id: 'routing',
+        heading: 'Cache-aware routing in ATOM Mesh',
+        paragraphs: [
+          'The distributed path repeats, in a different codebase, the pattern already visible in SGLang and Dynamo: routing has to know where state lives. ATOM’s router, ATOM Mesh, is a fork of SGLang’s router with most features removed — including SGLang’s cache-aware routing, which it turned out to need.',
+          'ATOM gained KV lifecycle events for cache-aware routers so the router can know where state lives at all, multi-node prefill and decode routing, and session-sticky data-parallel routing. The sticky policy is a two-sided compromise worth stating explicitly: a conversation returns to the healthy worker that owns its state, but idle assignments expire so that stickiness does not permanently unbalance the cluster on behalf of sessions that have gone away.',
+        ],
+        prs: [
+          { repo: 'ROCm/ATOM', number: 869 },
+          { repo: 'ROCm/ATOM', number: 919 },
+          { repo: 'ROCm/ATOM', number: 1699 },
+        ],
+      },
+      {
+        id: 'disaggregation',
+        heading: 'Disaggregation has to move whatever the model keeps',
+        paragraphs: [
+          'Disaggregation has to move whatever the model actually keeps, which is not always one uniform cache. DeepSeek-V4 transfers both buffers of its mixed FP8 and BF16 cache layout, and EAGLE disaggregation moves the draft model’s independent KV cache alongside the target cache — the same second-cache problem TensorRT-LLM and SGLang each had to solve.',
+          'Remote-KV admission and backpressure closes the loop by stopping the decode side from accepting more parked transfers than it can safely resume, which is the disaggregated form of accepting work you cannot finish.',
+        ],
+        prs: [
+          { repo: 'ROCm/ATOM', number: 1737 },
+          { repo: 'ROCm/ATOM', number: 1331 },
+          { repo: 'ROCm/ATOM', number: 1647 },
+        ],
+      },
+      {
+        id: 'long-prefill',
+        heading: 'Parallelism for long prefill',
+        paragraphs: [
+          'Long prefill receives parallelism that a fixed 8k prompt does not strongly exercise, because at 8k there is little to divide and TTFT is already short. Prefill context parallelism (PCP) splits DeepSeek-V4 query tokens across GPUs and reported 35 to 43% lower mean time to first token, with total throughput gains of up to about 49% at a 64,000-token input — a gain that grows with input length rather than with batch size.',
+          'Making that usable in practice required it to compose with everything else a session relies on, so decode context parallelism was made compatible with prefix caching, chunked prefill, and FP8 KV, and then extended to MTP. Parallelism that cannot coexist with the prefix cache would trade one long-context win for another.',
+          'Chunked pipeline-parallel prefill attacks the same problem from the memory side, replacing repeated tensor-parallel collectives with streamed layer-stage handoffs. Its GLM-5.2 result at high load is the most complete in this section: output throughput doubled, median time to first token fell from 28.6 seconds to 8.7 seconds, and each prefill GPU held 3.68 times as many KV blocks. That last figure is the one to read first, because capacity per prefill GPU decides how many long sessions can be in flight before the deployment hits the HBM cliff at all.',
+        ],
+        prs: [
+          { repo: 'ROCm/ATOM', number: 1220 },
+          { repo: 'ROCm/ATOM', number: 1701 },
+          { repo: 'ROCm/ATOM', number: 1746 },
+          { repo: 'ROCm/ATOM', number: 1552 },
+        ],
+      },
+    ],
+  },
+  {
+    slug: 'aiter',
+    name: 'ROCm AITER',
+    layer: 'kernels',
+    summary:
+      'Context-parallel process groups, 64-bit addressing for large cache pools, and persistent MLA decode kernels.',
+    lead: 'Long-context execution in ATOM, AMD vLLM, and AMD SGLang depends on matching lower-level AITER kernels, because a parallelism strategy at the engine layer is only real if the kernels can express it.',
+    highlights: [
+      { value: '>131k', label: 'tokens covered by widened row indexing' },
+      { value: '64-bit', label: 'addressing for pools above 4 GB' },
+      { value: '~150M', label: 'rows protected from silent wrong-row access' },
+    ],
+    sections: [
+      {
+        id: 'context-parallel',
+        heading: 'Context-parallel process groups',
+        paragraphs: [
+          'Prefill context-parallel process groups provide the extra query-sharding dimension that prefill context parallelism needs, and also widen fused-kernel row indexing for prompts above 131,000 tokens. Decode context parallelism (DCP) shards KV across the tensor-parallel GPUs already present, so a longer sequence or a larger batch fits without replicating the whole cache on every rank.',
+        ],
+        prs: [
+          { repo: 'ROCm/aiter', number: 3728 },
+          { repo: 'ROCm/aiter', number: 3267 },
+        ],
+      },
+      {
+        id: 'address-width',
+        heading: 'Address width: a failure short requests never reach',
+        paragraphs: [
+          'Large caches exposed a class of failure that short fixed requests essentially never reach: address width. A 32-bit offset is entirely adequate until a single cache pool crosses the boundary, at which point the arithmetic wraps and the kernel addresses the wrong row without any error being raised.',
+          'AITER added runtime 64-bit dispatch for batch prefill above 4 GB, 64-bit MLA offsets above 2 GB, and 64-bit addressing throughout DeepSeek-V4’s unified cache paths — the last preventing silent reads and writes to the wrong row in pools of roughly 150 million rows.',
+        ],
+        figure: {
+          key: 'aiter64BitOffsets',
+          alt: 'Two panels of a unified KV pool of about 150 million rows: with 32-bit offsets an access past 4 GB wraps to a different row; with 64-bit offsets the addressed row is the intended one.',
+          caption:
+            'A 32-bit offset is adequate until one cache pool crosses 4 GB, at which point the arithmetic wraps and the kernel reads or writes a different row without raising an error.',
+        },
+        prs: [
+          { repo: 'ROCm/aiter', number: 2893 },
+          { repo: 'ROCm/aiter', number: 4474 },
+          { repo: 'ROCm/aiter', number: 4680 },
+        ],
+      },
+      {
+        id: 'persistent-mla',
+        heading: 'Persistent MLA decode for common head packings',
+        paragraphs: [
+          'DeepSeek-V4 decode also gained a persistent MLA kernel for 64-head and 128-head MTP packings. Those two head counts are what ordinary decoding and speculative verification actually produce, so this gives the engine a dedicated long-context path for its common shapes instead of treating them as incidental variants of a kernel written for short contexts. It is the same argument as the vLLM AITER sparse-MLA selection: at long context the generic path is not a modest compromise, it is the wrong kernel.',
+        ],
+        prs: [{ repo: 'ROCm/aiter', number: 3459 }],
+      },
+    ],
+  },
+  {
+    slug: 'dynamo',
+    name: 'NVIDIA Dynamo',
+    layer: 'router',
+    summary:
+      'Batched KV matching, request-lease ownership, cheaper router state, and a leaner request plane.',
+    lead: 'A good part of the NVIDIA submission uses the Dynamo inference orchestration and router system. Its AgentX series shows that the distributed serving layer can become the bottleneck once engine kernels improve: the router’s work is proportional to the number and length of live prefixes rather than to the number of tokens generated, so many long, overlapping, long-lived sessions load it in a way fixed-shape traffic never does.',
+    highlights: [
+      { value: '+22.2%', label: 'median output throughput at concurrency 512' },
+      { value: '−23.71%', label: 'AgentX replay time with request leases (vLLM)' },
+      { value: '932 → 1,133', label: 'frontend requests per second' },
+    ],
+    sections: [
+      {
+        id: 'routing-cost',
+        heading: 'The cost of each routing decision',
+        paragraphs: [
+          'The first series of pull requests reduced the cost of each routing decision: less work on the lookup hot path, no redundant suffix invalidation, and finally batched KV matching, registration, ownership, and terminal dereferences, which reported a 22.2% median output-throughput gain at concurrency 512. Batching helps here for the same reason it helps in an engine: the per-item overhead was dominating the item.',
+        ],
+        prs: [
+          { repo: 'ai-dynamo/dynamo', number: 10540 },
+          { repo: 'ai-dynamo/dynamo', number: 10836 },
+          { repo: 'ai-dynamo/dynamo', number: 11095 },
+        ],
+      },
+      {
+        id: 'ownership',
+        heading: 'How ownership is represented',
+        paragraphs: [
+          'The second series changed how ownership is represented, which is the harder problem underneath. Every cached block needs to be attributed to the requests relying on it, so it is not freed while still in use and not pinned after everyone has finished. With thousands of concurrent sessions sharing overlapping prefixes, the bookkeeping itself becomes significant.',
+          'Dynamo moved from shared block chains to arena-level ownership counts and finally to backend-specific request leases, each step coarsening the unit being tracked. The lease design reduced AgentX replay time by 23.71% for the vLLM backend and 22.02% for SGLang, and lowered peak memory at the same time — a sign that the previous representation was the problem rather than the traffic.',
+        ],
+        figure: {
+          key: 'dynamoOwnershipTable',
+          alt: 'Before and after diagram of block ownership: before, each block holds its own reference counter and requests point at every block they use; after, ownership counts live in one table with rows, stamps, and parent rows, and each request keeps a bookmark.',
+          caption:
+            'Ownership representation, before and after. Counts move out of the individual blocks and into one table, so a request keeps a bookmark rather than a reference to every block it touches.',
+        },
+        prs: [
+          { repo: 'ai-dynamo/dynamo', number: 11503 },
+          { repo: 'ai-dynamo/dynamo', number: 11508 },
+          { repo: 'ai-dynamo/dynamo', number: 12329 },
+        ],
+      },
+      {
+        id: 'router-state',
+        heading: 'Router state that used to be small',
+        paragraphs: [
+          'Further router profiles removed costs with the same shape, where a periodic sweep or a full recomputation had been acceptable only because live state used to be small. Bucketed expiry pruning replaced a scan proportional to everything tracked and improved high-churn AgentX throughput by 13.7%. Delta-only suffix cleanup processes only what changed and absorbed about 28 times as many store and remove events in the same window. Compressed prompt paths cut frontend CPU by 35.3% and materially improved tail time to first token, which matters because prompts in this workload are long and largely repeated. Overload state is now tracked incrementally rather than recomputed.',
+          'One routing change is a deliberate trade rather than a pure win. Dynamo can now charge active decode requests in its routing score, so a worker already committed to long-running decodes looks more expensive than its queue depth alone suggests. That improved median AgentX latency at a small throughput cost in the reported tuning point — the kind of choice that only becomes visible when requests occupy a worker for a long time.',
+        ],
+        prs: [
+          { repo: 'ai-dynamo/dynamo', number: 10521 },
+          { repo: 'ai-dynamo/dynamo', number: 10676 },
+          { repo: 'ai-dynamo/dynamo', number: 11644 },
+          { repo: 'ai-dynamo/dynamo', number: 10645 },
+          { repo: 'ai-dynamo/dynamo', number: 12158 },
+        ],
+      },
+      {
+        id: 'request-plane',
+        heading: 'The request plane',
+        paragraphs: [
+          'The request plane was optimized next, because an agentic trace does not send one request and one response. It sends many related requests carrying largely identical prompts, and streams every token back as its own frame, so serialization and copying are paid per turn and per token rather than once. Switching to MessagePack request payloads improved throughput by 8.1% and reduced average time to first token by 9.7% in its AgentX test, and direct Python transcoding removed an intermediate value tree from that path entirely.',
+          'What followed is a sequence of changes that all remove a copy rather than speed one up: not copying MessagePack event payloads, not copying received ZeroMQ frames, and not paying full inter-token-latency metrics overhead on every token. The chat streaming hot path was shortened for the same reason. Individually these are unremarkable; multiplied by every streamed token of every concurrent session, they determine how many requests per second a frontend can sustain.',
+        ],
+        prs: [
+          { repo: 'ai-dynamo/dynamo', number: 10437 },
+          { repo: 'ai-dynamo/dynamo', number: 11104 },
+          { repo: 'ai-dynamo/dynamo', number: 11539 },
+          { repo: 'ai-dynamo/dynamo', number: 11574 },
+          { repo: 'ai-dynamo/dynamo', number: 11569 },
+          { repo: 'ai-dynamo/dynamo', number: 10433 },
+        ],
+      },
+      {
+        id: 'profiling',
+        heading: 'Costs that had nothing to do with moving data',
+        paragraphs: [
+          'High-concurrency profiling then found costs that had nothing to do with moving data. Static logging filters removed a shared span-matcher lock — a contention point rather than a volume problem — and raised reported frontend throughput from 932 to 1,133 requests per second. Simpler positional radix buckets reduced peak memory in the mocker by 5.51 GiB in a 32-worker run.',
+          'An open change flushes detokenization metrics once per response rather than updating cumulative counters on every streamed chunk, approximately halving frontend CPU time in its matched diagnostic profile. That last one is the clearest example of the category: the instrumentation was cheap per call and ruinous at one call per token.',
+        ],
+        prs: [
+          { repo: 'ai-dynamo/dynamo', number: 11820 },
+          { repo: 'ai-dynamo/dynamo', number: 12161 },
+          { repo: 'ai-dynamo/dynamo', number: 12999 },
+        ],
+      },
+    ],
+  },
+  {
+    slug: 'lmcache',
+    name: 'LMCache',
+    layer: 'kv-cache',
+    summary:
+      'Chunked external-cache loading, hybrid-group storage, AMD Instinct enablement, and DCP-aware offload.',
+    lead: 'LMCache is an open-source KV-cache layer that sits under inference engines such as vLLM, storing reusable KV chunks keyed by prefix hash across CPU DRAM, local NVMe, and remote backends including Mooncake, Redis, and S3. It can be used as an alternative to vLLM’s native offloading connectors.',
+    highlights: [
+      { value: '120 vs 28', label: 'requests completed before the old path deadlocked' },
+      { value: '~20×', label: 'less storage per token for hybrid groups' },
+      { value: '56/56', label: 'KV-transfer kernel tests passing on MI350X' },
+    ],
+    sections: [
+      {
+        id: 'chunked-loading',
+        heading: 'Chunked external-cache loading',
+        paragraphs: [
+          'LMCache’s multiprocess path was changed for the volume and shape of agentic cache movement, beginning with a failure that is not a slowdown but a stop. When each of many requests with contexts above 100,000 tokens reserves the blocks for its whole load before starting, the pool is exhausted by requests that are all waiting and none progressing.',
+          'Chunked external-cache loading reserves per chunk instead, so loads interleave and drain. At concurrency 32 the validation completed 120 requests where the old path deadlocked after 28, and concurrency 48 kept running with the KV pool 98.5% full.',
+        ],
+        figure: {
+          key: 'lmcacheTiers',
+          alt: 'Diagram of several vLLM instances above a shared LMCache layer with CPU DRAM, local NVMe through GDS, and a remote store tier, where a later request on another instance reads what an earlier one wrote.',
+          caption:
+            'LMCache sits under every instance as a hash-addressed KV layer across CPU DRAM, local NVMe, and a remote store, so a later request on a different instance can read what an earlier one wrote.',
+        },
+        prs: [{ repo: 'LMCache/LMCache', number: 3382 }],
+      },
+      {
+        id: 'moving-less',
+        heading: 'Moving less, and getting out of the way',
+        paragraphs: [
+          'The other changes reduce how much is moved and how often the runtime gets in the way. Storing only the useful portions of DeepSeek-V4’s hybrid groups cut storage per token by almost twenty times, and sliding-window prefetch now loads only the live window rather than window state that will never be read — the same reachability argument vLLM applied to offload, approached from the storage side.',
+          'One native transfer call per object group then removes repeated Python lock handoffs across staging copies and kernel launches, which is overhead proportional to the number of pieces rather than to the bytes in them.',
+        ],
+        prs: [
+          { repo: 'LMCache/LMCache', number: 3635 },
+          { repo: 'LMCache/LMCache', number: 3869 },
+          { repo: 'LMCache/LMCache', number: 3908 },
+        ],
+      },
+      {
+        id: 'lock-accounting',
+        heading: 'Open: hybrid lock accounting',
+        paragraphs: [
+          'Two current LMCache changes are especially specific to AgentX but remain open. The hybrid lock-accounting fix stops one request from releasing another request’s read locks on shared sliding-window or recurrent-state chunks. Reproducing it requires three things at once: several requests must share the same chunks, the accounting must be per-chunk rather than per-holder, and eviction must actually start.',
+          'Sustained Kimi-K3 runs with DRAM offload supplied all three and produced tens of thousands of warnings, corrupt generations, and eventually GPU crashes once eviction began. Anything short of a long, shared, memory-pressured run leaves the bug dormant.',
+        ],
+        prs: [{ repo: 'LMCache/LMCache', number: 4524 }],
+      },
+      {
+        id: 'amd-enablement',
+        heading: 'AMD Instinct enablement',
+        paragraphs: [
+          'A parallel line of work made all of the above reachable on AMD Instinct hardware. CacheBlend’s non-prefix reuse depended on FlashInfer, which is CUDA-only, so a Triton block-sparse attention backend reimplements the three kernels it needs: block-sparse attention with CSR indices and log-sum-exp output, causal prefill, and log-sum-exp output blending. It then routes to them automatically when ROCm is detected or FlashInfer is missing. ROCm Dockerfiles mirror the CUDA build and lightweight images. An AMD hipFile backend extends the GDS L1 slab-file tier, which reached storage only through NVIDIA cuFile, by binding ROCm’s hipFile through ctypes and dispatching on torch.version.hip; the cuFile path is unchanged.',
+          'Distribution was the remaining gap. CUDA users installed a prebuilt wheel while AMD users built from source. We worked with AMD to publish a prebuilt gfx942 and gfx950 wheel that closes it: it installs into the upstream image and passes all 56 KV-transfer kernel tests on MI350X, and it publishes to a GitHub release rather than PyPI so a plain pip install lmcache stays the CUDA build. A one-line follow-up marks the bind-mounted repository as a git safe directory, which only fails in CI because the container runs as root over a runner-owned checkout and the version introspection in setup.py refuses to read it.',
+        ],
+        prs: [
+          { repo: 'LMCache/LMCache', number: 3092 },
+          { repo: 'LMCache/LMCache', number: 3101 },
+          { repo: 'LMCache/LMCache', number: 3843 },
+          { repo: 'LMCache/LMCache', number: 4273 },
+          { repo: 'LMCache/LMCache', number: 4363 },
+        ],
+      },
+      {
+        id: 'dcp-offload',
+        heading: 'DCP-aware CPU offload',
+        paragraphs: [
+          'DCP-aware CPU offload resolves a straightforward incompatibility between two features that long contexts make mandatory together. With decode context parallelism enabled, each rank holds only a stride of the KV, so what any one rank could save is not a usable prefix. The fix gathers the strided shards before saving and redistributes them after loading. Without it, enabling context parallelism silently disables CPU cache hits for exactly the long prefixes that motivated both features. Its validation recorded more than 30,000 CPU hit events, with single-request loads reaching hundreds of thousands of tokens.',
+        ],
+        prs: [{ repo: 'LMCache/LMCache', number: 3561 }],
+      },
+    ],
+  },
+  {
+    slug: 'mooncake',
+    name: 'Mooncake',
+    layer: 'transfer',
+    summary:
+      'GPU-direct RDMA on ROCm through HIP dmabuf, plus a published ROCm wheel and release path.',
+    lead: 'Mooncake serves Moonshot’s Kimi production traffic along with production traffic at many labs, and is a transfer engine underneath disaggregated vLLM and SGLang configurations. Until recently its AMD support stopped short of both RDMA registration and installable packages.',
+    highlights: [
+      { value: 'HIP dmabuf', label: 'GPU-direct RDMA path on ROCm' },
+      { value: 'gfx942 + gfx950', label: 'covered by one architecture-agnostic wheel' },
+      { value: 'Python 3.10–3.13', label: 'tag-triggered publish matrix' },
+    ],
+    sections: [
+      {
+        id: 'rdma-registration',
+        heading: 'GPU-direct RDMA registration on ROCm',
+        paragraphs: [
+          'Registering GPU memory for RDMA on NVIDIA uses either the nvidia-peermem kernel module or an exported dmabuf file descriptor. AMD has no nvidia-peermem equivalent, so GPU-direct RDMA had no path at all and deployments fell back to staging KV through host DRAM.',
+          'A HIP dmabuf registration branch adds the mirror of the existing CUDA dmabuf path, exporting through ROCm instead of the CUDA handle call, and resolving the true allocation base first because caching allocators pack tensors at an offset inside a larger allocation. Host memory still registers directly.',
+        ],
+        prs: [{ repo: 'kvcache-ai/Mooncake', number: 2225 }],
+      },
+      {
+        id: 'packaging',
+        heading: 'Support that cannot be installed is not support',
+        paragraphs: [
+          'Mooncake published CUDA and MUSA wheels but no ROCm package, so AMD users built the engine from source inside every image. A ROCm wheel, CI, and release path publishes mooncake-transfer-engine-rocm to PyPI alongside them. This workstream came from Andy Luo, an AMD engineer, who noticed the pattern while dogfooding agentic workloads with AgentX: building Mooncake from source is not a first-class ROCm experience.',
+          'The transfer engine has no device kernels and does not depend on torch, so one architecture-agnostic wheel covers gfx942 and gfx950, and the ROCm runtime is bound at load time rather than vendored — which means the same wheel works unmodified in both the upstream vLLM ROCm image and the SGLang ROCm image. That was verified as a full cross product: MI300X and MI355X, each under vllm/vllm-openai-rocm and lmsysorg/sglang, running the master binary and a HIP buffer transfer test with data verification. The pull request adds a tag-triggered publish across Python 3.10 through 3.13.',
+          'An open follow-up adds a self-hosted two-node MI350X external prefill and decode tier so the ROCm disaggregated path is exercised on real hardware rather than only compiled. Together these mean an AMD AgentX run can now install the transfer engine and the KV-cache layer from published artifacts, into stock upstream images, and move KV directly between GPU memory and the fabric.',
+        ],
+        prs: [
+          { repo: 'kvcache-ai/Mooncake', number: 3184 },
+          { repo: 'kvcache-ai/Mooncake', number: 3338 },
+        ],
+        links: [
+          {
+            label: 'mooncake-transfer-engine on PyPI',
+            href: 'https://pypi.org/project/mooncake-transfer-engine/',
+          },
+        ],
+      },
+    ],
+  },
+];
+
+export const AGENTX_OPTIMIZATION_FRAMEWORKS = FRAMEWORKS;
+
+export const AGENTX_OPTIMIZATION_SLUGS: readonly string[] = FRAMEWORKS.map(
+  (framework) => framework.slug,
+);
+
+export function getOptimizationFramework(slug: string): OptimizationFramework | undefined {
+  return FRAMEWORKS.find((framework) => framework.slug === slug);
+}
+
+/** Total distinct upstream PRs referenced, used in the index-page copy. */
+export function countReferencedPrs(): number {
+  const seen = new Set<string>();
+  for (const section of OPTIMIZATIONS_OVERVIEW.sections) {
+    for (const pr of section.prs ?? []) seen.add(prUrl(pr));
+  }
+  for (const framework of FRAMEWORKS) {
+    for (const section of framework.sections) {
+      for (const pr of section.prs ?? []) seen.add(prUrl(pr));
+    }
+  }
+  return seen.size;
+}
diff --git a/packages/app/src/lib/agentx-telemetry-zh.ts b/packages/app/src/lib/agentx-telemetry-zh.ts
new file mode 100644
index 000000000..e3aa727d8
--- /dev/null
+++ b/packages/app/src/lib/agentx-telemetry-zh.ts
@@ -0,0 +1,168 @@
+/**
+ * Simplified Chinese port of `agentx-telemetry.ts`.
+ *
+ * Translations are keyed by section id and carry only the prose: figure
+ * assets, section ordering, and link targets come from the English guide, so
+ * the two languages cannot drift structurally. `agentx-telemetry.test.ts`
+ * fails if a section, paragraph, bullet, or highlight here does not line up
+ * with its English counterpart.
+ */
+
+import {
+  AGENTX_TELEMETRY_GUIDE,
+  type TelemetryFigureCopy,
+  type TelemetryGuide,
+  type TelemetrySection,
+} from './agentx-telemetry';
+import type { Locale } from './i18n';
+
+interface SectionTranslation {
+  heading: string;
+  paragraphs: readonly string[];
+  bullets?: readonly string[];
+  figure?: TelemetryFigureCopy;
+  /** Link labels, in the same order as the English section's `links`. */
+  links?: readonly string[];
+}
+
+const GUIDE_ZH = {
+  eyebrow: 'AgentX 教程',
+  title: '通过详细遥测数据解析智能体负载',
+  lead: 'AgentX 上的一个数据点，背后是数千个请求，跨越不断增长的对话、subagent、warmup 阶段、cache 状态，以及动态变化的在途负载。本文介绍如何把这个点展开来看。',
+  intro: [
+    'AgentX 需要的不只是新的基准测试框架和数据集。我们还花了一些时间重构 InferenceX 的可视化，让智能体测试结果更容易探索和理解。Pareto 曲线上的单个点会掩盖大量有价值的信息，因此 AgentX 图表上的每个点现在都是进入其背后完整运行数据的入口。',
+    '本教程会逐一介绍这个入口通向的内容：曲线本身如何构建、数据点 tooltip 暴露了哪些信息、详情页上的 11 张单点遥测图表、请求时间线，以及 AgentX 数据集页面上的单会话火焰图（flamegraph）。',
+  ],
+  highlights: [
+    '张单点遥测图表',
+    '种单点视图',
+    '个回放阶段',
+    '条曲线（每个 model、SKU 与引擎组合）',
+  ],
+  sections: {
+    'why-per-point-telemetry': {
+      heading: '为什么一个点不够用',
+      paragraphs: [
+        '固定序列长度的基准测试点，概括的是一个同质负载：每个请求的 input 和 output 长度都相同，因此用一个聚合吞吐量数字来描述整次运行是公平的。AgentX 的点则不然——它聚合了数千个请求，这些请求的 input 长度随对话延长而增长，subagent 会突发式到达，cache 状态也在整个回放过程中不断变化。',
+        '因此，两个聚合吞吐量几乎相同的点，底层行为可能截然不同：一个能稳定维持 prefix cache 命中率，另一个则在反复淘汰与重算。单点遥测数据正是让这种差异变得可见的手段。',
+      ],
+      figure: {
+        alt: 'Per-point 详情视图：上方是 input 与 output 序列长度分布，下方是 interactivity 随时间变化和 TTFT 随时间变化图表，顶部有 Per-point、Request timeline、Aggregates across configs 三个标签页以及 Profiling / Warmup 阶段切换。',
+        caption:
+          'Per-point 视图。三个标签页用于切换视图，Stage 开关在 warmup 与 profiling 阶段之间切换，每张图表都可展开到整行宽度。',
+      },
+    },
+    'one-curve-per-stack': {
+      heading: '每个 model、SKU 与推理引擎组合只画一条曲线',
+      paragraphs: [
+        '本次的一项重要改动，是曲线本身的构建方式。在此前版本的 InferenceX 中，开启和关闭 speculative decoding 的配置往往被画成两条独立曲线。我们正在放弃这种做法：前端现在会合并允许使用的推理优化，为每个 model、SKU 与推理引擎组合展示当前可达的最佳曲线。',
+        '因此，同一条曲线上的各个点可能使用了不同的优化技术与配置，包括 speculative decoding、disaggregation 或 KV cache offload。我们的目标是呈现每套软硬件栈在生产环境下能达到的最佳性能，而不是为每种可能的优化组合都单独画一条曲线。',
+        '我们仍然会暴露每个点的底层配置与来源信息。点击一个数据点会弹出 tooltip，其中详细列出产生该点的具体配置、运行元数据、指向公开可查的 CI 来源链接，以及 AgentX 特有的统计数据。从那里点击 “View charts” 链接，即可打开完整的单点详情页。',
+      ],
+      figure: {
+        alt: 'AgentX Pareto 曲线，横轴为 P90 interactivity、纵轴为单芯片 token 吞吐量，图上固定显示一个 tooltip，列出所选点的镜像、interactivity、吞吐量、芯片数量、并行方式、精度、cache 命中率、speculative decoding 与 token 数量，并附带 GitHub Actions 运行链接和 View charts 按钮。',
+        caption:
+          '数据点 tooltip。曲线上的每项结论都可追溯：容器镜像、并行方式、精度、cache 命中率、所用 speculator，以及产生这些数字的 CI 运行链接。',
+      },
+    },
+    'point-detail-page': {
+      heading: '单点详情页',
+      paragraphs: [
+        '单点详情视图可以更深入地观察所选的 AgentX 运行。这些指标让人更容易理解：为什么两个聚合吞吐量相近的点，在整个回放过程中的表现却可能截然不同。',
+        '该页面还将 warmup 与 profiling 数据分开。读者可以在两个阶段之间切换，分别查看系统在建立 cache 状态期间，以及在基准测试所用的 profiling 阶段中的行为。对外报告的结果只覆盖 profiling 窗口，因此 warmup 阶段正是观察 cache 填充行为及其代价的地方。',
+      ],
+      bullets: [
+        'Input 与 output 序列长度分布，可切换为直方图或在途平均值。',
+        'Interactivity 随时间变化，可选 P75 或 P90，并与其累计值对比。',
+        '首 token 延迟（TTFT）随时间变化，可在 TTFT 与端到端延迟之间切换。',
+        'KV cache 利用率随时间变化；当有多个引擎上报时会分别拆开显示。',
+        '请求队列深度，以及已完成请求数。',
+        '每个区间的 prefix cache 命中率。',
+        'Input 与 decode 吞吐量。',
+        '累计 prompt token 来源构成——prompt 中有多少来自 cache、多少来自重算。',
+        '累计唯一 input token 数随时间变化，以及在途唯一 input token 数与 KV cache pool 容量的对比。',
+      ],
+      figure: {
+        alt: 'B200 上 MiniMax-M3 FP4 vLLM 数据点的完整详情页，纵向排列 11 张遥测图表，从序列长度分布，到 KV cache 利用率、队列深度、prefix cache 命中率、吞吐量、prompt token 来源构成和唯一 input token 数。',
+        caption:
+          '单个数据点的完整详情页。页头显示 SKU、精度与推理引擎，同 SKU 配置导航器可在相邻配置间切换，所有图表都遵循当前选中的阶段。',
+      },
+    },
+    'kv-offload': {
+      heading: '如何解读启用 KV cache offload 的点',
+      paragraphs: [
+        '使用了 KV cache offload 的点，在主图上会额外套一圈虚线圆环，用于区分开启了 KV offload 的数据点。选中这类点后，详情页会显示 offload 类型、KV offload 引擎、芯片 cache 命中率和 CPU cache 命中率。',
+        '这样一来，无需为每种 offload 配置单独画一条曲线，也能看清 KV offload 在最佳曲线中的贡献——这与“把优化合并到同一条曲线”是同一个思路，只是应用在了效果最容易被忽略的那项优化上。',
+      ],
+    },
+    'request-timeline': {
+      heading: '请求时间线',
+      paragraphs: [
+        '另一项新功能是请求时间线。该视图展示所选 AgentX 运行中回放的每个请求，可以按 conversation 或按 worker 组织。Conversation 视图会把 subagent 归到对应的根会话之下，便于观察会话与 subagent 何时并行发生。Warmup 与 profiling 请求同样可以分开查看。',
+        '时间线中的每个请求都可点击，会直接跳转到 InferenceX 数据集页面上对应的会话与轮次。读者因此可以从 Pareto 曲线上的一个聚合点，一路追到实际被回放的那条匿名化请求。',
+      ],
+      figure: {
+        alt: '请求时间线视图：每个会话一行，subagent 行缩进显示在其根会话之下，每行用彩色色块表示运行期间回放的各个请求。',
+        caption:
+          '按 conversation 组织的请求时间线。Subagent 行位于其根会话之下，因此并行的 subagent 活动会在纵向上对齐呈现。',
+      },
+    },
+    flamegraph: {
+      heading: '单会话火焰图',
+      paragraphs: [
+        'AgentX 页面还提供了用于展示单个会话结构的火焰图。每个柱条代表一个轮次，并按该会话中最长轮次进行归一化缩放。柱条内部划分为 cached prefix token、uncached input token 和生成的 output token 三段。',
+        '这直观呈现了上下文如何在一段会话中逐步增长，以及每个请求中有多大比例可以从 KV cache 复用。Subagent 分组默认折叠，点击即可展开；左侧的彩色括号会把同一主 agent 或 subagent 作用域内、原始执行区间存在重叠的请求归为一组，让并行的工作在视觉上也呈现为并行。',
+      ],
+      figure: {
+        alt: '会话火焰图：每个轮次一个柱条，柱条按 cached prefix、uncached input 和 output 分段着色，折叠的 subagent 分组穿插在轮次之间，每行旁边标注了相对起始时间。',
+        caption:
+          '一段会话的火焰图。第 1 轮几乎全是 uncached input；到第 8 轮，柱条已主要由 cached prefix 构成——这正是 AgentX 要测量的 KV 复用模式。',
+      },
+    },
+  } as Readonly<Record<string, SectionTranslation>>,
+  ui: {
+    backToAgentX: '← AgentX 数据集',
+    onThisPage: '本页内容',
+    figureCta: '查看原图',
+    readMore: '阅读遥测数据教程',
+    openResults: '探索一个 AgentX 数据点',
+  },
+} as const;
+
+function localizeSection(section: TelemetrySection, zh: SectionTranslation): TelemetrySection {
+  return {
+    ...section,
+    heading: zh.heading,
+    paragraphs: zh.paragraphs,
+    bullets: zh.bullets ?? section.bullets,
+    figure: section.figure && zh.figure ? { ...section.figure, ...zh.figure } : section.figure,
+    links: section.links?.map((link, index) => ({
+      ...link,
+      label: zh.links?.[index] ?? link.label,
+    })),
+  };
+}
+
+/** The tutorial for a locale, with English structure and localized prose. */
+export function getTelemetryGuide(locale: Locale): TelemetryGuide {
+  if (locale === 'en') return AGENTX_TELEMETRY_GUIDE;
+  return {
+    ...AGENTX_TELEMETRY_GUIDE,
+    eyebrow: GUIDE_ZH.eyebrow,
+    title: GUIDE_ZH.title,
+    lead: GUIDE_ZH.lead,
+    intro: GUIDE_ZH.intro,
+    highlights: AGENTX_TELEMETRY_GUIDE.highlights.map((highlight, index) => ({
+      ...highlight,
+      label: GUIDE_ZH.highlights[index] ?? highlight.label,
+    })),
+    sections: AGENTX_TELEMETRY_GUIDE.sections.map((section) => {
+      const zh = GUIDE_ZH.sections[section.id];
+      return zh ? localizeSection(section, zh) : section;
+    }),
+    ui: GUIDE_ZH.ui,
+  };
+}
+
+/** Exported for the parity test, which walks the same keys the resolver reads. */
+export const TELEMETRY_ZH_INTERNALS = { GUIDE_ZH };
diff --git a/packages/app/src/lib/agentx-telemetry.test.ts b/packages/app/src/lib/agentx-telemetry.test.ts
new file mode 100644
index 000000000..844f9c11e
--- /dev/null
+++ b/packages/app/src/lib/agentx-telemetry.test.ts
@@ -0,0 +1,162 @@
+import { existsSync } from 'node:fs';
+import path from 'node:path';
+
+import { describe, expect, it } from 'vitest';
+
+import {
+  AGENTIC_DETAIL_SURFACE,
+  AGENTX_TELEMETRY_GUIDE,
+  AGENTX_TELEMETRY_SECTION_IDS,
+  TELEMETRY_FIGURES,
+} from './agentx-telemetry';
+import { getTelemetryGuide, TELEMETRY_ZH_INTERNALS } from './agentx-telemetry-zh';
+
+const CJK = /[一-鿿]/u;
+
+describe('AgentX telemetry guide', () => {
+  it('has unique, URL-safe section ids', () => {
+    expect(new Set(AGENTX_TELEMETRY_SECTION_IDS).size).toBe(AGENTX_TELEMETRY_SECTION_IDS.length);
+    for (const id of AGENTX_TELEMETRY_SECTION_IDS) expect(id).toMatch(/^[a-z0-9-]+$/u);
+  });
+
+  it('gives every section a heading and at least one paragraph', () => {
+    for (const section of AGENTX_TELEMETRY_GUIDE.sections) {
+      expect(section.heading.length, section.id).toBeGreaterThan(0);
+      expect(section.paragraphs.length, section.id).toBeGreaterThan(0);
+      for (const paragraph of section.paragraphs) {
+        expect(paragraph.length, section.id).toBeGreaterThan(40);
+      }
+    }
+  });
+
+  it('ships every figure it references, with usable next/image dimensions', () => {
+    const publicDir = path.resolve(import.meta.dirname, '../../public');
+    for (const section of AGENTX_TELEMETRY_GUIDE.sections) {
+      if (!section.figure) continue;
+      const asset = TELEMETRY_FIGURES[section.figure.key];
+      expect(asset, section.figure.key).toBeDefined();
+      expect(existsSync(path.join(publicDir, asset.src)), asset.src).toBe(true);
+      expect(asset.width).toBeGreaterThan(0);
+      expect(asset.height).toBeGreaterThan(0);
+      // Screen-reader copy has to describe the screenshot, not just name it.
+      expect(section.figure.alt.length, section.figure.key).toBeGreaterThan(40);
+      expect(section.figure.caption.length, section.figure.key).toBeGreaterThan(20);
+    }
+  });
+
+  it('uses every declared figure exactly once', () => {
+    const used = AGENTX_TELEMETRY_GUIDE.sections
+      .map((section) => section.figure?.key)
+      .filter((key): key is NonNullable<typeof key> => Boolean(key));
+    expect(new Set(used).size).toBe(used.length);
+    expect(new Set(used)).toEqual(new Set(Object.keys(TELEMETRY_FIGURES)));
+  });
+
+  it('links only to https references', () => {
+    for (const section of AGENTX_TELEMETRY_GUIDE.sections) {
+      for (const link of section.links ?? []) {
+        expect(link.href.startsWith('https://'), link.href).toBe(true);
+        expect(link.label.length).toBeGreaterThan(0);
+      }
+    }
+  });
+
+  // The prose cites these counts. AGENTIC_DETAIL_SURFACE mirrors what
+  // agentic-point-detail.tsx renders, so adding a chart there without
+  // updating the copy fails here rather than shipping a stale number.
+  it('keeps its numeric claims aligned with the detail-page surface', () => {
+    const byLabel = new Map(
+      AGENTX_TELEMETRY_GUIDE.highlights.map((highlight) => [highlight.label, highlight.value]),
+    );
+    expect(byLabel.get('per-point telemetry charts')).toBe(
+      String(AGENTIC_DETAIL_SURFACE.perPointCharts.length),
+    );
+    expect(byLabel.get('views per point')).toBe(String(AGENTIC_DETAIL_SURFACE.views));
+    expect(byLabel.get('replay stages')).toBe(String(AGENTIC_DETAIL_SURFACE.stages.length));
+
+    // The bullet list enumerates the same eleven charts, collapsed into nine
+    // lines because three pairs share a card. Guard the two that are easy to
+    // desynchronise instead of asserting a brittle line count.
+    const detail = AGENTX_TELEMETRY_GUIDE.sections.find(
+      (section) => section.id === 'point-detail-page',
+    );
+    expect(detail?.bullets?.length ?? 0).toBeGreaterThan(0);
+    expect(AGENTX_TELEMETRY_GUIDE.intro.join(' ')).toContain('eleven');
+    expect(AGENTIC_DETAIL_SURFACE.perPointCharts.length).toBe(11);
+  });
+
+  it('has no duplicate entries in the detail-page surface record', () => {
+    const charts = AGENTIC_DETAIL_SURFACE.perPointCharts;
+    expect(new Set(charts).size).toBe(charts.length);
+  });
+
+  it('carries no unresolved editorial notes from the source document', () => {
+    const prose = [
+      AGENTX_TELEMETRY_GUIDE.lead,
+      ...AGENTX_TELEMETRY_GUIDE.intro,
+      ...AGENTX_TELEMETRY_GUIDE.sections.flatMap((section) => [
+        section.heading,
+        ...section.paragraphs,
+        ...(section.bullets ?? []),
+      ]),
+    ].join('\n');
+    for (const marker of ['TO BE UPDATED', 'TODO', 'Source: SemiAnalysis']) {
+      expect(prose, marker).not.toContain(marker);
+    }
+  });
+});
+
+describe('AgentX telemetry Chinese port', () => {
+  it('returns the English guide untouched for the en locale', () => {
+    expect(getTelemetryGuide('en')).toBe(AGENTX_TELEMETRY_GUIDE);
+  });
+
+  it('translates the guide without dropping structure', () => {
+    const en = AGENTX_TELEMETRY_GUIDE;
+    const zh = getTelemetryGuide('zh');
+
+    expect(zh.title).not.toBe(en.title);
+    expect(CJK.test(zh.title)).toBe(true);
+    expect(CJK.test(zh.eyebrow)).toBe(true);
+    expect(CJK.test(zh.lead)).toBe(true);
+    expect(zh.intro).toHaveLength(en.intro.length);
+    for (const paragraph of zh.intro) expect(CJK.test(paragraph)).toBe(true);
+
+    expect(zh.highlights).toHaveLength(en.highlights.length);
+    for (const [index, highlight] of zh.highlights.entries()) {
+      // Values are numerals — they stay language-neutral; only labels translate.
+      expect(highlight.value).toBe(en.highlights[index].value);
+      expect(CJK.test(highlight.label), `highlight ${index}`).toBe(true);
+    }
+
+    expect(zh.sections).toHaveLength(en.sections.length);
+    for (const [index, section] of zh.sections.entries()) {
+      const source = en.sections[index];
+      expect(section.id).toBe(source.id);
+      expect(CJK.test(section.heading), section.id).toBe(true);
+      expect(section.paragraphs, section.id).toHaveLength(source.paragraphs.length);
+      for (const paragraph of section.paragraphs)
+        expect(CJK.test(paragraph), section.id).toBe(true);
+
+      expect(section.bullets?.length ?? 0, section.id).toBe(source.bullets?.length ?? 0);
+      for (const bullet of section.bullets ?? []) expect(CJK.test(bullet), section.id).toBe(true);
+
+      expect(section.figure?.key).toBe(source.figure?.key);
+      if (source.figure) {
+        expect(CJK.test(section.figure?.alt ?? ''), section.id).toBe(true);
+        expect(CJK.test(section.figure?.caption ?? ''), section.id).toBe(true);
+      }
+
+      expect(section.links?.map((link) => link.href)).toEqual(
+        source.links?.map((link) => link.href),
+      );
+    }
+
+    for (const value of Object.values(zh.ui)) expect(CJK.test(value)).toBe(true);
+  });
+
+  it('translates every section the English guide declares, and no others', () => {
+    const translated = Object.keys(TELEMETRY_ZH_INTERNALS.GUIDE_ZH.sections);
+    expect(new Set(translated)).toEqual(new Set(AGENTX_TELEMETRY_SECTION_IDS));
+  });
+});
diff --git a/packages/app/src/lib/agentx-telemetry.ts b/packages/app/src/lib/agentx-telemetry.ts
new file mode 100644
index 000000000..f95d271c1
--- /dev/null
+++ b/packages/app/src/lib/agentx-telemetry.ts
@@ -0,0 +1,241 @@
+/**
+ * AgentX detailed-telemetry tutorial — how to read the per-point views that
+ * sit behind every dot on an AgentX Pareto curve.
+ *
+ * English content lives here; the Simplified Chinese port is a 1:1 mirror in
+ * `agentx-telemetry-zh.ts`, keyed by the same section ids.
+ * `agentx-telemetry.test.ts` enforces that the two stay structurally
+ * identical, so a section added here without a translation fails CI.
+ *
+ * The counts in `highlights` describe the live UI, not the screenshots: the
+ * test asserts them against `AGENTIC_DETAIL_SURFACE`, which mirrors what
+ * `components/inference/agentic-point/agentic-point-detail.tsx` renders. If a
+ * chart or view is added there, update both and the copy that cites them.
+ */
+
+/**
+ * What the per-point detail page actually ships, as a machine-checkable
+ * record. Keeps the tutorial's numeric claims from drifting once the page
+ * grows another chart.
+ */
+export const AGENTIC_DETAIL_SURFACE = {
+  /** SegmentedToggle options: per-point, request timeline, aggregates. */
+  views: 3,
+  /** Cards in the per-point grid, in render order. */
+  perPointCharts: [
+    'input-sequence-length',
+    'output-sequence-length',
+    'interactivity-over-time',
+    'ttft-over-time',
+    'kv-cache-utilization',
+    'request-queue-depth',
+    'prefix-cache-hit-rate',
+    'throughput',
+    'prompt-token-source',
+    'cumulative-unique-input-tokens',
+    'inflight-unique-input-tokens',
+  ],
+  /** Replay stages the Stage toggle switches between. */
+  stages: ['profiling', 'warmup'],
+  /** Request-timeline groupings. */
+  timelineGroupings: ['by conversation', 'by worker'],
+} as const;
+
+/** Figures exported from the source document, served from /public. */
+export const TELEMETRY_FIGURES = {
+  perPointOverview: {
+    src: '/images/agentx-telemetry/per-point-overview.png',
+    width: 1110,
+    height: 687,
+  },
+  chartPointTooltip: {
+    src: '/images/agentx-telemetry/chart-point-tooltip.png',
+    width: 2048,
+    height: 1256,
+  },
+  pointDetailFull: {
+    src: '/images/agentx-telemetry/point-detail-full.png',
+    width: 1001,
+    height: 2048,
+  },
+  requestTimeline: {
+    src: '/images/agentx-telemetry/request-timeline.png',
+    width: 2048,
+    height: 1097,
+  },
+  flamegraph: {
+    src: '/images/agentx-telemetry/flamegraph.png',
+    width: 2048,
+    height: 879,
+  },
+} as const;
+
+export type TelemetryFigureKey = keyof typeof TELEMETRY_FIGURES;
+
+/** Prose that a translation replaces; the asset itself is language-neutral. */
+export interface TelemetryFigureCopy {
+  alt: string;
+  caption: string;
+}
+
+export interface TelemetryFigure extends TelemetryFigureCopy {
+  key: TelemetryFigureKey;
+}
+
+export interface TelemetryLink {
+  href: string;
+  label: string;
+}
+
+export interface TelemetrySection {
+  /** Anchor id — stable across languages, so deep links survive a locale switch. */
+  id: string;
+  heading: string;
+  paragraphs: readonly string[];
+  figure?: TelemetryFigure;
+  /** Bulleted detail, used where the source document enumerates rather than argues. */
+  bullets?: readonly string[];
+  links?: readonly TelemetryLink[];
+}
+
+export interface TelemetryHighlight {
+  value: string;
+  label: string;
+}
+
+export interface TelemetryGuide {
+  eyebrow: string;
+  title: string;
+  lead: string;
+  intro: readonly string[];
+  highlights: readonly TelemetryHighlight[];
+  sections: readonly TelemetrySection[];
+  /** Chrome around the article: back links, figure CTA, section nav. */
+  ui: {
+    backToAgentX: string;
+    onThisPage: string;
+    figureCta: string;
+    readMore: string;
+    openResults: string;
+  };
+}
+
+export const AGENTX_TELEMETRY_GUIDE: TelemetryGuide = {
+  eyebrow: 'AgentX tutorial',
+  title: 'Exploring Agentic Workloads: Detailed Telemetry',
+  lead: 'A single AgentX datapoint represents thousands of requests across growing conversations, subagents, warmup periods, cache states, and dynamically changing in-flight load. This is how to open one up.',
+  intro: [
+    'AgentX required more than a new benchmark harness and dataset. We also spent some time rebuilding parts of the InferenceX visualization to make agentic results easier to explore and digest. A single point on a Pareto curve can hide a lot of useful information, so every point on an AgentX chart is now a doorway into the run behind it.',
+    'This tutorial walks through what that doorway leads to: how the curves themselves are constructed, what the point tooltip exposes, the eleven per-point telemetry charts on the detail page, the request timeline, and the per-conversation flamegraph on the AgentX dataset pages.',
+  ],
+  highlights: [
+    { value: '11', label: 'per-point telemetry charts' },
+    { value: '3', label: 'views per point' },
+    { value: '2', label: 'replay stages' },
+    { value: '1', label: 'curve per model, SKU, and engine' },
+  ],
+  sections: [
+    {
+      id: 'why-per-point-telemetry',
+      heading: 'Why a single point is not enough',
+      paragraphs: [
+        'A fixed-sequence benchmark point is a summary of a homogeneous workload: every request has the same input and output length, so an aggregate throughput number describes the run fairly. An AgentX point is not that. It aggregates thousands of requests whose input lengths grow as conversations extend, whose subagents arrive in bursts, and whose cache state changes throughout the replay.',
+        'Two points with nearly identical aggregate throughput can therefore behave very differently under the hood — one sustaining a steady prefix-cache hit rate, the other repeatedly evicting and recomputing. Per-point telemetry is what makes that difference visible.',
+      ],
+      figure: {
+        key: 'perPointOverview',
+        alt: 'The per-point detail view showing input and output sequence length distributions above interactivity-over-time and TTFT-over-time charts, with Per-point, Request timeline, and Aggregates across configs tabs and a Profiling / Warmup stage toggle.',
+        caption:
+          'The per-point view. Three tabs select the view, a Stage toggle switches between the warmup and profiling phases, and each chart expands to full width.',
+      },
+    },
+    {
+      id: 'one-curve-per-stack',
+      heading: 'One curve per model, SKU, and inference engine',
+      paragraphs: [
+        'One of our major changes is how we construct the curves themselves. In previous versions of InferenceX, configurations with speculative decoding enabled and disabled were often displayed as separate curves. We are now moving away from this approach. The frontend combines allowed inference optimizations and displays the best available curve for each model, SKU, and inference engine combination.',
+        'Because of this, individual points along a single curve may use different optimization techniques and configurations, including speculative decoding, disaggregation, or KV cache offload. Our goal is to show the best production performance available from each hardware and software stack, rather than creating a separate curve for every possible combination of optimizations.',
+        'We still expose the underlying configuration and provenance for every point. Clicking a point shows a tooltip with a detailed view showing exactly which configuration produced it, along with the run metadata, links to the publicly viewable CI provenance, and AgentX specific statistics. From there, the "View charts" link opens the full point-detail page.',
+      ],
+      figure: {
+        key: 'chartPointTooltip',
+        alt: 'An AgentX Pareto curve of token throughput per chip against P90 interactivity, with a pinned tooltip listing the image, interactivity, throughput, chip count, parallelism, precision, cache hit rates, speculative decoding, and token counts for the selected point, plus a GitHub Actions run link and a View charts button.',
+        caption:
+          'The point tooltip. Every claim on the curve is traceable: container image, parallelism, precision, cache-hit rates, the speculator in use, and a link to the CI run that produced the numbers.',
+      },
+    },
+    {
+      id: 'point-detail-page',
+      heading: 'The point-detail page',
+      paragraphs: [
+        'The detailed point view provides a much deeper look into the selected AgentX run. These metrics make it easier to understand why two points with similar aggregate throughput may behave differently throughout the replay.',
+        'The page also separates warmup and profiling data. Readers can switch between the two phases to inspect how the system behaves while its cache state is being established and during the profiling period used for the benchmark run. Reported results cover the profiling window only, so the warmup stage is where cache-fill behavior — and the cost of establishing it — becomes visible.',
+      ],
+      bullets: [
+        'Input and output sequence length distributions, as a histogram or as an in-flight average.',
+        'Interactivity over time, at P75 or P90, against its cumulative value.',
+        'Time to first token (TTFT) over time, switchable between TTFT and end-to-end latency.',
+        'KV cache utilization over time, broken out by engine where more than one reports it.',
+        'Request queue depth, alongside the count of requests completed.',
+        'Prefix cache hit rate per interval.',
+        'Input and decode throughput.',
+        'Cumulative prompt-token source breakdown — how much of the prompt came from cache versus recomputation.',
+        'Total unique input tokens over time, and unique input tokens in flight against the KV cache pool size.',
+      ],
+      figure: {
+        key: 'pointDetailFull',
+        alt: 'The full point-detail page for a B200 MiniMax-M3 FP4 vLLM point, showing eleven stacked telemetry charts from sequence-length distributions through KV cache utilization, queue depth, prefix cache hit rate, throughput, prompt-token source breakdown, and unique input tokens.',
+        caption:
+          'The full detail page for one point. The header carries the SKU, precision, and engine, a sibling navigator moves between configurations in the same SKU, and every chart honors the selected stage.',
+      },
+    },
+    {
+      id: 'kv-offload',
+      heading: 'Reading KV cache offload points',
+      paragraphs: [
+        'Points using KV cache offload are surrounded by an additional dotted circle on the main chart, which is used to distinguish points with KV offload enabled. When one of these points is selected, the detail page shows the offload type, KV offload engine, chip cache-hit rate, and CPU cache-hit rate.',
+        'This makes it possible to see where KV offload contributes to the best curve without creating a separate curve for every offload configuration — the same principle behind combining optimizations into a single curve, applied to the one optimization whose effect is easiest to miss.',
+      ],
+    },
+    {
+      id: 'request-timeline',
+      heading: 'The request timeline',
+      paragraphs: [
+        'Another new feature is the request timeline. This view shows the individual requests replayed during a selected AgentX run and can be organized either by conversation or by worker. The conversation view groups subagents underneath their corresponding root conversation, making it easy to see when conversations and subagents overlap. Warmup and profiling requests can also still be viewed separately.',
+        'Each request in the timeline is clickable and links directly to the corresponding conversation and turns on the InferenceX datasets page. This allows readers to move from an aggregate point on the Pareto curve to the exact anonymized request that was replayed.',
+      ],
+      figure: {
+        key: 'requestTimeline',
+        alt: 'The request timeline view, with one row per conversation and indented subagent rows beneath their root conversation, each row showing colored bars for the individual requests replayed over the run duration.',
+        caption:
+          'The request timeline, grouped by conversation. Subagent rows sit under their root conversation, so overlapping subagent activity is visible as vertical alignment.',
+      },
+    },
+    {
+      id: 'flamegraph',
+      heading: 'The per-conversation flamegraph',
+      paragraphs: [
+        'The AgentX page also includes a flamegraph for visualizing the structure of an individual conversation. Each bar represents one turn and is scaled relative to the largest turn in that conversation. The bar is divided into cached prefix tokens, uncached input tokens, and generated output tokens.',
+        'This gives a visual representation of how the context grows throughout a conversation and how much of each request can be reused from KV cache. Subagent groups are collapsed by default and expand on click; a colored bracket on the left groups requests in the same main-agent or subagent scope whose original execution intervals overlapped, so parallel work reads as parallel.',
+      ],
+      figure: {
+        key: 'flamegraph',
+        alt: 'A conversation flamegraph with one bar per turn, each split into cached prefix, uncached input, and output segments, with collapsed subagent groups interleaved between turns and elapsed timestamps beside each row.',
+        caption:
+          'A conversation flamegraph. Turn 1 is almost entirely uncached input; by turn 8 the bar is mostly cached prefix — the KV-reuse pattern AgentX is built to measure.',
+      },
+    },
+  ],
+  ui: {
+    backToAgentX: '← AgentX datasets',
+    onThisPage: 'On this page',
+    figureCta: 'Open full size',
+    readMore: 'Read the telemetry tutorial',
+    openResults: 'Explore an AgentX point',
+  },
+};
+
+/** Section ids, for the on-this-page nav and the parity test. */
+export const AGENTX_TELEMETRY_SECTION_IDS = AGENTX_TELEMETRY_GUIDE.sections.map(
+  (section) => section.id,
+);
diff --git a/packages/app/src/lib/api-cache.test.ts b/packages/app/src/lib/api-cache.test.ts
index 93d6a1e73..4b28f5ec2 100644
--- a/packages/app/src/lib/api-cache.test.ts
+++ b/packages/app/src/lib/api-cache.test.ts
@@ -98,7 +98,7 @@ describe('cachedQuery', () => {
 
       const result = await wrapped();
       expect(result).toEqual({ answer: 42 });
-      expect(mockBlobGet).toHaveBeenCalledWith('blob-key');
+      expect(mockBlobGet).toHaveBeenCalledWith(expect.stringMatching(/^blob-key:v2:/u));
       expect(fn).not.toHaveBeenCalled();
       expect(mockBlobSet).not.toHaveBeenCalled();
     });
@@ -112,7 +112,9 @@ describe('cachedQuery', () => {
       const result = await wrapped();
       expect(result).toEqual({ big: 'payload' });
       expect(fn).toHaveBeenCalled();
-      expect(mockBlobSet).toHaveBeenCalledWith('miss-key', { big: 'payload' });
+      expect(mockBlobSet).toHaveBeenCalledWith(expect.stringMatching(/^miss-key:v2:/u), {
+        big: 'payload',
+      });
     });
 
     it('builds blob key from prefix and args', async () => {
@@ -122,8 +124,9 @@ describe('cachedQuery', () => {
       const wrapped = cachedQuery(fn, 'bench', { blobOnly: true });
 
       await wrapped('llama', '2025-01-01');
-      expect(mockBlobGet).toHaveBeenCalledWith('bench:llama:2025-01-01');
-      expect(mockBlobSet).toHaveBeenCalledWith('bench:llama:2025-01-01', ['llama', '2025-01-01']);
+      const key = mockBlobGet.mock.calls[0]![0];
+      expect(key).toMatch(/^bench:v2:[A-Za-z0-9_-]+$/u);
+      expect(mockBlobSet).toHaveBeenCalledWith(key, ['llama', '2025-01-01']);
     });
 
     it('hashes oversized argument lists into a bounded deterministic key', async () => {
@@ -139,11 +142,50 @@ describe('cachedQuery', () => {
       const firstKey = mockBlobGet.mock.calls[0]![0];
       const secondKey = mockBlobGet.mock.calls[1]![0];
       expect(firstKey).toBe(secondKey);
-      expect(firstKey).toMatch(/^derived-agentic-metrics-v7:sha256:[a-f0-9]{64}$/u);
+      expect(firstKey).toMatch(/^derived-agentic-metrics-v7:v2:sha256:[a-f0-9]{64}$/u);
       expect(firstKey.length).toBeLessThan(128);
       expect(mockBlobSet).toHaveBeenCalledWith(firstKey, 200);
     });
 
+    it('cannot collide arrays, objects, null, and undefined that stringify similarly', async () => {
+      mockBlobGet.mockResolvedValue(null);
+      mockBlobSet.mockResolvedValue(undefined);
+      const wrapped = cachedQuery((value: unknown) => Promise.resolve(value), 'structured', {
+        blobOnly: true,
+      });
+
+      await wrapped(['a:b']);
+      await wrapped(['a', 'b']);
+      await wrapped({ value: null });
+      await wrapped({ value: undefined });
+
+      expect(new Set(mockBlobGet.mock.calls.map(([key]) => key)).size).toBe(4);
+    });
+
+    it('serializes object properties deterministically', async () => {
+      mockBlobGet.mockResolvedValue(null);
+      mockBlobSet.mockResolvedValue(undefined);
+      const wrapped = cachedQuery((value: unknown) => Promise.resolve(value), 'structured', {
+        blobOnly: true,
+      });
+
+      await wrapped({ b: 2, a: 1 });
+      await wrapped({ a: 1, b: 2 });
+
+      expect(mockBlobGet.mock.calls[0]![0]).toBe(mockBlobGet.mock.calls[1]![0]);
+    });
+
+    it('rejects non-plain structured arguments instead of assigning a colliding key', async () => {
+      const wrapped = cachedQuery((value: unknown) => Promise.resolve(value), 'structured', {
+        blobOnly: true,
+      });
+
+      await expect(wrapped(new Date('2026-08-20T00:00:00Z'))).rejects.toThrow(
+        'Blob cache arguments must contain only arrays and plain objects',
+      );
+      expect(mockBlobGet).not.toHaveBeenCalled();
+    });
+
     it('serves a successful query result when the blob write fails', async () => {
       mockBlobGet.mockResolvedValue(null);
       mockBlobSet.mockRejectedValue(new Error('pathname is too long'));
@@ -159,13 +201,13 @@ describe('cachedQuery', () => {
       warn.mockRestore();
     });
 
-    it('uses bare prefix when no args are passed', async () => {
+    it('versions keys even when no arguments are passed', async () => {
       mockBlobGet.mockResolvedValue('cached');
       const fn = vi.fn(() => Promise.resolve('fresh'));
       const wrapped = cachedQuery(fn, 'no-args', { blobOnly: true });
 
       await wrapped();
-      expect(mockBlobGet).toHaveBeenCalledWith('no-args');
+      expect(mockBlobGet).toHaveBeenCalledWith(expect.stringMatching(/^no-args:v2:/u));
     });
 
     it('does not use unstable_cache in blobOnly mode', () => {
diff --git a/packages/app/src/lib/api-cache.ts b/packages/app/src/lib/api-cache.ts
index 34ab75ea7..67e50894b 100644
--- a/packages/app/src/lib/api-cache.ts
+++ b/packages/app/src/lib/api-cache.ts
@@ -54,19 +54,75 @@ interface CachedQueryOptions {
   tag?: string;
 }
 
-// Leave room for BLOB_CACHE_PREFIX and the `.json` suffix in the final Vercel
-// Blob pathname. Short keys keep their existing human-readable form (and warm
-// cache entries); large argument lists use a fixed-size digest instead.
+// Blob argument encoding is explicitly versioned. v1 used `args.join(':')`,
+// which was not injective for arrays, objects, null, and undefined. v2 keys
+// never read a v1 entry; the next ordinary prefix purge removes those orphaned
+// values along with every other cache generation.
+const BLOB_KEY_ENCODING_VERSION = 'v2';
 const MAX_INLINE_BLOB_KEY_BYTES = 512;
 
-function blobKeyForArgs(keyPrefix: string, args: unknown[]): string {
-  if (args.length === 0) return keyPrefix;
+function canonicalCacheValue(value: unknown, ancestors: Set<object>): unknown {
+  if (value === null) return ['null'];
+  if (value === undefined) return ['undefined'];
+
+  switch (typeof value) {
+    case 'string':
+    case 'boolean': {
+      return [typeof value, value];
+    }
+    case 'number': {
+      if (Number.isNaN(value)) return ['number', 'NaN'];
+      if (value === Number.POSITIVE_INFINITY) return ['number', 'Infinity'];
+      if (value === Number.NEGATIVE_INFINITY) return ['number', '-Infinity'];
+      if (Object.is(value, -0)) return ['number', '-0'];
+      return ['number', String(value)];
+    }
+    case 'bigint': {
+      return ['bigint', value.toString()];
+    }
+    case 'object': {
+      if (ancestors.has(value)) throw new TypeError('Blob cache arguments must not be cyclic');
+      if (
+        !Array.isArray(value) &&
+        Object.getPrototypeOf(value) !== Object.prototype &&
+        Object.getPrototypeOf(value) !== null
+      ) {
+        throw new TypeError('Blob cache arguments must contain only arrays and plain objects');
+      }
+      if (Object.getOwnPropertySymbols(value).length > 0) {
+        throw new TypeError('Blob cache arguments must not contain symbol keys');
+      }
+
+      ancestors.add(value);
+      let encoded: unknown;
+      if (Array.isArray(value)) {
+        encoded = ['array', value.map((item) => canonicalCacheValue(item, ancestors))];
+      } else {
+        // The prototype check above establishes the plain string-keyed object boundary.
+        const objectValue = value as Record<string, unknown>;
+        encoded = [
+          'object',
+          Object.keys(objectValue)
+            .toSorted()
+            .map((key) => [key, canonicalCacheValue(objectValue[key], ancestors)]),
+        ];
+      }
+      ancestors.delete(value);
+      return encoded;
+    }
+    default: {
+      throw new TypeError(`Unsupported Blob cache argument type: ${typeof value}`);
+    }
+  }
+}
 
-  const inlineKey = `${keyPrefix}:${args.join(':')}`;
+function blobKeyForArgs(keyPrefix: string, args: unknown[]): string {
+  const serialized = JSON.stringify(canonicalCacheValue(args, new Set()));
+  const inlineKey = `${keyPrefix}:${BLOB_KEY_ENCODING_VERSION}:${Buffer.from(serialized).toString('base64url')}`;
   if (Buffer.byteLength(inlineKey, 'utf8') <= MAX_INLINE_BLOB_KEY_BYTES) return inlineKey;
 
-  const digest = createHash('sha256').update(JSON.stringify(args)).digest('hex');
-  return `${keyPrefix}:sha256:${digest}`;
+  const digest = createHash('sha256').update(serialized).digest('hex');
+  return `${keyPrefix}:${BLOB_KEY_ENCODING_VERSION}:sha256:${digest}`;
 }
 
 /**
@@ -108,6 +164,21 @@ export function cachedQuery<T, Args extends unknown[]>(
   return (...args: Args): Promise<T> => nextCached(...args);
 }
 
+/**
+ * Cache compact, derived server payloads in Next's tagged cache.
+ *
+ * This deliberately has no Blob fallback: callers use it only after reducing
+ * raw database rows to a page-owned payload that stays below Next's cache
+ * limit. Arguments are part of the unstable_cache key, so callers should pass
+ * only canonical, meaningful selector values.
+ */
+export function cachedDerivedData<T, Args extends unknown[]>(
+  fn: (...args: Args) => Promise<T>,
+  keyPrefix: string,
+): (...args: Args) => Promise<T> {
+  return cachedQuery(fn, `derived:${keyPrefix}`);
+}
+
 /** Purge both unstable_cache (via revalidateTag) and blob storage. */
 export async function purgeAll(): Promise<number> {
   const deleted = await blobPurge();
diff --git a/packages/app/src/lib/api-documentation.ts b/packages/app/src/lib/api-documentation.ts
index 5167a85da..4a22d16c5 100644
--- a/packages/app/src/lib/api-documentation.ts
+++ b/packages/app/src/lib/api-documentation.ts
@@ -1,6 +1,8 @@
 import { DB_MODEL_TO_DISPLAY, DISPLAY_MODEL_TO_DB } from '@semianalysisai/inferencex-constants';
 import { COLLECTIVEX_VERSIONS } from '@semianalysisai/inferencex-db/collectivex/types';
 
+import { PUBLIC_API_ERRORS } from './public-api-errors';
+
 export type ApiDocumentationLocale = 'en' | 'zh';
 export type ApiGroupId = 'core' | 'external' | 'datasets' | 'collectivex' | 'diagnostics';
 export type ApiHttpMethod = 'GET';
@@ -231,6 +233,7 @@ const benchmarkRowSchema = objectSchemaWithOptional(
     conc: integerSchema,
     offload_mode: stringSchema,
     image: nullableStringSchema,
+    recipe_fingerprint: nullableStringSchema,
     metrics: metricMapSchema,
     workers: arraySchema(workerPowerSchema),
     date: { type: 'string', format: 'date' },
@@ -267,6 +270,7 @@ const benchmarkExample = [
     conc: 32,
     offload_mode: 'off',
     image: 'vllm/vllm-openai:v0.10.2',
+    recipe_fingerprint: '7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d7d72a33d',
     metrics: { median_ttft: 0.42, median_tpot: 0.018, tput_per_gpu: 128.4 },
     date: '2026-08-08',
     run_url: 'https://github.com/semianalysis/inference-benchmarks/actions/runs/123456789',
@@ -375,6 +379,10 @@ const submissionsSchema = objectSchema({
       is_multinode: booleanSchema,
       num_prefill_gpu: integerSchema,
       num_decode_gpu: integerSchema,
+      prefill_tp: integerSchema,
+      prefill_ep: integerSchema,
+      decode_tp: integerSchema,
+      decode_ep: integerSchema,
       date: { type: 'string', format: 'date' },
       total_datapoints: integerSchema,
       distinct_sequences: integerSchema,
@@ -429,49 +437,64 @@ const conversationItemSchema = objectSchema({
 const collectiveXDatasetSchema = objectSchema(
   {
     version: integerSchema,
-    run: objectSchema({
-      run_id: stringSchema,
-      run_attempt: integerSchema,
-      generated_at: { type: 'string', format: 'date-time' },
-      conclusion: nullableStringSchema,
-      source_sha: stringSchema,
-      requested_cases: integerSchema,
-      terminal_cases: integerSchema,
-      measured_cases: integerSchema,
-      unsupported_cases: integerSchema,
-      failed_cases: integerSchema,
-      requested_points: integerSchema,
-      terminal_points: integerSchema,
-      measured_points: integerSchema,
-      covered_skus: arraySchema(stringSchema),
-    }),
+    run: objectSchemaWithOptional(
+      {
+        run_id: stringSchema,
+        run_attempt: integerSchema,
+        generated_at: { type: 'string', format: 'date-time' },
+        conclusion: nullableStringSchema,
+        source_sha: stringSchema,
+        requested_cases: integerSchema,
+        terminal_cases: integerSchema,
+        measured_cases: integerSchema,
+        unsupported_cases: integerSchema,
+        failed_cases: integerSchema,
+        requested_points: integerSchema,
+        terminal_points: integerSchema,
+        measured_points: integerSchema,
+        covered_skus: arraySchema(stringSchema),
+        kv_requested_cases: integerSchema,
+        kv_measured_cases: integerSchema,
+      },
+      ['kv_requested_cases', 'kv_measured_cases'],
+    ),
     coverage: arraySchema(anyObjectSchema),
     series: arraySchema(anyObjectSchema),
+    kv: arraySchema(anyObjectSchema),
   },
   ['version', 'run', 'coverage', 'series'],
 );
-const collectiveRunSummarySchema = objectSchema({
-  run_id: stringSchema,
-  run_attempt: integerSchema,
-  generated_at: { type: 'string', format: 'date-time' },
-  conclusion: nullableStringSchema,
-  covered_skus: arraySchema(stringSchema),
-  requested_cases: integerSchema,
-  measured_cases: integerSchema,
-  requested_points: integerSchema,
-  terminal_points: integerSchema,
-  terminal_counts: objectSchema({
-    measured: integerSchema,
-    unsupported: integerSchema,
-    failed: integerSchema,
-  }),
-});
+const collectiveRunSummarySchema = objectSchemaWithOptional(
+  {
+    run_id: stringSchema,
+    run_attempt: integerSchema,
+    generated_at: { type: 'string', format: 'date-time' },
+    conclusion: nullableStringSchema,
+    covered_skus: arraySchema(stringSchema),
+    requested_cases: integerSchema,
+    measured_cases: integerSchema,
+    requested_points: integerSchema,
+    terminal_points: integerSchema,
+    terminal_counts: objectSchema({
+      measured: integerSchema,
+      unsupported: integerSchema,
+      failed: integerSchema,
+    }),
+    kv_cases: objectSchema({
+      requested: integerSchema,
+      measured: integerSchema,
+    }),
+  },
+  ['kv_cases'],
+);
 const percentileSchema = objectSchema({
   mean: numberSchema,
   p50: numberSchema,
   p75: numberSchema,
   p90: numberSchema,
+  p95: numberSchema,
   p99: numberSchema,
+  n: integerSchema,
 });
 const nullablePercentileSchema: ApiSchema = { oneOf: [percentileSchema, { type: 'null' }] };
 const idListSchema: ApiSchema = { type: 'string', pattern: '^\\d+(,\\d+)*$' };
@@ -592,16 +615,6 @@ export const apiOperations: readonly ApiOperation[] = [
         { type: 'string', format: 'date' },
         '2026-08-08',
       ),
-      parameter(
-        'benchmarkType',
-        'query',
-        false,
-        'string',
-        'Set to agentic_traces to scope per-run configuration coverage to Agentic Traces.',
-        '设为 agentic_traces 可将每次运行的配置覆盖限定为 Agentic Traces。',
-        { type: 'string', enum: ['agentic_traces'] },
-        'agentic_traces',
-      ),
       parameter(
         'exact',
         'query',
@@ -644,13 +657,13 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'The model is missing or unsupported.',
         '模型缺失或不受支持。',
-        'Unknown model',
+        PUBLIC_API_ERRORS.unknownModel,
       ),
       errorResponse(
         '500',
         'The benchmark query failed.',
         '基准查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'BenchmarkRows',
@@ -709,6 +722,16 @@ export const apiOperations: readonly ApiOperation[] = [
         { type: 'string', enum: ['agentic_traces'] },
         'agentic_traces',
       ),
+      parameter(
+        'view',
+        'query',
+        false,
+        'enum',
+        'calculator trims each row to the metrics the throughput calculator consumes, for a smaller payload. Omit for every stored metric, including measured power.',
+        'calculator 会将每行裁剪为吞吐量计算器所需的指标，以减小响应体积。省略则返回全部已存储指标，包括实测功率。',
+        { type: 'string', enum: ['calculator'] },
+        'calculator',
+      ),
     ],
     responses: [
       success('Historical benchmark rows.', '历史基准行。', benchmarkRowsSchema, benchmarkExample),
@@ -716,9 +739,14 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'Required parameters are missing or the model is unsupported.',
         '必填参数缺失或模型不受支持。',
-        'Missing required parameters',
+        PUBLIC_API_ERRORS.benchmarkHistoryParameters,
+      ),
+      errorResponse(
+        '500',
+        'The history query failed.',
+        '历史查询失败。',
+        PUBLIC_API_ERRORS.internal,
       ),
-      errorResponse('500', 'The history query failed.', '历史查询失败。', 'Internal server error'),
     ],
     responseShapeName: 'BenchmarkRows',
     curlUrl: `${API_BASE_URL}/api/v1/benchmarks/history?model=DeepSeek-R1-0528&isl=1024&osl=1024`,
@@ -746,6 +774,16 @@ export const apiOperations: readonly ApiOperation[] = [
         { type: 'string', format: 'date' },
         '2026-08-08',
       ),
+      parameter(
+        'benchmarkType',
+        'query',
+        false,
+        'string',
+        'Set to agentic_traces to scope per-run configuration coverage to Agentic Traces.',
+        '设为 agentic_traces 可将每次运行的配置覆盖限定为 Agentic Traces。',
+        { type: 'string', enum: ['agentic_traces'] },
+        'agentic_traces',
+      ),
     ],
     responses: [
       success(
@@ -774,13 +812,13 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'date is not YYYY-MM-DD.',
         'date 不是 YYYY-MM-DD。',
-        'Invalid date format',
+        PUBLIC_API_ERRORS.invalidDate,
       ),
       errorResponse(
         '500',
         'The workflow query failed.',
         '工作流查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'WorkflowInfo',
@@ -1041,6 +1079,10 @@ export const apiOperations: readonly ApiOperation[] = [
             is_multinode: false,
             num_prefill_gpu: 0,
             num_decode_gpu: 8,
+            prefill_tp: 1,
+            prefill_ep: 1,
+            decode_tp: 8,
+            decode_ep: 1,
             date: '2026-08-08',
             total_datapoints: 24,
             distinct_sequences: 3,
@@ -1424,9 +1466,12 @@ export const apiOperations: readonly ApiOperation[] = [
             terminal_points: 48,
             measured_points: 40,
             covered_skus: ['h200_sxm'],
+            kv_requested_cases: 4,
+            kv_measured_cases: 4,
           },
           coverage: [],
           series: [],
+          kv: [],
         },
       ),
       errorResponse(
@@ -1510,6 +1555,7 @@ export const apiOperations: readonly ApiOperation[] = [
               requested_points: 48,
               terminal_points: 48,
               terminal_counts: { measured: 40, unsupported: 8, failed: 0 },
+              kv_cases: { requested: 4, measured: 4 },
             },
           ],
           discovery_complete: true,
@@ -1599,9 +1645,12 @@ export const apiOperations: readonly ApiOperation[] = [
             terminal_points: 48,
             measured_points: 40,
             covered_skus: ['h200_sxm'],
+            kv_requested_cases: 4,
+            kv_measured_cases: 4,
           },
           coverage: [],
           series: [],
+          kv: [],
         },
       ),
       errorResponse(
@@ -1668,7 +1717,15 @@ export const apiOperations: readonly ApiOperation[] = [
         {
           '421': {
             id: 421,
-            isl: { mean: 18320, p50: 16440, p75: 20110, p90: 24880, p99: 31900 },
+            isl: {
+              mean: 18320,
+              p50: 16440,
+              p75: 20110,
+              p90: 24880,
+              p95: 27940,
+              p99: 31900,
+              n: 512,
+            },
             osl: null,
             kvCacheUtil: null,
             prefixCacheHitRate: null,
@@ -1679,13 +1736,13 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'ids is missing, malformed, or exceeds 200 unique IDs.',
         'ids 缺失、格式错误或超过 200 个唯一 ID。',
-        'Expected ids as comma-separated positive integers',
+        PUBLIC_API_ERRORS.idsRequired,
       ),
       errorResponse(
         '500',
         'The aggregate query failed.',
         '汇总查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'AgenticAggregateMap',
@@ -1740,11 +1797,15 @@ export const apiOperations: readonly ApiOperation[] = [
               decode_tp: 8,
               decode_ep: 1,
               decode_pp: null,
+              decode_dcp_size: 8,
+              decode_pcp_size: 1,
               decode_dp_attention: false,
               decode_num_workers: 1,
               prefill_tp: 8,
               prefill_ep: 1,
               prefill_pp: null,
+              prefill_dcp_size: 8,
+              prefill_pcp_size: 1,
               prefill_dp_attention: false,
               prefill_num_workers: 1,
               num_prefill_gpu: 0,
@@ -1763,15 +1824,20 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'id is missing or invalid.',
         'id 缺失或无效。',
-        'Expected a positive numeric id',
+        PUBLIC_API_ERRORS.idRequired,
       ),
       errorResponse(
         '404',
         'No benchmark result has this ID.',
         '没有使用此 ID 的基准结果。',
-        'Not found',
+        PUBLIC_API_ERRORS.notFound,
+      ),
+      errorResponse(
+        '500',
+        'The sibling query failed.',
+        '同组查询失败。',
+        PUBLIC_API_ERRORS.internal,
       ),
-      errorResponse('500', 'The sibling query failed.', '同组查询失败。', 'Internal server error'),
     ],
     responseShapeName: 'BenchmarkSiblings',
     curlUrl: `${API_BASE_URL}/api/v1/benchmark-siblings?id=421`,
@@ -1817,13 +1883,13 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'ids is missing, malformed, or exceeds 200 unique IDs.',
         'ids 缺失、格式错误或超过 200 个唯一 ID。',
-        'Expected ids as comma-separated positive integers',
+        PUBLIC_API_ERRORS.idsRequired,
       ),
       errorResponse(
         '500',
         'The derived metric query failed.',
         '派生指标查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'DerivedAgenticMetricMap',
@@ -1836,8 +1902,8 @@ export const apiOperations: readonly ApiOperation[] = [
     path: '/api/v1/request-timeline',
     summary: text('Read a request timeline', '读取请求时间线'),
     description: text(
-      'Returns a versioned benchmark window and per-request dispatch, acknowledgement, completion, token, phase, worker, and cancellation timing.',
-      '返回带版本的基准窗口，以及每个请求的调度、确认、完成、token、阶段、工作进程和取消时间。',
+      'Returns a versioned benchmark window and per-request replay identity, source provenance, dispatch, acknowledgement, completion, token, phase, worker, and cancellation timing.',
+      '返回带版本的基准窗口，以及每个请求的重放标识、来源溯源、调度、确认、完成、token、阶段、worker 和取消时间。',
     ),
     audience: 'public',
     stability: 'beta',
@@ -1863,32 +1929,41 @@ export const apiOperations: readonly ApiOperation[] = [
           endNs: integerSchema,
           durationS: numberSchema,
           requests: arraySchema(
-            objectSchema({
-              cid: stringSchema,
-              ti: integerSchema,
-              wid: stringSchema,
-              ad: integerSchema,
-              phase: stringSchema,
-              credit: integerSchema,
-              start: integerSchema,
-              ack: nullableNumberSchema,
-              end: integerSchema,
-              ttftMs: nullableNumberSchema,
-              tpotMs: nullableNumberSchema,
-              isl: nullableNumberSchema,
-              osl: nullableNumberSchema,
-              cancelled: booleanSchema,
-            }),
+            objectSchemaWithOptional(
+              {
+                cid: stringSchema,
+                ri: integerSchema,
+                ti: integerSchema,
+                srcTrace: stringSchema,
+                srcOuter: integerSchema,
+                srcInner: integerSchema,
+                srcKind: stringSchema,
+                wid: stringSchema,
+                ad: integerSchema,
+                phase: stringSchema,
+                credit: integerSchema,
+                start: integerSchema,
+                ack: nullableNumberSchema,
+                end: integerSchema,
+                ttftMs: nullableNumberSchema,
+                tpotMs: nullableNumberSchema,
+                isl: nullableNumberSchema,
+                osl: nullableNumberSchema,
+                cancelled: booleanSchema,
+              },
+              ['ri', 'srcTrace', 'srcOuter', 'srcInner', 'srcKind'],
+            ),
           ),
         }),
         {
-          version: 5,
+          version: 6,
           startNs: 1000000000,
           endNs: 2400000000,
           durationS: 1.4,
           requests: [
             {
               cid: 'trace-018',
+              ri: 0,
               ti: 0,
               wid: '7',
               ad: 0,
@@ -1910,19 +1985,19 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'id is missing or invalid.',
         'id 缺失或无效。',
-        'Expected a positive numeric id',
+        PUBLIC_API_ERRORS.idRequired,
       ),
       errorResponse(
         '404',
         'No timeline exists for this result.',
         '该结果没有时间线。',
-        'Not found',
+        PUBLIC_API_ERRORS.notFound,
       ),
       errorResponse(
         '500',
         'The timeline query failed.',
         '时间线查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'RequestTimeline',
@@ -1935,8 +2010,8 @@ export const apiOperations: readonly ApiOperation[] = [
     path: '/api/v1/server-log',
     summary: text('Read a benchmark server log', '读取基准服务器日志'),
     description: text(
-      'Returns the stored plain-text server log for one benchmark result ID.',
-      '返回某个基准结果 ID 已存储的纯文本服务器日志。',
+      'Returns one stored .log/.out file for a benchmark result ID. Use file with a name from server-log-files. Add offset or limit for a bounded chunk; add download=1 to stream the complete selected file as a text attachment.',
+      '返回某个基准结果 ID 已存储的一个 .log/.out 文件。file 应使用 server-log-files 返回的文件名。可添加 offset 或 limit 读取有界分块；添加 download=1 可将完整的当前文件作为文本附件流式下载。',
     ),
     audience: 'public',
     stability: 'beta',
@@ -1951,15 +2026,77 @@ export const apiOperations: readonly ApiOperation[] = [
         positiveIdSchema,
         421,
       ),
+      parameter(
+        'file',
+        'query',
+        false,
+        'string',
+        'Artifact-relative filename. Omit it to read the primary/legacy file.',
+        '产物内的相对文件名。省略时读取主文件或旧版日志文件。',
+        { type: 'string', minLength: 1, maxLength: 1024 },
+        'results/router.log',
+      ),
+      parameter(
+        'offset',
+        'query',
+        false,
+        'integer',
+        'Zero-based character offset. Supplying offset or limit enables chunked mode.',
+        '从零开始的字符偏移量。提供 offset 或 limit 后启用分块模式。',
+        { type: 'integer', minimum: 0, maximum: 2_000_000_000, default: 0 },
+        0,
+      ),
+      parameter(
+        'limit',
+        'query',
+        false,
+        'integer',
+        'Chunk size in characters, from 1 to 262144. Defaults to 65536.',
+        '分块字符数，范围为 1 到 262144，默认值为 65536。',
+        { type: 'integer', minimum: 1, maximum: 262_144, default: 65_536 },
+        65_536,
+      ),
+      parameter(
+        'download',
+        'query',
+        false,
+        'integer',
+        'Set to 1 to stream the complete selected file as a text attachment. Cannot be combined with offset or limit.',
+        '设为 1 时，将完整的当前文件作为文本附件流式下载。不能与 offset 或 limit 同时使用。',
+        { type: 'integer', enum: [1] },
+        1,
+      ),
     ],
     responses: [
       success(
-        'Benchmark result ID and server log text.',
-        '基准结果 ID 和服务器日志文本。',
-        objectSchema({ id: integerSchema, serverLog: stringSchema }),
-        { id: 421, serverLog: 'INFO engine initialized\nINFO benchmark complete' },
+        'Benchmark result ID and server log text, plus range metadata in chunked mode.',
+        '基准结果 ID 和服务器日志文本；分块模式还会返回范围元数据。',
+        {
+          oneOf: [
+            objectSchema({ id: integerSchema, serverLog: stringSchema }),
+            objectSchema({
+              id: integerSchema,
+              fileName: stringSchema,
+              serverLog: stringSchema,
+              offset: integerSchema,
+              nextOffset: nullableNumberSchema,
+            }),
+          ],
+        },
+        {
+          id: 421,
+          fileName: 'results/router.log',
+          serverLog: 'INFO router initialized\n',
+          offset: 0,
+          nextOffset: 24,
+        },
+      ),
+      errorResponse(
+        '400',
+        'id is missing or invalid.',
+        'id 缺失或无效。',
+        PUBLIC_API_ERRORS.idRequired,
       ),
-      errorResponse('400', 'id is missing or invalid.', 'id 缺失或无效。', 'Invalid id'),
       errorResponse(
         '404',
         'No server log exists for this result.',
@@ -1973,8 +2110,196 @@ export const apiOperations: readonly ApiOperation[] = [
         'Internal server error',
       ),
     ],
-    responseShapeName: 'ServerLog',
-    curlUrl: `${API_BASE_URL}/api/v1/server-log?id=421`,
+    responseShapeName: 'ServerLog | ServerLogChunk | text/plain attachment',
+    curlUrl: `${API_BASE_URL}/api/v1/server-log?id=421&file=results%2Frouter.log&offset=0&limit=65536`,
+  },
+  {
+    id: 'search-server-logs',
+    group: 'diagnostics',
+    method: 'GET',
+    path: '/api/v1/server-log-search',
+    summary: text('Search complete benchmark logs', '搜索完整基准测试日志'),
+    description: text(
+      'Runs a literal, case-insensitive search across every stored .log/.out file for one benchmark result. The bounded response contains contextual snippets without transferring complete log files.',
+      '对某个基准结果已存储的全部 .log/.out 文件执行不区分大小写的文本搜索。响应数量有上限，并返回上下文片段，无需传输完整日志文件。',
+    ),
+    audience: 'public',
+    stability: 'beta',
+    parameters: [
+      parameter(
+        'id',
+        'query',
+        true,
+        'integer',
+        'Positive benchmark result ID.',
+        '正整数基准结果 ID。',
+        positiveIdSchema,
+        421,
+      ),
+      parameter(
+        'q',
+        'query',
+        true,
+        'string',
+        'Literal search text, from 1 to 256 characters.',
+        '搜索文本，长度为 1 到 256 个字符，按原文匹配。',
+        { type: 'string', minLength: 1, maxLength: 256 },
+        'router ready',
+      ),
+      parameter(
+        'limit',
+        'query',
+        false,
+        'integer',
+        'Maximum matches to return, from 1 to 100. Defaults to 50.',
+        '最多返回的匹配数，范围为 1 到 100，默认值为 50。',
+        { type: 'integer', minimum: 1, maximum: 100, default: 50 },
+        50,
+      ),
+    ],
+    responses: [
+      success(
+        'Bounded matches with filenames, character offsets, and contextual text.',
+        '数量有上限的匹配结果，包含文件名、字符偏移量和上下文文本。',
+        objectSchema({
+          id: integerSchema,
+          query: stringSchema,
+          matches: arraySchema(
+            objectSchema({
+              fileName: stringSchema,
+              offset: integerSchema,
+              before: stringSchema,
+              match: stringSchema,
+              after: stringSchema,
+            }),
+          ),
+          truncated: booleanSchema,
+        }),
+        {
+          id: 421,
+          query: 'router ready',
+          matches: [
+            {
+              fileName: 'results/router.log',
+              offset: 128,
+              before: 'INFO ',
+              match: 'router ready',
+              after: ' on port 8000\n',
+            },
+          ],
+          truncated: false,
+        },
+      ),
+      errorResponse('400', 'The search parameters are invalid.', '搜索参数无效。', 'Invalid query'),
+      errorResponse(
+        '500',
+        'The complete-log search failed.',
+        '完整日志搜索失败。',
+        'Internal server error',
+      ),
+    ],
+    responseShapeName: 'ServerLogSearchResult',
+    curlUrl: `${API_BASE_URL}/api/v1/server-log-search?id=421&q=router%20ready&limit=50`,
+  },
+  {
+    id: 'get-server-log-files',
+    group: 'diagnostics',
+    method: 'GET',
+    path: '/api/v1/server-log-files',
+    summary: text('List stored benchmark log files', '列出已存储的基准测试日志文件'),
+    description: text(
+      'Lists every .log and .out filename retained from the matching server-log artifact. Paths are returned relative to the artifact root, with the primary file first.',
+      '列出匹配服务器日志产物中保留的所有 .log 和 .out 文件名。路径相对于产物根目录返回，主文件排在首位。',
+    ),
+    audience: 'public',
+    stability: 'beta',
+    parameters: [
+      parameter(
+        'id',
+        'query',
+        true,
+        'integer',
+        'Positive benchmark result ID.',
+        '正整数基准结果 ID。',
+        positiveIdSchema,
+        421,
+      ),
+    ],
+    responses: [
+      success(
+        'Artifact-relative .log/.out filenames.',
+        '相对于产物根目录的 .log/.out 文件名。',
+        arraySchema(stringSchema),
+        ['results/server.log', 'results/benchmark.log', 'results/router.log'],
+      ),
+      errorResponse(
+        '400',
+        'id is missing or invalid.',
+        'id 缺失或无效。',
+        PUBLIC_API_ERRORS.idRequired,
+      ),
+      errorResponse(
+        '404',
+        'No log bundle exists for this result.',
+        '该结果没有日志文件集合。',
+        PUBLIC_API_ERRORS.notFound,
+      ),
+      errorResponse(
+        '500',
+        'The log filename query failed.',
+        '日志文件名查询失败。',
+        PUBLIC_API_ERRORS.internal,
+      ),
+    ],
+    responseShapeName: 'ServerLogFileNames',
+    curlUrl: `${API_BASE_URL}/api/v1/server-log-files?id=421`,
+  },
+  {
+    id: 'get-log-availability',
+    group: 'diagnostics',
+    method: 'GET',
+    path: '/api/v1/log-availability',
+    summary: text('Check server-log availability', '检查服务器日志可用性'),
+    description: text(
+      'Returns only benchmark result IDs that have a stored server log. IDs are deduplicated and at most 500 are accepted.',
+      '仅返回具有已存储服务器日志的基准结果 ID。ID 会去重，最多接受 500 个。',
+    ),
+    audience: 'public',
+    stability: 'beta',
+    parameters: [
+      parameter(
+        'ids',
+        'query',
+        true,
+        'comma-separated integers',
+        'One to 500 positive benchmark result IDs.',
+        '1 到 500 个正整数基准结果 ID。',
+        idListSchema,
+        '421,422',
+      ),
+    ],
+    responses: [
+      success(
+        'Available result IDs mapped to true. Missing keys have no server log.',
+        '可用结果 ID 映射为 true。缺失的键表示没有服务器日志。',
+        mapSchema(booleanSchema),
+        { '421': true },
+      ),
+      errorResponse(
+        '400',
+        'ids is missing, malformed, or exceeds 500 unique IDs.',
+        'ids 缺失、格式错误或超过 500 个唯一 ID。',
+        'Expected ids as comma-separated positive integers',
+      ),
+      errorResponse(
+        '500',
+        'The server-log availability query failed.',
+        '服务器日志可用性查询失败。',
+        'Internal server error',
+      ),
+    ],
+    responseShapeName: 'LogAvailabilityMap',
+    curlUrl: `${API_BASE_URL}/api/v1/log-availability?ids=421,422`,
   },
   {
     id: 'get-trace-availability',
@@ -2011,13 +2336,13 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'ids is missing, malformed, or exceeds 500 unique IDs.',
         'ids 缺失、格式错误或超过 500 个唯一 ID。',
-        'Expected ids as comma-separated positive integers',
+        PUBLIC_API_ERRORS.idsRequired,
       ),
       errorResponse(
         '500',
         'The trace availability query failed.',
         '跟踪可用性查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'TraceAvailabilityMap',
@@ -2064,13 +2389,13 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'ids is missing, malformed, or exceeds 200 unique IDs.',
         'ids 缺失、格式错误或超过 200 个唯一 ID。',
-        'Expected ids as comma-separated positive integers',
+        PUBLIC_API_ERRORS.idsRequired,
       ),
       errorResponse(
         '500',
         'The histogram query failed.',
         '直方图查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'TraceHistogramMap',
@@ -2083,8 +2408,8 @@ export const apiOperations: readonly ApiOperation[] = [
     path: '/api/v1/trace-server-metrics',
     summary: text('Read trace server metrics', '读取跟踪服务器指标'),
     description: text(
-      'Returns point metadata and chart-ready time series for cache usage, queue depth, prefill and decode throughput, prompt-token sources, and metric sources.',
-      '返回点元数据，以及缓存使用率、队列深度、预填充和解码吞吐、提示 token 来源与指标来源的图表时间序列。',
+      'Returns point metadata and chart-ready aggregate time series for cache usage, queue depth, prefill and decode throughput, and prompt-token sources. metricSources contains source descriptors; source-specific arrays are loaded by the point-detail UI only when selected.',
+      '返回点元数据，以及缓存使用率、队列深度、预填充和解码吞吐、提示 token 来源的图表聚合时间序列。metricSources 仅包含来源描述信息；详情页只会在用户选中某个来源时加载其专属数组。',
     ),
     audience: 'public',
     stability: 'beta',
@@ -2152,19 +2477,19 @@ export const apiOperations: readonly ApiOperation[] = [
         '400',
         'id is missing or invalid.',
         'id 缺失或无效。',
-        'Expected a positive numeric id',
+        PUBLIC_API_ERRORS.idRequired,
       ),
       errorResponse(
         '404',
         'No server metrics exist for this result.',
         '该结果没有服务器指标。',
-        'Not found',
+        PUBLIC_API_ERRORS.notFound,
       ),
       errorResponse(
         '500',
         'The server metric query failed.',
         '服务器指标查询失败。',
-        'Internal server error',
+        PUBLIC_API_ERRORS.internal,
       ),
     ],
     responseShapeName: 'TraceServerMetrics',
diff --git a/packages/app/src/lib/api-route-catalog.test.ts b/packages/app/src/lib/api-route-catalog.test.ts
index 55c4ce175..ee843d9d2 100644
--- a/packages/app/src/lib/api-route-catalog.test.ts
+++ b/packages/app/src/lib/api-route-catalog.test.ts
@@ -15,6 +15,7 @@ import {
 import {
   apiContractSourceDigests,
   apiRouteCatalog,
+  stablePublicApiContracts,
   type ApiRouteHttpMethod,
   type BilingualReviewText,
 } from './api-route-catalog';
@@ -243,6 +244,38 @@ describe('API route catalog', () => {
       );
     }
   });
+
+  it('keeps stable public parameters, statuses, errors, auth, cache policy, and schemas explicit', () => {
+    const stableOperations = apiOperations.filter((operation) => operation.stability === 'stable');
+    expect(stablePublicApiContracts.map((contract) => contract.operationId).toSorted()).toEqual(
+      stableOperations.map((operation) => operation.id).toSorted(),
+    );
+
+    for (const contract of stablePublicApiContracts) {
+      const operation = stableOperations.find((candidate) => candidate.id === contract.operationId);
+      expect(
+        operation,
+        `${contract.operationId} contract metadata has no stable operation`,
+      ).toBeDefined();
+      expect(operation?.parameters.map((parameter) => parameter.name)).toEqual(contract.parameters);
+      expect(operation?.responses.map((response) => response.status)).toEqual(contract.statuses);
+      expect(operation?.responseShapeName).toBe(contract.responseShapeName);
+      const documentedErrors = operation?.responses
+        .filter((response) => response.status !== '200')
+        .map((response) => {
+          const example = response.example;
+          return example &&
+            typeof example === 'object' &&
+            'error' in example &&
+            typeof example.error === 'string'
+            ? example.error
+            : undefined;
+        });
+      expect(documentedErrors).toEqual(contract.errorExamples);
+      expect(contract.auth).toBe('none');
+      expect(['public-db-day', 'framework-release-hour']).toContain(contract.cachePolicy);
+    }
+  });
 });
 
 describe('published API documentation invariants', () => {
@@ -308,6 +341,60 @@ describe('published API documentation invariants', () => {
     }
   });
 
+  it('documents optional submission topology and CollectiveX KV fields', () => {
+    const submissions = apiOperations.find((operation) => operation.id === 'get-submissions');
+    const submissionSummary = submissions?.responses.find((response) => response.status === '200')
+      ?.schema.properties?.summary?.items?.properties;
+    expect(submissionSummary).toMatchObject({
+      prefill_tp: { type: 'integer' },
+      prefill_ep: { type: 'integer' },
+      decode_tp: { type: 'integer' },
+      decode_ep: { type: 'integer' },
+    });
+
+    const latest = apiOperations.find((operation) => operation.id === 'get-collectivex-latest');
+    const latestSchema = latest?.responses.find((response) => response.status === '200')?.schema;
+    expect(latestSchema?.properties?.run?.properties).toMatchObject({
+      kv_requested_cases: { type: 'integer' },
+      kv_measured_cases: { type: 'integer' },
+    });
+    expect(latestSchema?.properties?.kv).toMatchObject({ type: 'array' });
+    expect(latestSchema?.required).not.toContain('kv');
+
+    const runs = apiOperations.find((operation) => operation.id === 'list-collectivex-runs');
+    const runSummary = runs?.responses.find((response) => response.status === '200')?.schema
+      .properties?.runs?.items?.properties;
+    expect(runSummary?.kv_cases).toMatchObject({ type: 'object' });
+  });
+
+  it('uses canonical shared-route error examples', () => {
+    const bulkIds = [
+      'get-agentic-aggregates',
+      'get-derived-agentic-metrics',
+      'get-trace-availability',
+      'get-trace-histograms',
+    ];
+    const singleId = [
+      'get-benchmark-siblings',
+      'get-request-timeline',
+      'get-server-log',
+      'get-trace-server-metrics',
+    ];
+
+    for (const operationId of bulkIds) {
+      const operation = apiOperations.find((candidate) => candidate.id === operationId);
+      expect(operation?.responses.find((response) => response.status === '400')?.example).toEqual({
+        error: 'ids query param is required',
+      });
+    }
+    for (const operationId of singleId) {
+      const operation = apiOperations.find((candidate) => candidate.id === operationId);
+      expect(operation?.responses.find((response) => response.status === '400')?.example).toEqual({
+        error: 'id is required (benchmark_result_id)',
+      });
+    }
+  });
+
   it('documents every path parameter and at least one successful response', () => {
     for (const operation of apiOperations) {
       const pathNames = pathParameterNames(operation.path).toSorted();
diff --git a/packages/app/src/lib/api-route-catalog.ts b/packages/app/src/lib/api-route-catalog.ts
index 4ce362a30..a14b055c0 100644
--- a/packages/app/src/lib/api-route-catalog.ts
+++ b/packages/app/src/lib/api-route-catalog.ts
@@ -76,7 +76,7 @@ export const apiRouteCatalog = [
       en: 'UI-only overlay for unofficial workflow artifacts; upstream artifact availability and shape are not stable.',
       zh: '仅供界面叠加非官方工作流制品；上游制品的可用性和结构并不稳定。',
     },
-    sourceSha256: 'ef85fec8468757b0122c5fcebde78527c2efe62f9fedc95799cad56a457a8f04',
+    sourceSha256: '4a3f3da8399c741c26f0f502d44b1870a8ccdc05775edfd6ea3dee4e020df25c',
   },
   {
     source: 'src/app/api/v1/agentic-aggregates/route.ts',
@@ -84,7 +84,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-agentic-aggregates',
-    sourceSha256: '6a99fc156b02b86c5c26ca5ba6b05a280e064dd6123aaba528f124ff0f13c577',
+    sourceSha256: '7dab9929039926b2697d5db71c143eb39f94e795a0d83bb0348db8869f279a5f',
   },
   {
     source: 'src/app/api/v1/availability/route.ts',
@@ -92,7 +92,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-availability',
-    sourceSha256: '12bc0fa8930792897ed8300bd3be1333826b51903c04c6fbdec335f13d80ca7a',
+    sourceSha256: 'f1c5845a59cefbe03f57f471a5bb63b194a8c0aa5dccc188febdc177003310f1',
   },
   {
     source: 'src/app/api/v1/benchmark-siblings/route.ts',
@@ -100,7 +100,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-benchmark-siblings',
-    sourceSha256: '36b00e5d1c761e6a76e2c0454d9be4a15b7ccb7f3af5bdccd8ac61b6e5ff5312',
+    sourceSha256: '2b20de8b2b67ed53027eba478068f8f22ae7e3843ac38423e8ce6b27c1f74fd6',
   },
   {
     source: 'src/app/api/v1/benchmarks/route.ts',
@@ -108,7 +108,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'list-benchmarks',
-    sourceSha256: '37b5a31613a9c5a2e1de35758551dfdbbb8b920fcd6ae6baeedf973c8802bc2c',
+    sourceSha256: 'c6a5b78108b7e0d523b11590e1e34ef2e8c2d5457673eb41338d93d3d8f04909',
   },
   {
     source: 'src/app/api/v1/benchmarks/history/route.ts',
@@ -116,7 +116,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'list-benchmark-history',
-    sourceSha256: 'd4b3d2ad8ed6e35df70c6b651f71c9d86b6e1dad9c3eaf6eadc2a8591318d1fb',
+    sourceSha256: '42ba4d72298084ffe079ffcee7b6309ef3fb94d1df4769be4485d95097aa8cbe',
   },
   {
     source: 'src/app/api/v1/collectivex/latest/route.ts',
@@ -124,7 +124,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-collectivex-latest',
-    sourceSha256: '41c67c25c62ae2bad5f2ef16faf7a3e5a5c4440a53e3701ae68cf0fdb0973f18',
+    sourceSha256: '4d50dc7bdce61936c135c7a480fa048b97658737121ad9ff1b964d3db4cfa7ef',
   },
   {
     source: 'src/app/api/v1/collectivex/runs/route.ts',
@@ -132,7 +132,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'list-collectivex-runs',
-    sourceSha256: 'bce17c290c75ddae3300125b4cefec684f58748cbd524dd4b6916e2728660f23',
+    sourceSha256: '60548a817e1d408e3ed3ee993b48bcecfd5b1ddffe6dc554228eb131c43b1228',
   },
   {
     source: 'src/app/api/v1/collectivex/runs/[runId]/route.ts',
@@ -140,7 +140,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-collectivex-run',
-    sourceSha256: 'dabbf260d406f774a880ea4c19c25271775755107ada6d036e2cf6b13f818cbb',
+    sourceSha256: '911688cb21861d86639c1c75e9081c263433a6ab4d41959fb9a0e723e507733e',
   },
   {
     source: 'src/app/api/v1/collectivex/runs/[runId]/route.ts',
@@ -151,7 +151,7 @@ export const apiRouteCatalog = [
       en: 'Authenticated CollectiveX administration mutation; deletion is not part of the public read API.',
       zh: '需要身份验证的 CollectiveX 管理写操作；删除不属于公开只读 API。',
     },
-    sourceSha256: 'dabbf260d406f774a880ea4c19c25271775755107ada6d036e2cf6b13f818cbb',
+    sourceSha256: '911688cb21861d86639c1c75e9081c263433a6ab4d41959fb9a0e723e507733e',
   },
   {
     source: 'src/app/api/v1/datasets/route.ts',
@@ -159,7 +159,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'list-datasets',
-    sourceSha256: '22eb9d7bb5f91d64a19d43723c4be73c239513073819ad5af83864a3b4858505',
+    sourceSha256: '2a779884ef1a44c9a14f7f4a2005495a9809aa91448c1401674fd9d99b32c65e',
   },
   {
     source: 'src/app/api/v1/datasets/[slug]/route.ts',
@@ -202,7 +202,7 @@ export const apiRouteCatalog = [
       en: 'UI-only evaluation sample reader backed by live workflow artifacts with an unstable artifact contract.',
       zh: '仅供界面读取由实时工作流制品支持的评测样本；该制品契约不稳定。',
     },
-    sourceSha256: '7790c89e48b71f0cb2a819f34ec649ce300982fa41bba20a2b996b6c9768b6c7',
+    sourceSha256: 'd5b8c36466c5882fa253e653997c7c4dd181489d754aca6fef1f98aaa103cf65',
   },
   {
     source: 'src/app/api/v1/eval-samples/route.ts',
@@ -213,7 +213,7 @@ export const apiRouteCatalog = [
       en: 'UI drill-down for evaluation samples; its pagination and sample payload remain page-owned.',
       zh: '用于界面下钻评测样本；其分页和样本载荷仍由页面内部使用。',
     },
-    sourceSha256: '0e05c3d98153a53e5025e6c9973ac6a0eef294ad8cb6ac8748b9a1478067b233',
+    sourceSha256: '865f41e25148e30e5de094af98773eebdfc67a395cac7e1a4b75e9b96b23bf85',
   },
   {
     source: 'src/app/api/v1/evaluations/route.ts',
@@ -221,7 +221,7 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'list-evaluations',
-    sourceSha256: '8eab5b9d1df62b64b89458943d6e2182b62713c22421a30f427044f664b76b04',
+    sourceSha256: '6fac9705f7d595b00525639d861dad2cb67dcc51fea5ef2e69da7c2d61630f35',
   },
   {
     source: 'src/app/api/v1/feedback/route.ts',
@@ -262,7 +262,7 @@ export const apiRouteCatalog = [
       en: 'Secret-protected cache invalidation mutation for operators; it is not a public application contract.',
       zh: '供运维人员使用的密钥保护缓存失效写操作；它不是公开应用契约。',
     },
-    sourceSha256: 'b081b54a5960e6da0987e9ccee777bfbc40c03d0206823472ee7d0f0e969e8d7',
+    sourceSha256: 'eadfc008403b3a5321f9b857897ed5c9fce2de29dee34279fb559cf1b763c247',
   },
   {
     source: 'src/app/api/v1/latest-images/route.ts',
@@ -270,7 +270,15 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-latest-images',
-    sourceSha256: '67e50209f2a1200b56e418c9edc29eba955e6c1273319f98a2deb8657b79e895',
+    sourceSha256: 'ae0d5535af9f5bf8d287c04f915067c0e470ec907efb2a848cb6c35660770d2d',
+  },
+  {
+    source: 'src/app/api/v1/log-availability/route.ts',
+    path: '/api/v1/log-availability',
+    method: 'GET',
+    classification: 'published-read',
+    operationId: 'get-log-availability',
+    sourceSha256: '1dab7aee9d6adf3304807823dba7463f8ccc4575afb4506e93f390b042ff2ac9',
   },
   {
     source: 'src/app/api/v1/overview/route.ts',
@@ -278,10 +286,10 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'page-bff',
     exclusionReason: {
-      en: 'Page-owned BFF aggregation whose tier, comparison, and calculator projections are coupled to the overview UI.',
-      zh: '由页面拥有的 BFF 聚合；其档位、比较和计算器投影与概览界面紧密耦合。',
+      en: 'Page-owned BFF aggregation whose tier, comparison, row-scope, and calculator projections are coupled to the overview UI.',
+      zh: '由页面拥有的 BFF 聚合；其档位、比较、行范围和计算器投影与概览界面紧密耦合。',
     },
-    sourceSha256: '7c9830baae39d533ba7db36d44fe520f89cb36b32ffc54ef29404693d0b6b120',
+    sourceSha256: '499089d01754aa470d0ced953d9818d7d2e620dd3521685099924750249f13e2',
   },
   {
     source: 'src/app/api/v1/reliability/route.ts',
@@ -289,7 +297,18 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'list-reliability',
-    sourceSha256: 'fe295d04ecf45bfcbd5d59518eb9a5b04a7f460a95053d79f7f904502d39d189',
+    sourceSha256: 'ce1c5db78b47548beb77a69797f10fb33853cde01cea5c44675c8ad3519bcf20',
+  },
+  {
+    source: 'src/app/api/v1/request-chart-data/route.ts',
+    path: '/api/v1/request-chart-data',
+    method: 'GET',
+    classification: 'page-bff',
+    exclusionReason: {
+      en: 'Agentic point-detail BFF with a compact dictionary-encoded request projection coupled to the chart implementation.',
+      zh: '智能体数据点详情页专用 BFF；其字典编码的精简请求投影与图表实现紧密耦合。',
+    },
+    sourceSha256: '44f5a6830358417eb1402c948051fe13dd05bee392c0de7b1511654d94bb7c43',
   },
   {
     source: 'src/app/api/v1/request-timeline/route.ts',
@@ -299,13 +318,40 @@ export const apiRouteCatalog = [
     operationId: 'get-request-timeline',
     sourceSha256: '30659cb757b9fd23411757051b650cfb564016aa15c8b9fec48676b9591f01e9',
   },
+  {
+    source: 'src/app/api/v1/resident-sequence-lengths/route.ts',
+    path: '/api/v1/resident-sequence-lengths',
+    method: 'GET',
+    classification: 'ui-artifact-read',
+    exclusionReason: {
+      en: 'UI-only mergeable sequence-length sketches for the resident inference-chart point set.',
+      zh: '仅供界面合并当前推理图表数据点的序列长度 sketch。',
+    },
+    sourceSha256: '7a66a7116be3bf63a8c7fc1a54cc7eaf97405858e633daa0802cc2ae7cdbb37e',
+  },
+  {
+    source: 'src/app/api/v1/server-log-files/route.ts',
+    path: '/api/v1/server-log-files',
+    method: 'GET',
+    classification: 'published-read',
+    operationId: 'get-server-log-files',
+    sourceSha256: '367628818cb02273505c4826d35407e10c45ed4eced69b628da091cde79db0cf',
+  },
+  {
+    source: 'src/app/api/v1/server-log-search/route.ts',
+    path: '/api/v1/server-log-search',
+    method: 'GET',
+    classification: 'published-read',
+    operationId: 'search-server-logs',
+    sourceSha256: '029a578cc8417a3ff34b6585fcc91c568e9973a2c00aab025a11cdbe9a847cb3',
+  },
   {
     source: 'src/app/api/v1/server-log/route.ts',
     path: '/api/v1/server-log',
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-server-log',
-    sourceSha256: '0462679b059bb8e2c48507e147faa24389a76b04312f06dbf9ab3724352a1239',
+    sourceSha256: '13e61c279840bcc176bb75be1959cc417b711bcb329aaf278681416c078a9b69',
   },
   {
     source: 'src/app/api/v1/submissions/route.ts',
@@ -339,13 +385,24 @@ export const apiRouteCatalog = [
     operationId: 'get-trace-histograms',
     sourceSha256: '1157a59930b754787872b0d543167c917477be27e63a6fab0cc487ffab5d4825',
   },
+  {
+    source: 'src/app/api/v1/trace-server-metric-source/route.ts',
+    path: '/api/v1/trace-server-metric-source',
+    method: 'GET',
+    classification: 'page-bff',
+    exclusionReason: {
+      en: 'Agentic point-detail BFF that lazily returns the full time-series arrays for one UI-selected metric source.',
+      zh: '智能体数据点详情页专用 BFF；按界面选择按需返回单个指标来源的完整时间序列。',
+    },
+    sourceSha256: '6bf3444510f1451dff1c76414a257faec0c657d8a01ae821035a9e80addccd2c',
+  },
   {
     source: 'src/app/api/v1/trace-server-metrics/route.ts',
     path: '/api/v1/trace-server-metrics',
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-trace-server-metrics',
-    sourceSha256: 'b1bb5b859a28ec18abd0b4ddebf5e505974a08b7ed9693e5bab118f4533f3ce6',
+    sourceSha256: '365b428f2c32e5461cfe13966292eccecfb285e9bf348617383dda821619251f',
   },
   {
     source: 'src/app/api/v1/workflow-info/route.ts',
@@ -353,10 +410,169 @@ export const apiRouteCatalog = [
     method: 'GET',
     classification: 'published-read',
     operationId: 'get-workflow-info',
-    sourceSha256: 'b7b0f215e9bf2c766ce2d4c9a13090504c1704a67f697af3a2fd1f318de1e6d6',
+    sourceSha256: '8849e70ef8a890370cb13a57ef7792e5f8522a93a4fdff7fa811a9a58f5cac32',
   },
 ] as const satisfies readonly ApiRouteCatalogEntry[];
 
+export type PublicApiCachePolicy = 'public-db-day' | 'framework-release-hour';
+
+export interface StablePublicApiContract {
+  readonly operationId: string;
+  readonly parameters: readonly string[];
+  readonly statuses: readonly `${number}`[];
+  readonly auth: 'none';
+  readonly cachePolicy: PublicApiCachePolicy;
+  readonly errorExamples: readonly string[];
+  readonly responseShapeName: string;
+}
+
+/**
+ * Reviewable behavioral contract for stable public reads. This intentionally
+ * records behavior rather than defining DTOs: raw query-row types remain
+ * canonical in the database package.
+ */
+export const stablePublicApiContracts = [
+  {
+    operationId: 'get-availability',
+    parameters: [],
+    statuses: ['200', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Internal server error'],
+    responseShapeName: 'AvailabilityRows',
+  },
+  {
+    operationId: 'list-benchmarks',
+    parameters: ['model', 'date', 'exact', 'runId', 'exactRun'],
+    statuses: ['200', '400', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Unknown model', 'Internal server error'],
+    responseShapeName: 'BenchmarkRows',
+  },
+  {
+    operationId: 'list-benchmark-history',
+    parameters: ['model', 'isl', 'osl', 'benchmarkType', 'view'],
+    statuses: ['200', '400', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['model, isl, and osl are required', 'Internal server error'],
+    responseShapeName: 'BenchmarkRows',
+  },
+  {
+    operationId: 'get-workflow-info',
+    parameters: ['date', 'benchmarkType'],
+    statuses: ['200', '400', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Invalid date format (YYYY-MM-DD required)', 'Internal server error'],
+    responseShapeName: 'WorkflowInfo',
+  },
+  {
+    operationId: 'list-evaluations',
+    parameters: [],
+    statuses: ['200', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Internal server error'],
+    responseShapeName: 'EvaluationRows',
+  },
+  {
+    operationId: 'list-reliability',
+    parameters: [],
+    statuses: ['200', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Internal server error'],
+    responseShapeName: 'ReliabilityRows',
+  },
+  {
+    operationId: 'get-tco-feed',
+    parameters: [
+      'model',
+      'workloads',
+      'tiers',
+      'date',
+      'format',
+      'view',
+      'weights',
+      'workload_weights',
+      'alpha',
+    ],
+    statuses: ['200', '400', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: [
+      'Invalid tiers: expected comma-separated positive numbers',
+      'Internal server error',
+    ],
+    responseShapeName: 'TcoFeed',
+  },
+  {
+    operationId: 'get-submissions',
+    parameters: [],
+    statuses: ['200', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Internal server error'],
+    responseShapeName: 'Submissions',
+  },
+  {
+    operationId: 'get-framework-releases',
+    parameters: [],
+    statuses: ['200', '500'],
+    auth: 'none',
+    cachePolicy: 'framework-release-hour',
+    errorExamples: ['Internal server error'],
+    responseShapeName: 'FrameworkReleases',
+  },
+  {
+    operationId: 'get-latest-images',
+    parameters: [],
+    statuses: ['200', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Internal server error'],
+    responseShapeName: 'LatestImageRows',
+  },
+  {
+    operationId: 'list-datasets',
+    parameters: [],
+    statuses: ['200', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Internal server error'],
+    responseShapeName: 'DatasetRecords',
+  },
+  {
+    operationId: 'get-dataset',
+    parameters: ['slug'],
+    statuses: ['200', '404', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Not found', 'Internal server error'],
+    responseShapeName: 'DatasetDetail',
+  },
+  {
+    operationId: 'list-dataset-conversations',
+    parameters: ['slug', 'search', 'limit', 'offset', 'sort'],
+    statuses: ['200', '400', '404', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['search too long', 'Not found', 'Internal server error'],
+    responseShapeName: 'ConversationList',
+  },
+  {
+    operationId: 'get-dataset-conversation',
+    parameters: ['slug', 'convId'],
+    statuses: ['200', '404', '500'],
+    auth: 'none',
+    cachePolicy: 'public-db-day',
+    errorExamples: ['Not found', 'Internal server error'],
+    responseShapeName: 'ConversationDetail',
+  },
+] as const satisfies readonly StablePublicApiContract[];
+
 export interface ApiContractSourceDigest {
   /** Path relative to packages/app. */
   readonly source: string;
@@ -370,9 +586,73 @@ export interface ApiContractSourceDigest {
  * touching a route module. Digest changes require an explicit documentation review.
  */
 export const apiContractSourceDigests = [
+  {
+    source: 'src/lib/api-cache.ts',
+    sourceSha256: 'b710c4ce4c2dd0a6eb3b662c9e426e301aee3afe3d64fba35745b9323be39ddd',
+    reviewArea: {
+      en: 'Public CDN tags, cache lifetimes, Blob key dimensions, and purge behavior.',
+      zh: '公开 CDN 标签、缓存时长、Blob 键维度和清除行为。',
+    },
+  },
+  {
+    source: 'src/lib/blob-cache.ts',
+    sourceSha256: 'f15476f437c5ffea9d601d5ae1fa3dfafa77b1f53861ec02f0c359bf0f59c2ff',
+    reviewArea: {
+      en: 'Blob cache read, write, prefix migration, and purge behavior.',
+      zh: 'Blob 缓存读取、写入、前缀迁移和清除行为。',
+    },
+  },
+  {
+    source: 'src/lib/cached-read-route.ts',
+    sourceSha256: '3b2fdf31075919e08b34514b6434906f39d9d20bea996da7f59bf42cacf36715',
+    reviewArea: {
+      en: 'Shared parameterless cached-read fixture, response, and public-error behavior.',
+      zh: '共享无参数缓存读取的夹具、响应和公开错误行为。',
+    },
+  },
+  {
+    source: 'src/lib/bearer-auth.ts',
+    sourceSha256: 'f9bb8619b6b9c3017fa780a8956aa7584fdd54d89f86f8335c295077cab71f4d',
+    reviewArea: {
+      en: 'Byte-safe constant-time Bearer credential comparison used by administration routes.',
+      zh: '管理路由使用的字节安全恒定时间 Bearer 凭据比较。',
+    },
+  },
+  {
+    source: 'src/lib/public-api-errors.ts',
+    sourceSha256: '48301bb3567870921f5cae0e0a35240c9a687767b89f96141b36be41afbaa16a',
+    reviewArea: {
+      en: 'Canonical public JSON error strings shared by handlers and OpenAPI examples.',
+      zh: '处理程序与 OpenAPI 示例共享的规范公开 JSON 错误字符串。',
+    },
+  },
+  {
+    source: 'src/lib/eval-sample-params.ts',
+    sourceSha256: '60073c6a3dcd7b4e16d2024606a2c7f457c35d127d89afaf361d382e198e88b9',
+    reviewArea: {
+      en: 'Shared stored/live evaluation sample filter and pagination behavior.',
+      zh: '存储与实时评测样本共享的筛选和分页行为。',
+    },
+  },
+  {
+    source: 'src/lib/submissions-types.ts',
+    sourceSha256: '1665a6a65f061184458997069b3222c9243b993f12573d83b0d10e571046a4bc',
+    reviewArea: {
+      en: 'Submission response topology and canonical database row type exports.',
+      zh: '提交响应拓扑和规范数据库行类型导出。',
+    },
+  },
+  {
+    source: 'src/lib/benchmark-id.ts',
+    sourceSha256: '36f9adf8a92cf5830abc01d0a389aa21eca9f429f31a9231b5d77e108c985d35',
+    reviewArea: {
+      en: 'Client-side persisted benchmark identifier recognition.',
+      zh: '客户端持久化基准标识符识别。',
+    },
+  },
   {
     source: 'src/app/api/v1/id-routes.ts',
-    sourceSha256: '65a8ebf97d500fd4164e068dc7d303d82e74e5ed84301dcd19ef8d386317b721',
+    sourceSha256: '7dde26440ea11d775aa0daca0a6a06671bfdd8cdcb75979ca56eff3c2b90f78c',
     reviewArea: {
       en: 'Shared positive-ID and ID-list validation, status codes, and error payloads for diagnostic reads.',
       zh: '诊断读取共享的正整数 ID 与 ID 列表校验、状态码和错误载荷。',
@@ -380,7 +660,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: 'src/lib/api.ts',
-    sourceSha256: '03809377af6c2ee938169a065e06dccb25d983d7171a54b998ffa08dd970d306',
+    sourceSha256: '54e398a2041040ed7b66bd5c813041db26287111774dafd13a70d5e545fb24f3',
     reviewArea: {
       en: 'Public API client parameter serialization and TypeScript response contracts.',
       zh: '公开 API 客户端的参数序列化和 TypeScript 响应契约。',
@@ -388,7 +668,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: 'src/lib/overview-data.ts',
-    sourceSha256: '7757b32c1769bf4e25e8dd156ba049830ab25b8e3cf345d718cfad46adfd8e43',
+    sourceSha256: '8ca280fec5bbe633493cf021fee294927aead799742ac07ac7dbe8c5092e38f2',
     reviewArea: {
       en: 'Overview BFF tier, engine, comparison-window, reference, and model-scope parameters plus the OverviewPageData response shape.',
       zh: '概览 BFF 的档位、引擎、对比时间窗口、参考硬件和模型范围参数，以及 OverviewPageData 响应结构。',
@@ -404,7 +684,12 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../constants/src/models.ts',
-    sourceSha256: '50b87ca89143ede642f7d6c86bbfa7e42ce780c90572e242550d028b68263e04',
+    // Reviewed again for the upstream sync merge: adds MODEL_RELEASE_DATES
+    // entries for the fork-only DeepSeek-Coder-V2-Lite-Instruct and
+    // Gemma-4-31B models (see .claude/skills/fork-changelog/SKILL.md). No
+    // published model name, alias, or parameter enum is touched, and no
+    // endpoint exposes a release date, so the docs stand.
+    sourceSha256: '9330668d9a198d1e7a0bcb5d8d0f16373e056724218fa1b2ae40ef5e11b3aee0',
     reviewArea: {
       en: 'Published benchmark and TCO model names, aliases, and parameter enums.',
       zh: '已发布基准与 TCO 模型名称、别名和参数枚举。',
@@ -420,15 +705,15 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/etl/compute-request-timeline.ts',
-    sourceSha256: '7cf9d1ab318ca9c21d9200ca4347437ce9f7bb3159bedb29a169cdf6ba46baae',
+    sourceSha256: '377d6f6cab10d6d7d59e7021d16e2ca872a2bfff6d28030966e1b36189d6bbc0',
     reviewArea: {
-      en: 'Request timeline contract version and event timing field semantics.',
-      zh: '请求时间线契约版本和事件计时字段语义。',
+      en: 'Request timeline contract version, replay identity, source provenance, and event timing semantics.',
+      zh: '请求时间线契约版本、重放标识、来源溯源和事件计时语义。',
     },
   },
   {
     source: '../db/src/queries/agentic-aggregates.ts',
-    sourceSha256: 'b0f4be6c39fe440df99db687b4b6deeed58897bf20325770631680f6e41aa304',
+    sourceSha256: 'fae8d19971730132cb30cd781f677562bfc6328b1f4e35a8268a8391ad187c18',
     reviewArea: {
       en: 'Agentic aggregate percentile keys, nullability, and ID-keyed response shape.',
       zh: '智能体汇总百分位字段、可空性和按 ID 索引的响应结构。',
@@ -436,7 +721,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/benchmark-siblings.ts',
-    sourceSha256: '5a2b7e902c3003a0b3dbf3f402cc4b6b40648a1e014a218ab39dcbfd35c80f42',
+    sourceSha256: '07d3d1bf93820091d1014b14bf954555edcec422c05e575ad87142f9845e4156',
     reviewArea: {
       en: 'Benchmark sibling SKU metadata and sibling navigation row shape.',
       zh: '基准同组 SKU 元数据和同组导航行结构。',
@@ -444,7 +729,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/benchmarks.ts',
-    sourceSha256: 'bb2f2cd28d8e4cea7b556561235da43d6a33363dc2c6d1f0b13408193d04298e',
+    sourceSha256: '486e34d55275170c7e0544af24c151199752eb5628191d38606b1ecec289dfdf',
     reviewArea: {
       en: 'Benchmark row fields and latest, exact-run, history, and TCO query semantics.',
       zh: '基准行字段以及最新、精确运行、历史和 TCO 查询语义。',
@@ -460,7 +745,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/datasets.ts',
-    sourceSha256: '05b0f990f8aeb1061fb0bfe7204f1c015fc55b0a541249188ffe740edf2a5af6',
+    sourceSha256: '34aba7d420ce651b6f04269a73466c3635d36a1a8c7eb63ae8e0f6b9e43b8685',
     reviewArea: {
       en: 'Dataset registry, detail, conversation index, pagination, and conversation structure responses.',
       zh: '数据集目录、详情、会话索引、分页和会话结构响应。',
@@ -492,7 +777,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/request-timeline.ts',
-    sourceSha256: '5373c839e1290d5ce1fde420041028325418597e0c911b212368f2f839756fa2',
+    sourceSha256: 'e27e2c421d94d55f8178756d31e9791d4143d33e4086f767e4ea15132a7ad708',
     reviewArea: {
       en: 'Request timeline metadata, request event records, units, and nullable timings.',
       zh: '请求时间线元数据、请求事件记录、单位和可空计时字段。',
@@ -500,10 +785,10 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/server-logs.ts',
-    sourceSha256: '44a4a55283fe4952e07df4034abf6a211d055225aee510ef3e6a945c72b63b55',
+    sourceSha256: '7a6f6b0b3c60e1713f73f6d543745931886654f060cf55d6dc4be9619a64b479',
     reviewArea: {
-      en: 'Server log lookup result fields and missing-record behavior.',
-      zh: '服务器日志查询结果字段和记录缺失行为。',
+      en: 'Log filename discovery, selected-file reads, bounded chunk metadata, complete-file search, availability, and legacy-schema fallback.',
+      zh: '日志文件名发现、指定文件读取、有界分块元数据、完整文件搜索、可用性以及旧版 schema 回退行为。',
     },
   },
   {
@@ -524,7 +809,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/trace-histograms.ts',
-    sourceSha256: '43d11e53abc4f9ba723c987424823a03fc8f5dde51cf11653ab8c0bf3a634f55',
+    sourceSha256: 'ec3b7358046ca2acdaaa56d8ecde2829ba8fe4f4b2b31c3795838ac85f59ca89',
     reviewArea: {
       en: 'Trace histogram input/output token arrays and ID-keyed response shape.',
       zh: '跟踪直方图输入/输出 token 数组和按 ID 索引的响应结构。',
@@ -532,7 +817,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/trace-server-metrics.ts',
-    sourceSha256: '06b5a9b92b90a5e695ab6c1dfb4b4d1240180f8d306ff4523ea8cc4269b0ddf3',
+    sourceSha256: 'da987d22521dc63da34dcacf3caaad6adb21aa6e5e7a65d1a4fa495a71e9f1f0',
     reviewArea: {
       en: 'Trace server metric metadata, time-series groups, source labels, and units.',
       zh: '跟踪服务器指标元数据、时间序列分组、来源标签和单位。',
@@ -540,7 +825,7 @@ export const apiContractSourceDigests = [
   },
   {
     source: '../db/src/queries/workflow-info.ts',
-    sourceSha256: 'b6611604d41ab69c00cb804ad255d4cbc9f70e41e84ad43330538558956f9eb6',
+    sourceSha256: '7e7d6fc965a47655fe9fa6feb6d8282eff58c2aedca1bcdb59ac1e8c91128d31',
     reviewArea: {
       en: 'Availability rows plus workflow runs, changelogs, configurations, and run coverage responses.',
       zh: '可用配置行以及工作流运行、变更记录、配置和运行覆盖响应。',
diff --git a/packages/app/src/lib/api.test.ts b/packages/app/src/lib/api.test.ts
index a6596f8e0..a51e21087 100644
--- a/packages/app/src/lib/api.test.ts
+++ b/packages/app/src/lib/api.test.ts
@@ -9,6 +9,8 @@ import {
   fetchCollectiveXRunList,
   fetchReliability,
   fetchEvaluations,
+  fetchFrameworkReleases,
+  fetchLatestImages,
 } from './api';
 import { buildRunSummary } from '@semianalysisai/inferencex-db/collectivex/reader';
 import { makeCollectiveXDataset } from '@/components/collectivex/test-fixture';
@@ -117,6 +119,36 @@ describe('fetchWorkflowInfo', () => {
       expect.objectContaining({}),
     );
   });
+
+  it('serializes the Agentic Traces workflow cache dimension', async () => {
+    mockOk({ runs: [], changelogs: [], configs: [], runConfigs: [] });
+    await fetchWorkflowInfo('2026-03-01', undefined, 'agentic_traces');
+    expect(mockFetch).toHaveBeenCalledWith(
+      '/api/v1/workflow-info?date=2026-03-01&benchmarkType=agentic_traces',
+      expect.objectContaining({}),
+    );
+  });
+});
+
+describe('simple read abort signals', () => {
+  it('forwards AbortSignal for latest images and framework releases', async () => {
+    const controller = new AbortController();
+    mockOk([]);
+    await fetchLatestImages(controller.signal);
+    mockOk({});
+    await fetchFrameworkReleases(controller.signal);
+
+    expect(mockFetch).toHaveBeenNthCalledWith(
+      1,
+      '/api/v1/latest-images',
+      expect.objectContaining({ signal: controller.signal }),
+    );
+    expect(mockFetch).toHaveBeenNthCalledWith(
+      2,
+      '/api/v1/framework-releases',
+      expect.objectContaining({ signal: controller.signal }),
+    );
+  });
 });
 
 describe('fetchAvailability', () => {
diff --git a/packages/app/src/lib/api.ts b/packages/app/src/lib/api.ts
index 0b2a03f71..d794648f1 100644
--- a/packages/app/src/lib/api.ts
+++ b/packages/app/src/lib/api.ts
@@ -9,54 +9,12 @@ import {
   type CollectiveXRunSummary,
   type CollectiveXVersion,
 } from '@/components/collectivex/types';
-import type { WorkerPower } from '@/components/inference/types';
+import type { BenchmarkRow } from '@semianalysisai/inferencex-db/queries/benchmarks';
+import type { EvalRow } from '@semianalysisai/inferencex-db/queries/evaluations';
 
 import type { SubmissionsResponse } from './submissions-types';
 
-export interface BenchmarkRow {
-  /** Stable per-point id from benchmark_results; used for agentic detail lookups. */
-  id: number;
-  hardware: string;
-  framework: string;
-  model: string;
-  precision: string;
-  spec_method: string;
-  disagg: boolean;
-  is_multinode: boolean;
-  prefill_tp: number;
-  prefill_ep: number;
-  prefill_dp_attention: boolean;
-  prefill_num_workers: number;
-  decode_tp: number;
-  decode_ep: number;
-  decode_dp_attention: boolean;
-  decode_num_workers: number;
-  num_prefill_gpu: number;
-  num_decode_gpu: number;
-  benchmark_type: string;
-  // Null for agentic_traces rows; numeric for single_turn fixed-seq rows.
-  isl: number | null;
-  osl: number | null;
-  conc: number;
-  /** KV-cache offload mode. Defaults to 'off' for fixed-sequence rows. */
-  offload_mode: string;
-  image: string | null;
-  metrics: Record<string, number>;
-  /**
-   * Per-worker measured power for multinode / disagg runs. The runner emits
-   * this as a JSONB sibling of the scalar metrics; the API layer surfaces it
-   * as a separate field here so the scalar `metrics` index signature can stay
-   * `Record<string, number>` and existing `m.x ?? 0` call sites keep narrowing
-   * cleanly. Undefined for single-node runs and any run predating
-   * aggregate_power.py.
-   */
-  workers?: WorkerPower[];
-  date: string;
-  /** Internal workflow identity used to keep merged agentic curves within one run. */
-  workflow_run_id?: number;
-  run_started_at?: string | null;
-  run_url: string | null;
-}
+export type { BenchmarkRow, EvalRow };
 
 export interface WorkflowRunRow {
   github_run_id: number;
@@ -76,6 +34,7 @@ export interface ChangelogRow {
   config_keys: string[];
   description: string;
   pr_link: string | null;
+  append_only?: boolean;
 }
 
 export interface DateConfigRow {
@@ -121,34 +80,6 @@ export interface ReliabilityRow {
   total: number;
 }
 
-export interface EvalRow {
-  id: number;
-  config_id: number;
-  hardware: string;
-  framework: string;
-  model: string;
-  precision: string;
-  spec_method: string;
-  disagg: boolean;
-  is_multinode: boolean;
-  prefill_tp: number;
-  prefill_ep: number;
-  prefill_dp_attention: boolean;
-  prefill_num_workers: number;
-  decode_tp: number;
-  decode_ep: number;
-  decode_dp_attention: boolean;
-  decode_num_workers: number;
-  num_prefill_gpu: number;
-  num_decode_gpu: number;
-  task: string;
-  date: string;
-  conc: number | null;
-  metrics: Record<string, number>;
-  timestamp: string;
-  run_url: string | null;
-}
-
 async function fetchJson<T>(url: string, signal?: AbortSignal): Promise<T> {
   const res = await fetch(url, { signal });
   if (!res.ok) throw new Error(`API error: ${res.status} ${res.statusText}`);
@@ -183,9 +114,20 @@ export function fetchBenchmarkHistory(
   osl: number,
   signal?: AbortSignal,
   benchmarkType?: 'agentic_traces',
+  /**
+   * Trims the response to the calculator's metric allowlist — a ~24% smaller
+   * payload. Callers that plot measured power must leave this off; those
+   * metrics are not in the allowlist.
+   */
+  view?: 'calculator',
 ) {
-  const params = new URLSearchParams({ model, isl: String(isl), osl: String(osl) });
+  const params = new URLSearchParams({
+    model,
+    isl: String(isl),
+    osl: String(osl),
+  });
   if (benchmarkType) params.set('benchmarkType', benchmarkType);
+  if (view) params.set('view', view);
   return fetchJson<BenchmarkRow[]>(`/api/v1/benchmarks/history?${params}`, signal);
 }
 
@@ -390,12 +332,12 @@ export interface LatestImageRow {
   date: string;
 }
 
-export function fetchLatestImages() {
-  return fetchJson<LatestImageRow[]>('/api/v1/latest-images');
+export function fetchLatestImages(signal?: AbortSignal) {
+  return fetchJson<LatestImageRow[]>('/api/v1/latest-images', signal);
 }
 
 export type FrameworkReleases = Record<string, string | null>;
 
-export function fetchFrameworkReleases() {
-  return fetchJson<FrameworkReleases>('/api/v1/framework-releases');
+export function fetchFrameworkReleases(signal?: AbortSignal) {
+  return fetchJson<FrameworkReleases>('/api/v1/framework-releases', signal);
 }
diff --git a/packages/app/src/lib/bearer-auth.test.ts b/packages/app/src/lib/bearer-auth.test.ts
new file mode 100644
index 000000000..b37eb76f1
--- /dev/null
+++ b/packages/app/src/lib/bearer-auth.test.ts
@@ -0,0 +1,15 @@
+import { describe, expect, it } from 'vitest';
+
+import { bearerMatches } from './bearer-auth';
+
+describe('bearerMatches', () => {
+  it('matches only the complete Bearer header', () => {
+    expect(bearerMatches('Bearer secret', 'secret')).toBe(true);
+    expect(bearerMatches('secret', 'secret')).toBe(false);
+    expect(bearerMatches('Bearer other', 'secret')).toBe(false);
+  });
+
+  it('returns false instead of throwing for equal-length strings with different byte lengths', () => {
+    expect(bearerMatches('Bearer é', 'xx')).toBe(false);
+  });
+});
diff --git a/packages/app/src/lib/bearer-auth.ts b/packages/app/src/lib/bearer-auth.ts
new file mode 100644
index 000000000..7bd540fbd
--- /dev/null
+++ b/packages/app/src/lib/bearer-auth.ts
@@ -0,0 +1,8 @@
+import { timingSafeEqual } from 'node:crypto';
+
+/** Compare a complete Authorization header with a Bearer secret using its encoded bytes. */
+export function bearerMatches(header: string, secret: string): boolean {
+  const provided = Buffer.from(header);
+  const expected = Buffer.from(`Bearer ${secret}`);
+  return provided.length === expected.length && timingSafeEqual(provided, expected);
+}
diff --git a/packages/app/src/lib/benchmark-api-view.test.ts b/packages/app/src/lib/benchmark-api-view.test.ts
index 168435f2f..291ee0f7c 100644
--- a/packages/app/src/lib/benchmark-api-view.test.ts
+++ b/packages/app/src/lib/benchmark-api-view.test.ts
@@ -46,6 +46,35 @@ describe('toCalculatorBenchmarkRows', () => {
     ]);
   });
 
+  it('keeps all three cache tiers — the trim cannot know which one a row will use', () => {
+    // `cacheHitRateOf` picks between external and CPU per row, so the allowlist
+    // has to pass all three through or the choice is made for it by the trim.
+    // This runs on every calculator response, agentic included.
+    const cached = toCalculatorBenchmarkRows(
+      [
+        {
+          benchmark_type: 'agentic_traces',
+          isl: null,
+          osl: null,
+          metrics: {
+            tput_per_gpu: 100,
+            server_gpu_cache_hit_rate: 0.77,
+            server_external_cache_hit_rate: 0.06,
+            server_cpu_cache_hit_rate: 0.055,
+            theoretical_cache_hit_rate: 0.95,
+          },
+        },
+      ],
+      'agentic-traces',
+    );
+    expect(cached[0].metrics).toEqual({
+      tput_per_gpu: 100,
+      server_gpu_cache_hit_rate: 0.77,
+      server_external_cache_hit_rate: 0.06,
+      server_cpu_cache_hit_rate: 0.055,
+    });
+  });
+
   it('keeps the agentic percentile inputs used for interpolation', () => {
     expect(toCalculatorBenchmarkRows(rows, 'agentic-traces')).toEqual([
       {
diff --git a/packages/app/src/lib/benchmark-api-view.ts b/packages/app/src/lib/benchmark-api-view.ts
index 69859541f..3d2c9b23f 100644
--- a/packages/app/src/lib/benchmark-api-view.ts
+++ b/packages/app/src/lib/benchmark-api-view.ts
@@ -6,6 +6,15 @@ const CALCULATOR_METRIC_KEYS = new Set([
   'output_tput_per_gpu',
   'prefill_pp',
   'decode_pp',
+  // Cached-input billing in the Fleet Lifecycle section reads these. Only
+  // agentic rows carry them, but the trim runs on every calculator response, so
+  // omitting them here would strip the metric before interpolation ever sees it.
+  'server_gpu_cache_hit_rate',
+  'server_external_cache_hit_rate',
+  // Read only when no external rate is reported, but the trim cannot know that
+  // per row — so it has to survive the allowlist either way. See
+  // `cacheHitRateOf` for why the two are not simply summed.
+  'server_cpu_cache_hit_rate',
   ...['median', 'p75', 'p90'].flatMap((percentile) =>
     ['intvty', 'itl', 'full_response_itl', 'e2el', 'ttlt'].map(
       (metric) => `${percentile}_${metric}`,
diff --git a/packages/app/src/lib/benchmark-run-selection.ts b/packages/app/src/lib/benchmark-run-selection.ts
index f7332b96c..2a72d0598 100644
--- a/packages/app/src/lib/benchmark-run-selection.ts
+++ b/packages/app/src/lib/benchmark-run-selection.ts
@@ -10,27 +10,39 @@ export interface BenchmarkSeriesRow {
   date: string;
   workflow_run_id?: number;
   run_started_at?: string | null;
+  curve_date?: string;
+  curve_workflow_run_id?: number;
+  curve_run_started_at?: string | null;
 }
 
+export const benchmarkCurveDate = (row: BenchmarkSeriesRow): string => row.curve_date ?? row.date;
+
+export const benchmarkCurveWorkflowRunId = (row: BenchmarkSeriesRow): number | undefined =>
+  row.curve_workflow_run_id ?? row.workflow_run_id;
+
+export const benchmarkCurveRunStartedAt = (row: BenchmarkSeriesRow): string | null | undefined =>
+  row.curve_run_started_at ?? row.run_started_at;
+
 const seriesKey = (row: BenchmarkSeriesRow): string => {
   const specMethod = row.benchmark_type === 'agentic_traces' ? '' : row.spec_method;
   return `${row.hardware}|${row.framework}|${specMethod}|${row.disagg}|${row.precision}|${row.offload_mode ?? 'off'}`;
 };
 
 function isLaterRun(candidate: BenchmarkSeriesRow, current: BenchmarkSeriesRow): boolean {
-  const startedAt = candidate.run_started_at ?? '';
-  const currentStartedAt = current.run_started_at ?? '';
+  const startedAt = benchmarkCurveRunStartedAt(candidate) ?? '';
+  const currentStartedAt = benchmarkCurveRunStartedAt(current) ?? '';
   return (
     startedAt > currentStartedAt ||
     (startedAt === currentStartedAt &&
-      (candidate.workflow_run_id ?? Number.NEGATIVE_INFINITY) >
-        (current.workflow_run_id ?? Number.NEGATIVE_INFINITY))
+      (benchmarkCurveWorkflowRunId(candidate) ?? Number.NEGATIVE_INFINITY) >
+        (benchmarkCurveWorkflowRunId(current) ?? Number.NEGATIVE_INFINITY))
   );
 }
 
 function isWinningRun(row: BenchmarkSeriesRow, winner: BenchmarkSeriesRow): boolean {
   return (
-    row.run_started_at === winner.run_started_at && row.workflow_run_id === winner.workflow_run_id
+    benchmarkCurveRunStartedAt(row) === benchmarkCurveRunStartedAt(winner) &&
+    benchmarkCurveWorkflowRunId(row) === benchmarkCurveWorkflowRunId(winner)
   );
 }
 
@@ -40,12 +52,12 @@ export function dedupeRowsToLatestPerConfig<T extends BenchmarkSeriesRow>(rows:
   for (const row of rows) {
     const key = seriesKey(row);
     const current = winnerPerGroup.get(key);
-    if (!current || row.date > current.date) {
+    if (!current || benchmarkCurveDate(row) > benchmarkCurveDate(current)) {
       winnerPerGroup.set(key, row);
       continue;
     }
     if (
-      row.date === current.date &&
+      benchmarkCurveDate(row) === benchmarkCurveDate(current) &&
       row.benchmark_type === 'agentic_traces' &&
       isLaterRun(row, current)
     ) {
@@ -54,7 +66,7 @@ export function dedupeRowsToLatestPerConfig<T extends BenchmarkSeriesRow>(rows:
   }
   return rows.filter((row) => {
     const winner = winnerPerGroup.get(seriesKey(row));
-    if (!winner || row.date !== winner.date) return false;
+    if (!winner || benchmarkCurveDate(row) !== benchmarkCurveDate(winner)) return false;
     return row.benchmark_type !== 'agentic_traces' || isWinningRun(row, winner);
   });
 }
@@ -64,13 +76,13 @@ export function dedupeAgenticHistoryRuns<T extends BenchmarkSeriesRow>(rows: T[]
   const winnerPerDateAndSeries = new Map<string, T>();
   for (const row of rows) {
     if (row.benchmark_type !== 'agentic_traces') continue;
-    const key = `${row.date}|${seriesKey(row)}`;
+    const key = `${benchmarkCurveDate(row)}|${seriesKey(row)}`;
     const current = winnerPerDateAndSeries.get(key);
     if (!current || isLaterRun(row, current)) winnerPerDateAndSeries.set(key, row);
   }
   return rows.filter((row) => {
     if (row.benchmark_type !== 'agentic_traces') return true;
-    const winner = winnerPerDateAndSeries.get(`${row.date}|${seriesKey(row)}`);
+    const winner = winnerPerDateAndSeries.get(`${benchmarkCurveDate(row)}|${seriesKey(row)}`);
     return winner !== undefined && isWinningRun(row, winner);
   });
 }
diff --git a/packages/app/src/lib/benchmark-transform.test.ts b/packages/app/src/lib/benchmark-transform.test.ts
index 35f1eb7ef..fac8734d8 100644
--- a/packages/app/src/lib/benchmark-transform.test.ts
+++ b/packages/app/src/lib/benchmark-transform.test.ts
@@ -68,6 +68,40 @@ function makeRow(overrides: Partial<BenchmarkRow> = {}): BenchmarkRow {
 }
 
 describe('rowToAggDataEntry', () => {
+  it('preserves DCP and PCP metrics for point tooltips', () => {
+    const entry = rowToAggDataEntry(
+      makeRow({
+        metrics: {
+          prefill_dcp_size: 2,
+          decode_dcp_size: 8,
+          prefill_pcp_size: 4,
+          decode_pcp_size: 1,
+        },
+      }),
+    );
+
+    expect(entry.prefill_dcp_size).toBe(2);
+    expect(entry.decode_dcp_size).toBe(8);
+    expect(entry.prefill_pcp_size).toBe(4);
+    expect(entry.decode_pcp_size).toBe(1);
+  });
+
+  it('mirrors aggregate artifact DCP and PCP widths into both topology roles', () => {
+    const entry = rowToAggDataEntry(
+      makeRow({
+        metrics: {
+          dcp_size: 8,
+          pcp_size: 1,
+        },
+      }),
+    );
+
+    expect(entry.prefill_dcp_size).toBe(8);
+    expect(entry.decode_dcp_size).toBe(8);
+    expect(entry.prefill_pcp_size).toBe(1);
+    expect(entry.decode_pcp_size).toBe(1);
+  });
+
   it('maps hardware and framework fields', () => {
     const entry = rowToAggDataEntry(makeRow());
     expect(entry.hw).toBe('h200');
@@ -197,6 +231,18 @@ describe('rowToAggDataEntry', () => {
     expect(entryNull.image).toBeUndefined();
   });
 
+  it('passes recipe fingerprint through to chart point identity', () => {
+    const entry = rowToAggDataEntry(
+      makeRow({
+        recipe_fingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+      }),
+    );
+
+    expect(entry.recipe_fingerprint).toBe(
+      'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+    );
+  });
+
   it('passes runtime cache metadata through to chart points', () => {
     const entry = rowToAggDataEntry(
       makeRow({
@@ -222,13 +268,31 @@ describe('rowToAggDataEntry', () => {
   it('passes through measured power telemetry fields when present', () => {
     const entry = rowToAggDataEntry(
       makeRow({
-        metrics: { tput_per_gpu: 100, avg_power_w: 685.5, joules_per_output_token: 8.4 },
+        metrics: {
+          tput_per_gpu: 100,
+          avg_power_w: 685.5,
+          joules_per_output_token: 8.4,
+        },
       }),
     );
     expect(entry.avg_power_w).toBe(685.5);
     expect(entry.joules_per_output_token).toBe(8.4);
   });
 
+  it('passes through versioned whole-deployment joules per successful query', () => {
+    const entry = rowToAggDataEntry(
+      makeRow({
+        metrics: {
+          power_valid: 1,
+          power_metric_schema_version: 2,
+          joules_per_successful_query: 1542.75,
+        },
+      }),
+    );
+
+    expect(entry.joules_per_successful_query).toBe(1542.75);
+  });
+
   it('leaves measured power fields undefined for rows that predate the metric', () => {
     // Distinguishing "no measurement" from "0 W" matters: createChartDataPoint
     // uses typeof===number to decide whether to emit the measuredAvgPower field.
@@ -240,13 +304,18 @@ describe('rowToAggDataEntry', () => {
   it('passes through multinode / disagg role-split power scalars when present', () => {
     const entry = rowToAggDataEntry(
       makeRow({
+        disagg: true,
         metrics: {
           tput_per_gpu: 100,
+          power_valid: 1,
+          power_metric_schema_version: 2,
           prefill_avg_power_w: 612.3,
           decode_avg_power_w: 701.5,
           joules_per_input_token: 1.2,
-          // disagg: joules_per_output_token IS the per-stage decode value.
           joules_per_output_token: 9.7,
+          joules_per_total_token: 0.8,
+          prefill_joules_per_input_token: 0.4,
+          decode_joules_per_output_token: 5.1,
         },
       }),
     );
@@ -254,13 +323,157 @@ describe('rowToAggDataEntry', () => {
     expect(entry.decode_avg_power_w).toBe(701.5);
     expect(entry.joules_per_input_token).toBe(1.2);
     expect(entry.joules_per_output_token).toBe(9.7);
+    expect(entry.joules_per_total_token).toBe(0.8);
+    expect(entry.prefill_joules_per_input_token).toBe(0.4);
+    expect(entry.decode_joules_per_output_token).toBe(5.1);
+  });
+
+  it('withholds ambiguous unversioned disagg joules while preserving role watts', () => {
+    const entry = rowToAggDataEntry(
+      makeRow({
+        disagg: true,
+        metrics: {
+          avg_power_w: 650,
+          joules_per_successful_query: 1542.75,
+          prefill_avg_power_w: 612.3,
+          decode_avg_power_w: 701.5,
+          joules_per_input_token: 1.2,
+          joules_per_output_token: 9.7,
+          joules_per_total_token: 0.8,
+        },
+      }),
+    );
+
+    expect(entry.avg_power_w).toBe(650);
+    expect(entry.prefill_avg_power_w).toBe(612.3);
+    expect(entry.decode_avg_power_w).toBe(701.5);
+    expect(entry.joules_per_input_token).toBeUndefined();
+    expect(entry.joules_per_successful_query).toBeUndefined();
+    expect(entry.joules_per_output_token).toBeUndefined();
+    expect(entry.joules_per_total_token).toBeUndefined();
+  });
+
+  it('withholds disagg joules stamped with an unrecognized future schema version', () => {
+    const entry = rowToAggDataEntry(
+      makeRow({
+        disagg: true,
+        metrics: {
+          power_valid: 1,
+          power_metric_schema_version: 3,
+          avg_power_w: 650,
+          prefill_avg_power_w: 612.3,
+          joules_per_input_token: 1.2,
+          joules_per_successful_query: 42.5,
+          joules_per_output_token: 9.7,
+          joules_per_total_token: 0.8,
+        },
+      }),
+    );
+
+    expect(entry.avg_power_w).toBe(650);
+    expect(entry.prefill_avg_power_w).toBe(612.3);
+    expect(entry.joules_per_input_token).toBeUndefined();
+    expect(entry.joules_per_successful_query).toBeUndefined();
+    expect(entry.joules_per_output_token).toBeUndefined();
+    expect(entry.joules_per_total_token).toBeUndefined();
+  });
+
+  it('treats explicit power_valid=0 as authoritative and scrubs all measured telemetry', () => {
+    const workers = [{ role: 'agg', worker_idx: 0, num_gpus: 8, avg_power_w: 685.5 }];
+    const row = makeRow({
+      metrics: {
+        power_valid: 0,
+        power_metric_schema_version: 2,
+        avg_power_w: 685.5,
+        joules_per_successful_query: 1542.75,
+        joules_per_input_token: 1.2,
+        joules_per_output_token: 8.4,
+        joules_per_total_token: 0.8,
+        prefill_avg_power_w: 612.3,
+        decode_avg_power_w: 701.5,
+        prefill_joules_per_input_token: 0.4,
+        decode_joules_per_output_token: 5.1,
+        avg_temp_c: 68.4,
+        peak_temp_c: 79.2,
+        avg_util_pct: 88.5,
+        avg_mem_used_mb: 71234.5,
+      },
+      workers,
+    });
+    const entry = rowToAggDataEntry(row);
+
+    expect(entry.avg_power_w).toBeUndefined();
+    expect(entry.joules_per_successful_query).toBeUndefined();
+    expect(entry.joules_per_input_token).toBeUndefined();
+    expect(entry.joules_per_output_token).toBeUndefined();
+    expect(entry.joules_per_total_token).toBeUndefined();
+    expect(entry.prefill_avg_power_w).toBeUndefined();
+    expect(entry.decode_avg_power_w).toBeUndefined();
+    expect(entry.prefill_joules_per_input_token).toBeUndefined();
+    expect(entry.decode_joules_per_output_token).toBeUndefined();
+    expect(entry.avg_temp_c).toBeUndefined();
+    expect(entry.peak_temp_c).toBeUndefined();
+    expect(entry.avg_util_pct).toBeUndefined();
+    expect(entry.avg_mem_used_mb).toBeUndefined();
+    expect(entry.workers).toBeUndefined();
+
+    const { chartData } = transformBenchmarkRows([row]);
+    const point = chartData.find((data) => data.length > 0)![0];
+    expect(point.measuredJPerSuccessfulQuery).toBeUndefined();
+    expect(point.measuredWhPerSuccessfulQuery).toBeUndefined();
+    expect(point.measuredPowerPercentTdp).toBeUndefined();
+  });
+
+  it('keeps measured telemetry compatible when a legacy row omits power_valid', () => {
+    const workers = [{ role: 'agg', worker_idx: 0, num_gpus: 8, avg_power_w: 560 }];
+    const row = makeRow({
+      metrics: {
+        avg_power_w: 560,
+        joules_per_successful_query: 1800,
+        avg_temp_c: 68.4,
+        peak_temp_c: 79.2,
+        avg_util_pct: 88.5,
+        avg_mem_used_mb: 71234.5,
+      },
+      workers,
+    });
+    const entry = rowToAggDataEntry(row);
+
+    expect(entry.avg_power_w).toBe(560);
+    expect(entry.joules_per_successful_query).toBe(1800);
+    expect(entry.avg_temp_c).toBe(68.4);
+    expect(entry.peak_temp_c).toBe(79.2);
+    expect(entry.avg_util_pct).toBe(88.5);
+    expect(entry.avg_mem_used_mb).toBe(71234.5);
+    expect(entry.workers).toEqual(workers);
+
+    const { chartData } = transformBenchmarkRows([row]);
+    const point = chartData.find((data) => data.length > 0)![0];
+    expect(point.measuredJPerSuccessfulQuery?.y).toBe(1800);
+    expect(point.measuredWhPerSuccessfulQuery?.y).toBe(0.5);
+    expect(point.measuredPowerPercentTdp?.y).toBe(80);
   });
 
   it('passes through per-worker measured power array intact', () => {
     const workers = [
-      { role: 'prefill' as const, worker_idx: 0, num_gpus: 4, avg_power_w: 588.4 },
-      { role: 'prefill' as const, worker_idx: 1, num_gpus: 4, avg_power_w: 601.2 },
-      { role: 'decode' as const, worker_idx: 0, num_gpus: 8, avg_power_w: 712.1 },
+      {
+        role: 'prefill' as const,
+        worker_idx: 0,
+        num_gpus: 4,
+        avg_power_w: 588.4,
+      },
+      {
+        role: 'prefill' as const,
+        worker_idx: 1,
+        num_gpus: 4,
+        avg_power_w: 601.2,
+      },
+      {
+        role: 'decode' as const,
+        worker_idx: 0,
+        num_gpus: 8,
+        avg_power_w: 712.1,
+      },
       { role: 'frontend' as const, worker_idx: 0, num_gpus: 0, avg_power_w: 0 },
     ];
     const entry = rowToAggDataEntry(makeRow({ workers }));
@@ -429,7 +642,12 @@ describe('transformBenchmarkRows', () => {
 
   it('labels M3 mtp configs with the "M3 EAGLE" suffix', () => {
     const rows = [
-      makeRow({ model: 'minimaxm3', hardware: 'h100', framework: 'vllm', spec_method: 'mtp' }),
+      makeRow({
+        model: 'minimaxm3',
+        hardware: 'h100',
+        framework: 'vllm',
+        spec_method: 'mtp',
+      }),
     ];
     const { hardwareConfig } = transformBenchmarkRows(rows);
     const entry = hardwareConfig['h100_vllm_mtp'];
@@ -439,7 +657,12 @@ describe('transformBenchmarkRows', () => {
 
   it('keeps the generic MTP suffix for non-M3 mtp configs', () => {
     const rows = [
-      makeRow({ model: 'dsr1', hardware: 'h200', framework: 'sglang', spec_method: 'mtp' }),
+      makeRow({
+        model: 'dsr1',
+        hardware: 'h200',
+        framework: 'sglang',
+        spec_method: 'mtp',
+      }),
     ];
     const { hardwareConfig } = transformBenchmarkRows(rows);
     const entry = hardwareConfig['h200_sglang_mtp'];
@@ -1032,9 +1255,21 @@ describe('transformBenchmarkRows — dp_attention narrowing', () => {
 
 describe('mergeRunScopedRows', () => {
   const vllmRun = (over: Partial<BenchmarkRow> = {}) =>
-    makeRow({ model: 'dsv4', hardware: 'b300', framework: 'vllm', precision: 'fp4', ...over });
+    makeRow({
+      model: 'dsv4',
+      hardware: 'b300',
+      framework: 'vllm',
+      precision: 'fp4',
+      ...over,
+    });
   const sglangBase = (over: Partial<BenchmarkRow> = {}) =>
-    makeRow({ model: 'dsv4', hardware: 'b300', framework: 'sglang', precision: 'fp4', ...over });
+    makeRow({
+      model: 'dsv4',
+      hardware: 'b300',
+      framework: 'sglang',
+      precision: 'fp4',
+      ...over,
+    });
 
   it('pins configs the run covers to the run rows, replacing base rows', () => {
     const runRows = [vllmRun({ id: 10, conc: 32 }), vllmRun({ id: 11, conc: 64 })];
@@ -1068,6 +1303,39 @@ describe('mergeRunScopedRows', () => {
     expect(merged.map((r) => r.id).toSorted((a, b) => a - b)).toEqual([10, 90, 91, 92]);
   });
 
+  it('carries forward sibling topologies when an append run touches only one', () => {
+    const runRows = [vllmRun({ id: 10, decode_tp: 8, conc: 128 })];
+    const baseRows = [
+      vllmRun({ id: 90, decode_tp: 8, conc: 64 }),
+      vllmRun({ id: 91, decode_tp: 4, conc: 32 }),
+    ];
+
+    const merged = mergeRunScopedRows(runRows, baseRows);
+
+    expect(merged.map((r) => r.id).toSorted((a, b) => a - b)).toEqual([10, 91]);
+  });
+
+  it('carries forward sibling recipe variants at the same topology and concurrency', () => {
+    const runRows = [
+      vllmRun({
+        id: 10,
+        conc: 64,
+        recipe_fingerprint: 'bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb',
+      }),
+    ];
+    const baseRows = [
+      vllmRun({
+        id: 90,
+        conc: 64,
+        recipe_fingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+      }),
+    ];
+
+    const merged = mergeRunScopedRows(runRows, baseRows);
+
+    expect(merged.map((r) => r.id).toSorted((a, b) => a - b)).toEqual([10, 90]);
+  });
+
   it('scopes per benchmark_type — an agentic run does not hide fixed-seq carry-forward', () => {
     const runRows = [vllmRun({ id: 10, benchmark_type: 'agentic_traces' })];
     const baseRows = [
@@ -1090,7 +1358,14 @@ describe('rowToAggDataEntry — agentic interactivity invariant', () => {
   // interactivity selector is slow-tail, so we always derive intvty = 1/itl and
   // discard the artifact value. Mirrors the ingest mapper + backfill.
   const agentic = (metrics: Record<string, number>) =>
-    rowToAggDataEntry(makeRow({ benchmark_type: 'agentic_traces', isl: null, osl: null, metrics }));
+    rowToAggDataEntry(
+      makeRow({
+        benchmark_type: 'agentic_traces',
+        isl: null,
+        osl: null,
+        metrics,
+      }),
+    );
 
   it('overrides an artifact-supplied (fast-tail) *_intvty with 1/*_itl', () => {
     const entry = agentic({
diff --git a/packages/app/src/lib/benchmark-transform.ts b/packages/app/src/lib/benchmark-transform.ts
index a3b6cbfdc..4036af80c 100644
--- a/packages/app/src/lib/benchmark-transform.ts
+++ b/packages/app/src/lib/benchmark-transform.ts
@@ -4,18 +4,31 @@
 
 import { DB_MODEL_TO_DISPLAY } from '@semianalysisai/inferencex-constants';
 
-import chartDefinitions from '@/components/inference/inference-chart-config.json';
+import chartDefinitions from '@/components/inference/metric-registry';
 import type {
   AggDataEntry,
   ChartDefinition,
   HardwareConfig,
   InferenceData,
 } from '@/components/inference/types';
-import { createChartDataPoint, getHardwareKey } from '@/lib/chart-utils';
+import {
+  buildDerivedChartFields,
+  createChartDataPoint,
+  getHardwareKey,
+  type DerivedChartFields,
+} from '@/lib/chart-utils';
 import { getHardwareConfig } from '@/lib/constants';
 import { isPersistedBenchmarkId } from '@/lib/benchmark-id';
 import type { BenchmarkRow } from '@/lib/api';
 
+/**
+ * Producer schema version whose unprefixed `joules_per_*` fields are documented
+ * as whole-deployment energy. Mirrors `POWER_METRIC_SCHEMA_VERSION` in the
+ * runner's `utils/aggregate_power.py`; bump both together when the semantics of
+ * those fields change again.
+ */
+const WHOLE_DEPLOYMENT_ENERGY_SCHEMA_VERSION = 2;
+
 /**
  * Agentic trace-replay runs (`benchmark_type === 'agentic_traces'`) emit ttft/ttlt/itl
  * but not the intvty/e2el/tpot keys the chart pipeline expects. Bridge them here:
@@ -95,6 +108,16 @@ export function rowToAggDataEntry(row: BenchmarkRow): AggDataEntry {
   const aggregateDpAttention = aggregateUsesPrefill
     ? row.prefill_dp_attention
     : row.decode_dp_attention;
+  // An explicit invalid verdict is authoritative. Legacy rows without the
+  // verdict remain eligible so historical single-node measurements do not
+  // disappear. Unversioned disaggregated joules are withheld because their
+  // unprefixed fields changed from role-local to whole-deployment semantics.
+  const measuredPowerValid = m.power_valid !== 0;
+  // Match the version exactly rather than `>= 2`: an open bound would silently
+  // admit a future schema whose semantics changed again, which is the exact
+  // failure that made versioning necessary in the first place.
+  const hasWholeDeploymentEnergySemantics =
+    !row.disagg || m.power_metric_schema_version === WHOLE_DEPLOYMENT_ENERGY_SCHEMA_VERSION;
   // Prefer the dedicated column (added in migration 004); fall back to the
   // legacy stash inside `metrics` for any rows ingested before that column
   // existed.
@@ -115,6 +138,7 @@ export function rowToAggDataEntry(row: BenchmarkRow): AggDataEntry {
   return {
     rawMetricKeys: Object.keys(m),
     id: isPersistedBenchmarkId(numericId) ? numericId : undefined,
+    recipe_fingerprint: row.recipe_fingerprint ?? undefined,
     hw: row.hardware,
     framework: row.framework,
     model: DB_MODEL_TO_DISPLAY[row.model] ?? row.model,
@@ -168,26 +192,46 @@ export function rowToAggDataEntry(row: BenchmarkRow): AggDataEntry {
     // Measured GPU telemetry (runner's aggregate_power.py). Left undefined for
     // rows predating the field so downstream chart code can distinguish
     // "no measurement" from "0 W" via createChartDataPoint's typeof guard.
-    avg_power_w: m.avg_power_w,
-    joules_per_output_token: m.joules_per_output_token,
-    joules_per_total_token: m.joules_per_total_token,
-    // Multinode / disagg-only role splits — same undefined-for-legacy pattern.
-    // (disagg's decode-only J/output is carried by joules_per_output_token above,
-    // which the runner overrides to the per-stage value — no separate _decode key.)
-    prefill_avg_power_w: m.prefill_avg_power_w,
-    decode_avg_power_w: m.decode_avg_power_w,
-    joules_per_input_token: m.joules_per_input_token,
+    power_valid: m.power_valid,
+    power_metric_schema_version: m.power_metric_schema_version,
+    avg_power_w: measuredPowerValid ? m.avg_power_w : undefined,
+    joules_per_successful_query:
+      measuredPowerValid && hasWholeDeploymentEnergySemantics
+        ? m.joules_per_successful_query
+        : undefined,
+    joules_per_output_token:
+      measuredPowerValid && hasWholeDeploymentEnergySemantics
+        ? m.joules_per_output_token
+        : undefined,
+    joules_per_total_token:
+      measuredPowerValid && hasWholeDeploymentEnergySemantics
+        ? m.joules_per_total_token
+        : undefined,
+    // Role power remains unambiguous across schema versions. Version 2 also
+    // publishes explicit role energy alongside whole-deployment joules.
+    prefill_avg_power_w: measuredPowerValid ? m.prefill_avg_power_w : undefined,
+    decode_avg_power_w: measuredPowerValid ? m.decode_avg_power_w : undefined,
+    joules_per_input_token:
+      measuredPowerValid && hasWholeDeploymentEnergySemantics
+        ? m.joules_per_input_token
+        : undefined,
+    prefill_joules_per_input_token: measuredPowerValid
+      ? m.prefill_joules_per_input_token
+      : undefined,
+    decode_joules_per_output_token: measuredPowerValid
+      ? m.decode_joules_per_output_token
+      : undefined,
     // Cluster-wide GPU telemetry beyond power. Emitted when the perfmon CSVs
     // include the corresponding sample columns; left undefined otherwise so
     // the chart layer can distinguish "no measurement" from a real zero.
-    avg_temp_c: m.avg_temp_c,
-    peak_temp_c: m.peak_temp_c,
-    avg_util_pct: m.avg_util_pct,
-    avg_mem_used_mb: m.avg_mem_used_mb,
+    avg_temp_c: measuredPowerValid ? m.avg_temp_c : undefined,
+    peak_temp_c: measuredPowerValid ? m.peak_temp_c : undefined,
+    avg_util_pct: measuredPowerValid ? m.avg_util_pct : undefined,
+    avg_mem_used_mb: measuredPowerValid ? m.avg_mem_used_mb : undefined,
     // Per-worker measured power. Surfaced on BenchmarkRow as a sibling of the
     // scalar `metrics` dict (see api.ts). Narrow defensively so a malformed
     // payload can't poison downstream consumers.
-    workers: Array.isArray(row.workers) ? row.workers : undefined,
+    workers: measuredPowerValid && Array.isArray(row.workers) ? row.workers : undefined,
     disagg: row.disagg,
     num_prefill_gpu: row.num_prefill_gpu,
     num_decode_gpu: row.num_decode_gpu,
@@ -208,6 +252,14 @@ export function rowToAggDataEntry(row: BenchmarkRow): AggDataEntry {
     decode_tp: row.disagg ? row.decode_tp : aggregateTp,
     decode_ep: row.disagg ? row.decode_ep : aggregateEp,
     decode_pp: row.disagg ? m.decode_pp : aggregatePp,
+    // Context-parallel widths are emitted by the runtime into metrics JSONB.
+    // Preserve both role-shaped values even for aggregate deployments: the
+    // tooltip collapses the transport-only role names for aggregate serving,
+    // while disaggregated serving displays the values per role.
+    prefill_dcp_size: m.prefill_dcp_size ?? m.dcp_size,
+    decode_dcp_size: m.decode_dcp_size ?? m.dcp_size,
+    prefill_pcp_size: m.prefill_pcp_size ?? m.pcp_size,
+    decode_pcp_size: m.decode_pcp_size ?? m.pcp_size,
     decode_dp_attention: row.disagg ? row.decode_dp_attention : aggregateDpAttention,
     decode_num_workers: row.decode_num_workers,
     image: row.image ?? undefined,
@@ -238,6 +290,7 @@ interface PreparedEntry {
   entry: AggDataEntry;
   hwKey: string;
   date: string;
+  derivedFields: DerivedChartFields;
 }
 
 /**
@@ -249,14 +302,32 @@ export function withPercentile(key: string, percentile: string): string {
   return key.replace(/^(?:mean|median|p75|p90|p95|p99|p99\.9)_/u, `${percentile}_`);
 }
 
-// Replacement granularity for single-run scoping: the changelog config_key
-// tuple (model-precision-hardware-framework) plus benchmark_type AND offload_mode.
-// benchmark_type keeps an agentic-only run from hiding the same config's
-// fixed-seq carry-forward; offload_mode keeps a run that produced only one
-// offload variant (e.g. offload=on) from claiming — and thereby suppressing —
-// the other variant's (offload=off) base rows, which are a distinct series.
+// Replacement granularity for single-run scoping is an exact generated topology.
+// An append-only run may touch one TP/EP search-space row while the displayed
+// curve also contains sibling topologies from the preceding snapshot.
 const runScopeKey = (r: BenchmarkRow): string =>
-  `${r.model}|${r.precision}|${r.hardware}|${r.framework}|${r.benchmark_type}|${r.offload_mode ?? 'off'}`;
+  JSON.stringify([
+    r.model,
+    r.precision,
+    r.hardware,
+    r.framework,
+    r.spec_method,
+    r.disagg,
+    r.is_multinode,
+    r.prefill_tp,
+    r.prefill_ep,
+    r.prefill_dp_attention,
+    r.prefill_num_workers,
+    r.decode_tp,
+    r.decode_ep,
+    r.decode_dp_attention,
+    r.decode_num_workers,
+    r.benchmark_type,
+    r.isl,
+    r.osl,
+    r.offload_mode ?? 'off',
+    r.recipe_fingerprint ?? null,
+  ]);
 
 /**
  * Merge run-scoped benchmark rows with the normal latest-per-config rows.
@@ -269,8 +340,8 @@ const runScopeKey = (r: BenchmarkRow): string =>
  * e.g. selecting one of two same-day vLLM runs made the day's SGLang curve
  * vanish because it lived in a different workflow run.
  *
- * Run rows win for every (model, precision, hardware, framework,
- * benchmark_type) group they cover; base rows fill in the rest.
+ * Run rows win for every exact generated topology they cover; base rows fill
+ * in sibling topologies and unrelated series.
  */
 export function mergeRunScopedRows(
   runRows: BenchmarkRow[],
@@ -316,7 +387,12 @@ export function transformBenchmarkRows(
       if (hwConfig) gpuConfig[hwKey] = { ...hwConfig, name: hwKey };
     }
 
-    prepared[i] = { entry, hwKey, date: row.date };
+    prepared[i] = {
+      entry,
+      hwKey,
+      date: row.date,
+      derivedFields: buildDerivedChartFields(entry, hwKey),
+    };
   }
 
   // Phase 2: Build chart data per chart definition (reusing prepared entries)
@@ -324,13 +400,14 @@ export function transformBenchmarkRows(
     const xKey = withPercentile(chartDef.x, percentile);
     const groupedByHw: Record<string, InferenceData[]> = {};
 
-    for (const { entry, hwKey, date } of prepared) {
+    for (const { entry, hwKey, date, derivedFields } of prepared) {
       const dataPoint = createChartDataPoint(
         date,
         entry,
         xKey as keyof AggDataEntry,
         chartDef.y as keyof AggDataEntry,
         hwKey,
+        derivedFields,
       );
 
       if (!groupedByHw[hwKey]) groupedByHw[hwKey] = [];
diff --git a/packages/app/src/lib/blog-mdx.ts b/packages/app/src/lib/blog-mdx.ts
new file mode 100644
index 000000000..1251f3cfc
--- /dev/null
+++ b/packages/app/src/lib/blog-mdx.ts
@@ -0,0 +1,65 @@
+import { compileMDX } from 'next-mdx-remote/rsc';
+import rehypeShikiFromHighlighter from '@shikijs/rehype/core';
+import rehypeKatex from 'rehype-katex';
+import remarkGfm from 'remark-gfm';
+import remarkMath from 'remark-math';
+import { createHighlighterCore } from 'shiki/core';
+import { createOnigurumaEngine } from 'shiki/engine/oniguruma';
+import type { HighlighterCore } from 'shiki/types';
+
+import { createMdxComponents } from '@/components/blog/mdx-components';
+import type { BlogLocale } from '@/lib/blog';
+
+let highlighterPromise: Promise<HighlighterCore> | null = null;
+
+function getHighlighter() {
+  if (!highlighterPromise) {
+    highlighterPromise = createHighlighterCore({
+      themes: [import('shiki/themes/github-dark.mjs'), import('shiki/themes/github-light.mjs')],
+      langs: [
+        import('shiki/langs/typescript.mjs'),
+        import('shiki/langs/javascript.mjs'),
+        import('shiki/langs/python.mjs'),
+        import('shiki/langs/bash.mjs'),
+        import('shiki/langs/json.mjs'),
+        import('shiki/langs/yaml.mjs'),
+        import('shiki/langs/css.mjs'),
+        import('shiki/langs/html.mjs'),
+        import('shiki/langs/tsx.mjs'),
+        import('shiki/langs/jsx.mjs'),
+        import('shiki/langs/sql.mjs'),
+        import('shiki/langs/go.mjs'),
+        import('shiki/langs/rust.mjs'),
+      ],
+      engine: createOnigurumaEngine(import('shiki/wasm')),
+    });
+  }
+  return highlighterPromise;
+}
+
+export async function compileBlogMdx(source: string, locale: BlogLocale = 'en') {
+  const highlighter = await getHighlighter();
+
+  return compileMDX({
+    source,
+    components: createMdxComponents(locale),
+    options: {
+      mdxOptions: {
+        // Keep prices like $1.95/GPU/hr as prose; math must use `$$` delimiters.
+        remarkPlugins: [remarkGfm, [remarkMath, { singleDollarTextMath: false }]],
+        rehypePlugins: [
+          // A malformed expression renders inline in red instead of failing the whole build.
+          [rehypeKatex, { strict: false }],
+          [
+            rehypeShikiFromHighlighter,
+            highlighter,
+            {
+              themes: { dark: 'github-dark', light: 'github-light' },
+              defaultColor: false,
+            },
+          ],
+        ],
+      },
+    },
+  });
+}
diff --git a/packages/app/src/lib/blog.test.ts b/packages/app/src/lib/blog.test.ts
index ba06ed637..1f289329c 100644
--- a/packages/app/src/lib/blog.test.ts
+++ b/packages/app/src/lib/blog.test.ts
@@ -15,6 +15,7 @@ import {
   getAllPosts,
   getPostBySlug,
   getReadingTime,
+  getWordCount,
   hasZhTranslation,
   slugify,
   smartTruncate,
@@ -163,6 +164,16 @@ describe('slugify', () => {
   });
 });
 
+describe('getWordCount', () => {
+  it('counts whitespace-delimited English words', () => {
+    expect(getWordCount('one two three four five')).toBe(5);
+  });
+
+  it('counts no-whitespace Han text by character and ignores punctuation', () => {
+    expect(getWordCount('这是中文正文。')).toBe(6);
+  });
+});
+
 describe('getReadingTime', () => {
   it('returns 1 for short content', () => {
     expect(getReadingTime('hello world')).toBe(1);
@@ -686,11 +697,15 @@ describe('buildBlogPostingJsonLd', () => {
   });
 
   it('uses the zh language tag and /zh canonical URL for the Chinese page', () => {
-    const jsonLd = buildBlogPostingJsonLd(SAMPLE_META, 'a b c', 'zh') as Record<string, any>;
+    const jsonLd = buildBlogPostingJsonLd(SAMPLE_META, '这是中文正文。', 'zh') as Record<
+      string,
+      any
+    >;
     expect(jsonLd.inLanguage).toBe('zh-CN');
     expect(jsonLd.url).toBe(`${SITE_URL}/zh/blog/gb200-vs-mi355x`);
     expect(jsonLd.mainEntityOfPage['@id']).toBe(`${SITE_URL}/zh/blog/gb200-vs-mi355x`);
     expect(jsonLd.image).toBe(`${SITE_URL}/zh/blog/gb200-vs-mi355x/opengraph-image`);
+    expect(jsonLd.wordCount).toBe(6);
   });
 });
 
diff --git a/packages/app/src/lib/blog.ts b/packages/app/src/lib/blog.ts
index 444a2585e..7664b8ee4 100644
--- a/packages/app/src/lib/blog.ts
+++ b/packages/app/src/lib/blog.ts
@@ -50,10 +50,25 @@ export function slugify(raw: string): string {
 }
 
 const CJK_CHAR_REGEX = /\p{Script=Han}/gu;
+const WORD_TOKEN_REGEX = /[\p{L}\p{N}]/u;
 
-export function getReadingTime(content: string): number {
+function countContentUnits(content: string): { cjkChars: number; words: number } {
   const cjkChars = content.match(CJK_CHAR_REGEX)?.length ?? 0;
-  const words = content.replaceAll(CJK_CHAR_REGEX, ' ').trim().split(/\s+/u).filter(Boolean).length;
+  const nonCjkContent = content.replaceAll(CJK_CHAR_REGEX, ' ').trim();
+  const words = nonCjkContent
+    ? nonCjkContent.split(/\s+/u).filter((token) => WORD_TOKEN_REGEX.test(token)).length
+    : 0;
+  return { cjkChars, words };
+}
+
+/** Counts whitespace-delimited words and each Han character as one word. */
+export function getWordCount(content: string): number {
+  const { cjkChars, words } = countContentUnits(content);
+  return words + cjkChars;
+}
+
+export function getReadingTime(content: string): number {
+  const { cjkChars, words } = countContentUnits(content);
   return Math.max(1, Math.ceil(words / WORDS_PER_MINUTE + cjkChars / CJK_CHARS_PER_MINUTE));
 }
 
@@ -260,7 +275,7 @@ export function buildBlogPostingJsonLd(meta: BlogPostMeta, raw: string, locale:
     url,
     mainEntityOfPage: { '@type': 'WebPage', '@id': url },
     inLanguage: locale === 'zh' ? ZH_LANG_TAG : 'en-US',
-    wordCount: raw.trim().split(/\s+/u).length,
+    wordCount: getWordCount(raw),
     timeRequired: `PT${meta.readingTime}M`,
   };
 }
diff --git a/packages/app/src/lib/cached-derived-data.test.ts b/packages/app/src/lib/cached-derived-data.test.ts
new file mode 100644
index 000000000..655fef01d
--- /dev/null
+++ b/packages/app/src/lib/cached-derived-data.test.ts
@@ -0,0 +1,37 @@
+import { describe, expect, it, vi } from 'vitest';
+
+vi.mock('next/cache', () => ({
+  revalidateTag: vi.fn(),
+  unstable_cache: vi.fn((source: (...args: unknown[]) => Promise<unknown>) => {
+    const valuesByArguments = new Map<string, unknown>();
+    return async (...args: unknown[]): Promise<unknown> => {
+      const key = JSON.stringify(args);
+      if (valuesByArguments.has(key)) return valuesByArguments.get(key);
+
+      const value = await source(...args);
+      valuesByArguments.set(key, value);
+      return value;
+    };
+  }),
+}));
+
+vi.mock('./blob-cache', () => ({
+  blobGet: vi.fn(),
+  blobSet: vi.fn(),
+  blobPurge: vi.fn(),
+}));
+
+import { cachedDerivedData } from './api-cache';
+
+describe('cachedDerivedData performance contract', () => {
+  it('reads its source once for repeated identical inputs', async () => {
+    const sourceRead = vi.fn((ids: number[]) => Promise.resolve({ count: ids.length }));
+    const readDerivedData = cachedDerivedData(sourceRead, 'source-read-count');
+
+    await expect(readDerivedData([7, 11])).resolves.toEqual({ count: 2 });
+    await expect(readDerivedData([7, 11])).resolves.toEqual({ count: 2 });
+
+    expect(sourceRead).toHaveBeenCalledTimes(1);
+    expect(sourceRead).toHaveBeenCalledWith([7, 11]);
+  });
+});
diff --git a/packages/app/src/lib/cached-read-route.test.ts b/packages/app/src/lib/cached-read-route.test.ts
new file mode 100644
index 000000000..4e2d84bd4
--- /dev/null
+++ b/packages/app/src/lib/cached-read-route.test.ts
@@ -0,0 +1,26 @@
+import { describe, expect, it, vi } from 'vitest';
+
+vi.mock('@semianalysisai/inferencex-db/connection', () => ({ FIXTURES_MODE: true }));
+vi.mock('./api-cache', () => ({
+  cachedJson: (data: unknown) => Response.json(data),
+  cachedQuery: (fetch: () => Promise<unknown>) => fetch,
+}));
+
+import { cachedReadRoute } from './cached-read-route';
+
+describe('cachedReadRoute fixture mode', () => {
+  it('serves the fixture without calling the database reader', async () => {
+    const fetch = vi.fn().mockRejectedValue(new Error('database must not be read'));
+    const GET = cachedReadRoute({
+      cacheKey: 'fixture-contract',
+      fetch,
+      fixture: () => ({ source: 'fixture' }),
+      logLabel: 'fixture contract',
+    });
+
+    const response = await GET();
+    expect(response.status).toBe(200);
+    expect(await response.json()).toEqual({ source: 'fixture' });
+    expect(fetch).not.toHaveBeenCalled();
+  });
+});
diff --git a/packages/app/src/lib/cached-read-route.ts b/packages/app/src/lib/cached-read-route.ts
new file mode 100644
index 000000000..0ecb06468
--- /dev/null
+++ b/packages/app/src/lib/cached-read-route.ts
@@ -0,0 +1,30 @@
+import { FIXTURES_MODE } from '@semianalysisai/inferencex-db/connection';
+
+import { cachedJson, cachedQuery } from './api-cache';
+import { PUBLIC_API_ERRORS, publicApiError } from './public-api-errors';
+
+interface CachedReadRouteOptions<T> {
+  readonly cacheKey: string;
+  readonly fetch: () => Promise<T>;
+  readonly logLabel: string;
+  readonly blobOnly?: boolean;
+  readonly fixture?: () => T;
+}
+
+/** Build the standard parameterless, public, cached database-read handler. */
+export function cachedReadRoute<T>(options: CachedReadRouteOptions<T>): () => Promise<Response> {
+  const getCached = cachedQuery(options.fetch, options.cacheKey, {
+    blobOnly: options.blobOnly,
+  });
+
+  return async () => {
+    if (FIXTURES_MODE && options.fixture) return cachedJson(options.fixture());
+
+    try {
+      return cachedJson(await getCached());
+    } catch (error) {
+      console.error(`Error fetching ${options.logLabel}:`, error);
+      return publicApiError(PUBLIC_API_ERRORS.internal, 500);
+    }
+  };
+}
diff --git a/packages/app/src/lib/chart-utils.test.ts b/packages/app/src/lib/chart-utils.test.ts
index b5d667cba..f3114f760 100644
--- a/packages/app/src/lib/chart-utils.test.ts
+++ b/packages/app/src/lib/chart-utils.test.ts
@@ -1,4 +1,6 @@
 import { describe, it, expect, vi } from 'vitest';
+
+import { USD_TO_CNY } from '@semianalysisai/inferencex-constants';
 import iwanthue from 'iwanthue';
 
 import type * as ConstantsModule from '@/lib/constants';
@@ -10,9 +12,7 @@ import {
   getHardwareKey,
   normalizeEvalHardwareKey,
   createChartDataPoint,
-  calculateRoofline,
-  computeAllRooflines,
-  markRooflinePoints,
+  buildDerivedChartFields,
   paretoFrontUpperRight,
   paretoFrontLowerRight,
   paretoFrontLowerLeft,
@@ -28,7 +28,7 @@ vi.mock('@/lib/constants', async (importOriginal) => {
   return {
     ...actual,
     getHardwareConfig: vi.fn(() => ({ label: 'H100', suffix: '' })),
-    getGpuSpecs: vi.fn(() => ({ power: 700, costh: 2.8, costn: 1.4, costr: 0.7 })),
+    getGpuSpecs: vi.fn(() => ({ power: 700, tdp: 700, costh: 2.8, costn: 1.4, costr: 0.7 })),
   };
 });
 
@@ -148,70 +148,6 @@ function paretoPt(x: number, y: number, overrides: Partial<InferenceData> = {}):
 // assertion helper, extracts {x, y} from each point
 const xy = (pts: InferenceData[]) => pts.map((p) => ({ x: p.x, y: p.y }));
 
-/**
- * Creates a minimal InferenceData point with all optional roofline fields populated.
- */
-function fullPt(
-  x: number,
-  hwKey: string,
-  vals: {
-    tpPerGpuY: number;
-    costhY?: number;
-    costhOutputY?: number;
-    costnOutputY?: number;
-    costrOutputY?: number;
-    jTotalY?: number;
-    jOutputY?: number;
-    jInputY?: number;
-    outputTputY?: number;
-    inputTputY?: number;
-    inputTputPerMwY?: number;
-    outputTputPerMwY?: number;
-  },
-): InferenceData {
-  return {
-    date: '2025-01-15',
-    x,
-    y: 0,
-    tp: 8,
-    conc: 64,
-    hwKey,
-    precision: 'fp8',
-    tpPerGpu: { y: vals.tpPerGpuY, roof: false },
-    tpPerMw: { y: 5, roof: false },
-    costh: { y: vals.costhY ?? 1, roof: false },
-    costn: { y: 1.5, roof: false },
-    costr: { y: 1.2, roof: false },
-    costhi: { y: 2, roof: false },
-    costni: { y: 2.5, roof: false },
-    costri: { y: 2.2, roof: false },
-    ...(vals.costhOutputY === undefined
-      ? {}
-      : { costhOutput: { y: vals.costhOutputY, roof: false } }),
-    ...(vals.costnOutputY === undefined
-      ? {}
-      : { costnOutput: { y: vals.costnOutputY, roof: false } }),
-    ...(vals.costrOutputY === undefined
-      ? {}
-      : { costrOutput: { y: vals.costrOutputY, roof: false } }),
-    ...(vals.jTotalY === undefined ? {} : { jTotal: { y: vals.jTotalY, roof: false } }),
-    ...(vals.jOutputY === undefined ? {} : { jOutput: { y: vals.jOutputY, roof: false } }),
-    ...(vals.jInputY === undefined ? {} : { jInput: { y: vals.jInputY, roof: false } }),
-    ...(vals.outputTputY === undefined
-      ? {}
-      : { outputTputPerGpu: { y: vals.outputTputY, roof: false } }),
-    ...(vals.inputTputY === undefined
-      ? {}
-      : { inputTputPerGpu: { y: vals.inputTputY, roof: false } }),
-    ...(vals.inputTputPerMwY === undefined
-      ? {}
-      : { inputTputPerMw: { y: vals.inputTputPerMwY, roof: false } }),
-    ...(vals.outputTputPerMwY === undefined
-      ? {}
-      : { outputTputPerMw: { y: vals.outputTputPerMwY, roof: false } }),
-  } as InferenceData;
-}
-
 // ===========================================================================
 // buildAvailabilityHwKey
 // ===========================================================================
@@ -523,334 +459,6 @@ describe('getNestedYValue', () => {
   });
 });
 
-// ===========================================================================
-// calculateRoofline
-// ===========================================================================
-describe('calculateRoofline', () => {
-  it('returns an empty array for empty input', () => {
-    expect(calculateRoofline([], 'tpPerGpu.y', 'upper_right')).toEqual([]);
-  });
-
-  it('extracts nested y-value (dot notation) and maps it onto y before computing the front', () => {
-    // A(x=1, tpPerGpu.y=50), B(x=2, tpPerGpu.y=80), C(x=3, tpPerGpu.y=60)
-    // upper_right front → A and B; each result point's y is the tpPerGpu.y value
-    const points = [
-      pt(1, 999, 'h100', { tpPerGpuY: 50 }), // y=999 should be ignored
-      pt(2, 999, 'h100', { tpPerGpuY: 80 }),
-      pt(3, 999, 'h100', { tpPerGpuY: 60 }),
-    ];
-    const front = calculateRoofline(points, 'tpPerGpu.y', 'upper_right');
-    expect(front).toHaveLength(2);
-    expect(front[0].y).toBe(50); // remapped from tpPerGpu.y, not the original y=999
-    expect(front[1].y).toBe(80);
-  });
-
-  it('works with a flat (non-nested) y key', () => {
-    const points = [pt(1, 5, 'h100'), pt(2, 8, 'h100'), pt(3, 6, 'h100')];
-    const front = calculateRoofline(points, 'y', 'upper_right');
-    expect(front).toHaveLength(2);
-    expect(front[0].y).toBe(5);
-    expect(front[1].y).toBe(8);
-  });
-
-  it('excludes x <= 0 points (e.g. interactivity = 0) from the frontier', () => {
-    // A degenerate x=0 point with the highest tpPerGpu would otherwise anchor the
-    // upper_right front as its leftmost point and show up as "optimal".
-    const points = [
-      pt(0, 999, 'h100', { tpPerGpuY: 999 }), // interactivity = 0 → degenerate
-      pt(1, 999, 'h100', { tpPerGpuY: 50 }),
-      pt(2, 999, 'h100', { tpPerGpuY: 80 }),
-    ];
-    const front = calculateRoofline(points, 'tpPerGpu.y', 'upper_right');
-    expect(front.every((p) => p.x > 0)).toBe(true);
-    expect(front.some((p) => p.x === 0)).toBe(false);
-    // real front is the two positive-x points
-    expect(front.map((p) => p.y)).toEqual([50, 80]);
-  });
-
-  it('excludes non-finite / negative x from the frontier', () => {
-    const points = [
-      pt(Number.NaN, 999, 'h100', { tpPerGpuY: 999 }),
-      pt(-1, 999, 'h100', { tpPerGpuY: 999 }),
-      pt(3, 999, 'h100', { tpPerGpuY: 40 }),
-    ];
-    const front = calculateRoofline(points, 'tpPerGpu.y', 'upper_right');
-    expect(front).toHaveLength(1);
-    expect(front[0].x).toBe(3);
-  });
-
-  it.each([
-    ['upper_right', 2], // A(1,50) and B(2,80) both on front
-    ['upper_left', 1], // B(2,80) dominates A(1,50) → only B kept
-    ['lower_left', 1], // A(1,50) is lowest; B(2,80) has higher y → only A kept
-    ['lower_right', 1], // sorted desc by x: B(2,80) first, A(1,50) next (50<80 → add); but wait
-  ] as const)('routes direction "%s" to the correct Pareto function', (direction, _) => {
-    // Just verify the function doesn't throw and returns an array for each direction
-    const points = [pt(1, 0, 'h100', { tpPerGpuY: 50 }), pt(2, 0, 'h100', { tpPerGpuY: 80 })];
-    const front = calculateRoofline(points, 'tpPerGpu.y', direction);
-    expect(Array.isArray(front)).toBe(true);
-  });
-
-  it('delegates to paretoFrontUpperRight for "upper_right" (specific assertions)', () => {
-    const rPoints = [
-      paretoPt(1, 99, { tpPerGpu: { y: 10, roof: false } }),
-      paretoPt(2, 99, { tpPerGpu: { y: 30, roof: false } }),
-      paretoPt(3, 99, { tpPerGpu: { y: 20, roof: false } }),
-    ];
-    // after remapping y: (1,10),(2,30),(3,20)
-    // upper-right front: (1,10),(2,30) — (3,20) dropped because y=20<maxY=30
-    const result = calculateRoofline(rPoints, 'tpPerGpu.y', 'upper_right');
-    expect(xy(result)).toEqual([
-      { x: 1, y: 10 },
-      { x: 2, y: 30 },
-    ]);
-  });
-
-  it('delegates to paretoFrontUpperLeft for "upper_left" (specific assertions)', () => {
-    const rPoints = [
-      paretoPt(1, 99, { tpPerGpu: { y: 10, roof: false } }),
-      paretoPt(2, 99, { tpPerGpu: { y: 30, roof: false } }),
-      paretoPt(3, 99, { tpPerGpu: { y: 20, roof: false } }),
-    ];
-    // remapped: (1,10),(2,30),(3,20)
-    // upper-left: (1,10) push, (2,30): y=30>=10, pop (1,10), push (2,30),
-    //             (3,20): y=20<30, push → [(2,30),(3,20)]
-    const result = calculateRoofline(rPoints, 'tpPerGpu.y', 'upper_left');
-    expect(xy(result)).toEqual([
-      { x: 2, y: 30 },
-      { x: 3, y: 20 },
-    ]);
-  });
-
-  it('delegates to paretoFrontLowerLeft for "lower_left" (specific assertions)', () => {
-    const rPoints = [
-      paretoPt(1, 99, { tpPerGpu: { y: 10, roof: false } }),
-      paretoPt(2, 99, { tpPerGpu: { y: 30, roof: false } }),
-      paretoPt(3, 99, { tpPerGpu: { y: 20, roof: false } }),
-    ];
-    // lower-left keeps new minimums: (1,10) min=10; (2,30) skip; (3,20) skip → [(1,10)]
-    const result = calculateRoofline(rPoints, 'tpPerGpu.y', 'lower_left');
-    expect(xy(result)).toEqual([{ x: 1, y: 10 }]);
-  });
-
-  it('delegates to paretoFrontLowerRight for "lower_right" (specific assertions)', () => {
-    const rPoints = [
-      paretoPt(1, 99, { tpPerGpu: { y: 10, roof: false } }),
-      paretoPt(2, 99, { tpPerGpu: { y: 30, roof: false } }),
-      paretoPt(3, 99, { tpPerGpu: { y: 20, roof: false } }),
-    ];
-    // sorted x desc: (3,20),(2,30),(1,10)
-    // (3,20): min=20; (2,30): skip; (1,10): y=10<20, new min → [(3,20),(1,10)]
-    const result = calculateRoofline(rPoints, 'tpPerGpu.y', 'lower_right');
-    expect(xy(result)).toEqual([
-      { x: 3, y: 20 },
-      { x: 1, y: 10 },
-    ]);
-  });
-
-  it('returns empty array for an unrecognised roofline direction', () => {
-    const rPoints = [paretoPt(1, 5), paretoPt(2, 3)];
-    const result = calculateRoofline(rPoints, 'tpPerGpu.y', 'diagonal' as any);
-    expect(result).toEqual([]);
-  });
-
-  it('does not mutate the original point objects — only the remapped copies', () => {
-    const original = paretoPt(1, 99, { tpPerGpu: { y: 10, roof: false } });
-    calculateRoofline([original], 'tpPerGpu.y', 'upper_right');
-    // the original point's y should still be 99, not 10
-    expect(original.y).toBe(99);
-  });
-});
-
-// ===========================================================================
-// computeAllRooflines
-// ===========================================================================
-describe('computeAllRooflines', () => {
-  const chartDef: ChartDefinition = {
-    chartType: 'e2e',
-    heading: 'Test',
-    x: 'median_e2el',
-    x_label: 'E2E Latency',
-    y: 'tput_per_gpu',
-    y_tpPerGpu: 'tpPerGpu.y',
-    y_tpPerGpu_roofline: 'upper_right',
-    y_costh: 'costh.y',
-    y_costh_roofline: 'lower_left',
-  };
-
-  const groupedData = {
-    h100: [
-      pt(1, 0, 'h100', { tpPerGpuY: 50, costhY: 2 }),
-      pt(2, 0, 'h100', { tpPerGpuY: 80, costhY: 1.5 }),
-      pt(3, 0, 'h100', { tpPerGpuY: 60, costhY: 1.8 }),
-    ],
-  };
-
-  it('returns a roofline entry for each hardware group', () => {
-    const result = computeAllRooflines(groupedData, chartDef);
-    expect(result).toHaveProperty('h100');
-  });
-
-  it('computes y_tpPerGpu roofline (upper_right)', () => {
-    const result = computeAllRooflines(groupedData, chartDef);
-    const front = result.h100.y_tpPerGpu;
-    // points A(1,50) and B(2,80) form the upper_right front
-    expect(front).toHaveLength(2);
-    expect(front[0].x).toBe(1);
-    expect(front[0].y).toBe(50);
-    expect(front[1].x).toBe(2);
-    expect(front[1].y).toBe(80);
-  });
-
-  it('computes y_costh roofline (lower_left)', () => {
-    const result = computeAllRooflines(groupedData, chartDef);
-    const front = result.h100.y_costh;
-    // costh: A(x=1,2.0), B(x=2,1.5), C(x=3,1.8)
-    // lower_left sorted by x: A(1,2.0) added, B(1.5<2.0) added, C(1.8>1.5) skipped
-    expect(front).toHaveLength(2);
-    expect(front[0].x).toBe(1);
-    expect(front[0].y).toBe(2);
-    expect(front[1].x).toBe(2);
-    expect(front[1].y).toBe(1.5);
-  });
-
-  it('skips metrics not defined in the chartDef', () => {
-    const result = computeAllRooflines(groupedData, chartDef);
-    // y_tpPerMw is not defined in chartDef → should be undefined or empty
-    expect(result.h100.y_tpPerMw).toBeUndefined();
-  });
-
-  it('handles multiple hardware groups independently', () => {
-    const multiGrouped = {
-      h100: [pt(1, 0, 'h100', { tpPerGpuY: 50 }), pt(2, 0, 'h100', { tpPerGpuY: 80 })],
-      h200: [pt(1, 0, 'h200', { tpPerGpuY: 100 }), pt(2, 0, 'h200', { tpPerGpuY: 70 })],
-    };
-    const result = computeAllRooflines(multiGrouped, chartDef);
-    // h100 upper_right: [A(1,50), B(2,80)]
-    expect(result.h100.y_tpPerGpu).toHaveLength(2);
-    // h200 upper_right: only A(1,100) since B has lower y
-    expect(result.h200.y_tpPerGpu).toHaveLength(1);
-    expect(result.h200.y_tpPerGpu[0].x).toBe(1);
-  });
-});
-
-// ===========================================================================
-// markRooflinePoints
-// ===========================================================================
-describe('markRooflinePoints', () => {
-  // 3 points for h100: A(x=1, tpPerGpu.y=50), B(x=2, tpPerGpu.y=80), C(x=3, tpPerGpu.y=60)
-  // upper_right front = [A, B]  → A.tpPerGpu.roof=true, B.tpPerGpu.roof=true, C.tpPerGpu.roof=false
-  const chartDef: ChartDefinition = {
-    chartType: 'e2e',
-    heading: 'Test',
-    x: 'median_e2el',
-    x_label: 'E2E Latency',
-    y: 'tput_per_gpu',
-    y_tpPerGpu: 'tpPerGpu.y',
-    y_tpPerGpu_roofline: 'upper_right',
-  };
-
-  const pointA = pt(1, 0, 'h100', { tpPerGpuY: 50 });
-  const pointB = pt(2, 0, 'h100', { tpPerGpuY: 80 });
-  const pointC = pt(3, 0, 'h100', { tpPerGpuY: 60 });
-
-  const groupedData = { h100: [pointA, pointB, pointC] };
-
-  function buildRooflines() {
-    return computeAllRooflines(groupedData, chartDef);
-  }
-
-  it('sets tpPerGpu.roof=true for points on the upper_right roofline', () => {
-    const rooflines = buildRooflines();
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    const a = marked.find((p) => p.x === 1)!;
-    const b = marked.find((p) => p.x === 2)!;
-    expect(a.tpPerGpu.roof).toBe(true);
-    expect(b.tpPerGpu.roof).toBe(true);
-  });
-
-  it('sets tpPerGpu.roof=false for points NOT on the roofline', () => {
-    const rooflines = buildRooflines();
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    const c = marked.find((p) => p.x === 3)!;
-    expect(c.tpPerGpu.roof).toBe(false);
-  });
-
-  it('resets roof flags to false before re-evaluating (no stale state)', () => {
-    // manually pre-set a roof flag to true on a point that's NOT on the front
-    const contaminated = { ...pointC, tpPerGpu: { y: 60, roof: true } };
-    const dirtyGroup = { h100: [pointA, pointB, contaminated] };
-    const rooflines = computeAllRooflines(dirtyGroup, chartDef);
-    const marked = markRooflinePoints(dirtyGroup, rooflines, chartDef);
-
-    const c = marked.find((p) => p.x === 3)!;
-    expect(c.tpPerGpu.roof).toBe(false);
-  });
-
-  it('returns all input points (none are filtered out)', () => {
-    const rooflines = buildRooflines();
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-    expect(marked).toHaveLength(3);
-  });
-
-  it('handles multiple hardware groups, marking each independently', () => {
-    const chartDefMulti: ChartDefinition = {
-      ...chartDef,
-      y_tpPerGpu: 'tpPerGpu.y',
-      y_tpPerGpu_roofline: 'upper_right',
-    };
-    const h200A = pt(1, 0, 'h200', { tpPerGpuY: 100 });
-    const h200B = pt(2, 0, 'h200', { tpPerGpuY: 70 });
-    const multiGrouped = {
-      h100: [pointA, pointB, pointC],
-      h200: [h200A, h200B],
-    };
-    const rooflines = computeAllRooflines(multiGrouped, chartDefMulti);
-    const marked = markRooflinePoints(multiGrouped, rooflines, chartDefMulti);
-
-    // h200: A(x=1,y=100) on front, B(x=2,y=70) not
-    const h200Points = marked.filter((p) => p.hwKey === 'h200');
-    const h200Marked_A = h200Points.find((p) => p.x === 1)!;
-    const h200Marked_B = h200Points.find((p) => p.x === 2)!;
-    expect(h200Marked_A.tpPerGpu.roof).toBe(true);
-    expect(h200Marked_B.tpPerGpu.roof).toBe(false);
-  });
-
-  it('does not mark costh.roof when y_costh is not in chartDef', () => {
-    const rooflines = buildRooflines();
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-    // costh roofline not in this chartDef → costh.roof should remain false
-    for (const p of marked) {
-      expect(p.costh.roof).toBe(false);
-    }
-  });
-
-  it('marks costh.roof when y_costh_roofline is defined in chartDef', () => {
-    const chartDefWithCost: ChartDefinition = {
-      ...chartDef,
-      y_costh: 'costh.y',
-      y_costh_roofline: 'lower_left',
-    };
-    // costh values: A(x=1, costh.y=2.0), B(x=2, costh.y=1.5), C(x=3, costh.y=1.8)
-    const pA = pt(1, 0, 'h100', { tpPerGpuY: 50, costhY: 2 });
-    const pB = pt(2, 0, 'h100', { tpPerGpuY: 80, costhY: 1.5 });
-    const pC = pt(3, 0, 'h100', { tpPerGpuY: 60, costhY: 1.8 });
-    const group = { h100: [pA, pB, pC] };
-    const rooflines = computeAllRooflines(group, chartDefWithCost);
-    const marked = markRooflinePoints(group, rooflines, chartDefWithCost);
-
-    // lower_left: A(1,2.0) and B(2,1.5) are on the front
-    const mA = marked.find((p) => p.x === 1)!;
-    const mB = marked.find((p) => p.x === 2)!;
-    const mC = marked.find((p) => p.x === 3)!;
-    expect(mA.costh.roof).toBe(true);
-    expect(mB.costh.roof).toBe(true);
-    expect(mC.costh.roof).toBe(false);
-  });
-});
-
 // ---------------------------------------------------------------------------
 // Pareto front x-ordering — gradient labels require ascending x
 // ---------------------------------------------------------------------------
@@ -1068,6 +676,13 @@ describe('createChartDataPoint', () => {
     expect(point.tpPerGpu).toEqual({ y: 2000, roof: false });
   });
 
+  it('computes token revenue per GPU hour at the normalized $1/M token price', () => {
+    const e = entry({ tput_per_gpu: 2000 });
+    const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
+    // 2,000 tok/s/GPU × 3,600 s/hr ÷ 1,000,000 tok × $1/M tok = $7.20/GPU/hr.
+    expect(point.tokenRevenuePerGpuHour).toEqual({ y: 7.2, roof: false });
+  });
+
   it('sets outputTputPerGpu when output_tput_per_gpu > 0', () => {
     const e = entry({ output_tput_per_gpu: 800 });
     const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
@@ -1123,14 +738,26 @@ describe('createChartDataPoint', () => {
     expect(point.outputTputPerMw).toBeUndefined();
   });
 
-  it('computes cost fields (costh, costn, costr) from hardware config and throughput', () => {
-    // tokensPerHour = (tput_per_gpu * 3600) / 1_000_000 = (1000 * 3600) / 1e6 = 3.6
-    // costh.y = hwConfig.costh / tokensPerHour = 2.8 / 3.6
+  it('keeps cost-per-million fields and adds total tokens-per-dollar fields', () => {
     const e = entry({ tput_per_gpu: 1000 });
     const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
     expect(point.costh.y).toBeCloseTo(2.8 / 3.6, 5);
     expect(point.costn.y).toBeCloseTo(1.4 / 3.6, 5);
     expect(point.costr.y).toBeCloseTo(0.7 / 3.6, 5);
+    expect(point.tokensPerDollarH!.y).toBeCloseTo(3_600_000 / 2.8, 5);
+    expect(point.tokensPerDollarN!.y).toBeCloseTo(3_600_000 / 1.4, 5);
+    expect(point.tokensPerDollarR!.y).toBeCloseTo(3_600_000 / 0.7, 5);
+  });
+
+  it('prices the same tokens in yuan at the pinned FX rate', () => {
+    const e = entry({ tput_per_gpu: 1000 });
+    const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
+    // ¥ metrics are the $ metrics over USD_TO_CNY — the same tokens, priced in
+    // the other currency, so the two must stay in exact proportion.
+    expect(point.tokensPerRmbH!.y).toBeCloseTo(3_600_000 / (2.8 * USD_TO_CNY), 5);
+    expect(point.tokensPerRmbN!.y).toBeCloseTo(3_600_000 / (1.4 * USD_TO_CNY), 5);
+    expect(point.tokensPerRmbR!.y).toBeCloseTo(3_600_000 / (0.7 * USD_TO_CNY), 5);
+    expect(point.tokensPerRmbH!.y * USD_TO_CNY).toBeCloseTo(point.tokensPerDollarH!.y, 5);
   });
 
   it('sets cost fields to 0 when throughput is 0', () => {
@@ -1141,22 +768,24 @@ describe('createChartDataPoint', () => {
     expect(point.costr.y).toBe(0);
   });
 
-  it('computes output cost fields when output_tput_per_gpu > 0', () => {
+  it('adds output tokens-per-dollar fields without replacing output cost fields', () => {
     const e = entry({ output_tput_per_gpu: 500 });
-    // outputTokensPerHour = (500 * 3600) / 1e6 = 1.8
+    // outputTokensPerHour = 500 * 3600 = 1,800,000
     const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
     expect(point.costhOutput!.y).toBeCloseTo(2.8 / 1.8, 5);
-    expect(point.costnOutput!.y).toBeCloseTo(1.4 / 1.8, 5);
-    expect(point.costrOutput!.y).toBeCloseTo(0.7 / 1.8, 5);
+    expect(point.outputTokensPerDollarH!.y).toBeCloseTo(1_800_000 / 2.8, 5);
+    expect(point.outputTokensPerDollarN!.y).toBeCloseTo(1_800_000 / 1.4, 5);
+    expect(point.outputTokensPerDollarR!.y).toBeCloseTo(1_800_000 / 0.7, 5);
   });
 
-  it('computes input cost fields when input_tput_per_gpu > 0', () => {
+  it('adds input tokens-per-dollar fields without replacing input cost fields', () => {
     const e = entry({ input_tput_per_gpu: 200 });
-    // inputTokensPerHour = (200 * 3600) / 1e6 = 0.72
+    // inputTokensPerHour = 200 * 3600 = 720,000
     const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
     expect(point.costhi.y).toBeCloseTo(2.8 / 0.72, 5);
-    expect(point.costni.y).toBeCloseTo(1.4 / 0.72, 5);
-    expect(point.costri.y).toBeCloseTo(0.7 / 0.72, 5);
+    expect(point.inputTokensPerDollarH!.y).toBeCloseTo(720_000 / 2.8, 5);
+    expect(point.inputTokensPerDollarN!.y).toBeCloseTo(720_000 / 1.4, 5);
+    expect(point.inputTokensPerDollarR!.y).toBeCloseTo(720_000 / 0.7, 5);
   });
 
   it('narrows dp_attention string "true" to boolean true', () => {
@@ -1220,6 +849,47 @@ describe('createChartDataPoint', () => {
   });
 });
 
+describe('buildDerivedChartFields', () => {
+  it('matches full inference formulas while emitting only requested history fields', () => {
+    const e = entry({
+      tput_per_gpu: 900,
+      output_tput_per_gpu: 600,
+      input_tput_per_gpu: 300,
+    });
+    const fullPoint = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
+    const historicalFields = buildDerivedChartFields(e, 'h100', [
+      'outputTputPerGpu',
+      'costhOutput',
+    ]);
+
+    expect(historicalFields).toEqual({
+      outputTputPerGpu: fullPoint.outputTputPerGpu,
+      costhOutput: fullPoint.costhOutput,
+    });
+  });
+
+  it('selectively derives normalized token revenue for historical trends', () => {
+    const historicalFields = buildDerivedChartFields(entry({ tput_per_gpu: 1250 }), 'h100', [
+      'tokenRevenuePerGpuHour',
+    ]);
+
+    expect(historicalFields).toEqual({
+      tokenRevenuePerGpuHour: { y: 4.5, roof: false },
+    });
+  });
+
+  it('preserves missing measured metrics in selective output', () => {
+    const historicalFields = buildDerivedChartFields(entry(), 'h100', [
+      'tpPerGpu',
+      'measuredJPerTotalToken',
+    ]);
+
+    expect(historicalFields.tpPerGpu).toEqual({ y: 1000, roof: false });
+    expect(historicalFields.measuredJPerTotalToken).toBeUndefined();
+    expect(Object.keys(historicalFields)).toEqual(['tpPerGpu']);
+  });
+});
+
 // ===========================================================================
 // getHardwareKey — additional edge cases
 // ===========================================================================
@@ -1333,6 +1003,29 @@ describe('createChartDataPoint measured power fields', () => {
     expect(point.measuredJPerOutputToken!.y).toBe(8.4);
   });
 
+  it('derives J/query, Wh/query, and percent TDP from validated source fields', () => {
+    const e = entry({ avg_power_w: 560, joules_per_successful_query: 1800 });
+    const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
+
+    expect(point.measuredJPerSuccessfulQuery?.y).toBe(1800);
+    expect(point.measuredWhPerSuccessfulQuery?.y).toBe(0.5);
+    expect(point.measuredPowerPercentTdp?.y).toBe(80);
+  });
+
+  it('omits derived query and TDP axes when their inputs are absent', () => {
+    const point = createChartDataPoint(
+      '2025-01-01',
+      entry(),
+      'median_e2el',
+      'tput_per_gpu',
+      'h100',
+    );
+
+    expect(point.measuredJPerSuccessfulQuery).toBeUndefined();
+    expect(point.measuredWhPerSuccessfulQuery).toBeUndefined();
+    expect(point.measuredPowerPercentTdp).toBeUndefined();
+  });
+
   it('omits both fields when neither is on the entry', () => {
     // Legacy runs predating aggregate_power.py.
     const point = createChartDataPoint(
@@ -1551,7 +1244,7 @@ describe('createChartDataPoint output cost edge cases', () => {
     expect(point.costri.y).toBe(0);
   });
 
-  it('computes all 9 cost fields correctly for a point with all throughput types', () => {
+  it('computes all 9 added tokens-per-dollar fields for a point with all throughput types', () => {
     const e = entry({
       tput_per_gpu: 1000,
       output_tput_per_gpu: 500,
@@ -1559,386 +1252,20 @@ describe('createChartDataPoint output cost edge cases', () => {
     });
     const point = createChartDataPoint('2025-01-01', e, 'median_e2el', 'tput_per_gpu', 'h100');
 
-    // Total: tokensPerHour = (1000 * 3600) / 1e6 = 3.6
-    expect(point.costh.y).toBeCloseTo(2.8 / 3.6, 5);
-    expect(point.costn.y).toBeCloseTo(1.4 / 3.6, 5);
-    expect(point.costr.y).toBeCloseTo(0.7 / 3.6, 5);
-
-    // Output: outputTokensPerHour = (500 * 3600) / 1e6 = 1.8
-    expect(point.costhOutput!.y).toBeCloseTo(2.8 / 1.8, 5);
-    expect(point.costnOutput!.y).toBeCloseTo(1.4 / 1.8, 5);
-    expect(point.costrOutput!.y).toBeCloseTo(0.7 / 1.8, 5);
-
-    // Input: inputTokensPerHour = (200 * 3600) / 1e6 = 0.72
-    expect(point.costhi.y).toBeCloseTo(2.8 / 0.72, 5);
-    expect(point.costni.y).toBeCloseTo(1.4 / 0.72, 5);
-    expect(point.costri.y).toBeCloseTo(0.7 / 0.72, 5);
-  });
-});
-
-// ===========================================================================
-// computeAllRooflines — additional edge cases
-// ===========================================================================
-describe('computeAllRooflines edge cases', () => {
-  it('returns empty object for empty groupedData', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_tpPerGpu: 'tpPerGpu.y',
-      y_tpPerGpu_roofline: 'upper_right',
-    } as any;
-    const result = computeAllRooflines({}, chartDef);
-    expect(Object.keys(result)).toHaveLength(0);
-  });
-
-  it('skips metrics where roofline direction is not defined', () => {
-    // chartDef has y_tpPerGpu defined but no y_tpPerGpu_roofline
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_tpPerGpu: 'tpPerGpu.y',
-      // intentionally no y_tpPerGpu_roofline
-    } as any;
-    const groupedData = {
-      h100: [pt(1, 50, 'h100', { tpPerGpuY: 50 })],
-    };
-    const result = computeAllRooflines(groupedData, chartDef);
-    expect(result.h100.y_tpPerGpu).toBeUndefined();
-  });
-
-  it('computes energy roofline (jTotal) when defined in chartDef', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_jTotal: 'jTotal.y',
-      y_jTotal_roofline: 'lower_left',
-    } as any;
-    const p1 = pt(1, 0, 'h100');
-    (p1 as any).jTotal = { y: 5, roof: false };
-    const p2 = pt(2, 0, 'h100');
-    (p2 as any).jTotal = { y: 3, roof: false };
-    const p3 = pt(3, 0, 'h100');
-    (p3 as any).jTotal = { y: 4, roof: false };
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const result = computeAllRooflines(groupedData, chartDef);
-    // lower_left front: p1(1,5.0) added, p2(2,3.0) < 5.0 so added, p3(3,4.0) > 3.0 so skipped
-    expect(result.h100.y_jTotal).toHaveLength(2);
-    expect(result.h100.y_jTotal[0].y).toBe(5);
-    expect(result.h100.y_jTotal[1].y).toBe(3);
-  });
-
-  it('handles chartDef with no roofline keys at all', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      // No y_*_roofline keys at all
-    } as any;
-    const groupedData = {
-      h100: [pt(1, 50, 'h100')],
-    };
-    const result = computeAllRooflines(groupedData, chartDef);
-    // All metric rooflines should be undefined
-    for (const metric of [
-      'y_tpPerGpu',
-      'y_costh',
-      'y_costn',
-      'y_costr',
-      'y_tpPerMw',
-      'y_jTotal',
-    ] as const) {
-      expect((result as any).h100[metric]).toBeUndefined();
-    }
-  });
-});
-
-// ===========================================================================
-// markRooflinePoints — energy and output cost field marking
-// ===========================================================================
-describe('markRooflinePoints energy and output fields', () => {
-  it('marks jTotal.roof for points on the energy roofline', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_jTotal: 'jTotal.y',
-      y_jTotal_roofline: 'lower_left',
-    } as any;
-
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50, jTotalY: 10 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80, jTotalY: 5 });
-    const p3 = fullPt(3, 'h100', { tpPerGpuY: 60, jTotalY: 8 });
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    // lower_left: p1(1,10), p2(2,5) on front; p3(3,8) > 5 so not on front
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-    const m3 = marked.find((p) => p.x === 3)!;
-    expect(m1.jTotal!.roof).toBe(true);
-    expect(m2.jTotal!.roof).toBe(true);
-    expect(m3.jTotal!.roof).toBe(false);
-  });
-
-  it('marks costhOutput.roof for points on the output cost roofline', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_costhOutput: 'costhOutput.y',
-      y_costhOutput_roofline: 'lower_left',
-    } as any;
-
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50, costhOutputY: 3 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80, costhOutputY: 1.5 });
-    const p3 = fullPt(3, 'h100', { tpPerGpuY: 60, costhOutputY: 2.5 });
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-    const m3 = marked.find((p) => p.x === 3)!;
-    expect(m1.costhOutput!.roof).toBe(true);
-    expect(m2.costhOutput!.roof).toBe(true);
-    expect(m3.costhOutput!.roof).toBe(false);
-  });
-
-  it('marks outputTputPerGpu.roof for points on the output throughput roofline', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_outputTputPerGpu: 'outputTputPerGpu.y',
-      y_outputTputPerGpu_roofline: 'upper_right',
-    } as any;
-
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50, outputTputY: 300 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80, outputTputY: 600 });
-    const p3 = fullPt(3, 'h100', { tpPerGpuY: 60, outputTputY: 400 });
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    // upper_right: p1(1,300) and p2(2,600) on front; p3(3,400) < 600 so not on front
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-    const m3 = marked.find((p) => p.x === 3)!;
-    expect(m1.outputTputPerGpu!.roof).toBe(true);
-    expect(m2.outputTputPerGpu!.roof).toBe(true);
-    expect(m3.outputTputPerGpu!.roof).toBe(false);
-  });
-
-  it('marks inputTputPerGpu.roof for points on the input throughput roofline', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_inputTputPerGpu: 'inputTputPerGpu.y',
-      y_inputTputPerGpu_roofline: 'upper_right',
-    } as any;
-
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50, inputTputY: 200 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80, inputTputY: 500 });
-    const p3 = fullPt(3, 'h100', { tpPerGpuY: 60, inputTputY: 350 });
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-    const m3 = marked.find((p) => p.x === 3)!;
-    expect(m1.inputTputPerGpu!.roof).toBe(true);
-    expect(m2.inputTputPerGpu!.roof).toBe(true);
-    expect(m3.inputTputPerGpu!.roof).toBe(false);
-  });
-
-  it('marks inputTputPerMw.roof for points on the input per-MW roofline', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_inputTputPerMw: 'inputTputPerMw.y',
-      y_inputTputPerMw_roofline: 'upper_right',
-    } as any;
-
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50, inputTputPerMwY: 100 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80, inputTputPerMwY: 400 });
-    const p3 = fullPt(3, 'h100', { tpPerGpuY: 60, inputTputPerMwY: 250 });
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-    const m3 = marked.find((p) => p.x === 3)!;
-    expect(m1.inputTputPerMw!.roof).toBe(true);
-    expect(m2.inputTputPerMw!.roof).toBe(true);
-    expect(m3.inputTputPerMw!.roof).toBe(false);
-  });
-
-  it('marks outputTputPerMw.roof for points on the output per-MW roofline', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_outputTputPerMw: 'outputTputPerMw.y',
-      y_outputTputPerMw_roofline: 'upper_right',
-    } as any;
-
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50, outputTputPerMwY: 200 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80, outputTputPerMwY: 500 });
-    const p3 = fullPt(3, 'h100', { tpPerGpuY: 60, outputTputPerMwY: 350 });
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-    const m3 = marked.find((p) => p.x === 3)!;
-    expect(m1.outputTputPerMw!.roof).toBe(true);
-    expect(m2.outputTputPerMw!.roof).toBe(true);
-    expect(m3.outputTputPerMw!.roof).toBe(false);
-  });
-
-  it('handles points missing optional roofline fields gracefully', () => {
-    // Points without jTotal, jOutput, jInput, outputTputPerGpu, inputTputPerGpu
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_jTotal: 'jTotal.y',
-      y_jTotal_roofline: 'lower_left',
-      y_outputTputPerGpu: 'outputTputPerGpu.y',
-      y_outputTputPerGpu_roofline: 'upper_right',
-    } as any;
-
-    // Points WITHOUT jTotal or outputTputPerGpu fields
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80 });
-
-    const groupedData = { h100: [p1, p2] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    // Should not crash; optional fields stay undefined
-    expect(marked).toHaveLength(2);
-    expect(marked[0].jTotal).toBeUndefined();
-    expect(marked[0].outputTputPerGpu).toBeUndefined();
-  });
-
-  it('marks jOutput.roof and jInput.roof independently', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_jOutput: 'jOutput.y',
-      y_jOutput_roofline: 'lower_left',
-      y_jInput: 'jInput.y',
-      y_jInput_roofline: 'lower_left',
-    } as any;
-
-    // p1: low jOutput (on front), high jInput (not on front alone)
-    // p2: high jOutput (not on front), low jInput (on front)
-    const p1 = fullPt(1, 'h100', { tpPerGpuY: 50, jOutputY: 2, jInputY: 10 });
-    const p2 = fullPt(2, 'h100', { tpPerGpuY: 80, jOutputY: 5, jInputY: 4 });
-
-    const groupedData = { h100: [p1, p2] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-
-    // jOutput lower_left: p1(1,2.0) on front, p2(2,5.0) > 2.0 so NOT on front
-    expect(m1.jOutput!.roof).toBe(true);
-    expect(m2.jOutput!.roof).toBe(false);
-
-    // jInput lower_left: p1(1,10.0) on front, p2(2,4.0) < 10.0 so on front
-    expect(m1.jInput!.roof).toBe(true);
-    expect(m2.jInput!.roof).toBe(true);
-  });
-
-  it('marks costnOutput and costrOutput independently', () => {
-    const chartDef = {
-      chartType: 'e2e',
-      heading: 'Test',
-      x: 'median_e2el',
-      x_label: 'E2E Latency',
-      y: 'tput_per_gpu',
-      y_costnOutput: 'costnOutput.y',
-      y_costnOutput_roofline: 'lower_left',
-      y_costrOutput: 'costrOutput.y',
-      y_costrOutput_roofline: 'lower_left',
-    } as any;
-
-    const p1 = fullPt(1, 'h100', {
-      tpPerGpuY: 50,
-      costnOutputY: 3,
-      costrOutputY: 2,
-    });
-    const p2 = fullPt(2, 'h100', {
-      tpPerGpuY: 80,
-      costnOutputY: 1.5,
-      costrOutputY: 1,
-    });
-    const p3 = fullPt(3, 'h100', {
-      tpPerGpuY: 60,
-      costnOutputY: 2.5,
-      costrOutputY: 1.5,
-    });
-
-    const groupedData = { h100: [p1, p2, p3] };
-    const rooflines = computeAllRooflines(groupedData, chartDef);
-    const marked = markRooflinePoints(groupedData, rooflines, chartDef);
-
-    // lower_left: both costnOutput and costrOutput front = [p1, p2], p3 off front
-    const m1 = marked.find((p) => p.x === 1)!;
-    const m2 = marked.find((p) => p.x === 2)!;
-    const m3 = marked.find((p) => p.x === 3)!;
+    // Total: tokensPerHour = 1000 * 3600 = 3,600,000
+    expect(point.tokensPerDollarH!.y).toBeCloseTo(3_600_000 / 2.8, 5);
+    expect(point.tokensPerDollarN!.y).toBeCloseTo(3_600_000 / 1.4, 5);
+    expect(point.tokensPerDollarR!.y).toBeCloseTo(3_600_000 / 0.7, 5);
 
-    expect(m1.costnOutput!.roof).toBe(true);
-    expect(m2.costnOutput!.roof).toBe(true);
-    expect(m3.costnOutput!.roof).toBe(false);
+    // Output: outputTokensPerHour = 500 * 3600 = 1,800,000
+    expect(point.outputTokensPerDollarH!.y).toBeCloseTo(1_800_000 / 2.8, 5);
+    expect(point.outputTokensPerDollarN!.y).toBeCloseTo(1_800_000 / 1.4, 5);
+    expect(point.outputTokensPerDollarR!.y).toBeCloseTo(1_800_000 / 0.7, 5);
 
-    expect(m1.costrOutput!.roof).toBe(true);
-    expect(m2.costrOutput!.roof).toBe(true);
-    expect(m3.costrOutput!.roof).toBe(false);
+    // Input: inputTokensPerHour = 200 * 3600 = 720,000
+    expect(point.inputTokensPerDollarH!.y).toBeCloseTo(720_000 / 2.8, 5);
+    expect(point.inputTokensPerDollarN!.y).toBeCloseTo(720_000 / 1.4, 5);
+    expect(point.inputTokensPerDollarR!.y).toBeCloseTo(720_000 / 0.7, 5);
   });
 });
 
@@ -2326,7 +1653,7 @@ describe('metricTitle', () => {
     y: 'tput_per_gpu',
     y_tpPerGpu_title: 'Token Throughput per GPU',
     y_tpPerGpu_titleZh: '每 GPU token 吞吐量',
-    y_costh_title: 'Cost per Million Total Tokens (Owning - Hyperscaler)',
+    y_tokensPerDollarH_title: 'Total Tokens per $1 (Owning - Hyperscaler)',
   } as ChartDefinition;
 
   it('returns English title for locale en', () => {
@@ -2338,8 +1665,8 @@ describe('metricTitle', () => {
   });
 
   it('falls back to English when Zh field is missing', () => {
-    expect(metricTitle(chartDef, 'y_costh', 'zh')).toBe(
-      'Cost per Million Total Tokens (Owning - Hyperscaler)',
+    expect(metricTitle(chartDef, 'y_tokensPerDollarH', 'zh')).toBe(
+      'Total Tokens per $1 (Owning - Hyperscaler)',
     );
   });
 
@@ -2357,7 +1684,7 @@ describe('metricLabel', () => {
     y: 'tput_per_gpu',
     y_tpPerGpu_label: 'Token Throughput per GPU (tok/s/gpu)',
     y_tpPerGpu_labelZh: '每 GPU token 吞吐量（tok/s/gpu）',
-    y_costh_label: 'Cost per Million Total Tokens ($)',
+    y_tokensPerDollarH_label: 'Total Tokens per $1 (tok/$)',
   } as ChartDefinition;
 
   it('returns English label for locale en', () => {
@@ -2369,7 +1696,7 @@ describe('metricLabel', () => {
   });
 
   it('falls back to English when Zh field is missing', () => {
-    expect(metricLabel(chartDef, 'y_costh', 'zh')).toBe('Cost per Million Total Tokens ($)');
+    expect(metricLabel(chartDef, 'y_tokensPerDollarH', 'zh')).toBe('Total Tokens per $1 (tok/$)');
   });
 
   it('returns empty string for unknown metric', () => {
diff --git a/packages/app/src/lib/chart-utils.ts b/packages/app/src/lib/chart-utils.ts
index 087b3a114..7d3271656 100644
--- a/packages/app/src/lib/chart-utils.ts
+++ b/packages/app/src/lib/chart-utils.ts
@@ -4,10 +4,19 @@
  * They do NOT import Node.js-specific modules (fs, path) or build-time dependencies.
  */
 
-import { resolveFrameworkAlias } from '@semianalysisai/inferencex-constants';
+import { resolveFrameworkAlias, USD_TO_CNY } from '@semianalysisai/inferencex-constants';
 import iwanthue from 'iwanthue';
 
-import type { AggDataEntry, ChartDefinition, InferenceData } from '@/components/inference/types';
+import type {
+  AggDataEntry,
+  ChartDefinition,
+  InferenceData,
+  YAxisMetricKey,
+} from '@/components/inference/types';
+import {
+  BENCHMARK_METRIC_CONFIG_KEYS,
+  type BenchmarkMetricKey,
+} from '@/components/inference/metric-registry';
 import { getGpuSpecs, isKnownGpu } from '@/lib/constants';
 import { getVendor, type Vendor } from '@/lib/dynamic-colors';
 import type { Locale } from '@/lib/i18n';
@@ -25,6 +34,7 @@ import type { Locale } from '@/lib/i18n';
 const BANNED_HUE_TEST: Record<Vendor, ((hue: number) => boolean) | null> = {
   nvidia: (hue) => hue >= 320 || hue <= 40, // red/rose/pink zone
   amd: (hue) => hue >= 120 && hue <= 195, // green zone
+  teacup: (hue) => hue < 170 || hue > 300, // keep the blue/cyan zone
   unknown: null,
 };
 
@@ -38,6 +48,7 @@ const PREFERRED_ZONE: Record<
 > = {
   nvidia: { hmin: 100, hmax: 195 }, // greens/teals
   amd: { hmin: 20, hmax: 50, cmin: 70, lmin: 50 }, // vivid reds/oranges
+  teacup: { hmin: 190, hmax: 280 }, // cyans/blues
   unknown: null,
 };
 
@@ -160,38 +171,10 @@ export const generateHighContrastColors = (
 };
 
 /**
- * Defines all possible Y-axis metrics that can be used for chart generation,
- * including base metrics and calculated roofline metrics.
+ * Metrics backed by the benchmark transform. Custom user metrics are derived
+ * later from user-entered costs/power and are excluded by the registry.
  */
-export const Y_AXIS_METRICS = [
-  'y',
-  'y_tpPerGpu',
-  'y_inputTputPerGpu',
-  'y_outputTputPerGpu',
-  'y_tpPerMw',
-  'y_inputTputPerMw',
-  'y_outputTputPerMw',
-  'y_costh',
-  'y_costn',
-  'y_costr',
-  'y_costhOutput',
-  'y_costnOutput',
-  'y_costrOutput',
-  'y_costhi',
-  'y_costni',
-  'y_costri',
-  'y_jTotal',
-  'y_jOutput',
-  'y_jInput',
-  // Measured power / energy (sourced from runner's aggregate_power.py output;
-  // distinct from the spec-sheet TDP-derived jTotal/jOutput/jInput above).
-  'y_measuredAvgPower',
-  'y_measuredPrefillAvgPower',
-  'y_measuredDecodeAvgPower',
-  'y_measuredJPerOutputToken',
-  'y_measuredJPerTotalToken',
-  'y_measuredJPerInputToken',
-] as const;
+export const Y_AXIS_METRICS = ['y', ...BENCHMARK_METRIC_CONFIG_KEYS] as const;
 
 export type YAxisMetric = (typeof Y_AXIS_METRICS)[number];
 
@@ -300,48 +283,241 @@ export function buildAvailabilityHwKey(
   return hwKey;
 }
 
+export type DerivedMetricKey = BenchmarkMetricKey;
+export type DerivedChartFields = Pick<InferenceData, DerivedMetricKey>;
+
+const chartMetric = (y: number): { y: number; roof: boolean } => ({ y, roof: false });
+
 /**
- * Creates a single InferenceData point from an AggDataEntry.
- * Spreads all AggDataEntry fields through automatically, then overrides
- * with chart-specific derived fields (coordinates, costs, roofline metrics).
+ * Builds benchmark-derived metric fields. Passing `requestedMetrics` keeps
+ * lightweight consumers lightweight while sharing the exact formulas used by
+ * full inference points.
  */
-export function createChartDataPoint(
-  date: string,
+export function buildDerivedChartFields(
   entry: AggDataEntry,
-  xKey: keyof AggDataEntry,
-  yKey: keyof AggDataEntry,
   currentHwKey: string,
-): InferenceData {
-  const yValue = (entry[yKey] ?? 0) as number;
-  const xValue = (entry[xKey] ?? 0) as number;
+): DerivedChartFields;
+export function buildDerivedChartFields(
+  entry: AggDataEntry,
+  currentHwKey: string,
+  requestedMetrics: readonly DerivedMetricKey[],
+): Partial<DerivedChartFields>;
+export function buildDerivedChartFields(
+  entry: AggDataEntry,
+  currentHwKey: string,
+  requestedMetrics?: readonly DerivedMetricKey[],
+): Partial<DerivedChartFields> {
+  const requested = requestedMetrics ? new Set<DerivedMetricKey>(requestedMetrics) : null;
+  const wants = (key: DerivedMetricKey) => requested === null || requested.has(key);
   const specs = getGpuSpecs(currentHwKey);
   const hardwarePower = specs.power;
   const tputPerGpu = entry.tput_per_gpu ?? 0;
   const outputTputPerGpu = entry.output_tput_per_gpu ?? 0;
   const inputTputPerGpu = entry.input_tput_per_gpu ?? 0;
+  const tokensPerHour = tputPerGpu * 3600;
+  const outputTokensPerHour = outputTputPerGpu * 3600;
+  const inputTokensPerHour = inputTputPerGpu * 3600;
+  const millionTokensPerHour = tokensPerHour / 1_000_000;
+  const millionOutputTokensPerHour = outputTokensPerHour / 1_000_000;
+  const millionInputTokensPerHour = inputTokensPerHour / 1_000_000;
+  const fields: Partial<DerivedChartFields> = {};
+
+  if (wants('tpPerGpu')) fields.tpPerGpu = chartMetric(tputPerGpu);
+  if (wants('outputTputPerGpu') && outputTputPerGpu) {
+    fields.outputTputPerGpu = chartMetric(outputTputPerGpu);
+  }
+  if (wants('inputTputPerGpu') && inputTputPerGpu) {
+    fields.inputTputPerGpu = chartMetric(inputTputPerGpu);
+  }
+  if (wants('tokenRevenuePerGpuHour')) {
+    // At $1 per million total tokens, million tokens per GPU hour is
+    // numerically equal to gross token revenue in $/GPU/hr.
+    fields.tokenRevenuePerGpuHour = chartMetric(millionTokensPerHour);
+  }
+  if (wants('tpPerMw')) fields.tpPerMw = chartMetric((tputPerGpu * 1000) / hardwarePower);
+  if (wants('inputTputPerMw') && inputTputPerGpu) {
+    fields.inputTputPerMw = chartMetric(
+      hardwarePower ? (inputTputPerGpu * 1000) / hardwarePower : 0,
+    );
+  }
+  if (wants('outputTputPerMw') && outputTputPerGpu) {
+    fields.outputTputPerMw = chartMetric(
+      hardwarePower ? (outputTputPerGpu * 1000) / hardwarePower : 0,
+    );
+  }
 
-  const tokensPerHour = (tputPerGpu * 3600) / 1000000;
-  const outputTokensPerHour = (outputTputPerGpu * 3600) / 1000000;
-  const inputTokensPerHour = (inputTputPerGpu * 3600) / 1000000;
+  if (wants('costh')) {
+    fields.costh = chartMetric(
+      hardwarePower && millionTokensPerHour ? specs.costh / millionTokensPerHour : 0,
+    );
+  }
+  if (wants('costn')) {
+    fields.costn = chartMetric(
+      hardwarePower && millionTokensPerHour ? specs.costn / millionTokensPerHour : 0,
+    );
+  }
+  if (wants('costr')) {
+    fields.costr = chartMetric(
+      hardwarePower && millionTokensPerHour ? specs.costr / millionTokensPerHour : 0,
+    );
+  }
+  if (wants('costhOutput')) {
+    fields.costhOutput = chartMetric(
+      hardwarePower && millionOutputTokensPerHour ? specs.costh / millionOutputTokensPerHour : 0,
+    );
+  }
+  if (wants('costnOutput')) {
+    fields.costnOutput = chartMetric(
+      hardwarePower && millionOutputTokensPerHour ? specs.costn / millionOutputTokensPerHour : 0,
+    );
+  }
+  if (wants('costrOutput')) {
+    fields.costrOutput = chartMetric(
+      hardwarePower && millionOutputTokensPerHour ? specs.costr / millionOutputTokensPerHour : 0,
+    );
+  }
+  if (wants('costhi')) {
+    fields.costhi = chartMetric(
+      hardwarePower && millionInputTokensPerHour ? specs.costh / millionInputTokensPerHour : 0,
+    );
+  }
+  if (wants('costni')) {
+    fields.costni = chartMetric(
+      hardwarePower && millionInputTokensPerHour ? specs.costn / millionInputTokensPerHour : 0,
+    );
+  }
+  if (wants('costri')) {
+    fields.costri = chartMetric(
+      hardwarePower && millionInputTokensPerHour ? specs.costr / millionInputTokensPerHour : 0,
+    );
+  }
 
+  if (wants('tokensPerDollarH')) {
+    fields.tokensPerDollarH = chartMetric(specs.costh ? tokensPerHour / specs.costh : 0);
+  }
+  if (wants('tokensPerDollarN')) {
+    fields.tokensPerDollarN = chartMetric(specs.costn ? tokensPerHour / specs.costn : 0);
+  }
+  if (wants('tokensPerDollarR')) {
+    fields.tokensPerDollarR = chartMetric(specs.costr ? tokensPerHour / specs.costr : 0);
+  }
+  if (wants('outputTokensPerDollarH')) {
+    fields.outputTokensPerDollarH = chartMetric(
+      specs.costh ? outputTokensPerHour / specs.costh : 0,
+    );
+  }
+  if (wants('outputTokensPerDollarN')) {
+    fields.outputTokensPerDollarN = chartMetric(
+      specs.costn ? outputTokensPerHour / specs.costn : 0,
+    );
+  }
+  if (wants('outputTokensPerDollarR')) {
+    fields.outputTokensPerDollarR = chartMetric(
+      specs.costr ? outputTokensPerHour / specs.costr : 0,
+    );
+  }
+  if (wants('inputTokensPerDollarH')) {
+    fields.inputTokensPerDollarH = chartMetric(specs.costh ? inputTokensPerHour / specs.costh : 0);
+  }
+  if (wants('inputTokensPerDollarN')) {
+    fields.inputTokensPerDollarN = chartMetric(specs.costn ? inputTokensPerHour / specs.costn : 0);
+  }
+  if (wants('inputTokensPerDollarR')) {
+    fields.inputTokensPerDollarR = chartMetric(specs.costr ? inputTokensPerHour / specs.costr : 0);
+  }
+
+  if (wants('tokensPerRmbH')) {
+    fields.tokensPerRmbH = chartMetric(
+      specs.costh ? tokensPerHour / (specs.costh * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('tokensPerRmbN')) {
+    fields.tokensPerRmbN = chartMetric(
+      specs.costn ? tokensPerHour / (specs.costn * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('tokensPerRmbR')) {
+    fields.tokensPerRmbR = chartMetric(
+      specs.costr ? tokensPerHour / (specs.costr * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('outputTokensPerRmbH')) {
+    fields.outputTokensPerRmbH = chartMetric(
+      specs.costh ? outputTokensPerHour / (specs.costh * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('outputTokensPerRmbN')) {
+    fields.outputTokensPerRmbN = chartMetric(
+      specs.costn ? outputTokensPerHour / (specs.costn * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('outputTokensPerRmbR')) {
+    fields.outputTokensPerRmbR = chartMetric(
+      specs.costr ? outputTokensPerHour / (specs.costr * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('inputTokensPerRmbH')) {
+    fields.inputTokensPerRmbH = chartMetric(
+      specs.costh ? inputTokensPerHour / (specs.costh * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('inputTokensPerRmbN')) {
+    fields.inputTokensPerRmbN = chartMetric(
+      specs.costn ? inputTokensPerHour / (specs.costn * USD_TO_CNY) : 0,
+    );
+  }
+  if (wants('inputTokensPerRmbR')) {
+    fields.inputTokensPerRmbR = chartMetric(
+      specs.costr ? inputTokensPerHour / (specs.costr * USD_TO_CNY) : 0,
+    );
+  }
+
+  if (wants('jTotal')) {
+    fields.jTotal = chartMetric(
+      hardwarePower && tputPerGpu ? (hardwarePower * 1000) / tputPerGpu : 0,
+    );
+  }
+  if (wants('jOutput') && outputTputPerGpu) {
+    fields.jOutput = chartMetric(hardwarePower ? (hardwarePower * 1000) / outputTputPerGpu : 0);
+  }
+  if (wants('jInput') && inputTputPerGpu) {
+    fields.jInput = chartMetric(hardwarePower ? (hardwarePower * 1000) / inputTputPerGpu : 0);
+  }
+
+  const measured = buildMeasuredPowerChartFields(entry, specs.tdp);
+  for (const [key, value] of Object.entries(measured) as [
+    keyof MeasuredPowerChartFields,
+    { y: number; roof: boolean },
+  ][]) {
+    if (wants(key)) fields[key] = value;
+  }
+
+  return fields;
+}
+
+/**
+ * Creates a single InferenceData point from an AggDataEntry.
+ * Spreads all AggDataEntry fields through automatically, then overrides
+ * with chart-specific coordinates and canonical derived fields.
+ */
+export function createChartDataPoint(
+  date: string,
+  entry: AggDataEntry,
+  xKey: keyof AggDataEntry,
+  yKey: keyof AggDataEntry,
+  currentHwKey: string,
+  derivedFields: DerivedChartFields = buildDerivedChartFields(entry, currentHwKey),
+): InferenceData {
   return {
-    // Spread all AggDataEntry fields (raw stats, metadata, etc.)
     ...entry,
-
-    // Chart-specific overrides
     date,
-    x: xValue,
-    y: yValue,
+    x: (entry[xKey] ?? 0) as number,
+    y: (entry[yKey] ?? 0) as number,
     hwKey: currentHwKey,
-    // Total GPU count. Disagg rows carry explicit per-role GPU counts; for
-    // aggregated rows the world size is tp × pp (pp undefined/≤1 ⇒ ×1, so
-    // pre-PP rows are unchanged).
     tp: entry.disagg
       ? entry.num_prefill_gpu + entry.num_decode_gpu
       : entry.tp * (entry.pp && entry.pp > 1 ? entry.pp : 1),
     image: entry.image ?? undefined,
-
-    // Narrow boolean | string fields to boolean
     dp_attention:
       entry.dp_attention !== null && entry.dp_attention !== undefined
         ? entry.dp_attention === true || entry.dp_attention === 'true'
@@ -358,120 +534,80 @@ export function createChartDataPoint(
       entry.is_multinode !== null && entry.is_multinode !== undefined
         ? Boolean(entry.is_multinode)
         : undefined,
-
-    // Disagg fields: only set when active
     disagg: entry.disagg || undefined,
     num_prefill_gpu: entry.disagg ? entry.num_prefill_gpu : undefined,
     num_decode_gpu: entry.disagg ? entry.num_decode_gpu : undefined,
+    ...derivedFields,
+  };
+}
 
-    // Roofline metric fields
-    tpPerGpu: { y: tputPerGpu, roof: false },
-    ...(outputTputPerGpu ? { outputTputPerGpu: { y: outputTputPerGpu, roof: false } } : {}),
-    ...(inputTputPerGpu ? { inputTputPerGpu: { y: inputTputPerGpu, roof: false } } : {}),
-    tpPerMw: { y: (tputPerGpu * 1000) / hardwarePower, roof: false },
-    ...(inputTputPerGpu
-      ? {
-          inputTputPerMw: {
-            y: hardwarePower ? (inputTputPerGpu * 1000) / hardwarePower : 0,
-            roof: false,
-          },
-        }
-      : {}),
-    ...(outputTputPerGpu
-      ? {
-          outputTputPerMw: {
-            y: hardwarePower ? (outputTputPerGpu * 1000) / hardwarePower : 0,
-            roof: false,
-          },
-        }
-      : {}),
-
-    // Cost fields (combined throughput)
-    costh: {
-      y: hardwarePower && tokensPerHour ? specs.costh / tokensPerHour : 0,
-      roof: false,
-    },
-    costn: {
-      y: hardwarePower && tokensPerHour ? specs.costn / tokensPerHour : 0,
-      roof: false,
-    },
-    costr: {
-      y: hardwarePower && tokensPerHour ? specs.costr / tokensPerHour : 0,
-      roof: false,
-    },
-
-    // Cost per million output tokens
-    costhOutput: {
-      y: hardwarePower && outputTokensPerHour ? specs.costh / outputTokensPerHour : 0,
-      roof: false,
-    },
-    costnOutput: {
-      y: hardwarePower && outputTokensPerHour ? specs.costn / outputTokensPerHour : 0,
-      roof: false,
-    },
-    costrOutput: {
-      y: hardwarePower && outputTokensPerHour ? specs.costr / outputTokensPerHour : 0,
-      roof: false,
-    },
-
-    // Cost per million input tokens
-    costhi: {
-      y: hardwarePower && inputTokensPerHour ? specs.costh / inputTokensPerHour : 0,
-      roof: false,
-    },
-    costni: {
-      y: hardwarePower && inputTokensPerHour ? specs.costn / inputTokensPerHour : 0,
-      roof: false,
-    },
-    costri: {
-      y: hardwarePower && inputTokensPerHour ? specs.costr / inputTokensPerHour : 0,
-      roof: false,
-    },
-
-    // All-in provisioned Joules per token: J/token = W/GPU / tok/s/gpu
-    // hardwarePower is in kW, so multiply by 1000 to get watts
-    jTotal: {
-      y: hardwarePower && tputPerGpu ? (hardwarePower * 1000) / tputPerGpu : 0,
-      roof: false,
-    },
-    ...(outputTputPerGpu
-      ? {
-          jOutput: {
-            y: hardwarePower && outputTputPerGpu ? (hardwarePower * 1000) / outputTputPerGpu : 0,
-            roof: false,
-          },
-        }
-      : {}),
-    ...(inputTputPerGpu
-      ? {
-          jInput: {
-            y: hardwarePower && inputTputPerGpu ? (hardwarePower * 1000) / inputTputPerGpu : 0,
-            roof: false,
-          },
-        }
-      : {}),
-
-    // Measured power / energy from runner's aggregate_power.py. Gated on the
-    // raw fields existing so points from runs predating the measurement land
-    // without these keys and the chart correctly filters them out.
+type MeasuredPowerChartFields = Partial<
+  Pick<
+    InferenceData,
+    | 'measuredAvgPower'
+    | 'measuredPrefillAvgPower'
+    | 'measuredDecodeAvgPower'
+    | 'measuredJPerOutputToken'
+    | 'measuredJPerTotalToken'
+    | 'measuredJPerInputToken'
+    | 'measuredJPerSuccessfulQuery'
+    | 'measuredWhPerSuccessfulQuery'
+    | 'measuredPowerPercentTdp'
+  >
+>;
+
+/** Builds optional runner-telemetry fields for the canonical derived builder. */
+function buildMeasuredPowerChartFields(
+  entry: AggDataEntry,
+  tdpWatts: number,
+): MeasuredPowerChartFields {
+  return {
     ...(typeof entry.avg_power_w === 'number'
-      ? { measuredAvgPower: { y: entry.avg_power_w, roof: false } }
+      ? { measuredAvgPower: chartMetric(entry.avg_power_w) }
       : {}),
     ...(typeof entry.prefill_avg_power_w === 'number'
-      ? { measuredPrefillAvgPower: { y: entry.prefill_avg_power_w, roof: false } }
+      ? { measuredPrefillAvgPower: chartMetric(entry.prefill_avg_power_w) }
       : {}),
     ...(typeof entry.decode_avg_power_w === 'number'
-      ? { measuredDecodeAvgPower: { y: entry.decode_avg_power_w, roof: false } }
+      ? { measuredDecodeAvgPower: chartMetric(entry.decode_avg_power_w) }
       : {}),
     ...(typeof entry.joules_per_output_token === 'number'
-      ? { measuredJPerOutputToken: { y: entry.joules_per_output_token, roof: false } }
+      ? { measuredJPerOutputToken: chartMetric(entry.joules_per_output_token) }
       : {}),
     ...(typeof entry.joules_per_total_token === 'number'
-      ? { measuredJPerTotalToken: { y: entry.joules_per_total_token, roof: false } }
+      ? { measuredJPerTotalToken: chartMetric(entry.joules_per_total_token) }
       : {}),
     ...(typeof entry.joules_per_input_token === 'number'
-      ? { measuredJPerInputToken: { y: entry.joules_per_input_token, roof: false } }
+      ? { measuredJPerInputToken: chartMetric(entry.joules_per_input_token) }
+      : {}),
+    ...(typeof entry.joules_per_successful_query === 'number'
+      ? {
+          measuredJPerSuccessfulQuery: chartMetric(entry.joules_per_successful_query),
+          measuredWhPerSuccessfulQuery: chartMetric(entry.joules_per_successful_query / 3600),
+        }
       : {}),
+    ...(typeof entry.avg_power_w === 'number' && tdpWatts > 0
+      ? { measuredPowerPercentTdp: chartMetric((entry.avg_power_w / tdpWatts) * 100) }
+      : {}),
+  };
+}
+/**
+ * Remaps a chart-ready point onto a resolved metric and x-axis field. Official
+ * and overlay pipelines share this coordinate step while retaining separate
+ * clipping and frontier policies.
+ */
+export function remapInferencePoint(
+  point: InferenceData,
+  metricKey: YAxisMetricKey,
+  xAxisField: keyof AggDataEntry,
+): InferenceData {
+  const metric = point[metricKey];
+  const xCandidate = (point as Partial<AggDataEntry>)[xAxisField];
+  return {
+    ...point,
+    x: typeof xCandidate === 'number' ? xCandidate : point.x,
+    y: metric?.y ?? point.y,
+    roof: metric?.roof ?? false,
   };
 }
 
@@ -633,249 +769,6 @@ export const paretoFrontForDirection = (
   dir: ParetoDirection,
 ): ((points: InferenceData[]) => InferenceData[]) => PARETO_BY_DIRECTION[dir];
 
-/**
- * Calculates the roofline for a given set of points.
- */
-export const calculateRoofline = (
-  points: InferenceData[],
-  yKey:
-    | keyof InferenceData
-    | `tpPerGpu.y`
-    | `outputTputPerGpu.y`
-    | `inputTputPerGpu.y`
-    | `tpPerMw.y`
-    | `inputTputPerMw.y`
-    | `outputTputPerMw.y`
-    | `costh.y`
-    | `costn.y`
-    | `costr.y`
-    | `costhOutput.y`
-    | `costnOutput.y`
-    | `costrOutput.y`
-    | `costhi.y`
-    | `costni.y`
-    | `costri.y`
-    | `jTotal.y`
-    | `jOutput.y`
-    | `jInput.y`
-    | `measuredAvgPower.y`
-    | `measuredPrefillAvgPower.y`
-    | `measuredDecodeAvgPower.y`
-    | `measuredJPerOutputToken.y`
-    | `measuredJPerTotalToken.y`
-    | `measuredJPerInputToken.y`,
-  rooflineDirection: 'upper_right' | 'upper_left' | 'lower_left' | 'lower_right',
-): InferenceData[] => {
-  // Exclude degenerate x <= 0 points (see isFrontierEligible) so they never
-  // anchor a roofline or show up in the optimal-only view.
-  const pointsForRoofline = points.filter(isFrontierEligible).map((p) => {
-    const yValue = getNestedYValue(p, yKey);
-    return { ...p, y: yValue };
-  });
-
-  switch (rooflineDirection) {
-    case 'upper_right': {
-      return paretoFrontUpperRight(pointsForRoofline);
-    }
-    case 'upper_left': {
-      return paretoFrontUpperLeft(pointsForRoofline);
-    }
-    case 'lower_left': {
-      return paretoFrontLowerLeft(pointsForRoofline);
-    }
-    case 'lower_right': {
-      return paretoFrontLowerRight(pointsForRoofline);
-    }
-    default: {
-      return [];
-    }
-  }
-};
-
-/**
- * Computes all relevant rooflines for a given set of grouped data points.
- */
-export function computeAllRooflines(
-  groupedData: Record<string, InferenceData[]>,
-  chartDef: ChartDefinition,
-): Record<string, Record<YAxisMetric, InferenceData[]>> {
-  const computedRooflines: Record<string, Record<YAxisMetric, InferenceData[]>> = {};
-
-  for (const hw of Object.keys(groupedData)) {
-    computedRooflines[hw] = {} as Record<YAxisMetric, InferenceData[]>;
-    for (const chartDefYKey of Y_AXIS_METRICS) {
-      const actualDataYKey = chartDef[chartDefYKey as keyof ChartDefinition];
-      const rooflineDirectionKey = `${chartDefYKey}_roofline` as keyof ChartDefinition;
-      const rooflineDirection = chartDef[rooflineDirectionKey] as
-        | 'upper_right'
-        | 'upper_left'
-        | 'lower_left'
-        | 'lower_right'
-        | undefined;
-
-      if (actualDataYKey && rooflineDirection) {
-        computedRooflines[hw][chartDefYKey] = calculateRoofline(
-          groupedData[hw],
-          actualDataYKey as
-            | keyof InferenceData
-            | `tpPerGpu.y`
-            | `outputTputPerGpu.y`
-            | `inputTputPerGpu.y`
-            | `tpPerMw.y`
-            | `inputTputPerMw.y`
-            | `outputTputPerMw.y`
-            | `costh.y`
-            | `costn.y`
-            | `costr.y`
-            | `costhOutput.y`
-            | `costnOutput.y`
-            | `costrOutput.y`
-            | `costhi.y`
-            | `costni.y`
-            | `costri.y`
-            | `jTotal.y`
-            | `jOutput.y`
-            | `jInput.y`
-            | `measuredAvgPower.y`
-            | `measuredPrefillAvgPower.y`
-            | `measuredDecodeAvgPower.y`
-            | `measuredJPerOutputToken.y`
-            | `measuredJPerTotalToken.y`
-            | `measuredJPerInputToken.y`,
-          rooflineDirection,
-        );
-      }
-    }
-  }
-  return computedRooflines;
-}
-
-/**
- * Marks data points as being "on the roofline".
- */
-export function markRooflinePoints(
-  groupedData: Record<string, InferenceData[]>,
-  computedRooflines: Record<string, Record<YAxisMetric, InferenceData[]>>,
-  chartDef: ChartDefinition,
-): InferenceData[] {
-  const finalProcessedData: InferenceData[] = [];
-
-  for (const hwKey of Object.keys(groupedData)) {
-    for (const point of groupedData[hwKey]) {
-      const newPoint = { ...point };
-      newPoint.tpPerGpu.roof = false;
-      if (newPoint.outputTputPerGpu) {
-        newPoint.outputTputPerGpu.roof = false;
-      }
-      if (newPoint.inputTputPerGpu) {
-        newPoint.inputTputPerGpu.roof = false;
-      }
-      newPoint.tpPerMw.roof = false;
-      if (newPoint.inputTputPerMw) newPoint.inputTputPerMw.roof = false;
-      if (newPoint.outputTputPerMw) newPoint.outputTputPerMw.roof = false;
-      newPoint.costh.roof = false;
-      newPoint.costn.roof = false;
-      newPoint.costr.roof = false;
-      if (newPoint.costhOutput) newPoint.costhOutput.roof = false;
-      if (newPoint.costnOutput) newPoint.costnOutput.roof = false;
-      if (newPoint.costrOutput) newPoint.costrOutput.roof = false;
-      newPoint.costhi.roof = false;
-      newPoint.costni.roof = false;
-      newPoint.costri.roof = false;
-      if (newPoint.jTotal) newPoint.jTotal.roof = false;
-      if (newPoint.jOutput) newPoint.jOutput.roof = false;
-      if (newPoint.jInput) newPoint.jInput.roof = false;
-      if (newPoint.measuredAvgPower) newPoint.measuredAvgPower.roof = false;
-      if (newPoint.measuredPrefillAvgPower) newPoint.measuredPrefillAvgPower.roof = false;
-      if (newPoint.measuredDecodeAvgPower) newPoint.measuredDecodeAvgPower.roof = false;
-      if (newPoint.measuredJPerOutputToken) newPoint.measuredJPerOutputToken.roof = false;
-      if (newPoint.measuredJPerTotalToken) newPoint.measuredJPerTotalToken.roof = false;
-      if (newPoint.measuredJPerInputToken) newPoint.measuredJPerInputToken.roof = false;
-
-      for (const chartDefYKey of Y_AXIS_METRICS) {
-        const rooflinePoints = computedRooflines[hwKey]?.[chartDefYKey];
-        if (!rooflinePoints) {
-          continue;
-        }
-
-        const actualDataYKey = chartDef[chartDefYKey as keyof ChartDefinition];
-        if (!actualDataYKey) {
-          continue;
-        }
-
-        const onCurrentRoofline = rooflinePoints.some(
-          (rooflinePoint) =>
-            rooflinePoint.x === point.x &&
-            rooflinePoint.y === getNestedYValue(point, actualDataYKey as string) &&
-            rooflinePoint.hwKey === point.hwKey,
-        );
-
-        if (chartDefYKey === 'y_tpPerGpu') {
-          newPoint.tpPerGpu.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_outputTputPerGpu') {
-          if (newPoint.outputTputPerGpu) {
-            newPoint.outputTputPerGpu.roof = onCurrentRoofline;
-          }
-        } else if (chartDefYKey === 'y_inputTputPerGpu') {
-          if (newPoint.inputTputPerGpu) {
-            newPoint.inputTputPerGpu.roof = onCurrentRoofline;
-          }
-        } else if (chartDefYKey === 'y_tpPerMw') {
-          newPoint.tpPerMw.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_inputTputPerMw') {
-          if (newPoint.inputTputPerMw) newPoint.inputTputPerMw.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_outputTputPerMw') {
-          if (newPoint.outputTputPerMw) newPoint.outputTputPerMw.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costh') {
-          newPoint.costh.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costn') {
-          newPoint.costn.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costr') {
-          newPoint.costr.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costhOutput') {
-          if (newPoint.costhOutput) newPoint.costhOutput.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costnOutput') {
-          if (newPoint.costnOutput) newPoint.costnOutput.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costrOutput') {
-          if (newPoint.costrOutput) newPoint.costrOutput.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costhi') {
-          newPoint.costhi.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costni') {
-          newPoint.costni.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_costri') {
-          newPoint.costri.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_jTotal' && newPoint.jTotal) {
-          newPoint.jTotal.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_jOutput' && newPoint.jOutput) {
-          newPoint.jOutput.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_jInput' && newPoint.jInput) {
-          newPoint.jInput.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_measuredAvgPower' && newPoint.measuredAvgPower) {
-          newPoint.measuredAvgPower.roof = onCurrentRoofline;
-        } else if (
-          chartDefYKey === 'y_measuredPrefillAvgPower' &&
-          newPoint.measuredPrefillAvgPower
-        ) {
-          newPoint.measuredPrefillAvgPower.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_measuredDecodeAvgPower' && newPoint.measuredDecodeAvgPower) {
-          newPoint.measuredDecodeAvgPower.roof = onCurrentRoofline;
-        } else if (
-          chartDefYKey === 'y_measuredJPerOutputToken' &&
-          newPoint.measuredJPerOutputToken
-        ) {
-          newPoint.measuredJPerOutputToken.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_measuredJPerTotalToken' && newPoint.measuredJPerTotalToken) {
-          newPoint.measuredJPerTotalToken.roof = onCurrentRoofline;
-        } else if (chartDefYKey === 'y_measuredJPerInputToken' && newPoint.measuredJPerInputToken) {
-          newPoint.measuredJPerInputToken.roof = onCurrentRoofline;
-        }
-      }
-      finalProcessedData.push(newPoint);
-    }
-  }
-  return finalProcessedData;
-}
-
 // ---------------------------------------------------------------------------
 // Locale-aware metric label/title helpers
 // ---------------------------------------------------------------------------
diff --git a/packages/app/src/lib/chip-pages-zh.ts b/packages/app/src/lib/chip-pages-zh.ts
new file mode 100644
index 000000000..50c9c18cf
--- /dev/null
+++ b/packages/app/src/lib/chip-pages-zh.ts
@@ -0,0 +1,311 @@
+/**
+ * Simplified Chinese content for the /zh/chips page family.
+ *
+ * 1:1 sibling of `chip-pages.ts` (see AGENTS.md "Chinese Website Pages" —
+ * every English change there must be mirrored here in the same PR). Numbers
+ * are generated from the same registries, so only prose is translated.
+ */
+import {
+  buildChipVsHighlights,
+  type ChipFaqItem,
+  type ChipPageEntry,
+  type ChipVsHighlight,
+  type ChipVsPage,
+  getAllChipPages,
+  getChipHw,
+  getChipPage,
+  getChipSpec,
+  leadingNumber,
+} from './chip-pages';
+
+export interface ChipPageTranslation {
+  /** Chinese meta-description base. */
+  summary: string;
+  /** Translated overview paragraphs (same count as English). */
+  overview: readonly string[];
+  benchmarkContext: string;
+  /** Chinese-searcher keywords (SKUs stay in English). */
+  keywords: readonly string[];
+}
+
+const translations: Readonly<Record<string, ChipPageTranslation>> = {
+  h100: {
+    summary:
+      'NVIDIA H100 规格、云端价格与实时 LLM 推理基准测试：80 GB HBM3、3.35 TB/s 内存带宽、FP8 Tensor Core、NVLink 4.0，每日在 vLLM、SGLang 与 TensorRT-LLM 上持续测量。',
+    overview: [
+      'NVIDIA H100 SXM 是 Hopper 一代的数据中心加速器，支撑了第一波大规模 LLM 部署。每颗芯片配备 80 GB HBM3（带宽 3.35 TB/s）和第四代 Tensor Core，稠密 FP8 算力达 1,979 TFLOP/s；8 颗芯片通过 NVLink 4.0 组成节点，单芯片互联带宽 450 GB/s。',
+      'H100 仍是衡量其他加速器的基准线：它拥有所有 AI 芯片中最成熟的软件生态、最广泛的云端供应，以及 NVIDIA 产品线中最低的小时租价。它没有 FP4 Tensor Core，因此在适用 NVFP4 的量化推理场景中会落后于更新的 Blackwell 产品。',
+    ],
+    benchmarkContext:
+      'InferenceX 在 vLLM、SGLang 和 TensorRT-LLM 上对 H100 进行持续测试，覆盖固定序列服务与长上下文 agentic trace，发布吞吐量-交互性 Pareto 前沿、每百万 token 成本和每 token 能耗，并与每一代新芯片并列展示，让升级收益始终有数可查。',
+    keywords: [
+      'NVIDIA H100',
+      'H100 规格',
+      'H100 每小时价格',
+      'H100 推理基准测试',
+      'H100 FP8 算力',
+      'H100 内存带宽',
+      'Hopper 芯片',
+      'H100 对比 B200',
+    ],
+  },
+  h200: {
+    summary:
+      'NVIDIA H200 规格、云端价格与实时 LLM 推理基准测试：141 GB HBM3e、4.8 TB/s 内存带宽、FP8 Tensor Core 与 NVLink 4.0，每日与 Blackwell 及 AMD Instinct 对比测量。',
+    overview: [
+      'NVIDIA H200 SXM 是 Hopper 的大内存版本：稠密 FP8 算力与 H100 相同（1,979 TFLOP/s），但配备 141 GB HBM3e，带宽 4.8 TB/s。大 76% 的显存和高 43% 的带宽对长上下文服务和大 KV cache 最为关键，而这恰恰是 H100 最先遇到瓶颈的场景。',
+      '在内存受限的 decode 工作负载中，H200 以相近的小时租价提供明显高于 H100 的单芯片 token 吞吐量，这也是 Blackwell 上市后它仍是 Hopper 出货主力的原因。与 H100 一样，它不支持 FP4，推理精度以 FP8 和 INT4 仅权重量化为主。',
+    ],
+    benchmarkContext:
+      'InferenceX 每日在 vLLM、SGLang 和 TensorRT-LLM 上对 H200 进行基准测试，对比页面将 H200 FP8 与 B200 NVFP4、AMD MI325X 并列呈现，让 Hopper 到 Blackwell、NVIDIA 到 AMD 的取舍以实测数据为准。',
+    keywords: [
+      'NVIDIA H200',
+      'H200 规格',
+      'H200 每小时价格',
+      'H200 推理基准测试',
+      'H200 内存带宽',
+      'H200 对比 H100',
+      'H200 对比 B200',
+      'HBM3e 芯片',
+    ],
+  },
+  b200: {
+    summary:
+      'NVIDIA B200 规格、云端价格与实时 LLM 推理基准测试：180 GB HBM3e、8 TB/s 带宽、9,000 稠密 FP4 TFLOP/s 与 NVLink 5.0，每日在 vLLM、SGLang 与 TensorRT-LLM 上持续测量。',
+    overview: [
+      'NVIDIA B200 是 Blackwell 一代的主力数据中心芯片：180 GB 可用 HBM3e（带宽 8 TB/s）、单芯片 900 GB/s 的 NVLink 5.0，Tensor Core 每时钟吞吐量翻倍于 Hopper 并新增 FP4。稠密算力 9,000 FP4 / 4,500 FP8 TFLOP/s，在主导 LLM decode 的内存带宽受限场景中，单颗 B200 的表现超过两颗 H100。',
+      'NVFP4 推理正是在 B200 上成为主流：前沿开源模型发布 FP4 checkpoint，精度与 FP8 相差无几，单芯片吞吐量却接近翻倍。它 1,000 W 的 TDP 和约 1.7 kW 的整机功耗使小时租价远高于 Hopper，因此升级决策取决于每美元性能而非峰值算力。',
+    ],
+    benchmarkContext:
+      'B200 是 InferenceX 上测试最密集的芯片之一：每日在 vLLM、SGLang、TensorRT-LLM 和 Dynamo 分离式推理上运行固定序列扫描与 AgentX agentic 编码 trace，每美元性能与精度对比页面持续追踪每个模型上 NVFP4 与 FP8 的差距。',
+    keywords: [
+      'NVIDIA B200',
+      'B200 规格',
+      'B200 每小时价格',
+      'B200 推理基准测试',
+      'B200 NVFP4',
+      'B200 对比 H100',
+      'B200 对比 MI355X',
+      'Blackwell 芯片规格',
+    ],
+  },
+  b300: {
+    summary:
+      'NVIDIA B300（Blackwell Ultra）规格、云端价格与实时 LLM 推理基准测试：268 GB 可用 HBM3e、13,500 稠密 FP4 TFLOP/s、800 Gbit/s scale-out 网络，每日与 B200、GB300 NVL72 和 MI355X 对比测量。',
+    overview: [
+      'NVIDIA B300 是 Blackwell Ultra 的 SXM 形态：268 GB 可用 HBM3e（带宽 8 TB/s），稠密 FP4 算力 13,500 TFLOP/s，比 B200 高 1.5 倍，FP8 与 BF16 则保持 B200 水平。对推理而言更重要的是大 49% 的显存：同样的 8 芯片节点能容纳更大的 KV 工作集和更大的模型。',
+      'B300 还通过 ConnectX-8 将单芯片 scale-out 带宽翻倍至 800 Gbit/s，这对 prefill/decode 分离和 wide expert parallelism 尤为重要。它定位于 B200 与机柜级 GB300 NVL72 之间：NVLink 5.0 域规模同为 8，但拥有 Ultra 级的显存和 FP4 算力。',
+    ],
+    benchmarkContext:
+      'InferenceX 以与 B200 相同的每日节奏在 vLLM、SGLang 和 TensorRT-LLM 上测试 B300，包括 AgentX 长上下文 agentic trace：268 GB HBM3e 能让 KV 工作集常驻显存，而较小的芯片则被迫抢占或 offload。',
+    keywords: [
+      'NVIDIA B300',
+      'B300 规格',
+      'Blackwell Ultra',
+      'B300 每小时价格',
+      'B300 推理基准测试',
+      'B300 FP4 算力',
+      'B300 对比 B200',
+      'B300 显存',
+    ],
+  },
+  'gb200-nvl72': {
+    summary:
+      'NVIDIA GB200 NVL72 规格、整机柜价格与实时 LLM 推理基准测试：72 颗 Blackwell 芯片组成单一 NVLink 域，单芯片 186 GB HBM3e、900 GB/s scale-up 带宽，每日在分离式 vLLM、SGLang 与 Dynamo TRT-LLM 上测量。',
+    overview: [
+      'NVIDIA GB200 NVL72 是机柜级系统：72 颗 Blackwell 芯片与 36 颗 Grace CPU 通过第五代 NVLink 组成单一 72 芯片域，单芯片单向带宽 900 GB/s。机柜内部没有 scale-out 网络，因为机柜本身就是计算单元，汇聚约 13.4 TB HBM3e。',
+      '这种一跳全互联的 fabric 正是 GB200 NVL72 在大型 MoE 模型 wide expert parallelism 推理中占优的原因：专家层可以分布到数十颗芯片上而无需经过 InfiniBand 或以太网。Dynamo、SGLang PD 等 prefill/decode 分离方案能自然映射到这个域上。',
+    ],
+    benchmarkContext:
+      'InferenceX 以多节点分离式配置测试 GB200 NVL72，包括前沿 MoE 模型上的 wide-EP vLLM、SGLang 与 Dynamo TRT-LLM 运行，并将结果归一化为单芯片吞吐量，使机柜级数据可与 8 芯片 HGX 节点直接对比。',
+    keywords: [
+      'NVIDIA GB200 NVL72',
+      'GB200 NVL72 规格',
+      'GB200 价格',
+      'GB200 NVL72 基准测试',
+      'NVL72 机柜',
+      'GB200 对比 B200',
+      '机柜级 AI 系统',
+      'GB200 NVL72 功耗',
+    ],
+  },
+  'gb300-nvl72': {
+    summary:
+      'NVIDIA GB300 NVL72 规格、整机柜价格与实时 LLM 推理基准测试：72 颗 Blackwell Ultra 芯片，单芯片 278 GB HBM3e（整柜约 20 TB）、15,000 稠密 FP4 TFLOP/s，每日在分离式推理方案上测量。',
+    overview: [
+      'NVIDIA GB300 NVL72 是 Blackwell Ultra 机柜：72 颗芯片各配 278 GB 可用 HBM3e，整柜约 20 TB 显存，单芯片稠密 FP4 算力 15,000 TFLOP/s，NVLink 5.0 域与 GB200 NVL72 相同为 72 路。FP8 与 BF16 吞吐量沿用 GB200，Ultra 的提升集中在 FP4 和显存。',
+      '单芯片显存的增加在机柜尺度上被进一步放大：更大的前沿 MoE 模型、更长的上下文和更大的 KV 工作集都能常驻而不外溢；相应地每颗芯片 TDP 提高到 1,400 W 以支撑更大的 HBM 堆栈。在 Vera Rubin 之前，GB300 NVL72 是 NVIDIA 推理产品线的当前顶点。',
+    ],
+    benchmarkContext:
+      'InferenceX 在前沿 MoE 模型上以分离式与 wide-EP 配置测试 GB300 NVL72；AgentX agentic 编码基准将这个机柜作为参考上限，用来衡量 AMD MI355X ATOM 结果与更小的 NVIDIA 节点。',
+    keywords: [
+      'NVIDIA GB300 NVL72',
+      'GB300 NVL72 规格',
+      'GB300 价格',
+      'GB300 NVL72 基准测试',
+      'Blackwell Ultra 机柜',
+      'GB300 对比 GB200',
+      'GB300 FP4 算力',
+      'GB300 NVL72 功耗',
+    ],
+  },
+  mi300x: {
+    summary:
+      'AMD Instinct MI300X 规格、云端价格与实时 LLM 推理基准测试：192 GB HBM3、5.3 TB/s 带宽、2,615 稠密 FP8 TFLOP/s、全互联 Infinity Fabric，每日在 ROCm 版 vLLM 与 SGLang 上测量。',
+    overview: [
+      'AMD Instinct MI300X 是让 AMD 真正跻身 LLM 推理市场的 CDNA 3 加速器：192 GB HBM3（带宽 5.3 TB/s），显存超过 H100 的两倍，稠密 FP8 算力 2,615 TFLOP/s。8 颗芯片通过 Infinity Fabric 全互联组网，无需交换机，单芯片 scale-up 带宽 448 GB/s。',
+      '显存优势让 MI300X 能用更少的芯片部署模型，或在每个副本中容纳更大的 KV cache，小时租价也明显低于同级 NVIDIA 产品。软件是历史上的短板：ROCm 版 vLLM 和 SGLang 已大幅缩小差距，而差距究竟缩小了多少，公开基准测试记录里一目了然。',
+    ],
+    benchmarkContext:
+      'MI300X 是最早加入 InferenceMAX 的芯片之一，至今仍在 ROCm vLLM 与 SGLang 的持续扫描中运行，留下了所有加速器中最长的公开软件进步曲线之一：同一颗芯片，随着 kernel 与调度器的改进被连续测量了数月。',
+    keywords: [
+      'AMD MI300X',
+      'MI300X 规格',
+      'MI300X 每小时价格',
+      'MI300X 推理基准测试',
+      'MI300X 对比 H100',
+      'AMD Instinct 芯片',
+      'CDNA 3',
+      'ROCm 推理',
+    ],
+  },
+  mi325x: {
+    summary:
+      'AMD Instinct MI325X 规格、云端价格与实时 LLM 推理基准测试：256 GB HBM3e、6 TB/s 带宽、2,615 稠密 FP8 TFLOP/s、全互联 Infinity Fabric，每日在 ROCm 版 vLLM 与 SGLang 上测量。',
+    overview: [
+      'AMD Instinct MI325X 是 CDNA 3 的大内存版本：稠密 FP8 算力与 MI300X 相同（2,615 TFLOP/s），但配备 256 GB HBM3e，带宽 6 TB/s，是同代产品中最大的显存。一个 8 芯片全互联节点拥有超过 2 TB HBM，足以在节点内部署超大模型。',
+      'MI325X 以容量经济性对标 H200：更大的单芯片显存和更低的小时租价，换取的是 NVIDIA 软件生态的成熟度。对内存受限的 decode 和长上下文服务而言，它的单位价格带宽在 CDNA 4 之前的产品中名列前茅。',
+    ],
+    benchmarkContext:
+      'InferenceX 在 ROCm vLLM 与 SGLang 上以共享模型集测试 MI325X，对比页面将它与 H200、MI355X 并列，使 NVIDIA 与 AMD 之间、代际之间的差距都处于持续测量之下。',
+    keywords: [
+      'AMD MI325X',
+      'MI325X 规格',
+      'MI325X 每小时价格',
+      'MI325X 推理基准测试',
+      'MI325X 对比 H200',
+      'MI325X 显存',
+      'CDNA 3 芯片',
+      'MI325X 对比 MI300X',
+    ],
+  },
+  mi355x: {
+    summary:
+      'AMD Instinct MI355X 规格、云端价格与实时 LLM 推理基准测试：288 GB HBM3e、8 TB/s 带宽、10,066 稠密 FP4 TFLOP/s，AMD 首款支持 FP4 的芯片，每日在 ROCm 版 vLLM、SGLang 与 ATOM 上测量。',
+    overview: [
+      'AMD Instinct MI355X 是 CDNA 4 旗舰，也是 AMD 首款具备 FP4 Tensor 吞吐能力的加速器：稠密算力 10,066 FP4 / 5,033 FP8 TFLOP/s，配备 288 GB HBM3e（带宽 8 TB/s），单芯片显存为当前业界最大。8 颗芯片通过第五代 Infinity Fabric 全互联组网，单芯片带宽 538 GB/s。',
+      'MI355X 在带宽上对齐 B200 级别、显存反超，而小时租价明显更低，是 AMD 迄今在每美元性能上发起的最有力挑战。1,400 W 的 TDP 和 ROCm 推理软件栈的成熟度则是实时数据持续量化的两项代价。',
+    ],
+    benchmarkContext:
+      'MI355X 是 InferenceX 上追踪最密集的 AMD 芯片：每日运行 vLLM、SGLang 与 ATOM，AgentX agentic 编码 trace 与 GB300 NVL72、B300 直接对垒，并留下了最快的公开软件进步曲线之一，包括模型发布数周内的数量级提升。',
+    keywords: [
+      'AMD MI355X',
+      'MI355X 规格',
+      'MI355X 每小时价格',
+      'MI355X 推理基准测试',
+      'MI355X 对比 B200',
+      'MI355X FP4',
+      'CDNA 4 芯片',
+      'MI355X 内存带宽',
+    ],
+  },
+};
+
+export function getZhChipTranslation(slug: string): ChipPageTranslation | undefined {
+  return translations[slug];
+}
+
+export function getAllZhChipSlugs(): readonly string[] {
+  return Object.keys(translations);
+}
+
+/** Localized labels for the data-derived versus highlight rows. */
+export const CHIP_VS_HIGHLIGHT_LABELS_ZH: Readonly<Record<ChipVsHighlight['key'], string>> = {
+  memory: '单芯片显存',
+  memoryBandwidth: '内存带宽',
+  fp8: '稠密 FP8 算力',
+  fp4: '稠密 FP4 算力',
+  tdp: 'TDP',
+  costNeocloud: '小时租价（neocloud 档）',
+  scaleUpWorldSize: 'Scale-up 域规模',
+};
+
+const NOT_SUPPORTED_ZH = '不支持';
+const CHIP_COUNT_PATTERN = /^(?<count>\d+) chips$/u;
+
+/** Localize the generated English cell values ("Not supported", "72 chips") for zh rendering. */
+export function localizeVsHighlightValueZh(value: string): string {
+  if (value === 'Not supported') return NOT_SUPPORTED_ZH;
+  const chipCount = CHIP_COUNT_PATTERN.exec(value)?.groups?.count;
+  if (chipCount) return `${chipCount} 芯片`;
+  return value;
+}
+
+/** Chinese FAQ, generated from the same registries as the English one. */
+export function buildZhChipFaq(entry: ChipPageEntry): readonly ChipFaqItem[] {
+  const spec = getChipSpec(entry);
+  const hw = getChipHw(entry);
+  const domainGb = Math.round(leadingNumber(spec.memory) * spec.scaleUpWorldSize).toLocaleString(
+    'en-US',
+  );
+  return [
+    {
+      question: `${entry.label} 云端租用每小时多少钱？`,
+      answer: `按 SemiAnalysis AI Cloud TCO 模型，${entry.label} 在超大规模云约 $${hw.costh.toFixed(2)}/小时，neocloud 档约 $${hw.costn.toFixed(2)}/小时，零售档约 $${hw.costr.toFixed(2)}/小时。InferenceX 的每美元性能页面即使用这些费率将实测吞吐量换算为每百万 token 成本。`,
+    },
+    {
+      question: `${entry.label} 有多少显存？`,
+      answer: `${entry.label} 单芯片提供 ${spec.memory} 可用 ${spec.memoryType}，内存带宽 ${spec.memoryBandwidth}；一个 ${spec.scaleUpWorldSize} 芯片的 ${spec.scaleUpTech} 域合计约 ${domainGb} GB。`,
+    },
+    {
+      question: `${entry.label} 的功耗是多少？`,
+      answer: `${entry.label} 单芯片 TDP 为 ${hw.tdp.toLocaleString('en-US')} W，计入主机 CPU、网卡与散热分摊后整机约 ${hw.power} kW。InferenceX 的每 token 能耗计算采用整机功耗口径。`,
+    },
+    {
+      question: `${entry.label} 支持 FP4 吗？`,
+      answer: spec.fp4
+        ? `支持。${entry.label} 稠密 FP4 算力达 ${spec.fp4.toLocaleString('en-US')} TFLOP/s（FP8 为 ${spec.fp8.toLocaleString('en-US')}），InferenceX 的精度对比页面持续追踪 FP4 与 FP8 的推理精度和吞吐量差异。`
+        : `不支持。${entry.label} 最高支持 FP8，稠密算力 ${spec.fp8.toLocaleString('en-US')} TFLOP/s；FP4 推理需要更新一代的芯片。`,
+    },
+    {
+      question: `${entry.label} 的 LLM 推理速度有多快？`,
+      answer: `这取决于模型、推理引擎、精度和交互性目标，因此 InferenceX 为 ${entry.label} 发布持续更新的吞吐量-交互性 Pareto 前沿，而非单一数字。实时仪表板和对比页面展示每个覆盖模型上的最新结果。`,
+    },
+  ];
+}
+
+/** Chinese FAQ for a versus page. */
+export function buildZhChipVsFaq(page: ChipVsPage): readonly ChipFaqItem[] {
+  const highlights = buildChipVsHighlights(page);
+  const memory = highlights.find((h) => h.key === 'memory');
+  const cost = highlights.find((h) => h.key === 'costNeocloud');
+  const fp8 = highlights.find((h) => h.key === 'fp8');
+  return [
+    {
+      question: `${page.a.label} 和 ${page.b.label} 谁的显存更大？`,
+      answer: `${page.a.label} 单芯片为 ${memory?.aValue}，${page.b.label} 为 ${memory?.bValue}，容量比约 ${memory?.ratio}。`,
+    },
+    {
+      question: `${page.a.label} 和 ${page.b.label} 的价格怎么比？`,
+      answer: `按 SemiAnalysis TCO 模型 neocloud 档费率，${page.a.label} 约 ${cost?.aValue}，${page.b.label} 约 ${cost?.bValue}。单看小时租价容易误导：每美元性能对比页面会用实测吞吐量除以这些费率。`,
+    },
+    {
+      question: `${page.a.label} 的 LLM 推理速度比 ${page.b.label} 快吗？`,
+      answer: `纸面上 ${page.a.label} 的稠密 FP8 算力是 ${page.b.label} 的 ${fp8?.ratio}，但实际交付的 token 吞吐量取决于模型、推理引擎、精度和交互性目标。InferenceX 每日在完全相同的工作负载上测量这两款芯片，最新结果请见实时对比页面。`,
+    },
+  ];
+}
+
+/** Parity guard used by tests: every English chip page must have a zh translation. */
+export function assertZhChipParity(): void {
+  for (const entry of getAllChipPages()) {
+    const translation = translations[entry.slug];
+    if (!translation) throw new Error(`Missing zh translation for chip page ${entry.slug}`);
+    if (translation.overview.length !== entry.overview.length) {
+      throw new Error(`zh overview paragraph count mismatch for ${entry.slug}`);
+    }
+  }
+  for (const slug of Object.keys(translations)) {
+    if (!getChipPage(slug)) throw new Error(`zh translation for unknown chip page ${slug}`);
+  }
+}
diff --git a/packages/app/src/lib/chip-pages.test.ts b/packages/app/src/lib/chip-pages.test.ts
new file mode 100644
index 000000000..ac0a1fca5
--- /dev/null
+++ b/packages/app/src/lib/chip-pages.test.ts
@@ -0,0 +1,176 @@
+import { describe, expect, it } from 'vitest';
+
+import { GPU_SPECS } from '@/lib/gpu-specs';
+import { getPostBySlug } from '@/lib/blog';
+import { getGlossaryEntry } from '@/lib/glossary';
+import { HW_REGISTRY } from '@semianalysisai/inferencex-constants';
+
+import {
+  buildChipFaq,
+  buildChipVsFaq,
+  buildChipVsHighlights,
+  CHIP_VS_HIGHLIGHT_LABELS_EN,
+  CHIP_VS_PAIRS,
+  chipVsSlug,
+  getAllChipPages,
+  getAllChipRouteSlugs,
+  getAllChipVsPages,
+  getChipPage,
+  getChipVsPage,
+} from './chip-pages';
+import {
+  assertZhChipParity,
+  buildZhChipFaq,
+  buildZhChipVsFaq,
+  CHIP_VS_HIGHLIGHT_LABELS_ZH,
+  getAllZhChipSlugs,
+  getZhChipTranslation,
+  localizeVsHighlightValueZh,
+} from './chip-pages-zh';
+
+const EM_OR_EN_DASH = /[—–]/u;
+
+describe('chip page registry integrity', () => {
+  it('resolves every hwKey in HW_REGISTRY and every specName in GPU_SPECS', () => {
+    for (const entry of getAllChipPages()) {
+      expect(HW_REGISTRY[entry.hwKey], `hwKey ${entry.hwKey}`).toBeDefined();
+      expect(
+        GPU_SPECS.find((spec) => spec.name === entry.specName),
+        `specName ${entry.specName}`,
+      ).toBeDefined();
+    }
+  });
+
+  it('has unique slugs across chip and versus pages', () => {
+    const slugs = getAllChipRouteSlugs();
+    expect(new Set(slugs).size).toBe(slugs.length);
+  });
+
+  it('covers 9 chips and 12 versus pages (44 localized URLs with the index)', () => {
+    expect(getAllChipPages()).toHaveLength(9);
+    expect(getAllChipVsPages()).toHaveLength(12);
+    // index + chips + versus, each in en and zh
+    expect(2 * (1 + getAllChipRouteSlugs().length)).toBe(44);
+  });
+
+  it('only pairs known chips and never a chip with itself', () => {
+    for (const { a, b } of CHIP_VS_PAIRS) {
+      expect(a).not.toBe(b);
+      expect(getChipPage(a), `pair member ${a}`).toBeDefined();
+      expect(getChipPage(b), `pair member ${b}`).toBeDefined();
+      expect(getChipVsPage(chipVsSlug(a, b))).toBeDefined();
+    }
+  });
+
+  it('links only to blog posts and glossary terms that exist', () => {
+    for (const entry of getAllChipPages()) {
+      for (const slug of entry.relatedBlogSlugs) {
+        expect(getPostBySlug(slug), `blog ${slug} on ${entry.slug}`).not.toBeNull();
+      }
+      for (const slug of entry.relatedGlossarySlugs) {
+        expect(getGlossaryEntry(slug), `glossary ${slug} on ${entry.slug}`).toBeDefined();
+      }
+      for (const slug of entry.relatedChipSlugs) {
+        expect(getChipPage(slug), `related chip ${slug} on ${entry.slug}`).toBeDefined();
+      }
+    }
+  });
+
+  it('builds a ratio-bearing highlight table for every versus page', () => {
+    for (const page of getAllChipVsPages()) {
+      const highlights = buildChipVsHighlights(page);
+      expect(highlights.length).toBeGreaterThanOrEqual(5);
+      for (const row of highlights) {
+        expect(CHIP_VS_HIGHLIGHT_LABELS_EN[row.key]).toBeDefined();
+        expect(CHIP_VS_HIGHLIGHT_LABELS_ZH[row.key]).toBeDefined();
+        expect(row.aValue.length).toBeGreaterThan(0);
+        expect(row.bValue.length).toBeGreaterThan(0);
+        // Only fp4 may omit the ratio (when one side lacks FP4 support).
+        if (row.key !== 'fp4') {
+          expect(row.ratio, `${page.slug} ${row.key} ratio`).toBeDefined();
+        }
+      }
+    }
+  });
+});
+
+describe('chip page content quality', () => {
+  it('has non-empty keyword-bearing prose and FAQs on every English page', () => {
+    for (const entry of getAllChipPages()) {
+      expect(entry.summary.length).toBeGreaterThan(80);
+      expect(entry.overview.length).toBeGreaterThanOrEqual(2);
+      expect(entry.keywords.length).toBeGreaterThanOrEqual(5);
+      const faq = buildChipFaq(entry);
+      expect(faq.length).toBeGreaterThanOrEqual(4);
+      for (const item of faq) {
+        expect(item.question.length).toBeGreaterThan(0);
+        expect(item.answer.length).toBeGreaterThan(0);
+      }
+    }
+    for (const page of getAllChipVsPages()) {
+      expect(buildChipVsFaq(page).length).toBeGreaterThanOrEqual(3);
+    }
+  });
+
+  it('contains no em or en dashes in any locale (glossary style rule)', () => {
+    for (const entry of getAllChipPages()) {
+      const zh = getZhChipTranslation(entry.slug);
+      const texts = [
+        entry.summary,
+        ...entry.overview,
+        entry.benchmarkContext,
+        ...buildChipFaq(entry).flatMap((item) => [item.question, item.answer]),
+        ...(zh ? [zh.summary, ...zh.overview, zh.benchmarkContext] : []),
+        ...buildZhChipFaq(entry).flatMap((item) => [item.question, item.answer]),
+      ];
+      for (const text of texts) {
+        expect(text).not.toMatch(EM_OR_EN_DASH);
+      }
+    }
+    for (const page of getAllChipVsPages()) {
+      for (const item of [...buildChipVsFaq(page), ...buildZhChipVsFaq(page)]) {
+        expect(item.question).not.toMatch(EM_OR_EN_DASH);
+        expect(item.answer).not.toMatch(EM_OR_EN_DASH);
+      }
+    }
+  });
+});
+
+describe('Chinese chip page parity', () => {
+  it('ships a zh translation for every English page and nothing extra', () => {
+    expect(() => assertZhChipParity()).not.toThrow();
+    expect(getAllZhChipSlugs().length).toBe(getAllChipPages().length);
+  });
+
+  it('keeps SKUs in English while translating the prose', () => {
+    for (const entry of getAllChipPages()) {
+      const zh = getZhChipTranslation(entry.slug);
+      expect(zh).toBeDefined();
+      if (!zh) continue;
+      expect(zh.summary).toMatch(/[\u4E00-\u9FFF]/u);
+      expect(zh.summary).toContain(entry.label);
+      expect(zh.keywords.length).toBeGreaterThanOrEqual(5);
+      // zh-copy rule: the standalone English noun "Chip" must not appear.
+      for (const text of [zh.summary, ...zh.overview, zh.benchmarkContext]) {
+        expect(text).not.toMatch(/(?<![A-Za-z])[Cc]hips?(?![A-Za-z])/u);
+      }
+    }
+  });
+
+  it('localizes generated FAQ answers and highlight values', () => {
+    expect(localizeVsHighlightValueZh('Not supported')).toBe('不支持');
+    expect(localizeVsHighlightValueZh('72 chips')).toBe('72 芯片');
+    expect(localizeVsHighlightValueZh('8 TB/s')).toBe('8 TB/s');
+    for (const entry of getAllChipPages()) {
+      for (const item of buildZhChipFaq(entry)) {
+        expect(item.question).toMatch(/[\u4E00-\u9FFF]/u);
+        expect(item.answer).toMatch(/[\u4E00-\u9FFF]/u);
+      }
+    }
+    for (const page of getAllChipVsPages()) {
+      for (const item of buildZhChipVsFaq(page)) {
+        expect(item.answer).toMatch(/[\u4E00-\u9FFF]/u);
+      }
+    }
+  });
+});
diff --git a/packages/app/src/lib/chip-pages.ts b/packages/app/src/lib/chip-pages.ts
new file mode 100644
index 000000000..43696f9e9
--- /dev/null
+++ b/packages/app/src/lib/chip-pages.ts
@@ -0,0 +1,578 @@
+/**
+ * Chip pages — static, indexable SEO surfaces at /chips, /chips/[slug] and
+ * /chips/[a]-vs-[b].
+ *
+ * Content strategy: every page is fully static (no DB) and joins two existing
+ * sources of truth so numbers can never drift from the rest of the site:
+ *
+ * - `GPU_SPECS` (src/lib/gpu-specs.ts) — memory, bandwidth, dense TFLOP/s,
+ *   interconnect and topology details shown on the Chip Specs tab.
+ * - `HW_REGISTRY` (@semianalysisai/inferencex-constants) — TDP, all-in power
+ *   and the $/chip/hr tiers used by the TCO calculator and per-dollar pages.
+ *
+ * FAQ answers and versus-page copy are generated from those structures at
+ * render time; only the overview prose is hand-authored. The Simplified
+ * Chinese siblings live in `chip-pages-zh.ts` and MUST be updated in the same
+ * PR as this file (see AGENTS.md "Chinese Website Pages").
+ */
+import { HW_REGISTRY, type HwEntry } from '@semianalysisai/inferencex-constants';
+
+import { GPU_SPECS, type GpuSpec } from '@/lib/gpu-specs';
+
+export interface ChipFaqItem {
+  question: string;
+  answer: string;
+}
+
+export interface ChipPageEntry {
+  /** URL slug under /chips/ */
+  slug: string;
+  /** Key into HW_REGISTRY (TDP, all-in power, $/hr tiers) */
+  hwKey: string;
+  /** `name` of the matching GPU_SPECS row */
+  specName: string;
+  /** Full display title, e.g. "NVIDIA B200" */
+  title: string;
+  /** Short label used in tables and versus copy, e.g. "B200" */
+  label: string;
+  /** Meta-description base; supporters line is appended by the page. */
+  summary: string;
+  /** Hand-authored overview paragraphs (keyword-bearing prose). */
+  overview: readonly string[];
+  /** How InferenceX benchmarks this chip. */
+  benchmarkContext: string;
+  /** Meta keywords; searcher-phrased, includes "GPU" variants. */
+  keywords: readonly string[];
+  /** Blog posts under content/blog that feature this chip. */
+  relatedBlogSlugs: readonly string[];
+  /** Glossary entries that explain the concepts named in the prose. */
+  relatedGlossarySlugs: readonly string[];
+  /** Other chip pages worth crosslinking. */
+  relatedChipSlugs: readonly string[];
+}
+
+const INFERENCEMAX = 'inferencemax-open-source-inference-benchmarking';
+const INFERENCEX_V2 = 'inferencex-v2-nvidia-blackwell-vs-amd-vs-hopper';
+
+const entries = [
+  {
+    slug: 'h100',
+    hwKey: 'h100',
+    specName: 'H100 SXM',
+    title: 'NVIDIA H100 SXM',
+    label: 'H100',
+    summary:
+      'NVIDIA H100 specs, cloud pricing and live LLM inference benchmarks: 80 GB HBM3, 3.35 TB/s bandwidth, FP8 tensor cores, NVLink 4.0, measured daily on vLLM, SGLang and TensorRT-LLM.',
+    overview: [
+      'The NVIDIA H100 SXM is the Hopper-generation datacenter accelerator that powered the first wave of large-scale LLM deployment. Each chip pairs 80 GB of HBM3 at 3.35 TB/s with fourth-generation tensor cores that reach 1,979 dense FP8 TFLOP/s, connected in 8-chip nodes over NVLink 4.0 at 450 GB/s per chip.',
+      'H100 remains the baseline other accelerators are judged against: it has the deepest software support of any AI chip, wide cloud availability, and the lowest hourly rates of the NVIDIA lineup. It lacks FP4 tensor cores, so newer Blackwell parts pull ahead on quantized serving workloads where NVFP4 applies.',
+    ],
+    benchmarkContext:
+      'InferenceX runs H100 continuously on vLLM, SGLang and TensorRT-LLM across fixed-sequence serving and long-context agentic traces, publishing throughput-versus-interactivity Pareto frontiers, cost per million tokens and energy per token alongside every newer chip so the upgrade math stays visible.',
+    keywords: [
+      'NVIDIA H100',
+      'H100 GPU',
+      'H100 specs',
+      'H100 price per hour',
+      'H100 cloud pricing',
+      'H100 inference benchmark',
+      'H100 FP8 TFLOPS',
+      'H100 memory bandwidth',
+      'Hopper GPU',
+      'H100 vs B200',
+    ],
+    relatedBlogSlugs: [
+      INFERENCEMAX,
+      'qwen3-5-397b-agentx-b300-fp4-vs-h100',
+      'b200-minimax-m2-5-vllm-nvfp4-vs-h100-fp8-perf-per-dollar',
+    ],
+    relatedGlossarySlugs: ['fp8', 'high-bandwidth-memory', 'nvlink', 'tensor-parallelism'],
+    relatedChipSlugs: ['h200', 'b200', 'mi300x'],
+  },
+  {
+    slug: 'h200',
+    hwKey: 'h200',
+    specName: 'H200 SXM',
+    title: 'NVIDIA H200 SXM',
+    label: 'H200',
+    summary:
+      'NVIDIA H200 specs, cloud pricing and live LLM inference benchmarks: 141 GB HBM3e, 4.8 TB/s bandwidth, FP8 tensor cores and NVLink 4.0, measured daily against Blackwell and AMD Instinct.',
+    overview: [
+      'The NVIDIA H200 SXM is the memory-upgraded Hopper part: the same 1,979 dense FP8 TFLOP/s as H100, but with 141 GB of HBM3e at 4.8 TB/s. The 76% larger memory pool and 43% higher bandwidth matter most for long-context serving and large KV caches, where H100 runs out of headroom first.',
+      'For memory-bound decode workloads the H200 delivers materially better tokens per second per chip than H100 at a similar hourly rate, which is why it stayed the volume Hopper SKU once Blackwell arrived. Like H100 it has no FP4 support, so FP8 and INT4 weight-only quantization are its serving precisions.',
+    ],
+    benchmarkContext:
+      'InferenceX benchmarks H200 daily on vLLM, SGLang and TensorRT-LLM, and its compare pages line H200 FP8 results up against B200 NVFP4 and AMD MI325X so the Hopper-to-Blackwell and NVIDIA-to-AMD tradeoffs are measured rather than asserted.',
+    keywords: [
+      'NVIDIA H200',
+      'H200 GPU',
+      'H200 specs',
+      'H200 price per hour',
+      'H200 inference benchmark',
+      'H200 memory bandwidth',
+      'H200 vs H100',
+      'H200 vs B200',
+      'HBM3e GPU',
+      'Hopper H200 cloud pricing',
+    ],
+    relatedBlogSlugs: [
+      'b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar',
+      'b200-nvfp4-vs-h200-int4-kimi-k2-vllm-perf-per-dollar',
+      INFERENCEX_V2,
+    ],
+    relatedGlossarySlugs: ['high-bandwidth-memory', 'memory-bandwidth', 'kv-cache', 'fp8'],
+    relatedChipSlugs: ['h100', 'b200', 'mi325x'],
+  },
+  {
+    slug: 'b200',
+    hwKey: 'b200',
+    specName: 'B200 SXM',
+    title: 'NVIDIA B200',
+    label: 'B200',
+    summary:
+      'NVIDIA B200 specs, cloud pricing and live LLM inference benchmarks: 180 GB HBM3e, 8 TB/s bandwidth, 9,000 dense FP4 TFLOP/s and NVLink 5.0, measured daily on vLLM, SGLang and TensorRT-LLM.',
+    overview: [
+      'The NVIDIA B200 is the volume Blackwell datacenter chip: 180 GB of usable HBM3e at 8 TB/s, NVLink 5.0 at 900 GB/s per chip, and tensor cores that double Hopper throughput per clock while adding FP4. At 9,000 dense FP4 TFLOP/s and 4,500 FP8, a single B200 more than doubles H100 compute in half the memory-bandwidth-bound regimes that dominate LLM decode.',
+      'B200 is where NVFP4 serving became mainstream: frontier open models ship FP4 checkpoints that keep accuracy within noise of FP8 while nearly doubling throughput per chip. Its 1,000 W TDP and roughly 1.7 kW all-in power draw price it well above Hopper per hour, so performance per dollar, not peak TFLOP/s, decides the upgrade.',
+    ],
+    benchmarkContext:
+      'B200 is one of the most heavily benchmarked chips on InferenceX: daily fixed-sequence sweeps and AgentX agentic-coding traces on vLLM, SGLang, TensorRT-LLM and Dynamo disaggregated serving, with per-dollar and precision compare pages tracking NVFP4 versus FP8 on every covered model.',
+    keywords: [
+      'NVIDIA B200',
+      'B200 GPU',
+      'B200 specs',
+      'B200 price per hour',
+      'B200 cloud pricing',
+      'B200 inference benchmark',
+      'B200 NVFP4',
+      'B200 vs H100',
+      'B200 vs MI355X',
+      'Blackwell GPU specs',
+    ],
+    relatedBlogSlugs: [
+      'b200-glm5-nvfp4-vs-h200-fp8-3-6x-perf-per-dollar',
+      'sglang-0-5-6-b200-deepseek-r1-fp4-up-to-1-8x',
+      'deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set',
+    ],
+    relatedGlossarySlugs: ['nvfp4', 'fp4', 'nvlink', 'quantization'],
+    relatedChipSlugs: ['b300', 'gb200-nvl72', 'h100', 'mi355x'],
+  },
+  {
+    slug: 'b300',
+    hwKey: 'b300',
+    specName: 'B300 SXM',
+    title: 'NVIDIA B300',
+    label: 'B300',
+    summary:
+      'NVIDIA B300 (Blackwell Ultra) specs, cloud pricing and live LLM inference benchmarks: 268 GB usable HBM3e, 13,500 dense FP4 TFLOP/s, 800 Gbit/s scale-out, measured daily against B200, GB300 NVL72 and MI355X.',
+    overview: [
+      'The NVIDIA B300 is the Blackwell Ultra SXM part: 268 GB of usable HBM3e at 8 TB/s and 13,500 dense FP4 TFLOP/s, a 1.5x FP4 uplift over B200 while FP8 and BF16 stay at B200 levels. The 49% larger memory pool is the bigger serving story, holding larger KV working sets and bigger models per 8-chip node.',
+      'B300 also doubles scale-out networking to 800 Gbit/s per chip via ConnectX-8, which matters for disaggregated prefill/decode and wide expert parallelism. It slots between B200 and the rack-scale GB300 NVL72: same NVLink 5.0 world size of 8, but Ultra-class memory and FP4 compute.',
+    ],
+    benchmarkContext:
+      'InferenceX runs B300 on the same daily cadence as B200 across vLLM, SGLang and TensorRT-LLM, including AgentX long-context agentic traces where its 268 GB of HBM3e keeps KV working sets resident that force smaller chips to preempt or offload.',
+    keywords: [
+      'NVIDIA B300',
+      'B300 GPU',
+      'B300 specs',
+      'Blackwell Ultra',
+      'B300 price per hour',
+      'B300 inference benchmark',
+      'B300 FP4 TFLOPS',
+      'B300 vs B200',
+      'B300 memory',
+      'B300 vs MI355X',
+    ],
+    relatedBlogSlugs: [
+      'deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set',
+      'minimax-m3-agentx-b300-trtllm-tp2-vs-the-field',
+      'qwen3-5-397b-agentx-b300-fp4-vs-h100',
+    ],
+    relatedGlossarySlugs: ['fp4', 'kv-cache', 'disaggregated-inference', 'scale-up-vs-scale-out'],
+    relatedChipSlugs: ['b200', 'gb300-nvl72', 'mi355x'],
+  },
+  {
+    slug: 'gb200-nvl72',
+    hwKey: 'gb200',
+    specName: 'GB200 NVL72',
+    title: 'NVIDIA GB200 NVL72',
+    label: 'GB200 NVL72',
+    summary:
+      'NVIDIA GB200 NVL72 specs, rack pricing and live LLM inference benchmarks: 72 Blackwell chips in one NVLink domain, 186 GB HBM3e per chip, 900 GB/s scale-up, measured daily on disaggregated vLLM, SGLang and Dynamo TRT-LLM.',
+    overview: [
+      'The NVIDIA GB200 NVL72 is a rack-scale system: 72 Blackwell chips and 36 Grace CPUs joined by fifth-generation NVLink into a single 72-chip domain with 900 GB/s of unidirectional bandwidth per chip. There is no scale-out network inside the rack because the rack itself is the compute unit, with roughly 13.4 TB of pooled HBM3e.',
+      'That one-hop, all-to-all fabric is why GB200 NVL72 dominates wide-expert-parallel serving of large MoE models: expert layers can be spread across dozens of chips without crossing InfiniBand or Ethernet. Disaggregated prefill/decode stacks such as Dynamo and SGLang PD map naturally onto the domain.',
+    ],
+    benchmarkContext:
+      'InferenceX benchmarks GB200 NVL72 with multi-node disaggregated configurations, including wide-EP vLLM, SGLang and Dynamo TRT-LLM runs on frontier MoE models, and normalizes results as throughput per chip so rack-scale numbers stay comparable with 8-chip HGX nodes.',
+    keywords: [
+      'NVIDIA GB200 NVL72',
+      'GB200 NVL72 specs',
+      'GB200 price',
+      'GB200 NVL72 benchmark',
+      'NVL72 rack',
+      'GB200 vs B200',
+      'NVLink 72 GPU domain',
+      'GB200 inference performance',
+      'rack-scale AI system',
+      'GB200 NVL72 power consumption',
+    ],
+    relatedBlogSlugs: [
+      'gb200-nvl72-kimi-k2-5-vllm-wide-ep-3x-vs-b200',
+      'gb200-nvl72-vs-b200-disagg-deepseek-r1-fp4-dynamo-trt',
+      'vera-rubin-nvl72-vs-gb200-nvl72-inference',
+    ],
+    relatedGlossarySlugs: [
+      'wide-expert-parallelism',
+      'disaggregated-inference',
+      'nvlink',
+      'scale-up-vs-scale-out',
+    ],
+    relatedChipSlugs: ['gb300-nvl72', 'b200', 'b300'],
+  },
+  {
+    slug: 'gb300-nvl72',
+    hwKey: 'gb300',
+    specName: 'GB300 NVL72',
+    title: 'NVIDIA GB300 NVL72',
+    label: 'GB300 NVL72',
+    summary:
+      'NVIDIA GB300 NVL72 specs, rack pricing and live LLM inference benchmarks: 72 Blackwell Ultra chips, 278 GB HBM3e each (20 TB per rack), 15,000 dense FP4 TFLOP/s per chip, measured daily on disaggregated serving stacks.',
+    overview: [
+      'The NVIDIA GB300 NVL72 is the Blackwell Ultra rack: 72 chips with 278 GB of usable HBM3e each, roughly 20 TB of pooled memory per rack, and 15,000 dense FP4 TFLOP/s per chip on the same 72-way NVLink 5.0 domain as GB200 NVL72. FP8 and BF16 throughput carry over from GB200; the Ultra uplift is FP4 and memory.',
+      'The extra memory per chip compounds at rack scale: bigger frontier MoE models, longer contexts and larger KV working sets fit without spilling, and each chip runs a higher 1,400 W TDP to feed the larger stacks. GB300 NVL72 is the current top of the NVIDIA inference lineup ahead of Vera Rubin.',
+    ],
+    benchmarkContext:
+      'InferenceX runs GB300 NVL72 on frontier MoE models with disaggregated and wide-EP configurations, and its AgentX agentic-coding lane uses the rack as the reference ceiling that AMD MI355X ATOM results and smaller NVIDIA nodes are measured against.',
+    keywords: [
+      'NVIDIA GB300 NVL72',
+      'GB300 NVL72 specs',
+      'GB300 price',
+      'GB300 NVL72 benchmark',
+      'Blackwell Ultra rack',
+      'GB300 vs GB200',
+      'GB300 memory per GPU',
+      'GB300 FP4 TFLOPS',
+      'GB300 inference performance',
+      'GB300 NVL72 power',
+    ],
+    relatedBlogSlugs: [
+      'gb300-nvl72-vs-gb200-nvl72-dsv4-pro-vllm-fp4',
+      'deepseek-v4-pro-agentx-gb200-vs-gb300-disagg',
+      'kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72',
+    ],
+    relatedGlossarySlugs: ['fp4', 'wide-expert-parallelism', 'kv-cache', 'disaggregated-inference'],
+    relatedChipSlugs: ['gb200-nvl72', 'b300', 'mi355x'],
+  },
+  {
+    slug: 'mi300x',
+    hwKey: 'mi300x',
+    specName: 'MI300X',
+    title: 'AMD Instinct MI300X',
+    label: 'MI300X',
+    summary:
+      'AMD Instinct MI300X specs, cloud pricing and live LLM inference benchmarks: 192 GB HBM3, 5.3 TB/s bandwidth, 2,615 dense FP8 TFLOP/s, full-mesh Infinity Fabric, measured daily on vLLM and SGLang with ROCm.',
+    overview: [
+      'The AMD Instinct MI300X is the CDNA 3 accelerator that made AMD a serious LLM inference vendor: 192 GB of HBM3 at 5.3 TB/s, more than double H100 memory, with 2,615 dense FP8 TFLOP/s. Eight chips connect in a full-mesh Infinity Fabric node with no switches, 448 GB/s of scale-up bandwidth per chip.',
+      'The memory advantage lets MI300X serve models in fewer chips or hold larger KV caches per replica, and its hourly pricing sits well below comparable NVIDIA parts. Software is the historical caveat: ROCm builds of vLLM and SGLang have closed much of the gap, and the public benchmark record tracks exactly how far.',
+    ],
+    benchmarkContext:
+      'MI300X was one of the original InferenceMAX chips and still runs in the continuous sweep on ROCm vLLM and SGLang, giving one of the longest public software-progress curves of any accelerator: the same chip, measured for months, as kernels and schedulers improved.',
+    keywords: [
+      'AMD MI300X',
+      'MI300X specs',
+      'MI300X price per hour',
+      'MI300X inference benchmark',
+      'MI300X vs H100',
+      'AMD Instinct GPU',
+      'CDNA 3',
+      'ROCm inference',
+      'MI300X memory',
+      'MI300X vLLM performance',
+    ],
+    relatedBlogSlugs: [INFERENCEMAX, INFERENCEX_V2, 'mi355x-kimi-k2-5-vllm-aiter-7x-speedup'],
+    relatedGlossarySlugs: ['rocm', 'high-bandwidth-memory', 'vllm', 'sglang'],
+    relatedChipSlugs: ['mi325x', 'mi355x', 'h100'],
+  },
+  {
+    slug: 'mi325x',
+    hwKey: 'mi325x',
+    specName: 'MI325X',
+    title: 'AMD Instinct MI325X',
+    label: 'MI325X',
+    summary:
+      'AMD Instinct MI325X specs, cloud pricing and live LLM inference benchmarks: 256 GB HBM3e, 6 TB/s bandwidth, 2,615 dense FP8 TFLOP/s, full-mesh Infinity Fabric, measured daily on ROCm vLLM and SGLang.',
+    overview: [
+      'The AMD Instinct MI325X is the memory-bumped CDNA 3 part: the same 2,615 dense FP8 TFLOP/s as MI300X but with 256 GB of HBM3e at 6 TB/s, the largest memory pool of its generation. An 8-chip full-mesh node carries over 2 TB of HBM, enough to serve very large models without leaving the node.',
+      'MI325X competes on capacity economics against H200: more memory per chip and lower hourly rates, traded against the NVIDIA software ecosystem. For memory-bound decode and long-context serving its bandwidth-per-dollar is among the best of the pre-CDNA 4 field.',
+    ],
+    benchmarkContext:
+      'InferenceX benchmarks MI325X on ROCm vLLM and SGLang across the shared model set, with compare pages pairing it against H200 and MI355X so both the NVIDIA-versus-AMD and generation-over-generation deltas stay continuously measured.',
+    keywords: [
+      'AMD MI325X',
+      'MI325X specs',
+      'MI325X price per hour',
+      'MI325X inference benchmark',
+      'MI325X vs H200',
+      'MI325X memory',
+      'AMD Instinct MI325X',
+      'CDNA 3 GPU',
+      'ROCm benchmark',
+      'MI325X vs MI300X',
+    ],
+    relatedBlogSlugs: [INFERENCEX_V2, INFERENCEMAX, 'mi355x-qwen3-5-sglang-v0-5-12-up-to-17x'],
+    relatedGlossarySlugs: ['memory-bandwidth', 'kv-cache', 'rocm', 'expert-parallelism'],
+    relatedChipSlugs: ['mi300x', 'mi355x', 'h200'],
+  },
+  {
+    slug: 'mi355x',
+    hwKey: 'mi355x',
+    specName: 'MI355X',
+    title: 'AMD Instinct MI355X',
+    label: 'MI355X',
+    summary:
+      'AMD Instinct MI355X specs, cloud pricing and live LLM inference benchmarks: 288 GB HBM3e, 8 TB/s bandwidth, 10,066 dense FP4 TFLOP/s, the first AMD chip with FP4, measured daily on ROCm vLLM, SGLang and ATOM.',
+    overview: [
+      'The AMD Instinct MI355X is the CDNA 4 flagship and the first AMD accelerator with FP4 tensor throughput: 10,066 dense FP4 and 5,033 FP8 TFLOP/s with 288 GB of HBM3e at 8 TB/s, the largest single-chip memory in the field. Eight chips form a full-mesh node over fifth-generation Infinity Fabric at 538 GB/s per chip.',
+      'MI355X matches B200-class bandwidth and exceeds its memory while pricing meaningfully lower per hour, making it the sharpest performance-per-dollar challenge AMD has mounted. Its 1,400 W TDP and the maturity of ROCm serving stacks are the tradeoffs the live data quantifies.',
+    ],
+    benchmarkContext:
+      'MI355X is the most intensively tracked AMD chip on InferenceX: daily vLLM, SGLang and ATOM runs, AgentX agentic-coding traces against GB300 NVL72 and B300, and some of the fastest published software-progress curves, including order-of-magnitude gains within weeks of a model release.',
+    keywords: [
+      'AMD MI355X',
+      'MI355X specs',
+      'MI355X price per hour',
+      'MI355X inference benchmark',
+      'MI355X vs B200',
+      'MI355X FP4',
+      'CDNA 4 GPU',
+      'MI355X memory bandwidth',
+      'MI355X vLLM SGLang',
+      'AMD FP4 GPU',
+    ],
+    relatedBlogSlugs: [
+      'mi355x-deepseek-v4-pro-sglang-110x-in-26-days',
+      'mi355x-glm5-fp8-sglang-40-cheaper-than-b200',
+      'deepseek-v4-pro-agentx-mi355x-vs-b200-august',
+    ],
+    relatedGlossarySlugs: ['fp4', 'rocm', 'memory-bandwidth', 'pareto-frontier'],
+    relatedChipSlugs: ['mi325x', 'b200', 'b300', 'gb300-nvl72'],
+  },
+] as const satisfies readonly ChipPageEntry[];
+
+export type ChipSlug = (typeof entries)[number]['slug'];
+
+/** Curated versus pairs. Order = page narrative order (a is the newer/leading chip). */
+export const CHIP_VS_PAIRS = [
+  { a: 'b200', b: 'h100' },
+  { a: 'b200', b: 'h200' },
+  { a: 'b300', b: 'b200' },
+  { a: 'gb200-nvl72', b: 'b200' },
+  { a: 'gb300-nvl72', b: 'gb200-nvl72' },
+  { a: 'gb300-nvl72', b: 'b300' },
+  { a: 'mi355x', b: 'mi325x' },
+  { a: 'mi325x', b: 'mi300x' },
+  { a: 'mi300x', b: 'h100' },
+  { a: 'mi355x', b: 'b200' },
+  { a: 'mi355x', b: 'b300' },
+  { a: 'mi325x', b: 'h200' },
+] as const satisfies readonly { a: ChipSlug; b: ChipSlug }[];
+
+export function getAllChipPages(): readonly ChipPageEntry[] {
+  return entries;
+}
+
+export function getChipPage(slug: string): ChipPageEntry | undefined {
+  return entries.find((entry) => entry.slug === slug);
+}
+
+export function getChipSpec(entry: ChipPageEntry): GpuSpec {
+  const spec = GPU_SPECS.find((row) => row.name === entry.specName);
+  if (!spec) throw new Error(`No GPU_SPECS row named ${entry.specName}`);
+  return spec;
+}
+
+export function getChipHw(entry: ChipPageEntry): HwEntry {
+  const hw = HW_REGISTRY[entry.hwKey];
+  if (!hw) throw new Error(`No HW_REGISTRY entry for ${entry.hwKey}`);
+  return hw;
+}
+
+export function chipVsSlug(a: ChipSlug, b: ChipSlug): string {
+  return `${a}-vs-${b}`;
+}
+
+export interface ChipVsPage {
+  slug: string;
+  a: ChipPageEntry;
+  b: ChipPageEntry;
+}
+
+export function getAllChipVsPages(): readonly ChipVsPage[] {
+  return CHIP_VS_PAIRS.map(({ a, b }) => {
+    const pageA = getChipPage(a);
+    const pageB = getChipPage(b);
+    if (!pageA || !pageB) throw new Error(`Versus pair references unknown chip: ${a} vs ${b}`);
+    return { slug: chipVsSlug(a, b), a: pageA, b: pageB };
+  });
+}
+
+export function getChipVsPage(slug: string): ChipVsPage | undefined {
+  return getAllChipVsPages().find((page) => page.slug === slug);
+}
+
+/** Every slug served by /chips/[slug] (chip pages first, then versus pages). */
+export function getAllChipRouteSlugs(): readonly string[] {
+  return [...entries.map((entry) => entry.slug), ...getAllChipVsPages().map((page) => page.slug)];
+}
+
+/** Parse "186 GB" / "8 TB/s" / "3.35 TB/s" style values to a number in the given unit. */
+export function leadingNumber(value: string): number {
+  const match = /^(?<num>[\d.]+)/u.exec(value.trim());
+  if (!match?.groups?.num) throw new Error(`Cannot parse numeric prefix from "${value}"`);
+  return Number(match.groups.num);
+}
+
+const formatRatio = (x: number, y: number): string => `${(x / y).toFixed(2).replace(/0$/u, '')}x`;
+
+export interface ChipVsHighlight {
+  /** Stable key so the zh sibling can localize the label. */
+  key: 'memory' | 'memoryBandwidth' | 'fp8' | 'fp4' | 'tdp' | 'costNeocloud' | 'scaleUpWorldSize';
+  aValue: string;
+  bValue: string;
+  /** a over b, e.g. "1.5x" — omitted when either side lacks the number. */
+  ratio?: string;
+}
+
+/** Data-derived comparison rows shared by the EN and ZH versus pages. */
+export function buildChipVsHighlights(page: ChipVsPage): readonly ChipVsHighlight[] {
+  const specA = getChipSpec(page.a);
+  const specB = getChipSpec(page.b);
+  const hwA = getChipHw(page.a);
+  const hwB = getChipHw(page.b);
+
+  return [
+    {
+      key: 'memory',
+      aValue: `${specA.memory} ${specA.memoryType}`,
+      bValue: `${specB.memory} ${specB.memoryType}`,
+      ratio: formatRatio(leadingNumber(specA.memory), leadingNumber(specB.memory)),
+    },
+    {
+      key: 'memoryBandwidth',
+      aValue: specA.memoryBandwidth,
+      bValue: specB.memoryBandwidth,
+      ratio: formatRatio(
+        leadingNumber(specA.memoryBandwidth),
+        leadingNumber(specB.memoryBandwidth),
+      ),
+    },
+    {
+      key: 'fp8',
+      aValue: `${specA.fp8.toLocaleString('en-US')} TFLOP/s`,
+      bValue: `${specB.fp8.toLocaleString('en-US')} TFLOP/s`,
+      ratio: formatRatio(specA.fp8, specB.fp8),
+    },
+    {
+      key: 'fp4',
+      aValue: specA.fp4 ? `${specA.fp4.toLocaleString('en-US')} TFLOP/s` : 'Not supported',
+      bValue: specB.fp4 ? `${specB.fp4.toLocaleString('en-US')} TFLOP/s` : 'Not supported',
+      ...(specA.fp4 && specB.fp4 ? { ratio: formatRatio(specA.fp4, specB.fp4) } : {}),
+    },
+    {
+      key: 'tdp',
+      aValue: `${hwA.tdp.toLocaleString('en-US')} W`,
+      bValue: `${hwB.tdp.toLocaleString('en-US')} W`,
+      ratio: formatRatio(hwA.tdp, hwB.tdp),
+    },
+    {
+      key: 'costNeocloud',
+      aValue: `$${hwA.costn.toFixed(2)}/hr`,
+      bValue: `$${hwB.costn.toFixed(2)}/hr`,
+      ratio: formatRatio(hwA.costn, hwB.costn),
+    },
+    {
+      key: 'scaleUpWorldSize',
+      aValue: `${specA.scaleUpWorldSize} chips`,
+      bValue: `${specB.scaleUpWorldSize} chips`,
+      ratio: formatRatio(specA.scaleUpWorldSize, specB.scaleUpWorldSize),
+    },
+  ];
+}
+
+/** English FAQ, generated from the spec/pricing registries so it cannot drift. */
+export function buildChipFaq(entry: ChipPageEntry): readonly ChipFaqItem[] {
+  const spec = getChipSpec(entry);
+  const hw = getChipHw(entry);
+  return [
+    {
+      question: `How much does ${entry.label} cost per hour in the cloud?`,
+      answer:
+        `The SemiAnalysis AI Cloud TCO model rates ${entry.label} at about $${hw.costh.toFixed(2)}/hr at hyperscalers, ` +
+        `$${hw.costn.toFixed(2)}/hr at neoclouds and $${hw.costr.toFixed(2)}/hr at the retail tier. ` +
+        `InferenceX performance-per-dollar pages use these rates to turn measured throughput into $/M tokens.`,
+    },
+    {
+      question: `How much memory does ${entry.label} have?`,
+      answer:
+        `${entry.label} has ${spec.memory} of usable ${spec.memoryType} per chip with ${spec.memoryBandwidth} of memory bandwidth. ` +
+        `A ${spec.scaleUpWorldSize}-chip ${spec.scaleUpTech} domain pools ${Math.round(leadingNumber(spec.memory) * spec.scaleUpWorldSize).toLocaleString('en-US')} GB.`,
+    },
+    {
+      question: `What is the power consumption of ${entry.label}?`,
+      answer:
+        `${entry.label} has a ${hw.tdp.toLocaleString('en-US')} W TDP per chip, and about ${hw.power} kW all-in per chip ` +
+        `once the host CPU, NICs and cooling share are included. InferenceX uses the all-in figure for energy-per-token math.`,
+    },
+    {
+      question: `Does ${entry.label} support FP4?`,
+      answer: spec.fp4
+        ? `Yes. ${entry.label} reaches ${spec.fp4.toLocaleString('en-US')} dense FP4 TFLOP/s (${spec.fp8.toLocaleString('en-US')} at FP8), and InferenceX tracks FP4-versus-FP8 serving accuracy and throughput on its precision compare pages.`
+        : `No. ${entry.label} tops out at FP8 with ${spec.fp8.toLocaleString('en-US')} dense TFLOP/s; FP4 serving requires a newer chip generation.`,
+    },
+    {
+      question: `How fast is ${entry.label} for LLM inference?`,
+      answer:
+        `It depends on the model, framework, precision and interactivity target, so InferenceX publishes continuously refreshed ` +
+        `throughput-versus-interactivity Pareto frontiers for ${entry.label} instead of a single number. The live dashboard and ` +
+        `compare pages show current results on every covered model.`,
+    },
+  ];
+}
+
+/** English FAQ for a versus page, generated from the same registries. */
+export function buildChipVsFaq(page: ChipVsPage): readonly ChipFaqItem[] {
+  const highlights = buildChipVsHighlights(page);
+  const memory = highlights.find((h) => h.key === 'memory');
+  const cost = highlights.find((h) => h.key === 'costNeocloud');
+  const fp8 = highlights.find((h) => h.key === 'fp8');
+  return [
+    {
+      question: `Which has more memory, ${page.a.label} or ${page.b.label}?`,
+      answer: `${page.a.label} offers ${memory?.aValue} per chip versus ${memory?.bValue} on ${page.b.label} (${memory?.ratio} the capacity).`,
+    },
+    {
+      question: `How do ${page.a.label} and ${page.b.label} prices compare?`,
+      answer:
+        `At the neocloud tier the SemiAnalysis TCO model rates ${page.a.label} at ${cost?.aValue} versus ${cost?.bValue} for ${page.b.label}. ` +
+        `Hourly price alone is misleading; the per-dollar compare pages divide measured throughput by these rates.`,
+    },
+    {
+      question: `Is ${page.a.label} faster than ${page.b.label} for LLM inference?`,
+      answer:
+        `On paper ${page.a.label} has ${fp8?.ratio} the dense FP8 compute of ${page.b.label}, but delivered tokens per second depend on the model, ` +
+        `framework, precision and interactivity target. InferenceX measures both chips daily on identical workloads; see the live compare pages for current results.`,
+    },
+  ];
+}
+
+/** English labels for the data-derived versus highlight rows. */
+export const CHIP_VS_HIGHLIGHT_LABELS_EN: Readonly<Record<ChipVsHighlight['key'], string>> = {
+  memory: 'Memory per chip',
+  memoryBandwidth: 'Memory bandwidth',
+  fp8: 'Dense FP8 compute',
+  fp4: 'Dense FP4 compute',
+  tdp: 'TDP',
+  costNeocloud: 'Hourly rate (neocloud tier)',
+  scaleUpWorldSize: 'Scale-up world size',
+};
diff --git a/packages/app/src/lib/client-navigation.ts b/packages/app/src/lib/client-navigation.ts
index ea5017927..9b910a02e 100644
--- a/packages/app/src/lib/client-navigation.ts
+++ b/packages/app/src/lib/client-navigation.ts
@@ -7,6 +7,7 @@ interface RouterLike {
 }
 
 export const CLIENT_SEARCH_CHANGE_EVENT = 'inferencex:client-search-change';
+export const CLIENT_PATHNAME_CHANGE_EVENT = 'inferencex:client-pathname-change';
 
 /** Keep persistent layout controls in sync when an App Router transition only
  *  changes search params and therefore reuses the root layout. */
@@ -15,6 +16,65 @@ export function notifyClientSearchChange(href: string): void {
   window.dispatchEvent(new CustomEvent(CLIENT_SEARCH_CHANGE_EVENT, { detail: search }));
 }
 
+/**
+ * Replace only the current document's query string without involving the App
+ * Router. Next patches `window.history.replaceState`; calling the pristine
+ * prototype method avoids an RSC navigation for client-owned URL state.
+ */
+export function replaceClientSearch(searchParams: URLSearchParams): void {
+  const search = searchParams.toString();
+  const href = `${window.location.pathname}${search ? `?${search}` : ''}${window.location.hash}`;
+  History.prototype.replaceState.call(window.history, window.history.state, '', href);
+  notifyClientSearchChange(href);
+}
+
+/**
+ * Replace only the current document's pathname (keeping search and hash)
+ * without involving the App Router — same pristine-prototype trick as
+ * `replaceClientSearch`. Used by per-model dashboard routes so switching
+ * models rewrites `/historical/kimi-k3` in the address bar without an RSC
+ * navigation or component remount. `usePathname` intentionally keeps the
+ * server-rendered value; route resolution is prefix-based, so nav highlight,
+ * providers, and share scopes are unaffected. Controls that must reflect the
+ * live address bar (the header language toggle) subscribe via
+ * `useClientPathname`, which listens for CLIENT_PATHNAME_CHANGE_EVENT.
+ */
+export function replaceClientPathname(pathname: string): void {
+  const href = `${pathname}${window.location.search}${window.location.hash}`;
+  History.prototype.replaceState.call(window.history, window.history.state, '', href);
+  clientPathnameOverride = pathname;
+  window.dispatchEvent(new CustomEvent(CLIENT_PATHNAME_CHANGE_EVENT, { detail: pathname }));
+}
+
+let clientPathnameOverride: string | null = null;
+
+/** The last pathname written by `replaceClientPathname`, or null. Consumers
+ *  (useClientPathname) honor it only while the address bar still matches, so
+ *  a later App Router navigation naturally retires a stale override. */
+export function getClientPathnameOverride(): string | null {
+  return clientPathnameOverride;
+}
+
+/**
+ * Shallow-replace the current pathname, preserving the query string (minus
+ * `dropParams`) and hash. Unlike `replaceClientSearch`, this goes through the
+ * Next-patched `window.history.replaceState` on purpose: the pathname is
+ * changing, so the App Router must adopt the new URL (`usePathname` updates)
+ * — but as a same-document rewrite, without an RSC fetch or scroll reset.
+ * Router state is carried over so Back/Forward keep working.
+ */
+export function replaceRouterPathname(target: string, dropParams: readonly string[] = []): void {
+  if (window.location.pathname === target) return;
+  const search = new URLSearchParams(window.location.search);
+  for (const param of dropParams) search.delete(param);
+  const qs = search.toString();
+  window.history.replaceState(
+    window.history.state,
+    '',
+    `${target}${qs ? `?${qs}` : ''}${window.location.hash}`,
+  );
+}
+
 /**
  * The first dashboard transition can request the route without committing the
  * URL change. Repeating the same app-router push after the route payload has
diff --git a/packages/app/src/lib/collectivex-lazy-ingest.test.ts b/packages/app/src/lib/collectivex-lazy-ingest.test.ts
index 20bf58293..85265b022 100644
--- a/packages/app/src/lib/collectivex-lazy-ingest.test.ts
+++ b/packages/app/src/lib/collectivex-lazy-ingest.test.ts
@@ -3,16 +3,14 @@ import { afterAll, beforeEach, describe, expect, it, vi } from 'vitest';
 
 import { makeRawMatrix, makeRawShard } from '@/components/collectivex/test-fixture';
 
-const { mockGetStates, mockInsert, mockRefresh, mockGetDb, mockGetWriteDb } = vi.hoisted(() => ({
+const { mockGetStates, mockInsert, mockRefresh, mockGetWriteDb } = vi.hoisted(() => ({
   mockGetStates: vi.fn(),
   mockInsert: vi.fn(),
   mockRefresh: vi.fn(),
-  mockGetDb: vi.fn(() => 'mock-sql'),
   mockGetWriteDb: vi.fn(() => 'mock-write-sql'),
 }));
 
 vi.mock('@semianalysisai/inferencex-db/connection', () => ({
-  getCollectiveXDb: mockGetDb,
   getCollectiveXWriteDb: mockGetWriteDb,
 }));
 
@@ -460,6 +458,7 @@ describe('ensureCollectiveXRun', () => {
     mockFetch.mockResolvedValueOnce(Response.json(runObject()));
 
     await ensureCollectiveXRun(1, '160');
+    expect(mockGetStates).toHaveBeenCalledWith('mock-write-sql', ['160']);
 
     expect(mockFetch).toHaveBeenCalledTimes(1);
     expect(mockInsert).not.toHaveBeenCalled();
diff --git a/packages/app/src/lib/collectivex-lazy-ingest.ts b/packages/app/src/lib/collectivex-lazy-ingest.ts
index 6e5d7055b..7ca178b40 100644
--- a/packages/app/src/lib/collectivex-lazy-ingest.ts
+++ b/packages/app/src/lib/collectivex-lazy-ingest.ts
@@ -31,7 +31,7 @@ import {
   selectShardArtifacts,
 } from '@semianalysisai/inferencex-db/collectivex/artifact-selection';
 import type { CollectiveXVersion } from '@semianalysisai/inferencex-db/collectivex/types';
-import { getCollectiveXDb, getCollectiveXWriteDb } from '@semianalysisai/inferencex-db/connection';
+import { getCollectiveXWriteDb } from '@semianalysisai/inferencex-db/connection';
 import {
   getCollectiveXRunStates,
   insertCollectiveXRun,
@@ -492,7 +492,7 @@ async function considerCandidate(
   version: CollectiveXVersion,
   token: string,
 ): Promise<CandidateResult> {
-  const states = await getCollectiveXRunStates(getCollectiveXDb(), [String(run.id)]);
+  const states = await getCollectiveXRunStates(getCollectiveXWriteDb(), [String(run.id)]);
   const known = states[String(run.id)];
   if (known) {
     if (known.version !== version || known.state !== 'live') return 'skip';
@@ -576,7 +576,7 @@ export function ensureCollectiveXRun(version: CollectiveXVersion, runId: string)
     if (!Number.isSafeInteger(numericId) || numericId <= 0) {
       throw new CollectiveXSweepError('not-found', 'invalid run id');
     }
-    const states = await getCollectiveXRunStates(getCollectiveXDb(), [runId]);
+    const states = await getCollectiveXRunStates(getCollectiveXWriteDb(), [runId]);
     const known = states[runId];
     if (known && (known.state !== 'live' || known.version !== version)) {
       // Tombstoned or cross-version rows both read as absent to the caller.
diff --git a/packages/app/src/lib/compare-agentx.test.ts b/packages/app/src/lib/compare-agentx.test.ts
new file mode 100644
index 000000000..9f314f7a9
--- /dev/null
+++ b/packages/app/src/lib/compare-agentx.test.ts
@@ -0,0 +1,74 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  agentxDashboardHref,
+  AGENTX_NEW_MODEL_DISPLAY_NAMES,
+  comparisonPairHref,
+  comparisonScenarioForModel,
+  FEATURED_AGENTX_MODELS,
+} from './compare-agentx';
+import { COMPARE_MODEL_SLUGS } from './compare-slug';
+
+describe('AgentX comparison links', () => {
+  it('keeps the live AgentX models in the editorial order', () => {
+    expect(FEATURED_AGENTX_MODELS.map((model) => model.slug)).toEqual([
+      'kimi-k3',
+      'deepseek-v4',
+      'glm-5-2',
+      'minimax-m3',
+      'qwen-3-5',
+      'qwen-3-8-flash-next',
+    ]);
+  });
+
+  it('marks exactly the featured models as NEW for the dashboard selector', () => {
+    expect(AGENTX_NEW_MODEL_DISPLAY_NAMES).toEqual(
+      new Set([
+        'Kimi-K3',
+        'DeepSeek-V4-Pro',
+        'GLM-5.2',
+        'MiniMax-M3',
+        'Qwen-3.5-397B-A17B',
+        'Qwen3.8-Flash-Next',
+      ]),
+    );
+  });
+
+  it('opens the bare model subroute — Agentic + Optimal Only are already the defaults', () => {
+    expect(agentxDashboardHref('en', FEATURED_AGENTX_MODELS[0])).toBe('/inference/kimi-k3');
+    expect(agentxDashboardHref('zh', FEATURED_AGENTX_MODELS[1])).toBe('/zh/inference/deepseek-v4');
+  });
+
+  it('routes every featured model to a registered inference page without query params', () => {
+    for (const model of FEATURED_AGENTX_MODELS) {
+      expect(agentxDashboardHref('en', model)).toBe(`/inference/${model.slug}`);
+    }
+  });
+
+  it('uses AgentX for supported models and 8K→1K for the rest', () => {
+    const deepSeekV4 = COMPARE_MODEL_SLUGS.find((model) => model.slug === 'deepseek-v4')!;
+    const deepSeekR1 = COMPARE_MODEL_SLUGS.find((model) => model.slug === 'deepseek-r1')!;
+
+    expect(comparisonScenarioForModel(deepSeekV4)).toEqual({
+      label: 'AgentX',
+      sequence: 'agentic-traces',
+    });
+    expect(comparisonPairHref('en', 'deepseek-v4-h100-vs-h200', deepSeekV4)).toBe(
+      '/compare/deepseek-v4-h100-vs-h200/agentic',
+    );
+    expect(comparisonScenarioForModel(deepSeekR1)).toEqual({
+      label: '8K→1K',
+      sequence: '8k/1k',
+    });
+    expect(comparisonPairHref('zh', 'deepseek-r1-h100-vs-h200', deepSeekR1)).toBe(
+      '/zh/compare/deepseek-r1-h100-vs-h200/8k-1k',
+    );
+    // The per-dollar catalog links the same workloads under its own family.
+    expect(
+      comparisonPairHref('en', 'deepseek-v4-h100-vs-h200', deepSeekV4, 'compare-per-dollar'),
+    ).toBe('/compare-per-dollar/deepseek-v4-h100-vs-h200/agentic');
+    expect(
+      comparisonPairHref('zh', 'deepseek-r1-h100-vs-h200', deepSeekR1, 'compare-per-dollar'),
+    ).toBe('/zh/compare-per-dollar/deepseek-r1-h100-vs-h200/8k-1k');
+  });
+});
diff --git a/packages/app/src/lib/compare-agentx.ts b/packages/app/src/lib/compare-agentx.ts
new file mode 100644
index 000000000..0914c1c9a
--- /dev/null
+++ b/packages/app/src/lib/compare-agentx.ts
@@ -0,0 +1,81 @@
+import { scenarioSegmentForSequence } from '@/lib/compare-scenario-route';
+import { COMPARE_MODEL_SLUGS, type CompareModelSlug } from '@/lib/compare-slug';
+import { getInferenceModelBySlug, inferenceModelPath } from '@/lib/inference-model-slug';
+
+const FEATURED_AGENTX_MODEL_SLUGS = [
+  'kimi-k3',
+  'deepseek-v4',
+  'glm-5-2',
+  'minimax-m3',
+  // Editorial call: Flash Next sits below Qwen 3.5, the family's flagship row.
+  'qwen-3-5',
+  'qwen-3-8-flash-next',
+] as const;
+
+const FEATURED_AGENTX_MODEL_SET = new Set<string>(FEATURED_AGENTX_MODEL_SLUGS);
+
+export interface ComparisonScenario {
+  label: 'AgentX' | '8K→1K';
+  sequence: 'agentic-traces' | '8k/1k';
+}
+
+export const FEATURED_AGENTX_MODELS: readonly CompareModelSlug[] = FEATURED_AGENTX_MODEL_SLUGS.map(
+  (slug) => {
+    const model = COMPARE_MODEL_SLUGS.find((candidate) => candidate.slug === slug);
+    if (!model) throw new Error(`Missing AgentX comparison model: ${slug}`);
+    return model;
+  },
+);
+
+/**
+ * `Model` enum values (the dashboard's model identifiers) for the featured
+ * AgentX set. The landing/compare hero ledger and the /inference model
+ * selector both mark these models with a NEW badge, so the badge follows the
+ * single featured list instead of maintaining a second one.
+ */
+export const AGENTX_NEW_MODEL_DISPLAY_NAMES: ReadonlySet<string> = new Set(
+  FEATURED_AGENTX_MODELS.map((model) => model.displayName),
+);
+
+export function agentxDashboardHref(locale: 'en' | 'zh', model: CompareModelSlug): string {
+  // The model rides in the path — the indexable `/inference/<model>` subroute
+  // — so these hero links point crawlers at the canonical model page instead
+  // of a `?g_model=` variant of the base dashboard. Models without a
+  // registered inference page (none of the featured set today) fall back to
+  // the query form.
+  //
+  // No `i_seq` / `i_optimal` params: every model here has AgentX data, and the
+  // dashboard already defaults such models to the Agentic scenario
+  // (resolveEffectiveSequence) with "Optimal Only" on (i_optimal !== '0'), so
+  // the bare model page IS the AgentX optimal-only view — and the canonical,
+  // shareable address for it.
+  const entry = getInferenceModelBySlug(model.slug);
+  const path = entry ? inferenceModelPath(entry.slug) : '/inference';
+  const localizedPath = locale === 'zh' ? `/zh${path}` : path;
+  if (entry) return localizedPath;
+  const query = new URLSearchParams({ g_model: model.displayName });
+  return `${localizedPath}?${query}`;
+}
+
+export function comparisonScenarioForModel(model: CompareModelSlug): ComparisonScenario {
+  return FEATURED_AGENTX_MODEL_SET.has(model.slug)
+    ? { label: 'AgentX', sequence: 'agentic-traces' }
+    : { label: '8K→1K', sequence: '8k/1k' };
+}
+
+/**
+ * Catalog card link for a comparison pair, pointing at the workload the model
+ * actually has data for. The workload rides in the path (`/…/<slug>/agentic`)
+ * rather than `?i_seq=`, so a card links to a real, shareable address for that
+ * scenario instead of a query-string variant of the default view.
+ */
+export function comparisonPairHref(
+  locale: 'en' | 'zh',
+  slug: string,
+  model: CompareModelSlug,
+  family: 'compare' | 'compare-per-dollar' = 'compare',
+): string {
+  const path = locale === 'zh' ? `/zh/${family}/${slug}` : `/${family}/${slug}`;
+  const segment = scenarioSegmentForSequence(comparisonScenarioForModel(model).sequence);
+  return segment ? `${path}/${segment}` : path;
+}
diff --git a/packages/app/src/lib/compare-og-fonts/NotoSansCJKsc-Compare-Bold.otf b/packages/app/src/lib/compare-og-fonts/NotoSansCJKsc-Compare-Bold.otf
new file mode 100644
index 000000000..13f3728d0
Binary files /dev/null and b/packages/app/src/lib/compare-og-fonts/NotoSansCJKsc-Compare-Bold.otf differ
diff --git a/packages/app/src/lib/compare-og-fonts/NotoSansCJKsc-Compare-Regular.otf b/packages/app/src/lib/compare-og-fonts/NotoSansCJKsc-Compare-Regular.otf
new file mode 100644
index 000000000..9bdec59a0
Binary files /dev/null and b/packages/app/src/lib/compare-og-fonts/NotoSansCJKsc-Compare-Regular.otf differ
diff --git a/packages/app/src/lib/compare-og.tsx b/packages/app/src/lib/compare-og.tsx
new file mode 100644
index 000000000..07bc317a9
--- /dev/null
+++ b/packages/app/src/lib/compare-og.tsx
@@ -0,0 +1,158 @@
+import { ImageResponse } from 'next/og';
+
+import {
+  BG,
+  BLUE,
+  getCjkFonts,
+  getLogoSrc,
+  getTiles,
+  MUTED,
+  PANEL_BG,
+  TEXT,
+  TITLE,
+} from '@/lib/og-assets';
+
+interface CompareOgSize {
+  width: number;
+  height: number;
+}
+
+interface CompareOgOptions {
+  eyebrow: string;
+  title: string;
+  titleSize: number;
+  footer: string;
+  language: 'en' | 'zh';
+  size: CompareOgSize;
+}
+
+/** Shared compare OG layout with optional bundled Simplified Chinese fonts. */
+export async function renderCompareOg({
+  eyebrow,
+  title,
+  titleSize,
+  footer,
+  language,
+  size,
+}: CompareOgOptions): Promise<ImageResponse> {
+  const [logoSrc, tiles, fonts] = await Promise.all([
+    getLogoSrc(),
+    getTiles(),
+    language === 'zh' ? getCjkFonts() : Promise.resolve([]),
+  ]);
+
+  return new ImageResponse(
+    <div
+      style={{
+        display: 'flex',
+        width: '100%',
+        height: '100%',
+        backgroundColor: BG,
+        color: TEXT,
+        overflow: 'hidden',
+        ...(language === 'zh' ? { fontFamily: 'Noto Sans SC' } : {}),
+      }}
+    >
+      <div
+        style={{
+          display: 'flex',
+          width: 195,
+          height: '100%',
+          backgroundColor: PANEL_BG,
+          position: 'relative',
+        }}
+      >
+        {tiles.map((tile, index) => {
+          if (!tile) return null;
+          const row = Math.floor(index / 2);
+          const column = index % 2;
+          return (
+            <img
+              key={index}
+              src={tile.src}
+              style={{
+                position: 'absolute',
+                left: 12 + column * 90,
+                top: 12 + row * 104,
+                width: 78,
+                height: 86,
+                borderRadius: 4,
+                objectFit: 'cover',
+                ...(tile.rotate ? { transform: `rotate(${tile.rotate}deg)` } : {}),
+              }}
+            />
+          );
+        })}
+        <div
+          style={{
+            position: 'absolute',
+            right: 0,
+            top: 0,
+            width: 3,
+            height: '100%',
+            backgroundColor: BLUE,
+            display: 'flex',
+          }}
+        />
+      </div>
+
+      <div
+        style={{
+          display: 'flex',
+          flexDirection: 'column',
+          justifyContent: 'space-between',
+          flex: 1,
+          padding: '48px 55px 20px 55px',
+        }}
+      >
+        <div
+          style={{
+            display: 'flex',
+            flexDirection: 'column',
+            gap: 24,
+            flex: 1,
+            justifyContent: 'center',
+          }}
+        >
+          <div
+            style={{
+              fontSize: 24,
+              color: MUTED,
+              letterSpacing: '0.18em',
+              textTransform: 'uppercase',
+              display: 'flex',
+            }}
+          >
+            {eyebrow}
+          </div>
+          <div
+            style={{
+              fontSize: titleSize,
+              fontWeight: 800,
+              color: TITLE,
+              lineHeight: 1.1,
+              display: 'flex',
+            }}
+          >
+            {title}
+          </div>
+        </div>
+
+        <div
+          style={{
+            display: 'flex',
+            justifyContent: 'space-between',
+            alignItems: 'center',
+          }}
+        >
+          <span style={{ fontSize: 26, color: MUTED, display: 'flex' }}>{footer}</span>
+          {logoSrc && <img src={logoSrc} height={72} />}
+        </div>
+      </div>
+    </div>,
+    {
+      ...size,
+      ...(fonts.length > 0 ? { fonts } : {}),
+    },
+  );
+}
diff --git a/packages/app/src/lib/compare-page-data.server.test.ts b/packages/app/src/lib/compare-page-data.server.test.ts
new file mode 100644
index 000000000..8bc9f7944
--- /dev/null
+++ b/packages/app/src/lib/compare-page-data.server.test.ts
@@ -0,0 +1,114 @@
+import { readFileSync } from 'node:fs';
+
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+const mocks = vi.hoisted(() => ({
+  loadRows: vi.fn(),
+  compute: vi.fn(),
+  summarize: vi.fn((_: unknown, hardware: string) => ({ hardware })),
+}));
+
+vi.mock('next/cache', () => ({
+  revalidateTag: vi.fn(),
+  unstable_cache: (fn: (...args: unknown[]) => Promise<unknown>) => {
+    const entries = new Map<string, Promise<unknown>>();
+    return (...args: unknown[]) => {
+      const key = JSON.stringify(args);
+      const existing = entries.get(key);
+      if (existing) return existing;
+      const value = fn(...args);
+      entries.set(key, value);
+      return value;
+    };
+  },
+}));
+
+vi.mock('@/lib/benchmark-data.server', () => ({
+  getCachedBenchmarks: mocks.loadRows,
+}));
+
+vi.mock('@/lib/compare-pair-defaults', () => ({
+  pickPairDefaults: () => ({ sequence: '1k/1k', precision: 'fp8' }),
+}));
+
+vi.mock('@/lib/compare-ssr', () => ({
+  computeCompareTableData: mocks.compute,
+  dateRangeForPair: () => ({ oldest: '2026-01-01', newest: '2026-02-01' }),
+  summarize: mocks.summarize,
+}));
+
+import { getComparePageDerivedData, initialCompareBenchmarkRows } from './compare-page-data.server';
+
+const TABLE = {
+  defaultTargets: [10],
+  ssrRows: [],
+  interactivityRange: { min: 1, max: 2 },
+};
+
+describe('getComparePageDerivedData', () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+    mocks.loadRows.mockResolvedValue([{ id: 1, hardware: 'h100' }]);
+    mocks.compute.mockReturnValue(TABLE);
+  });
+
+  it('reuses one source read while preserving identical derived values', async () => {
+    const first = await getComparePageDerivedData(['model-a', 'model-b'], 'h100', 'h200');
+    const second = await getComparePageDerivedData(['model-a', 'model-b'], 'h100', 'h200');
+
+    expect(second).toEqual(first);
+    expect(mocks.compute).toHaveBeenCalledTimes(2);
+    expect(mocks.loadRows).toHaveBeenCalledWith(['model-a', 'model-b']);
+    expect(first.initialPairBenchmarkRows).toEqual([{ id: 1, hardware: 'h100' }]);
+  });
+
+  it('does not collide distinct meaningful selector states', async () => {
+    const defaultData = await getComparePageDerivedData(['model-c'], 'h100', 'h200');
+    const selectedData = await getComparePageDerivedData(
+      ['model-c'],
+      'h100',
+      'h200',
+      '8k/1k',
+      'bf16',
+    );
+
+    expect(defaultData.sequence).toBe('1k/1k');
+    expect(selectedData.sequence).toBe('8k/1k');
+    expect(selectedData.precision).toBe('bf16');
+    expect(mocks.loadRows).toHaveBeenCalledTimes(1);
+    expect(mocks.compute).toHaveBeenCalledTimes(2);
+    expect(mocks.compute).toHaveBeenLastCalledWith(
+      [{ id: 1, hardware: 'h100' }],
+      'h100',
+      'h200',
+      '8k/1k',
+      'bf16',
+    );
+  });
+
+  it('keeps the largest fixture GPU pair below the unstable_cache payload limit', () => {
+    const rows = JSON.parse(
+      readFileSync(new URL('../../cypress/fixtures/api/benchmarks.json', import.meta.url), 'utf8'),
+    ) as { hardware: string }[];
+    const countByHardware = new Map<string, number>();
+    for (const row of rows) {
+      countByHardware.set(row.hardware, (countByHardware.get(row.hardware) ?? 0) + 1);
+    }
+    const largestPair = [...countByHardware.entries()]
+      .toSorted((a, b) => b[1] - a[1])
+      .slice(0, 2)
+      .map(([hardware]) => hardware);
+    const pairRows = rows.filter((row) => largestPair.includes(row.hardware));
+
+    expect(Buffer.byteLength(JSON.stringify(pairRows))).toBeLessThan(2 * 1024 * 1024);
+  });
+});
+
+describe('initialCompareBenchmarkRows', () => {
+  it('never seeds slug rows under an overridden model key', () => {
+    const rows = [{ hardware: 'h100' }] as never[];
+
+    expect(initialCompareBenchmarkRows('Slug Model', 'Slug Model', rows)).toBe(rows);
+    expect(initialCompareBenchmarkRows('Slug Model', 'Override Model', rows)).toBeUndefined();
+  });
+});
diff --git a/packages/app/src/lib/compare-page-data.server.ts b/packages/app/src/lib/compare-page-data.server.ts
new file mode 100644
index 000000000..875dc47f8
--- /dev/null
+++ b/packages/app/src/lib/compare-page-data.server.ts
@@ -0,0 +1,109 @@
+import { FIXTURES_MODE } from '@semianalysisai/inferencex-db/connection';
+
+import type { BenchmarkRow } from '@/lib/api';
+import { cachedDerivedData } from '@/lib/api-cache';
+import { getCachedBenchmarks } from '@/lib/benchmark-data.server';
+import { pickPairDefaults } from '@/lib/compare-pair-defaults';
+import {
+  computeCompareTableData,
+  dateRangeForPair,
+  summarize,
+  type PairSummary,
+  type SsrInterpolatedRow,
+} from '@/lib/compare-ssr';
+
+export interface ComparePageDerivedData {
+  sequence: string | null;
+  precision: string | null;
+  summaryA: PairSummary;
+  summaryB: PairSummary;
+  defaultTargets: number[];
+  ssrRows: SsrInterpolatedRow[];
+  interactivityRange: { min: number; max: number };
+  oldest?: string;
+  newest?: string;
+  /** Full-shape rows limited to the fixed GPU pair for scoped chart hydration. */
+  initialPairBenchmarkRows: BenchmarkRow[];
+}
+
+export function initialCompareBenchmarkRows(
+  slugModel: string,
+  effectiveModel: string,
+  pairRows: BenchmarkRow[],
+): BenchmarkRow[] | undefined {
+  return effectiveModel === slugModel ? pairRows : undefined;
+}
+
+async function loadComparePairRows(
+  dbModelKeys: string[],
+  a: string,
+  b: string,
+): Promise<BenchmarkRow[]> {
+  const rows = await getCachedBenchmarks(dbModelKeys);
+  return rows.filter((row) => row.hardware === a || row.hardware === b);
+}
+
+const getCachedComparePairRows = cachedDerivedData(loadComparePairRows, 'compare-pair-rows-v1');
+
+function buildComparePageDerivedData(
+  rows: BenchmarkRow[],
+  a: string,
+  b: string,
+  requestedSequence: string | null,
+  requestedPrecision: string | null,
+  fallbackSequence: string | null,
+): Omit<ComparePageDerivedData, 'initialPairBenchmarkRows'> {
+  const defaults = pickPairDefaults(rows, a, b, fallbackSequence ?? undefined);
+  const sequence = requestedSequence ?? defaults.sequence;
+  const precision = requestedPrecision ?? defaults.precision;
+  const { defaultTargets, ssrRows, interactivityRange } = computeCompareTableData(
+    rows,
+    a,
+    b,
+    sequence,
+    precision,
+  );
+  const { oldest, newest } = dateRangeForPair(rows, a, b);
+
+  return {
+    sequence,
+    precision,
+    summaryA: summarize(rows, a),
+    summaryB: summarize(rows, b),
+    defaultTargets,
+    ssrRows,
+    interactivityRange,
+    oldest,
+    newest,
+  };
+}
+
+/**
+ * Assemble the compact server-owned comparison payload from one cached,
+ * pair-filtered row set. Selector transforms stay pure and cheap per request;
+ * caching them separately would either nest unstable_cache calls or duplicate
+ * the full pair payload in every sequence/precision entry.
+ */
+export async function getComparePageDerivedData(
+  dbModelKeys: readonly string[],
+  a: string,
+  b: string,
+  requestedSequence?: string | null,
+  requestedPrecision?: string | null,
+  fallbackSequence?: string | null,
+): Promise<ComparePageDerivedData> {
+  const keys = [...dbModelKeys];
+  const pairRowsLoader = FIXTURES_MODE ? loadComparePairRows : getCachedComparePairRows;
+  const initialPairBenchmarkRows = await pairRowsLoader(keys, a, b);
+  return {
+    ...buildComparePageDerivedData(
+      initialPairBenchmarkRows,
+      a,
+      b,
+      requestedSequence ?? null,
+      requestedPrecision ?? null,
+      fallbackSequence ?? null,
+    ),
+    initialPairBenchmarkRows,
+  };
+}
diff --git a/packages/app/src/lib/compare-pair-defaults.test.ts b/packages/app/src/lib/compare-pair-defaults.test.ts
index da81ca0e0..922cb5bd9 100644
--- a/packages/app/src/lib/compare-pair-defaults.test.ts
+++ b/packages/app/src/lib/compare-pair-defaults.test.ts
@@ -112,4 +112,72 @@ describe('pickPairDefaults', () => {
       precision: 'fp8',
     });
   });
+
+  it('recognizes AgentX rows without fixed input or output lengths', () => {
+    const rows: BenchmarkRow[] = [
+      makeRow({
+        hardware: 'h100',
+        benchmark_type: 'agentic_traces',
+        isl: null,
+        osl: null,
+        precision: 'fp8',
+      }),
+      makeRow({
+        hardware: 'gb200',
+        benchmark_type: 'agentic_traces',
+        isl: null,
+        osl: null,
+        precision: 'fp8',
+      }),
+    ];
+
+    expect(pickPairDefaults(rows, 'h100', 'gb200')).toEqual({
+      sequence: 'agentic-traces',
+      precision: 'fp8',
+    });
+  });
+
+  it('honors an available preferred scenario before overlap ranking', () => {
+    const rows: BenchmarkRow[] = [
+      makeRow({ hardware: 'h100', conc: 32 }),
+      makeRow({ hardware: 'h100', conc: 64 }),
+      makeRow({ hardware: 'gb200', conc: 32 }),
+      makeRow({ hardware: 'gb200', conc: 64 }),
+      makeRow({
+        hardware: 'h100',
+        benchmark_type: 'agentic_traces',
+        isl: null,
+        osl: null,
+        conc: 8,
+      }),
+      makeRow({
+        hardware: 'gb200',
+        benchmark_type: 'agentic_traces',
+        isl: null,
+        osl: null,
+        conc: 8,
+      }),
+    ];
+
+    expect(pickPairDefaults(rows, 'h100', 'gb200', 'agentic-traces')).toEqual({
+      sequence: 'agentic-traces',
+      precision: 'fp8',
+    });
+    expect(pickPairDefaults(rows, 'h100', 'gb200', '8k/1k')).toEqual({
+      sequence: '8k/1k',
+      precision: 'fp8',
+    });
+  });
+
+  it('falls back to the best available scenario when the preference has no rows', () => {
+    const rows: BenchmarkRow[] = [
+      makeRow({ hardware: 'h100', conc: 64 }),
+      makeRow({ hardware: 'gb200', conc: 64 }),
+    ];
+
+    expect(pickPairDefaults(rows, 'h100', 'gb200', 'agentic-traces')).toEqual({
+      sequence: '8k/1k',
+      precision: 'fp8',
+    });
+  });
 });
diff --git a/packages/app/src/lib/compare-pair-defaults.ts b/packages/app/src/lib/compare-pair-defaults.ts
index f5a37e1ff..fde1c53fa 100644
--- a/packages/app/src/lib/compare-pair-defaults.ts
+++ b/packages/app/src/lib/compare-pair-defaults.ts
@@ -1,4 +1,4 @@
-import { islOslToSequence } from '@semianalysisai/inferencex-constants';
+import { rowToSequence } from '@semianalysisai/inferencex-constants';
 import type { BenchmarkRow } from '@semianalysisai/inferencex-db/queries/benchmarks';
 
 // Picks the (sequence, precision) that maximises overlap of distinct
@@ -8,14 +8,14 @@ export function pickPairDefaults(
   rows: BenchmarkRow[],
   a: string,
   b: string,
+  preferredSequence?: string,
 ): { sequence: string | null; precision: string | null } {
   const tally = new Map<string, { both: number; either: number }>();
   const seenA = new Map<string, Set<string>>();
   const seenB = new Map<string, Set<string>>();
   for (const row of rows) {
     if (row.hardware !== a && row.hardware !== b) continue;
-    if (row.isl === null || row.osl === null) continue;
-    const seq = islOslToSequence(row.isl, row.osl);
+    const seq = rowToSequence(row);
     if (!seq) continue;
     const key = `${seq}|${row.precision}`;
     const variantId = `${row.framework}|${row.spec_method}|${row.conc}`;
@@ -36,6 +36,11 @@ export function pickPairDefaults(
   }
   if (tally.size === 0) return { sequence: null, precision: null };
   const best = [...tally.entries()].toSorted((left, right) => {
+    if (preferredSequence) {
+      const leftPreferred = left[0].startsWith(`${preferredSequence}|`);
+      const rightPreferred = right[0].startsWith(`${preferredSequence}|`);
+      if (leftPreferred !== rightPreferred) return leftPreferred ? -1 : 1;
+    }
     if (left[1].both !== right[1].both) return right[1].both - left[1].both;
     return right[1].either - left[1].either;
   })[0];
diff --git a/packages/app/src/lib/compare-scenario-route.test.ts b/packages/app/src/lib/compare-scenario-route.test.ts
new file mode 100644
index 000000000..ebbf3310a
--- /dev/null
+++ b/packages/app/src/lib/compare-scenario-route.test.ts
@@ -0,0 +1,79 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  compareScenarioPath,
+  isAgenticSequence,
+  SCENARIO_SEGMENT_TO_SEQUENCE,
+  SCENARIO_SEGMENTS,
+  scenarioSegmentForSequence,
+  sequenceForScenarioSegment,
+} from './compare-scenario-route';
+import { KNOWN_SEQUENCES } from './compare-ssr';
+
+describe('compare scenario segments', () => {
+  it('covers every sequence the query param accepts', () => {
+    // If a sequence can be expressed as `?i_seq=`, it needs a path segment
+    // too — otherwise a workload exists that the routes cannot address.
+    const mapped = new Set<string>(Object.values(SCENARIO_SEGMENT_TO_SEQUENCE));
+    expect(mapped).toEqual(new Set(KNOWN_SEQUENCES));
+  });
+
+  it('round-trips every segment through its sequence', () => {
+    for (const segment of SCENARIO_SEGMENTS) {
+      const sequence = sequenceForScenarioSegment(segment);
+      expect(sequence).not.toBeNull();
+      expect(scenarioSegmentForSequence(sequence!)).toBe(segment);
+    }
+  });
+
+  it('maps the two headline workloads', () => {
+    expect(sequenceForScenarioSegment('agentic')).toBe('agentic-traces');
+    expect(sequenceForScenarioSegment('8k-1k')).toBe('8k/1k');
+  });
+
+  it('resolves a capitalized segment rather than 404ing', () => {
+    expect(sequenceForScenarioSegment('Agentic')).toBe('agentic-traces');
+  });
+
+  it('rejects unknown segments so the route can 404', () => {
+    expect(sequenceForScenarioSegment('bogus')).toBeNull();
+    expect(sequenceForScenarioSegment('')).toBeNull();
+    // The raw sequence key is not a valid segment — a slash cannot live in one.
+    expect(sequenceForScenarioSegment('8k/1k')).toBeNull();
+  });
+
+  it('identifies the agentic sequence', () => {
+    expect(isAgenticSequence('agentic-traces')).toBe(true);
+    expect(isAgenticSequence('8k/1k')).toBe(false);
+    expect(isAgenticSequence(null)).toBe(false);
+    expect(isAgenticSequence(undefined)).toBe(false);
+  });
+});
+
+describe('compareScenarioPath', () => {
+  it('emits the bare slug URL for the default workload', () => {
+    // The default view keeps exactly one address rather than gaining a second,
+    // identical one at a scenario segment.
+    expect(
+      compareScenarioPath('/compare', 'dsv4-b200-vs-h200', 'agentic-traces', 'agentic-traces'),
+    ).toBe('/compare/dsv4-b200-vs-h200');
+  });
+
+  it('appends the segment for a non-default workload', () => {
+    expect(compareScenarioPath('/compare', 'dsv4-b200-vs-h200', '8k/1k', 'agentic-traces')).toBe(
+      '/compare/dsv4-b200-vs-h200/8k-1k',
+    );
+  });
+
+  it('falls back to the bare URL for an unmapped sequence', () => {
+    expect(compareScenarioPath('/compare', 'dsv4-b200-vs-h200', 'nonsense', 'agentic-traces')).toBe(
+      '/compare/dsv4-b200-vs-h200',
+    );
+  });
+
+  it('works for the other compare families', () => {
+    expect(
+      compareScenarioPath('/zh/compare-per-dollar', 'dsv4-b200-vs-h200', '1k/8k', 'agentic-traces'),
+    ).toBe('/zh/compare-per-dollar/dsv4-b200-vs-h200/1k-8k');
+  });
+});
diff --git a/packages/app/src/lib/compare-scenario-route.ts b/packages/app/src/lib/compare-scenario-route.ts
new file mode 100644
index 000000000..1fd3cf088
--- /dev/null
+++ b/packages/app/src/lib/compare-scenario-route.ts
@@ -0,0 +1,78 @@
+/**
+ * Scenario path segments for the compare families.
+ *
+ * The workload used to live only in `?i_seq=agentic-traces`. It is now also
+ * addressable as a path segment — `/compare/<slug>/agentic`,
+ * `/compare/<slug>/8k-1k` — so each workload has a real URL that can be
+ * linked, shared, and indexed on its own.
+ *
+ * The segment is the sequence key with `/` swapped for `-`, because a literal
+ * slash cannot live inside one path segment. Every `KNOWN_SEQUENCES` value has
+ * a segment, so the scheme is closed: there is no workload the query param can
+ * express that a route cannot.
+ *
+ * `/compare/<slug>` keeps working and keeps rendering the pair's default
+ * workload (agentic where the model has AgentX data, otherwise the fixed
+ * sequence the pair actually has rows for). Existing `?i_seq=` links keep
+ * working too; a scenario segment simply outranks the query param.
+ */
+
+import { Sequence } from '@/lib/data-mappings';
+
+/** Path segment → sequence key. Ordered as the selector lists them. */
+export const SCENARIO_SEGMENT_TO_SEQUENCE = {
+  agentic: Sequence.AgenticTraces,
+  '8k-1k': Sequence.EightK_OneK,
+  '1k-1k': Sequence.OneK_OneK,
+  '1k-8k': Sequence.OneK_EightK,
+} as const satisfies Record<string, Sequence>;
+
+export type ScenarioSegment = keyof typeof SCENARIO_SEGMENT_TO_SEQUENCE;
+
+const SEQUENCE_TO_SCENARIO_SEGMENT = new Map<string, ScenarioSegment>(
+  Object.entries(SCENARIO_SEGMENT_TO_SEQUENCE).map(([segment, sequence]) => [
+    sequence,
+    segment as ScenarioSegment,
+  ]),
+);
+
+export const SCENARIO_SEGMENTS = Object.keys(
+  SCENARIO_SEGMENT_TO_SEQUENCE,
+) as readonly ScenarioSegment[];
+
+/** `"agentic"` → `"agentic-traces"`; anything else → `null` (the route 404s). */
+export function sequenceForScenarioSegment(segment: string): Sequence | null {
+  return (
+    SCENARIO_SEGMENT_TO_SEQUENCE[segment as ScenarioSegment] ??
+    // A case-insensitive match still resolves rather than 404ing — the segment
+    // is user-typed often enough (shared links, docs) that a capitalized
+    // `/Agentic` should reach the page.
+    SCENARIO_SEGMENT_TO_SEQUENCE[segment.toLowerCase() as ScenarioSegment] ??
+    null
+  );
+}
+
+/** `"agentic-traces"` → `"agentic"`. Returns `null` for an unmapped sequence. */
+export function scenarioSegmentForSequence(sequence: string): ScenarioSegment | null {
+  return SEQUENCE_TO_SCENARIO_SEGMENT.get(sequence) ?? null;
+}
+
+export function isAgenticSequence(sequence: string | null | undefined): boolean {
+  return sequence === Sequence.AgenticTraces;
+}
+
+/**
+ * URL for a compare page at a given workload. Emits the bare slug URL when the
+ * workload is the pair's default, so the default view keeps exactly one
+ * address rather than gaining a second, identical one at a scenario segment.
+ */
+export function compareScenarioPath(
+  basePath: string,
+  slug: string,
+  sequence: string,
+  defaultSequence: string,
+): string {
+  const segment = scenarioSegmentForSequence(sequence);
+  if (!segment || sequence === defaultSequence) return `${basePath}/${slug}`;
+  return `${basePath}/${slug}/${segment}`;
+}
diff --git a/packages/app/src/lib/compare-slug.test.ts b/packages/app/src/lib/compare-slug.test.ts
index a04404bbc..bcbafd813 100644
--- a/packages/app/src/lib/compare-slug.test.ts
+++ b/packages/app/src/lib/compare-slug.test.ts
@@ -256,7 +256,7 @@ describe('compareModelDisplayLabel', () => {
       'Kimi K2.5/K2.6/K2.7-Code 1T — GB200 NVL72 vs MI355X',
     );
     expect(compareModelDisplayLabel(GLM_51, 'h100', 'h200')).toBe('GLM 5/5.1 — H100 vs H200');
-    expect(compareModelDisplayLabel(GLM_52, 'h100', 'h200')).toBe('GLM 5.2 — H100 vs H200');
+    expect(compareModelDisplayLabel(GLM_52, 'h100', 'h200')).toBe('GLM 5.3 744B — H100 vs H200');
   });
 });
 
@@ -273,6 +273,7 @@ describe('compareModelSeoName', () => {
     'glm-5-2': 'GLM-5.2',
     'minimax-m3': 'MiniMax M3',
     'minimax-m27': 'MiniMax M2.7',
+    'qwen-3-8-flash-next': 'Qwen3.8-Flash-Next',
     'qwen-3-5': 'Qwen3.5',
     'gptoss-120b': 'gpt-oss-120b',
     'llama-3-3-70b': 'Llama 3.3 70B',
diff --git a/packages/app/src/lib/compare-slug.ts b/packages/app/src/lib/compare-slug.ts
index e42ff186d..bb1b9257e 100644
--- a/packages/app/src/lib/compare-slug.ts
+++ b/packages/app/src/lib/compare-slug.ts
@@ -50,7 +50,7 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     slug: 'deepseek-v4',
     displayName: 'DeepSeek-V4-Pro',
     dbKeys: ['dsv4'],
-    label: 'DeepSeek V4 Pro 1.6T',
+    label: 'DeepSeekv4 Pro 0813 1.6T',
     seoName: 'DeepSeek V4 Pro',
   },
   {
@@ -104,7 +104,10 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     slug: 'glm-5-2',
     displayName: 'GLM-5.2',
     dbKeys: ['glm5.2'],
-    label: 'GLM 5.2',
+    // GLM-5.2 and GLM-5.3 share one data bucket (see MODEL_CONFIG); the card
+    // shows the current release. `slug`, `displayName`, and `dbKeys` stay on
+    // 5.2 so URLs and the `g_model` param keep resolving.
+    label: 'GLM 5.3 744B',
     seoName: 'GLM-5.2',
   },
   {
@@ -126,6 +129,15 @@ export const COMPARE_MODEL_SLUGS: CompareModelSlug[] = [
     // Primary version the slug canonicalizes to (M2.7).
     seoName: 'MiniMax M2.7',
   },
+  {
+    slug: 'qwen-3-8-flash-next',
+    displayName: 'Qwen3.8-Flash-Next',
+    dbKeys: ['qwen3.8next'],
+    // 176B total: a 125B main model plus a 51B n-gram embedding table, with 6B
+    // active per forward pass (MoE). The 4B MTP head sits outside that total.
+    label: 'Qwen 3.8 Flash Next 176B-A6B',
+    seoName: 'Qwen3.8-Flash-Next',
+  },
   {
     slug: 'qwen-3-5',
     displayName: 'Qwen-3.5-397B-A17B',
@@ -189,6 +201,20 @@ export function getCompareModelBySlug(slug: string): CompareModelSlug | null {
   return SLUG_TO_MODEL[canonical] ?? null;
 }
 
+const DISPLAY_NAME_TO_SLUG: Record<string, CompareModelSlug> = Object.fromEntries(
+  COMPARE_MODEL_SLUGS.map((m) => [m.displayName, m]),
+);
+
+/**
+ * Resolve a dashboard model (the `Model` enum value / `g_model` param) to its
+ * canonical model-page slug entry. Used by the `/model/[slug]` pages and by
+ * the dashboard's "learn more about the architecture" link. Returns null for
+ * models without a public slug (e.g. hidden models like Llama 3.1 70B).
+ */
+export function getModelSlugEntryForDisplayName(displayName: string): CompareModelSlug | null {
+  return DISPLAY_NAME_TO_SLUG[displayName] ?? null;
+}
+
 // ---------------------------------------------------------------------------
 // Slug parsing and canonicalization
 // ---------------------------------------------------------------------------
diff --git a/packages/app/src/lib/compare-ssr-zh.ts b/packages/app/src/lib/compare-ssr-zh.ts
index 0476f879f..36ced19ae 100644
--- a/packages/app/src/lib/compare-ssr-zh.ts
+++ b/packages/app/src/lib/compare-ssr-zh.ts
@@ -20,6 +20,7 @@ import {
   compareModelSeoName,
 } from '@/lib/compare-slug';
 import {
+  type AgenticScenarioIntro,
   bandFor,
   type CompareJsonLdVariant,
   computeCompareStat,
@@ -112,27 +113,27 @@ function fullSummaryZh(i: FullBoth): string {
       ? null
       : `${i.cheaper} 每 token 成本低 ${fmtPctDelta(i.costRatio)}`;
   const tputPart = i.tputTied
-    ? '每 Chip 吞吐量基本持平'
+    ? '每芯片吞吐量基本持平'
     : i.tputRatio === null
       ? null
-      : `${i.faster} 每 Chip 吞吐量高出 ${fmtPctDelta(i.tputRatio)}`;
+      : `${i.faster} 每芯片吞吐量高出 ${fmtPctDelta(i.tputRatio)}`;
   const both = [costPart, tputPart].filter(Boolean).join('；');
   return both.length > 0 ? both : '差距极小，难以判定优劣';
 }
 
 const FULL_BOTH_TEMPLATES_ZH: ((i: FullBoth) => string)[] = [
   (i) =>
-    `在 ${i.modelLabel} 上以 ${i.target} tok/s/user 交互性运行时，${i.aLabel} 吞吐量为 ${i.aValue.toFixed(0)} tok/s/Chip，每百万 token 成本 ${fmtCost(i.aCost)}；${i.bLabel} 吞吐量为 ${i.bValue.toFixed(0)} tok/s/Chip，成本 ${fmtCost(i.bCost)}。${fullSummaryZh(i)}。`,
+    `在 ${i.modelLabel} 上以 ${i.target} tok/s/user 交互性运行时，${i.aLabel} 吞吐量为 ${i.aValue.toFixed(0)} tok/s/chip，每百万 token 成本 ${fmtCost(i.aCost)}；${i.bLabel} 吞吐量为 ${i.bValue.toFixed(0)} tok/s/chip，成本 ${fmtCost(i.bCost)}。${fullSummaryZh(i)}。`,
   (i) =>
-    `${i.aLabel} 在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时达到 ${i.aValue.toFixed(0)} tok/s/Chip（每百万 token ${fmtCost(i.aCost)}）；${i.bLabel} 达到 ${i.bValue.toFixed(0)} tok/s/Chip（${fmtCost(i.bCost)}）。${fullSummaryZh(i)}。`,
+    `${i.aLabel} 在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时达到 ${i.aValue.toFixed(0)} tok/s/chip（每百万 token ${fmtCost(i.aCost)}）；${i.bLabel} 达到 ${i.bValue.toFixed(0)} tok/s/chip（${fmtCost(i.bCost)}）。${fullSummaryZh(i)}。`,
   (i) =>
-    `${i.modelLabel} 在 ${i.target} tok/s/user 交互性下的吞吐量：${i.aLabel} 为 ${i.aValue.toFixed(0)} tok/s/Chip，${i.bLabel} 为 ${i.bValue.toFixed(0)}。每百万 token 成本分别为 ${fmtCost(i.aCost)} 和 ${fmtCost(i.bCost)}。${fullSummaryZh(i)}。`,
+    `${i.modelLabel} 在 ${i.target} tok/s/user 交互性下的吞吐量：${i.aLabel} 为 ${i.aValue.toFixed(0)} tok/s/chip，${i.bLabel} 为 ${i.bValue.toFixed(0)}。每百万 token 成本分别为 ${fmtCost(i.aCost)} 和 ${fmtCost(i.bCost)}。${fullSummaryZh(i)}。`,
   (i) =>
-    `${i.aLabel} / ${i.bLabel} 在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行：${i.aValue.toFixed(0)} / ${i.bValue.toFixed(0)} tok/s/Chip，${fmtCost(i.aCost)} / ${fmtCost(i.bCost)} 每百万 token。${fullSummaryZh(i)}。`,
+    `${i.aLabel} / ${i.bLabel} 在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行：${i.aValue.toFixed(0)} / ${i.bValue.toFixed(0)} tok/s/chip，${fmtCost(i.aCost)} / ${fmtCost(i.bCost)} 每百万 token。${fullSummaryZh(i)}。`,
   (i) =>
-    `在 ${i.range} 交互性区间的${BAND_PHRASE_ZH[i.band]}，即 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时：${i.aLabel} 达到 ${i.aValue.toFixed(0)} tok/s/Chip（${fmtCost(i.aCost)}/百万 token），${i.bLabel} 达到 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}/百万）。${fullSummaryZh(i)}。`,
+    `在 ${i.range} 交互性区间的${BAND_PHRASE_ZH[i.band]}，即 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时：${i.aLabel} 达到 ${i.aValue.toFixed(0)} tok/s/chip（${fmtCost(i.aCost)}/百万 token），${i.bLabel} 达到 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}/百万）。${fullSummaryZh(i)}。`,
   (i) =>
-    `以 ${i.target} tok/s/user 为目标在 ${i.modelLabel} 上运行时，${i.aLabel} 产出 ${i.aValue.toFixed(0)} tok/s/Chip（每百万 token ${fmtCost(i.aCost)}），${i.bLabel} 产出 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}）。${fullSummaryZh(i)}。`,
+    `以 ${i.target} tok/s/user 为目标在 ${i.modelLabel} 上运行时，${i.aLabel} 产出 ${i.aValue.toFixed(0)} tok/s/chip（每百万 token ${fmtCost(i.aCost)}），${i.bLabel} 产出 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}）。${fullSummaryZh(i)}。`,
 ];
 
 const FULL_SINGLE_TEMPLATES_ZH: ((args: {
@@ -144,17 +145,25 @@ const FULL_SINGLE_TEMPLATES_ZH: ((args: {
   presentCost: number;
 }) => string)[] = [
   (i) =>
-    `在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时，${i.presentLabel} 吞吐量为 ${i.presentValue.toFixed(0)} tok/s/Chip，每百万 token 成本 ${fmtCost(i.presentCost)}；${i.missingLabel} 在此目标点没有基准测试数据。`,
+    `在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时，${i.presentLabel} 吞吐量为 ${i.presentValue.toFixed(0)} tok/s/chip，每百万 token 成本 ${fmtCost(i.presentCost)}；${i.missingLabel} 在此目标点没有基准测试数据。`,
   (i) =>
-    `${i.presentLabel} 在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时达到 ${i.presentValue.toFixed(0)} tok/s/Chip（每百万 token ${fmtCost(i.presentCost)}）。${i.missingLabel} 在此工作点没有数据。`,
+    `${i.presentLabel} 在 ${i.modelLabel} 上以 ${i.target} tok/s/user 运行时达到 ${i.presentValue.toFixed(0)} tok/s/chip（每百万 token ${fmtCost(i.presentCost)}）。${i.missingLabel} 在此工作点没有数据。`,
   (i) =>
-    `${i.presentLabel}：${i.presentValue.toFixed(0)} tok/s/Chip，每百万 token ${fmtCost(i.presentCost)}（${i.modelLabel} 上以 ${i.target} tok/s/user 运行）。${i.missingLabel} 在此点尚未测试。`,
+    `${i.presentLabel}：${i.presentValue.toFixed(0)} tok/s/chip，每百万 token ${fmtCost(i.presentCost)}（${i.modelLabel} 上以 ${i.target} tok/s/user 运行）。${i.missingLabel} 在此点尚未测试。`,
 ];
 
 // ---------------------------------------------------------------------------
 // compareTableNarrativeZh
 // ---------------------------------------------------------------------------
 
+/** 1:1 port of `AGENTIC_SCENARIO_INTRO` — see the English original for why. */
+export const AGENTIC_SCENARIO_INTRO_ZH: AgenticScenarioIntro = {
+  paragraph:
+    'AgentX 回放真实的编码 agent 会话，而不是固定长度的 prompt，因此上下文随轮次不断增长，每个请求的大部分内容由 cache 提供而无需重新计算。这让对比变成一个系统问题：跨节点的 KV 传输、prefix 感知路由与 cache 容量，都会与芯片原始吞吐量一同影响曲线。固定序列负载仍然是衡量 kernel 与芯片性能的干净基线，因此两种场景回答的是关于同一硬件的不同问题。',
+  linkLabel: '进一步了解 AgentX',
+  href: '/zh/agentx',
+};
+
 export function compareTableNarrativeZh(
   variant: CompareJsonLdVariant,
   modelLabel: string,
@@ -321,7 +330,7 @@ export function compareMetaDescriptionZh(
   if (!stat) return fallback;
 
   const tputClause =
-    stat.tputPct > 0 ? `${stat.faster} 每 Chip 吞吐量比 ${stat.slower} 高 ${stat.tputPct}%` : null;
+    stat.tputPct > 0 ? `${stat.faster} 每芯片吞吐量比 ${stat.slower} 高 ${stat.tputPct}%` : null;
   const costClause = stat.costPct > 0 ? `${stat.cheaper} 每 token 成本低 ${stat.costPct}%` : null;
 
   let core: string;
@@ -354,7 +363,7 @@ export function buildBreadcrumbJsonLdZh(
 ) {
   const indexUrl =
     variant === 'per-dollar' ? `${SITE_URL}/zh/compare-per-dollar` : `${SITE_URL}/zh/compare`;
-  const indexName = variant === 'per-dollar' ? 'Chip 每美元性能' : 'Chip 对比';
+  const indexName = variant === 'per-dollar' ? '芯片每美元性能' : '芯片对比';
   return {
     '@context': 'https://schema.org',
     '@type': 'BreadcrumbList',
@@ -388,7 +397,7 @@ export function buildJsonLdZh(
     variant === 'per-dollar' ? `${fullLabel} — 每美元性能` : `${fullLabel} 推理基准测试`;
   const itemListDescription =
     variant === 'per-dollar'
-      ? `${aLabel} 与 ${bLabel} 在 ${model.label} 上的每百万 token 成本。基于所属云服务商 TCO 归一化的 Chip 推理性能。`
+      ? `${aLabel} 与 ${bLabel} 在 ${model.label} 上的每百万 token 成本。基于所属云服务商 TCO 归一化的芯片推理性能。`
       : `${aLabel} 与 ${bLabel} 在 ${model.label} 上的正面 AI 推理基准测试对比。`;
   const datasetName =
     variant === 'per-dollar'
diff --git a/packages/app/src/lib/compare-ssr.test.ts b/packages/app/src/lib/compare-ssr.test.ts
index b9888abf4..051b5fdd1 100644
--- a/packages/app/src/lib/compare-ssr.test.ts
+++ b/packages/app/src/lib/compare-ssr.test.ts
@@ -7,7 +7,9 @@ import { COMPARE_MODEL_SLUGS } from './compare-slug';
 import {
   compareMetaDescription,
   computeCompareImageRows,
+  computeCompareTableData,
   KNOWN_MODELS,
+  KNOWN_SEQUENCES,
   META_DESCRIPTION_MAX,
   type SsrInterpolatedRow,
 } from './compare-ssr';
@@ -21,6 +23,10 @@ describe('compare URL validators', () => {
   it('accepts GLM-5.2 as a model override', () => {
     expect(KNOWN_MODELS.has('GLM-5.2')).toBe(true);
   });
+
+  it('accepts AgentX as a comparison scenario override', () => {
+    expect(KNOWN_SEQUENCES.has('agentic-traces')).toBe(true);
+  });
 });
 
 function stubRow(overrides: Partial<BenchmarkRow> = {}): BenchmarkRow {
@@ -69,6 +75,60 @@ function pairRows(): BenchmarkRow[] {
   ];
 }
 
+function agenticPairRows(): BenchmarkRow[] {
+  return [
+    stubRow({
+      hardware: 'h200',
+      benchmark_type: 'agentic_traces',
+      isl: null,
+      osl: null,
+      conc: 16,
+      metrics: { tput_per_gpu: 800, p90_itl: 0.1 },
+    }),
+    stubRow({
+      hardware: 'h200',
+      benchmark_type: 'agentic_traces',
+      isl: null,
+      osl: null,
+      conc: 32,
+      metrics: { tput_per_gpu: 600, p90_itl: 0.05 },
+    }),
+    stubRow({
+      hardware: 'b200',
+      benchmark_type: 'agentic_traces',
+      isl: null,
+      osl: null,
+      conc: 16,
+      metrics: { tput_per_gpu: 900, p90_itl: 0.1 },
+    }),
+    stubRow({
+      hardware: 'b200',
+      benchmark_type: 'agentic_traces',
+      isl: null,
+      osl: null,
+      conc: 32,
+      metrics: { tput_per_gpu: 700, p90_itl: 0.05 },
+    }),
+  ];
+}
+
+describe('computeCompareTableData', () => {
+  it('builds AgentX comparison rows from p90 full-response interactivity', () => {
+    const result = computeCompareTableData(
+      agenticPairRows(),
+      'h200',
+      'b200',
+      'agentic-traces',
+      'fp8',
+    );
+
+    expect(result.interactivityRange).toEqual({ min: 10, max: 20 });
+    expect(result.defaultTargets).toEqual([13, 15, 18]);
+    expect(result.ssrRows).toHaveLength(3);
+    expect(result.ssrRows.every((row) => row.a !== null && row.b !== null)).toBe(true);
+  });
+});
+
 describe('computeCompareImageRows', () => {
   const range = { min: 10, max: 40 };
 
@@ -282,7 +342,7 @@ describe('compareMetaDescriptionZh — structural 1:1 port', () => {
     expect(zh).toContain('B300');
     expect(zh).toContain('34%');
     expect(zh).toContain('12%');
-    expect(zh).toContain('每 Chip 吞吐量');
+    expect(zh).toContain('每芯片吞吐量');
     expect(zh).toContain('每 token 成本');
     expect(zh.length).toBeLessThanOrEqual(META_DESCRIPTION_MAX);
   });
diff --git a/packages/app/src/lib/compare-ssr.ts b/packages/app/src/lib/compare-ssr.ts
index 41afbd63b..a1a9971d4 100644
--- a/packages/app/src/lib/compare-ssr.ts
+++ b/packages/app/src/lib/compare-ssr.ts
@@ -14,6 +14,7 @@ import {
   AUTHOR_NAME,
   AUTHOR_URL,
   HW_REGISTRY,
+  rowToSequence,
   SITE_NAME,
   SITE_URL,
   sequenceToIslOsl,
@@ -48,6 +49,7 @@ export const KNOWN_MODELS = new Set([
   'DeepSeek-Coder-V2-Lite-Instruct',
   'gpt-oss-120b',
   'Qwen-3.5-397B-A17B',
+  'Qwen3.8-Flash-Next',
   'Kimi-K2.5',
   'Kimi-K3',
   'MiniMax-M2.5',
@@ -57,9 +59,29 @@ export const KNOWN_MODELS = new Set([
   'DeepSeek-V4-Pro',
   'Gemma-4-31B',
 ]);
-export const KNOWN_SEQUENCES = new Set(['1k/1k', '1k/8k', '8k/1k']);
+export const KNOWN_SEQUENCES = new Set(['1k/1k', '1k/8k', '8k/1k', 'agentic-traces']);
 export const KNOWN_PRECISIONS = new Set(['fp4', 'fp4fp8', 'fp8', 'bf16', 'int4', 'nvfp4', 'mxfp4']);
 
+/**
+ * Three-sentence framing shown above the narrative when a compare page is
+ * rendering the agentic workload. The agentic scenario is a different kind of
+ * measurement from a fixed-sequence run, and a reader who arrived from a GPU
+ * query ("b200 vs b300") has no reason to know that yet — so say it once,
+ * here, and hand off to /agentx for the rest.
+ */
+export interface AgenticScenarioIntro {
+  paragraph: string;
+  linkLabel: string;
+  href: string;
+}
+
+export const AGENTIC_SCENARIO_INTRO: AgenticScenarioIntro = {
+  paragraph:
+    'AgentX replays real coding-agent sessions rather than fixed-length prompts, so context grows turn over turn and most of each request is served from cache instead of being recomputed. That turns the comparison into a systems question: KV transfer between nodes, prefix-aware routing, and cache capacity all move the curve alongside raw chip throughput. Fixed-sequence workloads stay the clean baseline for kernel and silicon performance, so the two scenarios answer different questions about the same hardware.',
+  linkLabel: 'Learn more about AgentX',
+  href: '/agentx',
+};
+
 export function pickString(value: string | string[] | undefined): string | undefined {
   if (typeof value === 'string') return value;
   if (Array.isArray(value)) return value[0];
@@ -170,6 +192,74 @@ export function buildGpuDataPoints(
   return points;
 }
 
+function buildAgenticGpuDataPoints(
+  rows: BenchmarkRow[],
+  hw: string,
+  precision: string,
+  specMethod?: string,
+): GPUDataPoint[] {
+  const points: GPUDataPoint[] = [];
+  for (const row of rows) {
+    if (row.hardware !== hw) continue;
+    if (rowToSequence(row) !== 'agentic-traces') continue;
+    if (row.precision !== precision) continue;
+    if (specMethod !== undefined && row.spec_method !== specMethod) continue;
+
+    const entry = rowToAggDataEntry(row);
+    const hwKey = getHardwareKey(entry);
+    if (!getHardwareConfig(hwKey)) continue;
+
+    const interactivity = entry.p90_intvty;
+    if (!Number.isFinite(interactivity) || interactivity <= 0) continue;
+
+    const tput = entry.tput_per_gpu;
+    const outputTput = entry.output_tput_per_gpu || tput;
+    const inputTput = entry.input_tput_per_gpu;
+    const specs = getGpuSpecs(hwKey);
+    const power = specs.power;
+
+    points.push({
+      hwKey,
+      interactivity,
+      throughput: tput,
+      outputThroughput: outputTput,
+      inputThroughput: inputTput,
+      concurrency: row.conc,
+      tp: row.decode_tp,
+      precision: row.precision,
+      ep: row.decode_ep,
+      dp_attention: row.decode_dp_attention,
+      disagg: row.disagg,
+      costh: computeGpuCost(specs.costh, tput),
+      costn: computeGpuCost(specs.costn, tput),
+      costr: computeGpuCost(specs.costr, tput),
+      costhi: computeGpuCost(specs.costh, inputTput),
+      costni: computeGpuCost(specs.costn, inputTput),
+      costri: computeGpuCost(specs.costr, inputTput),
+      costhOutput: computeGpuCost(specs.costh, outputTput),
+      costnOutput: computeGpuCost(specs.costn, outputTput),
+      costrOutput: computeGpuCost(specs.costr, outputTput),
+      tpPerMw: power && power > 0 ? (tput * 1000) / power : 0,
+      inputTpPerMw: power && power > 0 ? (inputTput * 1000) / power : 0,
+      outputTpPerMw: power && power > 0 ? (outputTput * 1000) / power : 0,
+    });
+  }
+  return points;
+}
+
+function buildGpuDataPointsForSequence(
+  rows: BenchmarkRow[],
+  hw: string,
+  sequence: string,
+  precision: string,
+): GPUDataPoint[] {
+  if (sequence === 'agentic-traces') {
+    return buildAgenticGpuDataPoints(rows, hw, precision);
+  }
+  const islOsl = sequenceToIslOsl(sequence);
+  return islOsl ? buildGpuDataPoints(rows, hw, islOsl.isl, islOsl.osl, precision) : [];
+}
+
 function interactivityRangeOf(pts: GPUDataPoint[]): { min: number; max: number } | null {
   if (pts.length === 0) return null;
   let min = Infinity;
@@ -197,11 +287,8 @@ export function computeCompareTableData(
   const empty = { defaultTargets: [], ssrRows: [], interactivityRange: { min: 0, max: 100 } };
   if (!sequence || !precision) return empty;
 
-  const islOsl = sequenceToIslOsl(sequence);
-  if (!islOsl) return empty;
-
-  const pointsA = buildGpuDataPoints(rows, a, islOsl.isl, islOsl.osl, precision);
-  const pointsB = buildGpuDataPoints(rows, b, islOsl.isl, islOsl.osl, precision);
+  const pointsA = buildGpuDataPointsForSequence(rows, a, sequence, precision);
+  const pointsB = buildGpuDataPointsForSequence(rows, b, sequence, precision);
 
   if (pointsA.length === 0 && pointsB.length === 0) return empty;
 
@@ -271,11 +358,8 @@ export function computeCompareImageRows(
 ): SsrInterpolatedRow[] {
   if (!sequence || !precision || interactivityRange.max <= interactivityRange.min) return [];
 
-  const islOsl = sequenceToIslOsl(sequence);
-  if (!islOsl) return [];
-
-  const pointsA = buildGpuDataPoints(rows, a, islOsl.isl, islOsl.osl, precision);
-  const pointsB = buildGpuDataPoints(rows, b, islOsl.isl, islOsl.osl, precision);
+  const pointsA = buildGpuDataPointsForSequence(rows, a, sequence, precision);
+  const pointsB = buildGpuDataPointsForSequence(rows, b, sequence, precision);
   if (pointsA.length === 0 && pointsB.length === 0) return [];
 
   const sampleCount = 17;
diff --git a/packages/app/src/lib/compare-variant-ssr-zh.ts b/packages/app/src/lib/compare-variant-ssr-zh.ts
index 4d48ea869..b3b6fe8a2 100644
--- a/packages/app/src/lib/compare-variant-ssr-zh.ts
+++ b/packages/app/src/lib/compare-variant-ssr-zh.ts
@@ -50,10 +50,10 @@ function variantFullSummaryZh(i: VariantBoth): string {
       ? null
       : `${i.cheaper} 每 token 成本低 ${fmtPctDelta(i.costRatio)}`;
   const tputPart = i.tputTied
-    ? '每 Chip 吞吐量基本持平'
+    ? '每芯片吞吐量基本持平'
     : i.tputRatio === null
       ? null
-      : `${i.faster} 每 Chip 吞吐量高出 ${fmtPctDelta(i.tputRatio)}`;
+      : `${i.faster} 每芯片吞吐量高出 ${fmtPctDelta(i.tputRatio)}`;
   const both = [costPart, tputPart].filter(Boolean).join('；');
   return both.length > 0 ? both : '差距极小，难以判定优劣';
 }
@@ -64,13 +64,13 @@ function variantFullSummaryZh(i: VariantBoth): string {
 
 const PRECISION_BOTH_TEMPLATES_ZH: ((i: VariantBoth) => string)[] = [
   (i) =>
-    `在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 交互性运行时，${i.aLabel} 吞吐量为 ${i.aValue.toFixed(0)} tok/s/Chip，每百万 token 成本 ${fmtCost(i.aCost)}；${i.bLabel} 吞吐量为 ${i.bValue.toFixed(0)} tok/s/Chip，成本 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。低精度量化以模型精度换取吞吐量——请查看评估页面了解质量影响。`,
+    `在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 交互性运行时，${i.aLabel} 吞吐量为 ${i.aValue.toFixed(0)} tok/s/chip，每百万 token 成本 ${fmtCost(i.aCost)}；${i.bLabel} 吞吐量为 ${i.bValue.toFixed(0)} tok/s/chip，成本 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。低精度量化以模型精度换取吞吐量——请查看评估页面了解质量影响。`,
   (i) =>
-    `${i.aLabel} 在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时达到 ${i.aValue.toFixed(0)} tok/s/Chip（每百万 token ${fmtCost(i.aCost)}）；${i.bLabel} 达到 ${i.bValue.toFixed(0)} tok/s/Chip（${fmtCost(i.bCost)}）。${variantFullSummaryZh(i)}。量化级别的精度差异在评估页面中跟踪。`,
+    `${i.aLabel} 在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时达到 ${i.aValue.toFixed(0)} tok/s/chip（每百万 token ${fmtCost(i.aCost)}）；${i.bLabel} 达到 ${i.bValue.toFixed(0)} tok/s/chip（${fmtCost(i.bCost)}）。${variantFullSummaryZh(i)}。量化级别的精度差异在评估页面中跟踪。`,
   (i) =>
-    `${i.modelLabel}（${i.gpuLabel}）在 ${i.target} tok/s/user 交互性下的吞吐量：${i.aLabel} 为 ${i.aValue.toFixed(0)} tok/s/Chip，${i.bLabel} 为 ${i.bValue.toFixed(0)}。每百万 token 成本分别为 ${fmtCost(i.aCost)} 和 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。低精度带来的成本-吞吐量权衡只是全貌的一部分——请参阅评估页面的精度数据。`,
+    `${i.modelLabel}（${i.gpuLabel}）在 ${i.target} tok/s/user 交互性下的吞吐量：${i.aLabel} 为 ${i.aValue.toFixed(0)} tok/s/chip，${i.bLabel} 为 ${i.bValue.toFixed(0)}。每百万 token 成本分别为 ${fmtCost(i.aCost)} 和 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。低精度带来的成本-吞吐量权衡只是全貌的一部分——请参阅评估页面的精度数据。`,
   (i) =>
-    `在 ${i.range} 交互性区间的${BAND_PHRASE_ZH[i.band]}，即 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时：${i.aLabel} 达到 ${i.aValue.toFixed(0)} tok/s/Chip（${fmtCost(i.aCost)}/百万 token），${i.bLabel} 达到 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}/百万）。${variantFullSummaryZh(i)}。精度变更同时影响推理速度和模型质量——请查阅评估页面的精度基准测试。`,
+    `在 ${i.range} 交互性区间的${BAND_PHRASE_ZH[i.band]}，即 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时：${i.aLabel} 达到 ${i.aValue.toFixed(0)} tok/s/chip（${fmtCost(i.aCost)}/百万 token），${i.bLabel} 达到 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}/百万）。${variantFullSummaryZh(i)}。精度变更同时影响推理速度和模型质量——请查阅评估页面的精度基准测试。`,
 ];
 
 // ---------------------------------------------------------------------------
@@ -79,13 +79,13 @@ const PRECISION_BOTH_TEMPLATES_ZH: ((i: VariantBoth) => string)[] = [
 
 const SPEC_DECODE_BOTH_TEMPLATES_ZH: ((i: VariantBoth) => string)[] = [
   (i) =>
-    `在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 交互性运行时，${i.aLabel} 吞吐量为 ${i.aValue.toFixed(0)} tok/s/Chip，每百万 token 成本 ${fmtCost(i.aCost)}；${i.bLabel} 吞吐量为 ${i.bValue.toFixed(0)} tok/s/Chip，成本 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。投机解码通过接受草稿 token 来降低每 token 延迟——收益因工作负载和提示分布而异。`,
+    `在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 交互性运行时，${i.aLabel} 吞吐量为 ${i.aValue.toFixed(0)} tok/s/chip，每百万 token 成本 ${fmtCost(i.aCost)}；${i.bLabel} 吞吐量为 ${i.bValue.toFixed(0)} tok/s/chip，成本 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。投机解码通过接受草稿 token 来降低每 token 延迟——收益因工作负载和提示分布而异。`,
   (i) =>
-    `${i.aLabel} 在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时达到 ${i.aValue.toFixed(0)} tok/s/Chip（每百万 token ${fmtCost(i.aCost)}）；${i.bLabel} 达到 ${i.bValue.toFixed(0)} tok/s/Chip（${fmtCost(i.bCost)}）。${variantFullSummaryZh(i)}。草稿 token 的接受率决定了投机解码在给定并发水平下是否有效。`,
+    `${i.aLabel} 在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时达到 ${i.aValue.toFixed(0)} tok/s/chip（每百万 token ${fmtCost(i.aCost)}）；${i.bLabel} 达到 ${i.bValue.toFixed(0)} tok/s/chip（${fmtCost(i.bCost)}）。${variantFullSummaryZh(i)}。草稿 token 的接受率决定了投机解码在给定并发水平下是否有效。`,
   (i) =>
-    `${i.modelLabel}（${i.gpuLabel}）在 ${i.target} tok/s/user 交互性下的吞吐量：${i.aLabel} 为 ${i.aValue.toFixed(0)} tok/s/Chip，${i.bLabel} 为 ${i.bValue.toFixed(0)}。每百万 token 成本分别为 ${fmtCost(i.aCost)} 和 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。投机解码以额外的草稿 token 计算换取更少的解码步骤——收益取决于序列长度和批大小。`,
+    `${i.modelLabel}（${i.gpuLabel}）在 ${i.target} tok/s/user 交互性下的吞吐量：${i.aLabel} 为 ${i.aValue.toFixed(0)} tok/s/chip，${i.bLabel} 为 ${i.bValue.toFixed(0)}。每百万 token 成本分别为 ${fmtCost(i.aCost)} 和 ${fmtCost(i.bCost)}。${variantFullSummaryZh(i)}。投机解码以额外的草稿 token 计算换取更少的解码步骤——收益取决于序列长度和批大小。`,
   (i) =>
-    `在 ${i.range} 交互性区间的${BAND_PHRASE_ZH[i.band]}，即 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时：${i.aLabel} 达到 ${i.aValue.toFixed(0)} tok/s/Chip（${fmtCost(i.aCost)}/百万 token），${i.bLabel} 达到 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}/百万）。${variantFullSummaryZh(i)}。投机解码的收益因工作负载而异；短输出提示获益通常较小。`,
+    `在 ${i.range} 交互性区间的${BAND_PHRASE_ZH[i.band]}，即 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时：${i.aLabel} 达到 ${i.aValue.toFixed(0)} tok/s/chip（${fmtCost(i.aCost)}/百万 token），${i.bLabel} 达到 ${i.bValue.toFixed(0)}（${fmtCost(i.bCost)}/百万）。${variantFullSummaryZh(i)}。投机解码的收益因工作负载而异；短输出提示获益通常较小。`,
 ];
 
 // Single-side templates -- Chinese (shared by both kinds)
@@ -99,11 +99,11 @@ const VARIANT_SINGLE_TEMPLATES_ZH: ((args: {
   presentCost: number;
 }) => string)[] = [
   (i) =>
-    `在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时，${i.presentLabel} 吞吐量为 ${i.presentValue.toFixed(0)} tok/s/Chip，每百万 token 成本 ${fmtCost(i.presentCost)}；${i.missingLabel} 在此目标点没有基准测试数据。`,
+    `在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时，${i.presentLabel} 吞吐量为 ${i.presentValue.toFixed(0)} tok/s/chip，每百万 token 成本 ${fmtCost(i.presentCost)}；${i.missingLabel} 在此目标点没有基准测试数据。`,
   (i) =>
-    `${i.presentLabel} 在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时达到 ${i.presentValue.toFixed(0)} tok/s/Chip（每百万 token ${fmtCost(i.presentCost)}）。${i.missingLabel} 在此工作点没有数据。`,
+    `${i.presentLabel} 在 ${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行时达到 ${i.presentValue.toFixed(0)} tok/s/chip（每百万 token ${fmtCost(i.presentCost)}）。${i.missingLabel} 在此工作点没有数据。`,
   (i) =>
-    `${i.presentLabel}：${i.presentValue.toFixed(0)} tok/s/Chip，每百万 token ${fmtCost(i.presentCost)}（${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行）。${i.missingLabel} 在此点尚未测试。`,
+    `${i.presentLabel}：${i.presentValue.toFixed(0)} tok/s/chip，每百万 token ${fmtCost(i.presentCost)}（${i.modelLabel}（${i.gpuLabel}）上以 ${i.target} tok/s/user 运行）。${i.missingLabel} 在此点尚未测试。`,
 ];
 
 // ---------------------------------------------------------------------------
diff --git a/packages/app/src/lib/constants.test.ts b/packages/app/src/lib/constants.test.ts
index f561591b6..794810f67 100644
--- a/packages/app/src/lib/constants.test.ts
+++ b/packages/app/src/lib/constants.test.ts
@@ -128,6 +128,38 @@ describe('getHardwareConfig', () => {
     expect(config.gpu).toBe("NVIDIA 'Blackwell' GB200 Dynamo TRTLLM");
   });
 
+  it('labels the July Vera Rubin snapshot without the hosting provider', () => {
+    const config = getHardwareConfig('vr200_rubin-july');
+    expect(config.label).toBe('Vera Rubin');
+    expect(config.suffix).toBe('(July)');
+    expect(config.alwaysShowPrecision).toBe(true);
+    expect(getHardwareConfig('vr200_coreweave-vera-rubin').suffix).toBe('');
+  });
+
+  it('keeps the Teacup attribution in the Jalapeño display label', () => {
+    const config = getHardwareConfig('jalapeno_teacup');
+    expect(config.label).toBe('Jalapeño');
+    expect(config.suffix).toBe('(Teacup)');
+    expect(HW_REGISTRY.jalapeno.badgeLabel).toBe('Jalapeño (Teacup)');
+  });
+
+  it('uses the published Jalapeño and VR200 power and TCO assumptions', () => {
+    expect(getGpuSpecs('jalapeno_teacup')).toEqual({
+      tdp: 700,
+      power: 1.125,
+      costh: 1.47,
+      costn: 1.56,
+      costr: 1.79,
+    });
+    expect(getGpuSpecs('vr200_coreweave-vera-rubin')).toEqual({
+      tdp: 1800,
+      power: 3.3,
+      costh: 3.61,
+      costn: 3.61,
+      costr: 3.61,
+    });
+  });
+
   it('returns unknown config when base GPU is not recognised', () => {
     const config = getHardwareConfig('completelynew');
     expect(config.label).toBe('Unknown');
@@ -224,6 +256,7 @@ describe('getHardwareConfig', () => {
 describe('getGpuSpecs', () => {
   it('returns specs for a base GPU key', () => {
     const specs = getGpuSpecs('h100');
+    expect(specs.tdp).toBe(700);
     expect(specs.power).toBe(1.37);
     expect(specs.costh).toBe(1.17);
     expect(specs.costn).toBe(1.55);
@@ -244,6 +277,7 @@ describe('getGpuSpecs', () => {
   it('returns zero specs for unknown GPU', () => {
     const specs = getGpuSpecs('nonexistent');
     expect(specs.power).toBe(0);
+    expect(specs.tdp).toBe(0);
     expect(specs.costh).toBe(0);
     expect(specs.costn).toBe(0);
     expect(specs.costr).toBe(0);
@@ -253,6 +287,7 @@ describe('getGpuSpecs', () => {
     for (const [base, entry] of Object.entries(HW_REGISTRY)) {
       const result = getGpuSpecs(base);
       expect(result.power).toBe(entry.power);
+      expect(result.tdp).toBe(entry.tdp);
       expect(result.costh).toBe(entry.costh);
     }
   });
diff --git a/packages/app/src/lib/constants.ts b/packages/app/src/lib/constants.ts
index a720077d9..7955cdec1 100644
--- a/packages/app/src/lib/constants.ts
+++ b/packages/app/src/lib/constants.ts
@@ -15,13 +15,14 @@ export const TABLEAU_10 = [
 ] as const;
 
 export interface GpuSpecs {
+  tdp: number;
   power: number;
   costh: number;
   costn: number;
   costr: number;
 }
 
-const DEFAULT_SPECS: GpuSpecs = { power: 0, costh: 0, costn: 0, costr: 0 };
+const DEFAULT_SPECS: GpuSpecs = { tdp: 0, power: 0, costh: 0, costn: 0, costr: 0 };
 
 /**
  * Look up power/cost specs for a hardware key by extracting the base GPU name.
@@ -31,7 +32,13 @@ export function getGpuSpecs(hwKey: string): GpuSpecs {
   const base = hwKey.split(/[-_]/u)[0];
   const entry = HW_REGISTRY[base];
   if (!entry) return DEFAULT_SPECS;
-  return { power: entry.power, costh: entry.costh, costn: entry.costn, costr: entry.costr };
+  return {
+    tdp: entry.tdp,
+    power: entry.power,
+    costh: entry.costh,
+    costn: entry.costn,
+    costr: entry.costr,
+  };
 }
 
 /** Build the vendor prefix string for the `gpu` tooltip field. */
@@ -47,6 +54,8 @@ export interface HardwareEntry {
   suffix: string;
   gpu: string;
   framework?: string;
+  /** Keep precision in curve labels even when the filter selects one precision. */
+  alwaysShowPrecision?: boolean;
 }
 
 const UNKNOWN_HARDWARE: HardwareEntry = {
@@ -71,13 +80,18 @@ function buildHardwareEntry(hwKey: string, model?: string): HardwareEntry | null
   const parts = hwKey.split('_').slice(1);
   const label = reg.label;
   const gpuName = base.toUpperCase(); // always raw uppercase for gpu string
-  const partLabels = parts.map((p) => resolveFrameworkPartLabel(model, p));
+  const isJulyRubinSnapshot = base === 'vr200' && parts[0] === 'rubin-july';
+  const partLabels =
+    base === 'vr200' && parts[0] === 'coreweave-vera-rubin'
+      ? parts.slice(1).map((p) => resolveFrameworkPartLabel(model, p))
+      : parts.map((p) => resolveFrameworkPartLabel(model, p));
 
   return {
     name: hwKey.replaceAll('_', '-'),
     label,
     suffix: partLabels.length > 0 ? `(${partLabels.join(', ')})` : '',
     gpu: [getVendorPrefix(base), gpuName, ...partLabels].join(' '),
+    ...(isJulyRubinSnapshot ? { alwaysShowPrecision: true } : {}),
   };
 }
 
diff --git a/packages/app/src/lib/cypress-timings.test.ts b/packages/app/src/lib/cypress-timings.test.ts
new file mode 100644
index 000000000..107a3bd26
--- /dev/null
+++ b/packages/app/src/lib/cypress-timings.test.ts
@@ -0,0 +1,33 @@
+import { readdirSync, readFileSync } from 'node:fs';
+import { resolve } from 'node:path';
+
+import { describe, expect, it } from 'vitest';
+
+interface TimingEntry {
+  spec: string;
+  duration: number;
+}
+
+interface TimingFile {
+  durations: TimingEntry[];
+}
+
+const appRoot = resolve(import.meta.dirname, '../..');
+
+describe('Cypress split timing baseline', () => {
+  it('contains one positive timing for every current integration spec', () => {
+    const specs = readdirSync(resolve(appRoot, 'cypress/e2e'))
+      .filter((name) => name.endsWith('.cy.ts') || name.endsWith('.cy.tsx'))
+      .map((name) => `cypress/e2e/${name}`)
+      .toSorted();
+    const timings = (
+      JSON.parse(readFileSync(resolve(appRoot, 'timings.json'), 'utf8')) as TimingFile
+    ).durations;
+    expect(timings.map(({ spec }) => spec).toSorted()).toEqual(specs);
+    expect(new Set(timings.map(({ spec }) => spec)).size).toBe(timings.length);
+    for (const { duration } of timings) {
+      expect(Number.isFinite(duration)).toBe(true);
+      expect(duration).toBeGreaterThan(0);
+    }
+  });
+});
diff --git a/packages/app/src/lib/d3-chart/D3Chart/D3Chart.tsx b/packages/app/src/lib/d3-chart/D3Chart/D3Chart.tsx
index 2a18b9cdc..6553b1611 100644
--- a/packages/app/src/lib/d3-chart/D3Chart/D3Chart.tsx
+++ b/packages/app/src/lib/d3-chart/D3Chart/D3Chart.tsx
@@ -25,6 +25,9 @@ function D3ChartInner<T>(
     grabCursor = true,
     instructions,
     clipContent,
+    dataIdentity,
+    metricIdentity,
+    displayIdentity,
     xScale: xScaleConfig,
     yScale: yScaleConfig,
     xAxis,
@@ -37,6 +40,7 @@ function D3ChartInner<T>(
     noDataOverlay,
     caption,
     onRender,
+    onDisplayUpdate,
   }: D3ChartProps<T>,
   ref: React.ForwardedRef<D3ChartHandle>,
 ) {
@@ -106,6 +110,9 @@ function D3ChartInner<T>(
       grabCursor,
       instructions,
       clipContent,
+      dataIdentity,
+      metricIdentity,
+      displayIdentity,
       xScale: xScaleConfig,
       yScale: yScaleConfig,
       xAxis,
@@ -117,6 +124,7 @@ function D3ChartInner<T>(
       legendElement,
       noDataOverlay,
       onRender,
+      onDisplayUpdate,
     },
     {
       svgRef,
@@ -128,6 +136,7 @@ function D3ChartInner<T>(
       isPinned,
       pinTooltip,
       dismissTooltip,
+      hideTooltipElements,
       createRulers,
       attachHandlers,
     },
diff --git a/packages/app/src/lib/d3-chart/D3Chart/layer-renderer.test.ts b/packages/app/src/lib/d3-chart/D3Chart/layer-renderer.test.ts
new file mode 100644
index 000000000..b94616000
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/D3Chart/layer-renderer.test.ts
@@ -0,0 +1,285 @@
+// @vitest-environment jsdom
+import { describe, expect, it, vi } from 'vitest';
+import * as d3 from 'd3';
+
+import { setupChartStructure } from '../chart-setup';
+import type { LayerConfig, RenderContext } from './types';
+import { renderLayer, updateLayerForDisplay, updateLayerForMetric } from './layer-renderer';
+
+interface PointDatum {
+  id: string;
+  precision: string;
+  x: number;
+  y: number;
+}
+
+function scatterLayer(data: PointDatum[]): LayerConfig<PointDatum> {
+  return {
+    type: 'scatter',
+    data,
+    keyFn: (point) => point.id,
+    config: {
+      getColor: () => '#000',
+      selectedPrecisions: ['fp8'],
+    },
+  };
+}
+
+describe('updateLayerForMetric', () => {
+  it('updates bound coordinates by stable key without replacing joined point nodes', () => {
+    const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg');
+    const layout = setupChartStructure(svg, {
+      chartId: 'metric-phase-test',
+      containerWidth: 400,
+      containerHeight: 300,
+      margin: { top: 10, right: 10, bottom: 20, left: 30 },
+      watermark: 'none',
+    });
+    const xScale = d3.scaleLinear().domain([0, 10]).range([0, layout.width]);
+    const yScale = d3.scaleLinear().domain([0, 10]).range([layout.height, 0]);
+    const tooltipElement = document.createElement('div');
+    const ctx: RenderContext = {
+      layout,
+      tooltipElement,
+      xScale,
+      yScale,
+      width: layout.width,
+      height: layout.height,
+    };
+    const initial = scatterLayer([
+      { id: 'a', precision: 'fp8', x: 1, y: 2 },
+      { id: 'b', precision: 'fp8', x: 3, y: 4 },
+    ]);
+    renderLayer(initial, layout.zoomGroup, xScale, yScale, layout, ctx);
+    const before = layout.zoomGroup.selectAll<SVGGElement, PointDatum>('.dot-group').nodes();
+
+    const updated = scatterLayer([
+      { id: 'a', precision: 'fp8', x: 2, y: 8 },
+      { id: 'b', precision: 'fp8', x: 4, y: 6 },
+    ]);
+    updateLayerForMetric(updated, layout.zoomGroup, xScale, yScale, layout, ctx);
+    const after = layout.zoomGroup.selectAll<SVGGElement, PointDatum>('.dot-group').nodes();
+
+    expect(after).toHaveLength(before.length);
+    expect(after[0]).toBe(before[0]);
+    expect(after[1]).toBe(before[1]);
+    const boundData = layout.zoomGroup.selectAll<SVGGElement, PointDatum>('.dot-group').data();
+    expect(boundData.map((point) => point.y)).toEqual([8, 6]);
+    expect(after[0].getAttribute('transform')).toBe(`translate(${xScale(2)},${yScale(8)})`);
+  });
+
+  it('returns a custom metric selection for fresh tooltip handler attachment', () => {
+    const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg');
+    const layout = setupChartStructure(svg, {
+      chartId: 'custom-metric-selection-test',
+      containerWidth: 400,
+      containerHeight: 300,
+      margin: { top: 10, right: 10, bottom: 20, left: 30 },
+      watermark: 'none',
+    });
+    const xScale = d3.scaleLinear().domain([0, 10]).range([0, layout.width]);
+    const yScale = d3.scaleLinear().domain([0, 10]).range([layout.height, 0]);
+    const ctx: RenderContext = {
+      layout,
+      tooltipElement: document.createElement('div'),
+      xScale,
+      yScale,
+      width: layout.width,
+      height: layout.height,
+    };
+    const data = [{ id: 'a', precision: 'fp8', x: 4, y: 6 }];
+    const layer = {
+      type: 'custom',
+      render: (group: typeof layout.zoomGroup) =>
+        group
+          .selectAll<SVGCircleElement, PointDatum>('.metric-point')
+          .data(data)
+          .join('circle')
+          .attr('class', 'metric-point'),
+    } satisfies LayerConfig<PointDatum>;
+
+    const selection = updateLayerForMetric(layer, layout.zoomGroup, xScale, yScale, layout, ctx);
+
+    expect(selection?.node()).toBe(layout.zoomGroup.select('.metric-point').node());
+  });
+});
+
+describe('updateLayerForDisplay', () => {
+  it('recolors points, labels, and rooflines without replacing marks or changing zoomed geometry', () => {
+    const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg');
+    const layout = setupChartStructure(svg, {
+      chartId: 'display-phase-test',
+      containerWidth: 400,
+      containerHeight: 300,
+      margin: { top: 10, right: 10, bottom: 20, left: 30 },
+      watermark: 'none',
+    });
+    const xScale = d3.scaleLinear().domain([0, 10]).range([0, layout.width]);
+    const yScale = d3.scaleLinear().domain([0, 10]).range([layout.height, 0]);
+    const ctx: RenderContext = {
+      layout,
+      tooltipElement: document.createElement('div'),
+      xScale,
+      yScale,
+      width: layout.width,
+      height: layout.height,
+    };
+    const data = [{ id: 'a', precision: 'fp8', x: 1, y: 2 }];
+    const initialPointLayer = {
+      type: 'scatter',
+      data,
+      keyFn: (point: PointDatum) => point.id,
+      config: {
+        getColor: () => '#000000',
+        getLabelText: () => 'point',
+        foreground: '#111111',
+        selectedPrecisions: ['fp8'],
+      },
+    } satisfies LayerConfig<PointDatum>;
+    const initialRooflineLayer = {
+      type: 'roofline',
+      rooflines: {
+        series: [
+          { x: 1, y: 2 },
+          { x: 4, y: 5 },
+        ],
+      },
+      config: { getColor: () => '#000000' },
+    } satisfies LayerConfig<PointDatum>;
+    renderLayer(initialPointLayer, layout.zoomGroup, xScale, yScale, layout, ctx);
+    renderLayer(initialRooflineLayer, layout.zoomGroup, xScale, yScale, layout, ctx);
+
+    const point = layout.zoomGroup.select<SVGGElement>('.dot-group');
+    const roofline = layout.zoomGroup.select<SVGPathElement>('.roofline-path');
+    const pointNode = point.node();
+    const rooflineNode = roofline.node();
+    point.attr('transform', 'translate(123,45)');
+    roofline.attr('d', 'M 123 45 L 234 56');
+
+    updateLayerForDisplay(
+      {
+        ...initialPointLayer,
+        config: {
+          ...initialPointLayer.config,
+          getColor: () => '#ff0000',
+          foreground: '#eeeeee',
+          hideLabels: true,
+        },
+      },
+      layout.zoomGroup,
+      ctx,
+    );
+    updateLayerForDisplay(
+      {
+        ...initialRooflineLayer,
+        config: { getColor: () => '#00ff00' },
+      },
+      layout.zoomGroup,
+      ctx,
+    );
+
+    expect(point.node()).toBe(pointNode);
+    expect(roofline.node()).toBe(rooflineNode);
+    expect(point.attr('transform')).toBe('translate(123,45)');
+    expect(roofline.attr('d')).toBe('M 123 45 L 234 56');
+    expect(point.select('.visible-shape').attr('fill')).toBe('#ff0000');
+    expect(point.select('.point-label').attr('fill')).toBe('#eeeeee');
+    expect(point.select('.point-label').style('display')).toBe('none');
+    expect(roofline.attr('stroke')).toBe('#00ff00');
+  });
+
+  it('recolors horizontal bars without repeating their join or geometry writes', () => {
+    const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg');
+    const layout = setupChartStructure(svg, {
+      chartId: 'bar-display-phase-test',
+      containerWidth: 400,
+      containerHeight: 300,
+      margin: { top: 10, right: 10, bottom: 20, left: 30 },
+      watermark: 'none',
+    });
+    const xScale = d3.scaleLinear().domain([0, 100]).range([0, layout.width]);
+    const yScale = d3.scaleBand().domain(['a']).range([layout.height, 0]);
+    const ctx: RenderContext = {
+      layout,
+      tooltipElement: document.createElement('div'),
+      xScale,
+      yScale,
+      width: layout.width,
+      height: layout.height,
+    };
+    const data = [{ id: 'a', value: 42 }];
+    const initialLayer = {
+      type: 'horizontalBar',
+      data,
+      config: {
+        getY: (datum: (typeof data)[number]) => datum.id,
+        getX: (datum: (typeof data)[number]) => datum.value,
+        getColor: () => '#111111',
+        keyFn: (datum: (typeof data)[number]) => datum.id,
+      },
+    } satisfies LayerConfig<(typeof data)[number]>;
+
+    renderLayer(initialLayer, layout.zoomGroup, xScale, yScale, layout, ctx);
+    const bar = layout.zoomGroup.select<SVGRectElement>('.bar');
+    const barNode = bar.node();
+    const geometry = {
+      x: bar.attr('x'),
+      y: bar.attr('y'),
+      width: bar.attr('width'),
+      height: bar.attr('height'),
+    };
+
+    updateLayerForDisplay(
+      {
+        ...initialLayer,
+        config: { ...initialLayer.config, getColor: () => '#abcdef' },
+      },
+      layout.zoomGroup,
+      ctx,
+    );
+
+    expect(bar.node()).toBe(barNode);
+    expect(bar.attr('fill')).toBe('#abcdef');
+    expect({
+      x: bar.attr('x'),
+      y: bar.attr('y'),
+      width: bar.attr('width'),
+      height: bar.attr('height'),
+    }).toEqual(geometry);
+  });
+
+  it('runs only a custom layer display updater for display invalidation', () => {
+    const svg = document.createElementNS('http://www.w3.org/2000/svg', 'svg');
+    const layout = setupChartStructure(svg, {
+      chartId: 'custom-display-phase-test',
+      containerWidth: 400,
+      containerHeight: 300,
+      margin: { top: 10, right: 10, bottom: 20, left: 30 },
+      watermark: 'none',
+    });
+    const xScale = d3.scaleLinear().domain([0, 10]).range([0, layout.width]);
+    const yScale = d3.scaleLinear().domain([0, 10]).range([layout.height, 0]);
+    const ctx: RenderContext = {
+      layout,
+      tooltipElement: document.createElement('div'),
+      xScale,
+      yScale,
+      width: layout.width,
+      height: layout.height,
+    };
+    const render = vi.fn();
+    const onDisplayUpdate = vi.fn();
+    const customLayer = {
+      type: 'custom',
+      key: 'display-only',
+      displayIdentity: 'visible',
+      render,
+      onDisplayUpdate,
+    } satisfies LayerConfig<PointDatum>;
+
+    updateLayerForDisplay(customLayer, layout.zoomGroup, ctx);
+
+    expect(onDisplayUpdate).toHaveBeenCalledOnce();
+    expect(render).not.toHaveBeenCalled();
+  });
+});
diff --git a/packages/app/src/lib/d3-chart/D3Chart/layer-renderer.ts b/packages/app/src/lib/d3-chart/D3Chart/layer-renderer.ts
index 05f2cd366..10cb8c947 100644
--- a/packages/app/src/lib/d3-chart/D3Chart/layer-renderer.ts
+++ b/packages/app/src/lib/d3-chart/D3Chart/layer-renderer.ts
@@ -1,14 +1,27 @@
-import type * as d3 from 'd3';
+import * as d3 from 'd3';
 
 import type { ChartLayout, ContinuousScale } from '../types';
+import { CHART_TYPE, px } from '../typography';
 import { renderBars, updateBarsOnZoom } from '../layers/bars';
-import { renderHorizontalBars, updateHorizontalBarsOnZoom } from '../layers/horizontal-bars';
+import {
+  renderHorizontalBars,
+  updateHorizontalBarsForDisplay,
+  updateHorizontalBarsOnZoom,
+} from '../layers/horizontal-bars';
 import { renderPoints, updatePointsOnZoom } from '../layers/points';
 import { renderErrorBars, updateErrorBarsOnZoom } from '../layers/error-bars';
 import { renderLines, updateLinesOnZoom } from '../layers/lines';
-import { renderRooflines, updateRooflinesOnZoom } from '../layers/rooflines';
+import {
+  renderRooflines,
+  updateRooflinesForDisplay,
+  updateRooflinesOnZoom,
+} from '../layers/rooflines';
 import { renderBarLabels, updateBarLabelsOnZoom } from '../layers/bar-labels';
-import { renderScatterPoints, updateScatterPointsOnZoom } from '../layers/scatter-points';
+import {
+  renderScatterPoints,
+  updateScatterPointsForDisplay,
+  updateScatterPointsOnZoom,
+} from '../layers/scatter-points';
 import { renderRadar } from '../layers/radar';
 
 import type { BuiltScale } from './scale-builders';
@@ -124,46 +137,64 @@ export function renderLayer<T>(
 }
 
 /**
- * Update a single layer on zoom with new scales.
+ * Cheap geometry updates that should follow every zoom event.
  */
-export function updateLayerOnZoom<T>(
+export function updateLayerPositionOnZoom<T>(
   layer: LayerConfig<T>,
   group: d3.Selection<SVGGElement, unknown, null, undefined>,
   xScale: BuiltScale,
-  _yScale: BuiltScale,
   newXScale: BuiltScale,
   newYScale: BuiltScale,
   layout: ChartLayout,
-  ctx: ZoomContext,
 ): void {
   const { height } = layout;
-
   switch (layer.type) {
     case 'bar': {
       updateBarsOnZoom(group, newYScale as ContinuousScale, height, layer.config.getY);
       break;
     }
-
     case 'horizontalBar': {
       updateHorizontalBarsOnZoom(group, newXScale as ContinuousScale, layer.config.getX);
       break;
     }
-
     case 'point': {
       const { getX, getY, getCx, getCy } = layer.config;
       updatePointsOnZoom(
         group,
-        getX ? (d: any) => (newXScale as ContinuousScale)(getX(d)) : getCx,
-        getY ? (d: any) => (newYScale as ContinuousScale)(getY(d)) : getCy,
+        getX ? (d: T) => (newXScale as ContinuousScale)(getX(d)) : getCx,
+        getY ? (d: T) => (newYScale as ContinuousScale)(getY(d)) : getCy,
       );
       break;
     }
-
     case 'errorBar': {
       updateErrorBarsOnZoom(group, layer.config);
       break;
     }
+    case 'scatter': {
+      updateScatterPointsOnZoom(group, newXScale as ContinuousScale, newYScale as ContinuousScale);
+      break;
+    }
+    default: {
+      break;
+    }
+  }
+}
 
+/**
+ * Expensive path, label, and annotation work. The renderer coalesces this
+ * phase through one animation frame.
+ */
+export function updateLayerDecorationOnZoom<T>(
+  layer: LayerConfig<T>,
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  xScale: BuiltScale,
+  newXScale: BuiltScale,
+  newYScale: BuiltScale,
+  layout: ChartLayout,
+  ctx: ZoomContext,
+): void {
+  const { height } = layout;
+  switch (layer.type) {
     case 'line': {
       updateLinesOnZoom(
         group,
@@ -174,7 +205,6 @@ export function updateLayerOnZoom<T>(
       );
       break;
     }
-
     case 'roofline': {
       updateRooflinesOnZoom(
         group,
@@ -184,7 +214,6 @@ export function updateLayerOnZoom<T>(
       );
       break;
     }
-
     case 'barLabel': {
       updateBarLabelsOnZoom(
         group,
@@ -196,22 +225,90 @@ export function updateLayerOnZoom<T>(
       );
       break;
     }
-
-    case 'scatter': {
-      updateScatterPointsOnZoom(group, newXScale as ContinuousScale, newYScale as ContinuousScale);
+    case 'custom': {
+      layer.onZoom?.(group, ctx);
       break;
     }
-
-    case 'radar': {
-      // Radar charts don't support zoom
+    default: {
       break;
     }
+  }
+}
 
-    case 'custom': {
-      if (layer.onZoom) {
-        layer.onZoom(group, ctx);
-      }
-      break;
-    }
+/**
+ * Update metric coordinates without re-running the scatter data join.
+ */
+export function updateLayerForMetric<T>(
+  layer: LayerConfig<T>,
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  xScale: BuiltScale,
+  yScale: BuiltScale,
+  layout: ChartLayout,
+  ctx: RenderContext,
+): d3.Selection<any, any, any, any> | null {
+  if (layer.type !== 'scatter') {
+    return renderLayer(layer, group, xScale, yScale, layout, ctx);
+  }
+
+  const keyFor = (datum: (typeof layer.data)[number], index: number): string =>
+    layer.keyFn ? layer.keyFn(datum) : String(index);
+  const nextByKey = new Map<string, (typeof layer.data)[number]>();
+  layer.data.forEach((datum, index) => nextByKey.set(keyFor(datum, index), datum));
+  const selection = group
+    .selectAll<SVGGElement, (typeof layer.data)[number]>('.dot-group')
+    .each((bound, index) => {
+      const next = nextByKey.get(keyFor(bound, index));
+      if (next) Object.assign(bound, next);
+    });
+  const getLabelText = layer.config.getLabelText;
+  const foreground = layer.config.foreground;
+  if (getLabelText && foreground) {
+    selection.each(function (datum) {
+      const lines = getLabelText(datum).split('\n');
+      const firstDy = -(0.8 + (lines.length - 1) * 1.1);
+      d3.select(this)
+        .selectAll<SVGTextElement, boolean>('.point-label')
+        .data([true])
+        .join('text')
+        .attr('class', 'point-label')
+        .attr('text-anchor', 'middle')
+        .attr('fill', foreground)
+        .attr('font-size', px(CHART_TYPE.dataLabel))
+        .attr('font-weight', '700')
+        .attr('pointer-events', 'none')
+        .style('display', layer.config.hideLabels ? 'none' : '')
+        .style('opacity', layer.config.hideLabels ? 0 : 1)
+        .selectAll<SVGTSpanElement, string>('tspan')
+        .data(lines)
+        .join('tspan')
+        .attr('x', 0)
+        .attr('dy', (_line, index) => (index === 0 ? `${firstDy}em` : '1.1em'))
+        .text((line) => line);
+    });
+  } else {
+    selection.selectAll('.point-label').remove();
+  }
+  updateScatterPointsOnZoom(group, xScale as ContinuousScale, yScale as ContinuousScale);
+  return selection;
+}
+
+/** Update scale-neutral decoration without rebuilding joins, scales, or paths. */
+export function updateLayerForDisplay<T>(
+  layer: LayerConfig<T>,
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  ctx: RenderContext,
+): void {
+  if (layer.type === 'scatter') {
+    updateScatterPointsForDisplay(group, layer.config);
+    return;
+  }
+  if (layer.type === 'horizontalBar') {
+    updateHorizontalBarsForDisplay(group, layer.config);
+    return;
+  }
+  if (layer.type === 'roofline') {
+    updateRooflinesForDisplay(group, layer.config);
+    return;
   }
+  if (layer.type === 'custom') layer.onDisplayUpdate?.(group, ctx);
 }
diff --git a/packages/app/src/lib/d3-chart/D3Chart/types.ts b/packages/app/src/lib/d3-chart/D3Chart/types.ts
index 9aff5161b..b822d2d45 100644
--- a/packages/app/src/lib/d3-chart/D3Chart/types.ts
+++ b/packages/app/src/lib/d3-chart/D3Chart/types.ts
@@ -101,6 +101,12 @@ export interface CustomLayerConfig extends LayerBase {
         ctx: RenderContext,
       ) => d3.Selection<any, any, any, any> | void)
     | null;
+  /** Scale-neutral state for this layer; changes automatically invalidate phase 4. */
+  displayIdentity?: string;
+  /** Cheap decoration update run when this layer's display identity changes. */
+  onDisplayUpdate?:
+    | ((zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>, ctx: RenderContext) => void)
+    | null;
   onZoom?:
     | ((zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>, ctx: ZoomContext) => void)
     | null;
@@ -114,8 +120,8 @@ export type LayerConfig<T = any> =
   | LineLayerConfig
   | RooflineLayerConfig
   | BarLabelLayerConfig<T>
-  | ScatterLayerConfig<any>
-  | RadarLayerConfig<any>
+  | ScatterLayerConfig<T & { precision: string; x: number; y: number }>
+  | RadarLayerConfig<T>
   | CustomLayerConfig;
 
 // ---------------------------------------------------------------------------
@@ -203,6 +209,9 @@ export interface RenderContext {
   tooltipElement: HTMLDivElement;
   xScale: AnyScale | d3.ScaleTime<number, number>;
   yScale: AnyScale | d3.ScaleTime<number, number>;
+  /** Scales currently rendered after applying the active zoom transform. */
+  renderedXScale?: AnyScale | d3.ScaleTime<number, number>;
+  renderedYScale?: AnyScale | d3.ScaleTime<number, number>;
   width: number;
   height: number;
   /** Transition duration in ms for animated scale/domain changes. */
@@ -255,6 +264,15 @@ export interface D3ChartProps<T = any> {
   instructions?: string;
   /** When false, chart structure uses the root `g` group instead of clip-pathed zoomGroup. */
   clipContent?: boolean;
+  /**
+   * Stable identity for the set of rendered points. When supplied, coordinate
+   * changes update bound data in place instead of repeating the full join.
+   */
+  dataIdentity?: string;
+  /** Coordinate/scale invalidation key for in-place metric updates. */
+  metricIdentity?: string;
+  /** Display-only invalidation key for built-in layers and `onDisplayUpdate`. */
+  displayIdentity?: string;
 
   xScale?: ScaleConfig;
   yScale?: ScaleConfig;
@@ -273,4 +291,6 @@ export interface D3ChartProps<T = any> {
 
   /** Called after all layers render. Useful for one-off DOM manipulations. */
   onRender?: (ctx: RenderContext) => void;
+  /** Runs only for display invalidation, after structure/data/metric phases. */
+  onDisplayUpdate?: (ctx: RenderContext) => void;
 }
diff --git a/packages/app/src/lib/d3-chart/D3Chart/useD3ChartRenderer.test.ts b/packages/app/src/lib/d3-chart/D3Chart/useD3ChartRenderer.test.ts
new file mode 100644
index 000000000..ef58eb671
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/D3Chart/useD3ChartRenderer.test.ts
@@ -0,0 +1,189 @@
+// @vitest-environment jsdom
+import * as d3 from 'd3';
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+import {
+  animateTransitionGeometry,
+  captureTransitionGeometry,
+  createZoomFrameBatcher,
+  metricRenderCallbackContext,
+} from './useD3ChartRenderer';
+import type { RenderContext } from './types';
+
+function transitionDuration(element: SVGElement): number | undefined {
+  return Object.values(
+    (
+      element as SVGElement & {
+        __transition?: Record<string, { duration: number }>;
+      }
+    ).__transition ?? {},
+  )[0]?.duration;
+}
+
+afterEach(() => {
+  d3.select(document.body).selectAll('*').interrupt('data-update');
+  document.body.innerHTML = '';
+});
+
+describe('createZoomFrameBatcher', () => {
+  it('coalesces a burst and executes only its final transform work', () => {
+    const frames: FrameRequestCallback[] = [];
+    const requestFrame = vi.fn((callback: FrameRequestCallback) => {
+      frames.push(callback);
+      return frames.length;
+    });
+    const cancelFrame = vi.fn();
+    const batcher = createZoomFrameBatcher(requestFrame, cancelFrame);
+    const first = vi.fn();
+    const final = vi.fn();
+
+    batcher.schedule(first);
+    batcher.schedule(final);
+
+    expect(requestFrame).toHaveBeenCalledTimes(1);
+    frames[0](16);
+    expect(first).not.toHaveBeenCalled();
+    expect(final).toHaveBeenCalledTimes(1);
+  });
+
+  it('cancels pending work and permits a later frame', () => {
+    const frames: FrameRequestCallback[] = [];
+    const requestFrame = vi.fn((callback: FrameRequestCallback) => {
+      frames.push(callback);
+      return frames.length;
+    });
+    const cancelFrame = vi.fn();
+    const batcher = createZoomFrameBatcher(requestFrame, cancelFrame);
+    const cancelled = vi.fn();
+    const later = vi.fn();
+
+    batcher.schedule(cancelled);
+    batcher.cancel();
+    batcher.schedule(later);
+
+    expect(cancelFrame).toHaveBeenCalledWith(1);
+    expect(requestFrame).toHaveBeenCalledTimes(2);
+    frames[1](32);
+    expect(cancelled).not.toHaveBeenCalled();
+    expect(later).toHaveBeenCalledTimes(1);
+  });
+
+  it('flushes the latest pending work synchronously and cancels its frame', () => {
+    const frames: FrameRequestCallback[] = [];
+    const requestFrame = vi.fn((callback: FrameRequestCallback) => {
+      frames.push(callback);
+      return frames.length;
+    });
+    const cancelFrame = vi.fn();
+    const batcher = createZoomFrameBatcher(requestFrame, cancelFrame);
+    const first = vi.fn();
+    const final = vi.fn();
+
+    batcher.schedule(first);
+    batcher.schedule(final);
+    batcher.flush();
+
+    expect(cancelFrame).toHaveBeenCalledWith(1);
+    expect(first).not.toHaveBeenCalled();
+    expect(final).toHaveBeenCalledTimes(1);
+    frames[0](16);
+    expect(final).toHaveBeenCalledTimes(1);
+  });
+});
+
+describe('metricRenderCallbackContext', () => {
+  it('keeps base scales for decorations and exposes the zoomed rendered scales', () => {
+    const baseXScale = { id: 'base-x' };
+    const baseYScale = { id: 'base-y' };
+    const renderedXScale = { id: 'rendered-x' };
+    const renderedYScale = { id: 'rendered-y' };
+    const baseContext = {
+      layout: {},
+      tooltipElement: {},
+      xScale: baseXScale,
+      yScale: baseYScale,
+      width: 400,
+      height: 300,
+    };
+
+    const context = metricRenderCallbackContext(
+      baseContext as unknown as RenderContext,
+      renderedXScale as unknown as RenderContext['xScale'],
+      renderedYScale as unknown as RenderContext['yScale'],
+    );
+
+    expect(context.xScale).toBe(baseXScale);
+    expect(context.yScale).toBe(baseYScale);
+    expect(context.renderedXScale).toBe(renderedXScale);
+    expect(context.renderedYScale).toBe(renderedYScale);
+  });
+});
+
+describe('coordinate transitions', () => {
+  it('animates visible official geometry for 300ms while hidden and overlay marks snap', () => {
+    const svg = d3.select(document.body).append('svg');
+    const group = svg.append('g') as unknown as d3.Selection<SVGGElement, unknown, any, any>;
+    const point = (className: string, opacity?: number) => {
+      const selection = group
+        .append('g')
+        .attr('class', className)
+        .attr('transform', 'translate(0,0)');
+      if (opacity !== undefined) selection.style('opacity', opacity);
+      return selection.node()!;
+    };
+    const roofline = (className: string, opacity?: number) => {
+      const selection = group.append('path').attr('class', className).attr('d', 'M0,0L1,1');
+      if (opacity !== undefined) selection.style('opacity', opacity);
+      return selection.node()!;
+    };
+    const visibleOfficialPoint = point('dot-group');
+    const hiddenOfficialPoint = point('dot-group', 0);
+    const visibleOverlayPoint = point('unofficial-overlay-pt');
+    const hiddenOverlayPoint = point('unofficial-overlay-pt', 0);
+    const visibleOfficialRoofline = roofline('roofline-path');
+    const hiddenOfficialRoofline = roofline('roofline-path', 0);
+    const visibleOverlayRoofline = roofline('overlay-roofline-path');
+    const hiddenOverlayRoofline = roofline('overlay-roofline-path', 0);
+    const previous = captureTransitionGeometry(group);
+
+    expect(previous.transforms.size).toBe(1);
+    expect(previous.paths.size).toBe(1);
+
+    for (const element of [
+      visibleOfficialPoint,
+      hiddenOfficialPoint,
+      visibleOverlayPoint,
+      hiddenOverlayPoint,
+    ]) {
+      element.setAttribute('transform', 'translate(100,100)');
+    }
+    for (const element of [
+      visibleOfficialRoofline,
+      hiddenOfficialRoofline,
+      visibleOverlayRoofline,
+      hiddenOverlayRoofline,
+    ]) {
+      element.setAttribute('d', 'M100,100L200,200');
+    }
+
+    animateTransitionGeometry(group, previous, 300);
+
+    expect(visibleOfficialPoint.getAttribute('transform')).toBe('translate(0,0)');
+    expect(visibleOverlayPoint.getAttribute('transform')).toBe('translate(100,100)');
+    expect(hiddenOfficialPoint.getAttribute('transform')).toBe('translate(100,100)');
+    expect(hiddenOverlayPoint.getAttribute('transform')).toBe('translate(100,100)');
+    expect(visibleOfficialRoofline.getAttribute('d')).toBe('M0,0L1,1');
+    expect(visibleOverlayRoofline.getAttribute('d')).toBe('M100,100L200,200');
+    expect(hiddenOfficialRoofline.getAttribute('d')).toBe('M100,100L200,200');
+    expect(hiddenOverlayRoofline.getAttribute('d')).toBe('M100,100L200,200');
+
+    expect(transitionDuration(visibleOfficialPoint)).toBe(300);
+    expect(transitionDuration(visibleOverlayPoint)).toBeUndefined();
+    expect(transitionDuration(visibleOfficialRoofline)).toBe(300);
+    expect(transitionDuration(visibleOverlayRoofline)).toBeUndefined();
+    expect(transitionDuration(hiddenOfficialPoint)).toBeUndefined();
+    expect(transitionDuration(hiddenOverlayPoint)).toBeUndefined();
+    expect(transitionDuration(hiddenOfficialRoofline)).toBeUndefined();
+    expect(transitionDuration(hiddenOverlayRoofline)).toBeUndefined();
+  });
+});
diff --git a/packages/app/src/lib/d3-chart/D3Chart/useD3ChartRenderer.ts b/packages/app/src/lib/d3-chart/D3Chart/useD3ChartRenderer.ts
index fdcd415e4..e617c560a 100644
--- a/packages/app/src/lib/d3-chart/D3Chart/useD3ChartRenderer.ts
+++ b/packages/app/src/lib/d3-chart/D3Chart/useD3ChartRenderer.ts
@@ -2,14 +2,24 @@ import { useLayoutEffect, useRef } from 'react';
 import * as d3 from 'd3';
 import { getDomainAwareChartWatermark } from '@/lib/unofficial-domain';
 
-import { computeTooltipPosition } from '../layers/scatter-points';
+import {
+  computeTooltipPosition,
+  getTooltipContainerGeometry,
+  invalidateTooltipGeometry,
+} from '../layers/scatter-points';
 import { setupChartStructure } from '../chart-setup';
 import { renderAxes, renderGrid, type AnyScale } from '../chart-update';
 import type { ChartLayout, ContinuousScale } from '../types';
 
 import { buildScale, isBandScale, type BuiltScale } from './scale-builders';
-import { renderLayer, updateLayerOnZoom } from './layer-renderer';
-import type { AxisConfig, D3ChartProps, RenderContext, ZoomContext } from './types';
+import {
+  renderLayer,
+  updateLayerDecorationOnZoom,
+  updateLayerForDisplay,
+  updateLayerForMetric,
+  updateLayerPositionOnZoom,
+} from './layer-renderer';
+import type { AxisConfig, D3ChartProps, LayerConfig, RenderContext, ZoomContext } from './types';
 
 interface RendererDeps {
   svgRef: React.RefObject<SVGSVGElement | null>;
@@ -28,6 +38,10 @@ interface RendererDeps {
   isPinned: () => boolean;
   pinTooltip: (data: any, isOverlay?: boolean) => void;
   dismissTooltip: (clearPinnedPoint?: boolean) => void;
+  hideTooltipElements: (
+    tooltipRef: React.RefObject<HTMLDivElement | null>,
+    svgRef: React.RefObject<SVGSVGElement | null>,
+  ) => void;
   createRulers: (
     group: d3.Selection<SVGGElement, unknown, null, undefined>,
     rulerType: 'vertical' | 'horizontal' | 'crosshair' | 'none',
@@ -59,10 +73,151 @@ function resolveTickValues(
   if (!tickValues) return undefined;
   return typeof tickValues === 'function' ? tickValues(scale) : tickValues;
 }
+export interface ZoomFrameBatcher {
+  schedule: (work: () => void) => void;
+  flush: () => void;
+  cancel: () => void;
+}
+
+/** Coalesces expensive zoom work and always runs the latest submitted closure. */
+export function createZoomFrameBatcher(
+  requestFrame: (callback: FrameRequestCallback) => number,
+  cancelFrame: (id: number) => void,
+): ZoomFrameBatcher {
+  let frameId: number | null = null;
+  let latestWork: (() => void) | null = null;
+
+  return {
+    schedule(work) {
+      latestWork = work;
+      if (frameId !== null) return;
+      frameId = requestFrame(() => {
+        frameId = null;
+        const finalWork = latestWork;
+        latestWork = null;
+        finalWork?.();
+      });
+    },
+    flush() {
+      if (frameId !== null) cancelFrame(frameId);
+      frameId = null;
+      const finalWork = latestWork;
+      latestWork = null;
+      finalWork?.();
+    },
+    cancel() {
+      if (frameId !== null) cancelFrame(frameId);
+      frameId = null;
+      latestWork = null;
+    },
+  };
+}
+
+export interface TransitionGeometry {
+  transforms: Map<SVGGElement, string>;
+  paths: Map<SVGPathElement, string>;
+}
+
+const TRANSITION_POINT_SELECTOR = '.dot-group';
+const TRANSITION_ROOFLINE_SELECTOR = '.roofline-path';
+type GeometryVisibility = (element: SVGElement) => boolean;
+
+function isGeometryVisible(element: SVGElement): boolean {
+  return element.style.opacity !== '0' && element.style.display !== 'none';
+}
+
+function metricGeometryVisibility<T>(layers: LayerConfig<T>[]): GeometryVisibility {
+  const scatterLayer = layers.find((layer) => layer.type === 'scatter');
+  const getOpacity = scatterLayer?.type === 'scatter' ? scatterLayer.config.getOpacity : undefined;
+  if (!getOpacity) return isGeometryVisible;
+
+  return (element) =>
+    element.matches('.dot-group')
+      ? getOpacity(d3.select(element).datum() as Parameters<typeof getOpacity>[0]) > 0
+      : isGeometryVisible(element);
+}
+
+export function captureTransitionGeometry(
+  group: d3.Selection<SVGGElement, unknown, any, any>,
+): TransitionGeometry {
+  const transforms = new Map<SVGGElement, string>();
+  const paths = new Map<SVGPathElement, string>();
+  group.selectAll<SVGGElement, unknown>(TRANSITION_POINT_SELECTOR).each(function () {
+    if (isGeometryVisible(this)) transforms.set(this, this.getAttribute('transform') || '');
+  });
+  group.selectAll<SVGPathElement, unknown>(TRANSITION_ROOFLINE_SELECTOR).each(function () {
+    if (isGeometryVisible(this)) paths.set(this, this.getAttribute('d') || '');
+  });
+  return { transforms, paths };
+}
+
+export function animateTransitionGeometry(
+  group: d3.Selection<SVGGElement, unknown, any, any>,
+  previous: TransitionGeometry | null,
+  durationMs: number | undefined = 0,
+  isVisible: GeometryVisibility = isGeometryVisible,
+): void {
+  if (!previous || durationMs <= 0) return;
+  group.selectAll<SVGGElement, unknown>(TRANSITION_POINT_SELECTOR).each(function () {
+    if (!isVisible(this)) {
+      d3.select(this).interrupt('data-update');
+      return;
+    }
+    const oldPosition = previous.transforms.get(this);
+    const newPosition = this.getAttribute('transform');
+    if (oldPosition !== undefined && newPosition && oldPosition !== newPosition) {
+      this.setAttribute('transform', oldPosition);
+      d3.select(this).transition('data-update').duration(durationMs).attr('transform', newPosition);
+    }
+  });
+  group.selectAll<SVGPathElement, unknown>(TRANSITION_ROOFLINE_SELECTOR).each(function () {
+    if (!isVisible(this)) {
+      d3.select(this).interrupt('data-update');
+      return;
+    }
+    const oldPath = previous.paths.get(this);
+    const newPath = this.getAttribute('d');
+    if (oldPath !== undefined && newPath && oldPath !== newPath) {
+      this.setAttribute('d', oldPath);
+      d3.select(this).transition('data-update').duration(durationMs).attr('d', newPath);
+    }
+  });
+}
+export function metricRenderCallbackContext(
+  baseContext: RenderContext,
+  renderedXScale: RenderContext['xScale'],
+  renderedYScale: RenderContext['yScale'],
+): RenderContext {
+  return {
+    ...baseContext,
+    renderedXScale,
+    renderedYScale,
+  };
+}
+
+function customLayerDisplayIdentities<T>(layers: LayerConfig<T>[]): Map<string, string> {
+  const identities = new Map<string, string>();
+  layers.forEach((layer, index) => {
+    if (layer.type === 'custom' && layer.displayIdentity !== undefined) {
+      identities.set(layer.key ?? `custom:${index}`, layer.displayIdentity);
+    }
+  });
+  return identities;
+}
+
+function customLayerDisplayPlanIdentity<T>(layers: LayerConfig<T>[]): string {
+  return layers
+    .map((layer, index) =>
+      layer.type === 'custom' && layer.displayIdentity !== undefined
+        ? `${layer.key ?? `custom:${index}`}:${layer.displayIdentity}`
+        : '',
+    )
+    .join('|');
+}
 
 /**
- * Core render effect for D3Chart. Builds scales, renders structure/axes/grid/layers,
- * wires up tooltip and zoom handlers.
+ * Coordinates structure, stable-identity data joins, metric updates, display
+ * invalidation, tooltip wiring, and zoom work for D3Chart.
  */
 export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps): void {
   const {
@@ -71,6 +226,9 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
     margin = { top: 24, right: 10, bottom: 40, left: 60 },
     watermark = 'logo',
     clipContent = true,
+    dataIdentity,
+    metricIdentity,
+    displayIdentity,
     xScale: xScaleConfig,
     yScale: yScaleConfig,
     xAxis: xAxisConfig,
@@ -80,6 +238,7 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
     tooltip: tooltipConfig,
     transitionDuration,
     onRender,
+    onDisplayUpdate,
   } = props;
 
   const {
@@ -92,6 +251,7 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
     isPinned,
     pinTooltip,
     dismissTooltip,
+    hideTooltipElements,
     createRulers,
     attachHandlers,
   } = deps;
@@ -101,17 +261,90 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
   const layoutRef = useRef<ChartLayout | null>(null);
   const prevDataRef = useRef(data);
   const prevScalesRef = useRef({ xScaleConfig, yScaleConfig });
-
-  // useLayoutEffect ensures D3 renders synchronously before browser paint,
-  // preventing a frame where dots and lines are out of sync during y-axis metric changes.
+  const prevYAxisConfigRef = useRef(yAxisConfig);
+  const zoomFrameBatcherRef = useRef<ZoomFrameBatcher | null>(null);
+  if (zoomFrameBatcherRef.current === null) {
+    zoomFrameBatcherRef.current = createZoomFrameBatcher(
+      (callback) => requestAnimationFrame(callback),
+      (id) => cancelAnimationFrame(id),
+    );
+  }
+  const renderContextRef = useRef<RenderContext | null>(null);
+  const joinedIdentityRef = useRef<string | null>(null);
+  const lastMetricIdentityRef = useRef<string | undefined>(undefined);
+  const lastDisplayIdentityRef = useRef<string | undefined>(undefined);
+  const customLayerDisplayIdentitiesRef = useRef<Map<string, string>>(new Map());
+  const layersRef = useRef(layers);
+  const axesRef = useRef({ xAxisConfig, yAxisConfig });
+  const zoomConfigRef = useRef(zoomConfig);
+  layersRef.current = layers;
+  axesRef.current = { xAxisConfig, yAxisConfig };
+  zoomConfigRef.current = zoomConfig;
+  const displayCallbackRef = useRef(onDisplayUpdate);
+  displayCallbackRef.current = onDisplayUpdate;
+  const customDisplayPlanIdentity = customLayerDisplayPlanIdentity(layers);
+  const hasScales =
+    xScaleConfig !== null &&
+    xScaleConfig !== undefined &&
+    yScaleConfig !== null &&
+    yScaleConfig !== undefined;
+  const hasRenderableData = data.length > 0 || layers.some((layer) => layer.type === 'custom');
+  const dataJoinIdentity = dataIdentity ?? data;
+  const dataPhaseXScale = dataIdentity ? null : xScaleConfig;
+  const dataPhaseYScale = dataIdentity ? null : yScaleConfig;
+  const dataPhaseLayers = dataIdentity ? null : layers;
+  const dataPhaseTooltip = dataIdentity ? null : tooltipConfig;
+
+  // Phase 1: SVG structure. Metric changes may refresh labels here, but never
+  // touch joined marks.
   useLayoutEffect(() => {
-    if (!svgRef.current || !tooltipRef.current || dimensions.width === 0) return;
-    if (data.length === 0 && layers.every((layer) => layer.type !== 'custom')) {
+    if (!svgRef.current || dimensions.width === 0) return;
+    if (!hasRenderableData) {
       d3.select(svgRef.current).selectAll('*').remove();
       scalesRef.current = null;
       layoutRef.current = null;
+      renderContextRef.current = null;
       dismissTooltip(true);
-      prevDataRef.current = data;
+      return;
+    }
+
+    layoutRef.current = setupChartStructure(svgRef.current, {
+      chartId,
+      containerWidth: dimensions.width,
+      containerHeight: dimensions.height,
+      margin,
+      watermark: getDomainAwareChartWatermark(watermark, window.location.hostname),
+      xLabel: xAxisConfig?.label,
+      yLabel: yAxisConfig?.label,
+      clipContent,
+      hideAxes: !hasScales,
+    });
+  }, [
+    chartId,
+    dimensions.width,
+    dimensions.height,
+    margin.top,
+    margin.right,
+    margin.bottom,
+    margin.left,
+    watermark,
+    xAxisConfig?.label,
+    yAxisConfig?.label,
+    clipContent,
+    hasScales,
+    hasRenderableData,
+  ]);
+
+  // Phase 2: full data join. With a supplied identity this runs only when the
+  // point set changes, not when coordinates or display state change.
+  useLayoutEffect(() => {
+    if (
+      !svgRef.current ||
+      !tooltipRef.current ||
+      dimensions.width === 0 ||
+      !hasRenderableData ||
+      !layoutRef.current
+    ) {
       return;
     }
 
@@ -122,6 +355,7 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
       yScaleConfig !== prevScalesRef.current.yScaleConfig;
     prevDataRef.current = data;
     prevScalesRef.current = { xScaleConfig, yScaleConfig };
+    prevYAxisConfigRef.current = yAxisConfig;
 
     {
       if (!svgRef.current || !tooltipRef.current) return;
@@ -130,39 +364,14 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
       zoomTransformRef.current = d3.zoomTransform(svgRef.current);
 
       // ── Save old positions for animated transitions ──
-      const oldTransforms = new Map<SVGGElement, string>();
-      const oldPaths = new Map<SVGPathElement, string>();
-      if (transitionDuration && (dataChanged || scalesChanged)) {
-        const existingGroup = d3.select(svgRef.current).select('.zoom-group');
-        if (!existingGroup.empty()) {
-          existingGroup.selectAll<SVGGElement, unknown>('.dot-group').each(function () {
-            oldTransforms.set(this, this.getAttribute('transform') || '');
-          });
-          existingGroup.selectAll<SVGPathElement, unknown>('.roofline-path').each(function () {
-            oldPaths.set(this, this.getAttribute('d') || '');
-          });
-        }
-      }
-
-      // ── Structure setup ──
-      const hasScales =
-        xScaleConfig !== null &&
-        xScaleConfig !== undefined &&
-        yScaleConfig !== null &&
-        yScaleConfig !== undefined;
-      const layout = setupChartStructure(svgRef.current, {
-        chartId,
-        containerWidth: dimensions.width,
-        containerHeight: dimensions.height,
-        margin,
-        watermark: getDomainAwareChartWatermark(watermark, window.location.hostname),
-        xLabel: xAxisConfig?.label,
-        yLabel: yAxisConfig?.label,
-        clipContent,
-        hideAxes: !hasScales,
-      });
-      layoutRef.current = layout;
-
+      const existingGroup = d3.select(svgRef.current).select<SVGGElement>('.zoom-group');
+      const previousGeometry =
+        transitionDuration && (dataChanged || scalesChanged) && !existingGroup.empty()
+          ? captureTransitionGeometry(existingGroup)
+          : null;
+
+      const layout = layoutRef.current;
+      if (!layout) return;
       const { width, height, zoomGroup, g } = layout;
       const renderGroup = clipContent ? zoomGroup : g;
       const tooltip = d3.select(tooltipRef.current);
@@ -213,10 +422,13 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
         tooltipElement: tooltipRef.current,
         xScale,
         yScale,
+        renderedXScale: xScale,
+        renderedYScale: yScale,
         width,
         height,
         transitionDuration,
       };
+      renderContextRef.current = ctx;
 
       // ── Render layers ──
       const layerSelections: (d3.Selection<any, any, any, any> | null)[] = [];
@@ -224,6 +436,8 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
         const sel = renderLayer(layer, renderGroup, xScale, yScale, layout, ctx);
         layerSelections.push(sel);
       }
+      customLayerDisplayIdentitiesRef.current = customLayerDisplayIdentities(layers);
+      lastDisplayIdentityRef.current = displayIdentity;
 
       // Ensure points render above lines/rooflines on re-renders
       // (D3 enter appends new elements at the end, so new lines can end up after existing dots)
@@ -291,12 +505,13 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
                 .style('display', 'block')
                 .style('pointer-events', 'none')
                 .html(tooltipConfig.content(d, false));
+              invalidateTooltipGeometry(tooltip.node());
 
               // Position tooltip near mouse
-              const rect = containerEl.getBoundingClientRect();
-              const cmx = event.clientX - rect.left;
-              const cmy = event.clientY - rect.top;
-              const pos = computeTooltipPosition(cmx, cmy, tooltip, containerEl);
+              const geometry = getTooltipContainerGeometry(containerEl);
+              const cmx = event.clientX - geometry.bounds.left;
+              const cmy = event.clientY - geometry.bounds.top;
+              const pos = computeTooltipPosition(cmx, cmy, tooltip, containerEl, 10, geometry);
               tooltip.style('left', `${pos.left}px`).style('top', `${pos.top}px`);
 
               // Position rulers
@@ -333,11 +548,12 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
               if (!d) return;
 
               event.stopPropagation();
-              const rect = containerEl.getBoundingClientRect();
-              const cmx = event.clientX - rect.left;
-              const cmy = event.clientY - rect.top;
+              const geometry = getTooltipContainerGeometry(containerEl);
+              const cmx = event.clientX - geometry.bounds.left;
+              const cmy = event.clientY - geometry.bounds.top;
               tooltip.html(tooltipConfig.content(d, true));
-              const pos = computeTooltipPosition(cmx, cmy, tooltip, containerEl);
+              invalidateTooltipGeometry(tooltip.node());
+              const pos = computeTooltipPosition(cmx, cmy, tooltip, containerEl, 10, geometry);
               tooltip
                 .style('left', `${pos.left}px`)
                 .style('top', `${pos.top}px`)
@@ -406,90 +622,107 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
             customTransformStorage: zoomConfig.customTransformStorage,
             onZoom: (event: d3.D3ZoomEvent<SVGSVGElement, unknown>) => {
               const transform = event.transform;
+              const currentScales = scalesRef.current;
+              const currentLayout = layoutRef.current;
+              const currentCtx = renderContextRef.current;
+              if (!currentScales || !currentLayout || !currentCtx) return;
+              const { xScale: zoomXScale, yScale: zoomYScale } = currentScales;
+              const zoomLayout = currentLayout;
+              const zoomLayers = layersRef.current;
+              const { xAxisConfig: zoomXAxisConfig, yAxisConfig: zoomYAxisConfig } =
+                axesRef.current;
+              const currentZoomConfig = zoomConfigRef.current;
+              const currentZoomAxes = currentZoomConfig?.axes ?? 'both';
+              const zoomRenderGroup = clipContent ? zoomLayout.zoomGroup : zoomLayout.g;
+              const zoomBaseContext = currentCtx;
 
-              // Dismiss tooltip on zoom
               if (isPinned()) {
                 dismissTooltip(true);
-                tooltip
-                  .style('opacity', 0)
-                  .style('display', 'none')
-                  .style('pointer-events', 'none');
-                renderGroup.select('.ruler-group').style('display', 'none');
+                hideTooltipElements(tooltipRef, svgRef);
               }
 
-              // Compute new scales
-              let newXScale: BuiltScale = xScale;
-              let newYScale: BuiltScale = yScale;
-
-              if (zoomAxes === 'x' || zoomAxes === 'both') {
-                if (zoomConfig.rescaleX && !isBandScale(xScale)) {
-                  newXScale = zoomConfig.rescaleX(xScale as ContinuousScale, transform);
-                } else if (!isBandScale(xScale)) {
-                  newXScale = transform.rescaleX(xScale as any);
+              let newXScale: BuiltScale = zoomXScale;
+              let newYScale: BuiltScale = zoomYScale;
+              if (currentZoomAxes === 'x' || currentZoomAxes === 'both') {
+                if (currentZoomConfig?.rescaleX && !isBandScale(zoomXScale)) {
+                  newXScale = currentZoomConfig.rescaleX(zoomXScale as ContinuousScale, transform);
+                } else if (!isBandScale(zoomXScale)) {
+                  newXScale = transform.rescaleX(zoomXScale as ContinuousScale);
                 }
               }
-              if (zoomAxes === 'y' || zoomAxes === 'both') {
-                if (zoomConfig.rescaleY && !isBandScale(yScale)) {
-                  newYScale = zoomConfig.rescaleY(yScale as ContinuousScale, transform);
-                } else if (!isBandScale(yScale)) {
-                  newYScale = transform.rescaleY(yScale as any);
+              if (currentZoomAxes === 'y' || currentZoomAxes === 'both') {
+                if (currentZoomConfig?.rescaleY && !isBandScale(zoomYScale)) {
+                  newYScale = currentZoomConfig.rescaleY(zoomYScale as ContinuousScale, transform);
+                } else if (!isBandScale(zoomYScale)) {
+                  newYScale = transform.rescaleY(zoomYScale as ContinuousScale);
                 }
               }
 
-              // Update axes + grid
-              const xTickValues = resolveTickValues(xAxisConfig?.tickValues, newXScale as AnyScale);
-              const yTickValues = resolveTickValues(yAxisConfig?.tickValues, newYScale as AnyScale);
-              renderAxes(layout, newXScale as AnyScale, newYScale as any, {
-                xTickFormat: xAxisConfig?.tickFormat,
-                yTickFormat: yAxisConfig?.tickFormat,
-                xTickCount: xAxisConfig?.tickCount,
-                yTickCount: yAxisConfig?.tickCount,
-                xTickValues,
-                yTickValues,
-              });
-              if (xAxisConfig?.customize) {
-                xAxisConfig.customize(layout.xAxisGroup);
-              }
-              if (yAxisConfig?.customize) {
-                yAxisConfig.customize(layout.yAxisGroup);
-              }
-              renderGrid(
-                layout,
-                newXScale as AnyScale,
-                newYScale as any,
-                yAxisConfig?.tickCount ?? 5,
-                0,
-                xTickValues,
-                yTickValues,
-              );
-
-              // Update layers
-              const zoomCtx: ZoomContext = {
-                ...ctx,
+              const zoomContext: ZoomContext = {
+                ...zoomBaseContext,
                 newXScale,
                 newYScale,
                 transform,
               };
 
-              for (const layer of layers) {
-                updateLayerOnZoom(
+              for (const layer of zoomLayers) {
+                updateLayerPositionOnZoom(
                   layer,
-                  renderGroup,
-                  xScale,
-                  yScale,
+                  zoomRenderGroup,
+                  zoomXScale,
                   newXScale,
                   newYScale,
-                  layout,
-                  zoomCtx,
+                  zoomLayout,
                 );
               }
 
-              // Keep line labels above the points after the per-layer zoom
-              // updates re-touch the DOM (mirrors the full-render raise above).
-              renderGroup.selectAll('.line-label').raise();
-
-              // User callback
-              zoomConfig.onZoom?.(event, zoomCtx);
+              zoomFrameBatcherRef.current?.schedule(() => {
+                const updatesX = currentZoomAxes === 'x' || currentZoomAxes === 'both';
+                const updatesY = currentZoomAxes === 'y' || currentZoomAxes === 'both';
+                const xTickValues = updatesX
+                  ? resolveTickValues(zoomXAxisConfig?.tickValues, newXScale as AnyScale)
+                  : undefined;
+                const yTickValues = updatesY
+                  ? resolveTickValues(zoomYAxisConfig?.tickValues, newYScale as AnyScale)
+                  : undefined;
+                const zoomYAxisScale = newYScale as unknown as
+                  | ContinuousScale
+                  | d3.ScaleBand<string>;
+                renderAxes(zoomLayout, newXScale as AnyScale, zoomYAxisScale, {
+                  xTickFormat: zoomXAxisConfig?.tickFormat,
+                  yTickFormat: zoomYAxisConfig?.tickFormat,
+                  xTickCount: zoomXAxisConfig?.tickCount,
+                  yTickCount: zoomYAxisConfig?.tickCount,
+                  xTickValues,
+                  yTickValues,
+                  axes: currentZoomAxes,
+                });
+                if (updatesX) zoomXAxisConfig?.customize?.(zoomLayout.xAxisGroup);
+                if (updatesY) zoomYAxisConfig?.customize?.(zoomLayout.yAxisGroup);
+                renderGrid(
+                  zoomLayout,
+                  newXScale as AnyScale,
+                  zoomYAxisScale,
+                  zoomYAxisConfig?.tickCount ?? 5,
+                  0,
+                  xTickValues,
+                  yTickValues,
+                  currentZoomAxes,
+                );
+                for (const layer of zoomLayers) {
+                  updateLayerDecorationOnZoom(
+                    layer,
+                    zoomRenderGroup,
+                    zoomXScale,
+                    newXScale,
+                    newYScale,
+                    zoomLayout,
+                    zoomContext,
+                  );
+                }
+                zoomRenderGroup.selectAll('.line-label').raise();
+                currentZoomConfig?.onZoom?.(event, zoomContext);
+              });
             },
           },
         );
@@ -506,47 +739,245 @@ export function useD3ChartRenderer<T>(props: D3ChartProps<T>, deps: RendererDeps
           renderGroup.select('.ruler-group').style('display', 'none');
         }
       }
+      // A restored non-identity transform emits synchronously above, but its
+      // expensive path work is normally rAF-batched. Flush that replay before
+      // capturing transition targets or a running transition can overwrite
+      // zoomed paths with their base-scale `d` values.
+      zoomFrameBatcherRef.current?.flush();
 
-      // ── Animate from old positions to new positions ──
-      if (transitionDuration && (oldTransforms.size > 0 || oldPaths.size > 0)) {
-        // Scatter points: restore old position, then transition to current
-        renderGroup.selectAll<SVGGElement, unknown>('.dot-group').each(function () {
-          const oldPos = oldTransforms.get(this);
-          const newPos = this.getAttribute('transform');
-          if (oldPos !== undefined && newPos && oldPos !== newPos) {
-            this.setAttribute('transform', oldPos);
-            d3.select(this)
-              .transition('data-update')
-              .duration(transitionDuration)
-              .attr('transform', newPos);
-          }
-        });
-        // Roofline paths: restore old path, then transition to current
-        renderGroup.selectAll<SVGPathElement, unknown>('.roofline-path').each(function () {
-          const oldD = oldPaths.get(this);
-          const newD = this.getAttribute('d');
-          if (oldD !== undefined && newD && oldD !== newD) {
-            this.setAttribute('d', oldD);
-            d3.select(this).transition('data-update').duration(transitionDuration).attr('d', newD);
-          }
-        });
-      }
+      animateTransitionGeometry(renderGroup, previousGeometry, transitionDuration);
 
-      // ── User onRender callback ──
+      renderContextRef.current = ctx;
+      joinedIdentityRef.current = dataIdentity ?? null;
+      lastMetricIdentityRef.current = metricIdentity;
       onRender?.(ctx);
     }
+    return () => {
+      zoomFrameBatcherRef.current?.cancel();
+    };
     // We intentionally list specific deps rather than the entire props object.
   }, [
-    data,
-    dimensions,
+    dataJoinIdentity,
+    dimensions.width,
+    dimensions.height,
     chartId,
+    dataPhaseXScale,
+    dataPhaseYScale,
+    dataPhaseLayers,
+    zoomConfig?.enabled,
+    dataPhaseTooltip,
+    transitionDuration,
+    setupZoom,
+    watermark,
+    clipContent,
+    hasRenderableData,
+    hideTooltipElements,
+  ]);
+
+  // Phase 3: coordinate and scale updates. Bound scatter data is mutated by
+  // stable key, so metric-only changes skip enter/update/exit joins.
+  useLayoutEffect(() => {
+    if (
+      !dataIdentity ||
+      metricIdentity === undefined ||
+      joinedIdentityRef.current !== dataIdentity ||
+      lastMetricIdentityRef.current === metricIdentity ||
+      !svgRef.current ||
+      !tooltipRef.current ||
+      !layoutRef.current
+    ) {
+      return;
+    }
+
+    zoomFrameBatcherRef.current?.cancel();
+
+    const layout = layoutRef.current;
+    const { width, height } = layout;
+    const renderGroup = clipContent ? layout.zoomGroup : layout.g;
+    if (isPinned()) {
+      dismissTooltip(true);
+      hideTooltipElements(tooltipRef, svgRef);
+    }
+    const xScale = hasScales
+      ? buildScale(xScaleConfig!, [0, width])
+      : buildScale({ type: 'linear', domain: [0, 1] }, [0, width]);
+    const yScale = hasScales
+      ? buildScale(yScaleConfig!, [height, 0])
+      : buildScale({ type: 'linear', domain: [0, 1] }, [height, 0]);
+    scalesRef.current = { xScale, yScale };
+
+    const transform = d3.zoomTransform(svgRef.current);
+    const zoomAxes = zoomConfig?.axes ?? 'both';
+    let currentXScale = xScale;
+    let currentYScale = yScale;
+    if ((zoomAxes === 'x' || zoomAxes === 'both') && !isBandScale(xScale)) {
+      currentXScale = zoomConfig?.rescaleX
+        ? zoomConfig.rescaleX(xScale as ContinuousScale, transform)
+        : transform.rescaleX(xScale as ContinuousScale);
+    }
+    if ((zoomAxes === 'y' || zoomAxes === 'both') && !isBandScale(yScale)) {
+      currentYScale = zoomConfig?.rescaleY
+        ? zoomConfig.rescaleY(yScale as ContinuousScale, transform)
+        : transform.rescaleY(yScale as ContinuousScale);
+    }
+
+    if (hasScales) {
+      const yIsStatic =
+        zoomAxes === 'x' &&
+        prevScalesRef.current.yScaleConfig === yScaleConfig &&
+        prevYAxisConfigRef.current === yAxisConfig;
+      const updateAxes = yIsStatic ? 'x' : 'both';
+      const xTickValues = resolveTickValues(xAxisConfig?.tickValues, currentXScale as AnyScale);
+      const yTickValues = yIsStatic
+        ? undefined
+        : resolveTickValues(yAxisConfig?.tickValues, currentYScale as AnyScale);
+      const yAxisScale = currentYScale as unknown as ContinuousScale | d3.ScaleBand<string>;
+      renderGrid(
+        layout,
+        currentXScale as AnyScale,
+        yAxisScale,
+        yAxisConfig?.tickCount ?? 5,
+        0,
+        xTickValues,
+        yTickValues,
+        updateAxes,
+      );
+      renderAxes(layout, currentXScale as AnyScale, yAxisScale, {
+        xTickFormat: xAxisConfig?.tickFormat,
+        yTickFormat: yAxisConfig?.tickFormat,
+        xTickCount: xAxisConfig?.tickCount,
+        yTickCount: yAxisConfig?.tickCount,
+        xTickValues,
+        yTickValues,
+        axes: updateAxes,
+      });
+      xAxisConfig?.customize?.(layout.xAxisGroup);
+      if (!yIsStatic) yAxisConfig?.customize?.(layout.yAxisGroup);
+    }
+
+    const baseCtx: RenderContext = {
+      layout,
+      tooltipElement: tooltipRef.current,
+      xScale,
+      yScale,
+      renderedXScale: xScale,
+      renderedYScale: yScale,
+      width,
+      height,
+      transitionDuration,
+    };
+    const metricCtx: RenderContext = {
+      ...baseCtx,
+      xScale: currentXScale,
+      yScale: currentYScale,
+      renderedXScale: currentXScale,
+      renderedYScale: currentYScale,
+    };
+    const callbackCtx = metricRenderCallbackContext(baseCtx, currentXScale, currentYScale);
+    const previousGeometry =
+      (transitionDuration ?? 0) > 0 ? captureTransitionGeometry(renderGroup) : null;
+    const metricLayerSelections = layers.map((layer) =>
+      updateLayerForMetric(layer, renderGroup, currentXScale, currentYScale, layout, metricCtx),
+    );
+    animateTransitionGeometry(
+      renderGroup,
+      previousGeometry,
+      transitionDuration,
+      metricGeometryVisibility(layers),
+    );
+    customLayerDisplayIdentitiesRef.current = customLayerDisplayIdentities(layers);
+    lastDisplayIdentityRef.current = displayIdentity;
+    renderGroup.selectAll('.dot-group').raise();
+    renderGroup.selectAll('.point').raise();
+    renderGroup.selectAll('.line-label').raise();
+
+    if (tooltipConfig && !tooltipConfig.proximityHover) {
+      const attachIdx =
+        tooltipConfig.attachToLayer ??
+        metricLayerSelections.findIndex((selection) => selection !== null);
+      const targetSelection = attachIdx >= 0 ? metricLayerSelections[attachIdx] : null;
+      if (targetSelection) {
+        const rulers = createRulers(
+          renderGroup,
+          tooltipConfig.rulerType,
+          width,
+          height,
+          'var(--foreground)',
+        );
+        attachHandlers(
+          targetSelection,
+          {
+            rulerType: tooltipConfig.rulerType,
+            generateTooltipContent: tooltipConfig.content,
+            getRulerX: tooltipConfig.getRulerX,
+            getRulerY: tooltipConfig.getRulerY,
+            onHoverStart: tooltipConfig.onHoverStart,
+            onHoverEnd: tooltipConfig.onHoverEnd,
+            onPointClick: tooltipConfig.onPointClick,
+          },
+          svgRef.current.parentElement as HTMLDivElement,
+          d3.select(tooltipRef.current),
+          rulers,
+          xScale,
+          yScale,
+          svgRef,
+          zoomConfig?.axes,
+        );
+      }
+    }
+
+    renderContextRef.current = baseCtx;
+    lastMetricIdentityRef.current = metricIdentity;
+    prevScalesRef.current = { xScaleConfig, yScaleConfig };
+    prevYAxisConfigRef.current = yAxisConfig;
+    onRender?.(callbackCtx);
+  }, [
+    dataIdentity,
+    metricIdentity,
+    dimensions.width,
+    dimensions.height,
     xScaleConfig,
     yScaleConfig,
+    xAxisConfig,
+    yAxisConfig,
     layers,
-    zoomConfig?.enabled,
+    hideTooltipElements,
+    zoomConfig,
     tooltipConfig,
     transitionDuration,
-    setupZoom,
-    watermark,
+    onRender,
+    hasScales,
+    clipContent,
+    scalesRef,
+    svgRef,
+    tooltipRef,
+    createRulers,
+    attachHandlers,
+    isPinned,
+    dismissTooltip,
   ]);
+
+  // Phase 4: display-only invalidation. Charts can restyle existing marks
+  // without rebuilding structure, joins, scales, or paths.
+  useLayoutEffect(() => {
+    const ctx = renderContextRef.current;
+    if (!ctx) return;
+    const renderGroup = clipContent ? ctx.layout.zoomGroup : ctx.layout.g;
+    const previousCustomIdentities = customLayerDisplayIdentitiesRef.current;
+    const nextCustomIdentities = customLayerDisplayIdentities(layersRef.current);
+    const displayChanged = lastDisplayIdentityRef.current !== displayIdentity;
+    layersRef.current.forEach((layer, index) => {
+      if (layer.type === 'custom') {
+        if (layer.displayIdentity === undefined) return;
+        const key = layer.key ?? `custom:${index}`;
+        if (previousCustomIdentities.get(key) === layer.displayIdentity) return;
+      } else if (!displayChanged) {
+        return;
+      }
+      updateLayerForDisplay(layer, renderGroup, ctx);
+    });
+    customLayerDisplayIdentitiesRef.current = nextCustomIdentities;
+    lastDisplayIdentityRef.current = displayIdentity;
+    if (displayChanged) displayCallbackRef.current?.(ctx);
+  }, [displayIdentity, customDisplayPlanIdentity]);
 }
diff --git a/packages/app/src/lib/d3-chart/axis-labels.ts b/packages/app/src/lib/d3-chart/axis-labels.ts
index 0383944a9..94d10f30c 100644
--- a/packages/app/src/lib/d3-chart/axis-labels.ts
+++ b/packages/app/src/lib/d3-chart/axis-labels.ts
@@ -1,5 +1,7 @@
 import * as d3 from 'd3';
 
+import { CHART_TYPE, px } from './typography';
+
 export type LabelSplitMode = 'last-space' | 'newline' | 'parens';
 
 export interface TwoRowLabelOptions {
@@ -7,9 +9,9 @@ export interface TwoRowLabelOptions {
   split?: LabelSplitMode;
   /** Vertical offset in px. Default: 0 */
   yOffset?: number;
-  /** CSS font-size for the first row. Default: '12px' */
+  /** CSS font-size for the first row. Default: CHART_TYPE.axisLabel. */
   primarySize?: string;
-  /** CSS font-size for the second row. Default: '10px' */
+  /** CSS font-size for the second row. Default: CHART_TYPE.axisLabelSub. */
   secondarySize?: string;
 }
 
@@ -46,7 +48,12 @@ export function splitLabel(label: string, mode: LabelSplitMode): [string, string
 export function twoRowYAxisLabels(options: TwoRowLabelOptions | number = {}) {
   // Backwards compat: accept bare number as yOffset
   const opts = typeof options === 'number' ? { yOffset: options } : options;
-  const { split = 'last-space', yOffset = 0, primarySize = '12px', secondarySize = '10px' } = opts;
+  const {
+    split = 'last-space',
+    yOffset = 0,
+    primarySize = px(CHART_TYPE.axisLabel),
+    secondarySize = px(CHART_TYPE.axisLabelSub),
+  } = opts;
 
   return (axisGroup: d3.Selection<SVGGElement, unknown, null, undefined>) => {
     axisGroup.selectAll('.tick text').each(function () {
diff --git a/packages/app/src/lib/d3-chart/chart-setup.ts b/packages/app/src/lib/d3-chart/chart-setup.ts
index e4388f524..388772bb3 100644
--- a/packages/app/src/lib/d3-chart/chart-setup.ts
+++ b/packages/app/src/lib/d3-chart/chart-setup.ts
@@ -1,6 +1,7 @@
 import * as d3 from 'd3';
 
 import type { ChartLayout, ChartSetupConfig } from './types';
+import { CHART_TYPE, px } from './typography';
 import {
   createLogoWatermark,
   createUnofficialWatermark,
@@ -89,6 +90,9 @@ export function setupChartStructure(
 
     let zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>;
     if (clipContent) {
+      // `plot-bounds.ts` reconstructs this rect in viewport coordinates (from
+      // the .chart-root translate plus these dimensions) so callers can tell
+      // whether a point is actually visible. Keep the two in step.
       defs
         .append('clipPath')
         .attr('id', `clip-${config.chartId}`)
@@ -112,7 +116,7 @@ export function setupChartStructure(
         .attr('x', -(margin.top + height / 2))
         .attr('y', 14)
         .attr('text-anchor', 'middle')
-        .attr('font-size', '12px')
+        .attr('font-size', px(CHART_TYPE.axisLabel))
         .text(config.yLabel);
     }
 
@@ -123,7 +127,7 @@ export function setupChartStructure(
         .attr('x', margin.left + width / 2)
         .attr('y', containerHeight - 10)
         .attr('text-anchor', 'middle')
-        .attr('font-size', '12px')
+        .attr('font-size', px(CHART_TYPE.axisLabel))
         .text(config.xLabel);
     }
 
diff --git a/packages/app/src/lib/d3-chart/chart-update.test.ts b/packages/app/src/lib/d3-chart/chart-update.test.ts
index a8b3d0127..4fe1c9a15 100644
--- a/packages/app/src/lib/d3-chart/chart-update.test.ts
+++ b/packages/app/src/lib/d3-chart/chart-update.test.ts
@@ -300,6 +300,29 @@ describe('renderAxes', () => {
       expect(ticks1).toBeGreaterThan(0);
       expect(ticks2).toBeGreaterThan(0);
     });
+
+    it('leaves the Y-axis subtree untouched during an X-only update', () => {
+      const layout = makeLayout();
+      const xScale = d3.scaleLinear().domain([0, 10]).range([0, layout.width]);
+      const yScale = d3
+        .scaleBand()
+        .domain(['Alpha', 'Beta'])
+        .range([layout.height, 0])
+        .padding(0.1);
+
+      renderAxes(layout, xScale, yScale, {});
+      layout.yAxisGroup.select('.tick text').append('tspan').attr('data-static-label', 'true');
+      const yRoot = layout.yAxisGroup.node();
+      const yFirstChild = yRoot?.firstChild;
+      const yMarkup = yRoot?.innerHTML;
+
+      xScale.domain([2, 6]);
+      renderAxes(layout, xScale, yScale.domain(['Changed']), { axes: 'x' });
+
+      expect(layout.yAxisGroup.node()).toBe(yRoot);
+      expect(layout.yAxisGroup.node()?.firstChild).toBe(yFirstChild);
+      expect(layout.yAxisGroup.node()?.innerHTML).toBe(yMarkup);
+    });
   });
 });
 
@@ -520,6 +543,25 @@ describe('renderGrid', () => {
       expect(layout.gridGroup.selectAll('.grid-h').size()).toBe(1);
     });
 
+    it('leaves the horizontal grid subtree untouched during an X-only update', () => {
+      const layout = makeLayout();
+      const xScale = d3.scaleLinear().domain([0, 100]).range([0, layout.width]);
+      const yScale = d3.scaleLinear().domain([0, 50]).range([layout.height, 0]);
+
+      renderGrid(layout, xScale, yScale);
+      const horizontalGrid = layout.gridGroup.select<SVGGElement>('.grid-h').node();
+      const firstLine = horizontalGrid?.firstChild;
+      const markup = horizontalGrid?.innerHTML;
+
+      xScale.domain([25, 50]);
+      yScale.domain([500, 1000]);
+      renderGrid(layout, xScale, yScale, 5, 0, undefined, undefined, 'x');
+
+      expect(layout.gridGroup.select('.grid-h').node()).toBe(horizontalGrid);
+      expect(horizontalGrid?.firstChild).toBe(firstLine);
+      expect(horizontalGrid?.innerHTML).toBe(markup);
+    });
+
     it('updates line positions when scale domain changes', () => {
       const layout = makeLayout();
       const xScale = d3.scaleLinear().domain([0, 100]).range([0, layout.width]);
diff --git a/packages/app/src/lib/d3-chart/chart-update.ts b/packages/app/src/lib/d3-chart/chart-update.ts
index fb79ac270..77395b504 100644
--- a/packages/app/src/lib/d3-chart/chart-update.ts
+++ b/packages/app/src/lib/d3-chart/chart-update.ts
@@ -16,6 +16,8 @@ export interface AxisUpdateConfig {
   yTickSize?: number;
   /** When set, axes animate to new positions over this duration (ms). */
   transitionDuration?: number;
+  /** Limit the update to axes that actually changed. Defaults to both. */
+  axes?: 'x' | 'y' | 'both';
 }
 
 /** Render or update axes with current scales. */
@@ -34,45 +36,51 @@ export function renderAxes(
     yTickValues,
     yTickSize,
     transitionDuration,
+    axes = 'both',
   } = config;
   const dur = transitionDuration ?? 0;
+  const updateX = axes === 'x' || axes === 'both';
+  const updateY = axes === 'y' || axes === 'both';
 
-  // X axis
-  let xAxisGen: d3.Axis<d3.AxisDomain>;
-  if ('bandwidth' in xScale) {
-    xAxisGen = d3
-      .axisBottom(xScale as d3.ScaleBand<string>)
-      .tickSize(6) as unknown as d3.Axis<d3.AxisDomain>;
-  } else {
-    const gen = d3.axisBottom(xScale as ContinuousScale).tickSize(6);
-    if (xTickCount) gen.ticks(xTickCount);
-    if (xTickValues) {
-      gen.tickValues(visibleTickValues(xScale, xTickValues) as Iterable<d3.NumberValue>);
+  if (updateX) {
+    let xAxisGen: d3.Axis<d3.AxisDomain>;
+    if ('bandwidth' in xScale) {
+      xAxisGen = d3
+        .axisBottom(xScale as d3.ScaleBand<string>)
+        .tickSize(6) as unknown as d3.Axis<d3.AxisDomain>;
+    } else {
+      const gen = d3.axisBottom(xScale as ContinuousScale).tickSize(6);
+      if (xTickCount) gen.ticks(xTickCount);
+      if (xTickValues) {
+        gen.tickValues(visibleTickValues(xScale, xTickValues) as Iterable<d3.NumberValue>);
+      }
+      if (xTickFormat) gen.tickFormat(xTickFormat as any);
+      xAxisGen = gen as unknown as d3.Axis<d3.AxisDomain>;
     }
-    if (xTickFormat) gen.tickFormat(xTickFormat as any);
-    xAxisGen = gen as unknown as d3.Axis<d3.AxisDomain>;
-  }
 
-  const xTarget = dur > 0 ? layout.xAxisGroup.transition().duration(dur) : layout.xAxisGroup;
-  (xTarget as any).call(xAxisGen as any);
+    const xTarget = dur > 0 ? layout.xAxisGroup.transition().duration(dur) : layout.xAxisGroup;
+    (xTarget as any).call(xAxisGen as any);
+  }
 
-  // Y axis
-  if ('bandwidth' in yScale) {
-    const bandAxisGen = d3
-      .axisLeft(yScale as d3.ScaleBand<string>)
-      .tickSize(yTickSize ?? 0) as unknown as d3.Axis<d3.AxisDomain>;
-    if (yTickFormat) bandAxisGen.tickFormat(yTickFormat as any);
-    const yBandTarget = dur > 0 ? layout.yAxisGroup.transition().duration(dur) : layout.yAxisGroup;
-    (yBandTarget as any).call(bandAxisGen as any);
-  } else {
-    const yAxisGen = d3.axisLeft(yScale as ContinuousScale).tickSize(yTickSize ?? 6);
-    if (yTickCount) yAxisGen.ticks(yTickCount);
-    if (yTickValues) {
-      yAxisGen.tickValues(visibleTickValues(yScale, yTickValues) as Iterable<d3.NumberValue>);
+  if (updateY) {
+    if ('bandwidth' in yScale) {
+      const bandAxisGen = d3
+        .axisLeft(yScale as d3.ScaleBand<string>)
+        .tickSize(yTickSize ?? 0) as unknown as d3.Axis<d3.AxisDomain>;
+      if (yTickFormat) bandAxisGen.tickFormat(yTickFormat as any);
+      const yBandTarget =
+        dur > 0 ? layout.yAxisGroup.transition().duration(dur) : layout.yAxisGroup;
+      (yBandTarget as any).call(bandAxisGen as any);
+    } else {
+      const yAxisGen = d3.axisLeft(yScale as ContinuousScale).tickSize(yTickSize ?? 6);
+      if (yTickCount) yAxisGen.ticks(yTickCount);
+      if (yTickValues) {
+        yAxisGen.tickValues(visibleTickValues(yScale, yTickValues) as Iterable<d3.NumberValue>);
+      }
+      if (yTickFormat) yAxisGen.tickFormat(yTickFormat as any);
+      const yTarget = dur > 0 ? layout.yAxisGroup.transition().duration(dur) : layout.yAxisGroup;
+      (yTarget as any).call(yAxisGen as any);
     }
-    if (yTickFormat) yAxisGen.tickFormat(yTickFormat as any);
-    const yTarget = dur > 0 ? layout.yAxisGroup.transition().duration(dur) : layout.yAxisGroup;
-    (yTarget as any).call(yAxisGen as any);
   }
 }
 
@@ -85,90 +93,95 @@ export function renderGrid(
   transitionDuration = 0,
   xTickValues?: (number | Date)[],
   yTickValues?: (number | Date)[],
+  axes: 'x' | 'y' | 'both' = 'both',
 ): void {
   const { width, height, gridGroup } = layout;
   const dur = transitionDuration;
+  const updateX = axes === 'x' || axes === 'both';
+  const updateY = axes === 'y' || axes === 'both';
 
-  // Vertical grid lines — reuse existing group, join updates lines in place
-  let vGroup = gridGroup.select<SVGGElement>('.grid-v');
-  if (vGroup.empty()) vGroup = gridGroup.append('g').attr('class', 'grid-v');
+  if (updateX) {
+    let vGroup = gridGroup.select<SVGGElement>('.grid-v');
+    if (vGroup.empty()) vGroup = gridGroup.append('g').attr('class', 'grid-v');
 
-  if ('bandwidth' in xScale) {
-    const bandScale = xScale as d3.ScaleBand<string>;
-    vGroup
-      .selectAll<SVGLineElement, string>('line')
-      .data(bandScale.domain())
-      .join('line')
-      .attr('x1', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
-      .attr('x2', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
-      .attr('y1', 0)
-      .attr('y2', height);
-  } else {
-    const tickScale = xScale as { ticks: (count?: number) => number[]; (v: number): number };
-    const xTicks = xTickValues
-      ? (visibleTickValues(xScale, xTickValues) as number[])
-      : tickScale.ticks();
-    const vJoin = vGroup
-      .selectAll<SVGLineElement, number>('line')
-      .data(xTicks)
-      .join(
-        (enter) =>
-          enter
-            .append('line')
-            .attr('y1', 0)
-            .attr('y2', height)
-            .attr('x1', (d) => tickScale(d))
-            .attr('x2', (d) => tickScale(d)),
-        (update) => update,
-        (exit) => exit.remove(),
-      );
-    const vTarget = dur > 0 ? (vJoin as any).transition().duration(dur) : vJoin;
-    vTarget
-      .attr('x1', (d: number) => tickScale(d))
-      .attr('x2', (d: number) => tickScale(d))
-      .attr('y1', 0)
-      .attr('y2', height);
+    if ('bandwidth' in xScale) {
+      const bandScale = xScale as d3.ScaleBand<string>;
+      vGroup
+        .selectAll<SVGLineElement, string>('line')
+        .data(bandScale.domain())
+        .join('line')
+        .attr('x1', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
+        .attr('x2', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
+        .attr('y1', 0)
+        .attr('y2', height);
+    } else {
+      const tickScale = xScale as { ticks: (count?: number) => number[]; (v: number): number };
+      const xTicks = xTickValues
+        ? (visibleTickValues(xScale, xTickValues) as number[])
+        : tickScale.ticks();
+      const vJoin = vGroup
+        .selectAll<SVGLineElement, number>('line')
+        .data(xTicks)
+        .join(
+          (enter) =>
+            enter
+              .append('line')
+              .attr('y1', 0)
+              .attr('y2', height)
+              .attr('x1', (d) => tickScale(d))
+              .attr('x2', (d) => tickScale(d)),
+          (update) => update,
+          (exit) => exit.remove(),
+        );
+      const vTarget = dur > 0 ? (vJoin as any).transition().duration(dur) : vJoin;
+      vTarget
+        .attr('x1', (d: number) => tickScale(d))
+        .attr('x2', (d: number) => tickScale(d))
+        .attr('y1', 0)
+        .attr('y2', height);
+    }
   }
 
-  // Horizontal grid lines — reuse existing group
-  let hGroup = gridGroup.select<SVGGElement>('.grid-h');
-  if (hGroup.empty()) hGroup = gridGroup.append('g').attr('class', 'grid-h');
+  if (updateY) {
+    let hGroup = gridGroup.select<SVGGElement>('.grid-h');
+    if (hGroup.empty()) hGroup = gridGroup.append('g').attr('class', 'grid-h');
 
-  if ('bandwidth' in yScale) {
-    const bandScale = yScale as d3.ScaleBand<string>;
-    hGroup
-      .selectAll<SVGLineElement, string>('line')
-      .data(bandScale.domain())
-      .join('line')
-      .attr('x1', 0)
-      .attr('x2', width)
-      .attr('y1', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
-      .attr('y2', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
-      .style('stroke-width', 0.5);
-  } else {
-    const yTicks = yTickValues
-      ? (visibleTickValues(yScale, yTickValues) as number[])
-      : yScale.ticks(yTickCount ?? 5);
-    const hJoin = hGroup
-      .selectAll<SVGLineElement, number>('line')
-      .data(yTicks)
-      .join(
-        (enter) =>
-          enter
-            .append('line')
-            .attr('x1', 0)
-            .attr('x2', width)
-            .attr('y1', (d) => yScale(d))
-            .attr('y2', (d) => yScale(d)),
-        (update) => update,
-        (exit) => exit.remove(),
-      );
-    const hTarget = dur > 0 ? (hJoin as any).transition().duration(dur) : hJoin;
-    hTarget
-      .attr('x1', 0)
-      .attr('x2', width)
-      .attr('y1', (d: number) => yScale(d))
-      .attr('y2', (d: number) => yScale(d));
+    if ('bandwidth' in yScale) {
+      const bandScale = yScale as d3.ScaleBand<string>;
+      hGroup
+        .selectAll<SVGLineElement, string>('line')
+        .data(bandScale.domain())
+        .join('line')
+        .attr('x1', 0)
+        .attr('x2', width)
+        .attr('y1', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
+        .attr('y2', (d) => (bandScale(d) || 0) + bandScale.bandwidth() / 2)
+        .style('stroke-width', 0.5);
+    } else {
+      const yTicks = yTickValues
+        ? (visibleTickValues(yScale, yTickValues) as number[])
+        : yScale.ticks(yTickCount ?? 5);
+      const hJoin = hGroup
+        .selectAll<SVGLineElement, number>('line')
+        .data(yTicks)
+        .join(
+          (enter) =>
+            enter
+              .append('line')
+              .attr('x1', 0)
+              .attr('x2', width)
+              .attr('y1', (d) => yScale(d))
+              .attr('y2', (d) => yScale(d)),
+          (update) => update,
+          (exit) => exit.remove(),
+        );
+      const hTarget = dur > 0 ? (hJoin as any).transition().duration(dur) : hJoin;
+      hTarget
+        .attr('x1', 0)
+        .attr('x2', width)
+        .attr('y1', (d: number) => yScale(d))
+        .attr('y2', (d: number) => yScale(d));
+    }
   }
 }
 
diff --git a/packages/app/src/lib/d3-chart/index.ts b/packages/app/src/lib/d3-chart/index.ts
deleted file mode 100644
index 02c4e46a6..000000000
--- a/packages/app/src/lib/d3-chart/index.ts
+++ /dev/null
@@ -1,26 +0,0 @@
-export { setupChartStructure } from './chart-setup';
-export { renderAxes, renderGrid } from './chart-update';
-export type { ChartLayout, ChartMargin, ChartSetupConfig, ContinuousScale } from './types';
-export type { AnyScale, AxisUpdateConfig } from './chart-update';
-export * from './layers';
-export { contrastColors } from './contrast-colors';
-export { twoRowYAxisLabels, splitLabel } from './axis-labels';
-export type { TwoRowLabelOptions, LabelSplitMode } from './axis-labels';
-export {
-  measureTextWidth,
-  computeLeftMargin,
-  computeBottomMargin,
-  computeDynamicMargins,
-} from './dynamic-margins';
-export { D3Chart } from './D3Chart';
-export type {
-  D3ChartHandle,
-  D3ChartProps,
-  ScaleConfig,
-  LayerConfig,
-  AxisConfig,
-  ZoomConfig,
-  TooltipConfig,
-  RenderContext,
-  ZoomContext,
-} from './D3Chart';
diff --git a/packages/app/src/lib/d3-chart/layers/horizontal-bars.ts b/packages/app/src/lib/d3-chart/layers/horizontal-bars.ts
index 554f01d07..f240db205 100644
--- a/packages/app/src/lib/d3-chart/layers/horizontal-bars.ts
+++ b/packages/app/src/lib/d3-chart/layers/horizontal-bars.ts
@@ -46,3 +46,15 @@ export function updateHorizontalBarsOnZoom<T>(
 ): void {
   group.selectAll<SVGRectElement, T>('.bar').attr('width', (d) => Math.max(0, xScale(getX(d))));
 }
+
+/** Restyle existing bars without repeating their data join or geometry writes. */
+export function updateHorizontalBarsForDisplay<T>(
+  group: d3.Selection<SVGGElement, unknown, null, undefined>,
+  config: HorizontalBarConfig<T>,
+): void {
+  group
+    .selectAll<SVGRectElement, T>('.bar')
+    .attr('fill', (d) => config.getColor(d))
+    .attr('stroke', config.stroke ?? 'none')
+    .attr('stroke-width', config.strokeWidth ?? 0);
+}
diff --git a/packages/app/src/lib/d3-chart/layers/index.ts b/packages/app/src/lib/d3-chart/layers/index.ts
deleted file mode 100644
index 5af116653..000000000
--- a/packages/app/src/lib/d3-chart/layers/index.ts
+++ /dev/null
@@ -1,22 +0,0 @@
-export { renderBars, updateBarsOnZoom } from './bars';
-export type { BarConfig } from './bars';
-export { renderBarLabels, updateBarLabelsOnZoom } from './bar-labels';
-export type { BarLabelConfig } from './bar-labels';
-export { renderHorizontalBars, updateHorizontalBarsOnZoom } from './horizontal-bars';
-export type { HorizontalBarConfig } from './horizontal-bars';
-export { renderLines, updateLinesOnZoom } from './lines';
-export type { LineConfig } from './lines';
-export { renderPoints, updatePointsOnZoom } from './points';
-export type { PointConfig } from './points';
-export { renderErrorBars, updateErrorBarsOnZoom } from './error-bars';
-export type { ErrorBarConfig } from './error-bars';
-export {
-  renderScatterPoints,
-  attachScatterTooltipHandlers,
-  updateScatterPointsOnZoom,
-} from './scatter-points';
-export type { ScatterPointConfig } from './scatter-points';
-export { renderRooflines, updateRooflinesOnZoom } from './rooflines';
-export type { RooflineConfig } from './rooflines';
-export { renderRadar } from './radar';
-export type { RadarConfig, RadarDot } from './radar';
diff --git a/packages/app/src/lib/d3-chart/layers/perf-ruler.test.ts b/packages/app/src/lib/d3-chart/layers/perf-ruler.test.ts
new file mode 100644
index 000000000..d50fe1573
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/layers/perf-ruler.test.ts
@@ -0,0 +1,942 @@
+import { describe, expect, it } from 'vitest';
+
+import { createMockGroup, type MockSelection } from './test-helpers';
+import {
+  DEFAULT_LABEL_FONT_SIZE,
+  EMPTY_PERF_RULER_STATE,
+  MAX_PERF_RULERS,
+  clampIsoX,
+  clearPerfRulers,
+  computeIsoXRulerGeometry,
+  computePerfRulerLabelLayout,
+  computePerfRulerLabelLayouts,
+  deletePerfRuler,
+  formatPerfRatio,
+  intersectPathAtX,
+  isPerfRulerCurveVisible,
+  movePerfRulerIsoX,
+  nextPerfRulerState,
+  pathXExtent,
+  perfRulerCurveSet,
+  prunePerfRulers,
+  renderPerfRulers,
+  type PerfRulerEndInput,
+  type PerfRulerGeometry,
+  type PerfRulerLabelLayoutOptions,
+  type PerfRulerPathLike,
+  type PerfRulerRenderOptions,
+  type PerfRulerState,
+} from './perf-ruler';
+
+// ── Fixtures ─────────────────────────────────────────
+
+const ISO_X = 100;
+const END_A: PerfRulerEndInput = { py: 50, rawY: 400 };
+const END_B: PerfRulerEndInput = { py: 150, rawY: 197 };
+
+function makeOpts(overrides?: Partial<PerfRulerRenderOptions>): PerfRulerRenderOptions {
+  return {
+    color: 'var(--primary)',
+    ...overrides,
+  };
+}
+
+/**
+ * Synthetic polyline implementation of the SVGPathElement length API — jsdom
+ * has no getTotalLength/getPointAtLength, so intersection tests drive the
+ * binary search through this instead of a real path node.
+ */
+function polylinePath(vertices: { x: number; y: number }[]): PerfRulerPathLike {
+  const lengths: number[] = [0];
+  for (let i = 1; i < vertices.length; i++) {
+    lengths.push(
+      lengths[i - 1] +
+        Math.hypot(vertices[i].x - vertices[i - 1].x, vertices[i].y - vertices[i - 1].y),
+    );
+  }
+  const total = vertices.length > 0 ? (lengths.at(-1) ?? 0) : 0;
+  return {
+    getTotalLength: () => total,
+    getPointAtLength(length: number) {
+      const clamped = Math.min(Math.max(length, 0), total);
+      for (let i = 1; i < vertices.length; i++) {
+        if (clamped <= lengths[i]) {
+          const segLen = lengths[i] - lengths[i - 1];
+          const t = segLen === 0 ? 0 : (clamped - lengths[i - 1]) / segLen;
+          return {
+            x: vertices[i - 1].x + (vertices[i].x - vertices[i - 1].x) * t,
+            y: vertices[i - 1].y + (vertices[i].y - vertices[i - 1].y) * t,
+          };
+        }
+      }
+      return vertices.at(-1) ?? { x: Number.NaN, y: Number.NaN };
+    },
+  };
+}
+
+// ── formatPerfRatio ──────────────────────────────────────────────────
+
+describe('formatPerfRatio', () => {
+  it('formats small multiples with two decimals', () => {
+    expect(formatPerfRatio(2.0304)).toBe('2.03x');
+    expect(formatPerfRatio(1)).toBe('1.00x');
+  });
+
+  it('drops to one decimal at 10x and none at 100x', () => {
+    expect(formatPerfRatio(10.46)).toBe('10.5x');
+    expect(formatPerfRatio(123.4)).toBe('123x');
+  });
+
+  it('returns empty string for non-finite or non-positive ratios', () => {
+    expect(formatPerfRatio(Number.NaN)).toBe('');
+    expect(formatPerfRatio(Number.POSITIVE_INFINITY)).toBe('');
+    expect(formatPerfRatio(0)).toBe('');
+    expect(formatPerfRatio(-2)).toBe('');
+  });
+});
+
+// ── isPerfRulerCurveVisible ────────────────────────────────────────────────
+
+describe('isPerfRulerCurveVisible', () => {
+  it('treats legend-hidden curves (opacity 0) as invisible', () => {
+    expect(isPerfRulerCurveVisible('0')).toBe(false);
+    expect(isPerfRulerCurveVisible('0.0')).toBe(false);
+    expect(isPerfRulerCurveVisible(' 0 ')).toBe(false);
+  });
+
+  it('keeps hover-dimmed and fully opaque curves measurable', () => {
+    expect(isPerfRulerCurveVisible('0.15')).toBe(true);
+    expect(isPerfRulerCurveVisible('1')).toBe(true);
+  });
+
+  it('treats missing, empty, or unparseable opacity as visible', () => {
+    expect(isPerfRulerCurveVisible(null)).toBe(true);
+    expect(isPerfRulerCurveVisible(undefined)).toBe(true);
+    expect(isPerfRulerCurveVisible('')).toBe(true);
+    expect(isPerfRulerCurveVisible('inherit')).toBe(true);
+  });
+});
+
+// ── intersectPathAtX ─────────────────────────────────────────────────
+
+describe('intersectPathAtX', () => {
+  it('finds the mid-segment intersection on a single-segment path', () => {
+    const path = polylinePath([
+      { x: 0, y: 100 },
+      { x: 100, y: 0 },
+    ]);
+    const hit = intersectPathAtX(path, 50);
+    expect(hit).not.toBeNull();
+    expect(hit!.x).toBeCloseTo(50, 0);
+    expect(hit!.y).toBeCloseTo(50, 0);
+  });
+
+  it('interpolates within the correct segment of a multi-segment curve', () => {
+    const path = polylinePath([
+      { x: 0, y: 200 },
+      { x: 100, y: 100 },
+      { x: 300, y: 50 },
+    ]);
+    const hit = intersectPathAtX(path, 200);
+    expect(hit).not.toBeNull();
+    expect(hit!.x).toBeCloseTo(200, 0);
+    expect(hit!.y).toBeCloseTo(75, 1);
+  });
+
+  it('hits interior vertices exactly', () => {
+    const path = polylinePath([
+      { x: 0, y: 200 },
+      { x: 100, y: 100 },
+      { x: 300, y: 50 },
+    ]);
+    const hit = intersectPathAtX(path, 100);
+    expect(hit).not.toBeNull();
+    expect(hit!.x).toBeCloseTo(100, 0);
+    expect(hit!.y).toBeCloseTo(100, 0);
+  });
+
+  it('supports paths whose x decreases along their length', () => {
+    const path = polylinePath([
+      { x: 300, y: 10 },
+      { x: 100, y: 110 },
+    ]);
+    const hit = intersectPathAtX(path, 200);
+    expect(hit).not.toBeNull();
+    expect(hit!.x).toBeCloseTo(200, 0);
+    expect(hit!.y).toBeCloseTo(60, 0);
+  });
+
+  it('returns null when x lies outside the path x extent', () => {
+    const path = polylinePath([
+      { x: 100, y: 100 },
+      { x: 300, y: 50 },
+    ]);
+    expect(intersectPathAtX(path, 50)).toBeNull();
+    expect(intersectPathAtX(path, 350)).toBeNull();
+  });
+
+  it('still intersects at the endpoints (within half-pixel slack)', () => {
+    const path = polylinePath([
+      { x: 100, y: 100 },
+      { x: 300, y: 50 },
+    ]);
+    const atStart = intersectPathAtX(path, 100);
+    expect(atStart).not.toBeNull();
+    expect(atStart!.y).toBeCloseTo(100, 0);
+    const nearEnd = intersectPathAtX(path, 300.4);
+    expect(nearEnd).not.toBeNull();
+    expect(nearEnd!.y).toBeCloseTo(50, 0);
+  });
+
+  it('converges tightly on shallow curves approximated by many segments', () => {
+    // y = 10000 / x sampled on [50, 500] — a hyperbola like a latency curve.
+    const vertices = Array.from({ length: 91 }, (_v, i) => {
+      const x = 50 + i * 5;
+      return { x, y: 10000 / x };
+    });
+    const path = polylinePath(vertices);
+    const hit = intersectPathAtX(path, 250);
+    expect(hit).not.toBeNull();
+    expect(hit!.x).toBeCloseTo(250, 0);
+    expect(hit!.y).toBeCloseTo(40, 0);
+  });
+
+  it('returns a point on a vertical (constant-x) path instead of diverging', () => {
+    const path = polylinePath([
+      { x: 50, y: 0 },
+      { x: 50, y: 100 },
+    ]);
+    const hit = intersectPathAtX(path, 50);
+    expect(hit).not.toBeNull();
+    expect(hit!.x).toBe(50);
+  });
+
+  it('returns null for degenerate paths and non-finite x', () => {
+    expect(intersectPathAtX(polylinePath([{ x: 10, y: 10 }]), 10)).toBeNull();
+    expect(intersectPathAtX(polylinePath([]), 10)).toBeNull();
+    const path = polylinePath([
+      { x: 0, y: 0 },
+      { x: 100, y: 100 },
+    ]);
+    expect(intersectPathAtX(path, Number.NaN)).toBeNull();
+  });
+});
+
+// ── computeIsoXRulerGeometry ─────────────────────────────────────────
+
+describe('computeIsoXRulerGeometry', () => {
+  it('places the line at the iso-x, spanning the two intersection ys', () => {
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B);
+    expect(geometry).not.toBeNull();
+    expect(geometry!.x).toBe(100);
+    expect(geometry!.y1).toBe(50);
+    expect(geometry!.y2).toBe(150);
+  });
+
+  it('computes the ratio from raw y values, higher over lower', () => {
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B);
+    expect(geometry!.ratio).toBeCloseTo(400 / 197, 10);
+    expect(geometry!.ratioLabel).toBe('2.03x');
+  });
+
+  it('yields the same ratio regardless of end order (symmetric)', () => {
+    const swapped = computeIsoXRulerGeometry(ISO_X, END_B, END_A);
+    expect(swapped!.ratio).toBeCloseTo(400 / 197, 10);
+    expect(swapped!.y1).toBe(50);
+    expect(swapped!.y2).toBe(150);
+  });
+
+  it('handles the curves crossing at the iso-x (ratio 1, zero-height span)', () => {
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, { ...END_A });
+    expect(geometry).not.toBeNull();
+    expect(geometry!.ratio).toBe(1);
+    expect(geometry!.ratioLabel).toBe('1.00x');
+    expect(geometry!.y1).toBe(geometry!.y2);
+  });
+
+  it('returns null when either raw y is zero or negative', () => {
+    expect(computeIsoXRulerGeometry(ISO_X, END_A, { ...END_B, rawY: 0 })).toBeNull();
+    expect(computeIsoXRulerGeometry(ISO_X, { ...END_A, rawY: -5 }, END_B)).toBeNull();
+  });
+
+  it('returns null for non-finite inputs', () => {
+    expect(computeIsoXRulerGeometry(Number.NaN, END_A, END_B)).toBeNull();
+    expect(
+      computeIsoXRulerGeometry(ISO_X, END_A, { ...END_B, py: Number.POSITIVE_INFINITY }),
+    ).toBeNull();
+    expect(computeIsoXRulerGeometry(ISO_X, END_A, { ...END_B, rawY: Number.NaN })).toBeNull();
+  });
+});
+
+// ── computePerfRulerLabelLayout ──────────────────────────────────────────────────
+
+describe('computePerfRulerLabelLayout', () => {
+  const GEOMETRY = { x: 100, y1: 50, y2: 150, ratioLabel: '2.03x' };
+
+  it('places the label up-right of the line midpoint by default', () => {
+    const layout = computePerfRulerLabelLayout(GEOMETRY, { chartWidth: 800, chartHeight: 400 });
+    expect(layout.side).toBe(1);
+    expect(layout.textAnchor).toBe('start');
+    expect(layout.labelX).toBeGreaterThan(GEOMETRY.x);
+    expect(layout.labelY).toBeLessThan((GEOMETRY.y1 + GEOMETRY.y2) / 2);
+  });
+
+  it('flips to the left side near the right chart edge', () => {
+    const layout = computePerfRulerLabelLayout(GEOMETRY, { chartWidth: 125, chartHeight: 400 });
+    expect(layout.side).toBe(-1);
+    expect(layout.textAnchor).toBe('end');
+    expect(layout.labelX).toBeLessThan(GEOMETRY.x);
+  });
+
+  it('drops below the midpoint when the label would clip the top', () => {
+    const top = { x: 100, y1: 10, y2: 30, ratioLabel: '2.03x' };
+    const layout = computePerfRulerLabelLayout(top, { chartWidth: 800, chartHeight: 400 });
+    expect(layout.labelY).toBeGreaterThan((top.y1 + top.y2) / 2);
+  });
+
+  it('clamps the label inside the chart when above and below both clip', () => {
+    const top = { x: 100, y1: 10, y2: 30, ratioLabel: '2.03x' };
+    const layout = computePerfRulerLabelLayout(top, { chartWidth: 800, chartHeight: 70 });
+    const fontSize = DEFAULT_LABEL_FONT_SIZE;
+    expect(layout.labelY - fontSize / 2).toBeGreaterThanOrEqual(4);
+    expect(layout.labelY + fontSize / 2).toBeLessThanOrEqual(70);
+  });
+
+  it('ends the arrow curve horizontally next to the line midpoint', () => {
+    const layout = computePerfRulerLabelLayout(GEOMETRY, { chartWidth: 800, chartHeight: 400 });
+    const midY = (GEOMETRY.y1 + GEOMETRY.y2) / 2;
+    // Quadratic curve: M sx sy Q cx cy ex ey — the end point sits a few px
+    // right of the line at the midpoint height, tangent horizontal.
+    const numbers = layout.arrowPath.match(/-?\d+(?:\.\d+)?/g)!.map(Number);
+    const [sx, sy, cx, cy, ex, ey] = numbers;
+    expect(sx).toBeGreaterThan(GEOMETRY.x);
+    expect(sy).toBeLessThan(midY);
+    expect(cy).toBe(midY);
+    expect(cx).toBe(sx);
+    expect(ey).toBe(midY);
+    expect(ex).toBeGreaterThan(GEOMETRY.x);
+    expect(ex).toBeLessThan(sx);
+  });
+
+  it('points the arrowhead tip at the line, just off the stroke', () => {
+    const layout = computePerfRulerLabelLayout(GEOMETRY, { chartWidth: 800, chartHeight: 400 });
+    const midY = (GEOMETRY.y1 + GEOMETRY.y2) / 2;
+    const numbers = layout.arrowHeadPath.match(/-?\d+(?:\.\d+)?/g)!.map(Number);
+    const [tipX, tipY] = numbers;
+    expect(tipY).toBe(midY);
+    expect(Math.abs(tipX - GEOMETRY.x)).toBeLessThanOrEqual(6);
+    expect(layout.arrowHeadPath.endsWith('Z')).toBe(true);
+  });
+
+  it('respects a custom font size when checking the top clip', () => {
+    const nearTop = { x: 100, y1: 40, y2: 60, ratioLabel: '2.03x' };
+    const small = computePerfRulerLabelLayout(nearTop, { fontSize: 12 });
+    // 50 - 46 = 4 above the top — fits a 12px label (4 - 6 < 4 fails)…
+    // both sizes clip here, so both drop below; a taller chart midpoint
+    // stays above for both.
+    const tall = { x: 100, y1: 100, y2: 200, ratioLabel: '2.03x' };
+    expect(computePerfRulerLabelLayout(tall, { fontSize: 12 }).labelY).toBeLessThan(150);
+    expect(small.labelY).toBeGreaterThan(50);
+  });
+
+  it('anchors the × delete button just past the label, at the label height', () => {
+    const right = computePerfRulerLabelLayout(GEOMETRY, { chartWidth: 800, chartHeight: 400 });
+    expect(right.side).toBe(1);
+    expect(right.deleteX).toBeGreaterThan(right.labelX);
+    expect(right.deleteY).toBe(right.labelY);
+    const left = computePerfRulerLabelLayout(GEOMETRY, { chartWidth: 125, chartHeight: 400 });
+    expect(left.side).toBe(-1);
+    expect(left.deleteX).toBeLessThan(left.labelX);
+    expect(left.deleteY).toBe(left.labelY);
+  });
+});
+
+// ── computePerfRulerLabelLayouts ──────────────────────────────────
+
+describe('computePerfRulerLabelLayouts', () => {
+  const GEOMETRY = { x: 100, y1: 50, y2: 150, ratioLabel: '2.03x' };
+  const OPTS = { chartWidth: 800, chartHeight: 400 };
+
+  it('matches the single-label layout when there is no collision', () => {
+    const [only] = computePerfRulerLabelLayouts([GEOMETRY], OPTS);
+    expect(only).toEqual(computePerfRulerLabelLayout(GEOMETRY, OPTS));
+  });
+
+  it('passes null geometries through, keeping array positions aligned', () => {
+    const layouts = computePerfRulerLabelLayouts([null, GEOMETRY, null], OPTS);
+    expect(layouts).toHaveLength(3);
+    expect(layouts[0]).toBeNull();
+    expect(layouts[1]).not.toBeNull();
+    expect(layouts[2]).toBeNull();
+  });
+
+  it('nudges the second of two colliding labels vertically apart', () => {
+    const [first, second] = computePerfRulerLabelLayouts([GEOMETRY, { ...GEOMETRY }], OPTS);
+    expect(first).not.toBeNull();
+    expect(second).not.toBeNull();
+    expect(second!.labelY).not.toBe(first!.labelY);
+    expect(Math.abs(second!.labelY - first!.labelY)).toBeGreaterThanOrEqual(
+      DEFAULT_LABEL_FONT_SIZE + 8,
+    );
+    // The nudged label's arrow and × follow it.
+    expect(second!.deleteY).toBe(second!.labelY);
+    expect(second!.arrowPath).not.toBe(first!.arrowPath);
+  });
+
+  it('leaves far-apart labels untouched', () => {
+    const other = { x: 500, y1: 250, y2: 350, ratioLabel: '1.50x' };
+    const [first, second] = computePerfRulerLabelLayouts([GEOMETRY, other], OPTS);
+    expect(first).toEqual(computePerfRulerLabelLayout(GEOMETRY, OPTS));
+    expect(second).toEqual(computePerfRulerLabelLayout(other, OPTS));
+  });
+
+  it('keeps the below-midpoint collision fallback inside a short chart', () => {
+    // midY = 60 in a 100px chart: the above placement clips the top, and the
+    // naive below-midpoint fallback (midY + 46 = 106) would leave the plot.
+    const geometry = { x: 100, y1: 40, y2: 80, ratioLabel: '2.03x' };
+    const fontSize = DEFAULT_LABEL_FONT_SIZE;
+    const layouts = computePerfRulerLabelLayouts([geometry, { ...geometry }], {
+      chartWidth: 800,
+      chartHeight: 100,
+    });
+    for (const layout of layouts) {
+      expect(layout!.labelY + fontSize / 2).toBeLessThanOrEqual(100 - 4);
+      expect(layout!.labelY - fontSize / 2).toBeGreaterThanOrEqual(4);
+      expect(layout!.deleteY).toBe(layout!.labelY);
+    }
+  });
+
+  it('clamps downward nudges at the chart bottom, accepting residual overlap', () => {
+    // Three identical rulers near the top of a short chart: labels stack
+    // DOWNWARD, and the third label's second nudge (118 + 42 = 160) would
+    // escape a 140px chart without the clamp.
+    const geometry = { x: 100, y1: 10, y2: 50, ratioLabel: '2.03x' };
+    const fontSize = DEFAULT_LABEL_FONT_SIZE;
+    const chartHeight = 140;
+    const layouts = computePerfRulerLabelLayouts([geometry, { ...geometry }, { ...geometry }], {
+      chartWidth: 800,
+      chartHeight,
+    });
+    for (const layout of layouts) {
+      expect(layout!.labelY + fontSize / 2).toBeLessThanOrEqual(chartHeight - 4);
+      expect(layout!.labelY - fontSize / 2).toBeGreaterThanOrEqual(4);
+      expect(layout!.deleteY).toBe(layout!.labelY);
+    }
+    // The first two still resolve their collision; the third accepts overlap
+    // at the clamped bottom rather than leaving the plot.
+    expect(layouts[1]!.labelY).toBeGreaterThan(layouts[0]!.labelY);
+    expect(layouts[2]!.labelY).toBe(chartHeight - 4 - fontSize / 2);
+  });
+});
+
+// ── renderPerfRulers ───────────────────────────────────────────────────
+
+/** Render one ruler (id 1) with its layout computed like the caller does. */
+function renderSingle(
+  group: MockSelection,
+  geometry: PerfRulerGeometry | null,
+  opts?: Partial<PerfRulerRenderOptions>,
+  layoutOpts?: PerfRulerLabelLayoutOptions,
+): void {
+  const entries = [];
+  if (geometry) {
+    const [layout] = computePerfRulerLabelLayouts([geometry], layoutOpts);
+    entries.push({ id: 1, geometry, layout: layout! });
+  }
+  renderPerfRulers(group as any, entries, makeOpts(opts));
+}
+
+describe('renderPerfRulers', () => {
+  it('draws the ruler line, end caps, arrow, big ratio label, drag handle, and ×', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry);
+
+    const ruler = group.selectAll('.perf-ruler');
+    expect(ruler.elements).toHaveLength(1);
+    const children = ruler.elements[0].children.map((c) => String(c.attrs['class']));
+    expect(children).toContain('pr-line');
+    expect(children).toContain('pr-cap pr-cap-top');
+    expect(children).toContain('pr-cap pr-cap-bottom');
+    expect(children).toContain('pr-arrow');
+    expect(children).toContain('pr-arrow-head');
+    expect(children).toContain('pr-text pr-text-ratio');
+    expect(children).toContain('pr-drag');
+    expect(children).toContain('pr-delete no-export');
+    // The chip rect and secondary percent line are gone in the big-label
+    // design.
+    expect(children).not.toContain('pr-bg');
+    expect(children).not.toContain('pr-text pr-text-percent');
+  });
+
+  it('positions the vertical line and caps at the iso-x', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry, { capHalfWidth: 6 });
+
+    const ruler = group.selectAll('.perf-ruler');
+    const byClass = (cls: string) =>
+      ruler.elements[0].children.find((c) => String(c.attrs['class']) === cls)!;
+    const line = byClass('pr-line');
+    expect(line.attrs['x1']).toBe(100);
+    expect(line.attrs['x2']).toBe(100);
+    expect(line.attrs['y1']).toBe(50);
+    expect(line.attrs['y2']).toBe(150);
+    expect(line.attrs['stroke']).toBe('var(--primary)');
+
+    const capTop = byClass('pr-cap pr-cap-top');
+    expect(capTop.attrs['x1']).toBe(94);
+    expect(capTop.attrs['x2']).toBe(106);
+    expect(capTop.attrs['y1']).toBe(50);
+    expect(capTop.attrs['y2']).toBe(50);
+
+    const capBottom = byClass('pr-cap pr-cap-bottom');
+    expect(capBottom.attrs['y1']).toBe(150);
+    expect(capBottom.attrs['y2']).toBe(150);
+  });
+
+  it('overlays an invisible wide drag handle spanning the ruler line', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry);
+
+    const ruler = group.selectAll('.perf-ruler');
+    const drag = ruler.elements[0].children.find((c) => String(c.attrs['class']) === 'pr-drag')!;
+    expect(drag.attrs['x1']).toBe(100);
+    expect(drag.attrs['x2']).toBe(100);
+    expect(drag.attrs['y1']).toBe(50);
+    expect(drag.attrs['y2']).toBe(150);
+    expect(drag.attrs['stroke']).toBe('transparent');
+    expect(Number(drag.attrs['stroke-width'])).toBeGreaterThanOrEqual(12);
+    // Its own pointer-events overrides the group-level 'none' so the handle
+    // stays draggable while the visible marks never block clicks.
+    expect(drag.styles['pointer-events']).toBe('stroke');
+    expect(drag.styles['cursor']).toBe('ew-resize');
+  });
+
+  it('writes the big ratio label in the accent color with a readability halo', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry, undefined, { chartWidth: 800, chartHeight: 400 });
+
+    const ruler = group.selectAll('.perf-ruler');
+    const byClass = (cls: string) =>
+      ruler.elements[0].children.find((c) => String(c.attrs['class']) === cls)!;
+    const label = byClass('pr-text pr-text-ratio');
+    expect(label.textContent).toBe('2.03x');
+    expect(label.attrs['font-size']).toBe(`${DEFAULT_LABEL_FONT_SIZE}px`);
+    expect(label.attrs['font-weight']).toBe('800');
+    expect(label.attrs['fill']).toBe('var(--primary)');
+    // Halo: background-colored stroke painted UNDER the glyph fill.
+    expect(label.attrs['paint-order']).toBe('stroke');
+    expect(label.attrs['stroke']).toBe('var(--background)');
+    expect(Number(label.attrs['stroke-width'])).toBeGreaterThan(0);
+  });
+
+  it('honors a custom halo color and font size', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry, { halo: 'white', labelFontSize: 40 });
+
+    const ruler = group.selectAll('.perf-ruler');
+    const label = ruler.elements[0].children.find(
+      (c) => String(c.attrs['class']) === 'pr-text pr-text-ratio',
+    )!;
+    expect(label.attrs['stroke']).toBe('white');
+    expect(label.attrs['font-size']).toBe('40px');
+  });
+
+  it('draws the arrow curve and filled head in the accent color', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry, undefined, { chartWidth: 800, chartHeight: 400 });
+
+    const ruler = group.selectAll('.perf-ruler');
+    const byClass = (cls: string) =>
+      ruler.elements[0].children.find((c) => String(c.attrs['class']) === cls)!;
+    const layout = computePerfRulerLabelLayout(geometry, { chartWidth: 800, chartHeight: 400 });
+    const arrow = byClass('pr-arrow');
+    expect(arrow.attrs['d']).toBe(layout.arrowPath);
+    expect(arrow.attrs['stroke']).toBe('var(--primary)');
+    expect(arrow.attrs['fill']).toBe('none');
+    const head = byClass('pr-arrow-head');
+    expect(head.attrs['d']).toBe(layout.arrowHeadPath);
+    expect(head.attrs['fill']).toBe('var(--primary)');
+  });
+
+  it('places the label right of the line by default', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry, undefined, { chartWidth: 800 });
+
+    const ruler = group.selectAll('.perf-ruler');
+    const label = ruler.elements[0].children.find(
+      (c) => String(c.attrs['class']) === 'pr-text pr-text-ratio',
+    )!;
+    expect(Number(label.attrs['x'])).toBeGreaterThan(geometry.x);
+    expect(label.attrs['text-anchor']).toBe('start');
+  });
+
+  it('flips the label to the left near the right chart edge', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry, undefined, { chartWidth: 125 });
+
+    const ruler = group.selectAll('.perf-ruler');
+    const label = ruler.elements[0].children.find(
+      (c) => String(c.attrs['class']) === 'pr-text pr-text-ratio',
+    )!;
+    expect(Number(label.attrs['x'])).toBeLessThan(geometry.x);
+    expect(label.attrs['text-anchor']).toBe('end');
+  });
+
+  it('is idempotent: re-rendering keeps a single ruler group', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry);
+    renderSingle(group, { ...geometry, x: 200 });
+
+    const ruler = group.selectAll('.perf-ruler');
+    expect(ruler.elements).toHaveLength(1);
+    const line = ruler.elements[0].children.find((c) => String(c.attrs['class']) === 'pr-line')!;
+    expect(line.attrs['x1']).toBe(200);
+  });
+
+  it('renders one group per ruler and removes exited rulers', () => {
+    const group = createMockGroup();
+    const geometryA = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    const geometryB = computeIsoXRulerGeometry(300, { py: 40, rawY: 900 }, { py: 90, rawY: 300 })!;
+    const layouts = computePerfRulerLabelLayouts([geometryA, geometryB], { chartWidth: 800 });
+    const entries = [
+      { id: 1, geometry: geometryA, layout: layouts[0]! },
+      { id: 2, geometry: geometryB, layout: layouts[1]! },
+    ];
+    renderPerfRulers(group as any, entries, makeOpts());
+    expect(group.selectAll('.perf-ruler').elements).toHaveLength(2);
+
+    renderPerfRulers(group as any, entries.slice(0, 1), makeOpts());
+    expect(group.selectAll('.perf-ruler').elements).toHaveLength(1);
+  });
+
+  it('clears all rulers when the entry list is empty', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry);
+    renderSingle(group, null);
+
+    const ruler = group.selectAll('.perf-ruler');
+    expect(ruler.elements).toHaveLength(0);
+  });
+
+  it('renders nothing when called with no entries on an empty group', () => {
+    const group = createMockGroup();
+    renderSingle(group, null);
+    expect(group.selectAll('.perf-ruler').elements).toHaveLength(0);
+  });
+
+  it('disables pointer events so the ruler never blocks point clicks', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry);
+
+    const ruler = group.selectAll('.perf-ruler');
+    expect(ruler.elements[0].styles['pointer-events']).toBe('none');
+    // …while the big label opts back in so it can reveal the × on hover.
+    const label = ruler.elements[0].children.find(
+      (c) => String(c.attrs['class']) === 'pr-text pr-text-ratio',
+    )!;
+    expect(label.styles['pointer-events']).toBe('auto');
+  });
+
+  it('hides the × delete button by default and marks it no-export', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry);
+
+    const ruler = group.selectAll('.perf-ruler');
+    const del = ruler.elements[0].children.find(
+      (c) => String(c.attrs['class']) === 'pr-delete no-export',
+    )!;
+    // Hidden until hover; excluded from PNG exports via `no-export`.
+    expect(del.styles['display']).toBe('none');
+    expect(del.styles['pointer-events']).toBe('all');
+    expect(del.styles['cursor']).toBe('pointer');
+    const classes = String(del.attrs['class']).split(/\s+/u);
+    expect(classes).toContain('pr-delete');
+    expect(classes).toContain('no-export');
+    // Circular chip + × glyph in contrast colors.
+    const bg = del.children.find((c) => String(c.attrs['class']) === 'pr-delete-bg')!;
+    expect(bg.tag).toBe('circle');
+    expect(Number(bg.attrs['r'])).toBeGreaterThanOrEqual(9);
+    expect(bg.attrs['fill']).toBe('var(--primary)');
+    const x = del.children.find((c) => String(c.attrs['class']) === 'pr-delete-x')!;
+    expect(x.textContent).toBe('×');
+    expect(x.attrs['fill']).toBe('var(--background)');
+  });
+
+  it('invokes onDelete with the ruler id when the × is clicked', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    const deleted: number[] = [];
+    let stopped = 0;
+    renderSingle(group, geometry, { onDelete: (id) => deleted.push(id) });
+
+    const ruler = group.selectAll('.perf-ruler');
+    const del = ruler.elements[0].children.find(
+      (c) => String(c.attrs['class']) === 'pr-delete no-export',
+    )!;
+    const click = del.handlers?.['click'];
+    expect(click).toBeTypeOf('function');
+    click!({ stopPropagation: () => stopped++ }, del.datum);
+    expect(deleted).toEqual([1]);
+    // The click must not fall through to the chart underneath.
+    expect(stopped).toBe(1);
+  });
+
+  it('registers hover handlers that reveal the × per ruler group', () => {
+    const group = createMockGroup();
+    const geometry = computeIsoXRulerGeometry(ISO_X, END_A, END_B)!;
+    renderSingle(group, geometry);
+
+    const ruler = group.selectAll('.perf-ruler');
+    expect(ruler.elements[0].handlers?.['mouseenter']).toBeTypeOf('function');
+    expect(ruler.elements[0].handlers?.['mouseleave']).toBeTypeOf('function');
+  });
+});
+
+// ── nextPerfRulerState ───────────────────────────────────────
+
+/** Complete one measurement (two clicks) on a state, for test setup. */
+function complete(state: PerfRulerState, a: string, b: string, isoX: number): PerfRulerState {
+  return nextPerfRulerState(nextPerfRulerState(state, { curve: a, isoX }), {
+    curve: b,
+    isoX: isoX + 1,
+  });
+}
+
+describe('nextPerfRulerState', () => {
+  const EMPTY = EMPTY_PERF_RULER_STATE;
+
+  it('starts an in-progress draft on the first click', () => {
+    const state = nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: 40 });
+    expect(state.rulers).toEqual([]);
+    expect(state.draft).toEqual({ curve: 'curve-a', isoX: 40 });
+  });
+
+  it('moves the draft iso-x when the draft curve is clicked again', () => {
+    let state = nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: 40 });
+    state = nextPerfRulerState(state, { curve: 'curve-a', isoX: 72 });
+    expect(state.draft).toEqual({ curve: 'curve-a', isoX: 72 });
+    expect(state.rulers).toEqual([]);
+  });
+
+  it('completes a measurement at the DRAFT iso-x when a second curve is clicked', () => {
+    let state = nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: 40 });
+    state = nextPerfRulerState(state, { curve: 'curve-b', isoX: 55 });
+    expect(state.rulers).toEqual([{ id: 1, curveA: 'curve-a', curveB: 'curve-b', isoX: 40 }]);
+    expect(state.draft).toBeNull();
+    expect(state.nextId).toBe(2);
+  });
+
+  it('starts a NEW ruler on the next click after a completed measurement', () => {
+    let state = complete(EMPTY, 'curve-a', 'curve-b', 40);
+    // Clicking curve-a again does NOT retarget ruler 1 — it drafts ruler 2.
+    state = nextPerfRulerState(state, { curve: 'curve-a', isoX: 90 });
+    expect(state.rulers).toHaveLength(1);
+    expect(state.rulers[0].isoX).toBe(40);
+    expect(state.draft).toEqual({ curve: 'curve-a', isoX: 90 });
+    // …and completing it appends a second measurement with a fresh id.
+    state = nextPerfRulerState(state, { curve: 'curve-c', isoX: 95 });
+    expect(state.rulers.map((r) => r.id)).toEqual([1, 2]);
+    expect(state.rulers[1]).toEqual({ id: 2, curveA: 'curve-a', curveB: 'curve-c', isoX: 90 });
+  });
+
+  it('accumulates measurements up to the cap, then drops the OLDEST', () => {
+    let state = EMPTY;
+    for (let i = 0; i < MAX_PERF_RULERS + 2; i++) {
+      state = complete(state, `a${i}`, `b${i}`, i * 10);
+    }
+    expect(state.rulers).toHaveLength(MAX_PERF_RULERS);
+    // The two oldest (a0/b0, a1/b1) were dropped; ids keep counting up.
+    expect(state.rulers[0].curveA).toBe('a2');
+    expect(state.rulers.at(-1)!.curveA).toBe(`a${MAX_PERF_RULERS + 1}`);
+    expect(state.nextId).toBe(MAX_PERF_RULERS + 3);
+  });
+
+  it('returns the same reference for no-op clicks', () => {
+    expect(nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: Number.NaN })).toBe(EMPTY);
+    const drafted = nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: 40 });
+    expect(nextPerfRulerState(drafted, { curve: 'curve-a', isoX: 40 })).toBe(drafted);
+    expect(nextPerfRulerState(drafted, { curve: 'curve-b', isoX: Number.NaN })).toBe(drafted);
+  });
+
+  it('stores the CLAMPED iso-x when completion lands outside the pair overlap', () => {
+    const calls: [string, string, number][] = [];
+    const drafted = nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: 500 });
+    const state = nextPerfRulerState(drafted, { curve: 'curve-b', isoX: 510 }, (a, b, isoX) => {
+      calls.push([a, b, isoX]);
+      return 120; // curve B's span ends at 120 — clamp pulls the ruler back in.
+    });
+    // The clamp sees the DRAFT's iso-x (measurements complete at the draft).
+    expect(calls).toEqual([['curve-a', 'curve-b', 500]]);
+    expect(state.rulers).toEqual([{ id: 1, curveA: 'curve-a', curveB: 'curve-b', isoX: 120 }]);
+    expect(state.draft).toBeNull();
+  });
+
+  it('rejects completion and keeps the draft anchored when the pair has NO overlap', () => {
+    const drafted = nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: 40 });
+    const state = nextPerfRulerState(drafted, { curve: 'curve-b', isoX: 55 }, () => null);
+    // No measurement was appended, no id burned, the draft survives — the
+    // user can pick a different second curve. Same reference: React bails.
+    expect(state).toBe(drafted);
+    expect(state.rulers).toEqual([]);
+    expect(state.draft).toEqual({ curve: 'curve-a', isoX: 40 });
+  });
+
+  it('does not consult the clamp for draft-start or draft-move clicks', () => {
+    let calls = 0;
+    const clamp = () => {
+      calls++;
+      return null;
+    };
+    const drafted = nextPerfRulerState(EMPTY, { curve: 'curve-a', isoX: 40 }, clamp);
+    expect(drafted.draft).toEqual({ curve: 'curve-a', isoX: 40 });
+    const moved = nextPerfRulerState(drafted, { curve: 'curve-a', isoX: 60 }, clamp);
+    expect(moved.draft).toEqual({ curve: 'curve-a', isoX: 60 });
+    expect(calls).toBe(0);
+  });
+});
+
+// ── movePerfRulerIsoX / deletePerfRuler / clearPerfRulers ───────────────
+
+describe('movePerfRulerIsoX', () => {
+  const state = complete(EMPTY_PERF_RULER_STATE, 'curve-a', 'curve-b', 40);
+
+  it('moves only the targeted ruler', () => {
+    const two = complete(state, 'curve-c', 'curve-d', 80);
+    const moved = movePerfRulerIsoX(two, 1, 60);
+    expect(moved.rulers[0].isoX).toBe(60);
+    expect(moved.rulers[1].isoX).toBe(80);
+  });
+
+  it('returns the same reference for unknown ids, equal values, and non-finite x', () => {
+    expect(movePerfRulerIsoX(state, 99, 60)).toBe(state);
+    expect(movePerfRulerIsoX(state, 1, 40)).toBe(state);
+    expect(movePerfRulerIsoX(state, 1, Number.NaN)).toBe(state);
+  });
+});
+
+describe('deletePerfRuler', () => {
+  it('deletes just the targeted ruler by id', () => {
+    const two = complete(complete(EMPTY_PERF_RULER_STATE, 'a', 'b', 40), 'c', 'd', 80);
+    const afterDelete = deletePerfRuler(two, 1);
+    expect(afterDelete.rulers.map((r) => r.id)).toEqual([2]);
+    // Ids are never reused after a delete.
+    expect(afterDelete.nextId).toBe(3);
+  });
+
+  it('returns the same reference when the id is not found', () => {
+    const one = complete(EMPTY_PERF_RULER_STATE, 'a', 'b', 40);
+    expect(deletePerfRuler(one, 99)).toBe(one);
+  });
+});
+
+describe('clearPerfRulers', () => {
+  it('clears all rulers and the draft, preserving the id counter', () => {
+    let state = complete(EMPTY_PERF_RULER_STATE, 'a', 'b', 40);
+    state = nextPerfRulerState(state, { curve: 'c', isoX: 90 });
+    const cleared = clearPerfRulers(state);
+    expect(cleared.rulers).toEqual([]);
+    expect(cleared.draft).toBeNull();
+    expect(cleared.nextId).toBe(state.nextId);
+  });
+
+  it('returns the same reference when already empty', () => {
+    expect(clearPerfRulers(EMPTY_PERF_RULER_STATE)).toBe(EMPTY_PERF_RULER_STATE);
+  });
+});
+
+// ── prunePerfRulers / perfRulerCurveSet ─────────────────────────────
+
+describe('prunePerfRulers', () => {
+  it('removes only rulers whose curves left the data (per ruler)', () => {
+    const two = complete(complete(EMPTY_PERF_RULER_STATE, 'a', 'gone', 40), 'c', 'd', 80);
+    const pruned = prunePerfRulers(two, (curve) => curve !== 'gone');
+    expect(pruned.rulers.map((r) => r.id)).toEqual([2]);
+  });
+
+  it('drops the draft when its curve is gone', () => {
+    const drafted = nextPerfRulerState(EMPTY_PERF_RULER_STATE, { curve: 'gone', isoX: 40 });
+    const pruned = prunePerfRulers(drafted, () => false);
+    expect(pruned.draft).toBeNull();
+  });
+
+  it('returns the same reference when every curve still exists', () => {
+    const one = complete(EMPTY_PERF_RULER_STATE, 'a', 'b', 40);
+    expect(prunePerfRulers(one, () => true)).toBe(one);
+  });
+});
+
+describe('perfRulerCurveSet', () => {
+  it('collects every curve across rulers and the draft', () => {
+    let state = complete(EMPTY_PERF_RULER_STATE, 'a', 'b', 40);
+    state = nextPerfRulerState(state, { curve: 'c', isoX: 90 });
+    expect([...perfRulerCurveSet(state)].sort()).toEqual(['a', 'b', 'c']);
+  });
+
+  it('is empty for the empty state', () => {
+    expect(perfRulerCurveSet(EMPTY_PERF_RULER_STATE).size).toBe(0);
+  });
+});
+
+// ── pathXExtent ─────────────────────────────────────────────
+
+describe('pathXExtent', () => {
+  it('returns the min/max x of the path endpoints', () => {
+    const path = polylinePath([
+      { x: 100, y: 100 },
+      { x: 300, y: 50 },
+    ]);
+    expect(pathXExtent(path)).toEqual({ min: 100, max: 300 });
+  });
+
+  it('supports paths whose x decreases along their length', () => {
+    const path = polylinePath([
+      { x: 300, y: 10 },
+      { x: 100, y: 110 },
+    ]);
+    expect(pathXExtent(path)).toEqual({ min: 100, max: 300 });
+  });
+
+  it('returns null for degenerate paths', () => {
+    expect(pathXExtent(polylinePath([]))).toBeNull();
+    expect(pathXExtent(polylinePath([{ x: 10, y: 10 }]))).toBeNull();
+  });
+});
+
+// ── clampIsoX ───────────────────────────────────────────────
+
+describe('clampIsoX', () => {
+  const A = { min: 100, max: 300 };
+  const B = { min: 200, max: 500 };
+
+  it('clamps to the overlapping range of two curves', () => {
+    expect(clampIsoX(150, A, B)).toBe(200);
+    expect(clampIsoX(250, A, B)).toBe(250);
+    expect(clampIsoX(450, A, B)).toBe(300);
+  });
+
+  it('clamps to a single curve when the second is absent', () => {
+    expect(clampIsoX(50, A)).toBe(100);
+    expect(clampIsoX(350, A, null)).toBe(300);
+    expect(clampIsoX(200, A)).toBe(200);
+  });
+
+  it('returns null when the ranges do not overlap', () => {
+    expect(clampIsoX(250, { min: 100, max: 150 }, { min: 200, max: 300 })).toBeNull();
+  });
+
+  it('returns null for a non-finite x', () => {
+    expect(clampIsoX(Number.NaN, A, B)).toBeNull();
+  });
+});
diff --git a/packages/app/src/lib/d3-chart/layers/perf-ruler.ts b/packages/app/src/lib/d3-chart/layers/perf-ruler.ts
new file mode 100644
index 000000000..c44143529
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/layers/perf-ruler.ts
@@ -0,0 +1,723 @@
+import type * as d3 from 'd3';
+
+/**
+ * Perf ruler layer — an ISO-X (iso-interactivity) measurement ruler between
+ * two curves. The user clicks two curves; the ruler is a vertical line at a
+ * freely chosen iso-x whose BOTH ends are interpolated on the two curves'
+ * rendered roofline paths at that x (neither end needs to be a data point).
+ * The ruler is draggable horizontally to fine-tune the iso-x. Short
+ * horizontal end caps mark both ends and a big annotation-style label
+ * (e.g. "2.03x") sits in open chart space with a curved arrow pointing at
+ * the ruler line; the ratio compares the two RAW y-values at that x.
+ * Multiple rulers can coexist (capped at {@link MAX_PERF_RULERS}); each is
+ * individually draggable and deletable via a hover × button.
+ *
+ * Pure module: intersection search and geometry math are separated from
+ * rendering so all three are unit testable (see perf-ruler.test.ts).
+ * Rendering follows the narrow-mutation rules from docs/d3-charts.md — a
+ * single keyed join, texts written before any measurement, rects sized last.
+ */
+
+type GroupSelection = d3.Selection<SVGGElement, unknown, null, undefined>;
+
+/** One end of the ruler: a curve's interpolated position at the iso-x. */
+export interface PerfRulerEndInput {
+  /** Pixel y of the curve at the iso-x (from the rendered path). */
+  py: number;
+  /** Raw data-space y at that intersection (yScale.invert of `py`). */
+  rawY: number;
+}
+
+export interface PerfRulerGeometry {
+  /** Ruler line x pixel position: the iso-x through the x scale. */
+  x: number;
+  /** Top pixel y of the ruler span. */
+  y1: number;
+  /** Bottom pixel y of the ruler span. */
+  y2: number;
+  /** Performance multiple: higher raw y over lower raw y (>= 1). */
+  ratio: number;
+  /** Formatted ratio, e.g. "2.03x". */
+  ratioLabel: string;
+}
+
+/** Format a performance multiple like "2.03x" (fewer decimals as it grows). */
+export function formatPerfRatio(ratio: number): string {
+  if (!Number.isFinite(ratio) || ratio <= 0) return '';
+  if (ratio >= 100) return `${Math.round(ratio)}x`;
+  if (ratio >= 10) return `${ratio.toFixed(1)}x`;
+  return `${ratio.toFixed(2)}x`;
+}
+
+/**
+ * Minimal path interface needed by {@link intersectPathAtX}. Satisfied by a
+ * real `SVGPathElement`; unit tests supply a synthetic polyline
+ * implementation, since jsdom has no path-length support.
+ */
+export interface PerfRulerPathLike {
+  getTotalLength: () => number;
+  getPointAtLength: (length: number) => { x: number; y: number };
+}
+
+/**
+ * Find the point on `path` at horizontal pixel position `x` via binary search
+ * over the path-length parameter. Valid for paths whose x is monotonic along
+ * their length — true for every roofline: they are Pareto frontiers sorted by
+ * x and rendered with `d3.curveMonotoneX`, which preserves x-monotonicity.
+ *
+ * Works in RENDERED pixel space: the roofline `d` attribute is rewritten with
+ * the current scales on every zoom/metric pass before this layer runs, so the
+ * intersection matches the drawn curve exactly at any zoom level. Returns
+ * null when `x` lies outside the path's x extent (curve B does not span the
+ * anchor's x) or the path is degenerate.
+ */
+export function intersectPathAtX(
+  path: PerfRulerPathLike,
+  x: number,
+  opts?: { tolerance?: number; maxIterations?: number },
+): { x: number; y: number } | null {
+  if (!Number.isFinite(x)) return null;
+  const total = path.getTotalLength();
+  if (!Number.isFinite(total) || total <= 0) return null;
+
+  const start = path.getPointAtLength(0);
+  const end = path.getPointAtLength(total);
+  const ascending = end.x >= start.x;
+  const minX = Math.min(start.x, end.x);
+  const maxX = Math.max(start.x, end.x);
+  // Half-pixel slack so an anchor sitting exactly on a curve endpoint (a
+  // shared frontier point) still intersects despite float noise.
+  const slack = 0.5;
+  if (x < minX - slack || x > maxX + slack) return null;
+
+  const tolerance = opts?.tolerance ?? 0.25;
+  const maxIterations = opts?.maxIterations ?? 48;
+  let lo = 0;
+  let hi = total;
+  let best = Math.abs(start.x - x) <= Math.abs(end.x - x) ? start : end;
+  for (let i = 0; i < maxIterations; i++) {
+    const mid = (lo + hi) / 2;
+    const p = path.getPointAtLength(mid);
+    if (Math.abs(p.x - x) < Math.abs(best.x - x)) best = p;
+    if (Math.abs(p.x - x) <= tolerance) return { x: p.x, y: p.y };
+    if (ascending === p.x < x) lo = mid;
+    else hi = mid;
+  }
+  // Interval exhausted without hitting tolerance (extremely steep segment) —
+  // the closest sample seen is still visually on the curve.
+  return { x: best.x, y: best.y };
+}
+
+/**
+ * Compute iso-x ruler geometry from the two curves' interpolated positions
+ * at the iso-x pixel `x`. Returns null for degenerate inputs: non-finite
+ * coordinates or non-positive raw y values (a ratio over a zero/negative
+ * value is meaningless, and log scales cannot place them). The ratio is
+ * symmetric: higher raw y over lower raw y, regardless of click order.
+ */
+export function computeIsoXRulerGeometry(
+  x: number,
+  a: PerfRulerEndInput,
+  b: PerfRulerEndInput,
+): PerfRulerGeometry | null {
+  const inputs = [x, a.py, a.rawY, b.py, b.rawY];
+  if (!inputs.every((value) => Number.isFinite(value))) return null;
+  if (a.rawY <= 0 || b.rawY <= 0) return null;
+
+  const hi = Math.max(a.rawY, b.rawY);
+  const lo = Math.min(a.rawY, b.rawY);
+  const ratio = hi / lo;
+
+  return {
+    x,
+    y1: Math.min(a.py, b.py),
+    y2: Math.max(a.py, b.py),
+    ratio,
+    ratioLabel: formatPerfRatio(ratio),
+  };
+}
+
+/**
+ * Whether a rendered curve is visible given its inline `opacity` style.
+ * Legend and precision toggles hide curves with `opacity: 0` while leaving
+ * the path in the DOM (so it can fade back in) — those curves must be
+ * neither clickable nor measurable. Hover dimming uses small non-zero
+ * values and stays measurable; a missing/empty opacity means fully visible.
+ */
+export function isPerfRulerCurveVisible(opacity: string | null | undefined): boolean {
+  if (opacity === null || opacity === undefined) return true;
+  const trimmed = opacity.trim();
+  if (trimmed === '') return true;
+  const value = Number(trimmed);
+  return Number.isNaN(value) ? true : value > 0;
+}
+
+/** Inclusive pixel x range, e.g. the horizontal extent of a rendered path. */
+export interface PerfRulerXRange {
+  min: number;
+  max: number;
+}
+
+/**
+ * Horizontal pixel extent of a rendered path. Valid for paths whose x is
+ * monotonic along the length parameter (rooflines: pareto frontiers sorted
+ * by x). Returns null for empty/degenerate paths.
+ */
+export function pathXExtent(path: PerfRulerPathLike): PerfRulerXRange | null {
+  const total = path.getTotalLength();
+  if (!Number.isFinite(total) || total <= 0) return null;
+  const start = path.getPointAtLength(0);
+  const end = path.getPointAtLength(total);
+  if (!Number.isFinite(start.x) || !Number.isFinite(end.x)) return null;
+  return { min: Math.min(start.x, end.x), max: Math.max(start.x, end.x) };
+}
+
+/**
+ * Clamp an iso-x pixel position to the overlapping x range of two curves so
+ * a drag can never leave the span where both intersections exist. Returns
+ * null when the ranges do not overlap (no valid iso-x at all) or the input
+ * is not finite.
+ */
+export function clampIsoX(
+  x: number,
+  a: PerfRulerXRange,
+  b?: PerfRulerXRange | null,
+): number | null {
+  if (!Number.isFinite(x)) return null;
+  const min = b ? Math.max(a.min, b.min) : a.min;
+  const max = b ? Math.min(a.max, b.max) : a.max;
+  if (min > max) return null;
+  return Math.min(Math.max(x, min), max);
+}
+
+/** One perf-ruler click: a curve identity plus the click's data-space x. */
+export interface PerfRulerCurveClick {
+  curve: string;
+  isoX: number;
+}
+
+/** One completed measurement: a curve pair plus its iso-x in DATA space. */
+export interface PerfRulerMeasurement {
+  /** Stable id — render join key, drag identity, and delete target. */
+  id: number;
+  curveA: string;
+  curveB: string;
+  /** Iso-x in DATA space, so it survives zoom and metric changes. */
+  isoX: number;
+}
+
+/**
+ * Multi-ruler state: completed measurements accumulate in `rulers` while at
+ * most one in-progress `draft` (curve A + iso-x, awaiting curve B) exists.
+ * `nextId` is a monotonic counter so deleted ids are never reused.
+ */
+export interface PerfRulerState {
+  rulers: PerfRulerMeasurement[];
+  draft: PerfRulerCurveClick | null;
+  nextId: number;
+}
+
+/**
+ * Ruler cap — completing a measurement beyond this silently drops the
+ * OLDEST ruler (keeps the chart readable while never rejecting the newest
+ * measurement the user just placed).
+ */
+export const MAX_PERF_RULERS = 8;
+
+export const EMPTY_PERF_RULER_STATE: PerfRulerState = { rulers: [], draft: null, nextId: 1 };
+
+/**
+ * Click-transition table for multi-ruler mode (clicks land on CURVES,
+ * either on a widened curve hit stroke or via a data point standing in for
+ * its curve at that point's x):
+ *
+ * | State      | Click on…              | Result                             |
+ * |------------|------------------------|------------------------------------|
+ * | no draft   | any curve              | new draft: curve A, iso-x = click  |
+ * | draft on A | curve A again          | MOVE the draft iso-x to the click  |
+ * | draft on A | a different curve B    | measurement COMPLETE at the draft  |
+ * |            |                        | iso-x — appended to `rulers`, the  |
+ * |            |                        | next click starts a fresh draft    |
+ *
+ * Completed rulers are never retargeted by clicks — clicking any curve
+ * after completion starts a NEW ruler (move/reset semantics apply only to
+ * the in-progress draft). Dragging a completed ruler moves its own iso-x
+ * (see {@link movePerfRulerIsoX}); toggling the mode off clears everything.
+ * Returns `prev` (same reference) for no-ops so React state can bail out.
+ *
+ * `clampToOverlap` (optional) constrains a COMPLETING measurement's iso-x
+ * to the overlapping x-range of its curve pair (the caller owns the
+ * pixel-space paths/scales, so it supplies the clamp). Returning null
+ * means the pair shares NO x-range at all — the measurement is rejected
+ * and the draft stays anchored so the user can pick a different second
+ * curve. Without the clamp, an out-of-range iso-x would append a ruler
+ * that never renders and has no drag handle to recover it: stuck in
+ * state, counted against the cap, removable only by clear/toggle-off.
+ */
+export function nextPerfRulerState(
+  prev: PerfRulerState,
+  click: PerfRulerCurveClick,
+  clampToOverlap?: (curveA: string, curveB: string, isoX: number) => number | null,
+): PerfRulerState {
+  if (!Number.isFinite(click.isoX)) return prev;
+  if (!prev.draft) return { ...prev, draft: { curve: click.curve, isoX: click.isoX } };
+  if (click.curve === prev.draft.curve) {
+    return prev.draft.isoX === click.isoX
+      ? prev
+      : { ...prev, draft: { curve: prev.draft.curve, isoX: click.isoX } };
+  }
+  let isoX = prev.draft.isoX;
+  if (clampToOverlap) {
+    const clamped = clampToOverlap(prev.draft.curve, click.curve, isoX);
+    if (clamped === null || !Number.isFinite(clamped)) return prev;
+    isoX = clamped;
+  }
+  const measurement: PerfRulerMeasurement = {
+    id: prev.nextId,
+    curveA: prev.draft.curve,
+    curveB: click.curve,
+    isoX,
+  };
+  const rulers = [...prev.rulers, measurement];
+  while (rulers.length > MAX_PERF_RULERS) rulers.shift();
+  return { rulers, draft: null, nextId: prev.nextId + 1 };
+}
+
+/** Move one completed ruler's iso-x (drag commit). No-ops return `prev`. */
+export function movePerfRulerIsoX(prev: PerfRulerState, id: number, isoX: number): PerfRulerState {
+  if (!Number.isFinite(isoX)) return prev;
+  const index = prev.rulers.findIndex((ruler) => ruler.id === id);
+  if (index === -1 || prev.rulers[index].isoX === isoX) return prev;
+  const rulers = [...prev.rulers];
+  rulers[index] = { ...rulers[index], isoX };
+  return { ...prev, rulers };
+}
+
+/** Delete one ruler by id. Unknown ids return `prev` (same reference). */
+export function deletePerfRuler(prev: PerfRulerState, id: number): PerfRulerState {
+  const rulers = prev.rulers.filter((ruler) => ruler.id !== id);
+  return rulers.length === prev.rulers.length ? prev : { ...prev, rulers };
+}
+
+/** Clear all rulers and the draft. Already-empty state returns `prev`. */
+export function clearPerfRulers(prev: PerfRulerState): PerfRulerState {
+  if (prev.rulers.length === 0 && prev.draft === null) return prev;
+  return { rulers: [], draft: null, nextId: prev.nextId };
+}
+
+/**
+ * Prune rulers whose curves left the underlying data entirely (per-ruler:
+ * a ruler survives while BOTH its curves still exist). Hidden-but-present
+ * curves are the caller's visibility concern, not pruning — pass an
+ * existence predicate, not a visibility one. Returns `prev` on no-ops.
+ */
+export function prunePerfRulers(
+  prev: PerfRulerState,
+  curveExists: (curve: string) => boolean,
+): PerfRulerState {
+  const rulers = prev.rulers.filter(
+    (ruler) => curveExists(ruler.curveA) && curveExists(ruler.curveB),
+  );
+  const draft = prev.draft && curveExists(prev.draft.curve) ? prev.draft : null;
+  if (rulers.length === prev.rulers.length && draft === prev.draft) return prev;
+  return { ...prev, rulers, draft };
+}
+
+/** Every curve referenced by any ruler or the draft (hit-halo styling). */
+export function perfRulerCurveSet(state: PerfRulerState): Set<string> {
+  const curves = new Set<string>();
+  for (const ruler of state.rulers) {
+    curves.add(ruler.curveA);
+    curves.add(ruler.curveB);
+  }
+  if (state.draft) curves.add(state.draft.curve);
+  return curves;
+}
+
+export interface PerfRulerRenderOptions {
+  /** Stroke for the ruler line, caps, arrow, and label (accent color). */
+  color: string;
+  /**
+   * Halo stroke painted behind the big label (paint-order: stroke) so it
+   * stays readable over chart content. Default `var(--background)`.
+   */
+  halo?: string;
+  /** Half-length of the horizontal end caps in px. Default 6. */
+  capHalfWidth?: number;
+  /** Big ratio label font size in px. Default 32. */
+  labelFontSize?: number;
+  /** Called with the ruler's id when its hover × button is clicked. */
+  onDelete?: (id: number) => void;
+}
+
+export const DEFAULT_LABEL_FONT_SIZE = 32;
+const LABEL_OFFSET_X = 46;
+const LABEL_OFFSET_Y = 46;
+const ARROW_TIP_GAP = 5;
+const ARROW_HEAD_LENGTH = 9;
+const ARROW_HEAD_HALF_WIDTH = 4.5;
+const DELETE_GAP = 18;
+const LABEL_COLLISION_NUDGE_TRIES = 4;
+
+export interface PerfRulerLabelLayoutOptions {
+  /** Chart inner width; the label flips to the left side when clipping. */
+  chartWidth?: number;
+  /** Chart inner height; keeps the below-placement inside the chart. */
+  chartHeight?: number;
+  /** Label font size in px. Default {@link DEFAULT_LABEL_FONT_SIZE}. */
+  fontSize?: number;
+}
+
+export interface PerfRulerLabelLayout {
+  /** Text anchor-point x (near edge of the label, facing the line). */
+  labelX: number;
+  /** Text center y (render with dominant-baseline: central). */
+  labelY: number;
+  /** 'start' when the label sits right of the line, 'end' when left. */
+  textAnchor: 'start' | 'end';
+  /** +1 = label right of the ruler line, -1 = left. */
+  side: 1 | -1;
+  /** Curved arrow from under the label to the ruler-line midpoint. */
+  arrowPath: string;
+  /** Filled arrowhead triangle pointing at the ruler-line midpoint. */
+  arrowHeadPath: string;
+  /** Center of the × delete button, just past the label's far end. */
+  deleteX: number;
+  deleteY: number;
+}
+
+/** Estimated label width without a DOM (unit-testable placement math). */
+function estimateLabelWidth(label: string, fontSize: number): number {
+  return label.length * fontSize * 0.6;
+}
+
+/** Build the arrow + delete positions for a chosen label placement. */
+function buildLabelLayout(
+  geometry: Pick<PerfRulerGeometry, 'x' | 'y1' | 'y2' | 'ratioLabel'>,
+  side: 1 | -1,
+  labelY: number,
+  fontSize: number,
+): PerfRulerLabelLayout {
+  const midY = (geometry.y1 + geometry.y2) / 2;
+  const labelX = geometry.x + side * LABEL_OFFSET_X;
+  const above = labelY < midY;
+  // Quarter-curve: leaves the label vertically, ends horizontally at the
+  // back of the arrowhead so the head points straight at the line.
+  const startX = labelX + side * 4;
+  const startY = labelY + (above ? 1 : -1) * (fontSize / 2 + 6);
+  const tipX = geometry.x + side * ARROW_TIP_GAP;
+  const backX = tipX + side * ARROW_HEAD_LENGTH;
+  const arrowPath = `M ${startX} ${startY} Q ${startX} ${midY} ${backX} ${midY}`;
+  const arrowHeadPath =
+    `M ${tipX} ${midY} ` +
+    `L ${backX} ${midY - ARROW_HEAD_HALF_WIDTH} ` +
+    `L ${backX} ${midY + ARROW_HEAD_HALF_WIDTH} Z`;
+  const estimatedWidth = estimateLabelWidth(geometry.ratioLabel, fontSize);
+  return {
+    labelX,
+    labelY,
+    textAnchor: side === 1 ? 'start' : 'end',
+    side,
+    arrowPath,
+    arrowHeadPath,
+    deleteX: labelX + side * (estimatedWidth + DELETE_GAP),
+    deleteY: labelY,
+  };
+}
+
+/**
+ * Lay out the big annotation label and its arrow (mock-up: a large "2x"
+ * next to the ruler with a curved arrow pointing at the vertical line).
+ * The label prefers the upper-right of the line midpoint (open chart
+ * space), flips horizontally when it would clip the right edge with more
+ * room on the left, and drops below the midpoint when it would clip the
+ * top. The arrow is a quarter-curve whose end tangent is horizontal, so
+ * the arrowhead points squarely at the line's midpoint. Pure — unit
+ * testable without a DOM (text width is estimated from the label length).
+ */
+export function computePerfRulerLabelLayout(
+  geometry: Pick<PerfRulerGeometry, 'x' | 'y1' | 'y2' | 'ratioLabel'>,
+  opts?: PerfRulerLabelLayoutOptions,
+): PerfRulerLabelLayout {
+  const fontSize = opts?.fontSize ?? DEFAULT_LABEL_FONT_SIZE;
+  const midY = (geometry.y1 + geometry.y2) / 2;
+  const estimatedWidth = estimateLabelWidth(geometry.ratioLabel, fontSize);
+  const requiredRoom = LABEL_OFFSET_X + estimatedWidth + 8;
+
+  let side: 1 | -1 = 1;
+  if (opts?.chartWidth !== undefined) {
+    const roomRight = opts.chartWidth - geometry.x;
+    if (roomRight < requiredRoom && geometry.x > roomRight) side = -1;
+  }
+
+  let labelY = midY - LABEL_OFFSET_Y;
+  if (labelY - fontSize / 2 < 4) {
+    labelY = midY + LABEL_OFFSET_Y;
+    if (opts?.chartHeight !== undefined && labelY + fontSize / 2 > opts.chartHeight - 4) {
+      // No room below either — clamp the above-placement inside the chart.
+      labelY = Math.max(fontSize / 2 + 4, midY - LABEL_OFFSET_Y);
+    }
+  }
+
+  return buildLabelLayout(geometry, side, labelY, fontSize);
+}
+
+/**
+ * Clamp a label center-y so the glyph box (plus 4px padding) stays inside
+ * [0, chartHeight]. The top clamp always applies; the bottom clamp only
+ * when the chart height is known. Top wins if the chart is shorter than
+ * one label.
+ */
+function clampLabelY(y: number, fontSize: number, chartHeight?: number): number {
+  const min = fontSize / 2 + 4;
+  const max = chartHeight === undefined ? Number.POSITIVE_INFINITY : chartHeight - 4 - fontSize / 2;
+  return Math.max(Math.min(y, max), min);
+}
+
+/** Horizontal pixel span of a laid-out label (estimated, no DOM). */
+function labelSpan(
+  layout: PerfRulerLabelLayout,
+  label: string,
+  fontSize: number,
+): { x0: number; x1: number } {
+  const width = estimateLabelWidth(label, fontSize);
+  return layout.textAnchor === 'start'
+    ? { x0: layout.labelX, x1: layout.labelX + width }
+    : { x0: layout.labelX - width, x1: layout.labelX };
+}
+
+/**
+ * Lay out labels for MULTIPLE rulers at once with cheap collision handling:
+ * each label after the first is nudged vertically (away from its ruler's
+ * midpoint, up to a few steps) while its estimated box overlaps an already
+ * placed label. Best-effort — after the nudge budget, residual overlap is
+ * accepted. Null geometries pass through as null (hidden rulers keep their
+ * slot so results align with the input array).
+ */
+export function computePerfRulerLabelLayouts(
+  geometries: (Pick<PerfRulerGeometry, 'x' | 'y1' | 'y2' | 'ratioLabel'> | null)[],
+  opts?: PerfRulerLabelLayoutOptions,
+): (PerfRulerLabelLayout | null)[] {
+  const fontSize = opts?.fontSize ?? DEFAULT_LABEL_FONT_SIZE;
+  const placed: { x0: number; x1: number; y: number }[] = [];
+  return geometries.map((geometry) => {
+    if (!geometry) return null;
+    let layout = computePerfRulerLabelLayout(geometry, opts);
+    const midY = (geometry.y1 + geometry.y2) / 2;
+    const overlapsPlaced = (candidate: PerfRulerLabelLayout): boolean => {
+      const span = labelSpan(candidate, geometry.ratioLabel, fontSize);
+      return placed.some(
+        (box) =>
+          span.x0 < box.x1 && box.x0 < span.x1 && Math.abs(candidate.labelY - box.y) < fontSize + 8,
+      );
+    };
+    for (
+      let attempt = 0;
+      attempt < LABEL_COLLISION_NUDGE_TRIES && overlapsPlaced(layout);
+      attempt++
+    ) {
+      // Nudge away from the midpoint in the label's current direction; when
+      // that would clip the top, restart below the midpoint instead.
+      const direction = layout.labelY < midY ? -1 : 1;
+      let nudgedY = layout.labelY + direction * (fontSize + 10);
+      if (direction === -1 && nudgedY - fontSize / 2 < 4) nudgedY = midY + LABEL_OFFSET_Y;
+      // Every nudge (and the below-midpoint fallback) re-applies the same
+      // vertical clamp as the initial placement — a label and its × chip
+      // must never escape the plot, even at the cost of residual overlap.
+      nudgedY = clampLabelY(nudgedY, fontSize, opts?.chartHeight);
+      // Clamped into a wall: no further progress is possible this direction.
+      if (nudgedY === layout.labelY) break;
+      layout = buildLabelLayout(geometry, layout.side, nudgedY, fontSize);
+    }
+    const span = labelSpan(layout, geometry.ratioLabel, fontSize);
+    placed.push({ x0: span.x0, x1: span.x1, y: layout.labelY });
+    return layout;
+  });
+}
+
+/** One ruler ready to render: id (join key) + geometry + label layout. */
+export interface PerfRulerRenderEntry {
+  id: number;
+  geometry: PerfRulerGeometry;
+  layout: PerfRulerLabelLayout;
+}
+
+/** Delay before the hover × hides — bridges the label → × pointer gap. */
+const DELETE_HIDE_DELAY_MS = 250;
+const hideTimers = new WeakMap<Element, ReturnType<typeof setTimeout>>();
+
+/** Toggle one ruler group's hover affordances (× button + line emphasis). */
+function setPerfRulerHover(node: Element, hovered: boolean): void {
+  const deleteButton = node.querySelector<SVGGElement>('.pr-delete');
+  if (deleteButton) deleteButton.style.display = hovered ? '' : 'none';
+  const line = node.querySelector<SVGLineElement>('.pr-line');
+  line?.setAttribute('stroke-width', hovered ? '3' : '2');
+}
+
+/**
+ * Render ALL perf rulers inside `group` as one keyed join (join key: ruler
+ * id), so rulers enter/exit independently and drags stay bound to the right
+ * ruler. Idempotent — safe to call from render, zoom, and display passes.
+ *
+ * Each ruler group is pointer-events: none so it never intercepts point
+ * clicks; only its drag handle, its big label, and its × button opt back in.
+ * Hovering the label (or drag handle) reveals a circular × delete button
+ * next to the label — the × carries the `no-export` class, which
+ * useChartExport strips from PNG exports.
+ */
+export function renderPerfRulers(
+  group: GroupSelection,
+  entries: PerfRulerRenderEntry[],
+  opts: PerfRulerRenderOptions,
+): void {
+  const selection = group
+    .selectAll<SVGGElement, PerfRulerRenderEntry>('.perf-ruler')
+    .data(entries, (entry) => String(entry.id));
+
+  selection.exit().remove();
+
+  const entered = selection.enter().append('g').attr('class', 'perf-ruler');
+  entered
+    .append('line')
+    .attr('class', 'pr-line')
+    .attr('stroke-width', 2)
+    .attr('stroke-dasharray', '5 4');
+  entered.append('line').attr('class', 'pr-cap pr-cap-top').attr('stroke-width', 2);
+  entered.append('line').attr('class', 'pr-cap pr-cap-bottom').attr('stroke-width', 2);
+  // Curved annotation arrow + filled head pointing at the line midpoint.
+  // The head is a plain filled triangle (no <marker>), so it needs no defs
+  // ids, inherits nothing chart-specific, and PNG-exports like any path.
+  entered
+    .append('path')
+    .attr('class', 'pr-arrow')
+    .attr('fill', 'none')
+    .attr('stroke-width', 2.5)
+    .attr('stroke-linecap', 'round');
+  entered.append('path').attr('class', 'pr-arrow-head').attr('stroke', 'none');
+  // Big annotation-style ratio label with a halo stroke painted UNDER the
+  // glyph fill (paint-order) so it reads over dense chart content without a
+  // chip rect. It is hoverable (pointer-events overrides the group's none)
+  // so it can reveal this ruler's × delete button.
+  entered
+    .append('text')
+    .attr('class', 'pr-text pr-text-ratio')
+    .attr('dominant-baseline', 'central')
+    .attr('font-weight', '800')
+    .attr('paint-order', 'stroke')
+    .attr('stroke-linejoin', 'round')
+    .style('pointer-events', 'auto')
+    .style('cursor', 'default');
+  // Wide invisible drag handle over the line — the caller attaches d3.drag
+  // to it to move this ruler's iso-x. Its own pointer-events overrides the
+  // group's `none` (pointer-events inherits).
+  entered
+    .append('line')
+    .attr('class', 'pr-drag')
+    .attr('stroke', 'transparent')
+    .attr('stroke-width', 16)
+    .style('pointer-events', 'stroke')
+    .style('cursor', 'ew-resize');
+  // Circular × delete button, revealed on hover. `no-export` keeps it out
+  // of PNG exports (useChartExport hides/filters that class).
+  const enteredDelete = entered
+    .append('g')
+    .attr('class', 'pr-delete no-export')
+    .style('display', 'none')
+    .style('pointer-events', 'all')
+    .style('cursor', 'pointer');
+  enteredDelete.append('circle').attr('class', 'pr-delete-bg').attr('r', 10);
+  enteredDelete
+    .append('text')
+    .attr('class', 'pr-delete-x')
+    .attr('text-anchor', 'middle')
+    .attr('dominant-baseline', 'central')
+    .attr('font-size', '12px')
+    .attr('font-weight', '700')
+    .text('×');
+
+  const merged = entered.merge(selection).style('pointer-events', 'none');
+
+  const capHalfWidth = opts.capHalfWidth ?? 6;
+  const fontSize = opts.labelFontSize ?? DEFAULT_LABEL_FONT_SIZE;
+
+  merged
+    .select<SVGLineElement>('.pr-line')
+    .attr('x1', (entry) => entry.geometry.x)
+    .attr('x2', (entry) => entry.geometry.x)
+    .attr('y1', (entry) => entry.geometry.y1)
+    .attr('y2', (entry) => entry.geometry.y2)
+    .attr('stroke', opts.color);
+  merged
+    .select<SVGLineElement>('.pr-cap-top')
+    .attr('x1', (entry) => entry.geometry.x - capHalfWidth)
+    .attr('x2', (entry) => entry.geometry.x + capHalfWidth)
+    .attr('y1', (entry) => entry.geometry.y1)
+    .attr('y2', (entry) => entry.geometry.y1)
+    .attr('stroke', opts.color);
+  merged
+    .select<SVGLineElement>('.pr-cap-bottom')
+    .attr('x1', (entry) => entry.geometry.x - capHalfWidth)
+    .attr('x2', (entry) => entry.geometry.x + capHalfWidth)
+    .attr('y1', (entry) => entry.geometry.y2)
+    .attr('y2', (entry) => entry.geometry.y2)
+    .attr('stroke', opts.color);
+  merged
+    .select<SVGLineElement>('.pr-drag')
+    .attr('x1', (entry) => entry.geometry.x)
+    .attr('x2', (entry) => entry.geometry.x)
+    .attr('y1', (entry) => entry.geometry.y1)
+    .attr('y2', (entry) => entry.geometry.y2);
+
+  merged
+    .select<SVGPathElement>('.pr-arrow')
+    .attr('d', (entry) => entry.layout.arrowPath)
+    .attr('stroke', opts.color);
+  merged
+    .select<SVGPathElement>('.pr-arrow-head')
+    .attr('d', (entry) => entry.layout.arrowHeadPath)
+    .attr('fill', opts.color);
+  merged
+    .select<SVGTextElement>('.pr-text-ratio')
+    .attr('x', (entry) => entry.layout.labelX)
+    .attr('y', (entry) => entry.layout.labelY)
+    .attr('text-anchor', (entry) => entry.layout.textAnchor)
+    .attr('font-size', `${fontSize}px`)
+    .attr('fill', opts.color)
+    .attr('stroke', opts.halo ?? 'var(--background)')
+    .attr('stroke-width', 5)
+    .text((entry) => entry.geometry.ratioLabel);
+
+  merged
+    .select<SVGGElement>('.pr-delete')
+    .attr('transform', (entry) => `translate(${entry.layout.deleteX}, ${entry.layout.deleteY})`)
+    .on('click', (event: MouseEvent, entry: PerfRulerRenderEntry) => {
+      event.stopPropagation();
+      opts.onDelete?.(entry.id);
+    });
+  merged.select<SVGCircleElement>('.pr-delete-bg').attr('fill', opts.color);
+  merged.select<SVGTextElement>('.pr-delete-x').attr('fill', opts.halo ?? 'var(--background)');
+
+  // Hover reveal: the group itself is not hit-testable, but mouseenter /
+  // mouseleave fire on it when the pointer enters/leaves its interactive
+  // children (label, drag handle, ×). The pointer crosses dead space
+  // between the label and the ×, so hide on a short delay and cancel the
+  // timer when hover resumes.
+  merged
+    .on('mouseenter', (event: MouseEvent) => {
+      const node = event.currentTarget as Element;
+      const timer = hideTimers.get(node);
+      if (timer !== undefined) {
+        clearTimeout(timer);
+        hideTimers.delete(node);
+      }
+      setPerfRulerHover(node, true);
+    })
+    .on('mouseleave', (event: MouseEvent) => {
+      const node = event.currentTarget as Element;
+      const timer = setTimeout(() => {
+        hideTimers.delete(node);
+        setPerfRulerHover(node, false);
+      }, DELETE_HIDE_DELAY_MS);
+      hideTimers.set(node, timer);
+    });
+}
diff --git a/packages/app/src/lib/d3-chart/layers/radar.ts b/packages/app/src/lib/d3-chart/layers/radar.ts
index 64a881159..8d19336fe 100644
--- a/packages/app/src/lib/d3-chart/layers/radar.ts
+++ b/packages/app/src/lib/d3-chart/layers/radar.ts
@@ -1,5 +1,7 @@
 import * as d3 from 'd3';
 
+import { CHART_TYPE, px } from '../typography';
+
 export interface RadarConfig<T = any> {
   /** Labels for each axis spoke. */
   axes: { label: string; unit?: string }[];
@@ -95,7 +97,7 @@ export function renderRadar<T>(
     .attr('class', 'radar-grid-label')
     .attr('x', 4)
     .attr('y', (d) => -rPerLevel * d + 4)
-    .attr('font-size', '9px')
+    .attr('font-size', px(CHART_TYPE.micro))
     .attr('fill', 'var(--border)')
     .attr('opacity', 0.6)
     .text((d) => `${Math.round((d / levels) * 100)}%`);
@@ -130,7 +132,7 @@ export function renderRadar<T>(
       return cos > 0 ? 'start' : 'end';
     })
     .attr('dominant-baseline', 'central')
-    .attr('font-size', '10px')
+    .attr('font-size', px(CHART_TYPE.dataLabel))
     .attr('font-weight', '500')
     .attr('fill', 'var(--foreground)')
     .text((d) => d.label);
diff --git a/packages/app/src/lib/d3-chart/layers/rooflines.ts b/packages/app/src/lib/d3-chart/layers/rooflines.ts
index 5513ef121..0c5cfb409 100644
--- a/packages/app/src/lib/d3-chart/layers/rooflines.ts
+++ b/packages/app/src/lib/d3-chart/layers/rooflines.ts
@@ -72,6 +72,23 @@ export function renderRooflines<T extends { x: number; y: number }>(
   });
 }
 
+/** Restyle existing rooflines without rejoining entries or rewriting path geometry. */
+export function updateRooflinesForDisplay(
+  zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+  config: RooflineConfig,
+): void {
+  const { getColor, getOpacity, strokeWidth = 2, strokeDasharray } = config;
+  zoomGroup
+    .selectAll<SVGPathElement, RooflineEntry<{ x: number; y: number }>>('.roofline-path')
+    .attr('stroke', (d) => getColor(d.key))
+    .attr('stroke-width', strokeWidth)
+    .attr('stroke-dasharray', strokeDasharray ?? null)
+    .each(function (d) {
+      const opacity = getOpacity?.(d.key);
+      if (opacity !== undefined) d3.select(this).style('opacity', opacity);
+    });
+}
+
 /** Update roofline paths on zoom with new scales. */
 export function updateRooflinesOnZoom<T extends { x: number; y: number }>(
   zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
diff --git a/packages/app/src/lib/d3-chart/layers/scatter-points.test.ts b/packages/app/src/lib/d3-chart/layers/scatter-points.test.ts
index 8b691ee44..7eb74bdb8 100644
--- a/packages/app/src/lib/d3-chart/layers/scatter-points.test.ts
+++ b/packages/app/src/lib/d3-chart/layers/scatter-points.test.ts
@@ -1,10 +1,15 @@
 // @vitest-environment jsdom
 import * as d3 from 'd3';
-import { describe, expect, it } from 'vitest';
+import { describe, expect, it, vi } from 'vitest';
 
 import type { ShapeKey } from '@/lib/chart-rendering';
 
-import { computeTooltipPosition, renderScatterPoints, syncPointShape } from './scatter-points';
+import {
+  computeTooltipPosition,
+  invalidateTooltipGeometry,
+  renderScatterPoints,
+  syncPointShape,
+} from './scatter-points';
 
 interface TestPoint {
   hwKey: string;
@@ -165,6 +170,78 @@ describe('syncPointShape', () => {
 });
 
 describe('computeTooltipPosition', () => {
+  it('caches same-frame geometry and invalidates it by frame and content lifecycle', () => {
+    let nextFrame: FrameRequestCallback | undefined;
+    vi.stubGlobal(
+      'requestAnimationFrame',
+      vi.fn((callback: FrameRequestCallback) => {
+        nextFrame = callback;
+        return 1;
+      }),
+    );
+
+    const tooltipNode = document.createElement('div');
+    const tooltipBounds = vi.fn(() => ({
+      width: 100,
+      height: 80,
+      left: 0,
+      top: 0,
+      right: 100,
+      bottom: 80,
+      x: 0,
+      y: 0,
+      toJSON: () => ({}),
+    }));
+    Object.defineProperty(tooltipNode, 'getBoundingClientRect', { value: tooltipBounds });
+
+    const container = document.createElement('div');
+    const containerBounds = vi.fn(() => ({
+      width: 800,
+      height: 600,
+      left: 100,
+      top: 50,
+      right: 900,
+      bottom: 650,
+      x: 100,
+      y: 50,
+      toJSON: () => ({}),
+    }));
+    const clientWidthRead = vi.fn(() => 800);
+    const clientHeightRead = vi.fn(() => 600);
+    Object.defineProperties(container, {
+      clientWidth: { get: clientWidthRead },
+      clientHeight: { get: clientHeightRead },
+      getBoundingClientRect: { value: containerBounds },
+    });
+
+    const tooltip = d3.select(tooltipNode);
+    computeTooltipPosition(20, 30, tooltip, container);
+    tooltipNode.style.left = '120px';
+    computeTooltipPosition(25, 35, tooltip, container);
+
+    expect(containerBounds).toHaveBeenCalledTimes(1);
+    expect(clientWidthRead).toHaveBeenCalledTimes(1);
+    expect(clientHeightRead).toHaveBeenCalledTimes(1);
+    expect(tooltipBounds).toHaveBeenCalledTimes(1);
+
+    nextFrame!(0);
+    computeTooltipPosition(30, 40, tooltip, container);
+    expect(containerBounds).toHaveBeenCalledTimes(2);
+    expect(tooltipBounds).toHaveBeenCalledTimes(2);
+
+    tooltipNode.innerHTML = '<strong>updated content</strong>';
+    invalidateTooltipGeometry(tooltipNode);
+    computeTooltipPosition(35, 45, tooltip, container);
+    expect(containerBounds).toHaveBeenCalledTimes(2);
+    expect(tooltipBounds).toHaveBeenCalledTimes(3);
+
+    tooltipNode.style.width = '200px';
+    computeTooltipPosition(40, 50, tooltip, container);
+    expect(tooltipBounds).toHaveBeenCalledTimes(4);
+    nextFrame!(16);
+    vi.unstubAllGlobals();
+  });
+
   it('keeps a tall pinned tooltip inside the visible viewport', () => {
     const tooltipNode = document.createElement('div');
     document.body.append(tooltipNode);
diff --git a/packages/app/src/lib/d3-chart/layers/scatter-points.ts b/packages/app/src/lib/d3-chart/layers/scatter-points.ts
index cf5516be6..d2061cc3b 100644
--- a/packages/app/src/lib/d3-chart/layers/scatter-points.ts
+++ b/packages/app/src/lib/d3-chart/layers/scatter-points.ts
@@ -6,10 +6,10 @@ import {
   getShapeConfig,
   getShapeKeyForPrecision,
   applyNormalState,
-  applyHoverState,
 } from '@/lib/chart-rendering';
 
 import type { ContinuousScale } from '../types';
+import { CHART_TYPE, px } from '../typography';
 
 export interface ScatterPointConfig<T> {
   getColor: (d: T) => string;
@@ -107,35 +107,6 @@ export function renderScatterPoints<T extends { precision: string; x: number; y:
   // Visible shape is created (or swapped, if selectedPrecisions changed) in the
   // merged update pass below.
 
-  // Label (enter only). Multi-line labels are passed as `\n`-separated strings;
-  // we anchor the entire stack via the FIRST tspan's `dy` so getBBox() doesn't
-  // pick up the text element's own (unused) y=0 origin. The first tspan is
-  // raised so the LAST line baseline lands ~8px above the point; subsequent
-  // tspans cascade down by 1.1em.
-  if (!config.hideLabels && config.getLabelText && config.foreground) {
-    const labelGetter = config.getLabelText;
-    entered.each(function (d) {
-      const lines = labelGetter(d).split('\n');
-      const text = d3
-        .select(this)
-        .append('text')
-        .attr('class', 'point-label')
-        .attr('text-anchor', 'middle')
-        .attr('fill', config.foreground!)
-        .attr('font-size', '10px')
-        .attr('font-weight', '700')
-        .attr('pointer-events', 'none');
-      const firstDy = -(0.8 + (lines.length - 1) * 1.1);
-      lines.forEach((line, i) => {
-        text
-          .append('tspan')
-          .attr('x', 0)
-          .attr('dy', i === 0 ? `${firstDy}em` : '1.1em')
-          .text(line);
-      });
-    });
-  }
-
   // Exit: remove stale points
   selection.exit().remove();
 
@@ -166,10 +137,9 @@ export function renderScatterPoints<T extends { precision: string; x: number; y:
     syncPointShape(g, shapeKey, config.getColor(d));
   });
 
-  // Update labels: use data join so labels are created/removed properly on toggle.
-  // Anchor the stack via the first tspan (NOT the text dy — that doesn't shift the
-  // bbox cleanly when there are tspan children).
-  if (!config.hideLabels && config.getLabelText && config.foreground) {
+  // Labels stay joined across display toggles. The display phase changes only
+  // visibility, while data and metric phases may update their text.
+  if (config.getLabelText && config.foreground) {
     const labelGetter = config.getLabelText;
     points.each(function (d) {
       const lines = labelGetter(d).split('\n');
@@ -181,7 +151,7 @@ export function renderScatterPoints<T extends { precision: string; x: number; y:
         .attr('class', 'point-label')
         .attr('text-anchor', 'middle')
         .attr('fill', config.foreground!)
-        .attr('font-size', '10px')
+        .attr('font-size', px(CHART_TYPE.dataLabel))
         .attr('font-weight', '700')
         .attr('pointer-events', 'none');
       const firstDy = -(0.8 + (lines.length - 1) * 1.1);
@@ -193,6 +163,10 @@ export function renderScatterPoints<T extends { precision: string; x: number; y:
         .attr('dy', (_l, i) => (i === 0 ? `${firstDy}em` : '1.1em'))
         .text((l) => l);
     });
+    points
+      .selectAll('.point-label')
+      .style('display', config.hideLabels ? 'none' : '')
+      .style('opacity', config.hideLabels ? 0 : 1);
   } else {
     points.selectAll('.point-label').remove();
   }
@@ -201,105 +175,129 @@ export function renderScatterPoints<T extends { precision: string; x: number; y:
 }
 
 /**
- * Attach scatter point tooltip handlers (hover, click, pin).
- * Works for both ScatterGraph and GPUGraph.
+ * Restyle existing scatter marks without rejoining data or rewriting positions.
+ * Used by the display phase for palette, visibility, and label toggles.
  */
-export function attachScatterTooltipHandlers<
-  T extends { precision: string; x: number; y: number; hwKey?: string | number },
+export function updateScatterPointsForDisplay<
+  T extends { precision: string; x: number; y: number },
 >(
-  points: d3.Selection<SVGGElement, T, SVGGElement, unknown>,
-  config: {
-    xScale: ContinuousScale;
-    yScale: ContinuousScale;
-    svgRef: React.RefObject<SVGSVGElement | null> | React.RefObject<SVGSVGElement>;
-    tooltip: d3.Selection<HTMLDivElement | null, unknown, null, undefined>;
-    container: HTMLElement;
-    rulerGroup: d3.Selection<SVGGElement, unknown, null, undefined>;
-    verticalRuler?: d3.Selection<SVGLineElement, unknown, null, undefined>;
-    horizontalRuler?: d3.Selection<SVGLineElement, unknown, null, undefined>;
-    isPinned: () => boolean;
-    pinTooltip: (point: T, isOverlay: boolean) => void;
-    generateTooltipContent: (d: T, pinned: boolean) => string;
-    trackEvent?: (hw: string, x: number, y: number) => void;
-    /** Called after a point is clicked and tooltip is pinned. Use to attach handlers to tooltip buttons. */
-    onPointClick?: (
-      d: T,
-      tooltip: d3.Selection<HTMLDivElement | null, unknown, null, undefined>,
-    ) => void;
-    /** Ref to current scales — when provided, avoids stale-closure bugs after scale recalculation */
-    scalesRef?: React.RefObject<{ xScale: ContinuousScale; yScale: ContinuousScale } | null>;
-    /** Selected precisions, in selection order; controls hover shape key. */
-    selectedPrecisions?: readonly string[];
-  },
+  zoomGroup: d3.Selection<SVGGElement, unknown, null, undefined>,
+  config: ScatterPointConfig<T>,
 ): void {
-  const {
-    xScale,
-    yScale,
-    svgRef,
-    tooltip,
-    container,
-    rulerGroup,
-    verticalRuler,
-    horizontalRuler,
-    isPinned,
-    pinTooltip,
-    generateTooltipContent,
-    trackEvent,
-    onPointClick,
-    scalesRef,
-    selectedPrecisions,
-  } = config;
-
+  const points = zoomGroup.selectAll<SVGGElement, T>('.dot-group');
+  points.each(function (d) {
+    const point = d3.select(this);
+    if (config.getOpacity) point.style('opacity', config.getOpacity(d));
+    if (config.getPointerEvents) point.style('pointer-events', config.getPointerEvents(d));
+    const shapeKey = config.getShapeKey
+      ? config.getShapeKey(d)
+      : resolveShapeKey(d.precision, config.selectedPrecisions);
+    syncPointShape(
+      point as unknown as d3.Selection<SVGGElement, unknown, null, undefined>,
+      shapeKey,
+      config.getColor(d),
+    );
+  });
   points
-    .on('mouseenter', function (_event, d) {
-      if (isPinned()) return;
-      applyHoverState(
-        d3.select(this).select('.visible-shape') as any,
-        resolveShapeKey(d.precision, selectedPrecisions),
-      );
-      tooltip.style('opacity', 1).style('display', 'block').style('pointer-events', 'none');
-      const curXScale = scalesRef?.current?.xScale ?? xScale;
-      const curYScale = scalesRef?.current?.yScale ?? yScale;
-      const ct = d3.zoomTransform(svgRef.current!);
-      rulerGroup.style('display', 'block');
-      verticalRuler
-        ?.attr('x1', ct.rescaleX(curXScale)(d.x))
-        .attr('x2', ct.rescaleX(curXScale)(d.x));
-      horizontalRuler
-        ?.attr('y1', ct.rescaleY(curYScale)(d.y))
-        .attr('y2', ct.rescaleY(curYScale)(d.y));
-      tooltip.html(generateTooltipContent(d, false));
-    })
-    .on('mousemove', (event) => {
-      if (isPinned()) return;
-      const [mx, my] = d3.pointer(event, container);
-      const pos = computeTooltipPosition(mx, my, tooltip, container);
-      tooltip.style('left', `${pos.left}px`).style('top', `${pos.top}px`);
-    })
-    .on('mouseleave', function (_event, d) {
-      if (isPinned()) return;
-      applyNormalState(
-        d3.select(this).select('.visible-shape') as any,
-        resolveShapeKey(d.precision, selectedPrecisions),
-      );
-      tooltip.style('opacity', 0).style('display', 'none');
-      rulerGroup.style('display', 'none');
-    })
-    .on('click', (event, d) => {
-      event.stopPropagation();
-      const [mx, my] = d3.pointer(event, container);
-      tooltip.html(generateTooltipContent(d, true));
-      const pos = computeTooltipPosition(mx, my, tooltip, container);
-      tooltip
-        .style('left', `${pos.left}px`)
-        .style('top', `${pos.top}px`)
-        .style('opacity', 1)
-        .style('display', 'block')
-        .style('pointer-events', 'auto');
-      pinTooltip(d, false);
-      onPointClick?.(d, tooltip);
-      trackEvent?.(String(d.hwKey), d.x, d.y);
-    });
+    .selectAll('.point-label')
+    .attr('fill', config.foreground ?? null)
+    .style('display', config.hideLabels ? 'none' : '')
+    .style('opacity', config.hideLabels ? 0 : 1);
+}
+
+export interface TooltipContainerGeometry {
+  bounds: DOMRect;
+  width: number;
+  height: number;
+}
+
+interface TooltipGeometry {
+  width: number;
+  height: number;
+  sizeStyle: string;
+  className: string | null;
+}
+
+let containerGeometryCache = new WeakMap<HTMLElement, TooltipContainerGeometry>();
+let tooltipGeometryCache = new WeakMap<HTMLDivElement, TooltipGeometry>();
+let geometryResetPending = false;
+
+function scheduleGeometryReset(): void {
+  if (geometryResetPending) return;
+  geometryResetPending = true;
+  const reset = () => {
+    containerGeometryCache = new WeakMap();
+    tooltipGeometryCache = new WeakMap();
+    geometryResetPending = false;
+  };
+  if (typeof requestAnimationFrame === 'function') requestAnimationFrame(reset);
+  else setTimeout(reset, 0);
+}
+
+/**
+ * Invalidate dimensions after synchronously replacing tooltip content.
+ * Style/class changes are detected from their non-layout attributes; callers
+ * use this helper when content and positioning happen in one event.
+ */
+export function invalidateTooltipGeometry(tooltip: HTMLDivElement | null): void {
+  if (tooltip) tooltipGeometryCache.delete(tooltip);
+}
+
+function tooltipSizeStyleSignature(tooltip: HTMLDivElement): string {
+  const { style } = tooltip;
+  return [
+    style.width,
+    style.height,
+    style.minWidth,
+    style.minHeight,
+    style.maxWidth,
+    style.maxHeight,
+    style.padding,
+    style.borderWidth,
+    style.boxSizing,
+    style.font,
+    style.fontSize,
+    style.fontWeight,
+    style.lineHeight,
+    style.letterSpacing,
+    style.whiteSpace,
+  ].join('|');
+}
+
+function getTooltipGeometry(tooltip: HTMLDivElement): TooltipGeometry {
+  const sizeStyle = tooltipSizeStyleSignature(tooltip);
+  const className = tooltip.getAttribute('class');
+  const cached = tooltipGeometryCache.get(tooltip);
+  if (cached && cached.sizeStyle === sizeStyle && cached.className === className) return cached;
+  const bounds = tooltip.getBoundingClientRect();
+  const geometry = {
+    width: bounds.width || tooltip.offsetWidth,
+    height: bounds.height || tooltip.offsetHeight,
+    sizeStyle,
+    className,
+  };
+  tooltipGeometryCache.set(tooltip, geometry);
+  scheduleGeometryReset();
+  return geometry;
+}
+
+/**
+ * Return container geometry shared by every tooltip position calculation in
+ * the current animation frame. Pointer events can fire more often than paint;
+ * coalescing these layout reads avoids measuring the same chart repeatedly.
+ */
+export function getTooltipContainerGeometry(container: HTMLElement): TooltipContainerGeometry {
+  const cached = containerGeometryCache.get(container);
+  if (cached) return cached;
+
+  const geometry = {
+    bounds: container.getBoundingClientRect(),
+    width: container.clientWidth,
+    height: container.clientHeight,
+  };
+  containerGeometryCache.set(container, geometry);
+  scheduleGeometryReset();
+  return geometry;
 }
 
 /**
@@ -326,19 +324,25 @@ export function computeTooltipPosition(
     | d3.Selection<HTMLDivElement, unknown, null, undefined>,
   container: HTMLElement,
   offset = 10,
+  containerGeometry?: TooltipContainerGeometry,
 ): { left: number; top: number } {
   const node = tooltip.node();
   if (!node) return { left: mx + offset, top: my + offset };
 
-  // Ensure tooltip is measurable
-  node.style.display = 'block';
+  // Ensure the tooltip is measurable. Moving from hidden to shown changes its
+  // size lifecycle, so discard any dimensions measured before that style write.
+  if (node.style.display !== 'block') {
+    node.style.display = 'block';
+    invalidateTooltipGeometry(node);
+  }
 
-  // Force reflow so we get real dimensions
-  const tw = node.getBoundingClientRect().width || node.offsetWidth;
-  const th = node.getBoundingClientRect().height || node.offsetHeight;
-  const rect = container.getBoundingClientRect();
-  const cw = container.clientWidth;
-  const ch = container.clientHeight;
+  const tooltipGeometry = getTooltipGeometry(node);
+  const tw = tooltipGeometry.width;
+  const th = tooltipGeometry.height;
+  const geometry = containerGeometry ?? getTooltipContainerGeometry(container);
+  const rect = geometry.bounds;
+  const cw = geometry.width;
+  const ch = geometry.height;
   const EDGE_PAD = 4;
 
   // Prefer right of cursor; flip to left if no room.
diff --git a/packages/app/src/lib/d3-chart/layers/test-helpers.ts b/packages/app/src/lib/d3-chart/layers/test-helpers.ts
index c295d7900..86dfe36f1 100644
--- a/packages/app/src/lib/d3-chart/layers/test-helpers.ts
+++ b/packages/app/src/lib/d3-chart/layers/test-helpers.ts
@@ -5,7 +5,8 @@
  * selection that tracks attribute, style, and text calls. The mock supports
  * the full chaining API used by the layer modules: selectAll, data, join,
  * enter, append, exit, remove, merge, select, attr, style, text, each, datum,
- * empty, nodes, transition, duration.
+ * empty, nodes, transition, duration, on (handlers are recorded on the
+ * element, not dispatched — tests can invoke them directly).
  */
 
 // ─── Mock Element ────────────────────────────────────────────────────
@@ -19,6 +20,8 @@ export interface MockElement {
   children: MockElement[];
   datum: unknown;
   removed: boolean;
+  /** Event handlers recorded by `.on()`, keyed by event type. */
+  handlers?: Record<string, (...args: unknown[]) => void>;
 }
 
 function createMockElement(tag: string, datum?: unknown): MockElement {
@@ -72,6 +75,7 @@ export interface MockSelection<Datum = unknown> {
   size: () => number;
   transition: () => MockSelection<Datum>;
   duration: (ms: number) => MockSelection<Datum>;
+  on: (type: string, listener: (...args: unknown[]) => void) => MockSelection<Datum>;
 }
 
 // ─── Entry point ─────────────────────────────────────────────────────
@@ -183,7 +187,9 @@ function makeSel<D>(
       const store = getChildStore(parentEl);
       const existing = store.get(selectorStr)?.elements.filter((e) => !e.removed) ?? [];
 
-      const bound = makeSel(existing, newData, parentEl, selectorStr);
+      // The UPDATE selection holds only elements with a datum (by index,
+      // like d3) — surplus elements belong to exit(), not update/merge.
+      const bound = makeSel(existing.slice(0, newData.length), newData, parentEl, selectorStr);
 
       // Override enter/exit/join on the data-bound selection.
       bound.enter = () => {
@@ -302,6 +308,17 @@ function makeSel<D>(
       );
     },
 
+    // ── on ──
+    // Records the handler on each element (keyed by event type) without
+    // dispatching — tests can look it up and invoke it directly.
+    on(type: string, listener: (...args: unknown[]) => void): MockSelection<D> {
+      for (const el of elements) {
+        el.handlers ??= {};
+        el.handlers[type] = listener;
+      }
+      return sel;
+    },
+
     // ── Utility ──
     empty: () => elements.length === 0,
     nodes: () => elements,
@@ -364,7 +381,8 @@ function findChildBySelector(el: MockElement, selector: string): MockElement | u
     }
   }
 
-  // Fall back to scanning children by class attribute
+  // Fall back to scanning children by class attribute (breadth-first:
+  // direct children win over nested descendants, like d3's `select`).
   for (const child of el.children) {
     if (child.removed) continue;
     const classes = String(child.attrs['class'] || '').split(/\s+/u);
@@ -372,5 +390,10 @@ function findChildBySelector(el: MockElement, selector: string): MockElement | u
       return child;
     }
   }
+  for (const child of el.children) {
+    if (child.removed) continue;
+    const nested = findChildBySelector(child, selector);
+    if (nested) return nested;
+  }
   return undefined;
 }
diff --git a/packages/app/src/lib/d3-chart/overlay-x-marker.test.ts b/packages/app/src/lib/d3-chart/overlay-x-marker.test.ts
new file mode 100644
index 000000000..bae714afe
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/overlay-x-marker.test.ts
@@ -0,0 +1,24 @@
+// @vitest-environment jsdom
+
+import { describe, expect, it } from 'vitest';
+
+import { overlayMarkerPosition } from './overlay-x-marker';
+
+describe('overlayMarkerPosition', () => {
+  it.each([
+    ['translate(12.5,34)', { x: 12.5, y: 34 }],
+    ['translate(-8 9.25)', { x: -8, y: 9.25 }],
+  ])('reads the marker position from %s', (transform, expected) => {
+    const marker = document.createElementNS('http://www.w3.org/2000/svg', 'g');
+    marker.setAttribute('transform', transform);
+
+    expect(overlayMarkerPosition(marker)).toEqual(expected);
+  });
+
+  it('returns null for missing or non-numeric translations', () => {
+    const marker = document.createElementNS('http://www.w3.org/2000/svg', 'g');
+    expect(overlayMarkerPosition(marker)).toBeNull();
+    marker.setAttribute('transform', 'translate(nope,10)');
+    expect(overlayMarkerPosition(marker)).toBeNull();
+  });
+});
diff --git a/packages/app/src/lib/d3-chart/overlay-x-marker.ts b/packages/app/src/lib/d3-chart/overlay-x-marker.ts
new file mode 100644
index 000000000..81b337811
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/overlay-x-marker.ts
@@ -0,0 +1,94 @@
+import * as d3 from 'd3';
+import { invalidateTooltipGeometry } from '@/lib/d3-chart/layers/scatter-points';
+
+export interface OverlayTooltipHandle<T> {
+  isPinned: () => boolean;
+  pinTooltip: (data: T, isOverlay: boolean) => void;
+}
+
+export interface OverlayRulerLifecycle<T> {
+  show: (data: T, marker: SVGGElement) => void;
+  hide: () => void;
+}
+
+export interface OverlayXMarkerOptions<T> {
+  markerSelector: string;
+  normalPath: string;
+  hoverPath: string;
+  normalStrokeWidth?: number;
+  hoverStrokeWidth?: number;
+  tooltip: d3.Selection<HTMLDivElement, unknown, null, undefined>;
+  handle: OverlayTooltipHandle<T> | null;
+  content: (data: T, pinned: boolean) => string;
+  position: (event: PointerEvent | MouseEvent) => { left: number; top: number };
+  rulers?: OverlayRulerLifecycle<T>;
+  onClick?: (data: T) => void;
+}
+
+export function xMarkerPath(size: number, armScale = 1): string {
+  const arm = size * armScale;
+  return `M ${-arm} ${-arm} L ${arm} ${arm} M ${-arm} ${arm} L ${arm} ${-arm}`;
+}
+
+/** Current rendered translation of an overlay marker group. */
+export function overlayMarkerPosition(marker: Element): { x: number; y: number } | null {
+  const transform = marker.getAttribute('transform') ?? '';
+  const match = transform.match(/translate\((?<x>[^, ]+)[, ]+(?<y>[^) ]+)\)/u);
+  const x = Number(match?.groups?.x);
+  const y = Number(match?.groups?.y);
+  return Number.isFinite(x) && Number.isFinite(y) ? { x, y } : null;
+}
+
+export function attachOverlayXMarkerHandlers<T>(
+  points: d3.Selection<SVGGElement, T, SVGGElement, unknown>,
+  options: OverlayXMarkerOptions<T>,
+): void {
+  const normalStrokeWidth = options.normalStrokeWidth ?? 2.5;
+  const hoverStrokeWidth = options.hoverStrokeWidth ?? 3.5;
+  const positionTooltip = (event: PointerEvent | MouseEvent) => {
+    const position = options.position(event);
+    options.tooltip.style('left', `${position.left}px`).style('top', `${position.top}px`);
+  };
+
+  points
+    .on('mouseenter', function (_event, data) {
+      if (options.handle?.isPinned()) return;
+      d3.select(this)
+        .select(options.markerSelector)
+        .attr('d', options.hoverPath)
+        .attr('stroke-width', hoverStrokeWidth);
+      options.tooltip
+        .style('opacity', 1)
+        .style('display', 'block')
+        .style('pointer-events', 'none')
+        .html(options.content(data, false));
+      invalidateTooltipGeometry(options.tooltip.node());
+      options.rulers?.show(data, this);
+    })
+    .on('mousemove', (event) => {
+      if (options.handle?.isPinned()) return;
+      positionTooltip(event);
+    })
+    .on('mouseleave', function () {
+      if (options.handle?.isPinned()) return;
+      d3.select(this)
+        .select(options.markerSelector)
+        .attr('d', options.normalPath)
+        .attr('stroke-width', normalStrokeWidth);
+      options.tooltip.style('opacity', 0).style('display', 'none');
+      options.rulers?.hide();
+    })
+    .on('click', function (event, data) {
+      event.stopPropagation();
+      options.tooltip
+        .html(options.content(data, true))
+        .style('opacity', 1)
+        .style('display', 'block')
+        .style('pointer-events', 'auto');
+      invalidateTooltipGeometry(options.tooltip.node());
+      positionTooltip(event);
+      options.rulers?.show(data, this);
+      options.handle?.pinTooltip(data, true);
+      options.onClick?.(data);
+    });
+}
diff --git a/packages/app/src/lib/d3-chart/plot-bounds.test.ts b/packages/app/src/lib/d3-chart/plot-bounds.test.ts
new file mode 100644
index 000000000..ef381a192
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/plot-bounds.test.ts
@@ -0,0 +1,143 @@
+// @vitest-environment jsdom
+import { afterEach, describe, expect, it } from 'vitest';
+
+import { plotBounds } from './plot-bounds';
+
+const SVG_NS = 'http://www.w3.org/2000/svg';
+
+function svgEl<K extends string>(tag: K): Element {
+  return document.createElementNS(SVG_NS, tag);
+}
+
+function stubBox(element: Element, left: number, top: number, width: number, height: number) {
+  element.getBoundingClientRect = () =>
+    ({
+      left,
+      top,
+      width,
+      height,
+      right: left + width,
+      bottom: top + height,
+      x: left,
+      y: top,
+      toJSON: () => ({}),
+    }) as DOMRect;
+}
+
+/**
+ * The skeleton `setupChart` builds: a root group translated by the margins,
+ * and a clip rect sized to the plot area inside them.
+ */
+const CLIP_ID = 'clip-scatter-graph';
+
+function renderChartSvg({
+  box = { left: 100, top: 200, width: 800, height: 600 },
+  transform = 'translate(60,24)' as string | null,
+  clip = { width: 730, height: 516 } as { width: number; height: number } | null,
+  /** A same-SVG clipPath belonging to something else (overflow continuations). */
+  decoyClip = false,
+} = {}): Element {
+  const svg = svgEl('svg');
+  stubBox(svg, box.left, box.top, box.width, box.height);
+
+  const defs = svgEl('defs');
+  if (decoyClip) {
+    const other = svgEl('clipPath');
+    other.setAttribute('id', 'some-other-clip');
+    const otherRect = svgEl('rect');
+    otherRect.setAttribute('width', '1');
+    otherRect.setAttribute('height', '1');
+    other.append(otherRect);
+    defs.append(other);
+  }
+  if (clip) {
+    const clipPath = svgEl('clipPath');
+    clipPath.setAttribute('id', CLIP_ID);
+    const rect = svgEl('rect');
+    rect.setAttribute('width', String(clip.width));
+    rect.setAttribute('height', String(clip.height));
+    clipPath.append(rect);
+    defs.append(clipPath);
+  }
+  svg.append(defs);
+
+  const root = svgEl('g');
+  root.setAttribute('class', 'chart-root');
+  if (transform !== null) root.setAttribute('transform', transform);
+  const zoomGroup = svgEl('g');
+  zoomGroup.setAttribute('class', 'zoom-group');
+  if (clip) zoomGroup.setAttribute('clip-path', `url(#${CLIP_ID})`);
+  root.append(zoomGroup);
+  svg.append(root);
+
+  document.body.append(svg);
+  return svg;
+}
+
+afterEach(() => {
+  document.body.innerHTML = '';
+});
+
+describe('plotBounds', () => {
+  it('insets the SVG box by the chart margins', () => {
+    // 60px left / 24px top margins are the inference scatter's axis gutters.
+    const svg = renderChartSvg();
+
+    expect(plotBounds(svg)).toEqual({
+      left: 160,
+      top: 224,
+      right: 890,
+      bottom: 740,
+    });
+  });
+
+  it('is strictly smaller than the SVG box it is derived from', () => {
+    const svg = renderChartSvg();
+    const box = svg.getBoundingClientRect();
+    const bounds = plotBounds(svg)!;
+
+    expect(bounds.left).toBeGreaterThan(box.left);
+    expect(bounds.top).toBeGreaterThan(box.top);
+    expect(bounds.right).toBeLessThan(box.right);
+    expect(bounds.bottom).toBeLessThan(box.bottom);
+  });
+
+  it('accepts a space-separated translate', () => {
+    const svg = renderChartSvg({ transform: 'translate(60 24)' });
+    expect(plotBounds(svg)?.left).toBe(160);
+  });
+
+  it("resolves the chart's own clipPath, not whichever one comes first", () => {
+    // The overflow-continuation layer defines a clipPath per group, so
+    // "the first clipPath in the SVG" is not a safe assumption.
+    const svg = renderChartSvg({ decoyClip: true });
+    expect(plotBounds(svg)).toEqual({ left: 160, top: 224, right: 890, bottom: 740 });
+  });
+
+  it('returns null when the chart clips nothing', () => {
+    // clipContent: false — no clip-path on the zoom group, so nothing is
+    // painted away and the caller should fall back to the SVG box.
+    expect(plotBounds(renderChartSvg({ clip: null }))).toBeNull();
+  });
+
+  it('returns null when the referenced clipPath is missing', () => {
+    const svg = renderChartSvg();
+    svg.querySelector('clipPath')!.remove();
+    expect(plotBounds(svg)).toBeNull();
+  });
+
+  it('returns null before the skeleton has rendered', () => {
+    const svg = svgEl('svg');
+    stubBox(svg, 0, 0, 800, 600);
+    expect(plotBounds(svg)).toBeNull();
+  });
+
+  it('returns null for a missing or unreadable transform', () => {
+    expect(plotBounds(renderChartSvg({ transform: null }))).toBeNull();
+    expect(plotBounds(renderChartSvg({ transform: 'scale(2)' }))).toBeNull();
+  });
+
+  it('returns null for a degenerate clip rect', () => {
+    expect(plotBounds(renderChartSvg({ clip: { width: 0, height: 516 } }))).toBeNull();
+  });
+});
diff --git a/packages/app/src/lib/d3-chart/plot-bounds.ts b/packages/app/src/lib/d3-chart/plot-bounds.ts
new file mode 100644
index 000000000..bff8e25e0
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/plot-bounds.ts
@@ -0,0 +1,65 @@
+/**
+ * Viewport-space bounds of a chart's clip region.
+ *
+ * `setupChart` clips the zoom group to a rect covering the plot area only:
+ * `(0, 0, width, height)` inside `.chart-root`, which is itself translated by
+ * the chart margins. On the inference scatter those margins are 60px on the
+ * left and bottom — the axis gutters — so the SVG's own box is a sizeable
+ * superset of what the user can actually see.
+ *
+ * A zoom pushes points out of the plot, where the clip path paints them away
+ * while `getBoundingClientRect()` keeps reporting perfectly ordinary geometry.
+ * Anything asking "can the user see this point?" therefore has to compare
+ * against this rect, never against the SVG box.
+ *
+ * Deliberately free of d3 so callers outside the chart bundle (the anchored
+ * nudge) can import it without pulling d3 onto pages that have no chart.
+ */
+
+export interface PlotBounds {
+  left: number;
+  top: number;
+  right: number;
+  bottom: number;
+}
+
+/** `translate(x,y)` as written by `setupChart` — also tolerates space separators. */
+const TRANSLATE = /translate\(\s*(?<x>-?[\d.]+)[\s,]+(?<y>-?[\d.]+)\s*\)/u;
+/** The `clip-path="url(#…)"` reference `setupChart` puts on the zoom group. */
+const CLIP_URL = /url\(\s*#(?<id>[^)\s]+)\s*\)/u;
+
+/**
+ * @param svg the chart's `<svg>` element (`[data-testid="d3-chart-svg"]`).
+ * @returns the clip region in viewport coordinates, or `null` when the chart
+ *   does not clip its content (`clipContent: false`) or has not rendered its
+ *   skeleton yet. A `null` result means "nothing is being clipped away", so
+ *   callers should fall back to the SVG box rather than reject everything.
+ */
+export function plotBounds(svg: Element): PlotBounds | null {
+  const root = svg.querySelector('.chart-root');
+  const zoomGroup = svg.querySelector('.zoom-group');
+  if (!root || !zoomGroup) return null;
+
+  // Follow the zoom group's own reference rather than assuming the chart's is
+  // the only clipPath in the SVG — the overflow-continuation layer defines one
+  // per group, and future layers may too. No reference means `clipContent:
+  // false`, i.e. nothing is being clipped away.
+  const clipId = CLIP_URL.exec(zoomGroup.getAttribute('clip-path') ?? '')?.groups?.id;
+  if (!clipId) return null;
+  const clip = [...svg.querySelectorAll('clipPath')]
+    .find((candidate) => candidate.id === clipId)
+    ?.querySelector('rect');
+  if (!clip) return null;
+
+  const translate = TRANSLATE.exec(root.getAttribute('transform') ?? '')?.groups;
+  const width = Number(clip.getAttribute('width'));
+  const height = Number(clip.getAttribute('height'));
+  if (!translate || !(width > 0) || !(height > 0)) return null;
+
+  // No viewBox on the chart SVG, so one user unit is one CSS pixel and the
+  // margins can be added to the client rect directly.
+  const box = svg.getBoundingClientRect();
+  const left = box.left + Number(translate.x);
+  const top = box.top + Number(translate.y);
+  return { left, top, right: left + width, bottom: top + height };
+}
diff --git a/packages/app/src/lib/d3-chart/typography.ts b/packages/app/src/lib/d3-chart/typography.ts
new file mode 100644
index 000000000..a92caa5bd
--- /dev/null
+++ b/packages/app/src/lib/d3-chart/typography.ts
@@ -0,0 +1,43 @@
+/**
+ * Shared chart typography constants.
+ *
+ * Chart text is sized in TypeScript, not CSS custom properties, on purpose:
+ * the PNG export path (useChartExport) serializes the chart subtree with
+ * html-to-image, which cannot resolve `var(--*)` references, and its
+ * resolveCssVarsForExport() only bakes color-type presentation attributes
+ * (fill/stroke/etc.), not font-size. A `font-size: var(--…)` would silently
+ * collapse in exported images. Keep sizes here and interpolate px strings.
+ *
+ * OG-image renderers (Satori: opengraph-image.tsx, og-image-render.tsx,
+ * compare-og.tsx) are a separate fixed-canvas world and intentionally do not
+ * use these constants.
+ */
+export const CHART_TYPE = {
+  /** Axis titles (x-axis-label / y-axis-label) and axis tick primary rows. */
+  axisLabel: 12,
+  /** Secondary (muted) row of two-row axis labels. */
+  axisLabelSub: 10,
+  /** In-plot annotation text, first line. */
+  annotation: 11,
+  /** In-plot annotation text, secondary line. */
+  annotationSub: 10,
+  /** Small labels attached to marks (points, radar vertices, rulers). */
+  dataLabel: 10,
+  /** Smallest legible size, for dense diagram labels. */
+  micro: 9,
+  /** Unofficial-results watermark pattern text. */
+  watermark: 20,
+} as const;
+
+/** Format a CHART_TYPE size as a CSS px string for .attr('font-size', …). */
+export function px(size: number): string {
+  return `${size}px`;
+}
+
+/**
+ * Font stacks used when charts are rendered to a PNG. Shared with the export
+ * path so chart layers and useChartExport cannot drift apart.
+ */
+export const CHART_FONT_SANS =
+  'var(--font-dm-sans), -apple-system, BlinkMacSystemFont, "Segoe UI", sans-serif';
+export const CHART_FONT_MINECRAFT = 'var(--font-minecraft), "Monocraft", monospace';
diff --git a/packages/app/src/lib/d3-chart/watermark.ts b/packages/app/src/lib/d3-chart/watermark.ts
index 942006251..e7d5f5f7b 100644
--- a/packages/app/src/lib/d3-chart/watermark.ts
+++ b/packages/app/src/lib/d3-chart/watermark.ts
@@ -1,6 +1,7 @@
 import type * as d3 from 'd3';
 
 import type { ChartMargin } from './types';
+import { CHART_TYPE, px } from './typography';
 
 const UNOFFICIAL_WATERMARK_IMAGE_PATH = '/decorative/kanye-west.png';
 
@@ -97,7 +98,7 @@ export function createUnofficialWatermark(
       .attr('text-anchor', 'middle')
       .attr('dominant-baseline', 'middle')
       .attr('fill', '#dc2626')
-      .attr('font-size', '20px')
+      .attr('font-size', px(CHART_TYPE.watermark))
       .attr('font-weight', 'bold')
       .attr('opacity', 0.25);
 
diff --git a/packages/app/src/lib/dashboard-routes.test.ts b/packages/app/src/lib/dashboard-routes.test.ts
new file mode 100644
index 000000000..6302db740
--- /dev/null
+++ b/packages/app/src/lib/dashboard-routes.test.ts
@@ -0,0 +1,92 @@
+import { existsSync } from 'node:fs';
+
+import { describe, expect, it } from 'vitest';
+
+import {
+  DASHBOARD_ROUTES,
+  dashboardRouteForPathname,
+  dashboardShellCapabilitiesForPathname,
+  getDashboardRoute,
+  isDashboardRouteKey,
+} from './dashboard-routes';
+
+describe('dashboard route registry', () => {
+  it('has unique keys, navigation paths, and sitemap paths', () => {
+    expect(new Set(DASHBOARD_ROUTES.map((route) => route.key)).size).toBe(DASHBOARD_ROUTES.length);
+    expect(new Set(DASHBOARD_ROUTES.map((route) => route.path)).size).toBe(DASHBOARD_ROUTES.length);
+    expect(new Set(DASHBOARD_ROUTES.map((route) => route.canonicalPath)).size).toBe(
+      DASHBOARD_ROUTES.length,
+    );
+  });
+
+  it.each(DASHBOARD_ROUTES)('has English and Chinese page parity for "$key"', (route) => {
+    const englishPage = new URL(`../app/(dashboard)/${route.key}/page.tsx`, import.meta.url);
+    const chinesePage = new URL(`../app/zh/(dashboard)/${route.key}/page.tsx`, import.meta.url);
+
+    expect(existsSync(englishPage)).toBe(true);
+    expect(existsSync(chinesePage)).toBe(route.localeMirrored);
+  });
+
+  it('keeps reliability footer-only and indexable, and feedback out of the sitemap', () => {
+    expect(getDashboardRoute('reliability').navGroup).toBe('footer-only');
+    expect(getDashboardRoute('reliability').indexable).toBe(true);
+    expect(getDashboardRoute('feedback').indexable).toBe(false);
+  });
+
+  it('declares the standalone and self-seeded provider capabilities', () => {
+    expect(getDashboardRoute('collectivex').providers).toEqual({
+      globalFilters: false,
+      unofficialRuns: false,
+    });
+    expect(getDashboardRoute('calculator').providers).toEqual({
+      globalFilters: false,
+      unofficialRuns: true,
+    });
+  });
+
+  it.each(['/inference/agentic', '/zh/inference/agentic'])(
+    'keeps the agentic catalog standalone at %s',
+    (pathname) => {
+      expect(dashboardShellCapabilitiesForPathname(pathname)).toEqual({
+        providers: { globalFilters: false, unofficialRuns: false },
+        dashboardNudge: false,
+      });
+    },
+  );
+
+  it.each(['/inference/agentic/42', '/zh/inference/agentic/42'])(
+    'preserves standalone agentic detail behavior at %s',
+    (pathname) => {
+      expect(dashboardShellCapabilitiesForPathname(pathname)).toEqual({
+        providers: { globalFilters: false, unofficialRuns: false },
+        dashboardNudge: false,
+      });
+    },
+  );
+
+  it.each(['/collectivex', '/zh/collectivex'])(
+    'keeps the CollectiveX route free of dashboard providers and nudges at %s',
+    (pathname) => {
+      expect(dashboardShellCapabilitiesForPathname(pathname)).toEqual({
+        providers: { globalFilters: false, unofficialRuns: false },
+        dashboardNudge: false,
+      });
+    },
+  );
+
+  it.each(['/reliability', '/gpu-specs', '/submissions', '/feedback'])(
+    'keeps dashboard nudges active on provider-free route %s',
+    (pathname) => {
+      expect(dashboardShellCapabilitiesForPathname(pathname).dashboardNudge).toBe(true);
+    },
+  );
+
+  it('resolves English, Chinese, canonical, and dashboard child paths', () => {
+    expect(dashboardRouteForPathname('/evaluation')?.key).toBe('evaluation');
+    expect(dashboardRouteForPathname('/zh/evaluation')?.key).toBe('evaluation');
+    expect(dashboardRouteForPathname('/')?.key).toBe('inference');
+    expect(dashboardRouteForPathname('/zh/inference/agentic/42')?.key).toBe('inference');
+    expect(dashboardRouteForPathname('/compare/foo-vs-bar')).toBeUndefined();
+    expect(isDashboardRouteKey('not-a-tab')).toBe(false);
+  });
+});
diff --git a/packages/app/src/lib/dashboard-routes.ts b/packages/app/src/lib/dashboard-routes.ts
new file mode 100644
index 000000000..1ef6cb82e
--- /dev/null
+++ b/packages/app/src/lib/dashboard-routes.ts
@@ -0,0 +1,265 @@
+export type DashboardNavGroup = 'primary' | 'feature-gated' | 'footer-only';
+export type DashboardShareParamScope = 'g_' | 'i_' | 'e_' | 'r_' | 'c_';
+
+export interface DashboardProviderCapabilities {
+  readonly globalFilters: boolean;
+  readonly unofficialRuns: boolean;
+}
+
+export interface DashboardShellCapabilities {
+  readonly providers: DashboardProviderCapabilities;
+  readonly dashboardNudge: boolean;
+}
+
+interface DashboardRouteDefinition {
+  readonly key: string;
+  /** Route used for dashboard navigation. */
+  readonly path: `/${string}`;
+  /** SEO canonical and sitemap route. The inference dashboard canonicalizes to `/`. */
+  readonly canonicalPath: '/' | `/${string}`;
+  readonly navGroup: DashboardNavGroup;
+  readonly indexable: boolean;
+  readonly localeMirrored: boolean;
+  readonly providers: DashboardProviderCapabilities;
+  readonly shareParamScopes: readonly DashboardShareParamScope[];
+}
+
+const FILTERED_DASHBOARD_PROVIDERS = {
+  globalFilters: true,
+  unofficialRuns: true,
+} as const;
+
+const UNOFFICIAL_ONLY_DASHBOARD_PROVIDERS = {
+  globalFilters: false,
+  unofficialRuns: true,
+} as const;
+
+const STANDALONE_DASHBOARD_PROVIDERS = {
+  globalFilters: false,
+  unofficialRuns: false,
+} as const;
+
+interface DashboardShellCapabilityRoute extends DashboardShellCapabilities {
+  readonly path: `/${string}`;
+  readonly includeChildren: boolean;
+}
+
+/**
+ * Shell-only route overrides. These live beside the dashboard registry so
+ * provider and nudge decisions are data-driven rather than inferred in the
+ * React shell. Locale-prefixed paths are normalized by the resolver.
+ */
+export const DASHBOARD_SHELL_CAPABILITY_ROUTES = [
+  {
+    path: '/inference/agentic',
+    includeChildren: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    dashboardNudge: false,
+  },
+  {
+    path: '/collectivex',
+    includeChildren: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    dashboardNudge: false,
+  },
+] as const satisfies readonly DashboardShellCapabilityRoute[];
+
+/**
+ * Canonical, data-only registry for every dashboard route.
+ *
+ * Localized labels and metadata prose deliberately live in `tab-meta*.ts` and
+ * `tab-nav.tsx`; this module can be consumed by server and client code without
+ * importing React or locale content.
+ */
+export const DASHBOARD_ROUTES = [
+  {
+    key: 'inference',
+    path: '/inference',
+    canonicalPath: '/',
+    navGroup: 'primary',
+    indexable: true,
+    localeMirrored: true,
+    providers: FILTERED_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'evaluation',
+    path: '/evaluation',
+    canonicalPath: '/evaluation',
+    navGroup: 'primary',
+    indexable: true,
+    localeMirrored: true,
+    providers: FILTERED_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'e_'],
+  },
+  {
+    key: 'historical',
+    path: '/historical',
+    canonicalPath: '/historical',
+    navGroup: 'primary',
+    indexable: true,
+    localeMirrored: true,
+    providers: FILTERED_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'calculator',
+    path: '/calculator',
+    canonicalPath: '/calculator',
+    navGroup: 'primary',
+    indexable: true,
+    localeMirrored: true,
+    providers: UNOFFICIAL_ONLY_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_', 'c_'],
+  },
+  {
+    key: 'fleet',
+    path: '/fleet',
+    canonicalPath: '/fleet',
+    navGroup: 'primary',
+    indexable: true,
+    localeMirrored: true,
+    providers: UNOFFICIAL_ONLY_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_', 'c_'],
+  },
+  {
+    key: 'reliability',
+    path: '/reliability',
+    canonicalPath: '/reliability',
+    navGroup: 'footer-only',
+    indexable: true,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['r_'],
+  },
+  {
+    key: 'gpu-specs',
+    path: '/gpu-specs',
+    canonicalPath: '/gpu-specs',
+    navGroup: 'primary',
+    indexable: true,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'submissions',
+    path: '/submissions',
+    canonicalPath: '/submissions',
+    navGroup: 'primary',
+    indexable: true,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'collectivex',
+    path: '/collectivex',
+    canonicalPath: '/collectivex',
+    navGroup: 'feature-gated',
+    indexable: true,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'ai-chart',
+    path: '/ai-chart',
+    canonicalPath: '/ai-chart',
+    navGroup: 'feature-gated',
+    indexable: true,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'gpu-metrics',
+    path: '/gpu-metrics',
+    canonicalPath: '/gpu-metrics',
+    navGroup: 'feature-gated',
+    indexable: true,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'current-inferencex-image',
+    path: '/current-inferencex-image',
+    canonicalPath: '/current-inferencex-image',
+    navGroup: 'feature-gated',
+    indexable: true,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+  {
+    key: 'feedback',
+    path: '/feedback',
+    canonicalPath: '/feedback',
+    navGroup: 'feature-gated',
+    indexable: false,
+    localeMirrored: true,
+    providers: STANDALONE_DASHBOARD_PROVIDERS,
+    shareParamScopes: ['g_', 'i_'],
+  },
+] as const satisfies readonly DashboardRouteDefinition[];
+
+export type DashboardRoute = (typeof DASHBOARD_ROUTES)[number];
+export type DashboardRouteKey = DashboardRoute['key'];
+
+export const DASHBOARD_ROUTE_KEYS: readonly DashboardRouteKey[] = DASHBOARD_ROUTES.map(
+  (route) => route.key,
+);
+
+export function isDashboardRouteKey(value: string): value is DashboardRouteKey {
+  return DASHBOARD_ROUTE_KEYS.includes(value as DashboardRouteKey);
+}
+
+export function getDashboardRoute(key: DashboardRouteKey): DashboardRoute {
+  return DASHBOARD_ROUTES.find((route) => route.key === key)!;
+}
+
+/** Resolve a dashboard route from an English or `/zh` pathname, including child routes. */
+export function dashboardRouteForPathname(pathname: string): DashboardRoute | undefined {
+  const barePath = pathname.split(/[?#]/u, 1)[0] || '/';
+  const enPath = barePath === '/zh' ? '/' : barePath.replace(/^\/zh(?=\/)/u, '');
+  return DASHBOARD_ROUTES.find(
+    (route) =>
+      enPath === route.path ||
+      enPath.startsWith(`${route.path}/`) ||
+      enPath === route.canonicalPath,
+  );
+}
+
+const DEFAULT_DASHBOARD_SHELL_CAPABILITIES: DashboardShellCapabilities = {
+  providers: FILTERED_DASHBOARD_PROVIDERS,
+  dashboardNudge: true,
+};
+
+/** Resolve providers and dashboard nudges for an English or Chinese pathname. */
+export function dashboardShellCapabilitiesForPathname(
+  pathname: string,
+): DashboardShellCapabilities {
+  const barePath = pathname.split(/[?#]/u, 1)[0] || '/';
+  const enPath = barePath === '/zh' ? '/' : barePath.replace(/^\/zh(?=\/)/u, '');
+  const capabilityRoute = DASHBOARD_SHELL_CAPABILITY_ROUTES.find(
+    (route) =>
+      enPath === route.path || (route.includeChildren && enPath.startsWith(`${route.path}/`)),
+  );
+  if (capabilityRoute) {
+    return {
+      providers: capabilityRoute.providers,
+      dashboardNudge: capabilityRoute.dashboardNudge,
+    };
+  }
+
+  const dashboardRoute = dashboardRouteForPathname(enPath);
+  if (!dashboardRoute) return DEFAULT_DASHBOARD_SHELL_CAPABILITIES;
+  return {
+    providers: dashboardRoute.providers,
+    // Provider ownership and nudge ownership are independent. Production
+    // mounted dashboard nudges on standalone tabs too; only explicit shell
+    // overrides suppress them.
+    dashboardNudge: true,
+  };
+}
diff --git a/packages/app/src/lib/data-mappings.test.ts b/packages/app/src/lib/data-mappings.test.ts
index 85578bfc4..0c36ea9a0 100644
--- a/packages/app/src/lib/data-mappings.test.ts
+++ b/packages/app/src/lib/data-mappings.test.ts
@@ -4,6 +4,8 @@ import {
   getModelAndSequence,
   getModelAndSequenceFromArtifact,
   getModelLabel,
+  getOpenRouterModelId,
+  MODEL_OPTIONS,
   getModelExclusion,
   getSequenceDefaultExclusionGroup,
   getSequenceExclusion,
@@ -272,12 +274,12 @@ describe('getModelLabel', () => {
     expect(getModelLabel(Model.Llama3_3_70B)).toBe('Llama 3.3 70B Instruct');
     expect(getModelLabel(Model.Llama3_1_70B)).toBe('Llama 3.1 70B Instruct');
     expect(getModelLabel(Model.DeepSeek_R1)).toBe('DeepSeek R1 0528 671B');
-    expect(getModelLabel(Model.DeepSeek_V4_Pro)).toBe('DeepSeek V4 Pro 1.6T');
+    expect(getModelLabel(Model.DeepSeek_V4_Pro)).toBe('DeepSeek V4 Pro 0813 1.6T');
     expect(getModelLabel(Model.GptOss)).toBe('gpt-oss 120B');
     expect(getModelLabel(Model.Qwen3_5)).toBe('Qwen3.5 397B');
     expect(getModelLabel(Model.Kimi_K2_5)).toBe('Kimi K2.5/2.6/2.7-Code 1T');
     expect(getModelLabel(Model.GLM_5)).toBe('GLM5/5.1 744B');
-    expect(getModelLabel(Model.GLM_5_2)).toBe('GLM5.2');
+    expect(getModelLabel(Model.GLM_5_2)).toBe('GLM5.2/GLM5.3 744B');
     expect(getModelLabel(Model.MiniMax_M2_5)).toBe('MiniMax M2.5/2.7 230B');
   });
 
@@ -287,6 +289,15 @@ describe('getModelLabel', () => {
   });
 });
 
+describe('getOpenRouterModelId', () => {
+  it('maps every selectable dashboard model to an exact OpenRouter catalog id', () => {
+    for (const model of MODEL_OPTIONS) {
+      expect(getOpenRouterModelId(model), model).toMatch(/^[^/]+\/[^/]+$/u);
+    }
+    expect(getOpenRouterModelId(Model.DeepSeek_V4_Pro)).toBe('deepseek/deepseek-v4-pro-0813');
+  });
+});
+
 // ===========================================================================
 // getSequenceLabel
 // ===========================================================================
@@ -295,11 +306,11 @@ describe('getSequenceLabel', () => {
     expect(getSequenceLabel(Sequence.OneK_OneK)).toBe('1K / 1K');
     expect(getSequenceLabel(Sequence.OneK_EightK)).toBe('1K / 8K');
     expect(getSequenceLabel(Sequence.EightK_OneK)).toBe('8K / 1K');
-    expect(getSequenceLabel(Sequence.AgenticTraces)).toBe('Agentic Traces');
+    expect(getSequenceLabel(Sequence.AgenticTraces)).toBe('Agentic');
   });
 
   it('returns the Chinese agentic scenario label for the zh locale', () => {
-    expect(getSequenceLabel(Sequence.AgenticTraces, 'zh')).toBe('智能体轨迹');
+    expect(getSequenceLabel(Sequence.AgenticTraces, 'zh')).toBe('智能体');
   });
 
   it('falls back to the sequence value for unknown sequence', () => {
diff --git a/packages/app/src/lib/data-mappings.ts b/packages/app/src/lib/data-mappings.ts
index b2859e213..567e2298f 100644
--- a/packages/app/src/lib/data-mappings.ts
+++ b/packages/app/src/lib/data-mappings.ts
@@ -7,6 +7,7 @@ export enum Model {
   DeepSeek_Coder_V2_Lite = 'DeepSeek-Coder-V2-Lite-Instruct',
   GptOss = 'gpt-oss-120b',
   Qwen3_5 = 'Qwen-3.5-397B-A17B',
+  Qwen3_8_Flash_Next = 'Qwen3.8-Flash-Next',
   Kimi_K2_5 = 'Kimi-K2.5',
   Kimi_K3 = 'Kimi-K3',
   MiniMax_M2_5 = 'MiniMax-M2.5',
@@ -48,6 +49,15 @@ interface ModelConfig {
   label: string;
   prefix: string;
   category: CategoryTag;
+  /** Exact public model id in OpenRouter's `/api/v1/models` catalog. */
+  openRouterModelId?: string;
+  /**
+   * Filename under `public/logos/` for the model creator's logo, shown beside
+   * the model name in UI surfaces such as the inference chart caption. Absent =
+   * no logo rendered. Monochrome `currentColor`/black SVGs are inverted in dark
+   * mode by `ModelLogo`; colored raster assets (`.webp`) are shown as-is.
+   */
+  logo?: string;
   /**
    * Data-driven exclusion rules for this model (see `exclusion.ts`). Each spec
    * partitions matching config keys into comparability groups that can't share
@@ -65,13 +75,17 @@ interface ModelConfig {
  * Scoped to `hardware` for the same reason the STP rule below is: the guard
  * exists to stop two engines being read off one SKU's curve, not to stop a
  * chart holding two SKUs. B200 vLLM MTP next to B300 SGLang MTP compares
- * hardware, which is the point of the chart.
+ * hardware, which is the point of the chart. Where only one engine has a run
+ * for a SKU there is nothing to confuse, so that SKU is free to show it.
+ *
+ * ATOM is its own family and is deliberately NOT grouped with SGLang, so it
+ * stays comparable with everything — including both vLLM and SGLang on the
+ * SKUs where it runs.
  */
 const MTP_ENGINE_EXCLUSION: ExclusionSpec[] = [
   {
     suffix: '_mtp',
     stripPrefixes: ['dynamo-', 'mori-', 'llmd-', 'mooncake-'],
-    groupAliases: { atom: 'sglang' },
     scope: 'hardware',
   },
 ];
@@ -80,12 +94,16 @@ const MTP_ENGINE_EXCLUSION: ExclusionSpec[] = [
  * STP exclusion: unsuffixed standard-token configs for the same hardware SKU
  * can't mix engine families, because each engine tunes its serving path
  * differently. Different hardware may use different engines on one graph.
+ * `dynamo-`/`mori-`/`llmd-`/`mooncake-` are routers, not engines, so
+ * `dynamo-sglang` is SGLang and is guarded as such.
+ *
+ * ATOM is its own family and is not aliased to SGLang — it stays comparable
+ * with everything.
  */
 const STP_ENGINE_EXCLUSION: ExclusionSpec[] = [
   {
     suffix: null,
     stripPrefixes: ['dynamo-', 'mori-', 'llmd-', 'mooncake-'],
-    groupAliases: { atom: 'sglang' },
     scope: 'hardware',
   },
 ];
@@ -119,9 +137,11 @@ const GUARDED_ENGINE_FAMILIES = ['vllm', 'sglang'] as const;
 // duplication needed.
 const MODEL_CONFIG: Record<Model, ModelConfig> = {
   [Model.DeepSeek_V4_Pro]: {
-    label: 'DeepSeek V4 Pro 1.6T',
+    label: 'DeepSeek V4 Pro 0813 1.6T',
     prefix: 'dsv4',
     category: 'default',
+    openRouterModelId: 'deepseek/deepseek-v4-pro-0813',
+    logo: 'deepseek-color.svg',
     exclusion: MTP_ENGINE_EXCLUSION,
   },
   [Model.Kimi_K3]: {
@@ -130,6 +150,8 @@ const MODEL_CONFIG: Record<Model, ModelConfig> = {
     label: 'Kimi K3 2.8T',
     prefix: 'kimik3',
     category: 'default',
+    openRouterModelId: 'moonshotai/kimi-k3',
+    logo: 'kimi-color.svg',
   },
   [Model.Kimi_K2_5]: {
     // K2.5, K2.6, and K2.7-Code share an architecture, so the dropdown surfaces
@@ -145,36 +167,102 @@ const MODEL_CONFIG: Record<Model, ModelConfig> = {
     label: 'Kimi K2.5/2.6/2.7-Code 1T',
     prefix: 'kimik2.5',
     category: 'deprecated',
+    openRouterModelId: 'moonshotai/kimi-k2.7-code',
+    logo: 'kimi-color.svg',
   },
   [Model.MiniMax_M3]: {
     label: 'MiniMax M3 428B',
     prefix: 'minimaxm3',
     category: 'default',
+    openRouterModelId: 'minimax/minimax-m3',
+    logo: 'minimax-color.svg',
   },
   [Model.DeepSeek_R1]: {
     label: 'DeepSeek R1 0528 671B',
     prefix: 'dsr1',
     category: 'maintenance',
+    openRouterModelId: 'deepseek/deepseek-r1-0528',
+    logo: 'deepseek-color.svg',
+  },
+  [Model.GLM_5]: {
+    label: 'GLM5/5.1 744B',
+    prefix: 'glm5',
+    category: 'deprecated',
+    openRouterModelId: 'z-ai/glm-5.1',
+    logo: 'zai-color.svg',
+  },
+  // GLM-5.2 and GLM-5.3 share the same architecture and inference profile, so
+  // the selector presents both releases over the existing GLM-5.2 data bucket.
+  [Model.GLM_5_2]: {
+    label: 'GLM5.2/GLM5.3 744B',
+    prefix: 'glm5.2',
+    category: 'default',
+    openRouterModelId: 'z-ai/glm-5.3',
+    logo: 'zai-color.svg',
+  },
+  [Model.Qwen3_5]: {
+    label: 'Qwen3.5 397B',
+    prefix: 'qwen3.5',
+    category: 'default',
+    openRouterModelId: 'qwen/qwen3.5-397b-a17b',
+    logo: 'qwen-color.svg',
+  },
+  // 176B total: a 125B main model plus a 51B n-gram embedding table, 6B active
+  // per forward pass, and a separate 4B MTP head the parameter count excludes.
+  // Default alongside the other current models, so it appears in the /overview
+  // matrix from day zero. The matrix is built from DEFAULT_MODELS, so its
+  // fixed-sequence row stays empty until the sweep covers 8k1k as well as the
+  // agentic scenario.
+  [Model.Qwen3_8_Flash_Next]: {
+    label: 'Qwen3.8 Flash Next 176B',
+    prefix: 'qwen3.8next',
+    category: 'default',
+    openRouterModelId: 'qwen/qwen3.8-flash',
+    logo: 'qwen-color.svg',
+  },
+  [Model.GptOss]: {
+    label: 'gpt-oss 120B',
+    prefix: 'gptoss',
+    category: 'deprecated',
+    openRouterModelId: 'openai/gpt-oss-120b',
+    logo: 'openai.svg',
   },
   [Model.DeepSeek_Coder_V2_Lite]: {
     label: 'DeepSeek Coder V2 Lite 16B',
     prefix: 'dsv2lite',
     category: 'default',
+    openRouterModelId: 'deepseek/deepseek-coder-v2-lite-instruct',
+    logo: 'deepseek-color.svg',
   },
-  [Model.GLM_5]: { label: 'GLM5/5.1 744B', prefix: 'glm5', category: 'deprecated' },
-  [Model.GLM_5_2]: { label: 'GLM5.2', prefix: 'glm5.2', category: 'default' },
-  [Model.Qwen3_5]: { label: 'Qwen3.5 397B', prefix: 'qwen3.5', category: 'default' },
-  [Model.GptOss]: { label: 'gpt-oss 120B', prefix: 'gptoss', category: 'deprecated' },
   [Model.MiniMax_M2_5]: {
     // M2.5 and M2.7 share an architecture — same GLM5/5.1 pattern as Kimi.
     // Superseded by MiniMax M3, so it's deprecated (no longer actively benchmarked).
     label: 'MiniMax M2.5/2.7 230B',
     prefix: 'minimaxm2.5',
     category: 'deprecated',
+    openRouterModelId: 'minimax/minimax-m2.7',
+    logo: 'minimax-color.svg',
+  },
+  [Model.Llama3_3_70B]: {
+    label: 'Llama 3.3 70B Instruct',
+    prefix: '70b',
+    category: 'deprecated',
+    openRouterModelId: 'meta-llama/llama-3.3-70b-instruct',
+    logo: 'meta-color.svg',
+  },
+  [Model.Llama3_1_70B]: {
+    label: 'Llama 3.1 70B Instruct',
+    prefix: '',
+    category: 'hidden',
+    openRouterModelId: 'meta-llama/llama-3.1-70b-instruct',
+    logo: 'meta-color.svg',
+  },
+  [Model.Gemma4_31B]: {
+    label: 'Gemma 4 31B',
+    prefix: 'gemma4',
+    category: 'default',
+    openRouterModelId: 'google/gemma-4-31b-it',
   },
-  [Model.Llama3_3_70B]: { label: 'Llama 3.3 70B Instruct', prefix: '70b', category: 'deprecated' },
-  [Model.Llama3_1_70B]: { label: 'Llama 3.1 70B Instruct', prefix: '', category: 'hidden' },
-  [Model.Gemma4_31B]: { label: 'Gemma 4 31B', prefix: 'gemma4', category: 'default' },
 };
 
 function modelsByCategory(cat: CategoryTag): ReadonlySet<Model> {
@@ -215,6 +303,20 @@ export function getModelLabel(model: Model): string {
   return MODEL_CONFIG[model]?.label ?? model;
 }
 
+/** Exact OpenRouter catalog model used for live input/output pricing. */
+export function getOpenRouterModelId(model: Model): string | null {
+  return MODEL_CONFIG[model]?.openRouterModelId ?? null;
+}
+
+/**
+ * Filename under `public/logos/` for the model creator's logo, or null when the
+ * model has no configured logo. Callers render it via `ModelLogo`, which
+ * handles dark-mode inversion and load-failure fallback.
+ */
+export function getModelLogo(model: Model): string | null {
+  return MODEL_CONFIG[model]?.logo ?? null;
+}
+
 /**
  * Exclusion specs configured for a model (see `exclusion.ts`). Empty when the
  * model has no exclusion rules.
@@ -326,8 +428,8 @@ const SEQUENCE_CONFIG: Record<Sequence, SequenceConfig> = {
     exclusionFamilies: GUARDED_ENGINE_FAMILIES,
   },
   [Sequence.AgenticTraces]: {
-    label: 'Agentic Traces',
-    labelZh: '智能体轨迹',
+    label: 'Agentic',
+    labelZh: '智能体',
     compact: 'agentic',
     category: 'default',
     kind: 'agentic',
diff --git a/packages/app/src/lib/default-sequence.test.ts b/packages/app/src/lib/default-sequence.test.ts
index 8948bcd12..5dec6ea4b 100644
--- a/packages/app/src/lib/default-sequence.test.ts
+++ b/packages/app/src/lib/default-sequence.test.ts
@@ -34,7 +34,56 @@ describe('resolveEffectiveSequence', () => {
     });
   });
 
-  describe('honors a valid selection (rule 2a)', () => {
+  describe('availability-driven AgentX default (rule 2)', () => {
+    it('opens the Agentic scenario whenever the model has data and the user has not chosen', () => {
+      expect(
+        resolveEffectiveSequence({
+          selectedSequence: Sequence.EightK_OneK,
+          availableSequences: [Sequence.EightK_OneK, Sequence.AgenticTraces],
+          availabilityLoaded: true,
+          sequenceExplicit: false,
+        }),
+      ).toBe(Sequence.AgenticTraces);
+    });
+
+    it('yields to an explicit selection', () => {
+      // A shared `?i_seq=8k-1k` link (or a manual pick) must stay on 8K/1K.
+      expect(
+        resolveEffectiveSequence({
+          selectedSequence: Sequence.EightK_OneK,
+          availableSequences: [Sequence.EightK_OneK, Sequence.AgenticTraces],
+          availabilityLoaded: true,
+          sequenceExplicit: true,
+        }),
+      ).toBe(Sequence.EightK_OneK);
+    });
+
+    it('keeps the app default when the model has no Agentic data', () => {
+      expect(
+        resolveEffectiveSequence({
+          selectedSequence: Sequence.EightK_OneK,
+          availableSequences: [Sequence.EightK_OneK],
+          availabilityLoaded: true,
+          sequenceExplicit: false,
+        }),
+      ).toBe(Sequence.EightK_OneK);
+    });
+
+    it('does not apply before availability has loaded', () => {
+      // The static fallback list contains every scenario, so applying the
+      // AgentX preference here would fetch rows the model may not have.
+      expect(
+        resolveEffectiveSequence({
+          selectedSequence: Sequence.EightK_OneK,
+          availableSequences: [Sequence.EightK_OneK, Sequence.AgenticTraces],
+          availabilityLoaded: false,
+          sequenceExplicit: false,
+        }),
+      ).toBeNull();
+    });
+  });
+
+  describe('honors a valid selection (rule 3a)', () => {
     it('keeps AgenticTraces when the model actually has agentic data (dsr1 case)', () => {
       // DeepSeek-R1 in the seeded DB has both agentic and 8k/1k — an explicit
       // agentic selection (e.g. a shared ?i_seq= link) must survive.
@@ -58,7 +107,7 @@ describe('resolveEffectiveSequence', () => {
     });
   });
 
-  describe('fallback ordering when the selection is unavailable (rule 2b/2c)', () => {
+  describe('fallback ordering when the selection is unavailable (rule 3b/3c)', () => {
     it('for a fixed-seq-only model, an agentic selection falls back to 8k/1k, not the raw first entry (llama70b case)', () => {
       // Llama-3.3-70B has only 8k/1k in the seeded DB. An agentic selection is
       // unavailable, so it must resolve to a fixed-seq scenario — here the sole
diff --git a/packages/app/src/lib/default-sequence.ts b/packages/app/src/lib/default-sequence.ts
index 9c01f56ef..631045d1a 100644
--- a/packages/app/src/lib/default-sequence.ts
+++ b/packages/app/src/lib/default-sequence.ts
@@ -4,12 +4,12 @@ import { Sequence } from './data-mappings';
  * Effective-sequence resolution.
  *
  * `selectedSequence` defaults to {@link Sequence.EightK_OneK}, but the user may
- * have selected a scenario the model doesn't offer (e.g. Agentic Traces via a
+ * have selected a scenario the model doesn't offer (e.g. the Agentic scenario via a
  * shared `?i_seq=` link on a fixed-seq-only model). This helper turns the raw
  * user/default selection into the sequence the chart should actually render,
  * given what the selected model offers.
  *
- * Two rules, in order:
+ * Three rules, in order:
  *
  * 1. **Availability gate.** Until availability rows have loaded we do NOT know
  *    which sequences the model has. Resolving eagerly here would pick the static
@@ -21,33 +21,49 @@ import { Sequence } from './data-mappings';
  *    on a non-null result (a loading skeleton covers this window, which is
  *    short).
  *
- * 2. **Fallback ordering.** Once availability is known: keep the user's
- *    `selectedSequence` if the model has it. Otherwise fall back to a sensible
- *    fixed-seq scenario. `availableSequences[0]` follows DB row order, which can
- *    surface `1k/1k` even when `8k/1k` exists — but `8k/1k` is the app default
- *    scenario, so prefer it when present. Only if neither the selection nor
- *    `8k/1k` is available do we fall to `availableSequences[0]`.
+ * 2. **Availability-driven default.** `selectedSequence` starts at the app-wide default
+ *    (`8k/1k`) until the user actually picks something, so it can't distinguish
+ *    "the user wants 8K/1K" from "nobody has chosen yet". `sequenceExplicit`
+ *    makes that distinction: while it is false, any model with Agentic data
+ *    opens on that scenario instead of 8K/1K. Any explicit selection turns the
+ *    flag on and this rule stops applying.
+ *
+ * 3. **Fallback ordering.** Otherwise: keep the user's `selectedSequence` if the
+ *    model has it. Otherwise fall back to a sensible fixed-seq scenario.
+ *    `availableSequences[0]` follows DB row order, which can surface `1k/1k`
+ *    even when `8k/1k` exists — but `8k/1k` is the app default scenario, so
+ *    prefer it when present. Only if neither the selection nor `8k/1k` is
+ *    available do we fall to `availableSequences[0]`.
  */
 export function resolveEffectiveSequence({
   selectedSequence,
   availableSequences,
   availabilityLoaded,
+  sequenceExplicit = false,
 }: {
   selectedSequence: Sequence;
   availableSequences: Sequence[];
   availabilityLoaded: boolean;
+  /** Whether `selectedSequence` came from a real user/route choice. */
+  sequenceExplicit?: boolean;
 }): Sequence | null {
   // Rule 1: do not commit to a sequence before we know what the model has.
   if (!availabilityLoaded) return null;
 
-  // Rule 2a: honor the user's / default selection when the model supports it.
+  // Rule 2: nobody has chosen yet — prefer the real AgentX workload whenever
+  // availability confirms that the selected model has data for it.
+  if (!sequenceExplicit && availableSequences.includes(Sequence.AgenticTraces)) {
+    return Sequence.AgenticTraces;
+  }
+
+  // Rule 3a: honor the user's / default selection when the model supports it.
   if (availableSequences.includes(selectedSequence)) return selectedSequence;
 
-  // Rule 2b: prefer 8k/1k (the app default scenario) over whatever
+  // Rule 3b: prefer 8k/1k (the app default scenario) over whatever
   // availableSequences[0] happens to be (DB row order can yield 1k/1k).
   if (availableSequences.includes(Sequence.EightK_OneK)) return Sequence.EightK_OneK;
 
-  // Rule 2c: last resort — first available, or the selection itself if the model
+  // Rule 3c: last resort — first available, or the selection itself if the model
   // has no sequences at all (keeps the type non-null; downstream shows empty).
   return availableSequences[0] ?? selectedSequence;
 }
diff --git a/packages/app/src/lib/dynamic-colors.test.ts b/packages/app/src/lib/dynamic-colors.test.ts
index 52550a094..843c4bdc5 100644
--- a/packages/app/src/lib/dynamic-colors.test.ts
+++ b/packages/app/src/lib/dynamic-colors.test.ts
@@ -2,7 +2,11 @@ import { describe, expect, it } from 'vitest';
 
 import { VENDOR_OKLCH_ZONES } from '@semianalysisai/inferencex-constants';
 
-import { generateVendorColors, getVendor } from './dynamic-colors';
+import {
+  generateHighContrastGpuDateColors,
+  generateVendorColors,
+  getVendor,
+} from './dynamic-colors';
 
 function hueOf(color: string): number {
   const match = /^oklch\([\d.]+ [\d.]+ (?<hue>[\d.]+)\)$/.exec(color);
@@ -14,6 +18,7 @@ describe('getVendor', () => {
   it('classifies registered GPU base keys through GPU_VENDORS', () => {
     expect(getVendor('h100_vllm')).toBe('nvidia');
     expect(getVendor('mi300x_sglang')).toBe('amd');
+    expect(getVendor('jalapeno_teacup')).toBe('teacup');
   });
 
   it('classifies keys that lead with a literal vendor token', () => {
@@ -30,13 +35,19 @@ describe('getVendor', () => {
 
 describe('generateVendorColors', () => {
   it('places vendor-prefixed keys in their vendor hue zones', () => {
-    const colors = generateVendorColors(['nvidia_series-a', 'amd_series-b'], 'light');
+    const colors = generateVendorColors(
+      ['nvidia_series-a', 'amd_series-b', 'jalapeno_teacup'],
+      'light',
+    );
     const nvidia = VENDOR_OKLCH_ZONES.nvidia;
     const amd = VENDOR_OKLCH_ZONES.amd;
+    const teacup = VENDOR_OKLCH_ZONES.teacup;
     expect(hueOf(colors['nvidia_series-a'])).toBeGreaterThanOrEqual(nvidia.start);
     expect(hueOf(colors['nvidia_series-a'])).toBeLessThanOrEqual(nvidia.end);
     expect(hueOf(colors['amd_series-b'])).toBeGreaterThanOrEqual(amd.start);
     expect(hueOf(colors['amd_series-b'])).toBeLessThanOrEqual(amd.end);
+    expect(hueOf(colors.jalapeno_teacup)).toBeGreaterThanOrEqual(teacup.start);
+    expect(hueOf(colors.jalapeno_teacup)).toBeLessThanOrEqual(teacup.end);
   });
 
   it('keeps unclassifiable keys in the unknown zone', () => {
@@ -46,3 +57,60 @@ describe('generateVendorColors', () => {
     expect(hueOf(colors['mystery_series'])).toBeLessThanOrEqual(unknown.end);
   });
 });
+
+function lightnessOf(color: string): number {
+  const match = /^oklch\((?<l>[\d.]+) [\d.]+ [\d.]+\)$/u.exec(color);
+  expect(match, `expected oklch color, got ${color}`).not.toBeNull();
+  return Number(match!.groups!.l);
+}
+
+describe('generateHighContrastGpuDateColors', () => {
+  const bases = { h200_sglang: '#1f77b4', mi355x_vllm: '#d62728' };
+
+  it('gives every date of one GPU the same hue and chroma', () => {
+    const colors = generateHighContrastGpuDateColors(bases, 4, 'light');
+    const hues = [0, 1, 2, 3].map((di) => hueOf(colors[`${di}_h200_sglang`]));
+    expect(new Set(hues).size).toBe(1);
+  });
+
+  it('keeps different GPUs on different hues', () => {
+    const colors = generateHighContrastGpuDateColors(bases, 3, 'light');
+    expect(hueOf(colors['0_h200_sglang'])).not.toBeCloseTo(hueOf(colors['0_mi355x_vllm']), 0);
+  });
+
+  it('ramps lightness oldest → newest, matching the non-HC ramp direction', () => {
+    const colors = generateHighContrastGpuDateColors(bases, 3, 'light');
+    const ls = [0, 1, 2].map((di) => lightnessOf(colors[`${di}_h200_sglang`]));
+    expect(ls[0]).toBeGreaterThan(ls[1]);
+    expect(ls[1]).toBeGreaterThan(ls[2]);
+  });
+
+  it('separates consecutive dates enough to be noticeable', () => {
+    for (const theme of ['light', 'dark'] as const) {
+      const colors = generateHighContrastGpuDateColors(bases, 4, theme);
+      const ls = [0, 1, 2, 3].map((di) => lightnessOf(colors[`${di}_mi355x_vllm`]));
+      for (let i = 1; i < ls.length; i++) {
+        expect(ls[i - 1] - ls[i], `${theme} step ${i}`).toBeGreaterThan(0.08);
+      }
+    }
+  });
+
+  it('keeps the full span inside theme bounds even for a very dark base', () => {
+    const colors = generateHighContrastGpuDateColors({ gpu: '#0a0a1e' }, 2, 'dark');
+    const ls = [0, 1].map((di) => lightnessOf(colors[`${di}_gpu`]));
+    expect(Math.min(...ls)).toBeGreaterThanOrEqual(0.44);
+    expect(Math.max(...ls)).toBeLessThanOrEqual(0.95);
+    expect(ls[0] - ls[1]).toBeCloseTo(0.36, 2);
+  });
+
+  it('holds a single date at the base lightness', () => {
+    const colors = generateHighContrastGpuDateColors(bases, 1, 'light');
+    expect(Object.keys(colors)).toEqual(['0_h200_sglang', '0_mi355x_vllm']);
+  });
+
+  it('falls back to the base color when it cannot be parsed', () => {
+    const colors = generateHighContrastGpuDateColors({ gpu: 'var(--foreground)' }, 2, 'light');
+    expect(colors['0_gpu']).toBe('var(--foreground)');
+    expect(colors['1_gpu']).toBe('var(--foreground)');
+  });
+});
diff --git a/packages/app/src/lib/dynamic-colors.ts b/packages/app/src/lib/dynamic-colors.ts
index d788c8fd6..2054643c8 100644
--- a/packages/app/src/lib/dynamic-colors.ts
+++ b/packages/app/src/lib/dynamic-colors.ts
@@ -14,7 +14,7 @@ import { getModelSortIndex } from '@/lib/constants';
 // Vendor detection
 // ---------------------------------------------------------------------------
 
-export type Vendor = 'nvidia' | 'amd' | 'unknown';
+export type Vendor = 'nvidia' | 'amd' | 'teacup' | 'unknown';
 
 /** Determine vendor from a hardware key by looking up GPU_VENDORS. */
 export function getVendor(hwKey: string): Vendor {
@@ -22,10 +22,11 @@ export function getVendor(hwKey: string): Vendor {
   const base = hwKey.split('_')[0];
   // Keys whose dataset carries an explicit vendor (e.g. CollectiveX series) lead
   // with the vendor name itself rather than a registered GPU key.
-  if (base === 'nvidia' || base === 'amd') return base;
+  if (base === 'nvidia' || base === 'amd' || base === 'teacup') return base;
   const vendor = GPU_VENDORS[base];
   if (vendor === 'NVIDIA') return 'nvidia';
   if (vendor === 'AMD') return 'amd';
+  if (vendor === 'Teacup') return 'teacup';
   return 'unknown';
 }
 
@@ -166,3 +167,105 @@ export function generateGpuDateColors(
 
   return result;
 }
+
+// ---------------------------------------------------------------------------
+// High-contrast GPU × date gradient
+// ---------------------------------------------------------------------------
+
+/**
+ * Total OKLch lightness span swept across the compared dates in high-contrast
+ * mode. Wide on purpose: the whole point of the gradient is that consecutive
+ * dates read apart at a glance, so a subtle ramp would defeat it.
+ */
+const HC_DATE_L_SPAN = 0.36;
+
+/**
+ * Absolute lightness bounds for the ramp so neither end of the gradient washes
+ * out against the page background (dark themes need a higher floor).
+ */
+const HC_DATE_L_BOUNDS = {
+  light: { min: 0.34, max: 0.86 },
+  dark: { min: 0.44, max: 0.95 },
+} as const;
+
+/** sRGB channel (0–1) → linear-light. */
+function srgbToLinear(c: number): number {
+  return c <= 0.04045 ? c / 12.92 : ((c + 0.055) / 1.055) ** 2.4;
+}
+
+/** Parse `#rgb` / `#rrggbb` into OKLch `[L, C, H]`, or null when unparseable. */
+function hexToOklch(hex: string): [number, number, number] | null {
+  const raw = hex.trim().replace('#', '');
+  const full = raw.length === 3 ? [...raw].map((ch) => ch + ch).join('') : raw;
+  if (full.length !== 6 || !/^[0-9a-f]{6}$/iu.test(full)) return null;
+
+  const r = srgbToLinear(Number.parseInt(full.slice(0, 2), 16) / 255);
+  const g = srgbToLinear(Number.parseInt(full.slice(2, 4), 16) / 255);
+  const b = srgbToLinear(Number.parseInt(full.slice(4, 6), 16) / 255);
+
+  // Linear sRGB → OKLab (Björn Ottosson's matrices).
+  const l = Math.cbrt(0.4122214708 * r + 0.5363325363 * g + 0.0514459929 * b);
+  const m = Math.cbrt(0.2119034982 * r + 0.6806995451 * g + 0.1073969566 * b);
+  const s = Math.cbrt(0.0883024619 * r + 0.2817188376 * g + 0.6299787005 * b);
+
+  const okL = 0.2104542553 * l + 0.793617785 * m - 0.0040720468 * s;
+  const okA = 1.9779984951 * l - 2.428592205 * m + 0.4505937099 * s;
+  const okB = 0.0259040371 * l + 0.7827717662 * m - 0.808675766 * s;
+
+  const chroma = Math.hypot(okA, okB);
+  const hue = ((Math.atan2(okB, okA) * 180) / Math.PI + 360) % 360;
+  return [okL, chroma, hue];
+}
+
+/**
+ * Spread one high-contrast base color per GPU across the compared dates.
+ *
+ * High contrast previously handed every `date × GPU` series its own iwanthue
+ * hue, which severed the visual link between a hardware config's own runs. Here
+ * the hue/chroma stay pinned to the GPU (so a config keeps one identity) and
+ * only lightness ramps — oldest lightest → newest darkest, matching the
+ * non-high-contrast ramp in {@link generateGpuDateColors}.
+ *
+ * @param baseColors - hwKey → high-contrast base color (hex from iwanthue).
+ * @param dateCount  - Number of dates being compared.
+ * @param theme      - 'light' or 'dark'.
+ * @returns Map of `${date-index}_${hwKey}` → `oklch(L C H)` string. GPUs whose
+ *          base color can't be parsed keep that base color at every date index.
+ */
+export function generateHighContrastGpuDateColors(
+  baseColors: Record<string, string>,
+  dateCount: number,
+  theme: 'light' | 'dark',
+): Record<string, string> {
+  const result: Record<string, string> = {};
+  if (dateCount <= 0) return result;
+
+  const { min: lMin, max: lMax } = HC_DATE_L_BOUNDS[theme];
+
+  for (const [hwKey, base] of Object.entries(baseColors)) {
+    const oklch = hexToOklch(base);
+    if (!oklch) {
+      for (let di = 0; di < dateCount; di++) result[`${di}_${hwKey}`] = base;
+      continue;
+    }
+    const [baseL, chroma, hue] = oklch;
+
+    // Center the ramp on the base lightness, then slide (not squash) it back
+    // inside the theme bounds so the full span survives near the extremes.
+    const span = Math.min(HC_DATE_L_SPAN, lMax - lMin);
+    let top = Math.min(lMax, baseL + span / 2);
+    let bottom = top - span;
+    if (bottom < lMin) {
+      bottom = lMin;
+      top = bottom + span;
+    }
+
+    for (let di = 0; di < dateCount; di++) {
+      const lightness = dateCount <= 1 ? baseL : top - (di / (dateCount - 1)) * span;
+      result[`${di}_${hwKey}`] =
+        `oklch(${lightness.toFixed(3)} ${chroma.toFixed(3)} ${hue.toFixed(1)})`;
+    }
+  }
+
+  return result;
+}
diff --git a/packages/app/src/lib/eval-sample-params.test.ts b/packages/app/src/lib/eval-sample-params.test.ts
new file mode 100644
index 000000000..7b4a4b578
--- /dev/null
+++ b/packages/app/src/lib/eval-sample-params.test.ts
@@ -0,0 +1,32 @@
+import { describe, expect, it } from 'vitest';
+
+import { EVAL_SAMPLE_FILTER_ERROR, parseEvalSampleWindow } from './eval-sample-params';
+
+describe('parseEvalSampleWindow', () => {
+  it('shares stable defaults across stored and live readers', () => {
+    expect(parseEvalSampleWindow(new URLSearchParams())).toEqual({
+      filter: 'all',
+      offset: 0,
+      limit: 200,
+    });
+  });
+
+  it('normalizes non-finite pagination and clamps the window', () => {
+    expect(parseEvalSampleWindow(new URLSearchParams('offset=nan&limit=nan'))).toEqual({
+      filter: 'all',
+      offset: 0,
+      limit: 200,
+    });
+    expect(parseEvalSampleWindow(new URLSearchParams('offset=-8&limit=9999'))).toEqual({
+      filter: 'all',
+      offset: 0,
+      limit: 500,
+    });
+  });
+
+  it('returns the canonical filter error', () => {
+    expect(parseEvalSampleWindow(new URLSearchParams('filter=unknown'))).toEqual({
+      error: EVAL_SAMPLE_FILTER_ERROR,
+    });
+  });
+});
diff --git a/packages/app/src/lib/eval-sample-params.ts b/packages/app/src/lib/eval-sample-params.ts
new file mode 100644
index 000000000..a07b757e2
--- /dev/null
+++ b/packages/app/src/lib/eval-sample-params.ts
@@ -0,0 +1,36 @@
+export type EvalSampleFilter = 'all' | 'passed' | 'failed';
+
+export interface EvalSampleWindow {
+  readonly filter: EvalSampleFilter;
+  readonly offset: number;
+  readonly limit: number;
+}
+
+export const EVAL_SAMPLE_FILTER_ERROR = 'filter must be one of: all, passed, failed';
+
+const ALLOWED_FILTERS: Readonly<Record<EvalSampleFilter, true>> = {
+  all: true,
+  passed: true,
+  failed: true,
+};
+const DEFAULT_LIMIT = 200;
+const MAX_LIMIT = 500;
+
+/** Parse the pagination/filter contract shared by stored and live evaluation samples. */
+export function parseEvalSampleWindow(
+  params: URLSearchParams,
+): EvalSampleWindow | { readonly error: string } {
+  const filterParam = params.get('filter') ?? 'all';
+  if (!Object.hasOwn(ALLOWED_FILTERS, filterParam)) {
+    return { error: EVAL_SAMPLE_FILTER_ERROR };
+  }
+
+  const parsedOffset = Math.trunc(Number(params.get('offset') ?? '0'));
+  const offset = Number.isFinite(parsedOffset) ? Math.max(0, parsedOffset) : 0;
+  const parsedLimit = Math.trunc(Number(params.get('limit') ?? String(DEFAULT_LIMIT)));
+  const limit = Number.isFinite(parsedLimit)
+    ? Math.min(MAX_LIMIT, Math.max(1, parsedLimit || DEFAULT_LIMIT))
+    : DEFAULT_LIMIT;
+
+  return { filter: filterParam as EvalSampleFilter, offset, limit };
+}
diff --git a/packages/app/src/lib/glossary-zh.ts b/packages/app/src/lib/glossary-zh.ts
index aff5cf860..cbbb159ad 100644
--- a/packages/app/src/lib/glossary-zh.ts
+++ b/packages/app/src/lib/glossary-zh.ts
@@ -8,6 +8,7 @@ import {
 export const GLOSSARY_CATEGORY_LABELS_ZH: Readonly<Record<GlossaryCategory, string>> = {
   'Benchmark metrics': '基准指标',
   Serving: '推理服务',
+  'Agentic inference': '智能体推理',
   Parallelism: '并行策略',
   Hardware: '硬件',
   'Numerical precision': '数值精度',
@@ -39,18 +40,100 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '推理既是模型问题，也是系统问题。用户体验取决于延迟和交互性，运营成本则取决于吞吐量、利用率、功耗与硬件成本；只优化其中一个维度，往往会牺牲另一个维度。',
     benchmarkContext:
-      'InferenceX 测试完整的推理方案，因为芯片峰值规格无法代表实际服务性能。每条曲线都对应明确的模型、引擎、精度、并行策略、Chip 系统、序列长度和并发扫描。',
+      'InferenceX 测试完整的推理方案，因为芯片峰值规格无法代表实际服务性能。每条曲线都对应明确的模型、引擎、精度、并行策略、芯片系统、序列长度和并发扫描。',
+  },
+  'agentic-inference': {
+    term: '智能体推理',
+    aliases: ['agentic inference', 'AI 智能体推理', 'agent 推理'],
+    plainEnglish:
+      '智能体会通过多次模型请求完成一项任务，期间可能调用工具、保留会话状态，也可能把部分工作交给 subagent。',
+    definition:
+      '智能体推理是为 AI agent 提供模型服务的过程。此类 agent 会在多个轮次间保留状态、调用工具、复用不断增长的上下文，并可能并行运行 subagent。',
+    explanation:
+      '一次智能体会话会在模型请求、工具执行和等待之间交替。后续请求通常携带大部分历史对话，因此 prefix cache 和 KV cache 容量会直接影响速度与成本。并行 subagent 还会产生各自具有时间关系和上下文增长过程的请求分支。',
+    significance:
+      '固定 input 和 output 长度无法覆盖智能体带来的全部服务压力。长共享前缀会改变 cache 行为，工具等待会让流量呈现突发性，并行分支则会争用推理容量。同一组软硬件在这种请求模式下可能出现不同的性能排序。',
+    benchmarkContext:
+      'InferenceX 使用 AgentX 测量智能体推理。AgentX 与固定序列场景回答不同的容量问题，应分别比较。AgentX 采用闭环会话回放，同一会话中的后续请求会受前一轮完成时间影响。',
+  },
+  agentx: {
+    term: 'AgentX',
+    aliases: ['AgentX 基准测试', 'AgentX 场景'],
+    plainEnglish: 'AgentX 是 InferenceX 用来测试完整长上下文、多轮编码智能体会话的工作负载。',
+    definition:
+      'AgentX 是 InferenceX 的智能体推理基准测试场景，其工作负载形状来自自愿提供的编码智能体轨迹，并在移除原始内容后构建。',
+    explanation:
+      'AgentX 使用确定性合成 token 重建会话结构。回放会保留请求长度、轮次间隔、共享前缀增长、工具等待以及主 agent 与 subagent 的依赖关系；原始提示词、生成代码和工具 payload 不会进入测试数据。推理栈只接收请求模式。',
+    significance:
+      '长上下文考验 KV cache 容量，重复前缀考验 cache 复用，分支时序则考验请求调度。这些影响在短小独立请求中较少出现。最终曲线反映推理系统在智能体流量下的整体表现。',
+    benchmarkContext:
+      '具备对应数据的模型默认显示 Agentic 场景。比较 AgentX 结果时，应使用兼容设置下的其他 AgentX 运行，并同时查看吞吐量、延迟与交互性。固定序列场景适合分析传统请求流。',
+  },
+  'agentic-coding-workload': {
+    term: '智能体编码工作负载',
+    aliases: ['agentic coding workload', '编码智能体工作负载', '软件工程智能体工作负载'],
+    plainEnglish:
+      '编码智能体读取代码仓库、修改代码并运行工具，随后继续请求模型，直到完成任务；这一连串请求就是智能体编码工作负载。',
+    definition:
+      '智能体编码工作负载由软件 agent 产生，包含多轮模型生成、代码仓库检查、工具调用、代码修改以及委派给 subagent 的子任务。',
+    explanation:
+      '随着 agent 累积指令、文件、工具结果和历史回答，请求长度会持续增长。许多轮次会复用较大的共享前缀。工具执行造成不均匀的停顿，subagent 则可能生成时间重叠的请求分支，因此其流量形状与固定长度提示词不同。',
+    significance:
+      '编码智能体可能通过一串相互依赖的调用让推理系统持续工作数分钟甚至数小时。Cache 策略、调度公平性、内存容量和尾延迟都会影响任务进度，仅看峰值 decode 吞吐量无法描述这种体验。',
+    benchmarkContext:
+      'AgentX 使用从轨迹提取的请求形状与确定性合成内容表示该工作负载。它测量推理系统性能。模型能否完成编码任务需要单独做质量评估，因此质量分数与 AgentX 服务结果应分别解读。',
+  },
+  'trace-replay': {
+    term: '轨迹回放',
+    aliases: ['trace replay', '工作负载回放', '会话回放'],
+    plainEnglish:
+      '轨迹回放重现已记录会话的请求大小、先后关系与时间间隔，让基准测试按照原始工作负载的节奏发送请求。',
+    definition:
+      '轨迹回放是一种基准测试方法，它把记录下来的请求关系、长度和时间信息转化为可重复运行的系统工作负载。',
+    explanation:
+      '回放可以保留由主 agent 轮次、并行 subagent 分支和辅助请求组成的有向图。确定性合成 token 会替换私有内容，同时保留 token 数量与前缀关系。轮次间记录的停顿则重现 agent 使用工具或等待依赖项的时间。',
+    significance:
+      '这种方法能够表达独立提示词列表缺少的流量特征，也能使用相同会话形状重复比较不同软硬件。AgentX 会在公开回放数据前移除源会话内容。',
+    benchmarkContext:
+      'AgentX 通过 AIPerf 回放由轨迹衍生的会话。固定 seed 决定会话采样、起点和合成内容。对外结果只统计 cache warmup 后的 profiling 窗口，使多次运行聚焦于稳态服务表现。',
+  },
+  'closed-loop-benchmark': {
+    term: '闭环基准测试',
+    aliases: ['closed-loop benchmark', '闭环负载测试', '闭环工作负载'],
+    plainEnglish:
+      '在闭环基准测试中，每个模拟用户会等待当前步骤完成，再按照该会话的依赖关系发送下一步请求。',
+    definition:
+      '闭环基准测试中的客户端会在前一个依赖请求完成后生成新工作，同时遵循工作负载记录的等待时间和分支结构。',
+    explanation:
+      'Concurrency 表示活跃客户端或会话数量，同时存在的请求数会随时间变化。更快的系统更早完成轮次，因此会在同一个 profiling 窗口内发出更多请求。每条采样会话的推进速度取决于请求完成时间，实际请求组合可能有小幅变化。',
+    significance:
+      '这种负载模型符合交互式 agent 的运行方式，因为下一步动作依赖上一步结果。响应更快时，会话也会更快地产生后续工作，所以吞吐量与延迟相互关联。低并发运行的采样波动通常会比大型请求池更明显。',
+    benchmarkContext:
+      'AgentX 使用闭环 concurrency。该数值表示同时运行的 agent 客户端数量；request batch 会随着会话推进而变化。解读结果时需要结合吞吐量、首 token 延迟与交互性。',
+  },
+  subagent: {
+    term: '子智能体',
+    aliases: ['subagent', 'child agent', '委派智能体'],
+    plainEnglish: '子智能体由主 agent 启动，负责同一任务中的较小部分，并可能与其他工作同时运行。',
+    definition:
+      '子智能体是一次受委派的 agent 执行，拥有独立会话状态和模型请求，并通过任务关系与依赖关系连接到父会话。',
+    explanation:
+      '主 agent 可以启动一个或多个 subagent，稍后再使用它们的结果。Subagent 的请求可能与父会话或其他分支重叠，从而形成会话图中的分支。每个分支可以增长独立上下文，同时复用部分初始指令或代码仓库状态。',
+    significance:
+      'Subagent 会让智能体流量不再完全串行。一个用户任务可能在短时间内产生多条长上下文请求，调度策略会影响各分支的完成速度。系统总吞吐量上升时，单个分支仍可能等待更久。',
+    benchmarkContext:
+      'AgentX 会保留轨迹衍生工作负载中的 subagent 分支及其依赖关系。委派质量不在测试范围内。基准测试测量推理系统如何处理由此产生的并行请求、共享前缀和完成时序。',
   },
   'inference-engine': {
     term: '推理引擎',
     aliases: ['inference engine', '服务引擎', 'LLM 服务框架'],
     plainEnglish:
-      '推理引擎就像 AI 服务背后的交通调度员：它安排请求流转，并让 Chip 在正确时间执行正确任务。',
+      '推理引擎就像 AI 服务背后的交通调度员：它安排请求流转，并让芯片在正确时间执行正确任务。',
     definition: '推理引擎是将模型权重和用户请求转化为加速器上生成结果的软件运行时。',
     explanation:
-      '它负责请求调度、连续批处理、KV 缓存分配、分布式执行、内核选择与 token 采样。vLLM、SGLang 和 TensorRT-LLM 即使运行同一模型和 Chip，也会因调度器、内核与分布式策略不同而产生不同曲线。',
+      '它负责请求调度、连续批处理、KV 缓存分配、分布式执行、内核选择与 token 采样。vLLM、SGLang 和 TensorRT-LLM 即使运行同一模型和芯片，也会因调度器、内核与分布式策略不同而产生不同曲线。',
     significance:
-      '引擎版本和配置有时与 Chip 选择同样重要。一次调度器更新、融合注意力内核或模型专用路径修复，都可能在硬件不变时带来数倍性能变化。',
+      '引擎版本和配置有时与芯片选择同样重要。一次调度器更新、融合注意力内核或模型专用路径修复，都可能在硬件不变时带来数倍性能变化。',
     benchmarkContext:
       'InferenceX 将引擎和容器镜像记录为可复现方案的一部分，因此历史视图能够区分软件进步与芯片代际进步。',
   },
@@ -60,12 +143,12 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     plainEnglish: '吞吐量就是整个系统每秒一共能完成多少工作。',
     definition: '吞吐量是推理系统在所有活跃请求上生成 token 的总速率。',
     explanation:
-      'InferenceX 通常使用每 Chip 每秒 token 数进行归一化，便于比较不同规模的系统。提高批大小或并发往往能摊薄权重读取和计算成本，从而提高总吞吐量，但单个用户收到 token 的速度可能下降。',
+      'InferenceX 通常使用每芯片每秒 token 数进行归一化，便于比较不同规模的系统。提高批大小或并发往往能摊薄权重读取和计算成本，从而提高总吞吐量，但单个用户收到 token 的速度可能下降。',
     significance:
       '最大吞吐量不是完整的性能结论。某个点即使拥有最高 tok/s，也可能因为交互性过低而不适合实时产品；有效比较应在符合业务需求的延迟或交互性目标下进行。',
     benchmarkContext:
       'InferenceX 将吞吐量与交互性放在完整并发扫描中共同展示，并用 Pareto 前沿剔除两个轴上都更差的运行点。',
-    measurement: { label: '常用单位', value: 'token/秒/Chip（tok/s/Chip）' },
+    measurement: { label: '常用单位', value: 'tok/s/chip' },
   },
   interactivity: {
     term: '交互性',
@@ -77,7 +160,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '不同产品需要不同运行点。语音和交互式编程要求较高 token 速率，离线摘要则可以牺牲交互性换取更高总吞吐量；在交互性不一致时比较硬件很容易得出误导性结论。',
     benchmarkContext:
-      'InferenceX 将 tok/s/user 与吞吐量或成本一起绘制，并在等交互性表格中沿各自 Pareto 前沿插值，以固定用户体验。',
+      'InferenceX 将 tok/s/user 与吞吐量或成本一起绘制，并在等交互性表格中沿各自 Pareto 前沿插值，以固定用户体验。由于该坐标轴不计入首 token 之前的等待，agentic 图表还提供端到端归一化交互性，用同一单位把 TTFT 一并纳入。',
     measurement: { label: '常用单位', value: 'token/秒/用户（tok/s/user）' },
   },
   latency: {
@@ -135,14 +218,14 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     term: '批处理',
     aliases: ['batching', '连续批处理', '动态批处理'],
     plainEnglish:
-      '批处理就像让多名乘客坐同一辆巴士：Chip 一次处理多个请求，让每趟计算完成更多有效工作。',
+      '批处理就像让多名乘客坐同一辆巴士：芯片一次处理多个请求，让每趟计算完成更多有效工作。',
     definition: '批处理将多个请求的工作组合起来，使加速器能够一起处理它们的 token。',
     explanation:
-      '大型矩阵运算比大量微小运算更能发挥 Chip 效率。现代推理引擎采用连续批处理，请求到达和结束时动态加入或退出，无需等待固定批次全部完成。',
+      '大型矩阵运算比大量微小运算更能发挥芯片效率。现代推理引擎采用连续批处理，请求到达和结束时动态加入或退出，无需等待固定批次全部完成。',
     significance:
       '批处理是吞吐量与延迟核心权衡的来源。更大的有效批次能摊薄权重读取和内核启动开销，但通常会增加每位用户的 token 间隔。',
     benchmarkContext:
-      '并发量是批处理的输入，并不等同于某个固定内核批大小；并行策略、序列长度、请求完成时机和调度策略都会改变 Chip 实际看到的批形状。',
+      '并发量是批处理的输入，并不等同于某个固定内核批大小；并行策略、序列长度、请求完成时机和调度策略都会改变芯片实际看到的批形状。',
   },
   'pareto-frontier': {
     term: 'Pareto 前沿',
@@ -155,7 +238,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '前沿能避免噪声点或调优较差的点扭曲比较，并展示真正的权衡。沿曲线仍不存在普适赢家，最佳点取决于用户最低交互性或最高成本目标。',
     benchmarkContext:
-      'InferenceX 连接并发与配置扫描中的 Pareto 最优点，等交互性比较也沿这些前沿插值，避免用随意选择的原始点直接比较。',
+      'InferenceX 连接并发与配置扫描中的 Pareto 最优点，等交互性比较也沿这些前沿插值，避免用随意选择的原始点直接比较。每个 SKU 最佳配置视图现在会把允许启用的优化合并成一条曲线，按模型、芯片 SKU 和引擎各一条，因此同一条线上相邻的点可能在投机解码、分离式部署或 KV cache offload 上并不相同；每个点仍会展示产生它的具体配置。',
   },
   'iso-interactivity': {
     term: '等交互性',
@@ -167,7 +250,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '固定用户体验可以避免常见基准错误：某系统只有在让每个请求更慢时才达到更高吞吐量，却被错误地称为更快。',
     benchmarkContext:
-      'InferenceX 文章使用等交互性表格比较硬件、精度和软件；超出实测前沿的值会标记为不可达，而不会向观测区间之外外推。',
+      'InferenceX 文章使用等交互性表格比较硬件、精度和软件；超出实测前沿的值会标记为不可达，而不会向观测区间之外外推。前沿始终建立在吞吐量与交互性之上，每百万 token 成本和每 token 焦耳则由插值得到的吞吐量推导，而不是各自单独做样条：它们都是每芯片常数除以吞吐量，单独插值会破坏两个 knot 之间的这一恒等关系。',
   },
   'input-output-sequence-length': {
     term: '输入与输出序列长度',
@@ -181,7 +264,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '不同序列长度的结果不能直接互换。短聊天提示词上的最佳配置，在长上下文摘要或推理中可能排名不同，因为计算、容量与带宽压力都会变化。',
     benchmarkContext:
-      'InferenceX 在图表标签与方案描述中列出 ISL/OSL。只有先匹配工作负载形状，才能把差异归因于硬件或软件。',
+      'InferenceX 在图表标签与方案描述中列出 ISL/OSL。只有先匹配工作负载形状，才能把差异归因于硬件或软件。agentic 运行没有单一的 ISL/OSL 组合可供引用：长度近似服从对数正态分布，因此点详情视图会绘制拟合分布并给出分位数，p90 或 p99 输入长度可能是中位数的数倍。',
   },
   'cost-per-million-tokens': {
     term: '每百万 token 成本',
@@ -196,7 +279,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
       '成本曲线使用与吞吐曲线相同的并发扫描。在等交互性下，更低的 $/M 表示以更少建模成本提供相同流式体验。',
     measurement: {
       label: 'InferenceX 计算式',
-      value: '$/M = TCO（$/Chip 小时）× 1,000,000 /（3600 × tok/s/Chip）',
+      value: '$/M = TCO（$/chip-hour）× 1,000,000 /（3600 × tok/s/chip）',
     },
   },
   'performance-per-dollar': {
@@ -209,7 +292,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '芯片峰值 FLOPS 不能单独决定服务经济性；内存、网络、软件成熟度、数值精度和实际利用率都会影响最终比值。',
     benchmarkContext:
-      'InferenceX 在匹配交互性时比较 perf/$，并明确使用的 TCO 输入。该比值不能跨模型、序列长度、精度或延迟区间直接套用。',
+      'InferenceX 在匹配交互性时比较 perf/$，并明确使用的 TCO 输入。该比值不能跨模型、序列长度、精度或延迟区间直接套用。图表用每美元 token 数表达同一套经济性，它数值越大越好，也是默认的 Y 轴。',
   },
   'total-cost-of-ownership': {
     term: '总体拥有成本',
@@ -217,11 +300,11 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     plainEnglish: 'TCO 包含硬件采购，以及后续供电、制冷、网络和运维成本。',
     definition: '总体拥有成本（TCO）是基础设施在使用寿命内采购、部署和运营的综合成本估算。',
     explanation:
-      'Chip 采购价只是其中一项。TCO 模型还可包含主机、网络、供电、制冷、机房、融资、折旧、维护和预期利用率，并归一化为每 Chip 小时成本。',
+      '芯片采购价只是其中一项。TCO 模型还可包含主机、网络、供电、制冷、机房、融资、折旧、维护和预期利用率，并归一化为每芯片小时成本。',
     significance:
       'TCO 比标价更适合跨系统经济性比较，尤其是网络与电力基础设施不同的机架级产品；但它仍是模型，必须连同假设一起阅读。',
     benchmarkContext:
-      'InferenceX 将 SemiAnalysis AI Cloud TCO 输入与实测 tok/s/Chip 结合，从而区分每小时系统成本和决定该小时 token 产出的软硬件行为。',
+      'InferenceX 将 SemiAnalysis AI Cloud TCO 输入与实测 tok/s/chip 结合，从而区分每小时系统成本和决定该小时 token 产出的软硬件行为。',
   },
   'tokens-per-megawatt': {
     term: '每兆瓦 token 吞吐量',
@@ -233,7 +316,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '电力供应往往是新增 AI 部署的硬约束。每兆瓦生成更多 token 的系统，即使单个加速器功耗更高，也能在相同电力配额下服务更多需求。',
     benchmarkContext:
-      '比较 tokens/MW 时必须匹配模型、工作负载、精度与交互性，否则高吞吐低交互点可能看似高效，却无法满足目标用户体验。',
+      '比较 tokens/MW 时必须匹配模型、工作负载、精度与交互性，否则高吞吐低交互点可能看似高效，却无法满足目标用户体验。每 token 能耗表达的是同一份供电预算折算到单位输出上的结果；在遥测可信的前提下，InferenceX 还会给出加速器的实测能耗。',
     measurement: { label: '常用单位', value: '每单位配置市电兆瓦的 token/秒' },
   },
   prefill: {
@@ -246,7 +329,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '预填充与解码的资源特征不同。两者共享工作节点时，大型提示词任务会打断解码批次，使流式延迟更不稳定。',
     benchmarkContext:
-      '分离式方案将预填充放在独立 Chip 池。阅读结果时应检查预填充 TP、Chip 数、输入长度，以及 KV 状态是否需要跨网络传输到解码池。',
+      '分离式方案将预填充放在独立芯片池。阅读结果时应检查预填充 TP、芯片数、输入长度，以及 KV 状态是否需要跨网络传输到解码池。',
   },
   decode: {
     term: '解码',
@@ -258,7 +341,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '解码决定流式交互性，也常主导长输出成本。推测解码、MTP、量化和宽专家并行都试图减少每个有效 token 的工作量或耗时。',
     benchmarkContext:
-      'InferenceX 用 tok/s/user 与总 tok/s/Chip 展示不同并发下的解码性能。公平比较必须匹配输出长度、批形状、精度和并行策略。',
+      'InferenceX 用 tok/s/user 与总 tok/s/chip 展示不同并发下的解码性能。公平比较必须匹配输出长度、批形状、精度和并行策略。',
   },
   'kv-cache': {
     term: 'KV 缓存',
@@ -271,7 +354,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       'KV 缓存压力限制并发与长上下文服务。缓存量化、分页分配、潜在注意力、前缀复用和分离式传输系统都在降低其容量或移动成本。',
     benchmarkContext:
-      '除非方案另有说明，InferenceX 在随机数据比较中禁用前缀缓存，避免无关请求因偶然命中而获得不真实优势。',
+      '除非方案另有说明，InferenceX 在定长场景的随机数据比较中禁用前缀缓存，避免无关请求因偶然命中而获得不真实优势。AgentX 是有意为之的例外：它回放的会话本身就会复用前缀，因此缓存容量、淘汰策略和 offload 都属于该场景要测量的内容。',
   },
   'prefix-caching': {
     term: '前缀缓存',
@@ -284,15 +367,15 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '具有重复前缀的生产工作负载可能明显快于随机 token 基准；收益取决于命中率、缓存容量、淘汰策略与请求能否路由到持有所需状态的节点。',
     benchmarkContext:
-      'InferenceX 通常在随机数据集上禁用前缀缓存，避免把缓存策略混入完整提示词处理的测量。除非明确说明，应把结果视为无命中基线。',
+      'InferenceX 在定长场景的随机数据集上禁用前缀缓存，避免把缓存策略混入完整提示词处理的测量，因此那些数字应视为无命中基线。AgentX 则相反：命中率本身就是上报指标，会与该点所在的 offload 层级一并展示，因为复用正是这类工作负载的本质特征。',
   },
   'disaggregated-inference': {
     term: '分离式推理',
     aliases: ['disaggregated inference', 'PD 分离', '分离式预填充', 'disagg'],
-    plainEnglish: '分离式推理把“读提示词”和“写答案”交给两组 Chip，让每组都能针对自己的任务优化。',
+    plainEnglish: '分离式推理把“读提示词”和“写答案”交给两组芯片，让每组都能针对自己的任务优化。',
     definition: '分离式推理在不同工作池上运行预填充与解码，并在两者之间传输请求状态。',
     explanation:
-      '预填充通常偏计算密集，解码则常受内存带宽和通信限制。分离后，两侧可以采用不同 Chip 数、并行度、批策略和扩缩容方式。',
+      '预填充通常偏计算密集，解码则常受内存带宽和通信限制。分离后，两侧可以采用不同芯片数、并行度、批策略和扩缩容方式。',
     significance:
       '分离可隔离提示词峰值并提升吞吐量或 SLA 稳定性，但也增加路由与 KV 传输开销；网络薄弱或内核不成熟时，它可能反而慢于聚合式服务。',
     benchmarkContext:
@@ -310,7 +393,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       '加速取决于草稿 token 的接受数量，以及草稿与验证成本。稠密模型和 MoE 的表现可能不同，因为验证多个位置可能激活更多专家权重。',
     benchmarkContext:
-      '应在真实接受率下比较推测方案并验证模型质量。InferenceX 分开展示开启和关闭 MTP 的曲线，因为收益会随并发与交互性变化。',
+      '应在真实接受率下比较投机解码方案并验证模型质量。定长场景仍把投机解码作为曲线标识的一部分，因此开启和关闭 MTP 的方案会分开绘制；agentic 曲线则把它当作数据点级元数据并合并这些点，在 tooltip 中标明具体方式，因为 AgentX 按模型、芯片 SKU 和引擎给出可获得的最佳曲线。由于 AgentX 回放的内容是合成的，speculator 接受的 draft token 数会失真，因此运行时会套用一套按模型、speculator、draft 长度和思考模式在外部 agentic 编码数据集上采集的接受长度。',
   },
   'multi-token-prediction': {
     term: '多 token 预测',
@@ -341,7 +424,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   'tensor-parallelism': {
     term: '张量并行',
     aliases: ['tensor parallelism', 'TP'],
-    plainEnglish: '张量并行把一次大型计算拆给多张 Chip，让它们共同完成。',
+    plainEnglish: '张量并行把一次大型计算拆给多张芯片，让它们共同完成。',
     definition: '张量并行（TP）把单个张量运算和模型权重矩阵切分到多个加速器上。',
     explanation:
       '每一层由多个 rank 协同执行，部分结果需要通过集体通信合并，常见方式是在并行矩阵乘之后执行 all-reduce。',
@@ -353,13 +436,13 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   'expert-parallelism': {
     term: '专家并行',
     aliases: ['expert parallelism', 'EP'],
-    plainEnglish: '专家并行把模型中的不同“专家”分配给不同 Chip，再把每个 token 送到需要的专家。',
+    plainEnglish: '专家并行把模型中的不同“专家”分配给不同芯片，再把每个 token 送到需要的专家。',
     definition:
       '专家并行（EP）把 MoE 模型的专家分布到不同加速器，并将 token 路由到持有所选专家的 rank。',
     explanation:
-      'MoE 层对每个 token 只激活部分专家。EP 利用这种稀疏性，避免每张 Chip 存储和计算全部专家，但每个 MoE 层前后都要执行 dispatch 与 combine all-to-all。',
+      'MoE 层对每个 token 只激活部分专家。EP 利用这种稀疏性，避免每张芯片存储和计算全部专家，但每个 MoE 层前后都要执行 dispatch 与 combine all-to-all。',
     significance:
-      '更宽 EP 能减少每 Chip 专家权重占用，并改善解码批处理与容量；收益取决于路由均衡和互连能否足够快地移动 token。',
+      '更宽 EP 能减少每芯片专家权重占用，并改善解码批处理与容量；收益取决于路由均衡和互连能否足够快地移动 token。',
     benchmarkContext:
       'InferenceX 将 EP 宽度列为分布式方案的一部分。NVL72 可让远宽于传统八卡节点的专家组保持在 NVLink scale-up 域内。',
   },
@@ -379,20 +462,19 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   'wide-expert-parallelism': {
     term: '宽专家并行',
     aliases: ['wide expert parallelism', 'Wide EP'],
-    plainEnglish:
-      '宽专家并行把模型专家铺到大量 Chip 上，让每张 Chip 需要保存和移动的专家数据更少。',
+    plainEnglish: '宽专家并行把模型专家铺到大量芯片上，让每张芯片需要保存和移动的专家数据更少。',
     definition: '宽专家并行使用大量加速器 rank 构成 MoE 模型的专家并行组。',
     explanation:
-      '把数百个专家分散到更多 rank，可减少每张 Chip 需要存储和流式读取的专家权重；更大的同伴组也可形成更高效的专家批次，但 dispatch/combine 流量会扩展。',
+      '把数百个专家分散到更多 rank，可减少每张芯片需要存储和流式读取的专家权重；更大的同伴组也可形成更高效的专家批次，但 dispatch/combine 流量会扩展。',
     significance:
       'Wide EP 在高带宽 scale-up 网络中最有效。若流量跨越较慢的 scale-out 网络，同样的 all-to-all 可能成为瓶颈并抵消内存侧收益。',
     benchmarkContext:
-      'InferenceX 在机架级分离式方案中使用 Wide EP。比较时必须同时查看 EP 宽度、解码池大小与网络，而不能只看图例中的 Chip 型号。',
+      'InferenceX 在机架级分离式方案中使用 Wide EP。比较时必须同时查看 EP 宽度、解码池大小与网络，而不能只看图例中的芯片型号。',
   },
   'all-reduce': {
     term: 'All-reduce',
     aliases: ['全归约'],
-    plainEnglish: 'All-reduce 让每张 Chip 完成一部分计算，再合并结果并把完整答案发回所有 Chip。',
+    plainEnglish: 'All-reduce 让每张芯片完成一部分计算，再合并结果并把完整答案发回所有芯片。',
     definition:
       'All-reduce 是一种集体通信操作：合并所有参与 rank 的值，并把归约结果返回给每个 rank。',
     explanation:
@@ -405,32 +487,32 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   'all-to-all': {
     term: 'All-to-all',
     aliases: ['全交换'],
-    plainEnglish: 'All-to-all 是一次有组织的交换：每张 Chip 都向其他每张 Chip 发送不同的数据包。',
+    plainEnglish: 'All-to-all 是一次有组织的交换：每张芯片都向其他每张芯片发送不同的数据包。',
     definition: 'All-to-all 是每个参与 rank 向所有其他 rank 发送不同数据的集体通信模式。',
     explanation:
       '专家并行 MoE 层先用 all-to-all dispatch 把 token 发往所选专家，再用 combine 把专家输出送回；流量与不均衡程度取决于 token 路由。',
     significance:
       'All-to-all 比简单点对点传输更苛刻，EP 扩大后容易受网络限制。专用内核会重叠通信与计算并优化 token 打包。',
     benchmarkContext:
-      '机架级 NVLink 可让 Wide EP 的 all-to-all 留在 scale-up 域内；跨节点 InfiniBand 或 RoCE 方案需要面对远低得多的每 Chip scale-out 带宽。',
+      '机架级 NVLink 可让 Wide EP 的 all-to-all 留在 scale-up 域内；跨节点 InfiniBand 或 RoCE 方案需要面对远低得多的每芯片 scale-out 带宽。',
   },
   'scale-up-vs-scale-out': {
     term: 'Scale-up 与 scale-out 网络',
     aliases: ['纵向扩展域', '横向扩展网络'],
-    plainEnglish: 'Scale-up 是同一套 Chip 系统内部的超高速网络，scale-out 则连接不同服务器或机架。',
+    plainEnglish: 'Scale-up 是同一套芯片系统内部的超高速网络，scale-out 则连接不同服务器或机架。',
     definition:
       'Scale-up 网络连接同一紧耦合系统内的加速器，scale-out 网络则把多个系统或机架连接成更大集群。',
     explanation:
-      'NVLink 等 scale-up 网络为细粒度集体通信提供极高每 Chip 带宽和低延迟；InfiniBand 或 RoCE 等 scale-out 网络覆盖更多机器，但每加速器带宽通常更低。',
+      'NVLink 等 scale-up 网络为细粒度集体通信提供极高每芯片带宽和低延迟；InfiniBand 或 RoCE 等 scale-out 网络覆盖更多机器，但每加速器带宽通常更低。',
     significance:
       '分布式推理会跨越两个域。高频 TP/EP 集体通信尤其适合留在 scale-up 内，较粗粒度请求路由和部分预填充/解码传输则更能容忍 scale-out。',
     benchmarkContext:
-      'Chip 名称本身不能描述通信域。八卡节点中的 B200 与 GB200 NVL72 使用相关芯片，却拥有完全不同的 scale-up 组规模。',
+      '芯片名称本身不能描述通信域。八卡节点中的 B200 与 GB200 NVL72 使用相关芯片，却拥有完全不同的 scale-up 组规模。',
   },
   'high-bandwidth-memory': {
     term: '高带宽内存',
     aliases: ['high-bandwidth memory', 'HBM'],
-    plainEnglish: 'HBM 是紧挨 Chip 的一小池超高速内存，推理时模型权重和工作数据都要放在这里。',
+    plainEnglish: 'HBM 是紧挨芯片的一小池超高速内存，推理时模型权重和工作数据都要放在这里。',
     definition: '高带宽内存（HBM）是靠近加速器堆叠的内存，其带宽远高于传统服务器内存。',
     explanation:
       'HBM 存储模型权重、激活、工作区与 KV 缓存。容量决定哪些模型、批大小和并行布局能放入；带宽决定内存受限内核能多快读取这些状态。',
@@ -442,8 +524,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   'memory-bandwidth': {
     term: '内存带宽',
     aliases: ['memory bandwidth', 'HBM 带宽'],
-    plainEnglish:
-      '内存带宽就像向 Chip 计算单元供给数据的管道宽度；管道越宽，计算单元越不容易空等。',
+    plainEnglish: '内存带宽就像向芯片计算单元供给数据的管道宽度；管道越宽，计算单元越不容易空等。',
     definition: '内存带宽是数据在加速器内存与计算单元之间传输的速率。',
     explanation:
       '当移动所需字节比执行算术更耗时，内核就是内存带宽受限。LLM 解码经常处于该状态，因为每一步都要为较少的新 token 计算流式读取模型/专家权重和 KV 缓存。',
@@ -454,15 +535,15 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   },
   nvlink: {
     term: 'NVLink',
-    aliases: ['NVIDIA NVLink', 'Chip 高速互连'],
-    plainEnglish: 'NVLink 是 NVIDIA Chip 之间的高速公路，让多张 Chip 的协作远快于普通服务器网络。',
-    definition: 'NVLink 是 NVIDIA 用于 scale-up 域内 Chip 直接数据传输的高带宽加速器互连。',
+    aliases: ['NVIDIA NVLink', '芯片高速互连'],
+    plainEnglish: 'NVLink 是 NVIDIA 芯片之间的高速公路，让多张芯片的协作远快于普通服务器网络。',
+    definition: 'NVLink 是 NVIDIA 用于 scale-up 域内芯片直接数据传输的高带宽加速器互连。',
     explanation:
-      'NVSwitch 系统连接多个 NVLink 端点，使集体通信可覆盖八卡服务器，或在 NVL72 产品中覆盖 72 Chip 机架级域；该带宽不同于连接独立系统的 InfiniBand/Ethernet。',
+      'NVSwitch 系统连接多个 NVLink 端点，使集体通信可覆盖八卡服务器，或在 NVL72 产品中覆盖 72 芯片机架级域；该带宽不同于连接独立系统的 InfiniBand/Ethernet。',
     significance:
-      '大型 TP，尤其是 Wide EP，会在每个生成 token 上交换数据。把通信留在 NVLink 上，可让机架级方案显著快于通过 scale-out 连接的相似 Chip 数量。',
+      '大型 TP，尤其是 Wide EP，会在每个生成 token 上交换数据。把通信留在 NVLink 上，可让机架级方案显著快于通过 scale-out 连接的相似芯片数量。',
     benchmarkContext:
-      'InferenceX 同时比较节点级 Chip 与 NVL72。归因于单 Chip 算力前，应先理解系统拓扑与并行组宽度。',
+      'InferenceX 同时比较节点级芯片与 NVL72。归因于单芯片算力前，应先理解系统拓扑与并行组宽度。',
   },
   quantization: {
     term: '量化',
@@ -487,7 +568,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     significance:
       'FP8 在新一代 NVIDIA 与 AMD 加速器上支持广泛，常作为稳定低精度基线；真实性能取决于端到端内核覆盖，回退操作会抹平理论收益。',
     benchmarkContext:
-      'InferenceX 的 FP8 标签覆盖完整方案，检查点文件名只是其中一项。引擎、注意力后端、KV 缓存格式、Chip 代际和 MTP 设置都可能改变曲线。',
+      'InferenceX 的 FP8 标签覆盖完整方案，检查点文件名只是其中一项。引擎、注意力后端、KV 缓存格式、芯片代际和 MTP 设置都可能改变曲线。',
   },
   fp4: {
     term: 'FP4',
@@ -567,21 +648,21 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   cuda: {
     term: 'CUDA',
     aliases: ['NVIDIA CUDA'],
-    plainEnglish: 'CUDA 是让程序在 NVIDIA Chip 上运行的软件工具箱。',
-    definition: 'CUDA 是 NVIDIA 的 Chip 计算平台、编程模型、编译工具链与软件库生态。',
+    plainEnglish: 'CUDA 是让程序在 NVIDIA 芯片上运行的软件工具箱。',
+    definition: 'CUDA 是 NVIDIA 的芯片计算平台、编程模型、编译工具链与软件库生态。',
     explanation:
-      'LLM 引擎使用 CUDA 内核和库执行矩阵乘、注意力、集体通信、图捕获、内存管理与融合操作；容器、驱动、CUDA 和 Chip 架构版本必须兼容。',
+      'LLM 引擎使用 CUDA 内核和库执行矩阵乘、注意力、集体通信、图捕获、内存管理与融合操作；容器、驱动、CUDA 和芯片架构版本必须兼容。',
     significance:
-      '服务性能取决于芯片之上的软件。新内核、CUDA Graph、编译器专用化和库版本都能在 Chip 不变时移动基准曲线。',
+      '服务性能取决于芯片之上的软件。新内核、CUDA Graph、编译器专用化和库版本都能在芯片不变时移动基准曲线。',
     benchmarkContext:
       'InferenceX 固定容器镜像，从而固定具体 CUDA 栈。历史比较可隔离仅更新引擎镜像对相同硬件与配置的影响。',
   },
   rocm: {
     term: 'ROCm',
     aliases: ['AMD ROCm'],
-    plainEnglish: 'ROCm 是让 AI 和高性能程序在 AMD Chip 上运行的软件工具箱。',
+    plainEnglish: 'ROCm 是让 AI 和高性能程序在 AMD 芯片上运行的软件工具箱。',
     definition:
-      'ROCm 是 AMD 的开放 Chip 计算软件平台，包含运行时、编译器、通信库及优化数学和 AI 内核。',
+      'ROCm 是 AMD 的开放芯片计算软件平台，包含运行时、编译器、通信库及优化数学和 AI 内核。',
     explanation:
       'vLLM 与 SGLang 通过 ROCm、AMD 专用库和内核项目在 Instinct 加速器上运行。模型支持取决于兼容的注意力、MoE、量化、集体通信与图执行路径。',
     significance:
@@ -592,7 +673,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   vllm: {
     term: 'vLLM',
     aliases: ['开源 LLM 推理引擎'],
-    plainEnglish: 'vLLM 是开源软件，通过组织请求和 Chip 内存，让语言模型高效服务大量用户。',
+    plainEnglish: 'vLLM 是开源软件，通过组织请求和芯片内存，让语言模型高效服务大量用户。',
     definition:
       'vLLM 是开源 LLM 推理与服务引擎，重点提供高吞吐调度、高效 KV 缓存管理和广泛模型/硬件支持。',
     explanation:
@@ -609,7 +690,7 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
       'SGLang 是用于快速服务语言模型的开源软件，提供面向复杂 AI 工作负载的调度和优化功能。',
     definition: 'SGLang 是面向高性能 LLM 与多模态推理的开源服务引擎和语言模型编程系统。',
     explanation:
-      '服务运行时包含连续批处理、前缀感知调度、分布式并行、推测解码，以及面向 NVIDIA/AMD Chip 的多种注意力和 MoE 内核后端。',
+      '服务运行时包含连续批处理、前缀感知调度、分布式并行、推测解码，以及面向 NVIDIA/AMD 芯片的多种注意力和 MoE 内核后端。',
     significance:
       'SGLang 快速迭代的版本和模型专用内核可在硬件不变时显著改变吞吐量；低并发受调度开销影响，其他区间则由注意力、MoE 与通信内核主导。',
     benchmarkContext:
@@ -618,9 +699,9 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
   'tensorrt-llm': {
     term: 'TensorRT-LLM',
     aliases: ['TRT-LLM', 'TRTLLM'],
-    plainEnglish: 'TensorRT-LLM 是 NVIDIA 为自家 Chip 优化的 LLM 推理软件栈。',
+    plainEnglish: 'TensorRT-LLM 是 NVIDIA 为自家芯片优化的 LLM 推理软件栈。',
     definition:
-      'TensorRT-LLM 是 NVIDIA 用于在 NVIDIA Chip 上编译、优化和服务大语言模型的推理软件栈。',
+      'TensorRT-LLM 是 NVIDIA 用于在 NVIDIA 芯片上编译、优化和服务大语言模型的推理软件栈。',
     explanation:
       '它提供 NVIDIA 优化内核、量化路径、分布式执行和模型专用优化；既可作为服务后端，也可通过集成让其他引擎使用其衍生内核。',
     significance:
@@ -632,16 +713,887 @@ const translations: Readonly<Record<string, GlossaryTranslation>> = {
     term: 'NVIDIA Dynamo',
     aliases: ['Dynamo', '分布式推理框架'],
     plainEnglish:
-      'NVIDIA Dynamo 协调大量 Chip worker，负责路由请求、移动模型记忆，并把读提示词和写答案分配给合适的资源池。',
+      'NVIDIA Dynamo 协调大量芯片 worker，负责路由请求、移动模型记忆，并把读提示词和写答案分配给合适的资源池。',
     definition:
       'NVIDIA Dynamo 是用于编排请求路由、worker 池、KV 缓存移动和分离式服务的分布式推理框架。',
     explanation:
       'Dynamo 可把预填充与解码放在独立扩展的池中，并使用 vLLM 或 TensorRT-LLM 作为 worker 运行时。内核仍由这些引擎执行，Dynamo 负责外围数据与控制路径。',
     significance:
-      '机架级性能由单 Chip 运行时、路由、缓存传输、拓扑感知与池大小共同决定。这些因素决定 Wide EP 和分离式推理能否提升端到端性能。',
+      '机架级性能由单芯片运行时、路由、缓存传输、拓扑感知与池大小共同决定。这些因素决定 Wide EP 和分离式推理能否提升端到端性能。',
     benchmarkContext:
       'Dynamo vLLM、Dynamo TRT-LLM 标签同时标识编排层与执行引擎。InferenceX 文章还会明确预填充/解码拓扑，因为两种 Dynamo 配置可能表现完全不同。',
   },
+  'e2e-normalized-interactivity': {
+    term: '端到端归一化交互性',
+    aliases: ['E2E Normalized Interactivity', '归一化交互性', 'OSL/E2EL'],
+    plainEnglish:
+      '这个指标衡量一整条回答到达得有多快，既算首 token 之前的等待，也算之后的流式速度。',
+    definition:
+      '端到端归一化交互性是整条请求上的有效每用户 token 速率，即输出 token 数除以端到端延迟。',
+    explanation:
+      '把端到端延迟展开为 TTFT 加上输出长度乘以每输出 token 时间，该指标约等于 1 除以（token 间延迟加上 TTFT 除以输出 token 数），也就是常规交互性再加一项与首 token 等待成正比的惩罚。这里的“归一化”指按输出长度归一，既不是 0 到 1 的评分，也不是与其他系统对比后的相对值。',
+    significance:
+      '只看交互性，会奖励那些让用户先等很久、之后再快速输出的方案；只看 TTFT，则会奖励开头很快、随后拖沓的方案。把两者合成一个数字，可以暴露只在单一维度上好看的运行点。输出越短，TTFT 惩罚越明显，因为可供摊薄等待时间的 token 更少。',
+    benchmarkContext:
+      'InferenceX 把它作为 agentic 运行的实验性 X 轴模式，因此需要持久化的逐请求 trace，非官方运行 overlay 无法使用该视图。它是有意保留缺陷的：对高 TTFT 惩罚很重，也无法体现 prefill 与 decode 分离等优化的全部细节，所以 AgentX 提交仍会分别针对交互性和 TTFT 做优化。',
+    measurement: { label: '常用单位', value: 'token/秒/用户（tok/s/user）' },
+  },
+  'tokens-per-dollar': {
+    term: '每美元 token 数',
+    aliases: ['tokens per dollar', 'tok/$', '每 1 美元 token 数'],
+    plainEnglish: '每美元 token 数表示一美元基础设施支出能买到多少 token，数值越大说明系统越便宜。',
+    definition:
+      '每美元 token 数是某个配置在一单位建模成本下产出的 token 数量，即每 token 成本的倒数。',
+    explanation:
+      '它由每芯片吞吐量和建模的每芯片小时成本直接得出，因此与每百万 token 成本共用同一套假设，只是换成了人们规划容量时更习惯的方向。InferenceX 为总 token、输入 token 和输出 token 分别给出该指标，覆盖每种成本口径，并同时提供人民币与美元两种计价。',
+    significance:
+      '每百万 token 成本与每美元 token 数对系统的排序完全一致，但后者随硬件变好而升高，与吞吐量方向相同，因此同一张图里的坐标轴不会中途反向。该数值完全依赖背后的成本模型，脱离所声明的口径就不成立。',
+    benchmarkContext:
+      'InferenceX 推理图表默认的 Y 轴就是每 1 美元可购买的总 token 数。阅读时请对照图表上方的 TCO 行，并只在同一成本口径内比较：自有（超大规模费率）、自有（neocloud 费率）和 3 年租赁对同一颗芯片会给出不同结果。',
+    measurement: { label: '常用单位', value: '每 1 美元 token 数（tok/$）' },
+  },
+  'energy-per-token': {
+    term: '每 token 能耗',
+    aliases: ['energy per token', 'J/token', '每请求能耗'],
+    plainEnglish:
+      '每 token 能耗表示系统产出一个 token 要花多少电，是每 token 成本在电力侧的对应指标。',
+    definition:
+      '每 token 能耗是产出单位 token 所消耗的电能，可以按全量供电口径统计，也可以取加速器实测遥测值。',
+    explanation:
+      '两种口径回答的是不同问题，不能混用。全量供电口径把包含供电与制冷开销在内的设施功耗预算除以实测 token 速率；实测口径来自运行期间的加速器遥测，只覆盖芯片本身。InferenceX 还会给出每次成功请求的实测能耗，以及平均功耗占 TDP 的百分比。',
+    significance:
+      '新建部署的约束往往是电力而不是资本开支，每焦耳产出更多 token 的系统能在同样的用电额度下服务更多需求。TDP 占比则单独反映一套方案究竟把加速器压到了什么程度，这是仅看按 token 归一的数值看不出来的。',
+    benchmarkContext:
+      '比较前请先看清标签：全量供电口径与实测口径之间相差整个设施开销。当底层遥测无效或统计范围不明确时，InferenceX 会屏蔽实测能耗，因此数值缺失意味着该测量不可信，而不是这次运行不耗电。',
+    measurement: { label: '常用单位', value: '焦耳/token（J/tok）' },
+  },
+  'context-parallelism': {
+    term: '上下文并行',
+    aliases: ['context parallelism', 'PCP', 'DCP', '序列并行'],
+    plainEnglish:
+      '上下文并行把一条长提示词拆到多颗芯片上，让它们分担读取提示词和扫描注意力状态的工作。',
+    definition:
+      '上下文并行把 query token 切分到多个加速器上，用于 prefill 的形式称为 PCP，用于 decode 的形式称为 DCP。',
+    explanation:
+      'PCP 让每个 rank 处理一段 query，keys 和 values 以环形方式传递，从而并行化计算受限的 prefill，也避免某个 rank 独自承担整条长提示词。DCP 则切分 KV cache 本身，每个 rank 扫描各自分片，再以 flash-decode 方式合并部分注意力结果；由于 decode 受显存带宽限制，并行读取 KV 可以提高可达 token 速率。',
+    significance:
+      '张量并行会在每个 rank 上复制完整 KV cache，数据并行注意力则把会话绑定在持有其分片的 rank 上，两者在上下文达到几十万 token 时都难以扩展。上下文并行直接针对这一点，而且收益随输入长度增长，而不是随 batch 大小增长。',
+    benchmarkContext:
+      'InferenceX 在数据点 tooltip 和并行标签中与 TP、EP、DP 一起展示 DCP 与 PCP 的并行度。各厂商支持程度并不均衡：在 AgentX 1.0 结果发布时，vLLM 支持矩阵中 AMD 的注意力后端仍标为不支持，因此该技术仍构成 CUDA 实际优势的一部分。',
+  },
+  'kv-cache-offload': {
+    term: 'KV cache offload',
+    aliases: ['KV cache offload', 'CPU offload', 'KV 卸载'],
+    plainEnglish:
+      '把芯片放不下的注意力状态暂存到主机内存，长会话下一轮就能直接恢复，而不必整段重算。',
+    definition:
+      'KV cache offload 把 KV block 从加速器显存移到更慢的存储层（通常是主机 DRAM），并在后续请求复用该前缀时再读回来。',
+    explanation:
+      'offload 通常实现为写穿缓存：写入 HBM 缓存的前缀会同时写入较慢的一层，因此当 offload 池容量大致是 HBM 的 1.5 到 3 倍时效果最好。在 agentic 的上下文长度下，重新载入长前缀远比重算划算；但对短提示词结论相反，传输开销会超过它省下的 prefill。',
+    significance:
+      '长 agentic 会话超出 HBM 中 KV 容量的时间，远早于超出合理的 DRAM 预算，因此 offload 决定了有多少并发对话仍可恢复。它也会转移瓶颈：前缀能够留存之后，store 与 load 路径、传输批量化和索引记账才是值得优化的开销。',
+    benchmarkContext:
+      'InferenceX 会给每个使用了 offload 的数据点加上虚线光环，无论它是否位于 Pareto 前沿；点详情视图还会给出 offload 类型、引擎，以及芯片与 CPU 两侧的缓存命中率。offload 属于允许但可选的优化，因此同一条曲线上可以同时存在启用和未启用的数据点。',
+  },
+  'kv-cache-manager': {
+    term: 'KV cache 管理器',
+    aliases: ['KV cache manager', 'Mooncake', 'LMCache', 'HiCache'],
+    plainEnglish:
+      'KV cache 管理器负责把注意力状态存放在芯片之外，并决定哪些保留、哪些淘汰、什么时候取回。',
+    definition:
+      'KV cache 管理器是位于推理引擎之下的可插拔层，负责跨存储层级保存可复用的 KV block，并管理其放置、淘汰与传输。',
+    explanation:
+      '引擎对外提供 connector 接口，因此 Mooncake Store、LMCache、SGLang HiCache 等管理器可以服务不同运行时。管理器按前缀哈希为 block 建立索引，把它们放在主机 DRAM、本地 NVMe 或远端后端；实际的字节搬运由 Mooncake Transfer Engine、NIXL 等传输引擎完成。同一个引擎内部可以并存多条路径。',
+    significance:
+      '一旦工作负载大量复用前缀，这一层的正确性和记账就和内核速度同等重要。混合注意力模型更难处理：一个模型同时携带形状和生命周期都不同的多个 cache group，而假设单一 block 几何结构的 connector 无法描述它们。',
+    benchmarkContext:
+      'InferenceX 把 KV offload 引擎记录为运行元数据，并在 AgentX 点详情视图中展示。框架标签给出的是组合而不是单个引擎，因此一套方案会显示为 Mooncake ATOMesh 或 MoRI SGLang，而不只是引擎名。',
+  },
+  'kv-aware-routing': {
+    term: 'KV 感知路由',
+    aliases: ['KV-aware routing', '缓存感知路由', '会话亲和性'],
+    plainEnglish: 'KV 感知路由把请求发给已经持有该会话状态的 worker，而不是发给当前最空闲的那个。',
+    definition: 'KV 感知路由依据已缓存前缀状态所在的位置来选择 worker，而不是只看队列深度或负载。',
+    explanation:
+      '不带可复用历史的请求发给谁都行，此时只需考虑负载均衡；但带着数 MB 已缓存前缀的请求不同，把它发给没有该前缀的空闲 worker，就要为整条提示词再付一次代价。因此 router 会跟踪 cache 事件、按会话哈希到固定 worker，并让数据并行 rank 对其持有状态的会话保持粘性。',
+    significance:
+      '在数据并行注意力下，每个 rank 只拥有缓存池的一部分，长会话一旦落到错误的 rank 上就要全部重算，实测命中率会远低于理论上限。仅有亲和性也不够：不加约束的粘性会把负载集中到单个热点 worker，因此缓存均衡也必须进入路由打分。',
+    benchmarkContext:
+      '路由位于引擎之外，因此 InferenceX 把它视为方案的一部分：Dynamo vLLM、llm-d vLLM、Mooncake ATOMesh 这类标签同时标明编排层和运行时。它的开销正比于在线前缀的数量和长度，而不是生成的 token 数，所以在内核变快之后，它可能成为 agentic 流量下的瓶颈。',
+  },
+  tilert: {
+    term: 'TileRT',
+    aliases: ['TileRT engine', 'TileRT 引擎'],
+    plainEnglish:
+      'TileRT 是面向极低延迟单用户生成的推理运行时，它把模型编译成一个常驻程序，而不是许多次内核启动。',
+    definition:
+      'TileRT 是一款推理引擎，通过取消以单个 kernel 作为执行单元的范式，来实现超低延迟服务。',
+    explanation:
+      '常规运行时每个 decode 步骤都要依次派发一串 kernel，而在极小 batch 下，kernel 之间的启动与调度开销会盖过实际算术运算。常驻的 engine kernel 把工作保留在加速器上，这正是交互性坐标轴最右端得以企及的原因。',
+    significance:
+      '前沿曲线的高交互性一端与高吞吐一端是不同的工程问题，为其中一端调优的引擎很少能同时拿下另一端。对延迟敏感的产品来说，能达到每用户每秒数百 token 的方案很有价值，即便它的单芯片总吞吐量并不突出。',
+    benchmarkContext:
+      'InferenceX 将 TileRT 作为独立的框架标签，并在每个 SKU 最佳配置视图中特意保留它；否则只按吞吐量取胜的曲线会丢掉 TileRT 真正服务的那些运行点。比较时请在匹配交互性下进行，而不要只看峰值吞吐量。',
+  },
+  recipe: {
+    term: '测试配置',
+    aliases: ['recipe', '配置方案', '服务配置'],
+    plainEnglish:
+      '一套测试配置就是产生某条曲线的全部选择：模型、引擎、镜像、精度、并行策略和工作负载。',
+    definition:
+      '测试配置是指产生一条实测曲线所需的完整组合：模型、推理引擎与容器镜像、数值精度、并行策略、芯片系统和工作负载。',
+    explanation:
+      'InferenceX 上的每个数据点都属于某套测试配置，对同一套配置做并发扫描就画出它的曲线。改动其中任何一项都会得到另一套配置，而不是同一套配置的变体，因此升级引擎镜像会作为独立结果上报，而不会并入既有曲线。',
+    significance:
+      '芯片峰值规格无法描述服务性能，同一颗芯片在不同配置下可以相差数倍。把整套组合写清楚，结论才可核查：脱离配置的单个数字既无法复现，也无法与其他厂商公平比较。',
+    benchmarkContext:
+      'InferenceX 的测试配置主要跟随 vLLM 与 SGLang 官方 cookbook，并使用上游镜像，因此结果反映用户实际能部署的性能，而不是为基准测试特调过的镜像。数据点 tooltip 会展示背后的配置，并给出运行溯源链接。',
+  },
+  'tail-latency': {
+    term: '尾部延迟',
+    aliases: ['tail latency', 'p90', 'p99', '分位数延迟'],
+    plainEnglish:
+      '尾部延迟描述的是最慢的那部分请求，而不是典型请求，因为用户真正会注意到的正是这些。',
+    definition: '尾部延迟是请求分布高分位处的延迟，例如 p90 或 p99，而不是均值或中位数。',
+    explanation:
+      '分位数回答的问题与平均值不同。p90 为 5 秒意味着每 10 条请求中至少有 1 条等了这么久，而这条请求可能正是卡住某个 agent 继续执行的那一条。真实服务中的延迟分布高度偏斜，均值可能远低于尾部，从而把问题完全掩盖。',
+    significance:
+      '容量规划通常按分位数而不是平均值来定，因为一个平均很快、尾部很慢的服务，对用户来说依然是不合格的。某些优化还会在改善均值的同时恶化尾部，而单一汇总数字只会把它记成一次明确的胜利。',
+    benchmarkContext:
+      'InferenceX 上报的指标都带分位数限定，并在坐标轴上标明，因此 p90 TTFT 与平均 TTFT 不会混在同一次比较里。agentic 运行的分布尤其偏斜，因为端到端延迟随输出长度增长，最长的生成会主导尾部。',
+  },
+  'service-level-objective': {
+    term: '服务级目标',
+    aliases: ['service level objective', 'SLO', '延迟目标'],
+    plainEnglish: 'SLO 是一套部署必须兑现的性能承诺，例如十条请求中有九条要在一秒内返回首 token。',
+    definition: '服务级目标是对某个服务指标设定的目标值，通常表示为对延迟或交互性的分位数约束。',
+    explanation:
+      '一条有用的 SLO 会同时给出指标、分位数和阈值。此时服务容量就是系统在不违反该约束的前提下能维持的吞吐量，它小于峰值吞吐量，也是运营方唯一可以据以规划容量的数字。',
+    significance:
+      '吞吐量曲线上的每个点都是可达的，但只有一部分满足既定承诺。两套系统的峰值吞吐量可能很接近，而在交互性或首 token 约束下能保留多少吞吐量却相差悬殊。',
+    benchmarkContext:
+      'InferenceX 不强加统一的行业 SLO，因为可接受的目标因产品而异：交互式编程需要较高的 token 速率，批处理则可以容忍数秒的首 token 延迟。请按自己的阈值读取前沿曲线，而不要直接比较峰值。',
+  },
+  'acceptance-length': {
+    term: '接受长度',
+    aliases: ['acceptance length', 'AL', '接受率'],
+    plainEnglish:
+      '接受长度是指每次验证中完整模型实际认可的草稿 token 数，它决定了投机解码是否划算。',
+    definition: '接受长度是目标模型在一次验证中平均接受的投机草稿 token 数量。',
+    explanation:
+      '只有草稿能通过验证，投机解码才会省时间。接受长度接近 1 意味着起草与验证这套机制白跑了一趟；数值较高则能把一次昂贵的目标模型计算摊薄到多个输出 token 上。该数值取决于 speculator、草稿长度、模型本身以及正在生成的内容。',
+    significance:
+      '正因为接受率取决于内容，基准测试有可能在无意中左右结论。合成或匿名化文本对于在真实语料上训练的 speculator 而言属于分布外数据，实测接受率会朝任一方向偏离生产环境的真实值。',
+    benchmarkContext:
+      'AgentX 回放的是填充了合成 token 的匿名化 trace，因此不让接受率由这些内容自行决定。运行时改为套用一组固定接受长度，按模型、speculator、草稿长度和思考模式在外部 agentic 编码数据集上采集，从而保持厂商中立。',
+  },
+  'unofficial-run': {
+    term: '非官方运行',
+    aliases: ['unofficial run', '非官方 overlay', '社区运行'],
+    plainEnglish: '非官方运行是尚未入库到已发布数据集的基准运行，但仍可通过 URL 叠加绘制在图表上。',
+    definition:
+      '非官方运行是指通过运行 ID 以 overlay 形式加载进仪表板的 CI 基准运行，其数据不来自已发布的数据库。',
+    explanation:
+      '在页面 URL 中加上运行 ID，就会拉取该次运行，并用专门的 overlay 配色把它的数据点、roofline 和表格行画在官方数据旁边。overlay 走的是独立的渲染路径，因此可以按硬件类型开关，也可以按运行逐个关闭，而不影响下方的官方数据选择。',
+    significance:
+      '结果发布通常滞后于产出，而正在调优某套配置的贡献者，需要在决定是否入库之前先看到新曲线与当前前沿的对比。overlay 让尚未定稿的运行可被审阅，同时不赋予它已发布测量结果的地位。',
+    benchmarkContext:
+      'overlay 的配色来自按运行序号分配的调色板，而不是硬件配色，因此不会被误认为官方数据。部分视图依赖只在入库运行中持久化的数据，这些视图会明确说明 overlay 不可用，而不是默默略过。',
+  },
+  'chunked-prefill': {
+    term: '分块 prefill',
+    aliases: ['chunked prefill', '分块预填充'],
+    plainEnglish:
+      '分块 prefill 把长提示词分片读取，而不是一次性读完，这样其他用户在此期间仍能持续收到 token。',
+    definition:
+      '分块 prefill 把提示词处理切分为固定大小的 token 分块，由调度器与正在进行的 decode 工作交错执行。',
+    explanation:
+      '不切分的 prefill 会占用加速器直到整条提示词处理完，所有正在流式输出的用户都得排在后面。切分之后调度器可以交替执行，让 decode 在分块之间继续推进。分块大小是一个调优旋钮：分块越大 prefill 效率越高，分块越小对 decode 的打断越少。',
+    significance:
+      '这项技术把某一个用户的首 token 问题，转换成对所有人来说小而平稳的开销，通常是更划算的取舍。提示词越长它越重要：单条不切分的十万 token prefill 足以让整套部署卡住。',
+    benchmarkContext:
+      '分块大小属于测试配置的一部分，同一条曲线上不同点的取值也可能不同，因此吞吐量的跳变有可能来自重新调参而不是新硬件。长的 agentic 提示词让这个参数变得关键，而定长短提示词场景根本暴露不出这一点。',
+  },
+  roofline: {
+    term: 'Roofline 曲线',
+    aliases: ['roofline', '前沿包络', 'roofline 曲线'],
+    plainEnglish:
+      '在 InferenceX 上，roofline 是按硬件配置在其最优点上画出的外包络线，用来展示它达到过的边界。',
+    definition:
+      'InferenceX 上的 roofline 是按硬件配置绘制的 Pareto 包络曲线，穿过该配置在当前所选坐标轴下未被支配的数据点。',
+    explanation:
+      '哪个角算“最好”取决于所选指标：吞吐量对交互性取右上角，而成本类指标数值越低越好，因此包络会锚定在所选指标组合定义的那个角上。x 值退化的点不具备入选资格，但在显示全部数据点的视图中仍会绘制。',
+    significance:
+      '直接看一团原始数据点容易挑选有利结果，因为调优不佳的配置也会贡献一些运营方绝不会选择的点。包络线展示的是每套系统可达的边界，而容量决策正是依据这个形状做出的。',
+    benchmarkContext:
+      '这与 HPC 中经典的 roofline 模型不同：后者把可达 FLOPS 与算术强度放在一起，用于判断计算受限还是访存受限，仪表板只借用了“上界”这一图形表达。roofline 方向按指标配置，因此同一批数据点在不同坐标轴下会得到不同的包络。',
+  },
+  'arithmetic-intensity': {
+    term: '算术强度',
+    aliases: ['arithmetic intensity', '运算强度', '计算访存比'],
+    plainEnglish: '算术强度表示每搬运一个字节能做多少次运算，它决定瓶颈落在计算单元还是显存上。',
+    definition: '算术强度是一次计算所执行的算术运算次数与其在显存和计算单元之间搬运的字节数之比。',
+    explanation:
+      'prefill 会用同一份权重处理大量 token，每个载入的字节都被反复复用，因此通常是计算受限的。decode 每一步每条序列只产出一个 token，却仍要读取权重和 KV cache，搬运了大量数据却只做很少运算，因此通常受显存带宽限制。',
+    significance:
+      '两个阶段受限于芯片的不同部分，这解释了为什么峰值 FLOPS 亮眼的规格书在 decode 上可能令人失望，也解释了 batching 为何有效：把多条序列合并会提高算术强度，因为每次权重读取被更多 token 复用。',
+    benchmarkContext:
+      '这一区别解释了数据中反复出现的形态。低交互性的点跑大 batch、算术强度高，接近计算上限；高交互性一端跑小 batch，跟随显存带宽，因此带宽更充裕的硬件常常在这里胜出，尽管其峰值吞吐量更低。',
+  },
+  'prefix-cache-hit-rate': {
+    term: '前缀缓存命中率',
+    aliases: ['prefix cache hit rate', '缓存命中率', 'KV 复用率'],
+    plainEnglish:
+      '命中率是指提示词中有多少 token 直接来自缓存而无需重算，在长会话中这通常是绝大部分。',
+    definition:
+      '前缀缓存命中率是 prefill 阶段由已缓存 KV 状态满足、而非重新计算的输入 token 占比。',
+    explanation:
+      '命中率必须与产生它的缓存层级一起看才有意义，因为从加速器显存读取的 token 与从主机内存取回的 token，代价完全不同。它也不只取决于容量：淘汰策略可能丢掉仍会被用到的前缀，路由也可能把请求发到从未持有该前缀的 worker。',
+    significance:
+      '在多轮流量下，命中率基本决定了 prefill 的开销，因为每一轮都会把整段对话再加一点新内容重新发过来。一旦复用率很高，剩余的 prefill 工作就主要是真正的新 token，瓶颈也从计算转向缓存管理。',
+    benchmarkContext:
+      'AgentX 点详情视图会按缓存层级分别展示命中率随时间的变化，并给出 prompt token 来源拆分。如果一个数据点在低命中率下报出很高的总吞吐量，说明它做的 prefill 工作远多于缓存良好的部署。',
+  },
+  warmup: {
+    term: 'Warmup',
+    aliases: ['warmup', 'warmup 阶段', '缓存预热'],
+    plainEnglish: '性能采集开始前先做一轮预热，让系统进入稳态，而不是在缓存为空的状态下被打分。',
+    definition:
+      'warmup 是性能采集窗口之前的阶段，基准测试在此期间预热缓存并进入稳态，该阶段的请求不计入上报结果。',
+    explanation:
+      '冷启动系统在两方面都不像生产环境：缓存是空的，而且所有会话都从第 0 轮开始会造成同步的突发流量，这在真实部署中并不存在。因此 AgentX 会让每个对话从其历史中一个按 seed 选定的位置开始，先回放重建该状态所需的请求，再让每条回放通道继续前进若干请求，然后才开始测量。',
+    significance:
+      '测量窗口从哪里开始会改变结果。把预热过程算进去，前缀复用会显得比生产环境更差；完全不做预热，依赖缓存的方案则会在一个它们在生产中根本不会遇到的状态下被打分。',
+    benchmarkContext:
+      'AgentX 点详情视图会把两个阶段分开，便于分别查看遥测数据。warmup 请求的输出被限制为单个 token，因此它们的输出长度约为 1，交互性和 decode 曲线为空：单个 token 没有 token 间延迟。',
+  },
+  aiperf: {
+    term: 'AIPerf',
+    aliases: ['AIPerf', '回放框架', '负载生成器'],
+    plainEnglish:
+      'AIPerf 是发送基准流量的厂商中立客户端，负责重建录制的 agent 会话并记录每条请求的时序。',
+    definition: 'AIPerf 是驱动 AgentX 运行、面向服务端点的开源 HTTP 负载生成与回放框架。',
+    explanation:
+      '它把每个会话重建为有向无环图：节点是请求，边携带依赖请求发出前需要等待的延迟。这一结构可以复现主智能体轮次、并行且稍后汇合的子智能体分支、一次性的辅助请求，以及轮次之间的工具调用停顿，而这些都是扁平的提示词列表无法表达的。',
+    significance:
+      '客户端本身就是测量的一部分。无法表达依赖关系的框架，会把 agentic 工作负载当成互相独立的请求发出，从而抹掉这个场景本要考察的突发性与复用特征。保持厂商中立同样重要，可以避免负载生成器偏向某一套服务栈。',
+    benchmarkContext:
+      'seed 固定了采样哪些对话、每个对话从哪里开始，以及用于填充匿名化 block 的合成内容，因此同一套配置重跑会回放同样的工作负载。同一模型的所有提交都运行相同的框架次版本，以保证结果可比。',
+  },
+  'pipeline-parallelism': {
+    term: '流水线并行',
+    aliases: ['pipeline parallelism', 'PP', '层间并行'],
+    plainEnglish: '流水线并行让每颗芯片负责一部分层，沿链条传递激活值，而不是把每一层都切开。',
+    definition:
+      '流水线并行按层切分模型并分配到多个加速器上，每个阶段执行自己的层，并把激活值传给下一个阶段。',
+    explanation:
+      '通信是阶段边界处激活值的点对点交接，比张量并行所需的逐层集合通信便宜得多。代价是空闲：只有一条请求在途时，除当前阶段外其余阶段都在等待，只有持续的并发流量才能把流水线填满。',
+    significance:
+      '对最大的那批模型来说，这首先是容量手段，其次才是提速手段。有些前沿模型根本装不进单个节点，流水线并行才让它们可服务；当某项竞争性优化拒绝与任何方案组合时，它甚至是唯一选项。',
+    benchmarkContext:
+      'InferenceX 在数据点 tooltip 和并行标签中与 TP、EP、DP 一起展示流水线并行度，且仅在大于 1 时显示。可组合性与并行度同样重要：一种会导致投机解码无法启用的阶段切分，代价可能超过它节省的显存。',
+  },
+  'dp-attention': {
+    term: '数据并行注意力',
+    aliases: ['DP attention', 'DPA', '注意力数据并行'],
+    plainEnglish:
+      'DP attention 让每个 rank 处理各自的一批序列并持有自己的缓存，而不是每个 rank 都存同一份副本。',
+    definition:
+      '数据并行注意力让每个 rank 在互不重叠的序列集合上各自计算注意力，因此每个 rank 拥有 KV cache 的私有份额，而专家层仍然共享。',
+    explanation:
+      '张量并行的注意力按 head 切分，在 head 数较少时会在各 rank 之间复制 KV 状态，浪费容量。DP attention 改为把整条序列分配给某个 rank，从而避免这种重复。各 rank 仍共同参与 MoE 集合通信，因此注意力是本地的，而专家分发仍是全局的。',
+    significance:
+      '由于每个 rank 只拥有缓存池的私有一份，请求落在哪里就成了影响性能的关键：长会话一旦被路由到不持有其前缀的 rank，就要全部重算。此时实测命中率会远低于理论上限，而原因与缓存大小毫无关系。',
+    benchmarkContext:
+      'InferenceX 会在数据点 tooltip 的并行策略部分展示 DP attention。它是否有利取决于模型；当缓存局部性变成路由约束时，不启用它的配置有时反而占据前沿曲线。',
+  },
+  int4: {
+    term: 'INT4',
+    aliases: ['INT4', '4 位整数', 'W4A16'],
+    plainEnglish:
+      'INT4 用 4 位整数存放权重，把模型压小，让缺少原生 4 位浮点支持的硬件每个 token 少搬很多数据。',
+    definition:
+      'INT4 是主要用于权重量化的 4 位整数格式，通常为每一组数值配一个更高精度的缩放因子。',
+    explanation:
+      '整数格式的取值是均匀分布的，不像浮点那样疏密不一，因此 INT4 依赖分组缩放因子来跟踪每一块权重的局部数值范围。激活值通常保持更高精度，矩阵乘法在计算时实时反量化，这使得该技术更多是访存优化而不是算力优化。',
+    significance:
+      '它在缺少 4 位浮点硬件支持的平台上最有价值。在这类硬件上，INT4 是落地 4 位权重的现实路径，但通常比原生格式更依赖细致的校准，其精度必须实测验证而不能想当然。',
+    benchmarkContext:
+      'InferenceX 把 INT4 与 FP4、FP8、BF16 并列为独立的精度键，精度属于测试配置而不是显示选项。把 INT4 与原生 FP4 方案对比时，务必同时查看精度评估结果，因为两种格式并不等价。',
+  },
+  bf16: {
+    term: 'BF16',
+    aliases: ['BF16', 'bfloat16', 'brain float 16'],
+    plainEnglish: 'BF16 是多数模型训练所用的 16 位格式，用它做推理也是量化方案精度对照的基准。',
+    definition:
+      'BF16 是一种 16 位浮点格式，指数范围与 FP32 相同而尾数更短，广泛用于训练，也常作为未量化的推理基线。',
+    explanation:
+      '保留 FP32 的指数范围，使 BF16 能很好地容纳 transformer 激活值的数值分布，格式转换很少需要窄格式所依赖的缩放机制。它牺牲的是精度而不是范围，体积是 FP8 的两倍、4 位格式的四倍。',
+    significance:
+      '在基准测试中它通常扮演参照系。decode 阶段以权重读取为主，因此 BF16 方案每个 token 搬运的数据远多于量化方案，往往位于吞吐量曲线更低的位置，同时定义了其他方案所对照的精度水平。',
+    benchmarkContext:
+      'InferenceX 把 BF16 作为精度键，并在规格页给出每个加速器的 BF16 稠密峰值算力。量化方案会通过精度评估验证，而不是默认无损，这正是 BF16 对照有意义的前提。',
+  },
+  'kv-cache-quantization': {
+    term: 'KV cache 量化',
+    aliases: ['KV cache quantization', 'FP8 KV cache', '量化 KV'],
+    plainEnglish: '把对话缓存用更小的格式存放，让芯片装下更多上下文，并在生成时更快地把它读回来。',
+    definition:
+      'KV cache 量化用降精度格式存放注意力的 key 与 value 状态，与模型权重所用精度相互独立。',
+    explanation:
+      '权重精度与缓存精度是两个独立选择，一套方案可以用 FP8 权重搭配 BF16 缓存，反过来也可以。缓存位宽减半，加速器显存能容纳的 token 数大致翻倍，每个 decode 步骤要读取的字节数也减半，而这正是 decode 大部分时间在做的事。',
+    significance:
+      '在长上下文服务中，这往往比压缩权重更划算，因为高并发下最先耗尽显存的是缓存而不是权重。不同模型对精度的敏感度不同，量化 key 还是 value 也有差别，因此需要实测评估而不能一概而论。',
+    benchmarkContext:
+      '缓存精度属于测试配置的一部分，实践中还存在混合布局：有些模型会保留两份位宽不同的缓存缓冲区，分离式传输路径必须把它们成对搬运。阅读显存容量相关的结论时，请连同其背后的缓存格式一起看。',
+  },
+  'sliding-window-attention': {
+    term: '滑动窗口注意力',
+    aliases: ['sliding window attention', 'SWA', '局部注意力'],
+    plainEnglish:
+      '滑动窗口注意力让某些层只看最近一段 token，因此窗口填满之后它的缓存就不再继续增长。',
+    definition:
+      '滑动窗口注意力把每个 query 限制在固定长度的前文范围内，从而限定该层需要保留的 KV 状态。',
+    explanation:
+      '由于窗口长度固定，这类层的缓存会达到上限而不会随对话增长，较早的条目会随窗口推进被移出。模型通常把这类层与全注意力层交错排列，这样远距离信息在网络中仍有通路，而大多数层保持低成本。',
+    significance:
+      '有界的开销带来了分配器层面的问题。窗口页不断翻转，而持久的前缀页静止不动；当两者取自同一个池子时，短暂分配往往会把有价值的那份挤出去，于是长会话可能因为短命的窗口状态而丢失昂贵的全注意力历史。',
+    benchmarkContext:
+      '这些效应只在多轮流量中出现。单条短提示词既不会让窗口绕过前缀，也不会造成池子争用，因此窗口感知的淘汰、offload 和分叉处理都表现为由 AgentX 推动的引擎工作，而不是定长场景的测试结果。',
+  },
+  'hybrid-attention': {
+    term: '混合注意力',
+    aliases: ['hybrid attention', '混合缓存模型'],
+    plainEnglish:
+      '混合模型在不同层使用不同的注意力类型，因此它的缓存是几种不同状态，而不是一块统一的数据。',
+    definition:
+      '混合注意力模型交错使用不同类型的注意力层，从而产生形状和生命周期各不相同的多个 KV cache group。',
+    explanation:
+      '统一模型每个 token 只有一种缓存布局，用单一 block 几何结构就能描述所有需要保存的内容；混合模型则不然：全注意力层、窗口层，以及循环或压缩状态同时存在，各有各的占用、各有各的丢弃时机，能否重建也不一样。',
+    significance:
+      '在状态需要离开加速器之前，这一区别是看不出来的。假设单一布局的 connector 无法说明某个 block 属于哪一组，因此会话最长的那些模型一度反而完全用不了 offload。循环状态最棘手，因为它累积了此前的全部内容，无法从相邻 token 重算。',
+    benchmarkContext:
+      'InferenceX 测试矩阵中的前沿开放权重模型越来越多采用混合结构，因此引擎对混合 cache group 的支持本身就是测试内容的一部分。offload、分离式传输和前缀复用都必须逐组扩展，而不能从统一结构的情形直接继承。',
+  },
+  'linear-attention': {
+    term: '线性注意力',
+    aliases: ['linear attention', 'GatedDeltaNet', '常数状态注意力'],
+    plainEnglish:
+      '线性注意力保存一份固定大小的运行摘要，而不是全部历史 token，因此内存不会随对话增长。',
+    definition:
+      '线性注意力用一个大小恒定、随 token 到达而更新的循环状态，替代不断增长的 key 与 value 缓存。',
+    explanation:
+      '标准注意力保存的状态与序列长度成正比，且每一步都要重新读取。线性或门控循环层改为携带固定大小的状态，用对每个位置的精确回忆换取有界的内存占用。GatedDeltaNet 这类结构只在部分层这样做，其余层保留全注意力，以维持精确的远距离查找能力。',
+    significance:
+      '在长上下文下，节省的存储相当可观，但这种状态改变了缓存的含义：它无法像窗口尾部那样由周围 token 重建，一旦丢弃就只能重放整个序列，因此复用需要显式的检查点机制，而不是普通的 block 复用。',
+    benchmarkContext:
+      'InferenceX 服务的模型中已有采用这类层的，引擎对循环状态的检查点与传输支持属于测试配置的一部分。一个模型可以在 day 0 就能被服务，却仍不支持这类状态的复用，表现为重复轮次上出乎意料的高 prefill 开销。',
+  },
+  nvl72: {
+    term: 'NVL72',
+    aliases: ['NVL72', 'GB200 NVL72', 'GB300 NVL72', '机架级系统'],
+    plainEnglish:
+      'NVL72 是一个机架，其中 72 个加速器共享同一张高速网络，因而更像一台大机器而不是一个集群。',
+    definition:
+      'NVL72 是 NVIDIA 的机架级系统，把 72 个加速器放进同一个 NVLink scale-up 域，而不是分散在多个八芯片节点中。',
+    explanation:
+      '仪表板的规格数据记录为 NVLink 5.0、每颗芯片 900 GB/s 单向带宽、scale-up world size 为 72，并通过 NVSwitch 交换。常规节点把同样的带宽限定在八颗芯片之间，超出后就要退回更慢的 scale-out 网络，因此差别不在于原始速度，而在于换用另一种网络之前能触及多少颗芯片。',
+    significance:
+      '开销以集合通信为主的技术，在大规模 scale-up 域中经济性会发生变化。宽专家并行把专家分散到许多芯片上，每个 token 都要付出 all-to-all 流量；在 scale-up 带宽下这可以承受，在 scale-out 网络上往往不行。',
+    benchmarkContext:
+      '机架级优势并非自动成立。更高的单芯片成本必须被赚回来，而在 agentic 流量下，编排层可能比网络更早成为瓶颈；因此对于不怎么用到宽并行的模型，NVL72 配置在按 TCO 归一化的吞吐量上有时会落后于八芯片节点。',
+  },
+  atom: {
+    term: 'ATOM',
+    aliases: ['ATOM', 'AMD ATOM', 'ATOMesh'],
+    plainEnglish:
+      'ATOM 是 AMD 自研的推理引擎，是它在厂商专用运行时这条路线上的答案，而非上游开源引擎。',
+    definition:
+      'ATOM 是 AMD 面向 Instinct 加速器的推理引擎，与 ROCm 上的上游 vLLM、SGLang 并列，定位为厂商自有运行时。',
+    explanation:
+      '它对 AMD 的意义，与厂商运行时对 NVIDIA 的意义相同：针对自家硬件调优，且可以跑在上游引擎前面。它的 router 名为 ATOMesh，最初是 SGLang router 的一个分支。该引擎最初面向单轮服务设计，因此支持长上下文多轮场景需要对其缓存管理器和 kernel 做大量改造。',
+    significance:
+      '厂商引擎可以在开源栈跟上之前展示硬件的能力上限，这既是有价值的证据，也是不便直接照搬的建议。多数实验室部署的是上游引擎，因此只在厂商运行时下成立的结果，并不能代表这些用户实际拿到的性能。',
+    benchmarkContext:
+      'InferenceX 把 ATOM 作为独立的框架标签，避免与同一加速器上的 vLLM 或 SGLang 结果混为一谈。想知道硬件能做到什么，就拿它与其他厂商运行时比较；想知道客户今天能部署什么，就拿它与上游引擎比较。',
+  },
+  aiter: {
+    term: 'AITER',
+    aliases: ['AITER', 'AMD AITER', 'ROCm kernel 库'],
+    plainEnglish:
+      'AITER 是 AMD 调优过的 kernel 库，决定注意力和矩阵运算在 Instinct 芯片上究竟能跑多快。',
+    definition:
+      'AITER 是 AMD 面向 Instinct 加速器的优化 kernel 库，供运行在 ROCm 上的推理引擎调用。',
+    explanation:
+      '引擎负责表达策略，kernel 决定它是否够快。AITER 提供调优过的注意力、矩阵和融合算子，引擎会用它替代通用路径。这个分发决策本身也可调优，而且在某种 shape 上取胜的 kernel 在另一种 shape 上可能落败，因此选择可能取决于上下文长度，而不是固定不变。',
+    significance:
+      '并行策略只有在 kernel 能表达时才算数，这正是 AMD 上的上下文并行和长上下文稀疏注意力以 kernel 工作而非引擎工作形式出现的原因。超大缓存还会暴露短请求根本触及不到的问题，例如缓存池跨过某个容量边界后地址运算溢出，从而静默访问错误的行。',
+    benchmarkContext:
+      '该库位于测试配置所固定的容器镜像内部，因此 AITER 的改进可以在引擎版本和硬件都不变的情况下推动曲线上移。在统一 shape 上实测到的 kernel 级收益，未必能在 agentic trace 上留存，缓存与调度带来的波动可能将其淹没。',
+  },
+  flashinfer: {
+    term: 'FlashInfer',
+    aliases: ['FlashInfer', '注意力 kernel 库'],
+    plainEnglish: 'FlashInfer 是一个注意力 kernel 库，服务引擎直接调用它，而不必自己实现注意力。',
+    definition:
+      'FlashInfer 是开源的注意力 kernel 与后端库，供推理引擎在 prefill、decode 和投机验证中使用。',
+    explanation:
+      '推理特有的复杂性大多集中在注意力上：分页缓存、变长序列、分组 query head、稀疏模式，以及对草稿 token 的验证，都会改变 kernel 的形态。共享库让多个引擎复用同一份调优实现，引擎则按 shape 和硬件目标选择后端。',
+    significance:
+      '由于后端是被选择而不是固定的，kernel 的可用性就变成了可移植性问题。只为某一家厂商后端实现的功能，会让另一方退回通用路径；而在长上下文下，这不是小小的妥协，而是选错了 kernel。',
+    benchmarkContext:
+      '所使用的后端属于测试配置的一部分，改动它可以在硬件和引擎版本都不变的情况下推动曲线变化。正是这些 kernel 支持了循环状态的检查点，混合模型才得以参与前缀复用。',
+  },
+  'cuda-graphs': {
+    term: 'CUDA graph',
+    aliases: ['CUDA graphs', 'graph capture', 'full-graph 模式'],
+    plainEnglish: 'CUDA graph 把一整串芯片操作录制一次并整体回放，从而免去逐个启动它们的开销。',
+    definition:
+      'CUDA graph capture 把一串 kernel 启动及其依赖关系录制成可回放的图，从而一次性提交整段序列，而不是逐个算子启动。',
+    explanation:
+      '一个 decode 步骤会发出许多小 kernel，在小 batch 下它们之间的启动与调度开销可以与实际运算相当。把整步录制下来即可消除这部分逐次启动的开销。代价是图是固定的：shape 必须稳定，因此引擎会按分桶分别录制，并把真正动态的部分留在图之外。',
+    significance:
+      '这更多是延迟优化而不是吞吐量优化，恰恰在 batch 很小、交互性很高的场景最有价值。它也会与一切改变 shape 的因素相互作用，因此变长的 agentic 流量可能击垮过度特化的运行时，使其几乎为每个请求重新编译。',
+    benchmarkContext:
+      'graph 的使用方式取决于测试配置所固定的引擎镜像，因此它可以在硬件不变的情况下推动曲线变化。有些方案会录制稳定的算子，同时让依赖请求的注意力保持 eager 执行，这是在录制覆盖率与 shape 灵活性之间的有意折中。',
+  },
+  goodput: {
+    term: 'Goodput',
+    aliases: ['goodput', '有效吞吐量', 'SLO 约束吞吐量'],
+    plainEnglish:
+      'Goodput 只统计满足延迟目标的那部分工作量，看起来很快但赶不上截止时间的系统拿不到任何分数。',
+    definition:
+      'Goodput 指满足既定服务等级目标（SLO）的那部分吞吐量，例如首 token 时间上限或每用户每秒最低 token 数。',
+    explanation:
+      '原始吞吐量只关心服务器完成了多少请求，不管每个用户被服务得多慢。Goodput 先做一层过滤：只有满足运营方承诺的延迟或交互性约束的请求才被计入。两套吞吐量完全相同的系统，加上截止时间后 goodput 可能相差很大，因为一套在高负载下仍能压住延迟，另一套则让排队把所有请求都拖过界。',
+    significance:
+      '忽略 goodput 的容量规划要么多买要么超卖。若运营方按峰值吞吐量报价，但只有一半流量能在 SLO 内完成，实际需要的集群规模就是模型估算的两倍。goodput 是把基准测试曲线换算成部署真实承载用户数的那个数字。',
+    benchmarkContext:
+      'InferenceX 发布完整的吞吐量对交互性 Pareto 前沿，而不是单一 goodput 数字，读者可以套用自己的 SLO。像 TCO 计算器那样在固定交互性档位上读取前沿，本质上就是在该档位测量 goodput。',
+  },
+  'model-flops-utilization': {
+    term: '模型算力利用率',
+    aliases: ['MFU', 'model FLOPs utilization', 'MBU'],
+    plainEnglish: 'MFU 把模型实际完成的有效运算与芯片理论峰值运算量相除，得到一个效率百分比。',
+    definition:
+      '模型算力利用率（MFU）是模型逻辑上需要的浮点运算量与硬件在相同墙钟时间内理论峰值运算量之比。',
+    explanation:
+      '峰值 TFLOP/s 假设每个 tensor core 每个周期都在满负荷工作，这在实际推理服务中从不成立。kernel 启动间隙、访存停顿、通信等待和不完美的批处理都会让算力闲置，MFU 把这些损耗折算进一个数字。decode 通常是访存受限的，因此 decode 阶段的 MFU 天然偏低，此时带宽版本的 MBU 往往是更真实的效率指标。',
+    significance:
+      'MFU 把硬件能力与软件成熟度区分开：峰值算力巨大但 kernel 不行的芯片，可能输给算力较低但喂得饱运算单元的芯片。硬件不变而 MFU 上升，正是软件进步的特征，而推理性能的大部分收益正来自软件。',
+    benchmarkContext:
+      'InferenceX 跟踪的是每颗芯片随时间交付的 token 数而非直接报告 MFU。硬件不变而性能大幅提升的反复出现的模式，例如模型发布数周内的数量级改进，本质上就是软件把利用率追讨回来的过程。',
+  },
+  'memory-bound-vs-compute-bound': {
+    term: '访存受限与计算受限',
+    aliases: ['memory bound', 'compute bound', '带宽受限'],
+    plainEnglish: '当运算单元是瓶颈时工作负载是计算受限，当等待数据搬运是瓶颈时则是访存受限。',
+    definition:
+      '当 kernel 的运行时间由算术吞吐能力决定时称为计算受限，由操作数在内存与运算单元之间的搬运速度决定时称为访存受限。',
+    explanation:
+      '每个 kernel 都有算术强度，即运算次数与访问字节数之比。如果这个比值低于硬件的平衡点，内存系统会先于运算单元饱和。LLM 的 prefill 执行算术强度很高的大矩阵乘法，通常是计算受限；而 decode 为每个请求读取全部权重和 KV cache 只产出一个 token，通常是访存受限。',
+    significance:
+      '瓶颈资源决定了哪个硬件规格才重要。decode 的访存受限特性解释了为什么每次加速器发布都把 HBM 容量和带宽放在标题上，为什么量化通过减少搬运字节数来加速 decode，以及为什么算力一般但内存很快的芯片能在交互式服务中获胜。',
+    benchmarkContext:
+      'InferenceX 对并发做扫描，让系统在两种状态之间移动：低并发 decode 受带宽限制，高并发批处理则推向计算极限。每条吞吐量对交互性曲线的形状，反映了该配置发生状态切换的位置。',
+  },
+  'gpu-utilization': {
+    term: 'GPU 利用率',
+    aliases: ['GPU utilization', '芯片利用率', '加速器利用率'],
+    plainEnglish:
+      'GPU 利用率衡量加速器有多忙，但监控工具里常见的百分比可能很高，而真实效率仍然很低。',
+    definition:
+      'GPU 利用率指加速器用于有效工作的时间或能力占比，口径从粗糙的繁忙百分比到严格的模型算力利用率不等。',
+    explanation:
+      '基础监控工具里的利用率只说明有 kernel 驻留，不代表芯片被用得好，一个只占用单个计算单元的 kernel 也会显示为繁忙。更严格的度量把实际交付的运算量或带宽与硬件峰值对比。在推理服务中利用率还受流量影响：请求之间的空闲、低并发以及 agent 会话中的工具调用停顿，都会让花钱买来的芯片闲着。',
+    significance:
+      '集群经济性取决于利用率。批处理和调度做得好的部署与朴素部署之间，同样硬件上每 token 成本常常相差数倍，这正是推理软件和请求路由与芯片本身同样受重视的原因。',
+    benchmarkContext:
+      'InferenceX 报告每个交互性水平下每颗芯片的实际交付吞吐量而非利用率百分比，因此引擎、精度和并行方案之间的利用率差异，会直接表现为同一硬件上曲线之间的间距。',
+  },
+  'continuous-batching': {
+    term: '连续批处理',
+    aliases: ['continuous batching', 'in-flight batching', '动态批处理'],
+    plainEnglish:
+      '连续批处理让新请求在旧请求完成的瞬间就加入正在运行的批次，而不是等整个批次全部结束。',
+    definition:
+      '连续批处理是一种在每个生成步都能接纳和退出请求的调度技术，当各序列在不同时刻完成时始终保持批次饱满。',
+    explanation:
+      '静态批处理先凑齐一组请求、一起运行、一起返回，因此一个批次要等最慢的成员结束。LLM 输出长度差异巨大，这种方式浪费大量算力。连续批处理在每次迭代重组批次：产出最后一个 token 的序列立即离开，排队中的请求在下一步顶上空位，加速器始终保持饱和。',
+    significance:
+      '这是现代 LLM 服务的奠基性优化之一，也是开源引擎取代朴素部署方式的重要原因。它在给定延迟下成倍提高吞吐量，并与分页式 KV cache 内存天然配合，使得槽位复用非常廉价。',
+    benchmarkContext:
+      'InferenceX 测试的每个引擎，包括 vLLM、SGLang 和 TensorRT-LLM，都使用连续批处理。并发扫描测量的正是各调度器在批次填满过程中维持交互性的能力，引擎实现之间的差异在这里显形。',
+  },
+  'paged-attention': {
+    term: 'PagedAttention',
+    aliases: ['paged attention', '分页 KV cache', 'KV cache 分页'],
+    plainEnglish:
+      'PagedAttention 像虚拟内存分页一样，把 KV cache 存进固定大小的小块，避免缓存内存浪费在用不到的空间上。',
+    definition:
+      'PagedAttention 是一种 KV cache 管理技术，通过映射表寻址的固定大小块来分配缓存，而不是为每个请求预留一整块连续区域。',
+    explanation:
+      '按请求连续分配必须为最长可能输出预留空间，而其中大部分永远用不上。分页借用了操作系统的做法：缓存块随序列增长按需分配，序列结束立即释放，共享相同前缀的序列还能通过写时复制共用缓存块。碎片率降到接近零，同样的 HBM 里能容纳多得多的序列。',
+    significance:
+      '这个由 vLLM 提出的思想解锁了让连续批处理真正获益的批次规模，并成为各推理引擎的标准做法。约束长上下文和智能体工作负载并发能力的，是有效 KV 容量而不是裸内存大小。',
+    benchmarkContext:
+      'InferenceX 配置中的所有引擎都以分页或分块形式管理 KV 内存。AgentX 这类长上下文场景的高并发数据点之所以可达，正是因为分页让数百个会话增长收缩时缓存浪费保持很小。',
+  },
+  'radix-attention': {
+    term: 'RadixAttention',
+    aliases: ['radix attention', 'radix tree 缓存', '基数树前缀缓存'],
+    plainEnglish:
+      'RadixAttention 把完成请求的 KV cache 按 token 内容存进基数树，任何新请求都能复用最长匹配前缀。',
+    definition:
+      'RadixAttention 是 SGLang 的前缀缓存设计，请求结束后把 KV cache 条目保留在基数树中，让共享 token 前缀的请求之间自动复用。',
+    explanation:
+      '基数树按 token 内容索引缓存片段，一次查找就能定位新请求已被计算过的最长前缀。复用是自动且跨请求的：多轮对话、许多用户共享的系统提示词、从共同历史分叉出的 agent 分支都会命中相同的缓存节点。近期最少使用等淘汰策略约束树占用的内存上限。',
+    significance:
+      '前缀复用把重复的 prefill 计算变成缓存命中。在每一轮都重发不断增长历史的 agent 流量里，这部分节省可以主导端到端成本。把复用做成结构性能力而不是可选项，是 SGLang 在此类工作负载上表现出色的重要原因。',
+    benchmarkContext:
+      'InferenceX 的 AgentX 轨迹保留了轮次之间和子智能体分支之间真实的共享前缀结构，因此具备强 radix 式复用能力的引擎在智能体场景中的首 token 时间和吞吐量，会明显好于固定序列场景的预测。',
+  },
+  'draft-model': {
+    term: '草稿模型',
+    aliases: ['draft model', 'speculator', '投机草稿模型'],
+    plainEnglish:
+      '草稿模型是投机解码中那个又小又快的模型，先猜出若干后续 token，交给大模型一次性验证。',
+    definition:
+      '草稿模型是投机解码中的轻量提议组件，生成候选 token 序列，由目标模型在单次批量前向中完成验证。',
+    explanation:
+      '草稿有多种形式：同家族的独立小模型、像 EAGLE 那样训练到目标模型上的额外预测头，或某些模型自带的多 token 预测头。草稿廉价地抢先生成几个 token，目标模型一次性检查全部候选，接受的 token 一起输出。被拒绝时回退到目标模型的输出，因此结果与目标分布一致。',
+    significance:
+      '草稿质量决定接受长度，接受长度决定加速比。匹配良好的草稿能在小批次下把 decode 速度提高数倍，而不匹配或过于激进的草稿会浪费验证算力，高负载下甚至拖慢服务。',
+    benchmarkContext:
+      'InferenceX 记录每个结果背后的投机方法和接受长度，并发布用于复现的黄金接受长度分布，因为不切实际的接受率是基准测试数字脱离生产行为的经典方式。',
+  },
+  'offline-inference': {
+    term: '离线推理',
+    aliases: ['offline inference', 'batch inference', '批量推理'],
+    plainEnglish: '离线推理处理一大堆没有用户在等的请求，唯一目标是每美元最多 token，而不是延迟。',
+    definition:
+      '离线推理是没有交互截止时间的模型服务，请求成批处理，目标是总吞吐量和成本而不是单用户延迟。',
+    explanation:
+      '合成数据生成、文档处理、embedding 回填和评估扫描都不在乎单个请求何时返回。调度器因此可以放开手脚：跑内存允许的最大批次、按前缀复用最大化的方式排序请求、把加速器压在吞吐量极限上。在线服务处于同一权衡的另一端，用吞吐量换取每个用户不低于某个交互性下限。',
+    significance:
+      '同一套硬件在离线模式与紧延迟模式之间，每美元 token 数可以相差数倍，所以不说明工作点就报每百万 token 价格几乎没有意义。集群常常因此划分成不同的延迟档位。',
+    benchmarkContext:
+      'InferenceX 吞吐量对交互性曲线的最右端，即批次最大、单用户速度最低处，近似离线运行。在一条曲线的两端各读一次，就能看到该配置从在线到离线的完整成本区间。',
+  },
+  'context-window': {
+    term: '上下文窗口',
+    aliases: ['context window', '上下文长度', '长上下文'],
+    plainEnglish: '上下文窗口是模型一次能处理的最大 token 数，涵盖输入和到目前为止生成的全部内容。',
+    definition:
+      '上下文窗口是模型支持的最大序列长度，约束提示词、对话历史、检索材料与生成输出的 token 总数。',
+    explanation:
+      '注意力让每个 token 都能引用之前的 token，KV cache 为它们全部保存状态，所以更长的窗口意味着随长度增长的内存和计算开销。位置编码方案和训练长度决定可用窗口，服务栈则必须为其预算 KV 容量。现代前沿模型宣称几十万 token 的窗口，但序列接近上限时吞吐量和交互性都会下降。',
+    significance:
+      '长上下文让代码 agent、重检索流水线和文档分析成为可能，也让它们的服务成本变高。滑动窗口层、潜在注意力和线性注意力等架构响应，主要就是为了压弯窗口的成本曲线。',
+    benchmarkContext:
+      'InferenceX 从两个方向覆盖窗口：固定序列场景钉住输入输出长度，例如 8K 输入 1K 输出；AgentX 则回放上下文逐轮增长、逼近真实 agent 工作集的会话。',
+  },
+  'grouped-query-attention': {
+    term: '分组查询注意力',
+    aliases: ['GQA', 'grouped-query attention', 'multi-query attention', 'MQA'],
+    plainEnglish:
+      '分组查询注意力让多个查询头共享一组键值头，在几乎不损失质量的情况下缩小 KV cache。',
+    definition:
+      '分组查询注意力（GQA）是一种注意力变体，把查询头划分成若干组，每组共享一个键值头，从而降低每 token 的 KV cache 大小和带宽。',
+    explanation:
+      '标准多头注意力为每个头都保存键和值，缓存大小随头数增长。多查询注意力（MQA）把所有头折叠到一对键值上，最省内存但可能伤害质量。GQA 处于两者之间：一个模型可以用 8 个 KV 头服务 64 个查询头，把缓存缩小八倍。由于 decode 的主要开销就是读取 KV cache，这一节省直接转化为更快的 token 生成。',
+    significance:
+      'GQA 成为稠密开源模型的默认注意力布局，因为它正面攻击了 decode 真正受限的内存侧。它也为多头潜在注意力等更激进的 KV 压缩方案铺平了道路。',
+    benchmarkContext:
+      '注意力布局由各模型架构固定，因此 GQA 在 InferenceX 中表现为模型层面每 token KV 字节数的差异，进而塑造相同硬件和引擎版本下可达的并发与交互性。',
+  },
+  'active-parameters': {
+    term: '激活参数',
+    aliases: ['active parameters', 'activated parameters', '激活参数量'],
+    plainEnglish:
+      '激活参数是混合专家模型处理每个 token 时实际用到的权重，只占其庞大总规模的一小部分。',
+    definition:
+      '激活参数是稀疏模型中参与计算单个 token 的那部分权重，包括共享层加上路由器选中的专家。',
+    explanation:
+      '一个混合专家模型可能拥有上万亿总参数，但每个 token 只经过几百亿。每 token 计算量随激活参数量增长，这是稀疏前沿模型运行成本可以接受的原因。内存则是另一回事：每个专家都必须驻留在 HBM 里随时待命，所以容量需求和并行方案跟随总参数量，尽管算术量跟随激活参数量。',
+    significance:
+      '总参数与激活参数之分解释了现代推理服务的大部分经济性：万亿参数模型因此可以部署，专家并行因此要跨多颗芯片，按总参数量比较模型也因此说明不了服务成本。',
+    benchmarkContext:
+      'InferenceX 测试的 MoE 模型，包括 DeepSeek、Kimi、Qwen 和 MiniMax 家族，激活比都很低，其配置把专家分布到多个节点，正是因为总参数量决定了内存账单。',
+  },
+  'dense-model': {
+    term: '稠密模型',
+    aliases: ['dense model', '稠密 Transformer'],
+    plainEnglish:
+      '稠密模型对处理的每个 token 都动用全部参数，不像稀疏模型那样把 token 路由给少数专家。',
+    definition: '稠密模型是所有权重都参与每次前向计算的神经网络，每 token 计算量与总参数量成正比。',
+    explanation:
+      '稠密 Transformer 是更简单的设计：每一层用全部权重处理每个 token。这让行为可预测、并行策略直接、单位参数的质量表现强，但服务成本随规模线性增长。混合专家模型打破了这一关联，每个 token 只激活一小部分权重，因此最大的前沿模型是稀疏的，而中小模型往往保持稠密。',
+    significance:
+      '稠密与稀疏之选驱动服务策略。稠密模型占用更少芯片、避开专家路由的复杂性；稀疏模型用大得多的内存占用和更重的跨芯片通信，换取单位算力更高的质量。',
+    benchmarkContext:
+      'InferenceX 的覆盖以运营方实际部署的大型稀疏前沿模型为中心，Llama 级别的稠密基线则提供对照，展示没有专家路由时张量并行和内存压力的表现。',
+  },
+  'reasoning-model': {
+    term: 'Reasoning 模型',
+    aliases: ['reasoning model', '深度思考模型', '思维链模型'],
+    plainEnglish:
+      'Reasoning 模型在回答前生成很长的隐藏思维链，用额外的输出 token 换取困难问题上更好的结果。',
+    definition:
+      'Reasoning 模型是被训练成额外花费生成 token 来推演问题的 LLM，在给出最终答案之前或同时产出大量中间推理。',
+    explanation:
+      '除了扩展训练算力，reasoning 模型还扩展测试时算力：它们用 token 来思考。一个数学或编程问题可能触发数千 token 的内部推演，输出长度相对聊天模型爆炸式增长。对服务而言，负载大幅偏向 decode，每请求的 KV cache 驻留量膨胀，每用户每秒 token 数成为决定难题几秒还是几分钟出答案的指标。',
+    significance:
+      'Reasoning 把推理变成了扩展前沿：能力现在通过在服务时花更多算力来提升，成倍放大对 decode 吞吐量的需求。它重塑了硬件优先级，转向内存带宽和互连，也是智能体工作负载主导当前基准设计的核心原因。',
+    benchmarkContext:
+      'InferenceX 测试的前沿模型都具备 reasoning 能力，其场景反映了这类流量：固定序列测试中的长生成，以及 AgentX 中推演与工具调用在多轮之间交织的完整 agent 会话。',
+  },
+  tokenization: {
+    term: 'Tokenization',
+    aliases: ['分词', 'tokenizer', 'token', 'BPE'],
+    plainEnglish: 'Tokenization 把文本切分成模型实际读写的子词单元，所有性能和价格数字都以它计价。',
+    definition:
+      'Tokenization 是文本与模型处理的离散 token ID 之间的转换，使用通过字节对编码（BPE）等方法学得的固定词表。',
+    explanation:
+      'tokenizer 把常见词映射为单个 token，把较罕见的字符串拆成多个，英文平均大约四个字符一个 token。不同模型家族词表不同，同一段文本在不同模型下的 token 数可能差异明显。下游一切都以 token 计价：上下文窗口、KV cache 大小、吞吐量、每 token 延迟和每百万 token 价格，数的都是这个单位而不是字符或单词。',
+    significance:
+      'token 效率是一个隐藏的价格杠杆：同样内容用更少 token 表达的模型，在相同单价下更便宜。比较供应商或基准测试时不对 tokenizer 差异做归一化，会悄悄扭曲成本和速度结论。',
+    benchmarkContext:
+      'InferenceX 的指标以 token 计价，其 AgentX 轨迹用确定性合成 token 替换原始文本，同时保留每轮 token 数，因此回放的会话以与源工作负载相同的 token 算术压测服务系统。',
+  },
+  'sequence-parallelism': {
+    term: '序列并行',
+    aliases: ['sequence parallelism', 'SP', 'sequence parallel'],
+    plainEnglish: '序列并行把一条长序列切分到多颗芯片上，让一个请求的 token 由多个加速器同时处理。',
+    definition:
+      '序列并行是沿序列的 token 维度在设备间做切分的策略，为超长输入分摊激活内存和注意力计算。',
+    explanation:
+      '张量并行切分的是权重，序列并行切分的是序列本身：每颗芯片持有一段 token 及对应的激活和 KV 状态。注意力因此需要通信，一颗芯片上的查询必须与其他芯片上的键值相遇，ring 式注意力算法把这些通信与计算重叠。在推理中，密切相关的上下文并行正是几十万 token 上下文的 prefill 变得可行的原因。',
+    significance:
+      '当单个请求而非批次过大时，序列式切分就是答案：一次百万 token 的 prefill 可能超出任何单芯片的内存和时间预算。它把上下文长度从一堵硬墙变成一个可扩展维度，代价是互连流量。',
+    benchmarkContext:
+      'InferenceX 记录每个配置的完整并行方案。长上下文智能体场景正是序列和上下文切分选择连同互连质量，把单芯片规格相近的系统明显区分开的地方。',
+  },
+  'all-gather': {
+    term: 'All-gather',
+    aliases: ['allgather', '全收集'],
+    plainEnglish:
+      'All-gather 是一种群体通信操作，结束时每颗芯片都拿到最初分散在所有芯片上的完整数据。',
+    definition:
+      'All-gather 是一种集合通信操作，每个参与设备贡献自己的分片，所有设备都收到全部分片拼接后的结果。',
+    explanation:
+      '分片执行经常需要重组完整张量：某些张量并行布局中矩阵乘法前的权重分片，或需要完整隐藏状态的操作之前各设备的激活。All-gather 把每个分片送到每个 rank，通常按 ring 或 tree 调度执行，开销随张量大小和参与者数量增长。它是 reduce-scatter 的对偶操作，两者组合起来就是一次 all-reduce。',
+    significance:
+      '与 all-reduce、all-to-all 一起，all-gather 是决定并行方案能否扩展的少数几个集合操作之一。它的延迟处在每一使用它的层的关键路径上，这也是芯片间 scale-up 带宽被大力宣传的原因。',
+    benchmarkContext:
+      'InferenceX 测试的每个多芯片配置都通过其并行方案频繁调用集合操作，而 CollectiveX 工作流直接在多家厂商硬件上测量这类操作，让通信行为可以在完整模型运行之外被比较。',
+  },
+  'reduce-scatter': {
+    term: 'Reduce-scatter',
+    aliases: ['reduce scatter', '归约散播'],
+    plainEnglish:
+      'Reduce-scatter 把每颗芯片上对应位置的数据求和，然后让每颗芯片只保留合并结果中属于自己的那一片。',
+    definition:
+      'Reduce-scatter 是一种集合操作，对所有设备贡献的张量做逐元素归约，并把归约结果按分片分发，每个设备一片。',
+    explanation:
+      '当每个 rank 都为同一张量计算出部分结果时，这些部分结果必须相加。Reduce-scatter 完成求和后只把下一步需要的那一片交给每个 rank，避免把完整归约张量发给所有人的浪费。一次 all-reduce 恰好等于 reduce-scatter 加 all-gather，调度器会根据下一个操作实际消费什么来选择融合形式还是拆分形式。',
+    significance:
+      '当下游只需要分片时，用 reduce-scatter 代替完整 all-reduce 能把每个 rank 的接收数据量减半。在 NVL72 这种规模下，集合通信流量与模型本身争夺互连带宽，这一点尤为重要。把这些集合操作与计算重叠，是成熟推理栈的标志性能力。',
+    benchmarkContext:
+      'InferenceX 中带张量并行分片的配置在每一层 Transformer 都会触发归约类集合操作，而 CollectiveX 存在的意义正是在推理实际使用的消息尺寸上，跨厂商发布这些原语的测量结果。',
+  },
+  infiniband: {
+    term: 'InfiniBand',
+    aliases: ['IB', 'InfiniBand 网络', 'NDR InfiniBand'],
+    plainEnglish:
+      'InfiniBand 是连接 AI 集群中各台服务器的高速低延迟网络，承担 NVLink 够不到的跨节点流量。',
+    definition:
+      'InfiniBand 是原生支持 RDMA 的交换式网络，是大多数大型 NVIDIA 集群中节点之间的 scale-out 互连。',
+    explanation:
+      '节点或机柜内部，芯片通过 NVLink 这类 scale-up 链路通信；越过这个边界，流量就进入 scale-out 网络，InfiniBand 在这里与支持 RDMA 的 Ethernet 竞争。InfiniBand 提供微秒级延迟、当前世代每链路数百 Gb/s 的带宽，以及 SHARP 这类在网归约能力。多节点推理、prefill 与 decode 分离、宽专家并行都把跨节点集合通信和 KV 传输放在这张网络上。',
+    significance:
+      '一旦模型跨节点，网络就与算力一样成为一级性能要素。网络选型塑造集群成本和厂商绑定，InfiniBand 与 Ethernet 之争是 AI 基础设施领域最核心的竞争战场之一。',
+    benchmarkContext:
+      'InferenceX 的多节点配置，包括分离式和机柜级结果，都跑在所在集群的 scale-out 网络上，配置元数据会记录互连类型，因为它对跨节点数字的可复现性影响重大。',
+  },
+  rdma: {
+    term: '远程直接内存访问',
+    aliases: ['RDMA', 'RoCE', 'GPUDirect RDMA'],
+    plainEnglish:
+      'RDMA 让一台机器直接通过网络读写另一台机器的内存，两边的 CPU 都不用参与数据拷贝。',
+    definition:
+      '远程直接内存访问（RDMA）是一种网络能力，由网卡在两台机器的内存之间直接搬运数据，绕过操作系统和 CPU 的拷贝开销。',
+    explanation:
+      '常规网络栈在两端都要经过内核缓冲区拷贝数据，消耗 CPU 周期和延迟。RDMA 网卡在已注册的内存区域之间直接传输，GPUDirect 进一步让网卡直接写入加速器的 HBM。InfiniBand 原生内置 RDMA，RoCE 则把同一套语义跑在 Ethernet 上。NCCL、RCCL 等集合通信库，以及分离式服务中的 KV cache 传输路径，都建在这些原语之上。',
+    significance:
+      'RDMA 是整个分布式 AI 栈的地板：没有它，跨节点集合通信和缓存传输早在链路跑满之前就会卡在 CPU 上。RoCE 变体让基于 Ethernet 的集群能以更低成本与 InfiniBand 竞争。',
+    benchmarkContext:
+      'InferenceX 的每个多节点结果在集合通信之下都依赖 RDMA 传输，分离式配置还依赖它完成 prefill 到 decode 的 KV 搬运，因此传输层成熟度是区分相似集群结果的因素之一。',
+  },
+  ualink: {
+    term: 'UALink',
+    aliases: ['Ultra Accelerator Link', 'UALoE'],
+    plainEnglish:
+      'UALink 是机柜内加速器之间高速 scale-up 链路的开放行业标准，是生态对 NVLink 的回应。',
+    definition:
+      'UALink 是面向加速器之间 scale-up 通信的开放互连规范，让机柜内的大量芯片以交换级速度共享内存流量。',
+    explanation:
+      'NVLink 证明了用低延迟内存语义网络连接的一机柜芯片可以表现得像一颗巨型加速器，但它是私有的。UALink 联盟定义了开放的等价方案，让 NVIDIA 之外的厂商也能构建机柜级域。UALink over Ethernet（简称 UALoE）把协议跑在 Ethernet 交换上，AMD Helios 世代的机柜采用这条路线，组成结构上对标 NVL72 的 72 芯片 scale-up 域。',
+    significance:
+      'scale-up 域的大小日益决定服务架构，宽专家并行和分离式部署都希望几十颗芯片处在同一张快速网络内。开放标准决定机柜级推理是继续作为单厂商优势，还是成为整个生态的能力。',
+    benchmarkContext:
+      'InferenceX 在硬件覆盖中把 AMD MI455X 机柜这类 UALoE72 级系统与 NVL72 系统并列，机柜级网络上市后即可在相同模型工作负载上正面对比。',
+  },
+  tdp: {
+    term: '热设计功耗',
+    aliases: ['TDP', '整卡功耗', '全部包含功耗'],
+    plainEnglish:
+      'TDP 是芯片设计上可持续消耗并以热量形式散发的功率，是每份加速器规格表上的标题瓦数。',
+    definition:
+      '热设计功耗（TDP）是芯片被设计为可持续运行的最大功率包络，其散热系统必须能够持续把这些热量排出。',
+    explanation:
+      '现代加速器每颗芯片的功耗在千瓦上下甚至更高，整机柜系统乘上去就是十万瓦量级。只看 TDP 还会低估真实账单：内存、网络、CPU、电源转换损耗和散热开销都叠加在上面，因此每芯片的全部包含功耗明显高于芯片 TDP。数据中心容量按兆瓦出售，这些功率包络直接决定一个场地能容纳多少加速器。',
+    significance:
+      '电力已成为 AI 扩建的硬约束，在许多市场甚至排在资本之前。每芯片 TDP 的持续上升迫使行业转向液冷，也让每瓦性能与每美元性能一样，成为比较芯片世代的主要维度。',
+    benchmarkContext:
+      'InferenceX 用包含散热和基础设施开销的每芯片全部包含功耗来计算每 token 能耗和每兆瓦 token 数，PowerX 工作流正把这些指标从铭牌数值推向运行中的实测功耗。',
+  },
+  pue: {
+    term: '电源使用效率',
+    aliases: ['PUE', 'power usage effectiveness', '能源使用效率'],
+    plainEnglish: 'PUE 衡量数据中心每向内部计算设备输送一瓦电，总共要消耗多少瓦。',
+    definition:
+      '电源使用效率（PUE）是数据中心总功耗与 IT 设备功耗之比，1.0 意味着每一瓦都用于计算。',
+    explanation:
+      '制冷、电源转换和场地系统在服务器之外额外耗能。PUE 为 1.5 意味着每一瓦 IT 负载还要搭上半瓦开销，而现代超大规模 AI 设施借助液冷和高效配电把 PUE 压向 1.1 以下。AI 园区以数百兆瓦计，PUE 的微小差异就会摊出巨大的绝对能耗和成本。',
+    significance:
+      'PUE 把芯片级效率与场地经济性连接起来：芯片消耗的每一瓦在电表上都要乘以 PUE。当电力供给卡住 AI 扩建时，场地效率与芯片和软件一样进入竞争计算。',
+    benchmarkContext:
+      'InferenceX 每 token 能耗和每兆瓦 token 数背后的全部包含功耗，已按现代 AI 数据中心的 PUE 水平计入场地开销，因此其成本与能耗比较反映的是落地的场地经济性而不是裸芯片瓦数。',
+  },
+  int8: {
+    term: 'INT8',
+    aliases: ['8 位整数量化', 'W8A8'],
+    plainEnglish:
+      'INT8 用 8 位整数加缩放因子存储数值，内存相比 16 位格式减半，支持的硬件上算力翻倍。',
+    definition:
+      'INT8 是一种 8 位整数数值格式，搭配每张量或每通道的缩放因子，用于量化推理中的模型权重和激活。',
+    explanation:
+      '整数量化通过缩放因子（有时还有零点）把浮点值映射到 256 个均匀间隔的等级上。均匀间隔对离群值处理得很差，因此 SmoothQuant 等技术先把激活中的离群值迁移到权重里，再按 W8A8 模式对两侧一起量化。在较早的加速器世代，INT8 是 16 位以下的主要快速路径，而新芯片增加了 FP8，同样位宽下指数位带来更宽的动态范围。',
+    significance:
+      'INT8 定义了 LLM 量化的第一波主流实践，在没有 8 位浮点支持的硬件上仍然重要。INT8 与 FP8 的对比也恰好展示了量化的核心权衡：均匀精度对动态范围。',
+    benchmarkContext:
+      'InferenceX 为每个结果标注精度，其精度对比页面存在的原因就是格式切换曾大幅移动曲线。Blackwell 和 MI350 级硬件上的现代配置偏好 FP8 和 FP4 路径，整数格式出现在特定的权重量化配置中。',
+  },
+  'weight-only-quantization': {
+    term: '仅权重量化',
+    aliases: ['weight-only quantization', 'W4A16', 'AWQ', 'GPTQ'],
+    plainEnglish: '仅权重量化只把存储的模型权重压到低精度，运算本身仍在较高精度格式下进行。',
+    definition:
+      '仅权重量化把模型权重存为 4 位整数等低位格式，而激活和算术保持高精度，典型模式是 W4A16。',
+    explanation:
+      '权重是静态的，可以用 GPTQ、AWQ 等方法在离线阶段精心量化，选择能最小化质量损失的缩放和顺序。激活是动态的、更难压缩，保持 16 位可以绕开它们的离群值问题。服务时 kernel 在计算中即时反量化权重，因此访存流量减小，而乘加运算本身并没有变快。',
+    significance:
+      '由于 decode 是访存受限的，削减权重字节数直接加快 token 生成，也让更大的模型装进更少的芯片。仅权重方法让大型开源模型能在普通硬件上运行，在激活量化代价过高时仍是标准方案。',
+    benchmarkContext:
+      'InferenceX 在精度标签中区分仅权重配置与全低精度路径，因为 W4A16 配置和 NVFP4 配置对于是哪些硬件单元和带宽预算产生了曲线，做出的是很不一样的声明。',
+  },
+  'block-scaling': {
+    term: '块缩放',
+    aliases: ['block scaling', 'microscaling', 'MX 格式', '块浮点'],
+    plainEnglish: '块缩放给每一小组低精度数值配一个共享缩放因子，找回超小格式自身缺失的动态范围。',
+    definition:
+      '块缩放是一种量化结构，数值以极低位宽格式存储，每个固定大小的块共享一个较高精度的缩放因子。',
+    explanation:
+      '一个 4 位数只能表示少数几个量级，远不足以覆盖模型张量的动态范围。把数值分成 16 或 32 个元素左右的块并附上共享缩放，每个块就能把格式对准自己的量级。MX 标准格式如 MXFP4、MXFP8 使用 2 的幂缩放，NVFP4 则在 16 元素块上使用 FP8 缩放，粒度更细。',
+    significance:
+      '块缩放是 FP4 推理世代背后的关键思想：没有每块缩放，4 位浮点对前沿模型根本不可用。缩放格式和块大小的选择，如今是硬件厂商与量化方案之间真正的差异化点。',
+    benchmarkContext:
+      'InferenceX 在 Blackwell 和 MI355X 覆盖中的 NVFP4 与 MXFP4 结果都是块缩放格式，精度对比页面很大程度上就是为了展示这些配置相对同硬件 FP8 基线的得失。',
+  },
+  'flash-attention': {
+    term: 'FlashAttention',
+    aliases: ['flash attention', '融合注意力 kernel'],
+    plainEnglish:
+      'FlashAttention 在高速片上内存中分块计算精确注意力，避免生成那个让注意力又慢又耗内存的巨大中间矩阵。',
+    definition:
+      'FlashAttention 是一种注意力算法，把计算分块安排在片上 SRAM 中并增量式重缩放结果，不在 HBM 中实体化完整分数矩阵就得到精确注意力。',
+    explanation:
+      '朴素注意力要把随序列长度平方增长的分数矩阵写入主存再读回，速度由带宽而非算术决定。FlashAttention 把整个计算融合进一个 kernel，让键值块流过片上内存，用在线 softmax 保持结果精确。后续版本和各厂商实现把这一思想扩展到新硬件世代、新头布局和面向推理的 decode 路径。',
+    significance:
+      '这个 kernel 家族让长上下文真正可行，把注意力从长序列的主导成本变成可控的一环。它也是单个精心设计的 kernel 能撕动整个行业性能的最经典例子。',
+    benchmarkContext:
+      'InferenceX 测试的所有引擎都依赖这一脉络的融合注意力 kernel，NVIDIA 硬件上经由 FlashInfer，AMD 上经由 AITER，那里的 kernel 改进经常在硬件不变的情况下移动已发布的曲线。',
+  },
+  triton: {
+    term: 'Triton',
+    aliases: ['OpenAI Triton', 'Triton kernel 语言'],
+    plainEnglish:
+      'Triton 是基于 Python 的加速器 kernel 编写语言，让 ML 工程师不写底层代码也能接近手工调优的速度。',
+    definition:
+      'Triton 是开源的 kernel 编程语言和编译器，开发者用类 Python 代码编写高性能加速器 kernel，在维护后端的厂商之间可移植。',
+    explanation:
+      '传统上写出峰值性能的 kernel 需要 CUDA 或汇编级调优的厂商专家知识。Triton 抬高了抽象层级：开发者编写块级程序，编译器处理访存合并、分块和调度。NVIDIA、AMD 等厂商维护后端，同一份 kernel 源码可以面向多种架构。推理引擎用它实现融合算子、量化路径和没有厂商库可用的 MoE kernel。这个名字与 NVIDIA Triton Inference Server 重名，后者是另一个独立的模型服务产品。',
+    significance:
+      'Triton 降低了模型研究者与硬件性能之间的门槛，其跨厂商后端具有战略意义，因为用它写的 kernel 不会锁定在单一芯片家族上。厂商把 Triton 生态运营得好不好，已经成为其软件叙事的一部分。',
+    benchmarkContext:
+      'InferenceX 配置中的引擎在 CUDA、CUTLASS 和 AITER 代码之外还携带大量 Triton kernel，因此编译器与后端成熟度是 NVIDIA 和 AMD 系统上引擎版本之间曲线移动的隐形推力之一。',
+  },
+  cutlass: {
+    term: 'CUTLASS',
+    aliases: ['CuTe', 'CUDA 线性代数模板库'],
+    plainEnglish:
+      'CUTLASS 是 NVIDIA 的模板库，提供搭建接近硬件峰值速度的矩阵乘法 kernel 所需的积木。',
+    definition:
+      'CUTLASS 是 NVIDIA 开源的可组合 C++ 模板库，用于构建面向各代 GPU tensor core 的 GEMM 及相关 kernel。',
+    explanation:
+      '峰值矩阵乘法性能要求对 tensor core 指令、共享内存搬运和异步流水线的精确编排，而且每代架构都不同。CUTLASS 把这套编排封装成可组合的部件，其 CuTe 层描述数据布局，kernel 作者可以拼装出接近峰值的 GEMM，并在尾部融合偏置、激活或量化缩放，而不必从零开始。推理引擎中大量高性能 kernel 直接建在它之上。',
+    significance:
+      'CUTLASS 是 NVIDIA 教生态使用每一代新 tensor core 的地方，包括 Blackwell 上的 FP4 和 FP8 路径。其模式向引擎扩散的速度，是新芯片多快达到宣称性能的真实变量。',
+    benchmarkContext:
+      'InferenceX 在 NVIDIA 硬件上的结果背后，GEMM 和注意力 kernel 大量源于 CUTLASS 衍生代码，引擎镜像里这些库的版本升级是平台追踪到的曲线逐日变化的常见来源。',
+  },
+  nccl: {
+    term: 'NCCL',
+    aliases: ['RCCL', 'NVIDIA 集合通信库'],
+    plainEnglish:
+      'NCCL 是 NVIDIA 的集合通信库，在集合操作中负责芯片之间的数据搬运，RCCL 是 AMD 的对应实现。',
+    definition:
+      'NCCL 是 NVIDIA 的集合通信库，在节点内经 NVLink、节点间经 RDMA 网络实现 all-reduce、all-gather、all-to-all 等跨 GPU 操作。',
+    explanation:
+      '框架不直接操作互连，而是调用一个集合通信库，由它发现拓扑并为每种消息尺寸选择算法和信道调度。NVIDIA 系统由 NCCL 负责，AMD 维护接口对齐的 RCCL 服务 ROCm 平台。调优是针对具体网络的：ring 与 tree 算法、协议阈值、信道数都随拓扑变化，同一模型在两个集群上的通信行为可能大不相同。',
+    significance:
+      '每个多芯片推理和训练任务都站在这一层之上，集合通信的回退会悄无声息地向整个集群征税。NCCL 与 RCCL 的接口兼容性同样至关重要，引擎可以面向同一套集合通信 API 覆盖多家厂商。',
+    benchmarkContext:
+      'InferenceX 的多芯片配置在每个张量并行和专家并行层都在调用这些库，而 CollectiveX 在推理相关的消息尺寸上跨厂商直接测量底层集合通信性能，把网络行为与模型行为分离开。',
+  },
+  gemm: {
+    term: 'GEMM',
+    aliases: ['通用矩阵乘法', 'matrix multiply'],
+    plainEnglish: 'GEMM 即通用矩阵乘法，是训练和推理神经网络时占据绝大部分算术量的那一种计算。',
+    definition:
+      'GEMM 是通用的矩阵乘矩阵例程，神经网络中的线性层、注意力投影和专家计算最终都归结为它。',
+    explanation:
+      'Transformer 主要由线性变换堆叠而成，服务一个模型就是执行海量矩阵乘法。tensor core 专为加速它们而存在，峰值 TFLOP/s 规格也是针对这类稠密运算给出的。形状决定效率：prefill 产生接近方形的大矩阵乘法能喂饱算力，decode 产生的细长矩阵则受带宽限制。MoE 还带来分组 GEMM，把许多小型专家乘法合并成一次高效启动。',
+    significance:
+      'GEMM 效率是行业里一切性能主张的地基。在真实服务形状下，尤其是 decode 的细长矩阵上，实际交付与峰值 GEMM 吞吐量的差距，解释了规格表比例为何预测不了基准测试排名。',
+    benchmarkContext:
+      'InferenceX 每条曲线背后都是来自 CUTLASS、hipBLASLt 和 Triton 生成代码的 GEMM kernel 堆栈，量化配置最终也站在各厂商在其调度器产生的形状上执行低精度 GEMM 的能力之上。',
+  },
+  'kernel-fusion': {
+    term: 'Kernel 融合',
+    aliases: ['kernel fusion', '算子融合'],
+    plainEnglish:
+      'Kernel 融合把多个小的芯片操作合并成一个，中间数据留在快速内存里而不是反复经过 HBM。',
+    definition:
+      'Kernel 融合把多个连续操作合并为单次 kernel 启动，中间结果保留在寄存器或片上内存中，而不是在步骤之间写回主存。',
+    explanation:
+      '未融合的矩阵乘法、偏置加法、激活函数序列，每一步都把中间张量写入 HBM 再读回。融合成一个 kernel 就消除了这些往返和中间的启动开销。融合可以在库里手工实现，可以通过 CUTLASS 式的模板尾部实现，也可以由 torch.compile 和基于 Triton 的栈自动完成。FlashAttention 是这一思想最著名的单个例子。',
+    significance:
+      '在访存受限的推理中，消除中间流量比提升原始算术能力更值钱，因此融合是引擎手中最可靠的杠杆之一。引擎版本间的提速，很大一部分可以归结为更激进或更精准的融合。',
+    benchmarkContext:
+      'InferenceX 配置钉住的引擎镜像在不同版本和硬件后端上的融合库存不同，这是逐日追踪中硬件不变而曲线移动的常见原因：面向新模型和新精度的融合 kernel 落地了。',
+  },
 };
 
 const entries = getAllGlossaryEntries().map((entry) => {
diff --git a/packages/app/src/lib/glossary.test.ts b/packages/app/src/lib/glossary.test.ts
index 18eb6cc9c..e524c5805 100644
--- a/packages/app/src/lib/glossary.test.ts
+++ b/packages/app/src/lib/glossary.test.ts
@@ -55,6 +55,16 @@ describe('glossary content', () => {
     }
   });
 
+  it('defines the AgentX workload and its agentic inference concepts', () => {
+    expect(GLOSSARY_CATEGORIES).toContain('Agentic inference');
+    expect(getGlossaryEntry('agentic-inference')?.category).toBe('Agentic inference');
+    expect(getGlossaryEntry('agentx')?.definition).toContain('agentic inference benchmark');
+    expect(getGlossaryEntry('agentic-coding-workload')?.relatedTerms).toContain('agentx');
+    expect(getGlossaryEntry('trace-replay')?.benchmarkContext).toContain('AIPerf');
+    expect(getGlossaryEntry('closed-loop-benchmark')?.benchmarkContext).toContain('agent clients');
+    expect(getGlossaryEntry('subagent')?.relatedTerms).toContain('trace-replay');
+  });
+
   it('links every glossary source to a real article and covers the complete article library', () => {
     const entries = getAllGlossaryEntries();
     const referencedArticles = new Set(entries.flatMap((entry) => entry.articleSlugs));
@@ -107,6 +117,12 @@ describe('Chinese glossary content', () => {
   });
 
   it('resolves canonical slugs and walks the Chinese term order without gaps', () => {
+    expect(GLOSSARY_CATEGORY_LABELS_ZH['Agentic inference']).toBe('智能体推理');
+    expect(getZhGlossaryEntry('agentic-inference')?.term).toBe('智能体推理');
+    expect(getZhGlossaryEntry('agentx')?.term).toBe('AgentX');
+    expect(getZhGlossaryEntry('trace-replay')?.term).toBe('轨迹回放');
+    expect(getZhGlossaryEntry('closed-loop-benchmark')?.term).toBe('闭环基准测试');
+    expect(getZhGlossaryEntry('subagent')?.term).toBe('子智能体');
     expect(getZhGlossaryEntry('multi-token-prediction')?.term).toBe('多 token 预测');
     expect(getZhGlossaryEntry('not-a-real-term')).toBeUndefined();
 
diff --git a/packages/app/src/lib/glossary.ts b/packages/app/src/lib/glossary.ts
index cdacf6f0c..60e1a196f 100644
--- a/packages/app/src/lib/glossary.ts
+++ b/packages/app/src/lib/glossary.ts
@@ -1,6 +1,7 @@
 export const GLOSSARY_CATEGORIES = [
   'Benchmark metrics',
   'Serving',
+  'Agentic inference',
   'Parallelism',
   'Hardware',
   'Numerical precision',
@@ -46,6 +47,20 @@ const SGLANG_056 = 'sglang-0-5-6-b200-deepseek-r1-fp4-up-to-1-8x';
 const VR_RUBIN = 'vera-rubin-nvl72-vs-gb200-nvl72-inference';
 const KIMI_K3 = 'kimi-k3-the-manos-the-mythos-the';
 const TILERT = 'ultra-high-interactivity-on-nvidia';
+const AGENT_BENCHMARK = 'agentic-benchmark-agent-benchmark-guide';
+const AGENTIC_WORKLOADS = 'brief-overview-of-agentic-workloads';
+const AGENTX_V3 = 'agentx-inferencexv3-does-cuda-moat';
+const AGENTX_DSV4_MI355X_B200 = 'deepseek-v4-pro-agentx-mi355x-vs-b200-august';
+const AGENTX_DSV4_B200_B300 = 'deepseek-v4-pro-agentx-b200-vs-b300-kv-working-set';
+const AGENTX_DSV4_GB200_GB300 = 'deepseek-v4-pro-agentx-gb200-vs-gb300-disagg';
+const AGENTX_K3_ATOM = 'kimi-k3-agentx-mi355x-atom-vs-gb300-nvl72';
+const AGENTX_M3_TRT = 'minimax-m3-agentx-b300-trtllm-tp2-vs-the-field';
+const AGENTX_M3_RACK = 'minimax-m3-agentx-b200-b300-vs-rack-scale';
+const AGENTX_QWEN_SGLANG = 'qwen3-5-397b-agentx-nvidia-vs-amd-sglang';
+const AGENTX_QWEN_B300 = 'qwen3-5-397b-agentx-b300-fp4-vs-h100';
+const AGENTX_GLM_SGLANG = 'glm-5-3-agentx-nvidia-vs-amd-sglang-150-toks';
+const AGENTX_GLM_ATOM = 'glm-5-3-agentx-mi355x-atom-vs-gb300-nvl72';
+const JALAPENO = 'openai-jalapeno-better-than-nvidia';
 
 const entries = [
   {
@@ -66,6 +81,126 @@ const entries = [
     relatedTerms: ['inference-engine', 'prefill', 'decode', 'throughput', 'interactivity'],
     articleSlugs: [INFERENCEMAX, INFERENCEX_V2],
   },
+  {
+    slug: 'agentic-inference',
+    term: 'Agentic inference',
+    aliases: ['AI agent inference', 'agent inference'],
+    category: 'Agentic inference',
+    plainEnglish:
+      'Agentic inference serves an AI system that works through a task over many model requests, often using tools and delegating work along the way.',
+    definition:
+      'Agentic inference is model serving for agents that maintain state across multiple turns, call tools, reuse growing context, and may run subagents in parallel.',
+    explanation:
+      'A single agent session can alternate between model requests, tool execution, and waiting periods. Later requests often include much of the earlier conversation, so prefix caching and KV-cache capacity affect both speed and cost. Parallel subagents add branches with their own request timing and context growth.',
+    significance:
+      'Fixed input and output lengths miss several pressures created by agents. Long shared prefixes change cache behavior, tool delays make traffic bursty, and concurrent branches compete for serving capacity. Hardware and software can rank differently under this request pattern.',
+    benchmarkContext:
+      'InferenceX uses AgentX to measure agentic inference. Read AgentX results alongside fixed-sequence scenarios because they answer different capacity questions. AgentX reports the behavior of a closed-loop session replay instead of treating every request as an independent batch item.',
+    relatedTerms: ['agentx', 'agentic-coding-workload', 'subagent', 'prefix-caching', 'kv-cache'],
+    articleSlugs: [AGENTIC_WORKLOADS, AGENT_BENCHMARK, TILERT, VR_RUBIN, INFERENCEX_V2, AGENTX_V3],
+  },
+  {
+    slug: 'agentx',
+    term: 'AgentX',
+    aliases: ['AgentX benchmark', 'AgentX scenario'],
+    category: 'Agentic inference',
+    plainEnglish:
+      'AgentX is the InferenceX workload for testing how inference systems serve complete long-context, multi-turn coding-agent sessions.',
+    definition:
+      'AgentX is InferenceX’s agentic inference benchmark scenario, built from workload shapes derived from opt-in coding-agent traces after original content is removed.',
+    explanation:
+      'AgentX reconstructs session structure with deterministic synthetic tokens. Its replay keeps request lengths, turn timing, shared-prefix growth, tool pauses, and main-agent or subagent dependencies while excluding original prompts, generated code, and tool payloads. The serving stack receives the traffic pattern without receiving the source conversation.',
+    significance:
+      'Long contexts pressure KV-cache capacity, repeated prefixes reward effective cache reuse, and branch timing tests request scheduling. These effects are small or absent in short, independent requests. The resulting curve describes the complete serving system under agent traffic.',
+    benchmarkContext:
+      'The Agentic scenario appears by default for models with matching AgentX data. Compare its throughput, latency, and interactivity only with other AgentX runs at compatible settings. Use fixed-sequence scenarios when the target workload is a conventional request stream.',
+    relatedTerms: [
+      'agentic-inference',
+      'agentic-coding-workload',
+      'trace-replay',
+      'closed-loop-benchmark',
+      'subagent',
+    ],
+    articleSlugs: [AGENTIC_WORKLOADS, AGENT_BENCHMARK, TILERT, VR_RUBIN, AGENTX_V3],
+  },
+  {
+    slug: 'agentic-coding-workload',
+    term: 'Agentic coding workload',
+    aliases: ['coding-agent workload', 'software-engineering agent workload'],
+    category: 'Agentic inference',
+    plainEnglish:
+      'This is the request pattern created when a coding agent reads a repository, edits code, runs tools, and revisits the model until the task is done.',
+    definition:
+      'An agentic coding workload is a multi-turn inference workload produced by a software agent that combines model generation with repository inspection, tool calls, code changes, and delegated subtasks.',
+    explanation:
+      'Request sizes grow as the agent accumulates instructions, files, tool results, and earlier responses. Many turns reuse a large common prefix. Tool execution inserts uneven delays, while subagents can create overlapping request branches. These properties produce a different traffic shape from fixed-length prompt benchmarks.',
+    significance:
+      'Coding agents can keep a serving system busy for minutes or hours through a chain of dependent calls. Cache policy, scheduler fairness, memory capacity, and tail latency all affect task progress. Peak decode throughput alone cannot describe that behavior.',
+    benchmarkContext:
+      'AgentX represents this workload with trace-derived request shapes and deterministic synthetic content. It measures inference-system performance. Model coding quality requires a separate evaluation, so quality scores and AgentX serving results answer separate questions.',
+    relatedTerms: ['agentic-inference', 'agentx', 'subagent', 'prefix-caching', 'trace-replay'],
+    articleSlugs: [AGENTIC_WORKLOADS, AGENT_BENCHMARK, TILERT, VR_RUBIN, INFERENCEX_V2, AGENTX_V3],
+  },
+  {
+    slug: 'trace-replay',
+    term: 'Trace replay',
+    aliases: ['workload replay', 'session replay'],
+    category: 'Agentic inference',
+    plainEnglish:
+      'Trace replay recreates the timing and shape of recorded sessions so a benchmark sends requests like the original workload.',
+    definition:
+      'Trace replay is a benchmarking method that converts recorded request relationships, lengths, and timing into a repeatable workload for a system under test.',
+    explanation:
+      'A replay can preserve a directed graph of main-agent turns, parallel subagent branches, and auxiliary requests. Deterministic synthetic tokens replace private content while retaining token counts and prefix relationships. Recorded gaps between turns reproduce the periods when an agent was using tools or waiting on dependencies.',
+    significance:
+      'The method captures traffic features that a list of independent prompts cannot express. It also makes repeated hardware and software comparisons possible from the same session shapes. AgentX removes source-conversation content before publishing replay data.',
+    benchmarkContext:
+      'AgentX replays trace-derived sessions through AIPerf. A fixed seed selects sessions, starting points, and synthetic content. Reported results cover the profiling window after cache warmup, which keeps run-to-run comparisons focused on steady-state serving behavior.',
+    relatedTerms: ['agentx', 'closed-loop-benchmark', 'subagent', 'concurrency', 'kv-cache'],
+    articleSlugs: [AGENTIC_WORKLOADS, AGENT_BENCHMARK, TILERT, VR_RUBIN, AGENTX_V3],
+  },
+  {
+    slug: 'closed-loop-benchmark',
+    term: 'Closed-loop benchmark',
+    aliases: ['closed-loop load test', 'closed-loop workload'],
+    category: 'Agentic inference',
+    plainEnglish:
+      'In a closed-loop benchmark, each simulated user waits for one step to finish before sending the next step in that session.',
+    definition:
+      'A closed-loop benchmark generates new work from each client in response to completion of its previous dependent request, subject to the workload’s recorded delays and branch structure.',
+    explanation:
+      'Concurrency is the number of active clients or sessions; the simultaneous request count changes over time. Faster systems complete turns sooner and therefore issue more requests during the same profiling period. The exact request mix can vary slightly because progress through each sampled session depends on completion time.',
+    significance:
+      'This load model resembles interactive agents, where the next action depends on the previous result. Throughput and latency remain coupled: a faster response advances the session and creates later work sooner. Low-concurrency runs can show more sampling variation than large pooled runs.',
+    benchmarkContext:
+      'AgentX uses closed-loop concurrency. Its concurrency value is the number of agent clients; request batch size changes as the sessions advance. Read throughput, time to first token, and interactivity together.',
+    relatedTerms: ['agentx', 'trace-replay', 'concurrency', 'throughput', 'latency'],
+    articleSlugs: [AGENT_BENCHMARK, TILERT, INFERENCEX_V2, AGENTX_V3],
+  },
+  {
+    slug: 'subagent',
+    term: 'Subagent',
+    aliases: ['child agent', 'delegated agent'],
+    category: 'Agentic inference',
+    plainEnglish:
+      'A subagent is an additional agent started by a main agent to handle a smaller piece of the same task, sometimes at the same time as other work.',
+    definition:
+      'A subagent is a delegated agent execution with its own conversation state and model requests, connected to a parent session through task and dependency relationships.',
+    explanation:
+      'The main agent can launch one or more subagents and later consume their results. Their requests may overlap with the parent or with each other, creating branches in the session graph. Each branch can grow a separate context while sharing some initial instructions or repository state.',
+    significance:
+      'Subagents make agent traffic less sequential. A serving stack may receive bursts of long-context requests from one user task, and scheduler decisions affect how quickly branches finish. Aggregate throughput can rise while an individual branch waits longer for service.',
+    benchmarkContext:
+      'AgentX preserves subagent branches from the trace-derived workload and replays their dependencies. Delegation quality is outside its scope. The benchmark measures how the inference system serves the resulting parallel requests, shared prefixes, and completion timing.',
+    relatedTerms: [
+      'agentic-inference',
+      'agentic-coding-workload',
+      'agentx',
+      'trace-replay',
+      'concurrency',
+    ],
+    articleSlugs: [AGENTIC_WORKLOADS, TILERT, VR_RUBIN, AGENTX_V3, AGENTX_DSV4_GB200_GB300],
+  },
   {
     slug: 'inference-engine',
     term: 'Inference engine',
@@ -117,9 +252,15 @@ const entries = [
     significance:
       'Different products need different operating points. Voice and interactive coding demand high token rates, while offline summarization can trade interactivity for much more aggregate throughput. Comparing hardware at unmatched interactivity can therefore produce a misleading winner.',
     benchmarkContext:
-      'InferenceX plots tokens per second per user against throughput or cost. Iso-interactivity tables interpolate each system’s Pareto frontier at the same token rate so the comparison holds user experience constant.',
+      'InferenceX plots tokens per second per user against throughput or cost. Iso-interactivity tables interpolate each system’s Pareto frontier at the same token rate so the comparison holds user experience constant. Because this axis ignores the wait before the first token, agentic charts also offer E2E Normalized Interactivity, which folds TTFT into the same unit.',
     measurement: { label: 'Typical unit', value: 'tokens/second/user (tok/s/user)' },
-    relatedTerms: ['time-per-output-token', 'throughput', 'iso-interactivity', 'latency'],
+    relatedTerms: [
+      'time-per-output-token',
+      'throughput',
+      'iso-interactivity',
+      'e2e-normalized-interactivity',
+      'latency',
+    ],
     articleSlugs: [INFERENCEMAX, INFERENCEX_V2, MI355X_KIMI, TILERT],
   },
   {
@@ -137,7 +278,7 @@ const entries = [
       'Low latency can require smaller batches or more parallel resources, which may reduce hardware utilization and increase cost. Good serving design chooses a latency service level and then maximizes throughput within it.',
     benchmarkContext:
       'InferenceX exposes workload shape and concurrency alongside interactivity. This keeps a high-throughput batch point from being mistaken for a low-latency serving point.',
-    relatedTerms: ['time-to-first-token', 'time-per-output-token', 'interactivity', 'concurrency'],
+    relatedTerms: ['time-to-first-token', 'time-per-output-token', 'tail-latency', 'interactivity'],
     articleSlugs: [INFERENCEMAX, INFERENCEX_V2],
   },
   {
@@ -228,8 +369,14 @@ const entries = [
     significance:
       'The frontier prevents noisy or poorly tuned points from distorting comparisons and makes the real tradeoff visible. There is still no universal winner along the curve: the best point depends on the user’s minimum interactivity or maximum cost target.',
     benchmarkContext:
-      'InferenceX connects Pareto-optimal points from a concurrency and configuration sweep. Iso-interactivity comparisons interpolate along those frontiers because direct comparisons of arbitrary raw points can mislead.',
-    relatedTerms: ['throughput', 'interactivity', 'iso-interactivity', 'concurrency'],
+      'InferenceX connects Pareto-optimal points from a concurrency and configuration sweep. Iso-interactivity comparisons interpolate along those frontiers because direct comparisons of arbitrary raw points can mislead. Best-per-SKU views now merge the permitted optimizations into one curve per model, chip SKU, and engine, so adjacent points on a single line may differ in speculative decoding, disaggregation, or KV cache offload. Each point still exposes the exact configuration that produced it.',
+    relatedTerms: [
+      'throughput',
+      'interactivity',
+      'iso-interactivity',
+      'kv-cache-offload',
+      'concurrency',
+    ],
     articleSlugs: [INFERENCEMAX, INFERENCEX_V2, MI355X_GLM5],
   },
   {
@@ -245,9 +392,9 @@ const entries = [
     significance:
       'Holding user experience constant avoids a common benchmark error: declaring a high-throughput system faster when it reaches that throughput only by serving every request more slowly.',
     benchmarkContext:
-      'InferenceX articles use iso-interactivity tables for hardware, precision, and software comparisons. Values outside a measured frontier are marked unreachable and are not extrapolated beyond observed data.',
+      'InferenceX articles use iso-interactivity tables for hardware, precision, and software comparisons. Values outside a measured frontier are marked unreachable and are not extrapolated beyond observed data. The frontier is always built on throughput against interactivity; cost per million tokens and joules per token are then derived from the interpolated throughput rather than splined on their own, because each is a per-chip constant divided by that throughput and splining it separately would break the identity between knots.',
     relatedTerms: ['interactivity', 'pareto-frontier', 'throughput', 'performance-per-dollar'],
-    articleSlugs: [B200_GLM5, B200_MINIMAX, B200_KIMI, GB300_DSV4],
+    articleSlugs: [B200_GLM5, B200_MINIMAX, B200_KIMI, GB300_DSV4, AGENTX_GLM_SGLANG],
   },
   {
     slug: 'input-output-sequence-length',
@@ -264,8 +411,8 @@ const entries = [
     significance:
       'Results from different sequence lengths are not interchangeable. A configuration tuned for short chat prompts can rank differently on long-context summarization or reasoning because compute, memory capacity, and bandwidth pressure shift.',
     benchmarkContext:
-      'InferenceX includes ISL and OSL in chart labels and recipe descriptions. Compare systems on the same workload shape before attributing a difference to hardware or software.',
-    relatedTerms: ['prefill', 'decode', 'kv-cache', 'time-to-first-token'],
+      'InferenceX includes ISL and OSL in chart labels and recipe descriptions. Compare systems on the same workload shape before attributing a difference to hardware or software. Agentic runs have no single pair to quote: lengths follow a roughly lognormal distribution, so the point detail view plots the fitted distribution and reports percentiles, and a p90 or p99 input can be several times the median.',
+    relatedTerms: ['prefill', 'decode', 'kv-cache', 'agentx', 'time-to-first-token'],
     articleSlugs: [INFERENCEMAX, B200_GLM5, GB300_DSV4],
   },
   {
@@ -293,7 +440,7 @@ const entries = [
       'iso-interactivity',
       'performance-per-dollar',
     ],
-    articleSlugs: [INFERENCEX_V2, B200_KIMI, B200_GLM5, GB300_DSV4],
+    articleSlugs: [INFERENCEX_V2, B200_KIMI, B200_GLM5, GB300_DSV4, AGENTX_GLM_SGLANG],
   },
   {
     slug: 'performance-per-dollar',
@@ -309,14 +456,15 @@ const entries = [
     significance:
       'Peak chip FLOPS account for only part of serving economics. Memory, networking, software maturity, numerical precision, and achievable utilization all affect the measured output behind the ratio.',
     benchmarkContext:
-      'InferenceX compares perf/$ at matched interactivity and names the TCO inputs used. Ratios should not be carried across different model, sequence-length, precision, or latency regimes.',
+      'InferenceX compares perf/$ at matched interactivity and names the TCO inputs used. Ratios should not be carried across different model, sequence-length, precision, or latency regimes. The charts express the same economics as tokens per dollar, which reads in the higher-is-better direction and is the default y-axis.',
     relatedTerms: [
       'cost-per-million-tokens',
+      'tokens-per-dollar',
       'total-cost-of-ownership',
       'iso-interactivity',
       'throughput',
     ],
-    articleSlugs: [B200_GLM5, B200_MINIMAX, B200_KIMI, MI355X_GLM5],
+    articleSlugs: [B200_GLM5, B200_MINIMAX, B200_KIMI, MI355X_GLM5, AGENTX_DSV4_MI355X_B200],
   },
   {
     slug: 'total-cost-of-ownership',
@@ -339,7 +487,7 @@ const entries = [
       'tokens-per-megawatt',
       'throughput',
     ],
-    articleSlugs: [INFERENCEMAX, INFERENCEX_V2, GB200_R1, VR_RUBIN],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2, GB200_R1, VR_RUBIN, JALAPENO],
   },
   {
     slug: 'tokens-per-megawatt',
@@ -355,15 +503,16 @@ const entries = [
     significance:
       'Power availability is often the binding constraint on new AI deployments. A system that produces more tokens per provisioned megawatt can serve more demand from the same utility allocation even if its individual accelerators draw more power.',
     benchmarkContext:
-      'Compare tokens/MW at the same model, workload shape, precision, and interactivity. Otherwise a high-throughput low-interactivity point can appear efficient while failing the target user experience.',
+      'Compare tokens/MW at the same model, workload shape, precision, and interactivity. Otherwise a high-throughput low-interactivity point can appear efficient while failing the target user experience. Energy per token expresses the same provisioned budget per unit of output, and InferenceX additionally reports measured accelerator energy where the telemetry is trustworthy.',
     measurement: { label: 'Typical unit', value: 'tokens/second per provisioned utility MW' },
     relatedTerms: [
       'throughput',
+      'energy-per-token',
       'interactivity',
       'total-cost-of-ownership',
       'performance-per-dollar',
     ],
-    articleSlugs: [INFERENCEMAX, DEEPSEEK_V4, VR_RUBIN],
+    articleSlugs: [INFERENCEMAX, DEEPSEEK_V4, VR_RUBIN, JALAPENO],
   },
   {
     slug: 'prefill',
@@ -380,7 +529,13 @@ const entries = [
       'Prefill has a different resource profile from decode. When both share the same workers, large prompt jobs can interrupt decode batches and make streaming latency less predictable.',
     benchmarkContext:
       'Disaggregated recipes place prefill on a separately sized chip pool. When reading a result, check the prefill tensor parallelism, chip count, input length, and whether KV state must cross a network before decode.',
-    relatedTerms: ['decode', 'kv-cache', 'time-to-first-token', 'disaggregated-inference'],
+    relatedTerms: [
+      'decode',
+      'kv-cache',
+      'chunked-prefill',
+      'time-to-first-token',
+      'arithmetic-intensity',
+    ],
     articleSlugs: [INFERENCEX_V2, GB300_DSV4, GB200_KIMI],
   },
   {
@@ -415,15 +570,16 @@ const entries = [
     significance:
       'KV-cache pressure limits concurrency and long-context serving. Cache quantization, paged allocation, latent attention, prefix reuse, and disaggregated transfer systems all target its capacity or movement cost.',
     benchmarkContext:
-      'InferenceX disables prefix caching for random-data comparisons unless a recipe states otherwise. That keeps unrelated requests from receiving artificial cache hits and makes raw serving stacks easier to compare.',
+      'InferenceX disables prefix caching for fixed-sequence comparisons on random data unless a recipe states otherwise, which keeps unrelated requests from receiving artificial cache hits. AgentX is the deliberate exception: its replayed sessions are built to reuse prefixes, so cache capacity, eviction policy, and offload are part of what the scenario measures.',
     relatedTerms: [
       'prefill',
       'decode',
       'prefix-caching',
+      'kv-cache-offload',
       'multi-head-latent-attention',
       'high-bandwidth-memory',
     ],
-    articleSlugs: [INFERENCEX_V2, MI355X_KIMI, SGLANG_056, KIMI_K3],
+    articleSlugs: [INFERENCEX_V2, MI355X_KIMI, SGLANG_056, KIMI_K3, AGENTX_V3],
   },
   {
     slug: 'prefix-caching',
@@ -439,9 +595,16 @@ const entries = [
     significance:
       'Production workloads with repeated prefixes may outperform synthetic random-token benchmarks. The benefit depends on hit rate, cache capacity, eviction policy, and whether requests route to workers that hold the needed state.',
     benchmarkContext:
-      'InferenceX generally disables prefix caching on random datasets to isolate full prompt processing from cache policy. Treat benchmark cost as a no-hit baseline unless the recipe says otherwise.',
-    relatedTerms: ['kv-cache', 'prefill', 'time-to-first-token', 'nvidia-dynamo'],
-    articleSlugs: [INFERENCEX_V2, GB200_KIMI, KIMI_K3],
+      'InferenceX disables prefix caching on random fixed-sequence datasets to isolate full prompt processing from cache policy, so treat those numbers as a no-hit baseline. AgentX inverts this: hit rate is a reported quantity there, shown per point alongside the offload tier it was served from, because reuse is the defining property of the workload.',
+    relatedTerms: [
+      'kv-cache',
+      'kv-cache-offload',
+      'kv-aware-routing',
+      'prefill',
+      'time-to-first-token',
+      'nvidia-dynamo',
+    ],
+    articleSlugs: [AGENTIC_WORKLOADS, INFERENCEX_V2, GB200_KIMI, KIMI_K3, AGENTX_V3],
   },
   {
     slug: 'disaggregated-inference',
@@ -460,7 +623,16 @@ const entries = [
     benchmarkContext:
       'A disagg label identifies the serving layout, not its performance. Judge it from the prefill and decode world sizes, TP/EP layout, framework, network domain, and the interactivity range where its frontier leads.',
     relatedTerms: ['prefill', 'decode', 'kv-cache', 'nvidia-dynamo', 'wide-expert-parallelism'],
-    articleSlugs: [INFERENCEX_V2, GB200_R1, GB300_DSV4, GB200_KIMI, TILERT],
+    articleSlugs: [
+      INFERENCEX_V2,
+      GB200_R1,
+      GB300_DSV4,
+      GB200_KIMI,
+      TILERT,
+      AGENTX_V3,
+      AGENTX_DSV4_GB200_GB300,
+      JALAPENO,
+    ],
   },
   {
     slug: 'speculative-decoding',
@@ -476,9 +648,9 @@ const entries = [
     significance:
       'The speedup depends on how many draft tokens are accepted and on the cost of drafting and verification. Dense and MoE models can behave differently because verifying several positions may activate more expert weights.',
     benchmarkContext:
-      'Compare speculative recipes at realistic acceptance rates and verify model quality. InferenceX distinguishes MTP-enabled and disabled curves because the benefit changes across concurrency and interactivity.',
-    relatedTerms: ['multi-token-prediction', 'decode', 'batching', 'mixture-of-experts'],
-    articleSlugs: [INFERENCEX_V2, DEEPSEEK_V4, B200_GLM5, KIMI_K3],
+      'Fixed-sequence scenarios keep speculative decoding as part of a curve’s identity, so MTP-enabled and disabled recipes plot separately. Agentic curves instead treat it as point-level metadata and merge the points, with the method named in each tooltip, because AgentX reports the best available curve per model, chip SKU, and engine. Since replayed AgentX content is synthetic, a speculator would accept an unrepresentative number of draft tokens, so runs apply an acceptance length collected per model, speculator, draft length, and thinking mode on an external agentic coding dataset.',
+    relatedTerms: ['multi-token-prediction', 'acceptance-length', 'decode', 'batching', 'agentx'],
+    articleSlugs: [INFERENCEX_V2, DEEPSEEK_V4, B200_GLM5, KIMI_K3, AGENTX_V3],
   },
   {
     slug: 'multi-token-prediction',
@@ -654,7 +826,7 @@ const entries = [
     benchmarkContext:
       'System topology determines the communication domain. A B200 in an eight-chip node and a GB200 NVL72 expose related silicon through different scale-up group sizes.',
     relatedTerms: ['nvlink', 'wide-expert-parallelism', 'all-to-all', 'tensor-parallelism'],
-    articleSlugs: [INFERENCEX_V2, GB200_R1, GB200_KIMI],
+    articleSlugs: [INFERENCEX_V2, GB200_R1, GB200_KIMI, JALAPENO],
   },
   {
     slug: 'high-bandwidth-memory',
@@ -672,7 +844,7 @@ const entries = [
     benchmarkContext:
       'InferenceX hardware comparisons separate HBM capacity from bandwidth. For example, GB300’s larger capacity fits wider prefill/decode layouts than GB200 despite similar bandwidth per chip.',
     relatedTerms: ['memory-bandwidth', 'decode', 'kv-cache', 'quantization'],
-    articleSlugs: [GB300_DSV4, B200_KIMI, MI355X_DSV4],
+    articleSlugs: [GB300_DSV4, B200_KIMI, MI355X_DSV4, JALAPENO],
   },
   {
     slug: 'memory-bandwidth',
@@ -725,7 +897,7 @@ const entries = [
       'A nominal format alone says little about speed or quality. Conversion quality, model calibration, outliers, kernel maturity, and hardware support determine the result.',
     benchmarkContext:
       'InferenceX treats precision as a first-class recipe dimension and pairs throughput measurements with accuracy checks. Compare FP8, FP4, NVFP4, MXFP4, and INT4 only when the model, workload, engine, and quality bar are compatible.',
-    relatedTerms: ['fp8', 'fp4', 'nvfp4', 'mxfp4', 'high-bandwidth-memory'],
+    relatedTerms: ['fp8', 'fp4', 'int4', 'bf16', 'kv-cache-quantization'],
     articleSlugs: [INFERENCEX_V2, B200_KIMI, B200_GLM5, MI355X_DSV4],
   },
   {
@@ -762,7 +934,7 @@ const entries = [
     benchmarkContext:
       'InferenceX identifies concrete recipe formats such as NVFP4 and MXFP4 where possible and validates representative configurations. Each FP4 line still has its own numerical and operational behavior.',
     relatedTerms: ['quantization', 'nvfp4', 'mxfp4', 'fp8', 'memory-bandwidth'],
-    articleSlugs: [INFERENCEX_V2, B200_KIMI, MI355X_DSV4, SGLANG_056],
+    articleSlugs: [INFERENCEX_V2, B200_KIMI, MI355X_DSV4, SGLANG_056, AGENTX_QWEN_B300],
   },
   {
     slug: 'nvfp4',
@@ -798,7 +970,7 @@ const entries = [
     benchmarkContext:
       'InferenceX records MXFP4 as part of a complete engine and hardware recipe. Comparisons with NVFP4 or FP8 should use the same model, sequence length, quality requirements, and interactivity target.',
     relatedTerms: ['fp4', 'quantization', 'nvfp4', 'rocm', 'memory-bandwidth'],
-    articleSlugs: [MI355X_KIMI, INFERENCEX_V2],
+    articleSlugs: [MI355X_KIMI, INFERENCEX_V2, JALAPENO],
   },
   {
     slug: 'mixture-of-experts',
@@ -927,7 +1099,14 @@ const entries = [
     benchmarkContext:
       'InferenceX continuously reruns pinned SGLang recipes. Version-to-version curves show where a change affects performance across the operating range and reveal regressions or gains hidden by one peak point.',
     relatedTerms: ['inference-engine', 'eagle', 'vllm', 'rocm', 'cuda'],
-    articleSlugs: [SGLANG_056, B200_GLM5, MI355X_DSV4, MI355X_GLM5, MI355X_QWEN],
+    articleSlugs: [
+      SGLANG_056,
+      B200_GLM5,
+      MI355X_DSV4,
+      MI355X_GLM5,
+      MI355X_QWEN,
+      AGENTX_QWEN_SGLANG,
+    ],
   },
   {
     slug: 'tensorrt-llm',
@@ -945,7 +1124,7 @@ const entries = [
     benchmarkContext:
       'InferenceX includes direct TensorRT-LLM and Dynamo TensorRT-LLM configurations and also tracks cases where SGLang or vLLM uses a TRT-LLM-derived kernel backend.',
     relatedTerms: ['inference-engine', 'cuda', 'nvidia-dynamo', 'nvfp4', 'sglang'],
-    articleSlugs: [GB200_R1, INFERENCEX_V2, B200_GLM5, B200_MINIMAX],
+    articleSlugs: [GB200_R1, INFERENCEX_V2, B200_GLM5, B200_MINIMAX, AGENTX_M3_TRT],
   },
   {
     slug: 'nvidia-dynamo',
@@ -971,6 +1150,1299 @@ const entries = [
     ],
     articleSlugs: [GB200_R1, GB300_DSV4, GB200_KIMI, INFERENCEX_V2],
   },
+  {
+    slug: 'e2e-normalized-interactivity',
+    term: 'E2E Normalized Interactivity',
+    aliases: ['normalized interactivity', 'OSL/E2EL'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'This metric asks how fast a whole answer arrives, counting the wait before the first word as well as the streaming speed after it.',
+    definition:
+      'E2E Normalized Interactivity is the effective per-user token rate across a complete request: output tokens divided by end to end latency.',
+    explanation:
+      'Substituting end to end latency for time to first token plus output length times time per output token gives approximately 1 divided by the sum of inter-token latency and TTFT divided by output tokens. The result is ordinary interactivity plus a penalty proportional to first-token wait. Normalized here means normalized by output length, not scaled to a 0 to 1 score or measured against another system.',
+    significance:
+      'Interactivity alone rewards a recipe that streams quickly after making the user wait, and TTFT alone rewards one that starts fast and then crawls. Folding both into one number exposes operating points that look strong on a single axis. Short responses feel the TTFT penalty most, because there are fewer tokens to amortize the wait across.',
+    benchmarkContext:
+      'InferenceX exposes this as an experimental x-axis mode for agentic runs, which is why it needs persisted per-request traces and is unavailable for unofficial-run overlays. It is deliberately imperfect: it penalizes high TTFT heavily and does not capture every nuance of prefill and decode disaggregation, so AgentX submissions still optimize interactivity and TTFT separately.',
+    measurement: { label: 'Typical unit', value: 'tokens/second/user (tok/s/user)' },
+    relatedTerms: ['interactivity', 'time-to-first-token', 'time-per-output-token', 'latency'],
+    articleSlugs: [AGENTX_V3, AGENT_BENCHMARK, AGENTX_GLM_ATOM],
+  },
+  {
+    slug: 'tokens-per-dollar',
+    term: 'Tokens per dollar',
+    abbreviation: 'tok/$',
+    aliases: ['tokens per $1 USD', 'tokens per RMB'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'Tokens per dollar asks how many tokens one dollar of infrastructure spend buys, so a bigger number is a cheaper system.',
+    definition:
+      'Tokens per dollar is the count of tokens a configuration produces for one unit of modeled infrastructure cost, the reciprocal of cost per token.',
+    explanation:
+      'The figure follows directly from throughput per chip and the modeled cost per chip hour, so it carries the same assumptions as cost per million tokens while reading in the direction most people reason about capacity. InferenceX publishes it for total, input, and output tokens, against each cost basis it models, and in Chinese yuan alongside US dollars.',
+    significance:
+      'Cost per million tokens and tokens per dollar rank systems identically, but a metric that rises with better hardware sits the same way up as throughput, so a chart mixing the two no longer inverts halfway down the axis. The absolute value depends entirely on the cost model behind it, so it travels only with its stated basis.',
+    benchmarkContext:
+      'Total tokens per $1 USD is the default y-axis on the InferenceX inference charts. Read it against the TCO row shown above the chart, and compare only within one cost basis: owning at hyperscaler rates, owning at neocloud rates, and three year rental produce different numbers for identical silicon.',
+    measurement: { label: 'Typical unit', value: 'tokens per $1 USD (tok/$)' },
+    relatedTerms: [
+      'cost-per-million-tokens',
+      'performance-per-dollar',
+      'total-cost-of-ownership',
+      'throughput',
+    ],
+    articleSlugs: [AGENTX_V3, INFERENCEX_V2, B200_GLM5, AGENTX_QWEN_B300],
+  },
+  {
+    slug: 'energy-per-token',
+    term: 'Energy per token',
+    abbreviation: 'J/token',
+    aliases: ['joules per token', 'joules per query'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'Energy per token is how much electricity the system spends to produce one token, the power-side counterpart of cost per token.',
+    definition:
+      'Energy per token is the electrical energy consumed per token produced, reported either from all-in provisioned power or from measured accelerator telemetry.',
+    explanation:
+      'The two bases answer different questions and are not interchangeable. All-in provisioned figures divide a facility power budget, including power delivery and cooling overhead, by measured token rates. Measured figures come from accelerator telemetry during the run and describe the chips alone. InferenceX also reports measured energy per successful query and average power as a percentage of thermal design power.',
+    significance:
+      'Power, not capital, is often the binding constraint on new deployments, and a system that produces more tokens per joule serves more demand from the same utility allocation. The percentage of TDP figure separately reveals how hard a recipe actually drives its accelerators, which a token-normalized number alone hides.',
+    benchmarkContext:
+      'Read the label before comparing: all-in provisioned and measured values differ by the facility overhead between them. InferenceX withholds measured energy where the underlying telemetry is invalid or its scope is ambiguous, so a missing value means the measurement could not be trusted rather than that the run drew no power.',
+    measurement: { label: 'Typical unit', value: 'joules per token (J/tok)' },
+    relatedTerms: ['tokens-per-megawatt', 'throughput', 'total-cost-of-ownership', 'concurrency'],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2, AGENTX_V3, JALAPENO],
+  },
+  {
+    slug: 'context-parallelism',
+    term: 'Context parallelism',
+    abbreviation: 'CP',
+    aliases: ['PCP', 'DCP', 'sequence parallelism'],
+    category: 'Parallelism',
+    plainEnglish:
+      'Context parallelism splits one long prompt across several chips so they share the work of reading it and of scanning its stored attention state.',
+    definition:
+      'Context parallelism shards query tokens across accelerators, in a prefill form known as PCP and a decode form known as DCP.',
+    explanation:
+      'PCP gives each rank a chunk of the query while keys and values are passed around a ring, which parallelizes the compute-bound prefill and stops one rank absorbing an entire long prompt. DCP shards the KV cache itself, so every rank scans its own slice and the partial attention results merge flash-decode style. Because decode is memory-bandwidth bound, parallel KV reads raise the achievable token rate.',
+    significance:
+      'Tensor parallelism replicates the full KV cache on each rank and data-parallel attention pins a session to whichever rank owns its shard, so neither scales cleanly as contexts reach hundreds of thousands of tokens. Context parallelism attacks that directly, and its gain grows with input length rather than with batch size.',
+    benchmarkContext:
+      'InferenceX surfaces DCP and PCP degrees in point tooltips and parallelism labels alongside TP, EP, and DP. Support is uneven across vendors: the technique remains part of the practical CUDA advantage because the AMD attention backends were still listed as unsupported in the vLLM matrix at the time of the AgentX 1.0 results.',
+    relatedTerms: ['tensor-parallelism', 'data-parallelism', 'kv-cache', 'prefill', 'decode'],
+    articleSlugs: [AGENTX_V3, INFERENCEX_V2, AGENTX_M3_RACK],
+  },
+  {
+    slug: 'kv-cache-offload',
+    term: 'KV cache offload',
+    aliases: ['CPU offload', 'KV offloading'],
+    category: 'Serving',
+    plainEnglish:
+      'KV cache offload parks attention state the chips cannot hold in host memory, so a long session can be resumed instead of recomputed.',
+    definition:
+      'KV cache offload moves KV blocks out of accelerator memory into a slower tier, usually host DRAM, and loads them back when a later request reuses that prefix.',
+    explanation:
+      'Offload is usually a write-through cache: a prefix written to the HBM cache is also written to the slower tier, so it helps most when the offload pool is roughly one and a half to three times HBM capacity. Reloading a long prefix beats recomputing it by a wide margin at agentic context lengths, but the arithmetic reverses for short prompts, where the transfer costs more than the prefill it avoids.',
+    significance:
+      'Long agentic sessions exceed HBM KV capacity well before they exceed a plausible DRAM budget, so offload decides how many concurrent conversations stay resumable. It also shifts the bottleneck: once prefixes survive, store and load paths, transfer batching, and index bookkeeping become the costs worth optimizing.',
+    benchmarkContext:
+      'InferenceX rings every point that used KV offload with a dashed halo, whether or not it is Pareto optimal, and the point detail view names the offload type and engine alongside the chip and CPU cache hit rates. Offload is an allowed but optional optimization, so a single curve can mix points with and without it.',
+    relatedTerms: ['kv-cache', 'prefix-caching', 'kv-cache-manager', 'high-bandwidth-memory'],
+    articleSlugs: [AGENTX_V3, AGENTIC_WORKLOADS, KIMI_K3, AGENTX_DSV4_B200_B300],
+  },
+  {
+    slug: 'kv-cache-manager',
+    term: 'KV cache manager',
+    aliases: ['Mooncake', 'LMCache', 'HiCache'],
+    category: 'Software',
+    plainEnglish:
+      'A KV cache manager is the component that stores attention state outside the chips and decides what to keep, evict, and fetch back.',
+    definition:
+      'A KV cache manager is a pluggable layer beneath an inference engine that stores reusable KV blocks across memory tiers and manages their placement, eviction, and transfer.',
+    explanation:
+      'Engines expose a connector interface, so managers such as Mooncake Store, LMCache, and SGLang HiCache can serve different runtimes. The manager keys blocks by prefix hash and places them in host DRAM, local NVMe, or a remote backend, while a separate transfer engine such as Mooncake Transfer Engine or NIXL performs the byte movement. Several paths can coexist inside one engine.',
+    significance:
+      'Once a workload reuses prefixes heavily, correctness and accounting in this layer matter as much as kernel speed. Hybrid-attention models make it harder still, because a model carrying several cache groups with different shapes and lifetimes cannot be described by a connector that assumes one uniform block geometry.',
+    benchmarkContext:
+      'InferenceX records the KV offload backend as run metadata and shows it in the AgentX point detail view. Framework labels name the combination rather than the engine alone, so a recipe reads as Mooncake ATOMesh or MoRI SGLang instead of just its engine.',
+    relatedTerms: ['kv-cache-offload', 'kv-cache', 'prefix-caching', 'inference-engine'],
+    articleSlugs: [AGENTX_V3, KIMI_K3, INFERENCEX_V2],
+  },
+  {
+    slug: 'kv-aware-routing',
+    term: 'KV-aware routing',
+    aliases: ['cache-aware routing', 'session affinity'],
+    category: 'Serving',
+    plainEnglish:
+      'KV-aware routing sends a request to the worker that already holds its conversation state, instead of to whichever worker is least busy.',
+    definition:
+      'KV-aware routing selects a worker using where cached prefix state already resides, rather than on queue depth or load alone.',
+    explanation:
+      'A request carrying no reusable history can go anywhere, and load balancing is the only question worth asking. A request carrying megabytes of cached prefix is different: sending it to an idle worker that lacks that prefix pays for the whole prompt again. Routers therefore track cache events, hash sessions to consistent workers, and keep data-parallel ranks sticky to the sessions whose state they own.',
+    significance:
+      'Under data-parallel attention each rank owns a private slice of the cache pool, so a long session landing on the wrong rank recomputes everything and the measured hit rate collapses far below its theoretical ceiling. Affinity alone is not enough either, since unchecked stickiness concentrates load on one hot worker, so cache balance has to enter the routing score.',
+    benchmarkContext:
+      'Routing sits outside the engine, so InferenceX treats it as part of the recipe: labels such as Dynamo vLLM, llm-d vLLM, and Mooncake ATOMesh name the orchestration layer as well as the runtime. Its cost scales with the number and length of live prefixes rather than with tokens generated, which is why it can become the bottleneck on agentic traffic once kernels improve.',
+    relatedTerms: [
+      'prefix-caching',
+      'nvidia-dynamo',
+      'data-parallelism',
+      'disaggregated-inference',
+    ],
+    articleSlugs: [AGENTX_V3, AGENTIC_WORKLOADS, AGENTX_M3_RACK],
+  },
+  {
+    slug: 'tilert',
+    term: 'TileRT',
+    aliases: ['TileRT engine'],
+    category: 'Software',
+    plainEnglish:
+      'TileRT is an inference runtime built for very fast single-user generation, compiling a model into one resident program instead of many separate kernel launches.',
+    definition:
+      'TileRT is an inference engine that targets ultra-low-latency serving by abolishing the individual kernel as the unit of execution.',
+    explanation:
+      'A conventional runtime dispatches a sequence of kernels for every decode step, and at very small batch sizes the launch and scheduling overhead between them dominates the arithmetic. A persistent engine kernel keeps the work resident on the accelerator instead, which is what makes the far-right end of the interactivity axis reachable at all.',
+    significance:
+      'The high-interactivity corner of the frontier is a different engineering problem from the high-throughput corner, and an engine tuned for one rarely wins the other. Recipes that reach hundreds of tokens per second per user matter for latency-critical products even when their aggregate throughput per chip is unremarkable.',
+    benchmarkContext:
+      'InferenceX reports TileRT as its own framework label and deliberately retains it in best-per-SKU views, because a curve that only survives where it dominates on throughput would drop the operating points TileRT exists to serve. Compare it at matched interactivity rather than on peak throughput alone.',
+    relatedTerms: ['inference-engine', 'interactivity', 'decode', 'sglang'],
+    articleSlugs: [TILERT, AGENTX_V3],
+  },
+  {
+    slug: 'recipe',
+    term: 'Recipe',
+    aliases: ['configuration', 'serving recipe'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'A recipe is the complete set of choices behind one curve: which model, engine, image, precision, parallelism, and workload were run.',
+    definition:
+      'A recipe is the fully specified combination of model, inference engine and container image, numerical precision, parallelism strategy, chip system, and workload that produces one measured curve.',
+    explanation:
+      'Every point on InferenceX belongs to a recipe, and a concurrency sweep across one recipe traces out its curve. Changing any element produces a different recipe rather than a variation of the same one, which is why an engine image bump is reported as its own result rather than folded into an existing line.',
+    significance:
+      'Peak chip specifications do not describe serving performance, and the same silicon can differ by multiples across recipes. Naming the whole combination is what makes a claim checkable: a number without its recipe cannot be reproduced or fairly compared against another vendor.',
+    benchmarkContext:
+      'InferenceX benchmark configs mainly track the published vLLM and SGLang cookbooks on upstream images, so results reflect what users can actually deploy rather than images tuned for the benchmark. Point tooltips expose the recipe behind each point along with links to the run provenance.',
+    relatedTerms: ['inference-engine', 'pareto-frontier', 'concurrency', 'quantization'],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2, AGENTX_V3],
+  },
+  {
+    slug: 'tail-latency',
+    term: 'Tail latency',
+    aliases: ['p90', 'p99', 'percentile latency'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'Tail latency describes the slowest requests rather than the typical one, because the unlucky few are what users actually notice.',
+    definition:
+      'Tail latency is the latency at a high percentile of the request distribution, such as p90 or p99, rather than the mean or median.',
+    explanation:
+      'A percentile answers a different question from an average. A p90 of five seconds means one request in ten waited at least that long, and that request may be the one blocking an agent from continuing. Distributions in real serving are heavily skewed, so the mean can sit far below the tail and hide it completely.',
+    significance:
+      'Capacity planning is usually written against a percentile, not an average, because a service that is fast on average and slow at the tail still fails its users. Optimizations can also improve the mean while worsening the tail, which a single-number summary would report as an unambiguous win.',
+    benchmarkContext:
+      'InferenceX reports percentile-qualified metrics and labels the percentile on the axis, so p90 TTFT and mean TTFT are never mixed on one comparison. Agentic runs are especially skewed, because end to end latency scales with output length and the longest generations dominate the tail.',
+    relatedTerms: ['latency', 'time-to-first-token', 'interactivity', 'concurrency'],
+    articleSlugs: [AGENTX_V3, INFERENCEX_V2, AGENT_BENCHMARK],
+  },
+  {
+    slug: 'service-level-objective',
+    term: 'Service level objective',
+    abbreviation: 'SLO',
+    aliases: ['latency target', 'SLA target'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'An SLO is the performance promise a deployment has to keep, such as a first token within one second for nine requests in ten.',
+    definition:
+      'A service level objective is a stated target for a serving metric, usually expressed as a percentile bound on latency or interactivity.',
+    explanation:
+      'A useful SLO names a metric, a percentile, and a threshold together. Serving capacity is then whatever throughput the system sustains without breaching it, which is a smaller number than peak throughput and the only one an operator can safely provision against.',
+    significance:
+      'Every point on a throughput curve is reachable, but only part of the curve satisfies a given promise. Two systems can look close on peak throughput and differ sharply in how much of that throughput survives an interactivity or first-token bound.',
+    benchmarkContext:
+      'InferenceX does not impose one industry SLO, because acceptable targets differ by product: interactive coding needs a high token rate, while batch processing tolerates seconds of first-token delay. Read the frontier at your own threshold instead of comparing peak values.',
+    relatedTerms: ['latency', 'tail-latency', 'interactivity', 'iso-interactivity'],
+    articleSlugs: [AGENTX_V3, INFERENCEMAX, AGENT_BENCHMARK],
+  },
+  {
+    slug: 'acceptance-length',
+    term: 'Acceptance length',
+    abbreviation: 'AL',
+    aliases: ['acceptance rate', 'draft acceptance'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'Acceptance length is how many drafted tokens the full model actually approves per verification step, which is what decides whether speculation pays off.',
+    definition:
+      'Acceptance length is the average number of speculatively drafted tokens accepted by the target model in one verification pass.',
+    explanation:
+      'Speculative decoding only saves time when drafts survive verification. An acceptance length near one means the draft and verify machinery ran for nothing, while a high value amortizes one expensive target-model step across several emitted tokens. The value depends on the speculator, the draft length, the model, and the content being generated.',
+    significance:
+      'Because acceptance depends on content, a benchmark can accidentally decide the result. Synthetic or anonymized text is out of distribution for a speculator trained on real language, so measured acceptance drifts away from what production would see, in either direction.',
+    benchmarkContext:
+      'AgentX replays anonymized traces filled with synthetic tokens, so it does not let acceptance emerge from that content. Runs instead apply a fixed acceptance length collected per model, speculator, draft length, and thinking mode on an external agentic coding dataset, which keeps the comparison vendor neutral.',
+    relatedTerms: ['speculative-decoding', 'multi-token-prediction', 'eagle', 'agentx'],
+    articleSlugs: [AGENTX_V3, INFERENCEX_V2, DEEPSEEK_V4],
+  },
+  {
+    slug: 'unofficial-run',
+    term: 'Unofficial run',
+    aliases: ['unofficial overlay', 'community run'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'An unofficial run is a benchmark run that has not been ingested into the published dataset but can still be drawn on top of the charts from its URL.',
+    definition:
+      'An unofficial run is a CI benchmark run loaded into the dashboard as an overlay from its run identifier, rather than from the published database.',
+    explanation:
+      'Adding a run identifier to the page URL fetches that run and draws its points, rooflines, and rows alongside the official data in distinct overlay colors. The overlay is a separate rendering path, so it can be toggled per hardware type or dismissed per run without disturbing the official selection underneath.',
+    significance:
+      'Publishing a result usually lags producing it, and a contributor tuning a recipe needs to see the new curve against the current frontier before anyone decides to ingest it. The overlay makes an in-flight run reviewable without granting it the standing of a published measurement.',
+    benchmarkContext:
+      'Overlay colors come from a run-index palette rather than the hardware palette, so an overlay is never mistaken for official data. Some views require data that is only persisted for ingested runs, and those views state that overlays are unavailable rather than silently omitting them.',
+    relatedTerms: ['recipe', 'pareto-frontier', 'inference-engine', 'agentx'],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2],
+  },
+  {
+    slug: 'chunked-prefill',
+    term: 'Chunked prefill',
+    aliases: ['split prefill', 'piecewise prefill'],
+    category: 'Serving',
+    plainEnglish:
+      'Chunked prefill reads a long prompt in slices instead of all at once, so other users keep receiving tokens while it is being read.',
+    definition:
+      'Chunked prefill splits prompt processing into fixed-size token chunks that the scheduler interleaves with ongoing decode work.',
+    explanation:
+      'An unsplit prefill occupies the accelerator for as long as the whole prompt takes, and every user already streaming waits behind it. Splitting the prompt lets the scheduler alternate, so decode continues between chunks. The chunk size is a tuning knob: larger chunks prefill more efficiently, smaller chunks interrupt decode less.',
+    significance:
+      'The technique converts a first-token problem for one user into a small, steady tax on everyone else, which is usually the better trade. It matters far more as prompts grow, since a single unsplit hundred-thousand-token prefill would stall a deployment outright.',
+    benchmarkContext:
+      'Chunk size is part of the recipe and can differ across points on the same curve, so a jump in throughput may reflect a retune rather than new hardware. Long agentic prompts make the setting consequential in a way fixed short-prompt scenarios never expose.',
+    relatedTerms: ['prefill', 'decode', 'batching', 'time-to-first-token'],
+    articleSlugs: [AGENTX_V3, INFERENCEX_V2, SGLANG_056],
+  },
+  {
+    slug: 'roofline',
+    term: 'Roofline',
+    aliases: ['frontier envelope', 'roofline curve'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'On InferenceX a roofline is the outer envelope drawn through the best points of one hardware configuration, showing the edge of what it achieved.',
+    definition:
+      'A roofline on InferenceX is the Pareto envelope curve drawn per hardware configuration through its non-dominated points for the selected pair of axes.',
+    explanation:
+      'Which corner counts as best depends on the metrics: throughput against interactivity takes the upper right, while a cost axis is better when lower, so the envelope is anchored at whichever corner the selected metric pair defines. Points with a degenerate x value are excluded from eligibility, though they still render in the show-all view.',
+    significance:
+      'Reading a cloud of raw points invites cherry-picking, since a badly tuned configuration contributes points that no operator would ever choose. The envelope shows the achievable boundary per system, which is the shape a capacity decision is actually made against.',
+    benchmarkContext:
+      'This is not the classic roofline model from HPC, which plots attainable FLOPS against arithmetic intensity to expose a compute or memory bound. The dashboard borrows only the picture of an upper bound. Roofline direction is configured per metric, so the same points can produce a different envelope on a different axis.',
+    relatedTerms: ['pareto-frontier', 'iso-interactivity', 'throughput', 'recipe'],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2, B200_GLM5, AGENTX_GLM_ATOM],
+  },
+  {
+    slug: 'arithmetic-intensity',
+    term: 'Arithmetic intensity',
+    aliases: ['operational intensity', 'compute to memory ratio'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'Arithmetic intensity is how much math a computation does per byte it moves, which decides whether the chip or its memory is the limit.',
+    definition:
+      'Arithmetic intensity is the ratio of arithmetic operations performed to bytes moved between memory and the compute units.',
+    explanation:
+      'Prefill processes many tokens against one set of weights, so it reuses each loaded byte heavily and is usually compute bound. Decode emits one token per step per sequence and must still read the weights and the KV cache, so it moves a great deal of data for very little arithmetic and is usually memory bandwidth bound.',
+    significance:
+      'The two phases are limited by different parts of the chip, which is why a specification sheet with impressive peak FLOPS can disappoint on decode and why batching helps: grouping sequences raises intensity by reusing each weight read across more tokens.',
+    benchmarkContext:
+      'The split explains recurring shapes in the data. Low interactivity points run large batches at high intensity and approach compute limits, while the high interactivity end runs small batches and tracks memory bandwidth, so bandwidth-rich parts often win there despite lower peak throughput.',
+    relatedTerms: ['prefill', 'decode', 'memory-bandwidth', 'batching'],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2, TILERT, JALAPENO],
+  },
+  {
+    slug: 'prefix-cache-hit-rate',
+    term: 'Prefix cache hit rate',
+    aliases: ['cache hit rate', 'KV reuse rate'],
+    category: 'Serving',
+    plainEnglish:
+      'The hit rate is the share of prompt tokens served from cache instead of being recomputed, which on long sessions is most of the prompt.',
+    definition:
+      'Prefix cache hit rate is the fraction of input tokens satisfied from cached KV state rather than recomputed during prefill.',
+    explanation:
+      'A hit rate is only meaningful next to the tier that produced it, since a token served from accelerator memory and one fetched back from host memory cost very differently. It also depends on more than capacity: eviction can discard a prefix that is still wanted, and routing can send a request to a worker that never held it.',
+    significance:
+      'On multi-turn traffic the hit rate largely determines prefill cost, because each turn resends the whole conversation plus a little more. Once reuse is high, the remaining prefill work is dominated by genuinely new tokens, and the bottleneck moves from computation to cache management.',
+    benchmarkContext:
+      'The AgentX point view reports hit rate over time, separated by cache tier, alongside the prompt token source breakdown. A run that reports high aggregate throughput on a low hit rate is doing far more prefill work than a well-cached deployment would.',
+    relatedTerms: ['prefix-caching', 'kv-cache', 'prefill', 'agentx'],
+    articleSlugs: [AGENTX_V3, AGENTIC_WORKLOADS, AGENT_BENCHMARK, AGENTX_DSV4_B200_B300],
+  },
+  {
+    slug: 'warmup',
+    term: 'Warmup',
+    aliases: ['warmup phase', 'cache priming'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'Warmup is the priming pass before measurement starts, so the run is scored on a system in steady state rather than one with an empty cache.',
+    definition:
+      'Warmup is the phase before the profiling window in which a benchmark primes caches and reaches steady state, and whose requests are excluded from reported results.',
+    explanation:
+      'A cold system misrepresents production twice over: the cache is empty, and every session starting at turn zero creates a synchronized burst that no real deployment sees. AgentX therefore starts each conversation at a seeded point partway through its history, replays the requests needed to reconstruct that state, then advances each replay lane further before measurement begins.',
+    significance:
+      'Where the measurement window starts changes the result. Include the priming pass and prefix reuse looks worse than production; skip priming entirely and cache-dependent recipes are scored on a state they would never serve from.',
+    benchmarkContext:
+      'The AgentX point view separates the two phases, so telemetry can be inspected for either. Warmup requests are capped at a single output token, which is why their output length is about one and their interactivity and decode series are blank: one token has no inter-token latency.',
+    relatedTerms: ['prefix-caching', 'agentx', 'trace-replay', 'closed-loop-benchmark'],
+    articleSlugs: [AGENTX_V3, AGENT_BENCHMARK, AGENTIC_WORKLOADS],
+  },
+  {
+    slug: 'aiperf',
+    term: 'AIPerf',
+    aliases: ['replay harness', 'load generator'],
+    category: 'Agentic inference',
+    plainEnglish:
+      'AIPerf is the vendor-neutral client that sends the benchmark traffic, reconstructing recorded agent sessions and timing every request.',
+    definition:
+      'AIPerf is the open-source HTTP load generation and replay harness that drives AgentX runs against a serving endpoint.',
+    explanation:
+      'It reconstructs each session as a directed acyclic graph in which nodes are requests and edges carry the delay before a dependent request may be sent. That structure reproduces main-agent turns, parallel subagent branches that join later, one-off auxiliary requests, and the tool-use pauses between turns, none of which a flat list of prompts can express.',
+    significance:
+      'The client is part of the measurement. A harness that cannot express dependencies would issue an agentic workload as independent requests and erase exactly the burstiness and reuse the scenario exists to test. Keeping it vendor neutral also keeps the load generator from favoring any one serving stack.',
+    benchmarkContext:
+      'A seed fixes which conversations are sampled, where each starts, and the synthetic content used to fill anonymized blocks, so a rerun of the same recipe replays the same workload. All submissions for a model run the same harness minor version so results stay comparable.',
+    relatedTerms: ['trace-replay', 'agentx', 'closed-loop-benchmark', 'subagent'],
+    articleSlugs: [AGENTX_V3, AGENT_BENCHMARK, AGENTIC_WORKLOADS],
+  },
+  {
+    slug: 'pipeline-parallelism',
+    term: 'Pipeline parallelism',
+    abbreviation: 'PP',
+    aliases: ['layer parallelism', 'PP'],
+    category: 'Parallelism',
+    plainEnglish:
+      'Pipeline parallelism gives each chip a different slice of the layers, passing activations along the chain instead of splitting each layer.',
+    definition:
+      'Pipeline parallelism partitions a model by layer across accelerators, so each stage runs its own layers and forwards activations to the next.',
+    explanation:
+      'Communication is a point to point handoff of activations at stage boundaries, which is far cheaper than the per-layer collectives tensor parallelism requires. The cost is idle time: with one request in flight, every stage but the active one waits, and only a stream of concurrent work keeps the pipeline full.',
+    significance:
+      'For the largest models this is a capacity technique before it is a speed technique. Some frontier models do not fit in one node at all, and pipeline parallelism is what makes them servable, sometimes as the only option when a competing optimization refuses to compose with anything else.',
+    benchmarkContext:
+      'InferenceX reports the pipeline degree in point tooltips and parallelism labels alongside TP, EP, and DP, and only when it exceeds one. Composability matters as much as the degree: a stage split that blocks speculative decoding can cost more than the memory it saved.',
+    relatedTerms: [
+      'tensor-parallelism',
+      'expert-parallelism',
+      'data-parallelism',
+      'high-bandwidth-memory',
+    ],
+    articleSlugs: [AGENTX_V3, KIMI_K3, INFERENCEX_V2],
+  },
+  {
+    slug: 'dp-attention',
+    term: 'Data-parallel attention',
+    abbreviation: 'DPA',
+    aliases: ['DP attention', 'attention data parallelism'],
+    category: 'Parallelism',
+    plainEnglish:
+      'DP attention gives each rank its own slice of the attention work and its own cache, instead of every rank holding a copy of the same thing.',
+    definition:
+      'Data-parallel attention replicates attention computation per rank over disjoint sets of sequences, so each rank owns a private share of the KV cache while experts remain shared.',
+    explanation:
+      'Tensor-parallel attention splits heads and ends up replicating KV state across ranks, which wastes capacity when heads are few. DP attention avoids that duplication by assigning whole sequences to ranks. The ranks still participate together in the MoE collectives, so attention is local while expert dispatch stays global.',
+    significance:
+      'Because each rank owns a private slice of the pool, placement becomes correctness-adjacent for performance: a long session routed to a rank that does not hold its prefix recomputes everything. Measured hit rates can then land far below the theoretical ceiling for reasons that have nothing to do with cache size.',
+    benchmarkContext:
+      'InferenceX shows DP attention in the parallelism section of point tooltips. Whether it helps is model dependent, and configurations without it sometimes dominate the frontier when cache locality turns into a routing constraint.',
+    relatedTerms: [
+      'data-parallelism',
+      'tensor-parallelism',
+      'kv-cache',
+      'mixture-of-experts',
+      'prefix-caching',
+    ],
+    articleSlugs: [AGENTX_V3, INFERENCEX_V2, GB200_KIMI, AGENTX_M3_TRT],
+  },
+  {
+    slug: 'int4',
+    term: 'INT4',
+    aliases: ['4-bit integer', 'W4A16'],
+    category: 'Numerical precision',
+    plainEnglish:
+      'INT4 stores weights in four-bit integers, shrinking the model enough to move far less memory per token on hardware without native 4-bit floats.',
+    definition:
+      'INT4 is a four-bit integer format used mainly for weight quantization, typically with a higher-precision scale per group of values.',
+    explanation:
+      'Integer formats spread their values evenly, unlike floating point, so INT4 depends on grouped scaling factors to track the local range of each block of weights. Activations commonly stay at higher precision, and the matrix multiply dequantizes on the fly, which makes the technique a memory movement optimization more than an arithmetic one.',
+    significance:
+      'It matters most where 4-bit floating point has no hardware support. On such parts INT4 is the practical route to 4-bit weights, though it usually needs more calibration care than a native format and its accuracy has to be checked rather than assumed.',
+    benchmarkContext:
+      'InferenceX treats INT4 as its own precision key alongside FP4, FP8, and BF16, and precision is part of the recipe rather than a display option. Compare INT4 against a native FP4 recipe only with the accuracy evaluations in view, since the formats are not interchangeable.',
+    relatedTerms: ['quantization', 'fp4', 'fp8', 'memory-bandwidth'],
+    articleSlugs: [B200_KIMI, INFERENCEX_V2, MI355X_KIMI],
+  },
+  {
+    slug: 'bf16',
+    term: 'BF16',
+    aliases: ['bfloat16', 'brain float 16'],
+    category: 'Numerical precision',
+    plainEnglish:
+      'BF16 is the 16-bit format most models are trained in, and serving in it is the accuracy reference the quantized recipes are measured against.',
+    definition:
+      'BF16 is a 16-bit floating point format with the same exponent range as FP32 and a reduced mantissa, widely used for training and as an unquantized serving baseline.',
+    explanation:
+      'Keeping the FP32 exponent range makes BF16 tolerant of the value distributions that appear in transformer activations, so conversion rarely needs the scaling machinery narrower formats require. The tradeoff is precision rather than range, and the format is twice the size of FP8 and four times that of a 4-bit format.',
+    significance:
+      'Its role in a benchmark is usually as a reference point. Weight reads dominate decode, so a BF16 recipe moves far more memory per token than a quantized one and tends to sit lower on the throughput curve while defining the accuracy the others are compared against.',
+    benchmarkContext:
+      'InferenceX carries BF16 as a precision key and reports peak BF16 dense throughput per accelerator in the specs pages. Quantized recipes are validated with accuracy evaluations rather than assumed lossless, which is what makes a BF16 comparison meaningful.',
+    relatedTerms: ['quantization', 'fp8', 'fp4', 'memory-bandwidth'],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2, DEEPSEEK_V4],
+  },
+  {
+    slug: 'kv-cache-quantization',
+    term: 'KV cache quantization',
+    aliases: ['FP8 KV cache', 'quantized KV'],
+    category: 'Numerical precision',
+    plainEnglish:
+      'This stores the conversation cache in a smaller format, so a chip holds more context and reads it back faster during generation.',
+    definition:
+      'KV cache quantization stores attention key and value states in a reduced-precision format, independently of the precision used for model weights.',
+    explanation:
+      'Weight precision and cache precision are separate choices, and a recipe can serve FP8 weights with a BF16 cache or the reverse. Halving cache width roughly doubles the tokens that fit in accelerator memory and halves the bytes read per decode step, which is the operation decode spends most of its time on.',
+    significance:
+      'On long-context serving this often buys more than shrinking the weights, because at high concurrency the cache, not the weights, is what exhausts memory. Accuracy sensitivity differs by model and by which of keys and values is quantized, so it needs evaluation rather than a blanket assumption.',
+    benchmarkContext:
+      'Cache precision is part of the recipe, and mixed layouts exist in practice: some models keep two cache buffers at different widths, which disaggregated transfer paths then have to move as a pair. Read a memory-capacity claim together with the cache format behind it.',
+    relatedTerms: ['kv-cache', 'quantization', 'fp8', 'high-bandwidth-memory'],
+    articleSlugs: [INFERENCEX_V2, AGENTX_V3, GB300_DSV4],
+  },
+  {
+    slug: 'sliding-window-attention',
+    term: 'Sliding window attention',
+    abbreviation: 'SWA',
+    aliases: ['local attention', 'windowed attention'],
+    category: 'Model architecture',
+    plainEnglish:
+      'Sliding window attention lets a layer look only at a recent span of tokens, so its cache stops growing once the window is full.',
+    definition:
+      'Sliding window attention restricts each query to a fixed span of preceding tokens, bounding the KV state a layer must retain.',
+    explanation:
+      'Because the span is fixed, the cache for such a layer reaches a ceiling instead of growing with the conversation, and older entries fall out as the window advances. Models usually interleave these layers with full-attention layers, so long-range information still has a path through the network while most layers stay cheap.',
+    significance:
+      'The bounded cost comes with an allocator problem. Window pages turn over constantly while durable prefix pages sit still, and when both are drawn from one pool the transient allocation tends to evict the valuable one, so a long session can lose its expensive full-attention history to short-lived window state.',
+    benchmarkContext:
+      'These effects appear only on multi-turn traffic. A single short prompt never laps the window or contends for the pool, which is why window-aware eviction, offload, and branch handling show up as AgentX-driven engine work rather than as fixed-sequence results.',
+    relatedTerms: ['sparse-attention', 'kv-cache', 'prefix-caching', 'hybrid-attention'],
+    articleSlugs: [AGENTX_V3, KIMI_K3, B200_GLM5],
+  },
+  {
+    slug: 'hybrid-attention',
+    term: 'Hybrid attention',
+    aliases: ['mixed attention', 'hybrid cache model'],
+    category: 'Model architecture',
+    plainEnglish:
+      'A hybrid model mixes attention types across its layers, so its cache is several different kinds of state rather than one uniform block.',
+    definition:
+      'A hybrid attention model interleaves layers of different attention types, producing multiple KV cache groups with distinct shapes and lifetimes.',
+    explanation:
+      'A uniform model has one cache layout per token, so a single block geometry describes everything worth saving. A hybrid model does not: full-attention layers, windowed layers, and recurrent or compressor state coexist, each with its own footprint and its own rules about when it can be discarded and whether it can be rebuilt.',
+    significance:
+      'The distinction is invisible until state has to leave the accelerator. A connector that assumes one uniform layout cannot say which group a block belongs to, so the models with the longest sessions were for a time the ones that could not use offload at all. Recurrent state is the hardest case, since it accumulates everything before it and cannot be recomputed from neighboring tokens.',
+    benchmarkContext:
+      'Frontier open-weight models in the InferenceX matrix are increasingly hybrid, so engine support for hybrid cache groups is part of what a recipe is measuring. Offload, disaggregated transfer, and prefix reuse each had to be extended per group rather than inherited from the uniform case.',
+    relatedTerms: [
+      'sliding-window-attention',
+      'kv-cache',
+      'linear-attention',
+      'sparse-attention',
+      'prefix-caching',
+    ],
+    articleSlugs: [AGENTX_V3, KIMI_K3, DEEPSEEK_V4],
+  },
+  {
+    slug: 'linear-attention',
+    term: 'Linear attention',
+    aliases: ['GatedDeltaNet', 'recurrent attention', 'constant-state attention'],
+    category: 'Model architecture',
+    plainEnglish:
+      'Linear attention keeps a fixed-size running summary instead of every past token, so its memory does not grow as the conversation does.',
+    definition:
+      'Linear attention replaces the growing key and value cache with a recurrent state of constant size that is updated as tokens arrive.',
+    explanation:
+      'Standard attention stores state proportional to sequence length and rereads it every step. A linear or gated recurrent layer carries a fixed-size state instead, trading exact recall of every position for bounded memory. Architectures such as GatedDeltaNet apply this on a fraction of layers, leaving full attention elsewhere to preserve precise long-range lookup.',
+    significance:
+      'For long context the storage saving is substantial, but the state changes what caching means. It cannot be reconstructed from surrounding tokens the way a window tail can, so if it is dropped the only way back is replaying the sequence, and reuse requires an explicit checkpoint mechanism rather than ordinary block reuse.',
+    benchmarkContext:
+      'InferenceX serves models using these layers, and engine support for checkpointing and transferring recurrent state is part of the recipe. A model can be day-zero servable and still lack reuse of that state, which shows up as unexpectedly high prefill cost on repeated turns.',
+    relatedTerms: ['hybrid-attention', 'kv-cache', 'sparse-attention', 'prefill'],
+    articleSlugs: [AGENTX_V3, MI355X_QWEN, KIMI_K3, AGENTX_QWEN_SGLANG],
+  },
+  {
+    slug: 'nvl72',
+    term: 'NVL72',
+    aliases: ['GB200 NVL72', 'GB300 NVL72', 'rack-scale system'],
+    category: 'Hardware',
+    plainEnglish:
+      'NVL72 is a rack where 72 accelerators share one high-speed fabric, so they behave more like a single large machine than a cluster.',
+    definition:
+      'NVL72 is a rack-scale NVIDIA system that places 72 accelerators in a single NVLink scale-up domain rather than in separate eight-chip nodes.',
+    explanation:
+      'The dashboard specs record NVLink 5.0 at 900 GB/s per chip unidirectional across a scale-up world size of 72, switched through NVSwitch. A conventional node keeps that bandwidth among eight chips and falls back to slower scale-out networking beyond them, so the difference is not raw speed but how many chips are reachable before the fabric changes character.',
+    significance:
+      'Techniques whose cost is dominated by collectives change economics inside a large domain. Wide expert parallelism spreads experts across many chips and pays all-to-all traffic for every token, which is tolerable at scale-up bandwidth and often is not across a scale-out fabric.',
+    benchmarkContext:
+      'A rack-scale advantage is not automatic. Higher cost per chip has to be earned back, and on agentic traffic the orchestration layer can become the bottleneck before the fabric does, so NVL72 configurations sometimes trail eight-chip nodes on TCO-normalized throughput for models that do not exercise wide parallelism.',
+    relatedTerms: [
+      'nvlink',
+      'scale-up-vs-scale-out',
+      'wide-expert-parallelism',
+      'all-to-all',
+      'total-cost-of-ownership',
+    ],
+    articleSlugs: [GB200_R1, GB300_DSV4, GB200_KIMI, VR_RUBIN, AGENTX_K3_ATOM, JALAPENO],
+  },
+  {
+    slug: 'atom',
+    term: 'ATOM',
+    aliases: ['AMD ATOM', 'ATOMesh'],
+    category: 'Software',
+    plainEnglish:
+      'ATOM is AMD’s own inference engine, its answer to a vendor runtime rather than an upstream open-source one.',
+    definition:
+      'ATOM is AMD’s inference engine for Instinct accelerators, positioned as the vendor runtime alongside upstream vLLM and SGLang on ROCm.',
+    explanation:
+      'It occupies the same role for AMD that a vendor runtime does for NVIDIA: tuned for the vendor’s own hardware and free to move ahead of upstream engines. Its router, ATOMesh, began as a fork of the SGLang router. The engine was built for single-turn serving, so long-context multi-turn support required substantial changes to its cache manager and kernels.',
+    significance:
+      'A vendor engine can show what silicon is capable of before the open stack catches up, which makes it useful evidence and awkward guidance at the same time. Most labs deploy upstream engines, so a result that exists only under a vendor runtime does not describe what those users will get.',
+    benchmarkContext:
+      'InferenceX reports ATOM as its own framework label so it is never conflated with a vLLM or SGLang result on the same accelerator. Compare it to other vendor runtimes when asking what the hardware can do, and to upstream engines when asking what a customer can deploy today.',
+    relatedTerms: ['inference-engine', 'rocm', 'vllm', 'sglang', 'tensorrt-llm'],
+    articleSlugs: [AGENTX_V3, MI355X_DSV4, MI355X_GLM5, AGENTX_DSV4_MI355X_B200, AGENTX_K3_ATOM],
+  },
+  {
+    slug: 'aiter',
+    term: 'AITER',
+    aliases: ['AMD AITER', 'ROCm kernel library'],
+    category: 'Software',
+    plainEnglish:
+      'AITER is AMD’s tuned kernel library, the layer that decides how fast attention and matrix work actually run on Instinct chips.',
+    definition:
+      'AITER is AMD’s library of optimized kernels for Instinct accelerators, used by inference engines running on ROCm.',
+    explanation:
+      'Engines express a strategy; kernels decide whether it is fast. AITER supplies tuned attention, matrix, and fused operations, and an engine dispatches to it in place of a generic path. That dispatch decision is itself tunable, and a kernel that wins on one shape can lose on another, so selection may depend on context length rather than being fixed.',
+    significance:
+      'A parallelism strategy is only real if the kernels can express it, which is why context parallelism and long-context sparse attention on AMD arrived as kernel work rather than engine work. Very large caches also expose failures short requests never reach, such as address arithmetic that overflows once a pool crosses a size boundary and silently addresses the wrong row.',
+    benchmarkContext:
+      'The library sits inside the container image a recipe pins, so an AITER improvement can move a curve with no change to the engine version or the hardware. Kernel-level gains measured on uniform shapes do not always survive agentic traces, where cache and scheduling variance can swamp them.',
+    relatedTerms: ['rocm', 'inference-engine', 'sparse-attention', 'memory-bandwidth', 'vllm'],
+    articleSlugs: [MI355X_KIMI, AGENTX_V3, MI355X_DSV4],
+  },
+  {
+    slug: 'flashinfer',
+    term: 'FlashInfer',
+    aliases: ['attention kernel library'],
+    category: 'Software',
+    plainEnglish:
+      'FlashInfer is a library of attention kernels that serving engines call instead of writing their own attention implementations.',
+    definition:
+      'FlashInfer is an open-source library of attention kernels and backends used by inference engines for prefill, decode, and speculative verification.',
+    explanation:
+      'Attention is where most serving-specific complexity lives: paged caches, variable sequence lengths, grouped query heads, sparsity patterns, and verification of drafted tokens all reshape the kernel. A shared library lets several engines reuse one tuned implementation, and engines select a backend per shape and per hardware target.',
+    significance:
+      'Because backends are selected rather than fixed, kernel availability becomes a portability question. A feature implemented only for one vendor’s backend leaves the alternative running a generic path, which on long context is not a small compromise but the wrong kernel for the shape.',
+    benchmarkContext:
+      'The backend in use is part of the recipe, and a change to it can move a curve without any hardware or engine version change. Support for checkpointing recurrent state in these kernels is what allowed hybrid models to participate in prefix reuse at all.',
+    relatedTerms: ['inference-engine', 'sparse-attention', 'kv-cache', 'vllm', 'sglang'],
+    articleSlugs: [AGENTX_V3, SGLANG_056, INFERENCEX_V2],
+  },
+  {
+    slug: 'cuda-graphs',
+    term: 'CUDA graphs',
+    aliases: ['graph capture', 'full-graph mode'],
+    category: 'Software',
+    plainEnglish:
+      'CUDA graphs record a whole sequence of chip operations once and replay it as a unit, removing the per-step cost of launching each one.',
+    definition:
+      'CUDA graph capture records a sequence of kernel launches and their dependencies into a replayable graph, so the sequence is submitted once rather than launched operation by operation.',
+    explanation:
+      'A decode step issues many small kernels, and at small batch sizes the launch and scheduling overhead between them can rival the arithmetic. Capturing the step removes that per-launch cost. The catch is that a graph is fixed: shapes must be stable, so engines capture per bucket and leave genuinely dynamic work outside the graph.',
+    significance:
+      'This is a latency optimization more than a throughput one, and it matters most exactly where batches are small and interactivity is high. It also interacts with everything that changes shape, which is why variable-length agentic traffic can defeat a runtime that specializes too eagerly and recompiles for nearly every request it sees.',
+    benchmarkContext:
+      'Graph usage is part of the engine image a recipe pins, so it can move a curve with no change in hardware. Recipes may capture stable producers while leaving request-dependent attention eager, which is a deliberate compromise between capture coverage and shape flexibility.',
+    relatedTerms: ['cuda', 'decode', 'interactivity', 'inference-engine', 'batching'],
+    articleSlugs: [AGENTX_V3, TILERT, INFERENCEX_V2],
+  },
+  {
+    slug: 'goodput',
+    term: 'Goodput',
+    aliases: ['SLO-constrained throughput', 'useful throughput'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'Goodput counts only the work that meets your latency target, so a fast-looking system that misses deadlines gets no credit for it.',
+    definition:
+      'Goodput is the portion of throughput that satisfies a stated service level objective, such as a time to first token bound or a minimum tokens per second per user.',
+    explanation:
+      'Raw throughput rewards a server for finishing requests no matter how slowly each user was served. Goodput applies a filter first: a request only counts if it met the latency or interactivity constraint the operator promised. Two systems with identical throughput can have very different goodput once a deadline is applied, because one may hold latency flat under load while the other lets queueing push every request past the bound.',
+    significance:
+      'Capacity planning that ignores goodput overbuys or oversells. An operator who quotes peak throughput but serves half of it within the SLO needs twice the fleet they modeled. Goodput is the number that connects a benchmark curve to how many real users a deployment can actually carry.',
+    benchmarkContext:
+      'InferenceX publishes full throughput versus interactivity Pareto frontiers rather than a single goodput number, which lets readers apply their own SLO. Reading the frontier at a fixed interactivity tier, as the TCO calculator does, is exactly a goodput measurement at that tier.',
+    relatedTerms: [
+      'service-level-objective',
+      'throughput',
+      'interactivity',
+      'iso-interactivity',
+      'tail-latency',
+    ],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2],
+  },
+  {
+    slug: 'model-flops-utilization',
+    term: 'Model FLOPs utilization',
+    abbreviation: 'MFU',
+    aliases: ['MFU', 'model bandwidth utilization', 'MBU'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'MFU compares the useful math a model actually performed against the maximum the chip could theoretically perform, giving an efficiency percentage.',
+    definition:
+      'Model FLOPs utilization is the ratio of the floating point operations a model logically requires to the peak operations the hardware could deliver in the same wall clock time.',
+    explanation:
+      'Peak TFLOP/s figures assume every tensor core is busy every cycle, which never happens in serving. Kernel launch gaps, memory stalls, communication waits, and imperfect batching all leave compute idle. MFU folds all of that into one number. Decode is usually memory bound, so decode MFU is naturally low, and the bandwidth analog MBU is often the more honest efficiency measure for token generation.',
+    significance:
+      'MFU separates hardware capability from software maturity. A chip with enormous peak FLOPs but weak kernels can lose to a slower chip that keeps its units fed. Rising MFU on fixed hardware is the signature of software progress, which is where most inference performance gains come from.',
+    benchmarkContext:
+      'InferenceX tracks delivered tokens per chip over time rather than reporting MFU directly, and the repeated pattern of large gains on unchanged hardware, such as order of magnitude improvements within weeks of a model release, is utilization being recovered by better software.',
+    relatedTerms: ['arithmetic-intensity', 'roofline', 'memory-bandwidth', 'throughput'],
+    articleSlugs: [MI355X_DSV4, DEEPSEEK_V4, INFERENCEX_V2],
+  },
+  {
+    slug: 'memory-bound-vs-compute-bound',
+    term: 'Memory bound vs compute bound',
+    aliases: ['bandwidth bound', 'compute limited'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'A workload is compute bound when the math units are the bottleneck and memory bound when waiting on data movement is the bottleneck.',
+    definition:
+      'A kernel is compute bound when its runtime is set by arithmetic throughput and memory bound when its runtime is set by how fast operands move between memory and the compute units.',
+    explanation:
+      'Every kernel has an arithmetic intensity, the ratio of operations to bytes touched. If that ratio is below the hardware balance point, the memory system saturates before the math units do. LLM prefill runs large matrix multiplies with high intensity and tends to be compute bound, while decode reads the entire working set of weights and KV cache to produce one token per request and is usually memory bound.',
+    significance:
+      'The binding resource decides which hardware spec matters. Memory bound decode explains why HBM capacity and bandwidth headline every accelerator launch, why quantization speeds up decode by shrinking bytes moved, and why a chip with modest FLOPs but fast memory can win interactive serving.',
+    benchmarkContext:
+      'InferenceX sweeps concurrency, which walks a system between regimes: low concurrency decode is bandwidth limited while high concurrency batching pushes toward compute limits. The shape of each throughput versus interactivity curve reflects where that transition happens for a given recipe.',
+    relatedTerms: [
+      'arithmetic-intensity',
+      'roofline',
+      'memory-bandwidth',
+      'high-bandwidth-memory',
+      'decode',
+    ],
+    articleSlugs: [INFERENCEX_V2, TILERT],
+  },
+  {
+    slug: 'gpu-utilization',
+    term: 'GPU utilization',
+    aliases: ['chip utilization', 'accelerator utilization'],
+    category: 'Benchmark metrics',
+    plainEnglish:
+      'GPU utilization measures how busy an accelerator is, though the common percentage from monitoring tools can look high while real efficiency stays low.',
+    definition:
+      'GPU utilization is the share of time or capability an accelerator spends doing useful work, reported anywhere from coarse busy percentages to strict measures like model FLOPs utilization.',
+    explanation:
+      'The utilization number in basic monitoring tools only says a kernel was resident, not that it used the chip well. A kernel occupying one compute unit still reads as busy. Stricter measures compare delivered arithmetic or bandwidth against hardware peaks. In serving, utilization is also shaped by traffic: idle gaps between requests, low concurrency, and tool call pauses in agent sessions all leave paid-for silicon idle.',
+    significance:
+      'Fleet economics hinge on utilization. The gap between a well batched, well scheduled deployment and a naive one is often several times the cost per token on identical hardware, which is why serving software and request routing get as much attention as the chips themselves.',
+    benchmarkContext:
+      'InferenceX benchmarks report delivered throughput per chip at each interactivity level rather than a utilization percentage, so utilization differences between engines, precisions, and parallelism plans show up directly as separation between curves on the same hardware.',
+    relatedTerms: [
+      'model-flops-utilization',
+      'batching',
+      'concurrency',
+      'throughput',
+      'total-cost-of-ownership',
+    ],
+    articleSlugs: [INFERENCEMAX, AGENTIC_WORKLOADS],
+  },
+  {
+    slug: 'continuous-batching',
+    term: 'Continuous batching',
+    aliases: ['in-flight batching', 'dynamic batching', 'iteration-level scheduling'],
+    category: 'Serving',
+    plainEnglish:
+      'Continuous batching lets new requests join a running batch the moment old ones finish, instead of waiting for the whole batch to complete.',
+    definition:
+      'Continuous batching is a scheduling technique that admits and retires requests at every generation step, keeping the batch full as individual sequences finish at different times.',
+    explanation:
+      'Static batching waits to collect a group of requests, runs them together, and returns them together, so a batch runs as long as its slowest member. Because LLM outputs vary wildly in length, that wastes enormous capacity. Continuous batching reforms the batch every iteration: a sequence that emits its final token leaves immediately and a queued request takes its slot at the next step, so the accelerator stays saturated.',
+    significance:
+      'This is one of the foundational optimizations of modern LLM serving and a large part of why open source engines displaced naive deployment. It multiplies throughput at a given latency and pairs naturally with paged KV cache memory, which makes slot reuse cheap.',
+    benchmarkContext:
+      'Every engine InferenceX benchmarks, including vLLM, SGLang, and TensorRT-LLM, uses continuous batching. Concurrency sweeps measure how well each scheduler holds interactivity as the batch fills, which is where implementation differences between engines become visible.',
+    relatedTerms: ['batching', 'concurrency', 'paged-attention', 'inference-engine', 'throughput'],
+    articleSlugs: [INFERENCEMAX, SGLANG_056],
+  },
+  {
+    slug: 'paged-attention',
+    term: 'PagedAttention',
+    aliases: ['paged KV cache', 'KV cache paging'],
+    category: 'Serving',
+    plainEnglish:
+      'PagedAttention stores the KV cache in small fixed size blocks, like virtual memory pages, so cache memory is not wasted on unused space.',
+    definition:
+      'PagedAttention is a KV cache management technique that allocates cache in fixed size blocks addressed through a mapping table, rather than reserving one contiguous region per request.',
+    explanation:
+      'Contiguous per-request allocation must reserve space for the longest possible output, and most of that reservation is never used. Paging borrows the operating system playbook: cache blocks are allocated on demand as a sequence grows, freed the moment it ends, and shared between sequences with a common prefix through copy on write. Fragmentation drops to near zero, so far more sequences fit in the same HBM.',
+    significance:
+      'Introduced by vLLM, this idea unlocked the batch sizes that make continuous batching pay off and became standard across serving engines. Effective KV capacity, not raw memory size, is what bounds concurrency for long context and agentic workloads.',
+    benchmarkContext:
+      'All engines in InferenceX recipes manage KV memory in paged or block based form. High concurrency points on long context scenarios such as AgentX are only reachable because paging keeps cache waste small as hundreds of sessions grow and shrink.',
+    relatedTerms: ['kv-cache', 'kv-cache-manager', 'continuous-batching', 'prefix-caching', 'vllm'],
+    articleSlugs: [AGENT_BENCHMARK, INFERENCEMAX],
+  },
+  {
+    slug: 'radix-attention',
+    term: 'RadixAttention',
+    aliases: ['radix tree cache', 'radix prefix cache'],
+    category: 'Serving',
+    plainEnglish:
+      'RadixAttention keeps completed KV cache in a radix tree keyed by token content, so any new request can reuse the longest matching prefix.',
+    definition:
+      'RadixAttention is a prefix caching design from SGLang that retains KV cache entries in a radix tree after requests finish, enabling automatic reuse across requests that share token prefixes.',
+    explanation:
+      'A radix tree indexes cached segments by their token content, so lookup finds the longest previously computed prefix of an incoming request in one walk. Reuse is automatic and cross request: multi turn chats, system prompts shared by many users, and agent branches that fork from a common history all hit the same cached nodes. An eviction policy such as least recently used bounds the memory the tree holds.',
+    significance:
+      'Prefix reuse converts redundant prefill compute into cache hits, and in agent traffic where each turn resends a growing history the savings can dominate end to end cost. Making reuse structural rather than opt in is a large part of why SGLang performs well on such workloads.',
+    benchmarkContext:
+      'InferenceX AgentX traces preserve real shared prefix structure between turns and subagent branches, so engines with strong radix style reuse show materially better time to first token and throughput on the agentic scenario than the fixed sequence scenarios would predict.',
+    relatedTerms: [
+      'prefix-caching',
+      'prefix-cache-hit-rate',
+      'kv-cache',
+      'sglang',
+      'agentic-inference',
+    ],
+    articleSlugs: [AGENT_BENCHMARK, AGENTX_QWEN_SGLANG],
+  },
+  {
+    slug: 'draft-model',
+    term: 'Draft model',
+    aliases: ['draft head', 'speculator'],
+    category: 'Serving',
+    plainEnglish:
+      'A draft model is the small fast model in speculative decoding that guesses several upcoming tokens for the large model to verify in one pass.',
+    definition:
+      'A draft model is the lightweight proposal component of speculative decoding, producing candidate token sequences that the target model verifies in a single batched forward pass.',
+    explanation:
+      'Drafts take several forms: a separate small model from the same family, extra prediction heads trained onto the target model as in EAGLE style methods, or the multi token prediction heads some models ship with. The draft races ahead a few tokens cheaply, the target checks all of them at once, and accepted tokens are emitted together. Rejection falls back to the target model output, so results match the target distribution.',
+    significance:
+      'Draft quality sets acceptance length, and acceptance length sets the speedup. A well matched draft can multiply decode speed at low batch sizes, while a mismatched or overly aggressive draft wastes verify compute and can even slow serving under load.',
+    benchmarkContext:
+      'InferenceX records the speculative method and acceptance length behind each result and publishes golden acceptance length distributions for reproduction, because an unrealistic acceptance rate is a classic way a benchmark number stops describing production behavior.',
+    relatedTerms: [
+      'speculative-decoding',
+      'eagle',
+      'multi-token-prediction',
+      'acceptance-length',
+      'decode',
+    ],
+    articleSlugs: [SGLANG_056, GB300_DSV4],
+  },
+  {
+    slug: 'offline-inference',
+    term: 'Offline inference',
+    aliases: ['batch inference', 'offline batch serving'],
+    category: 'Serving',
+    plainEnglish:
+      'Offline inference processes a large pile of requests with no user waiting, so the only goal is maximum tokens per dollar, not latency.',
+    definition:
+      'Offline inference is model serving without an interactive deadline, where requests are processed in bulk and the objective is total throughput and cost rather than per-user latency.',
+    explanation:
+      'Synthetic data generation, document processing, embedding backfills, and evaluation sweeps do not care when any individual request returns. That frees the scheduler to run the largest batches the memory allows, order requests to maximize prefix reuse, and hold the accelerator at its throughput limit. Online serving lives at the opposite end of the same tradeoff, sacrificing throughput to keep every user above an interactivity floor.',
+    significance:
+      'The same hardware can differ by multiples in tokens per dollar between offline and tight latency operation, so quoting a single price per million tokens without stating the operating point is close to meaningless. Fleets often split into latency tiers for this reason.',
+    benchmarkContext:
+      'The right edge of an InferenceX throughput versus interactivity curve, where batch size is maximal and per-user speed is lowest, approximates offline operation. Reading one curve at both edges shows the full online to offline cost range for a recipe.',
+    relatedTerms: [
+      'throughput',
+      'interactivity',
+      'batching',
+      'cost-per-million-tokens',
+      'pareto-frontier',
+    ],
+    articleSlugs: [INFERENCEMAX, B200_MINIMAX],
+  },
+  {
+    slug: 'context-window',
+    term: 'Context window',
+    aliases: ['context length', 'max sequence length', 'long context'],
+    category: 'Model architecture',
+    plainEnglish:
+      'The context window is the maximum number of tokens a model can consider at once, covering both the input and everything generated so far.',
+    definition:
+      'The context window is the maximum sequence length a model supports, bounding the combined token count of prompt, conversation history, retrieved material, and generated output.',
+    explanation:
+      'Attention lets every token reference earlier tokens, and the KV cache holds state for all of them, so a longer window costs memory and compute that grow with length. Position encoding schemes and training length set the usable window, while serving stacks must budget KV capacity for it. Modern frontier models advertise windows of hundreds of thousands of tokens, but throughput and interactivity degrade as sequences approach those limits.',
+    significance:
+      'Long context is what makes coding agents, retrieval heavy pipelines, and document analysis workloads possible, and it is also what makes them expensive to serve. Architectural responses such as sliding window layers, latent attention, and linear attention exist mainly to bend the cost curve of the window.',
+    benchmarkContext:
+      'InferenceX covers the window from both directions: fixed sequence scenarios pin input and output lengths such as 8K in and 1K out, while AgentX replays sessions whose contexts grow turn by turn toward realistic agent working sets.',
+    relatedTerms: [
+      'kv-cache',
+      'input-output-sequence-length',
+      'sliding-window-attention',
+      'linear-attention',
+      'agentic-coding-workload',
+    ],
+    articleSlugs: [AGENTIC_WORKLOADS, KIMI_K3],
+  },
+  {
+    slug: 'grouped-query-attention',
+    term: 'Grouped-query attention',
+    abbreviation: 'GQA',
+    aliases: ['GQA', 'multi-query attention', 'MQA'],
+    category: 'Model architecture',
+    plainEnglish:
+      'Grouped-query attention lets several query heads share one set of key and value heads, shrinking the KV cache without giving up much quality.',
+    definition:
+      'Grouped-query attention is an attention variant where query heads are divided into groups that each share a single key and value head, reducing KV cache size and bandwidth per token.',
+    explanation:
+      'Standard multi-head attention stores keys and values for every head, so cache size scales with head count. Multi-query attention collapses all heads onto one KV pair, which is maximally cheap but can hurt quality. GQA sits between the two: a model might serve 64 query heads from 8 KV heads, cutting cache size eight fold. Because decode is dominated by reading the KV cache, the saving translates directly into faster token generation.',
+    significance:
+      'GQA became the default attention layout for dense open models because it attacks the memory side of decode where serving is actually bound. It also set the stage for more aggressive KV compression schemes such as multi-head latent attention.',
+    benchmarkContext:
+      'Attention layout is fixed by each model architecture, so GQA shows up in InferenceX through model level differences in KV bytes per token, which shape achievable concurrency and interactivity on identical hardware and engine versions.',
+    relatedTerms: [
+      'kv-cache',
+      'multi-head-latent-attention',
+      'decode',
+      'memory-bandwidth',
+      'kv-cache-quantization',
+    ],
+    articleSlugs: [INFERENCEX_V2, AGENT_BENCHMARK],
+  },
+  {
+    slug: 'active-parameters',
+    term: 'Active parameters',
+    aliases: ['activated parameters', 'active params'],
+    category: 'Model architecture',
+    plainEnglish:
+      'Active parameters are the weights a mixture of experts model actually uses for each token, a small slice of its much larger total size.',
+    definition:
+      'Active parameters are the subset of a sparse model, its shared layers plus the experts its router selects, that participate in computing any single token.',
+    explanation:
+      'A mixture of experts model might hold a trillion total parameters while routing each token through only a few tens of billions. Compute per token scales with the active count, which is why sparse frontier models can be affordable to run. Memory tells a different story: every expert must sit in HBM ready to be selected, so capacity requirements and parallelism plans follow total parameters even though arithmetic follows active ones.',
+    significance:
+      'The total versus active split explains most modern serving economics. It is why trillion parameter models are deployable at all, why expert parallelism across many chips exists, and why comparing models by total parameter count says little about their serving cost.',
+    benchmarkContext:
+      'The MoE models InferenceX benchmarks, including the DeepSeek, Kimi, Qwen, and MiniMax families, all have low active to total ratios, and their recipes spread experts across nodes precisely because total parameters set the memory bill.',
+    relatedTerms: [
+      'mixture-of-experts',
+      'expert-parallelism',
+      'wide-expert-parallelism',
+      'high-bandwidth-memory',
+    ],
+    articleSlugs: [DEEPSEEK_V4, KIMI_K3],
+  },
+  {
+    slug: 'dense-model',
+    term: 'Dense model',
+    aliases: ['dense transformer', 'dense LLM'],
+    category: 'Model architecture',
+    plainEnglish:
+      'A dense model applies every one of its parameters to every token it processes, unlike sparse models that route tokens to a few experts.',
+    definition:
+      'A dense model is a neural network in which all weights participate in every forward pass, so compute per token scales directly with total parameter count.',
+    explanation:
+      'Dense transformers are the simpler design: every layer processes every token with all of its weights. That makes their behavior predictable, their parallelism straightforward, and their quality per parameter strong, but serving cost grows linearly with size. Mixture of experts models break that link by activating a fraction of their weights per token, which is why the largest frontier models are sparse while small and mid size models often stay dense.',
+    significance:
+      'The dense versus sparse choice drives serving strategy. Dense models fit on fewer chips and avoid expert routing complexity, while sparse models buy more quality per unit of compute at the price of much larger memory footprints and heavier cross chip communication.',
+    benchmarkContext:
+      'InferenceX coverage centers on the large sparse models operators actually deploy at the frontier, and dense baselines such as Llama class models provide contrast in how tensor parallelism and memory pressure behave without expert routing.',
+    relatedTerms: ['mixture-of-experts', 'active-parameters', 'tensor-parallelism', 'quantization'],
+    articleSlugs: [INFERENCEMAX, INFERENCEX_V2],
+  },
+  {
+    slug: 'reasoning-model',
+    term: 'Reasoning model',
+    aliases: ['thinking model', 'test-time compute', 'chain-of-thought model'],
+    category: 'Model architecture',
+    plainEnglish:
+      'A reasoning model generates long hidden chains of thought before answering, trading extra output tokens for better results on hard problems.',
+    definition:
+      'A reasoning model is an LLM trained to spend additional generated tokens working through a problem, producing extended intermediate reasoning before or alongside its final answer.',
+    explanation:
+      'Instead of scaling only training compute, reasoning models scale test time compute: they think in tokens. A single math or coding query can trigger thousands of tokens of internal deliberation, so output lengths explode relative to chat models. For serving, that shifts load heavily toward decode, inflates KV cache residency per request, and makes tokens per second per user the metric that decides whether a hard query answers in seconds or minutes.',
+    significance:
+      'Reasoning turned inference into the scaling frontier: capability now improves by spending more at serving time, which multiplies demand for decode throughput. It reshaped hardware priorities toward memory bandwidth and interconnect, and it is a core reason agentic workloads dominate current benchmark design.',
+    benchmarkContext:
+      'The frontier models InferenceX benchmarks are reasoning capable, and its scenarios reflect their traffic: long generations in fixed sequence tests and full agent sessions in AgentX, where deliberation and tool use interleave over many turns.',
+    relatedTerms: ['decode', 'interactivity', 'agentic-inference', 'kv-cache', 'throughput'],
+    articleSlugs: [AGENTIC_WORKLOADS, GB200_R1],
+  },
+  {
+    slug: 'tokenization',
+    term: 'Tokenization',
+    aliases: ['tokenizer', 'token', 'byte pair encoding', 'BPE'],
+    category: 'Model architecture',
+    plainEnglish:
+      'Tokenization splits text into the sub word units a model actually reads and writes, and every performance and price number is quoted in them.',
+    definition:
+      'Tokenization is the conversion between text and the discrete token IDs a model processes, using a fixed vocabulary learned with methods such as byte pair encoding.',
+    explanation:
+      'A tokenizer maps common words to single tokens and rarer strings to several, with English averaging very roughly four characters per token. Vocabularies differ between model families, so identical text can produce meaningfully different token counts across models. Everything downstream is denominated in tokens: context windows, KV cache size, throughput, latency per token, and price per million tokens all count these units, not characters or words.',
+    significance:
+      'Token efficiency is a hidden price lever, since a model that needs fewer tokens for the same content is cheaper at an identical per token rate. Comparing providers or benchmarks without normalizing for tokenizer differences quietly distorts cost and speed conclusions.',
+    benchmarkContext:
+      'InferenceX metrics are token denominated, and its AgentX traces replace original text with deterministic synthetic tokens while preserving per turn token counts, so replayed sessions stress serving systems with the same token arithmetic as the source workload.',
+    relatedTerms: ['throughput', 'cost-per-million-tokens', 'context-window', 'trace-replay'],
+    articleSlugs: [AGENT_BENCHMARK, INFERENCEMAX],
+  },
+  {
+    slug: 'sequence-parallelism',
+    term: 'Sequence parallelism',
+    aliases: ['sequence parallel', 'SP'],
+    category: 'Parallelism',
+    plainEnglish:
+      'Sequence parallelism splits a single long sequence across chips, so the tokens of one request are processed by several accelerators at once.',
+    definition:
+      'Sequence parallelism is a strategy that partitions the token dimension of a sequence across devices, dividing activation memory and attention work for very long inputs.',
+    explanation:
+      'Tensor parallelism splits weights, while sequence parallelism splits the sequence itself: each chip holds a slice of the tokens and the associated activations and KV state. Attention then needs communication, since queries on one chip must meet keys and values on others, which ring style attention algorithms overlap with compute. In inference the closely related context parallel approach is what makes prefill of contexts with hundreds of thousands of tokens tractable.',
+    significance:
+      'Sequence style partitioning is the answer when one request, not the batch, is too large: a single million token prefill can exceed the memory and time budget of any one chip. It converts context length from a hard wall into a scaling dimension, at the price of interconnect traffic.',
+    benchmarkContext:
+      'InferenceX records the full parallelism plan of each recipe, and long context agentic scenarios are where sequence and context partitioning choices, together with interconnect quality, visibly separate systems with similar single chip specifications.',
+    relatedTerms: [
+      'context-parallelism',
+      'tensor-parallelism',
+      'context-window',
+      'prefill',
+      'nvlink',
+    ],
+    articleSlugs: [AGENT_BENCHMARK, VR_RUBIN],
+  },
+  {
+    slug: 'all-gather',
+    term: 'All-gather',
+    aliases: ['allgather', 'gather collective'],
+    category: 'Parallelism',
+    plainEnglish:
+      'All-gather is a group communication step where every chip ends up holding the combined data that started out split across all of them.',
+    definition:
+      'All-gather is a collective operation in which each participating device contributes its shard and every device receives the concatenation of all shards.',
+    explanation:
+      'Sharded execution constantly needs to reassemble full tensors: weight shards before a matrix multiply in some tensor parallel layouts, or per device activations before an operation that needs the whole hidden state. All-gather moves each shard to every rank, typically over a ring or tree schedule, and its cost grows with tensor size and the number of participants. It is the inverse companion of reduce-scatter, and the two composed together form an all-reduce.',
+    significance:
+      'Together with all-reduce and all-to-all, all-gather is one of the handful of collectives that decide whether a parallelism plan scales. Its latency sits on the critical path of every layer that uses it, which is why scale up bandwidth between chips is so heavily marketed.',
+    benchmarkContext:
+      'Every multi chip recipe InferenceX benchmarks exercises collectives through its parallelism plan, and the CollectiveX workstream measures operations like this directly across vendors so communication behavior can be compared outside full model runs.',
+    relatedTerms: [
+      'all-reduce',
+      'reduce-scatter',
+      'tensor-parallelism',
+      'nvlink',
+      'scale-up-vs-scale-out',
+    ],
+    articleSlugs: [VR_RUBIN, INFERENCEX_V2],
+  },
+  {
+    slug: 'reduce-scatter',
+    term: 'Reduce-scatter',
+    aliases: ['reduce scatter collective'],
+    category: 'Parallelism',
+    plainEnglish:
+      'Reduce-scatter sums matching data from every chip and leaves each chip holding just its own slice of the combined result.',
+    definition:
+      'Reduce-scatter is a collective operation that element wise reduces tensors contributed by all devices and distributes the reduced result in shards, one shard per device.',
+    explanation:
+      'When every rank computes a partial result for the same tensor, the partials must be summed. Reduce-scatter does the summation and hands each rank only the slice it will need next, avoiding the waste of giving everyone the full reduced tensor. An all-reduce is exactly a reduce-scatter followed by an all-gather, so schedulers choose between the fused and split forms depending on what the next operation actually consumes.',
+    significance:
+      'Using reduce-scatter instead of a full all-reduce halves the data each rank must receive when only a shard is needed, which matters at NVL72 scale where collective traffic competes with the model itself for interconnect bandwidth. Overlap of these collectives with compute is a defining quality of mature serving stacks.',
+    benchmarkContext:
+      'InferenceX recipes with tensor parallel sharding trigger reduction collectives in every transformer layer, and CollectiveX exists precisely to publish cross vendor measurements of these primitives at the message sizes inference actually uses.',
+    relatedTerms: ['all-reduce', 'all-gather', 'tensor-parallelism', 'nvlink'],
+    articleSlugs: [VR_RUBIN, GB200_KIMI],
+  },
+  {
+    slug: 'infiniband',
+    term: 'InfiniBand',
+    abbreviation: 'IB',
+    aliases: ['IB', 'InfiniBand networking', 'NDR InfiniBand'],
+    category: 'Hardware',
+    plainEnglish:
+      'InfiniBand is a high speed, low latency network fabric that connects servers in AI clusters, carrying traffic between nodes that NVLink cannot reach.',
+    definition:
+      'InfiniBand is a switched network fabric with native RDMA support, used as the scale out interconnect between nodes in most large NVIDIA based AI clusters.',
+    explanation:
+      'Inside a node or rack, chips talk over scale up links such as NVLink. Beyond that boundary, traffic crosses the scale out network, where InfiniBand competes with RDMA capable Ethernet. InfiniBand offers microsecond scale latency, hundreds of gigabits per second per link in current generations, and in network reduction features such as SHARP. Multi node inference, disaggregated prefill and decode, and wide expert parallelism all place their cross node collectives and KV transfers on this fabric.',
+    significance:
+      'Once a model spans nodes, the network joins the compute as a first order performance component. Fabric choice shapes cluster cost and vendor lock in, and the InfiniBand versus Ethernet contest is one of the central competitive battles in AI infrastructure.',
+    benchmarkContext:
+      'InferenceX multi node recipes, including disaggregated and rack scale results, run over the scale out fabric of the host cluster, and recipe metadata records the interconnect because it materially affects reproducibility of cross node numbers.',
+    relatedTerms: [
+      'rdma',
+      'nvlink',
+      'scale-up-vs-scale-out',
+      'disaggregated-inference',
+      'wide-expert-parallelism',
+    ],
+    articleSlugs: [GB200_R1, VR_RUBIN],
+  },
+  {
+    slug: 'rdma',
+    term: 'Remote direct memory access',
+    abbreviation: 'RDMA',
+    aliases: ['RDMA', 'RoCE', 'GPUDirect RDMA'],
+    category: 'Hardware',
+    plainEnglish:
+      'RDMA lets one machine read or write the memory of another directly over the network, without either CPU copying data along the way.',
+    definition:
+      'Remote direct memory access is a networking capability where the network adapter moves data straight between the memories of two machines, bypassing operating system and CPU copy overhead.',
+    explanation:
+      'A conventional network stack copies data through kernel buffers on both ends, burning CPU cycles and latency. RDMA adapters transfer directly between registered memory regions, and GPUDirect extends this so adapters write straight into accelerator HBM. InfiniBand has RDMA built in, while RoCE carries the same verbs over Ethernet. Collective libraries such as NCCL and RCCL, and KV cache transfer paths in disaggregated serving, are built on these primitives.',
+    significance:
+      'RDMA is the floor the whole distributed AI stack stands on: without it, cross node collectives and cache transfers would bottleneck on CPUs long before saturating the links. The RoCE variant is what allows Ethernet based clusters to compete with InfiniBand at lower cost.',
+    benchmarkContext:
+      'Every InferenceX multi node result depends on RDMA transports underneath its collectives and, for disaggregated recipes, underneath prefill to decode KV movement, so transport maturity is part of what separates otherwise similar cluster results.',
+    relatedTerms: [
+      'infiniband',
+      'nvlink',
+      'disaggregated-inference',
+      'kv-cache',
+      'scale-up-vs-scale-out',
+    ],
+    articleSlugs: [GB200_R1, AGENTX_DSV4_GB200_GB300],
+  },
+  {
+    slug: 'ualink',
+    term: 'UALink',
+    aliases: ['Ultra Accelerator Link', 'UALoE'],
+    category: 'Hardware',
+    plainEnglish:
+      'UALink is an open industry standard for the fast scale up links between accelerators in a rack, the ecosystem answer to NVLink.',
+    definition:
+      'UALink is an open interconnect specification for accelerator to accelerator scale up communication, letting many chips in a rack share memory traffic at switch fabric speeds.',
+    explanation:
+      'NVLink proved that a rack of chips joined by a low latency memory semantic fabric can behave like one giant accelerator, but it is proprietary. The UALink consortium defines an open equivalent so vendors beyond NVIDIA can build rack scale domains. UALink over Ethernet, shortened to UALoE, runs the protocol over Ethernet switching. AMD Helios generation racks adopt this path to form 72 chip scale up domains comparable in structure to NVL72.',
+    significance:
+      'Scale up domain size increasingly decides serving architecture, since wide expert parallelism and disaggregation want dozens of chips within one fast fabric. An open standard determines whether rack scale inference stays a single vendor advantage or becomes an ecosystem capability.',
+    benchmarkContext:
+      'InferenceX lists UALoE72 class systems such as AMD MI455X racks alongside NVL72 systems in its hardware coverage, so rack scale fabrics can be compared head to head on identical model workloads as they reach the market.',
+    relatedTerms: ['nvlink', 'nvl72', 'scale-up-vs-scale-out', 'wide-expert-parallelism'],
+    articleSlugs: [VR_RUBIN, AGENTX_V3],
+  },
+  {
+    slug: 'tdp',
+    term: 'Thermal design power',
+    abbreviation: 'TDP',
+    aliases: ['TDP', 'board power', 'all-in power'],
+    category: 'Hardware',
+    plainEnglish:
+      'TDP is the sustained power a chip is designed to draw and shed as heat, the headline wattage on every accelerator spec sheet.',
+    definition:
+      'Thermal design power is the maximum sustained power envelope a chip is engineered to operate within, which its cooling system must dissipate continuously.',
+    explanation:
+      'Modern accelerators run around or above the kilowatt mark per chip, and a full rack system multiplies that into six figures of watts. TDP alone also undersells the true bill: memory, networking, CPUs, power conversion losses, and cooling overhead stack on top, which is why all in power per chip is meaningfully higher than the chip TDP. Datacenter capacity is sold in megawatts, so these envelopes translate directly into how many accelerators a site can host.',
+    significance:
+      'Power has become the binding constraint of AI buildout, ahead of capital in many markets. Rising per chip TDP forced the shift to liquid cooling and made performance per watt, not just performance per dollar, a primary axis for comparing silicon generations.',
+    benchmarkContext:
+      'InferenceX derives energy per token and tokens per megawatt using per chip all in power figures that include cooling and infrastructure overhead above TDP, and the PowerX workstream is extending this from rated figures toward measured draw during runs.',
+    relatedTerms: ['energy-per-token', 'tokens-per-megawatt', 'pue', 'total-cost-of-ownership'],
+    articleSlugs: [INFERENCEX_V2, VR_RUBIN],
+  },
+  {
+    slug: 'pue',
+    term: 'Power usage effectiveness',
+    abbreviation: 'PUE',
+    aliases: ['PUE', 'datacenter efficiency ratio'],
+    category: 'Hardware',
+    plainEnglish:
+      'PUE measures how much total datacenter power is consumed for every watt that actually reaches the computing equipment inside it.',
+    definition:
+      'Power usage effectiveness is the ratio of total facility power to IT equipment power, where a value of 1.0 would mean every watt goes to computing.',
+    explanation:
+      'Cooling, power conversion, and facility systems consume energy on top of the servers themselves. A PUE of 1.5 means half again as much power is spent on overhead as on IT load, while modern hyperscale AI facilities push toward 1.1 and below with liquid cooling and high efficiency distribution. Because AI campuses are sized in hundreds of megawatts, small PUE differences move enormous absolute energy and cost figures.',
+    significance:
+      'PUE links chip level efficiency to facility economics: every watt a chip draws is multiplied by PUE at the utility meter. As power availability gates AI buildout, facility efficiency became part of the competitive calculus alongside silicon and software.',
+    benchmarkContext:
+      'The all in power figures behind InferenceX energy per token and tokens per megawatt metrics incorporate facility overhead consistent with modern AI datacenter PUE, so its cost and energy comparisons reflect delivered facility economics rather than bare chip wattage.',
+    relatedTerms: ['tdp', 'energy-per-token', 'tokens-per-megawatt', 'total-cost-of-ownership'],
+    articleSlugs: [INFERENCEX_V2, INFERENCEMAX],
+  },
+  {
+    slug: 'int8',
+    term: 'INT8',
+    aliases: ['8-bit integer quantization', 'W8A8'],
+    category: 'Numerical precision',
+    plainEnglish:
+      'INT8 stores numbers as 8 bit integers with a scale factor, halving memory versus 16 bit formats and doubling math rates on supporting hardware.',
+    definition:
+      'INT8 is an 8 bit integer numerical format used with per tensor or per channel scale factors to represent model weights and activations in quantized inference.',
+    explanation:
+      'Integer quantization maps floating point values onto 256 evenly spaced levels via a scale, and sometimes a zero point. Uniform spacing handles outliers poorly, so techniques such as SmoothQuant migrate activation outliers into weights before quantizing both sides, in the W8A8 pattern. On older accelerator generations INT8 was the primary fast path below 16 bits, while newer chips add FP8, whose exponent gives it a wider dynamic range at the same bit width.',
+    significance:
+      'INT8 defined the first mainstream wave of LLM quantization and remains relevant on hardware without floating point 8 bit support. The INT8 versus FP8 contrast also illustrates the core quantization tradeoff between uniform precision and dynamic range.',
+    benchmarkContext:
+      'InferenceX labels every result with its precision, and its comparison families exist because format changes moved curves so much. Modern recipes on Blackwell and MI350 class hardware favor FP8 and FP4 paths, with integer formats appearing in specific weight quantized configurations.',
+    relatedTerms: ['quantization', 'fp8', 'int4', 'weight-only-quantization', 'bf16'],
+    articleSlugs: [B200_KIMI, INFERENCEMAX],
+  },
+  {
+    slug: 'weight-only-quantization',
+    term: 'Weight-only quantization',
+    aliases: ['W4A16', 'AWQ', 'GPTQ'],
+    category: 'Numerical precision',
+    plainEnglish:
+      'Weight-only quantization compresses just the stored model weights to low precision while the math still runs in higher precision formats.',
+    definition:
+      'Weight-only quantization stores model weights in a low bit format such as 4 bit integers while keeping activations and arithmetic at higher precision, as in the W4A16 pattern.',
+    explanation:
+      'Weights are static and can be carefully quantized offline with methods such as GPTQ and AWQ, which choose scales and orderings that minimize quality loss. Activations are dynamic and harder to compress, so leaving them at 16 bits sidesteps their outlier problem. At serve time the kernel dequantizes weights on the fly, so memory traffic shrinks even though the multiply accumulate math itself does not get faster.',
+    significance:
+      'Because decode is memory bound, cutting weight bytes directly speeds up token generation and lets larger models fit on fewer chips. Weight-only methods made large open models runnable on modest hardware and remain the standard recipe when activation quantization would cost too much quality.',
+    benchmarkContext:
+      'InferenceX distinguishes weight-only configurations from full low precision paths in its precision labels, since a W4A16 recipe and an NVFP4 recipe make very different claims about which hardware units and bandwidth budgets produced a curve.',
+    relatedTerms: ['quantization', 'int4', 'int8', 'kv-cache-quantization', 'memory-bandwidth'],
+    articleSlugs: [B200_KIMI, MI355X_KIMI],
+  },
+  {
+    slug: 'block-scaling',
+    term: 'Block scaling',
+    aliases: ['microscaling', 'MX formats', 'block floating point'],
+    category: 'Numerical precision',
+    plainEnglish:
+      'Block scaling gives each small group of low precision numbers its own shared scale factor, recovering range that tiny formats lack on their own.',
+    definition:
+      'Block scaling is a quantization structure where values are stored in a very low bit format and each fixed size block of them shares one higher precision scale factor.',
+    explanation:
+      'A 4 bit number can represent only a handful of distinct magnitudes, far too few to span the dynamic range of model tensors. Grouping values into blocks of around 16 or 32 elements and attaching a shared scale lets each block center the format on its own magnitude. The MX standard formats such as MXFP4 and MXFP8 use power of two scales, while NVFP4 uses FP8 scales over 16 element blocks for finer granularity.',
+    significance:
+      'Block scaling is the enabling idea behind the FP4 generation of inference: without per block scales, 4 bit floating point would be unusable for frontier models. Scale format and block size choices are now genuine differentiators between hardware vendors and quantization recipes.',
+    benchmarkContext:
+      'The NVFP4 and MXFP4 results across InferenceX Blackwell and MI355X coverage are block scaled formats, and the compare-precision family exists largely to show what these recipes give up or gain against FP8 baselines on identical hardware.',
+    relatedTerms: ['nvfp4', 'mxfp4', 'fp4', 'fp8', 'quantization'],
+    articleSlugs: [B200_GLM5, GB300_DSV4],
+  },
+  {
+    slug: 'flash-attention',
+    term: 'FlashAttention',
+    aliases: ['flash attention kernel', 'fused attention'],
+    category: 'Software',
+    plainEnglish:
+      'FlashAttention computes exact attention in fast on-chip memory tiles, avoiding the huge intermediate matrix that made attention slow and memory hungry.',
+    definition:
+      'FlashAttention is an attention algorithm that tiles the computation through on-chip SRAM and rescales results incrementally, producing exact attention without materializing the full score matrix in HBM.',
+    explanation:
+      'Naive attention writes a score matrix that grows with the square of sequence length to main memory and reads it back, so bandwidth rather than arithmetic sets its speed. FlashAttention fuses the whole computation into one kernel that streams blocks of keys and values through on-chip memory, using an online softmax to keep results exact. Successive versions and vendor implementations extend the idea to new hardware generations, head layouts, and inference specific decode paths.',
+    significance:
+      'This kernel family is what made long context practical, turning attention from the dominant cost of long sequences into one manageable component. It is also the canonical example of how a single well engineered kernel can shift performance across the entire industry.',
+    benchmarkContext:
+      'Every serving engine InferenceX benchmarks relies on fused attention kernels descended from this line, through libraries such as FlashInfer on NVIDIA hardware and AITER on AMD, and kernel improvements there routinely move published curves without any hardware change.',
+    relatedTerms: ['flashinfer', 'aiter', 'memory-bandwidth', 'context-window', 'kernel-fusion'],
+    articleSlugs: [SGLANG_056, MI355X_QWEN],
+  },
+  {
+    slug: 'triton',
+    term: 'Triton',
+    aliases: ['OpenAI Triton', 'Triton kernel language'],
+    category: 'Software',
+    plainEnglish:
+      'Triton is a Python-based language for writing custom accelerator kernels, letting ML engineers get near hand-tuned speed without writing low level code.',
+    definition:
+      'Triton is an open source kernel programming language and compiler that lets developers write high performance accelerator kernels in Python-like code, portable across vendors that maintain backends.',
+    explanation:
+      'Writing peak performance kernels traditionally requires vendor specific expertise in CUDA or assembly level tuning. Triton raises the abstraction: the developer writes block level programs and the compiler handles memory coalescing, tiling, and scheduling. NVIDIA, AMD, and other vendors maintain backends, so one kernel source can target multiple architectures. Serving engines use it for fused operations, quantization paths, and MoE kernels where no vendor library fits. The name collides with NVIDIA Triton Inference Server, a separate model serving product.',
+    significance:
+      'Triton lowered the barrier between model researchers and hardware performance, and its cross vendor backends are strategically important because kernels written in it are not locked to one chip family. How well a vendor runs the Triton ecosystem has become part of its software story.',
+    benchmarkContext:
+      'Engines in InferenceX recipes ship substantial Triton kernel inventories alongside CUDA, CUTLASS, and AITER code, so compiler and backend maturity is one of the quiet forces behind curve movement between engine versions on both NVIDIA and AMD systems.',
+    relatedTerms: ['cuda', 'rocm', 'kernel-fusion', 'inference-engine', 'gemm'],
+    articleSlugs: [MI355X_DSV4, INFERENCEX_V2],
+  },
+  {
+    slug: 'cutlass',
+    term: 'CUTLASS',
+    aliases: ['CUDA templates for linear algebra', 'CuTe'],
+    category: 'Software',
+    plainEnglish:
+      'CUTLASS is an NVIDIA template library that provides the building blocks for writing matrix multiply kernels that approach peak hardware speed.',
+    definition:
+      'CUTLASS is an open source NVIDIA library of composable C++ templates for building GEMM and related kernels that target tensor cores across GPU generations.',
+    explanation:
+      'Peak matrix multiply performance demands precise choreography of tensor core instructions, shared memory movement, and asynchronous pipelines, and it changes with every architecture. CUTLASS packages that choreography as composable pieces, with its CuTe layer describing data layouts, so kernel authors assemble near peak GEMMs and fuse epilogues such as bias, activation, or quantization scaling instead of starting from scratch. Much of the high performance kernel work in serving engines builds on it directly.',
+    significance:
+      'CUTLASS is where NVIDIA teaches the ecosystem to use each new tensor core generation, including the FP4 and FP8 paths on Blackwell. The speed at which its patterns propagate into engines is a real component of how quickly new silicon reaches its advertised performance.',
+    benchmarkContext:
+      'The GEMM and attention kernels behind InferenceX results on NVIDIA hardware lean heavily on CUTLASS derived code, and version bumps of these libraries inside engine images are a recurring source of day over day curve movement the platform tracks.',
+    relatedTerms: ['cuda', 'gemm', 'kernel-fusion', 'flashinfer', 'nvfp4'],
+    articleSlugs: [SGLANG_056, GB300_DSV4],
+  },
+  {
+    slug: 'nccl',
+    term: 'NCCL',
+    abbreviation: 'NCCL',
+    aliases: ['NVIDIA Collective Communications Library', 'RCCL'],
+    category: 'Software',
+    plainEnglish:
+      'NCCL is the NVIDIA library that moves data between chips during collective operations, with RCCL as its AMD counterpart.',
+    definition:
+      'NCCL is the NVIDIA collective communications library implementing operations such as all-reduce, all-gather, and all-to-all across GPUs, over NVLink within nodes and RDMA fabrics between them.',
+    explanation:
+      'Frameworks do not talk to interconnects directly; they call a collectives library that discovers the topology and picks algorithms and channel schedules for each message size. NCCL handles that for NVIDIA systems, and AMD maintains RCCL with a matching interface for ROCm platforms. Tuning is fabric specific: ring versus tree algorithms, protocol thresholds, and channel counts all shift with topology, which is why the same model can communicate very differently on two clusters.',
+    significance:
+      'Every multi chip inference and training job stands on this layer, and a collectives regression can silently tax an entire fleet. Interface compatibility between NCCL and RCCL is also load bearing for portability, since engines can target one collectives API across vendors.',
+    benchmarkContext:
+      'InferenceX multi chip recipes exercise these libraries in every tensor parallel and expert parallel layer, and CollectiveX measures the underlying collective performance directly across vendors at inference relevant message sizes, separating fabric behavior from model behavior.',
+    relatedTerms: ['all-reduce', 'all-gather', 'all-to-all', 'nvlink', 'rdma'],
+    articleSlugs: [GB200_KIMI, VR_RUBIN],
+  },
+  {
+    slug: 'gemm',
+    term: 'GEMM',
+    abbreviation: 'GEMM',
+    aliases: ['general matrix multiply', 'matrix multiplication kernel'],
+    category: 'Software',
+    plainEnglish:
+      'GEMM is the general matrix multiply operation, the single computation that consumes most of the arithmetic in training and serving neural networks.',
+    definition:
+      'GEMM is the general matrix to matrix multiply routine, the core primitive that linear layers, attention projections, and expert computations in neural networks reduce to.',
+    explanation:
+      'Transformers are mostly stacks of linear transformations, so serving a model means executing enormous numbers of matrix multiplies. Tensor cores exist specifically to accelerate them, and peak TFLOP/s specifications are quoted for these dense operations. Shape determines efficiency: prefill produces large square-ish multiplies that saturate compute, while decode produces skinny ones that are bandwidth bound. MoE adds grouped GEMMs, where many small expert multiplies are batched into one efficient launch.',
+    significance:
+      'GEMM efficiency is the substrate of every performance claim in the industry. The gap between delivered and peak GEMM throughput at real serving shapes, especially the thin matrices of decode, explains much of why spec sheet ratios fail to predict benchmark rankings.',
+    benchmarkContext:
+      'Behind every InferenceX curve sits a stack of GEMM kernels from libraries such as CUTLASS, hipBLASLt, and Triton generated code, and quantized recipes ultimately stand on how well each vendor executes low precision GEMMs at the shapes its scheduler produces.',
+    relatedTerms: [
+      'cutlass',
+      'arithmetic-intensity',
+      'memory-bound-vs-compute-bound',
+      'mixture-of-experts',
+      'triton',
+    ],
+    articleSlugs: [INFERENCEX_V2, MI355X_GLM5],
+  },
+  {
+    slug: 'kernel-fusion',
+    term: 'Kernel fusion',
+    aliases: ['fused kernels', 'operator fusion'],
+    category: 'Software',
+    plainEnglish:
+      'Kernel fusion merges several small chip operations into one, so intermediate data stays in fast memory instead of bouncing through HBM.',
+    definition:
+      'Kernel fusion combines multiple consecutive operations into a single kernel launch, keeping intermediate results in registers or on-chip memory rather than writing them to main memory between steps.',
+    explanation:
+      'An unfused sequence like matrix multiply, bias add, and activation writes its intermediate tensor to HBM after each step and reads it back for the next. Fusing them into one kernel eliminates those round trips and the launch overhead between them. Fusion happens by hand in libraries, through template epilogues in CUTLASS style code, and automatically in compilers such as torch.compile and Triton based stacks. FlashAttention is the most famous single example of the idea.',
+    significance:
+      'In memory bound inference, removing intermediate traffic is worth more than raw arithmetic improvements, so fusion is one of the most reliable levers engines have. A large share of version over version engine speedups reduces to more aggressive or better targeted fusion.',
+    benchmarkContext:
+      'InferenceX recipes pin engine images whose fusion inventories differ by version and by hardware backend, which is a recurring reason day over day tracking shows curves moving on unchanged silicon as fused kernels land for new models and precisions.',
+    relatedTerms: [
+      'flash-attention',
+      'cuda-graphs',
+      'gemm',
+      'memory-bandwidth',
+      'inference-engine',
+    ],
+    articleSlugs: [MI355X_KIMI, SGLANG_056],
+  },
 ] as const satisfies readonly GlossaryEntry[];
 
 export type GlossaryPreview = Pick<
diff --git a/packages/app/src/lib/i18n.test.ts b/packages/app/src/lib/i18n.test.ts
index 7d89f66bc..3a01e8fd5 100644
--- a/packages/app/src/lib/i18n.test.ts
+++ b/packages/app/src/lib/i18n.test.ts
@@ -1,12 +1,14 @@
 import { describe, expect, it } from 'vitest';
 
 import { SITE_URL } from '@semianalysisai/inferencex-constants';
+import { DASHBOARD_ROUTES } from './dashboard-routes';
 
 import {
   enAlternates,
   hasZhSibling,
   isZhPathname,
   languageAlternates,
+  localePath,
   switchLocalePath,
   zhAlternates,
   zhPath,
@@ -23,6 +25,14 @@ describe('zhPath', () => {
   });
 });
 
+describe('localePath', () => {
+  it('keeps English paths and prefixes Chinese paths', () => {
+    expect(localePath('/about', 'en')).toBe('/about');
+    expect(localePath('/about', 'zh')).toBe('/zh/about');
+    expect(localePath('/', 'zh')).toBe('/zh');
+  });
+});
+
 describe('isZhPathname', () => {
   it('matches the zh root and zh children', () => {
     expect(isZhPathname('/zh')).toBe(true);
@@ -45,6 +55,13 @@ describe('hasZhSibling', () => {
     expect(hasZhSibling('/collectivex')).toBe(true);
   });
 
+  it.each(DASHBOARD_ROUTES.filter((route) => route.localeMirrored))(
+    'derives the mirrored dashboard route "$path" from the canonical registry',
+    ({ path }) => {
+      expect(hasZhSibling(path)).toBe(true);
+    },
+  );
+
   it('matches blog and compare child paths', () => {
     expect(hasZhSibling('/blog/some-post')).toBe(true);
     expect(hasZhSibling('/compare')).toBe(true);
@@ -57,8 +74,8 @@ describe('hasZhSibling', () => {
   });
 
   it('matches datasets, gated tabs, and agentic detail pages', () => {
-    expect(hasZhSibling('/datasets')).toBe(true);
-    expect(hasZhSibling('/datasets/some-set/conversations/abc123')).toBe(true);
+    expect(hasZhSibling('/agentx')).toBe(true);
+    expect(hasZhSibling('/agentx/some-set/conversations/abc123')).toBe(true);
     expect(hasZhSibling('/ai-chart')).toBe(true);
     expect(hasZhSibling('/current-inferencex-image')).toBe(true);
     expect(hasZhSibling('/feedback')).toBe(true);
@@ -107,8 +124,8 @@ describe('switchLocalePath', () => {
   });
 
   it('switches datasets pages within the language trees', () => {
-    expect(switchLocalePath('/datasets')).toBe('/zh/datasets');
-    expect(switchLocalePath('/zh/datasets/some-set')).toBe('/datasets/some-set');
+    expect(switchLocalePath('/agentx')).toBe('/zh/agentx');
+    expect(switchLocalePath('/zh/agentx/some-set')).toBe('/agentx/some-set');
   });
 
   it('falls back to the other homepage for unmirrored paths', () => {
diff --git a/packages/app/src/lib/i18n.ts b/packages/app/src/lib/i18n.ts
index 13f4e4e53..6f33e04a7 100644
--- a/packages/app/src/lib/i18n.ts
+++ b/packages/app/src/lib/i18n.ts
@@ -1,4 +1,5 @@
 import { SITE_URL } from '@semianalysisai/inferencex-constants';
+import { DASHBOARD_ROUTES } from '@/lib/dashboard-routes';
 
 /**
  * Minimal locale plumbing for the Simplified Chinese (/zh) page tree.
@@ -24,6 +25,11 @@ export function zhPath(enPath: string): string {
   return enPath === '/' ? ZH_PREFIX : `${ZH_PREFIX}${enPath}`;
 }
 
+/** Map an English internal path to the requested locale tree. */
+export function localePath(enPath: string, locale: Locale): string {
+  return locale === 'zh' ? zhPath(enPath) : enPath;
+}
+
 export function isZhPathname(pathname: string): boolean {
   return pathname === ZH_PREFIX || pathname.startsWith(`${ZH_PREFIX}/`);
 }
@@ -38,19 +44,9 @@ export const ZH_MIRRORED_ROUTES: readonly { path: string; exact?: boolean }[] =
   { path: '/', exact: true },
   { path: '/overview', exact: true },
   { path: '/api', exact: true },
-  { path: '/inference', exact: true },
-  { path: '/inference/agentic' },
-  { path: '/evaluation', exact: true },
-  { path: '/historical', exact: true },
-  { path: '/calculator', exact: true },
-  { path: '/reliability', exact: true },
-  { path: '/gpu-specs', exact: true },
-  { path: '/gpu-metrics', exact: true },
-  { path: '/collectivex', exact: true },
-  { path: '/submissions', exact: true },
-  { path: '/ai-chart', exact: true },
-  { path: '/current-inferencex-image', exact: true },
-  { path: '/feedback', exact: true },
+  ...DASHBOARD_ROUTES.filter((route) => route.localeMirrored).map((route) => ({
+    path: route.path,
+  })),
   { path: '/about', exact: true },
   { path: '/quotes', exact: true },
   { path: '/land-acknowledgement', exact: true },
@@ -60,7 +56,10 @@ export const ZH_MIRRORED_ROUTES: readonly { path: string; exact?: boolean }[] =
   { path: '/compare-spec-decode' },
   { path: '/blog' },
   { path: '/glossary' },
-  { path: '/datasets' },
+  { path: '/chips' },
+  { path: '/agentx' },
+  { path: '/rankings' },
+  { path: '/run' },
 ];
 
 export function hasZhSibling(enPathname: string): boolean {
diff --git a/packages/app/src/lib/inference-model-meta.ts b/packages/app/src/lib/inference-model-meta.ts
new file mode 100644
index 000000000..3649d26c5
--- /dev/null
+++ b/packages/app/src/lib/inference-model-meta.ts
@@ -0,0 +1,44 @@
+/**
+ * @file inference-model-meta.ts
+ * @description SEO copy for the `/inference/<model>` and `/zh/inference/<model>`
+ * pages, shared by both page files so the English and Chinese descriptions
+ * cannot drift structurally. Prose style mirrors `TAB_META.inference` /
+ * `TAB_META_ZH.inference` with the model name substituted in.
+ */
+import type { InferenceModelSlug } from '@/lib/inference-model-slug';
+
+export interface InferenceModelMeta {
+  title: string;
+  description: string;
+}
+
+/** `label` adds detail beyond `seoName` (e.g. parameter count) for most
+ * models, but for some they are identical — skip the parenthetical then so
+ * the description doesn't read "DeepSeek R1 (DeepSeek R1)". */
+function nameWithDetail(entry: InferenceModelSlug): string {
+  return entry.label === entry.seoName ? entry.seoName : `${entry.seoName} (${entry.label})`;
+}
+
+export function inferenceModelMeta(entry: InferenceModelSlug): InferenceModelMeta {
+  return {
+    // Lead with the model name — that's the phrase people search
+    // ("kimi k3 inference benchmark") and it must survive Google's ~60-char
+    // SERP truncation.
+    title: `${entry.seoName} Inference Benchmarks`,
+    description:
+      `Compare ${nameWithDetail(entry)} latency, throughput, cost, and ` +
+      'time-to-first-token across chips and serving frameworks. Every datapoint is ' +
+      'produced by a public, reproducible GitHub Actions run.',
+  };
+}
+
+export function inferenceModelMetaZh(entry: InferenceModelSlug): InferenceModelMeta {
+  // Model names, SKUs, and product names stay in English per the translation
+  // quality bar; the surrounding prose follows TAB_META_ZH.inference.
+  return {
+    title: `${entry.seoName} 推理基准测试`,
+    description:
+      `跨芯片与推理框架，对比 ${nameWithDetail(entry)} 推理的延迟、吞吐量、` +
+      '成本与首 token 延迟（TTFT）。每个数据点都来自公开的 GitHub Actions 运行，可复现、可审计。',
+  };
+}
diff --git a/packages/app/src/lib/inference-model-slug.test.ts b/packages/app/src/lib/inference-model-slug.test.ts
new file mode 100644
index 000000000..2562719f0
--- /dev/null
+++ b/packages/app/src/lib/inference-model-slug.test.ts
@@ -0,0 +1,168 @@
+import { describe, expect, it } from 'vitest';
+
+import { COMPARE_MODEL_SLUGS } from './compare-slug';
+import { getModelCategory, Model, MODEL_OPTIONS } from './data-mappings';
+import {
+  ACTIVE_INFERENCE_MODEL_SLUGS,
+  getInferenceModelBySlug,
+  INFERENCE_MODEL_ALIASES,
+  INFERENCE_MODEL_SLUGS,
+  inferenceModelForPathname,
+  inferenceModelPath,
+  inferenceModelRouteForSelection,
+  inferenceModelSlugForModel,
+} from './inference-model-slug';
+
+describe('INFERENCE_MODEL_SLUGS registry', () => {
+  it('covers every non-hidden dashboard model exactly once', () => {
+    const models = INFERENCE_MODEL_SLUGS.map((entry) => entry.model);
+    expect(new Set(models).size).toBe(models.length);
+    expect(new Set(models)).toEqual(new Set(MODEL_OPTIONS));
+  });
+
+  it('excludes hidden models', () => {
+    for (const entry of INFERENCE_MODEL_SLUGS) {
+      expect(getModelCategory(entry.model)).not.toBe('hidden');
+    }
+  });
+
+  it('has unique lowercase slugs shared with the compare vocabulary', () => {
+    const compareSlugs = new Set(COMPARE_MODEL_SLUGS.map((entry) => entry.slug));
+    const slugs = INFERENCE_MODEL_SLUGS.map((entry) => entry.slug);
+    expect(new Set(slugs).size).toBe(slugs.length);
+    for (const slug of slugs) {
+      expect(slug).toBe(slug.toLowerCase());
+      expect(compareSlugs.has(slug)).toBe(true);
+    }
+  });
+
+  it('keeps active (landing-promoted) entries a subset of the registry', () => {
+    for (const entry of ACTIVE_INFERENCE_MODEL_SLUGS) {
+      expect(getModelCategory(entry.model)).not.toBe('deprecated');
+      expect(INFERENCE_MODEL_SLUGS).toContain(entry);
+    }
+  });
+});
+
+describe('getInferenceModelBySlug', () => {
+  it('resolves canonical slugs', () => {
+    expect(getInferenceModelBySlug('kimi-k3')?.model).toBe(Model.Kimi_K3);
+    expect(getInferenceModelBySlug('deepseek-v4')?.model).toBe(Model.DeepSeek_V4_Pro);
+  });
+
+  it('is case-insensitive', () => {
+    expect(getInferenceModelBySlug('Kimi-K3')?.model).toBe(Model.Kimi_K3);
+  });
+
+  it('resolves g_model display names as aliases', () => {
+    for (const entry of INFERENCE_MODEL_SLUGS) {
+      expect(getInferenceModelBySlug(entry.model)?.slug).toBe(entry.slug);
+    }
+  });
+
+  it('resolves compare family aliases', () => {
+    expect(getInferenceModelBySlug('kimi')?.slug).toBe('kimi-k26');
+    expect(getInferenceModelBySlug('glm-5')?.slug).toBe('glm-5-1');
+  });
+
+  it('returns null for unknown and reserved segments', () => {
+    expect(getInferenceModelBySlug('agentic')).toBeNull();
+    expect(getInferenceModelBySlug('logs')).toBeNull();
+    expect(getInferenceModelBySlug('not-a-model')).toBeNull();
+  });
+
+  it('alias targets all resolve to canonical entries', () => {
+    for (const [alias, target] of Object.entries(INFERENCE_MODEL_ALIASES)) {
+      expect(getInferenceModelBySlug(alias)?.slug).toBe(target);
+    }
+  });
+});
+
+describe('inferenceModelSlugForModel / inferenceModelPath', () => {
+  it('round-trips every registry entry', () => {
+    for (const entry of INFERENCE_MODEL_SLUGS) {
+      expect(inferenceModelSlugForModel(entry.model)).toBe(entry.slug);
+      expect(inferenceModelPath(entry.slug)).toBe(`/inference/${entry.slug}`);
+    }
+  });
+
+  it('returns null for hidden models', () => {
+    expect(inferenceModelSlugForModel(Model.Llama3_1_70B)).toBeNull();
+  });
+});
+
+describe('inferenceModelForPathname', () => {
+  it('pins the model on English model pages', () => {
+    expect(inferenceModelForPathname('/inference/kimi-k3')).toBe(Model.Kimi_K3);
+    expect(inferenceModelForPathname('/inference/minimax-m3/')).toBe(Model.MiniMax_M3);
+  });
+
+  it('pins the model on /zh model pages', () => {
+    expect(inferenceModelForPathname('/zh/inference/kimi-k3')).toBe(Model.Kimi_K3);
+  });
+
+  it('resolves alias segments (pre-redirect client render)', () => {
+    expect(inferenceModelForPathname('/inference/Kimi-K3')).toBe(Model.Kimi_K3);
+    expect(inferenceModelForPathname('/inference/deepseek-v4-pro')).toBe(Model.DeepSeek_V4_Pro);
+    expect(inferenceModelForPathname('/inference/Kimi-K2.5')).toBe(Model.Kimi_K2_5);
+  });
+
+  it('ignores query strings and hashes', () => {
+    expect(inferenceModelForPathname('/inference/kimi-k3?i_seq=8k%2F1k#chart')).toBe(Model.Kimi_K3);
+  });
+
+  it('returns null off the model pages', () => {
+    expect(inferenceModelForPathname('/inference')).toBeNull();
+    expect(inferenceModelForPathname('/zh/inference')).toBeNull();
+    expect(inferenceModelForPathname('/inference/agentic')).toBeNull();
+    expect(inferenceModelForPathname('/inference/agentic/some-point')).toBeNull();
+    expect(inferenceModelForPathname('/inference/logs')).toBeNull();
+    expect(inferenceModelForPathname('/inference/kimi-k3/extra')).toBeNull();
+    expect(inferenceModelForPathname('/compare/kimi-k3-gb200-vs-mi355x')).toBeNull();
+    expect(inferenceModelForPathname('/')).toBeNull();
+    expect(inferenceModelForPathname('/zh')).toBeNull();
+    expect(inferenceModelForPathname('/inference/%E4%B8%8D%E5%AD%98%E5%9C%A8')).toBeNull();
+  });
+
+  it('returns null when usePathname yields no pathname (bare test renders)', () => {
+    expect(inferenceModelForPathname(null)).toBeNull();
+    expect(inferenceModelForPathname(undefined)).toBeNull();
+    expect(inferenceModelForPathname('')).toBeNull();
+  });
+});
+
+describe('inferenceModelRouteForSelection', () => {
+  it('rewrites model subroutes to the newly selected model', () => {
+    expect(inferenceModelRouteForSelection('/inference/kimi-k3', Model.DeepSeek_V4_Pro)).toBe(
+      '/inference/deepseek-v4',
+    );
+    expect(inferenceModelRouteForSelection('/zh/inference/kimi-k3', Model.DeepSeek_V4_Pro)).toBe(
+      '/zh/inference/deepseek-v4',
+    );
+  });
+
+  it('moves the base /inference page onto the model subroute', () => {
+    expect(inferenceModelRouteForSelection('/inference', Model.Kimi_K3)).toBe('/inference/kimi-k3');
+    expect(inferenceModelRouteForSelection('/zh/inference', Model.Kimi_K3)).toBe(
+      '/zh/inference/kimi-k3',
+    );
+  });
+
+  it('leaves every other surface alone', () => {
+    expect(inferenceModelRouteForSelection('/', Model.Kimi_K3)).toBeNull();
+    expect(inferenceModelRouteForSelection('/overview', Model.Kimi_K3)).toBeNull();
+    expect(inferenceModelRouteForSelection('/inference/agentic', Model.Kimi_K3)).toBeNull();
+    expect(inferenceModelRouteForSelection('/inference/agentic/point-1', Model.Kimi_K3)).toBeNull();
+    expect(
+      inferenceModelRouteForSelection('/compare/kimi-k3-h100-vs-h200', Model.Kimi_K3),
+    ).toBeNull();
+    expect(inferenceModelRouteForSelection('/zh', Model.Kimi_K3)).toBeNull();
+    expect(inferenceModelRouteForSelection(null, Model.Kimi_K3)).toBeNull();
+  });
+
+  it('keeps the same path when the selection already matches', () => {
+    expect(inferenceModelRouteForSelection('/inference/kimi-k3', Model.Kimi_K3)).toBe(
+      '/inference/kimi-k3',
+    );
+  });
+});
diff --git a/packages/app/src/lib/inference-model-slug.ts b/packages/app/src/lib/inference-model-slug.ts
new file mode 100644
index 000000000..c7db95177
--- /dev/null
+++ b/packages/app/src/lib/inference-model-slug.ts
@@ -0,0 +1,162 @@
+/**
+ * @file inference-model-slug.ts
+ * @description Model-slug registry for the indexable `/inference/<model>`
+ * subroutes.
+ *
+ * `/inference?g_model=Kimi-K3` keeps working as a share-link form, but query
+ * params are a weak indexing surface: crawlers treat them as views of one page
+ * and the canonical collapses them all to `/`. The `/inference/<model>` path
+ * form gives every model a stable, individually indexable URL with its own
+ * title, description, canonical, and hreflang pair — same playbook as the
+ * `/compare/<slug>` pages.
+ *
+ * The slug vocabulary is deliberately NOT a second naming scheme: entries are
+ * derived from `COMPARE_MODEL_SLUGS`, so `/inference/kimi-k3` and
+ * `/compare/kimi-k3-gb200-vs-mi355x` agree on what "kimi-k3" means. Aliases
+ * accept the compare aliases plus the lowercased `?g_model=` display names
+ * (e.g. `/inference/deepseek-v4-pro`), all 308-redirecting to the canonical
+ * slug so exactly one URL per model is ever indexed.
+ */
+import { COMPARE_MODEL_ALIASES, COMPARE_MODEL_SLUGS } from '@/lib/compare-slug';
+import { getModelCategory, Model } from '@/lib/data-mappings';
+
+export interface InferenceModelSlug {
+  /** Canonical URL slug, e.g. 'kimi-k3'. Lowercase, matches the compare slug. */
+  slug: string;
+  /** Model enum value — the same string `?g_model=` accepts. */
+  model: Model;
+  /** Human label for the page header and OG copy, e.g. 'Kimi K3 2.8T'. */
+  label: string;
+  /** Short, search-shaped model name for `<title>` tags, e.g. 'Kimi K3'. */
+  seoName: string;
+}
+
+const MODEL_VALUES = new Set<string>(Object.values(Model));
+
+/**
+ * Ordered like `COMPARE_MODEL_SLUGS` (flagship models first, per product
+ * spec). One entry per dashboard model option: compare slugs are finer-grained
+ * than the display-grouped dropdown (`kimi-k26` vs the retired `kimi-k25`
+ * alias), but each display name appears exactly once in the compare registry,
+ * so the mapping is 1:1. Hidden models get no page.
+ */
+export const INFERENCE_MODEL_SLUGS: readonly InferenceModelSlug[] = COMPARE_MODEL_SLUGS.filter(
+  (entry) => MODEL_VALUES.has(entry.displayName),
+)
+  .map((entry) => ({
+    slug: entry.slug,
+    model: entry.displayName as Model,
+    label: entry.label,
+    seoName: entry.seoName,
+  }))
+  .filter((entry) => getModelCategory(entry.model) !== 'hidden');
+
+const SLUG_TO_ENTRY: ReadonlyMap<string, InferenceModelSlug> = new Map(
+  INFERENCE_MODEL_SLUGS.map((entry) => [entry.slug, entry]),
+);
+
+const MODEL_TO_SLUG: ReadonlyMap<Model, string> = new Map(
+  INFERENCE_MODEL_SLUGS.map((entry) => [entry.model, entry.slug]),
+);
+
+/**
+ * Alias slugs that 308 to the canonical page. Two sources:
+ * - the compare aliases (family names and superseded versions), filtered to
+ *   targets that actually have an inference page;
+ * - the lowercased `Model` enum values, so a `?g_model=` value can be dropped
+ *   into the path verbatim (`/inference/Kimi-K2.5` → `/inference/kimi-k26`).
+ */
+export const INFERENCE_MODEL_ALIASES: Readonly<Record<string, string>> = {
+  ...Object.fromEntries(
+    Object.entries(COMPARE_MODEL_ALIASES).filter(([, target]) => SLUG_TO_ENTRY.has(target)),
+  ),
+  ...Object.fromEntries(
+    INFERENCE_MODEL_SLUGS.filter((entry) => entry.model.toLowerCase() !== entry.slug).map(
+      (entry) => [entry.model.toLowerCase(), entry.slug],
+    ),
+  ),
+};
+
+/**
+ * Resolve a raw path segment (any case, aliases allowed) to its registry
+ * entry. Returns null for unknown slugs — including the reserved static
+ * `/inference/*` segments (`agentic`, `logs`), which are not in the registry.
+ */
+export function getInferenceModelBySlug(rawSlug: string): InferenceModelSlug | null {
+  const lower = rawSlug.toLowerCase();
+  const canonical = INFERENCE_MODEL_ALIASES[lower] ?? lower;
+  return SLUG_TO_ENTRY.get(canonical) ?? null;
+}
+
+/** Canonical slug for a dashboard model, or null for hidden models. */
+export function inferenceModelSlugForModel(model: Model): string | null {
+  return MODEL_TO_SLUG.get(model) ?? null;
+}
+
+/** English path of a model page, e.g. '/inference/kimi-k3'. */
+export function inferenceModelPath(slug: string): string {
+  return `/inference/${slug}`;
+}
+
+/**
+ * Registry entries whose model is still actively benchmarked. Deprecated
+ * models keep their pages (historical data stays reachable and indexed) but
+ * are not promoted on the landing page.
+ */
+export const ACTIVE_INFERENCE_MODEL_SLUGS: readonly InferenceModelSlug[] =
+  INFERENCE_MODEL_SLUGS.filter((entry) => getModelCategory(entry.model) !== 'deprecated');
+
+/**
+ * Model pinned by the current pathname, or null when the path carries no
+ * model. Accepts the English and `/zh` trees and alias slugs (aliases redirect
+ * server-side, but the client provider may briefly render before the redirect
+ * resolves). Used by the dashboard shell to seed `GlobalFilterProvider` and by
+ * the provider's pathname effect on soft navigations; an explicit `?g_model=`
+ * param always wins over the path pin.
+ *
+ * Accepts null/undefined because `usePathname()` is typed to return null
+ * outside an app router (e.g. in unit tests that render the provider bare).
+ */
+export function inferenceModelForPathname(pathname: string | null | undefined): Model | null {
+  if (!pathname) return null;
+  const barePath = pathname.split(/[?#]/u, 1)[0] || '/';
+  const enPath = barePath === '/zh' ? '/' : barePath.replace(/^\/zh(?=\/)/u, '');
+  const match = /^\/inference\/(?<slug>[^/]+)\/?$/u.exec(enPath);
+  if (!match?.groups?.slug) return null;
+  let segment: string;
+  try {
+    segment = decodeURIComponent(match.groups.slug);
+  } catch {
+    return null;
+  }
+  return getInferenceModelBySlug(segment)?.model ?? null;
+}
+
+/**
+ * Pathname the inference dashboard should shallow-rewrite to when the user
+ * picks `model` from the selector, or null when the current path must be left
+ * alone. Only the inference tab itself qualifies — the base `/inference`
+ * page and the `/inference/<model>` subroutes (in both locale trees). Every
+ * other surface sharing the global model filter (overview, compare, the
+ * agentic catalog, …) keeps its own URL.
+ *
+ * Models without a registry page (hidden entries) fall back to the base
+ * `/inference` path rather than minting a URL that would 404 on reload.
+ */
+export function inferenceModelRouteForSelection(
+  pathname: string | null | undefined,
+  model: Model,
+): string | null {
+  if (!pathname) return null;
+  const barePath = pathname.split(/[?#]/u, 1)[0] || '/';
+  const isZh = barePath === '/zh' || barePath.startsWith('/zh/');
+  const enPath = isZh ? barePath.replace(/^\/zh(?=\/|$)/u, '') || '/' : barePath;
+  const onInferenceTab =
+    enPath === '/inference' ||
+    enPath === '/inference/' ||
+    inferenceModelForPathname(pathname) !== null;
+  if (!onInferenceTab) return null;
+  const slug = inferenceModelSlugForModel(model);
+  const target = slug ? inferenceModelPath(slug) : '/inference';
+  return isZh ? `/zh${target}` : target;
+}
diff --git a/packages/app/src/lib/model-architectures.test.ts b/packages/app/src/lib/model-architectures.test.ts
index efbf39e35..6cac00991 100644
--- a/packages/app/src/lib/model-architectures.test.ts
+++ b/packages/app/src/lib/model-architectures.test.ts
@@ -1,6 +1,7 @@
+import { MODEL_RELEASE_DATES, getModelReleaseDate } from '@semianalysisai/inferencex-constants';
 import { describe, expect, it } from 'vitest';
 
-import { Model } from '@/lib/data-mappings';
+import { MODEL_OPTIONS, Model } from '@/lib/data-mappings';
 
 import {
   type SubBlockFlow,
@@ -20,6 +21,48 @@ import {
   MODEL_ARCHITECTURES,
 } from './model-architectures';
 
+describe('MODEL_RELEASE_DATES coverage', () => {
+  it('has a sourced date for every model the user can select', () => {
+    // Without a date the Fleet Lifecycle axis falls back to the model's first
+    // benchmark run, which reads as "this is when the model appeared" when it is
+    // really "this is when we got to it" — for gpt-oss that gap is five weeks.
+    // A model whose release date genuinely cannot be sourced belongs in an
+    // explicit allowlist here, so the gap stays visible; do not delete the test.
+    const missing = MODEL_OPTIONS.filter((model) => !getModelReleaseDate(model));
+    expect(missing).toEqual([]);
+  });
+
+  it('anchors each model to its weights publication, not its first sweep', () => {
+    // Pinned so a silent revert fails. Every one of these predates the model's
+    // first InferenceX sweep, which is the invariant that caught the four wrong
+    // dates: a model cannot be benchmarked before its weights exist.
+    expect(getModelReleaseDate(Model.Kimi_K3)).toBe('2026-07-27');
+    expect(getModelReleaseDate(Model.GLM_5_2)).toBe('2026-06-13');
+    expect(getModelReleaseDate(Model.MiniMax_M3)).toBe('2026-06-07');
+    expect(getModelReleaseDate(Model.DeepSeek_V4_Pro)).toBe('2026-04-24');
+    expect(getModelReleaseDate(Model.GLM_5)).toBe('2026-02-13');
+    expect(getModelReleaseDate(Model.Qwen3_5)).toBe('2026-02-16');
+    expect(getModelReleaseDate(Model.MiniMax_M2_5)).toBe('2026-02-12');
+    expect(getModelReleaseDate(Model.Kimi_K2_5)).toBe('2026-01-27');
+    expect(getModelReleaseDate(Model.GptOss)).toBe('2025-08-05');
+    expect(getModelReleaseDate(Model.DeepSeek_R1)).toBe('2025-05-28');
+    expect(getModelReleaseDate(Model.Llama3_3_70B)).toBe('2024-12-06');
+    expect(getModelReleaseDate(Model.Llama3_1_70B)).toBe('2024-07-23');
+  });
+
+  it('names a real model, so no entry is dead weight', () => {
+    // The table is keyed by string because it lives in the constants package,
+    // below the `Model` enum. A key that matches no model never throws — the
+    // lookup returns null and every caller quietly falls back to its own
+    // default, so a typo here is invisible until someone notices the axis is
+    // anchored to the wrong date. This is the only place both are in scope.
+    const displayNames = new Set<string>(Object.values(Model));
+    for (const model of Object.keys(MODEL_RELEASE_DATES)) {
+      expect(displayNames.has(model), `${model} matches no Model display name`).toBe(true);
+    }
+  });
+});
+
 describe('MODEL_ARCHITECTURES', () => {
   it('has architecture data for all supported models', () => {
     const models = [
@@ -42,6 +85,28 @@ describe('MODEL_ARCHITECTURES', () => {
     }
   });
 
+  it('carries no release date of its own, so there is nothing to diverge from', () => {
+    // This is the guard on the refactor, not a style rule. When each entry kept
+    // its own `releaseDate`, this file and MODEL_RELEASE_DATES disagreed about
+    // DeepSeek-V4-Pro — one of them had it releasing six weeks after its own
+    // first benchmark run — and both were rendered to users. Re-adding the field
+    // here recreates that, so the test fails on the field itself rather than
+    // waiting for the values to drift.
+    for (const [model, arch] of Object.entries(MODEL_ARCHITECTURES)) {
+      expect(arch, model).toBeDefined();
+      expect(Object.keys(arch!), model).not.toContain('releaseDate');
+    }
+  });
+
+  it('sources every architecture caption from the one release-date table', () => {
+    // The diagram prints "Released by {developer} on {date}", so an entry with a
+    // developer and no date silently drops the whole caption.
+    const missing = Object.values(MODEL_ARCHITECTURES)
+      .filter((arch) => arch?.developer && !getModelReleaseDate(arch.model))
+      .map((arch) => arch!.model);
+    expect(missing).toEqual([]);
+  });
+
   it('ensures dense models have equal active and total params', () => {
     for (const arch of Object.values(MODEL_ARCHITECTURES)) {
       if (!arch) continue;
@@ -143,7 +208,9 @@ describe('getModelArchitecture', () => {
     expect(arch?.numKVHeads).toBe(8);
     expect(arch?.ffnDim).toBe(28672);
     expect(arch?.vocabSize).toBe(128256);
-    expect(arch?.releaseDate).toBe('2024-07-23');
+    // The release date is no longer carried here; it lives once in
+    // MODEL_RELEASE_DATES and is asserted in the constants package.
+    expect(getModelReleaseDate(Model.Llama3_1_70B)).toBe('2024-07-23');
   });
 
   it('returns architecture for DeepSeek R1 with MoE and MLA details', () => {
diff --git a/packages/app/src/lib/model-architectures.ts b/packages/app/src/lib/model-architectures.ts
index 43bd3022e..f30393bb2 100644
--- a/packages/app/src/lib/model-architectures.ts
+++ b/packages/app/src/lib/model-architectures.ts
@@ -90,8 +90,12 @@ export interface ModelArchitecture {
   contextWindow?: number;
   /** Special architectural features */
   features?: string[];
-  /** Release date (YYYY-MM-DD) */
-  releaseDate?: string;
+  // No `releaseDate` here, deliberately. A model's release date is the same fact
+  // the Fleet Lifecycle time axis anchors on, so it lives once in
+  // `MODEL_RELEASE_DATES` (`@semianalysisai/inferencex-constants`) and is read
+  // through `getModelReleaseDate(arch.model)`. While this file kept its own copy
+  // the two disagreed: DeepSeek-V4-Pro was listed here as releasing six weeks
+  // after its own first benchmark run, and both values were rendered to users.
   /** Developer/Organization */
   developer?: string;
   /** Link to model card or paper */
@@ -171,7 +175,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
       'Auxiliary-loss-free Load Balancing',
       'Multi-Token Prediction',
     ],
-    releaseDate: '2025-05-28',
     developer: 'DeepSeek',
     sourceUrl: 'https://huggingface.co/deepseek-ai/DeepSeek-R1-0528',
   },
@@ -237,7 +240,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
       'FP4 Experts + FP8 Mixed Precision',
       'Muon Optimizer',
     ],
-    releaseDate: '2026-06-08',
     developer: 'DeepSeek',
     sourceUrl: 'https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro',
   },
@@ -255,7 +257,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
     ffnDim: 28672,
     contextWindow: 128000,
     features: ['Grouped Query Attention', 'RoPE'],
-    releaseDate: '2024-12-06',
     developer: 'Meta',
     sourceUrl: 'https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct',
   },
@@ -273,7 +274,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
     ffnDim: 28672,
     contextWindow: 128000,
     features: ['Grouped Query Attention', 'RoPE'],
-    releaseDate: '2024-07-23',
     developer: 'Meta',
     sourceUrl: 'https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct',
   },
@@ -316,7 +316,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
       'YaRN RoPE (factor=32)',
       'MXFP4 Quantization',
     ],
-    releaseDate: '2025-06-13',
     developer: 'OpenAI',
     sourceUrl: 'https://huggingface.co/openai/gpt-oss-120b',
   },
@@ -338,7 +337,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
     denseFFNDim: 18432,
     contextWindow: 262144,
     features: ['Multi-head Latent Attention', 'DeepSeek-style MoE', 'YaRN RoPE'],
-    releaseDate: '2026-01-27',
     developer: 'Moonshot AI',
     sourceUrl: 'https://huggingface.co/moonshotai/Kimi-K2.5',
   },
@@ -407,7 +405,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
       'Native Multimodality (text/image/video)',
       'MXFP4 Quantization',
     ],
-    releaseDate: '2026-06-13',
     developer: 'Moonshot AI',
     sourceUrl: 'https://huggingface.co/moonshotai/Kimi-K3',
   },
@@ -435,7 +432,6 @@ export const MODEL_ARCHITECTURES: Partial<Record<Model, ModelArchitecture>> = {
       'Multi-Token Prediction (3 modules)',
       'FP8 Quantization',
     ],
-    releaseDate: '2025-10-25',
     developer: 'MiniMax',
     sourceUrl: 'https://huggingface.co/MiniMaxAI/MiniMax-M2',
   },
diff --git a/packages/app/src/lib/model-logos.test.ts b/packages/app/src/lib/model-logos.test.ts
new file mode 100644
index 000000000..7e984cc2a
--- /dev/null
+++ b/packages/app/src/lib/model-logos.test.ts
@@ -0,0 +1,75 @@
+import fs from 'node:fs';
+import path from 'node:path';
+
+import matter from 'gray-matter';
+import { describe, expect, it } from 'vitest';
+
+import { Model, getModelLogo } from '@/lib/data-mappings';
+import { MODEL_DEVELOPER_LOGOS, getModelDeveloperLogo, isMonochromeLogo } from '@/lib/model-logos';
+
+const MODELS_DIR = path.join(process.cwd(), 'content', 'models');
+const LOGOS_DIR = path.join(process.cwd(), 'public', 'logos');
+
+describe('model developer logos', () => {
+  it('maps every model-page developer to a logo', () => {
+    const developers = new Set(
+      fs
+        .readdirSync(MODELS_DIR)
+        .filter((f) => f.endsWith('.mdx'))
+        .map((f) => {
+          const { data } = matter(fs.readFileSync(path.join(MODELS_DIR, f), 'utf8'));
+          return (data as { developer: string }).developer;
+        }),
+    );
+    for (const developer of developers) {
+      expect(
+        getModelDeveloperLogo(developer),
+        `missing logo mapping for '${developer}'`,
+      ).toBeTruthy();
+    }
+  });
+
+  it('points every developer mapping at an existing file under public/logos', () => {
+    for (const [developer, logo] of Object.entries(MODEL_DEVELOPER_LOGOS)) {
+      expect(
+        fs.existsSync(path.join(LOGOS_DIR, logo)),
+        `logo file '${logo}' for '${developer}' not found in public/logos`,
+      ).toBe(true);
+    }
+  });
+
+  it('points every MODEL_CONFIG logo at an existing file under public/logos', () => {
+    for (const model of Object.values(Model)) {
+      const logo = getModelLogo(model);
+      if (!logo) continue;
+      expect(
+        fs.existsSync(path.join(LOGOS_DIR, logo)),
+        `logo file '${logo}' for model '${model}' not found in public/logos`,
+      ).toBe(true);
+    }
+  });
+
+  it('returns undefined for unknown developers', () => {
+    expect(getModelDeveloperLogo('Unknown Lab')).toBeUndefined();
+  });
+
+  it('uses the actual Z.ai mark for GLM models', () => {
+    expect(getModelLogo(Model.GLM_5)).toBe('zai-color.svg');
+    expect(getModelLogo(Model.GLM_5_2)).toBe('zai-color.svg');
+    expect(getModelDeveloperLogo('Z.ai (Zhipu AI)')).toBe('zai-color.svg');
+  });
+
+  it('keeps every model logo full-color except monochrome-by-design marks', () => {
+    // `currentColor` fills are the monochrome giveaway: a full-color mark
+    // ships its own brand fills and never inherits the text color.
+    for (const model of Object.values(Model)) {
+      const logo = getModelLogo(model);
+      if (!logo || !logo.endsWith('.svg') || isMonochromeLogo(logo)) continue;
+      const svg = fs.readFileSync(path.join(LOGOS_DIR, logo), 'utf8');
+      expect(
+        svg.includes('currentColor'),
+        `logo '${logo}' for model '${model}' is not full-color`,
+      ).toBe(false);
+    }
+  });
+});
diff --git a/packages/app/src/lib/model-logos.ts b/packages/app/src/lib/model-logos.ts
new file mode 100644
index 000000000..feeac9321
--- /dev/null
+++ b/packages/app/src/lib/model-logos.ts
@@ -0,0 +1,46 @@
+/**
+ * Logo files under `public/logos/` whose brand mark is plain black
+ * (`currentColor` SVGs): they need a dark-mode invert to stay visible.
+ * Full-color logos render as-is in both themes. Shared by every surface that
+ * renders model/developer or hardware-vendor logos (`/model` pages, inference
+ * chart captions, `/compare` pair cards).
+ *
+ * `amd.svg` is here because AMD's brand guidelines only permit the standard
+ * black logo or the reversed-out white logo — no color variant exists, so the
+ * dark-mode invert reproduces the official reversed treatment.
+ */
+export const MONOCHROME_LOGO_FILES: ReadonlySet<string> = new Set(['amd.svg', 'openai.svg']);
+
+/** Whether a logo file needs a dark-mode invert to stay visible. */
+export function isMonochromeLogo(file: string): boolean {
+  return MONOCHROME_LOGO_FILES.has(file);
+}
+
+/**
+ * Maps `/model/[slug]` frontmatter `developer` names to full-color logo files
+ * under `public/logos/`. Keys must match the `developer` field in
+ * `content/models/<slug>.mdx` exactly; a missing entry simply renders no logo
+ * (the pages treat the logo as optional decoration).
+ */
+export const MODEL_DEVELOPER_LOGOS: Record<string, string> = {
+  'Alibaba (Qwen)': 'qwen-color.svg',
+  DeepSeek: 'deepseek-color.svg',
+  Meta: 'meta-color.svg',
+  MiniMax: 'minimax-color.svg',
+  // Moonshot AI's model pages are all Kimi models; the Kimi product mark is
+  // the recognizable brand. `kimi-color.svg` sets the white-and-blue mark on
+  // its brand black rounded tile (the Kimi app icon), so it stays full-color
+  // AND visible on light backgrounds — unlike the bare color variant.
+  'Moonshot AI': 'kimi-color.svg',
+  // OpenAI's brand mark is monochrome by design — no color variant exists.
+  OpenAI: 'openai.svg',
+  // The Z.ai brand mark is monochrome by design — no color variant exists.
+  // `zai-color.svg` sets the white Z glyph on its brand black rounded tile
+  // (the Z.ai app icon), so it stays theme-independent like the Kimi mark.
+  'Z.ai (Zhipu AI)': 'zai-color.svg',
+};
+
+/** Logo filename under `/logos/` for a model developer, if one exists. */
+export function getModelDeveloperLogo(developer: string): string | undefined {
+  return MODEL_DEVELOPER_LOGOS[developer];
+}
diff --git a/packages/app/src/lib/model-pages.ts b/packages/app/src/lib/model-pages.ts
new file mode 100644
index 000000000..8d4d1cbe2
--- /dev/null
+++ b/packages/app/src/lib/model-pages.ts
@@ -0,0 +1,52 @@
+import fs from 'node:fs';
+import path from 'node:path';
+
+import matter from 'gray-matter';
+
+import { COMPARE_MODEL_SLUGS, type CompareModelSlug } from '@/lib/compare-slug';
+
+/**
+ * Frontmatter for `/model/[slug]` deep-dive pages. Content lives in
+ * `content/models/<slug>.mdx`, one file per canonical compare slug — the body
+ * is standard MDX (Overview, Architecture, Official vendor eval scores,
+ * Benchmark explanations) compiled with the same pipeline as blog posts.
+ */
+export interface ModelPageFrontmatter {
+  /** Page H1 / SEO title, e.g. 'Kimi K3'. */
+  title: string;
+  /** Model developer shown in the fact strip, e.g. 'Moonshot AI'. */
+  developer: string;
+  /** Human-readable release date(s); free-form to allow multi-version slugs. */
+  releaseDate: string;
+  /** Meta description and page subtitle. Keep ≤160 chars. */
+  description: string;
+}
+
+export interface ModelPage {
+  meta: ModelPageFrontmatter;
+  entry: CompareModelSlug;
+  raw: string;
+}
+
+const MODELS_DIR = path.join(process.cwd(), 'content', 'models');
+
+/**
+ * Canonical model-page slugs in the master compare-index order, filtered to
+ * slugs that actually have a content file. Drives `generateStaticParams` and
+ * the sitemap so a slug without content never produces a route.
+ */
+export function getModelPageSlugs(): string[] {
+  return COMPARE_MODEL_SLUGS.filter((m) =>
+    fs.existsSync(path.join(MODELS_DIR, `${m.slug}.mdx`)),
+  ).map((m) => m.slug);
+}
+
+/** Loads one model page. Returns null for unknown slugs or missing content. */
+export function getModelPage(slug: string): ModelPage | null {
+  const entry = COMPARE_MODEL_SLUGS.find((m) => m.slug === slug);
+  if (!entry) return null;
+  const filePath = path.join(MODELS_DIR, `${slug}.mdx`);
+  if (!fs.existsSync(filePath)) return null;
+  const { data, content } = matter(fs.readFileSync(filePath, 'utf8'));
+  return { meta: data as ModelPageFrontmatter, entry, raw: content };
+}
diff --git a/packages/app/src/lib/model-routes.test.ts b/packages/app/src/lib/model-routes.test.ts
new file mode 100644
index 000000000..48b649c32
--- /dev/null
+++ b/packages/app/src/lib/model-routes.test.ts
@@ -0,0 +1,172 @@
+import { describe, expect, it } from 'vitest';
+
+import { Model, MODEL_OPTIONS } from './data-mappings';
+import {
+  DEFAULT_ROUTE_MODEL,
+  MODEL_ROUTE_TABS,
+  MODEL_ROUTES,
+  modelRoutePath,
+  modelRoutePathnameRewrite,
+  pathWithSearchParams,
+  modelRouteSlug,
+  parseModelRoutePathname,
+  resolveModelRouteSlug,
+  routeModelForPathname,
+} from './model-routes';
+import { PARAM_DEFAULTS } from './url-state';
+
+describe('MODEL_ROUTES registry', () => {
+  it('covers every visible model with a unique canonical slug', () => {
+    expect(MODEL_ROUTES.map((route) => route.model)).toEqual([...MODEL_OPTIONS]);
+    const slugs = MODEL_ROUTES.map((route) => route.slug);
+    expect(new Set(slugs).size).toBe(slugs.length);
+  });
+
+  it('uses lowercase hyphenated URL-safe slugs', () => {
+    for (const route of MODEL_ROUTES) {
+      expect(route.slug).toMatch(/^[a-z0-9]+(?:-[a-z0-9]+)*$/u);
+    }
+  });
+
+  it('keeps the default route model in lockstep with the g_model URL default', () => {
+    // The pathname rewrite treats a bare /calculator or /historical as
+    // DEFAULT_ROUTE_MODEL; if the app-wide default drifted, plain page loads
+    // would rewrite the URL.
+    expect(PARAM_DEFAULTS.g_model).toBe(DEFAULT_ROUTE_MODEL);
+  });
+});
+
+describe('resolveModelRouteSlug', () => {
+  it('resolves canonical slugs without flagging an alias', () => {
+    const resolved = resolveModelRouteSlug('kimi-k3');
+    expect(resolved?.route.model).toBe(Model.Kimi_K3);
+    expect(resolved?.isAlias).toBe(false);
+  });
+
+  it('resolves compare aliases and flags them for redirect', () => {
+    const resolved = resolveModelRouteSlug('kimi');
+    expect(resolved?.route.slug).toBe('kimi-k26');
+    expect(resolved?.isAlias).toBe(true);
+  });
+
+  it('is case-insensitive but flags non-canonical casing', () => {
+    const resolved = resolveModelRouteSlug('Kimi-K3');
+    expect(resolved?.route.slug).toBe('kimi-k3');
+    expect(resolved?.isAlias).toBe(true);
+  });
+
+  it('returns null for unknown slugs', () => {
+    expect(resolveModelRouteSlug('not-a-model')).toBeNull();
+  });
+});
+
+describe('parseModelRoutePathname', () => {
+  it('parses bare and slugged English tab paths', () => {
+    expect(parseModelRoutePathname('/calculator')).toEqual({
+      tab: 'calculator',
+      zh: false,
+      slug: null,
+    });
+    expect(parseModelRoutePathname('/historical/kimi-k3')).toEqual({
+      tab: 'historical',
+      zh: false,
+      slug: 'kimi-k3',
+    });
+  });
+
+  it('parses /zh siblings', () => {
+    expect(parseModelRoutePathname('/zh/calculator/deepseek-r1')).toEqual({
+      tab: 'calculator',
+      zh: true,
+      slug: 'deepseek-r1',
+    });
+    expect(parseModelRoutePathname('/zh/historical')).toEqual({
+      tab: 'historical',
+      zh: true,
+      slug: null,
+    });
+  });
+
+  it('rejects other routes and deeper paths', () => {
+    expect(parseModelRoutePathname('/')).toBeNull();
+    expect(parseModelRoutePathname('/inference')).toBeNull();
+    expect(parseModelRoutePathname('/compare/kimi-k3-b200-vs-mi355x')).toBeNull();
+    expect(parseModelRoutePathname('/historical/kimi-k3/extra')).toBeNull();
+    expect(parseModelRoutePathname('/calculatorx')).toBeNull();
+  });
+});
+
+describe('routeModelForPathname', () => {
+  it('maps slugged paths (including aliases) to models', () => {
+    expect(routeModelForPathname('/historical/kimi-k3')).toBe(Model.Kimi_K3);
+    expect(routeModelForPathname('/zh/calculator/deepseek-r1')).toBe(Model.DeepSeek_R1);
+    expect(routeModelForPathname('/calculator/glm')).toBe(Model.GLM_5);
+  });
+
+  it('returns null for bare tabs, unknown slugs, and non-model routes', () => {
+    expect(routeModelForPathname('/historical')).toBeNull();
+    expect(routeModelForPathname('/historical/not-a-model')).toBeNull();
+    expect(routeModelForPathname('/inference')).toBeNull();
+    expect(routeModelForPathname(null)).toBeNull();
+    expect(routeModelForPathname(undefined)).toBeNull();
+  });
+});
+
+describe('modelRoutePathnameRewrite', () => {
+  it('rewrites slugged paths when the model changes', () => {
+    expect(modelRoutePathnameRewrite('/historical/kimi-k3', Model.DeepSeek_R1)).toBe(
+      '/historical/deepseek-r1',
+    );
+    expect(modelRoutePathnameRewrite('/zh/calculator/kimi-k3', Model.MiniMax_M3)).toBe(
+      '/zh/calculator/minimax-m3',
+    );
+  });
+
+  it('rewrites bare paths only for non-default models', () => {
+    expect(modelRoutePathnameRewrite('/calculator', DEFAULT_ROUTE_MODEL)).toBeNull();
+    expect(modelRoutePathnameRewrite('/calculator', Model.Kimi_K3)).toBe('/calculator/kimi-k3');
+    expect(modelRoutePathnameRewrite('/zh/historical', Model.Kimi_K3)).toBe(
+      '/zh/historical/kimi-k3',
+    );
+  });
+
+  it('keeps the slugged path when switching to the default model', () => {
+    expect(modelRoutePathnameRewrite('/historical/kimi-k3', DEFAULT_ROUTE_MODEL)).toBe(
+      `/historical/${modelRouteSlug(DEFAULT_ROUTE_MODEL)}`,
+    );
+  });
+
+  it('returns null when the path already matches', () => {
+    expect(modelRoutePathnameRewrite('/historical/kimi-k3', Model.Kimi_K3)).toBeNull();
+  });
+
+  it('never rewrites non-model routes or unknown slugs', () => {
+    expect(modelRoutePathnameRewrite('/inference', Model.Kimi_K3)).toBeNull();
+    expect(modelRoutePathnameRewrite('/', Model.Kimi_K3)).toBeNull();
+    expect(modelRoutePathnameRewrite('/historical/not-a-model', Model.Kimi_K3)).toBeNull();
+  });
+});
+
+describe('modelRoutePath', () => {
+  it('builds canonical English paths for both tabs', () => {
+    for (const tab of MODEL_ROUTE_TABS) {
+      expect(modelRoutePath(tab, 'kimi-k3')).toBe(`/${tab}/kimi-k3`);
+    }
+  });
+});
+
+describe('pathWithSearchParams', () => {
+  it('returns the bare path when there are no params', () => {
+    expect(pathWithSearchParams('/calculator/kimi-k26', {})).toBe('/calculator/kimi-k26');
+  });
+
+  it('keeps share-link params, including repeated keys and reserved characters', () => {
+    expect(
+      pathWithSearchParams('/calculator/kimi-k26', {
+        i_seq: '8k/1k',
+        unofficialruns: ['a', 'b'],
+        missing: undefined,
+      }),
+    ).toBe('/calculator/kimi-k26?i_seq=8k%2F1k&unofficialruns=a&unofficialruns=b');
+  });
+});
diff --git a/packages/app/src/lib/model-routes.ts b/packages/app/src/lib/model-routes.ts
new file mode 100644
index 000000000..4df6e3bc6
--- /dev/null
+++ b/packages/app/src/lib/model-routes.ts
@@ -0,0 +1,188 @@
+import {
+  COMPARE_MODEL_ALIASES,
+  COMPARE_MODEL_SLUGS,
+  type CompareModelSlug,
+} from '@/lib/compare-slug';
+import { Model, MODEL_OPTIONS } from '@/lib/data-mappings';
+
+// ---------------------------------------------------------------------------
+// Per-model dashboard routes: /calculator/<model> and /historical/<model>
+// ---------------------------------------------------------------------------
+//
+// The bare /calculator and /historical tabs render the DEFAULT model; the
+// `[model]` child routes render the same page seeded to a specific model so
+// every model gets its own indexable URL (title, description, canonical, and
+// hreflang all model-specific). Slugs deliberately REUSE the compare-page
+// model registry (`COMPARE_MODEL_SLUGS`) so `kimi-k3` means the same model in
+// `/compare/kimi-k3-b200-vs-mi355x`, `/calculator/kimi-k3`, and
+// `/historical/kimi-k3` — one slug vocabulary across the whole site.
+//
+// Unlike compare (whose slugs are finer-grained than the display dropdown,
+// e.g. distinct kimi-k25/k26 DB buckets), these routes address the dashboard's
+// display-grouped model selector, so exactly one canonical slug exists per
+// visible `Model` option. Aliases (`kimi`, `glm-5`, …) 308-redirect to the
+// canonical slug, mirroring compare behavior.
+
+/** Dashboard tabs that expose per-model child routes. */
+export const MODEL_ROUTE_TABS = ['calculator', 'historical'] as const;
+export type ModelRouteTab = (typeof MODEL_ROUTE_TABS)[number];
+
+const MODEL_ROUTE_TAB_PATHS: Record<ModelRouteTab, `/${string}`> = {
+  calculator: '/calculator',
+  historical: '/historical',
+};
+
+/**
+ * Model the bare tab route renders when no slug (and no `?g_model=`) is
+ * present. Must stay in lockstep with `PARAM_DEFAULTS.g_model` in
+ * `url-state.ts` and the `GlobalFilterProvider` initializer — the pathname
+ * rewrite below treats a bare path as this model, so drift would rewrite the
+ * URL on plain page loads. `model-routes.test.ts` pins the invariant.
+ */
+export const DEFAULT_ROUTE_MODEL: Model = Model.DeepSeek_V4_Pro;
+
+export interface ModelRoute {
+  /** Canonical URL segment, e.g. 'kimi-k3'. Lowercase, hyphen-separated. */
+  slug: string;
+  /** Display-grouped dashboard model this route seeds. */
+  model: Model;
+  /** Natural model name for SEO titles/descriptions, e.g. 'Kimi K3'. */
+  seoName: string;
+  /** Human label for headers, e.g. 'Kimi K3 2.8T'. */
+  label: string;
+}
+
+function compareEntryForModel(model: Model): CompareModelSlug | undefined {
+  return COMPARE_MODEL_SLUGS.find((entry) => entry.displayName === model);
+}
+
+/**
+ * One route per visible dashboard model. Derived (not hand-copied) from the
+ * compare registry so a new model added there automatically gets calculator
+ * and historical routes; the test suite fails if a `MODEL_OPTIONS` entry has
+ * no compare slug to derive from.
+ */
+export const MODEL_ROUTES: ModelRoute[] = MODEL_OPTIONS.flatMap((model) => {
+  const entry = compareEntryForModel(model);
+  if (!entry) return [];
+  return [{ slug: entry.slug, model, seoName: entry.seoName, label: entry.label }];
+});
+
+const ROUTE_BY_SLUG: ReadonlyMap<string, ModelRoute> = new Map(
+  MODEL_ROUTES.map((route) => [route.slug, route]),
+);
+const ROUTE_BY_MODEL: ReadonlyMap<Model, ModelRoute> = new Map(
+  MODEL_ROUTES.map((route) => [route.model, route]),
+);
+
+export interface ResolvedModelRoute {
+  route: ModelRoute;
+  /** True when the input was an alias or non-canonical casing — caller should
+   *  308 to the canonical slug. */
+  isAlias: boolean;
+}
+
+/** Resolve a URL segment (canonical slug, alias, or mixed case) to a model
+ *  route. Returns null for unknown slugs. */
+export function resolveModelRouteSlug(slug: string): ResolvedModelRoute | null {
+  const lower = slug.toLowerCase();
+  const canonical = COMPARE_MODEL_ALIASES[lower] ?? lower;
+  const route = ROUTE_BY_SLUG.get(canonical);
+  if (!route) return null;
+  return { route, isAlias: canonical !== slug };
+}
+
+/** Canonical slug for a model, or null for models without a route (hidden). */
+export function modelRouteSlug(model: Model): string | null {
+  return ROUTE_BY_MODEL.get(model)?.slug ?? null;
+}
+
+/** Canonical English path for a per-model tab page, e.g. '/historical/kimi-k3'. */
+export function modelRoutePath(tab: ModelRouteTab, slug: string): string {
+  return `${MODEL_ROUTE_TAB_PATHS[tab]}/${slug}`;
+}
+
+/**
+ * Reattach incoming search params when 308ing an alias slug to its canonical
+ * path, so share-link state (g_rundate, i_seq, c_price, unofficialruns, …)
+ * survives the redirect.
+ */
+export function pathWithSearchParams(
+  path: string,
+  sp: Record<string, string | string[] | undefined>,
+): string {
+  const params = new URLSearchParams();
+  for (const [key, value] of Object.entries(sp)) {
+    if (typeof value === 'string') params.append(key, value);
+    else if (Array.isArray(value)) for (const item of value) params.append(key, item);
+  }
+  const search = params.toString();
+  return search ? `${path}?${search}` : path;
+}
+
+export interface ParsedModelRoutePathname {
+  tab: ModelRouteTab;
+  /** True for /zh-prefixed pathnames. */
+  zh: boolean;
+  /** URL segment after the tab path, or null on the bare tab route. */
+  slug: string | null;
+}
+
+/**
+ * Parse an English or /zh pathname into its model-routed tab, if any.
+ * `/calculator` → slug null; `/zh/historical/kimi-k3` → slug 'kimi-k3';
+ * anything else (other tabs, deeper children) → null.
+ */
+export function parseModelRoutePathname(pathname: string): ParsedModelRoutePathname | null {
+  const barePath = pathname.split(/[?#]/u, 1)[0] || '/';
+  const zh = barePath === '/zh' || barePath.startsWith('/zh/');
+  const enPath = zh ? barePath.slice('/zh'.length) || '/' : barePath;
+  for (const tab of MODEL_ROUTE_TABS) {
+    const tabPath = MODEL_ROUTE_TAB_PATHS[tab];
+    if (enPath === tabPath) return { tab, zh, slug: null };
+    if (enPath.startsWith(`${tabPath}/`)) {
+      const rest = enPath.slice(tabPath.length + 1);
+      if (!rest || rest.includes('/')) return null;
+      return { tab, zh, slug: rest };
+    }
+  }
+  return null;
+}
+
+/**
+ * Model implied by a model-routed pathname, or null when the pathname is not
+ * a model-routed page or carries no (known) slug. Used to seed
+ * `GlobalFilterProvider` so `/historical/kimi-k3` server-renders and hydrates
+ * on Kimi K3 without waiting for a client effect.
+ */
+export function routeModelForPathname(pathname: string | null | undefined): Model | null {
+  if (!pathname) return null;
+  const parsed = parseModelRoutePathname(pathname);
+  if (!parsed?.slug) return null;
+  return resolveModelRouteSlug(parsed.slug)?.route.model ?? null;
+}
+
+/**
+ * Pathname the address bar should show for `model` while on a model-routed
+ * page, or null when no rewrite is needed. Pure so the model-switch URL
+ * behavior is unit-testable; the caller performs the actual
+ * `history.replaceState`.
+ *
+ * Rules:
+ * - Non-model-routed pages, unknown slugs, and slug-less models → null.
+ * - Bare tab path showing the default model → stays bare (plain page loads
+ *   must not rewrite the URL).
+ * - Any other mismatch (model switched, alias slug, legacy `?g_model=` link
+ *   resolved by the provider) → the canonical slugged path.
+ */
+export function modelRoutePathnameRewrite(pathname: string, model: Model): string | null {
+  const parsed = parseModelRoutePathname(pathname);
+  if (!parsed) return null;
+  if (parsed.slug && !resolveModelRouteSlug(parsed.slug)) return null;
+  const slug = modelRouteSlug(model);
+  if (!slug) return null;
+  if (parsed.slug === slug) return null;
+  if (!parsed.slug && model === DEFAULT_ROUTE_MODEL) return null;
+  const enPath = modelRoutePath(parsed.tab, slug);
+  return parsed.zh ? `/zh${enPath}` : enPath;
+}
diff --git a/packages/app/src/lib/nudges/agentic-point-coach-mark.test.ts b/packages/app/src/lib/nudges/agentic-point-coach-mark.test.ts
new file mode 100644
index 000000000..8eceb3f2f
--- /dev/null
+++ b/packages/app/src/lib/nudges/agentic-point-coach-mark.test.ts
@@ -0,0 +1,371 @@
+// @vitest-environment jsdom
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+import {
+  AGENTIC_POINT_ACTION_SELECTOR,
+  AGENTIC_POINT_SELECTOR,
+  getAgenticPointAnchorMutationRoot,
+  getAgenticPointAnchorRect,
+  resolveAgenticPointAnchor,
+} from './agentic-point-coach-mark';
+
+// jsdom gives every element a zero-size box, so the resolver's on-screen check
+// would reject everything. Stub the geometry each test cares about instead.
+function stubRect(element: Element, left: number, top: number, size = 12): void {
+  element.getBoundingClientRect = () =>
+    ({
+      left,
+      top,
+      width: size,
+      height: size,
+      right: left + size,
+      bottom: top + size,
+      x: left,
+      y: top,
+      toJSON: () => ({}),
+    }) as DOMRect;
+}
+
+interface PointSpec {
+  left: number;
+  top: number;
+  benchmarkType?: string;
+  hasTrace?: boolean;
+  traceAvailability?: 'pending' | 'resolved';
+  hidden?: boolean;
+  overlay?: boolean;
+  /** Add the marker child the real chart renders inside every `.dot-group`. */
+  withShape?: boolean;
+}
+
+function stubBox(element: Element, left: number, top: number, width: number, height: number) {
+  element.getBoundingClientRect = () =>
+    ({
+      left,
+      top,
+      width,
+      height,
+      right: left + width,
+      bottom: top + height,
+      x: left,
+      y: top,
+      toJSON: () => ({}),
+    }) as DOMRect;
+}
+
+const SVG_NS = 'http://www.w3.org/2000/svg';
+
+interface ChartSvgSpec {
+  /** The `<svg>` box — includes the axis gutters, so bigger than the plot. */
+  box: { left: number; top: number; width: number; height: number };
+  /** `.chart-root` translate, i.e. the left/top chart margins. */
+  margin: { left: number; top: number };
+  /** The clipPath rect `setupChart` sizes to the plot area. */
+  clip: { width: number; height: number };
+}
+
+/**
+ * The chart SVG skeleton `setupChart` builds: a root group translated by the
+ * margins plus a clip rect covering the plot area only. Built for real (rather
+ * than stubbing a single box) because the gap between the SVG box and the clip
+ * region is exactly what the clipping tests are about.
+ */
+function appendChartSvg(chart: Element, spec: ChartSvgSpec): void {
+  const svg = document.createElementNS(SVG_NS, 'svg');
+  svg.dataset.testid = 'd3-chart-svg';
+  stubBox(svg, spec.box.left, spec.box.top, spec.box.width, spec.box.height);
+
+  const clipId = 'clip-scatter-graph';
+  const defs = document.createElementNS(SVG_NS, 'defs');
+  const clipPath = document.createElementNS(SVG_NS, 'clipPath');
+  clipPath.setAttribute('id', clipId);
+  const clipRect = document.createElementNS(SVG_NS, 'rect');
+  clipRect.setAttribute('width', String(spec.clip.width));
+  clipRect.setAttribute('height', String(spec.clip.height));
+  clipPath.append(clipRect);
+  defs.append(clipPath);
+  svg.append(defs);
+
+  const root = document.createElementNS(SVG_NS, 'g');
+  root.setAttribute('class', 'chart-root');
+  root.setAttribute('transform', `translate(${spec.margin.left},${spec.margin.top})`);
+  const zoomGroup = document.createElementNS(SVG_NS, 'g');
+  zoomGroup.setAttribute('class', 'zoom-group');
+  zoomGroup.setAttribute('clip-path', `url(#${clipId})`);
+  root.append(zoomGroup);
+  svg.append(root);
+
+  chart.append(svg);
+}
+
+/**
+ * Build a 1000x600 chart container with the given points and return them in
+ * order. When `svgSpec` is set, a real chart SVG skeleton is added so the
+ * resolver can work out the clip region.
+ */
+function renderChart(points: PointSpec[], svgSpec?: ChartSvgSpec): Element[] {
+  document.body.innerHTML =
+    '<div data-coach-mark-root><div data-testid="scatter-graph"></div></div>';
+  const chart = document.querySelector('[data-testid="scatter-graph"]')!;
+  stubBox(chart, 0, 0, 1000, 600);
+  if (svgSpec) appendChartSvg(chart, svgSpec);
+
+  return points.map((spec) => {
+    const element = document.createElement('div');
+    element.className = spec.overlay ? 'unofficial-overlay-pt' : 'dot-group';
+    if (!spec.overlay) {
+      element.dataset.benchmarkType = spec.benchmarkType ?? 'agentic_traces';
+    }
+    if (spec.hasTrace) element.dataset.hasTrace = 'true';
+    if (spec.traceAvailability) {
+      element.dataset.traceAvailability = spec.traceAvailability;
+    }
+    if (spec.hidden) element.style.opacity = '0';
+    if (spec.withShape) {
+      const shape = document.createElement('div');
+      shape.className = 'visible-shape';
+      element.append(shape);
+      stubRect(shape, spec.left + 2, spec.top + 2, 8);
+    }
+    chart.append(element);
+    stubRect(element, spec.left, spec.top);
+    return element;
+  });
+}
+
+afterEach(() => {
+  document.body.innerHTML = '';
+  vi.restoreAllMocks();
+});
+
+describe('AGENTIC_POINT_SELECTOR', () => {
+  it('matches only official agentic points inside the scatter chart', () => {
+    renderChart([
+      { left: 100, top: 100 },
+      { left: 200, top: 100, benchmarkType: 'single_turn' },
+      { left: 300, top: 100, overlay: true },
+    ]);
+    const matched = [...document.querySelectorAll<HTMLElement>(AGENTIC_POINT_SELECTOR)];
+
+    expect(matched).toHaveLength(1);
+    expect(matched[0].dataset.benchmarkType).toBe('agentic_traces');
+  });
+
+  it('does not reach into the GPU comparison charts', () => {
+    document.body.innerHTML =
+      '<div data-testid="gpu-graph">' +
+      '<div class="dot-group" data-benchmark-type="agentic_traces"></div>' +
+      '</div>';
+    expect(document.querySelectorAll(AGENTIC_POINT_SELECTOR)).toHaveLength(0);
+  });
+});
+
+describe('AGENTIC_POINT_ACTION_SELECTOR', () => {
+  it('treats a click on an official OR an overlay point as taking the action', () => {
+    const [official, overlay] = renderChart([
+      { left: 100, top: 100 },
+      { left: 200, top: 100, overlay: true },
+    ]);
+
+    expect(official.closest(AGENTIC_POINT_ACTION_SELECTOR)).toBe(official);
+    expect(overlay.closest(AGENTIC_POINT_ACTION_SELECTOR)).toBe(overlay);
+  });
+});
+
+describe('resolveAgenticPointAnchor', () => {
+  it('returns null when the chart has not rendered', () => {
+    document.body.innerHTML = '';
+    expect(resolveAgenticPointAnchor()).toBeNull();
+  });
+
+  it('returns null when the chart has rendered but has no agentic points', () => {
+    renderChart([{ left: 100, top: 100, benchmarkType: 'single_turn' }]);
+    expect(resolveAgenticPointAnchor()).toBeNull();
+  });
+
+  it('picks the agentic point nearest the chart centre', () => {
+    // Chart centre is (500, 300) — the middle point is closest.
+    const [, middle] = renderChart([
+      { left: 60, top: 60 },
+      { left: 480, top: 280 },
+      { left: 900, top: 540 },
+    ]);
+    expect(resolveAgenticPointAnchor()).toBe(middle);
+  });
+
+  it('prefers a point with stored telemetry over a closer one without', () => {
+    const [, withTrace] = renderChart([
+      { left: 490, top: 290 },
+      { left: 300, top: 200, hasTrace: true },
+    ]);
+    expect(resolveAgenticPointAnchor()).toBe(withTrace);
+  });
+
+  it('falls back to any agentic point while the trace lookup is still pending', () => {
+    const [only] = renderChart([{ left: 480, top: 280 }]);
+    expect(resolveAgenticPointAnchor()).toBe(only);
+  });
+
+  it('does not point at a dead-end point after trace availability resolves empty', () => {
+    renderChart([{ left: 480, top: 280, traceAvailability: 'resolved' }]);
+    expect(resolveAgenticPointAnchor()).toBeNull();
+  });
+
+  it('scopes point geometry to each visible scatter chart', () => {
+    renderChart([]);
+    const firstChart = document.querySelector<HTMLElement>('[data-testid="scatter-graph"]')!;
+    stubBox(firstChart, 0, 1200, 1000, 600);
+
+    const secondChart = document.createElement('div');
+    secondChart.dataset.testid = 'scatter-graph';
+    stubBox(secondChart, 0, 0, 1000, 600);
+    const point = document.createElement('div');
+    point.className = 'dot-group';
+    point.dataset.benchmarkType = 'agentic_traces';
+    point.dataset.hasTrace = 'true';
+    stubRect(point, 480, 280);
+    secondChart.append(point);
+    document.body.append(secondChart);
+
+    expect(resolveAgenticPointAnchor()).toBe(point);
+  });
+
+  it('skips points hidden by a legend or precision filter', () => {
+    const [, visible] = renderChart([
+      { left: 500, top: 300, hidden: true },
+      { left: 200, top: 150 },
+    ]);
+    expect(resolveAgenticPointAnchor()).toBe(visible);
+  });
+
+  it('never anchors to an unofficial-run overlay marker', () => {
+    // Overlay runs have no stored trace, so their tooltip offers no
+    // "View charts" link — pointing at one would teach a dead end.
+    renderChart([{ left: 500, top: 300, overlay: true }]);
+    expect(resolveAgenticPointAnchor()).toBeNull();
+  });
+
+  it('points at the marker inside the group, not the label-inflated group box', () => {
+    // A `.dot-group` box also spans the hit area and the C=/DEP label above the
+    // dot, so its centre can sit well off the dot itself.
+    const [group] = renderChart([{ left: 480, top: 280, withShape: true }]);
+    const shape = group.querySelector('.visible-shape');
+
+    expect(shape).not.toBeNull();
+    expect(resolveAgenticPointAnchor()).toBe(shape);
+  });
+
+  it('falls back to the group when it has no marker child yet', () => {
+    const [group] = renderChart([{ left: 480, top: 280 }]);
+    expect(resolveAgenticPointAnchor()).toBe(group);
+  });
+
+  // The scatter's real geometry: a 900x520 SVG with 60px left / 24px top
+  // margins, clipped to the 830x436 plot inside them. The left 60px and the
+  // bottom 60px of the SVG are axis gutters — inside the SVG box, outside the
+  // clip.
+  const CHART_SVG = {
+    box: { left: 40, top: 40, width: 900, height: 520 },
+    margin: { left: 60, top: 24 },
+    clip: { width: 830, height: 436 },
+  };
+
+  it('ignores points that zooming has pushed outside the clipped plot area', () => {
+    // Outside the plot the chart's clip path hides a point, but its bounding
+    // box stays perfectly ordinary — only the clip region rules it out.
+    renderChart([{ left: 960, top: 300 }], CHART_SVG);
+    expect(resolveAgenticPointAnchor()).toBeNull();
+  });
+
+  it.each([
+    ['left gutter, over the y-axis labels', { left: 70, top: 300 }],
+    ['bottom gutter, under the x-axis labels', { left: 500, top: 505 }],
+  ])('ignores a zoomed point parked in the %s', (_label, spec) => {
+    // Regression: this used to be checked against the SVG's own box, which
+    // includes the 60px axis margins — so a point sitting over the axis labels
+    // was painted away by the clip path yet still accepted, aiming the pointer
+    // at something invisible.
+    renderChart([spec], CHART_SVG);
+    const svgBox = document.querySelector('[data-testid="d3-chart-svg"]')!.getBoundingClientRect();
+    // Guard the guard: the point really is inside the SVG, so this is not
+    // vacuously passing because it fell off the element altogether.
+    expect(spec.left).toBeGreaterThan(svgBox.left);
+    expect(spec.left).toBeLessThan(svgBox.right);
+    expect(spec.top).toBeGreaterThan(svgBox.top);
+    expect(spec.top).toBeLessThan(svgBox.bottom);
+
+    expect(resolveAgenticPointAnchor()).toBeNull();
+  });
+
+  it('still anchors to points that remain inside the plot area', () => {
+    const [inside] = renderChart(
+      [
+        { left: 400, top: 300 },
+        { left: 960, top: 300 },
+      ],
+      CHART_SVG,
+    );
+    expect(resolveAgenticPointAnchor()).toBe(inside);
+  });
+
+  it('falls back to the SVG box when the chart clips nothing', () => {
+    // clipContent: false leaves no clipPath, so nothing is painted away and
+    // the SVG box is the honest visibility test — not a reason to reject every
+    // point.
+    const [only] = renderChart([{ left: 70, top: 300 }]);
+    expect(resolveAgenticPointAnchor()).toBe(only);
+  });
+
+  it('bails out cheaply while the whole chart is below the fold', () => {
+    // The chart usually starts off-screen, and this runs on every scroll
+    // event — so an off-screen chart must be rejected without measuring points.
+    const [only] = renderChart([{ left: 480, top: 280 }]);
+    const chart = document.querySelector('[data-testid="scatter-graph"]')!;
+    stubBox(chart, 0, 5000, 1000, 600);
+    const measure = vi.spyOn(only, 'getBoundingClientRect');
+
+    expect(resolveAgenticPointAnchor()).toBeNull();
+    expect(measure, 'no per-point layout reads').not.toHaveBeenCalled();
+  });
+
+  it('returns null when every point is scrolled out of the viewport', () => {
+    renderChart([
+      { left: 400, top: -900 },
+      { left: 400, top: 4000 },
+    ]);
+    expect(resolveAgenticPointAnchor()).toBeNull();
+  });
+
+  it('validates a cached marker without measuring sibling points', () => {
+    const [first, selected, last] = renderChart([
+      { left: 100, top: 100, withShape: true },
+      { left: 480, top: 280, withShape: true },
+      { left: 900, top: 500, withShape: true },
+    ]);
+    const anchor = resolveAgenticPointAnchor();
+    expect(anchor).toBe(selected.querySelector('.visible-shape'));
+    const firstMeasure = vi.spyOn(first, 'getBoundingClientRect');
+    const lastMeasure = vi.spyOn(last, 'getBoundingClientRect');
+
+    expect(getAgenticPointAnchorRect(anchor!)).not.toBeNull();
+    expect(firstMeasure, 'first sibling skipped').not.toHaveBeenCalled();
+    expect(lastMeasure, 'last sibling skipped').not.toHaveBeenCalled();
+  });
+
+  it('rejects a cached marker after its point becomes hidden', () => {
+    const [group] = renderChart([{ left: 480, top: 280, withShape: true }]);
+    const anchor = resolveAgenticPointAnchor();
+    expect(anchor).not.toBeNull();
+
+    group.setAttribute('style', 'opacity: 0');
+    expect(getAgenticPointAnchorRect(anchor!)).toBeNull();
+  });
+
+  it('scopes mutation observation to the stable chart surface', () => {
+    const [group] = renderChart([{ left: 480, top: 280, withShape: true }], CHART_SVG);
+    const chartRoot = document.querySelector('[data-coach-mark-root]')!;
+    const anchor = group.querySelector('.visible-shape');
+
+    expect(getAgenticPointAnchorMutationRoot(anchor)).toBe(chartRoot);
+  });
+});
diff --git a/packages/app/src/lib/nudges/agentic-point-coach-mark.ts b/packages/app/src/lib/nudges/agentic-point-coach-mark.ts
new file mode 100644
index 000000000..16a6c04ca
--- /dev/null
+++ b/packages/app/src/lib/nudges/agentic-point-coach-mark.ts
@@ -0,0 +1,177 @@
+/**
+ * Anchor resolution for the "click a point for server metrics and logs"
+ * coach mark on the agentic inference chart.
+ *
+ * The affordance it teaches: clicking an agentic scatter point pins its
+ * tooltip, and the pinned tooltip carries a "View charts" link to
+ * `/inference/agentic/[id]`. Nothing on the chart hints at that, so first-time
+ * visitors never find it.
+ */
+
+import { plotBounds } from '@/lib/d3-chart/plot-bounds';
+
+import { isAnchorOnScreen, isAnchorWithin, viewportSize } from './anchor';
+
+export const AGENTIC_COACH_MARK_STORAGE_KEY = 'inferencex-agentic-point-coach-mark-dismissed';
+
+/**
+ * Dispatched by ScatterGraph after every full D3 render.
+ *
+ * The coach mark needs a painted point to anchor to, and the chart paints
+ * asynchronously (availability → benchmarks → derived metrics). A plain timer
+ * would fire into an empty chart and give up; re-attempting on each render
+ * means the tip appears as soon as there is something to point at. Exported so
+ * the dispatch site imports a stable constant, mirroring `GRADIENT_NUDGE_EVENT`.
+ */
+export const SCATTER_RENDERED_EVENT = 'inferencex:scatter-chart-rendered';
+
+/**
+ * Official agentic points on the main inference scatter chart.
+ *
+ * Deliberately scoped to `[data-testid="scatter-graph"]` and to `.dot-group`:
+ *  - the GPU comparison charts (`[data-testid="gpu-graph"]`) are a different
+ *    surface with their own layout;
+ *  - unofficial-run overlay markers are `.unofficial-overlay-pt`, and their
+ *    tooltip has no "View charts" link (overlay runs have no stored trace), so
+ *    pointing at one would teach a dead end.
+ */
+const SCATTER_GRAPH_SELECTOR = '[data-testid="scatter-graph"]';
+const COACH_MARK_ROOT_SELECTOR = '[data-coach-mark-root]';
+const AGENTIC_POINT_GROUP_SELECTOR = '.dot-group[data-benchmark-type="agentic_traces"]';
+
+export const AGENTIC_POINT_SELECTOR = `${SCATTER_GRAPH_SELECTOR} ${AGENTIC_POINT_GROUP_SELECTOR}`;
+
+/**
+ * Clicking any scatter point — official or overlay — means the user found the
+ * interaction, so the coach mark has done its job.
+ */
+export const AGENTIC_POINT_ACTION_SELECTOR =
+  '[data-testid="scatter-graph"] .dot-group, [data-testid="scatter-graph"] .unofficial-overlay-pt';
+
+function isRendered(element: Element): boolean {
+  // Hidden points stay in the DOM at opacity 0 with pointer-events off (see
+  // the ScatterGraph decoration effect) — they are not clickable, so they are
+  // not anchor candidates either.
+  const style = getComputedStyle(element);
+  return style.opacity !== '0' && style.pointerEvents !== 'none' && style.visibility !== 'hidden';
+}
+
+function agenticPointGroup(element: Element): SVGGElement | null {
+  const group = element.matches(AGENTIC_POINT_GROUP_SELECTOR)
+    ? element
+    : element.closest(AGENTIC_POINT_GROUP_SELECTOR);
+  return group as SVGGElement | null;
+}
+
+function canOpenTrace(group: SVGGElement): boolean {
+  return group.dataset.hasTrace === 'true' || group.dataset.traceAvailability !== 'resolved';
+}
+
+/**
+ * Validate one previously selected point without rescanning every series.
+ *
+ * D3 coordinate transitions mutate every point's `transform` each frame. The
+ * coach mark can keep following its chosen point while this stays non-null;
+ * filtering, clipping, or resolved-empty trace availability invalidates it and
+ * sends the caller back through `resolveAgenticPointAnchor`.
+ */
+export function getAgenticPointAnchorRect(element: Element): DOMRect | null {
+  if (typeof document === 'undefined') return null;
+  const group = agenticPointGroup(element);
+  const chart = group?.closest<HTMLElement>(SCATTER_GRAPH_SELECTOR);
+  if (!group || !chart || !isRendered(group) || !canOpenTrace(group)) return null;
+
+  const viewport = viewportSize();
+  const rect = element.getBoundingClientRect();
+  if (!isAnchorOnScreen(rect, viewport)) return null;
+
+  const svg = chart.querySelector('[data-testid="d3-chart-svg"]');
+  const plot =
+    (svg && plotBounds(svg)) ?? svg?.getBoundingClientRect() ?? chart.getBoundingClientRect();
+  return isAnchorWithin(rect, plot) ? rect : null;
+}
+
+/**
+ * Observe the stable chart owner so point mutations are local, while replacing
+ * the chart itself still invalidates the cached anchor.
+ */
+export function getAgenticPointAnchorMutationRoot(element: Element | null): Node | null {
+  if (typeof document === 'undefined') return null;
+  return (
+    element?.closest(COACH_MARK_ROOT_SELECTOR) ??
+    element?.closest(SCATTER_GRAPH_SELECTOR) ??
+    document.querySelector(COACH_MARK_ROOT_SELECTOR) ??
+    document.querySelector(SCATTER_GRAPH_SELECTOR)
+  );
+}
+
+/**
+ * Pick the agentic point to point at: the eligible candidate nearest the
+ * chart's centre, so the callout has room on every side and the choice is
+ * stable between re-renders of the same view.
+ *
+ * Prefers points with stored telemetry (`data-has-trace`) — those are the ones
+ * whose tooltip actually offers "View charts" — and falls back to any visible
+ * agentic point while the availability lookup is still in flight.
+ */
+export function resolveAgenticPointAnchor(): Element | null {
+  if (typeof document === 'undefined') return null;
+
+  const viewport = viewportSize();
+  const eligible: {
+    element: SVGGElement;
+    hasTrace: boolean;
+    traceAvailability: string | undefined;
+    distance: number;
+  }[] = [];
+
+  for (const chart of document.querySelectorAll<HTMLElement>('[data-testid="scatter-graph"]')) {
+    const chartRect = chart.getBoundingClientRect();
+    // Cheap bail-out. This runs on every scroll event, and the full scan is a
+    // layout read per point, so reject each off-screen chart first.
+    if (
+      chartRect.bottom <= 0 ||
+      chartRect.top >= viewport.height ||
+      chartRect.right <= 0 ||
+      chartRect.left >= viewport.width
+    ) {
+      continue;
+    }
+
+    const chartCx = chartRect.left + chartRect.width / 2;
+    const chartCy = chartRect.top + chartRect.height / 2;
+    // Zoomed points retain ordinary bounding boxes outside the clip region.
+    // Check against the plot rather than the SVG box, which includes gutters.
+    const svg = chart.querySelector('[data-testid="d3-chart-svg"]');
+    const plot = (svg && plotBounds(svg)) ?? svg?.getBoundingClientRect() ?? chartRect;
+
+    for (const element of chart.querySelectorAll<SVGGElement>(AGENTIC_POINT_GROUP_SELECTOR)) {
+      if (!isRendered(element)) continue;
+      const rect = element.getBoundingClientRect();
+      if (!isAnchorOnScreen(rect, viewport)) continue;
+      if (!isAnchorWithin(rect, plot)) continue;
+      const cx = rect.left + rect.width / 2;
+      const cy = rect.top + rect.height / 2;
+      eligible.push({
+        element,
+        hasTrace: element.dataset.hasTrace === 'true',
+        traceAvailability: element.dataset.traceAvailability,
+        distance: (cx - chartCx) ** 2 + (cy - chartCy) ** 2,
+      });
+    }
+  }
+
+  if (eligible.length === 0) return null;
+
+  const withTrace = eligible.filter((candidate) => candidate.hasTrace);
+  const pending = eligible.filter((candidate) => candidate.traceAvailability !== 'resolved');
+  const pool = withTrace.length > 0 ? withTrace : pending;
+  if (pool.length === 0) return null;
+  const chosen = pool.reduce((best, candidate) =>
+    candidate.distance < best.distance ? candidate : best,
+  ).element;
+
+  // Point at the marker, not at the group: a `.dot-group` also contains the
+  // hit area and point label, so its bounding box centre can sit off the dot.
+  return chosen.querySelector('.visible-shape') ?? chosen;
+}
diff --git a/packages/app/src/lib/nudges/anchor.test.ts b/packages/app/src/lib/nudges/anchor.test.ts
new file mode 100644
index 000000000..2bd9e27c8
--- /dev/null
+++ b/packages/app/src/lib/nudges/anchor.test.ts
@@ -0,0 +1,159 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  COACH_MARK_MARGIN,
+  anchorCentre,
+  computeCoachMarkPlacement,
+  isAnchorOnScreen,
+  isAnchorWithin,
+  type AnchorRect,
+  type Size,
+} from './anchor';
+
+const VIEWPORT: Size = { width: 1280, height: 800 };
+const CARD: Size = { width: 288, height: 96 };
+
+const point = (left: number, top: number, size = 12): AnchorRect => ({
+  left,
+  top,
+  width: size,
+  height: size,
+});
+
+function cardRect(placement: { left: number; top: number }, card: Size = CARD) {
+  return {
+    left: placement.left,
+    top: placement.top,
+    right: placement.left + card.width,
+    bottom: placement.top + card.height,
+  };
+}
+
+describe('anchorCentre', () => {
+  it('returns the geometric centre of the rect', () => {
+    expect(anchorCentre({ left: 100, top: 40, width: 20, height: 10 })).toEqual({ x: 110, y: 45 });
+  });
+});
+
+describe('isAnchorOnScreen', () => {
+  it('accepts a point well inside the viewport', () => {
+    expect(isAnchorOnScreen(point(600, 400), VIEWPORT)).toBe(true);
+  });
+
+  it('rejects a zero-size rect (element not rendered yet)', () => {
+    expect(isAnchorOnScreen({ left: 600, top: 400, width: 0, height: 0 }, VIEWPORT)).toBe(false);
+  });
+
+  it('rejects a point scrolled above the viewport', () => {
+    expect(isAnchorOnScreen(point(600, -400), VIEWPORT)).toBe(false);
+  });
+
+  it('rejects a point scrolled below the viewport', () => {
+    expect(isAnchorOnScreen(point(600, 1400), VIEWPORT)).toBe(false);
+  });
+
+  it('rejects a point hugging the viewport edge, where the pointer would run off-canvas', () => {
+    expect(isAnchorOnScreen(point(-2, 400), VIEWPORT)).toBe(false);
+    expect(isAnchorOnScreen(point(VIEWPORT.width - 4, 400), VIEWPORT)).toBe(false);
+  });
+});
+
+describe('isAnchorWithin', () => {
+  const plot = { left: 100, top: 100, right: 900, bottom: 500 };
+
+  it('accepts a point inside the bounds', () => {
+    expect(isAnchorWithin(point(500, 300), plot)).toBe(true);
+  });
+
+  it('rejects a point outside the bounds', () => {
+    // Zooming pushes points beyond the plot, where a clip path hides them —
+    // they keep an ordinary bounding box, so bounds are what rule them out.
+    expect(isAnchorWithin(point(1400, 300), plot)).toBe(false);
+    expect(isAnchorWithin(point(500, 40), plot)).toBe(false);
+  });
+
+  it('applies the inset to every edge', () => {
+    expect(isAnchorWithin(point(104, 300), plot)).toBe(true);
+    expect(isAnchorWithin(point(104, 300), plot, 40)).toBe(false);
+  });
+
+  it('rejects a zero-size rect', () => {
+    expect(isAnchorWithin({ left: 500, top: 300, width: 0, height: 0 }, plot)).toBe(false);
+  });
+});
+
+describe('computeCoachMarkPlacement', () => {
+  it('places the card below the anchor and centres it horizontally', () => {
+    const anchor = point(600, 300);
+    const placement = computeCoachMarkPlacement(anchor, CARD, VIEWPORT);
+    const { x: cx, y: cy } = anchorCentre(anchor);
+
+    expect(placement.side).toBe('below');
+    expect(placement.top).toBeGreaterThan(cy);
+    expect(placement.left + CARD.width / 2).toBeCloseTo(cx, 5);
+  });
+
+  it('flips above the anchor when the card would overflow the bottom', () => {
+    const anchor = point(600, 760);
+    const placement = computeCoachMarkPlacement(anchor, CARD, VIEWPORT);
+
+    expect(placement.side).toBe('above');
+    expect(cardRect(placement).bottom).toBeLessThan(anchorCentre(anchor).y);
+  });
+
+  it('always points at the anchor centre', () => {
+    const anchor = point(240, 180);
+    const placement = computeCoachMarkPlacement(anchor, CARD, VIEWPORT);
+    const centre = anchorCentre(anchor);
+
+    expect(placement.arrowX2).toBe(centre.x);
+    expect(placement.arrowY2).toBe(centre.y);
+  });
+
+  it('starts the pointer on the card edge that faces the anchor', () => {
+    const below = computeCoachMarkPlacement(point(600, 300), CARD, VIEWPORT);
+    expect(below.arrowY1).toBe(below.top);
+
+    const above = computeCoachMarkPlacement(point(600, 760), CARD, VIEWPORT);
+    expect(above.arrowY1).toBe(above.top + CARD.height);
+  });
+
+  it('keeps the pointer origin within the card, not at its corner', () => {
+    // An anchor hard against the left edge pushes the card right; the pointer
+    // must still start inside the card rather than beyond its corner.
+    const placement = computeCoachMarkPlacement(point(30, 300), CARD, VIEWPORT);
+    expect(placement.arrowX1).toBeGreaterThan(placement.left);
+    expect(placement.arrowX1).toBeLessThan(placement.left + CARD.width);
+  });
+
+  it.each([
+    ['top-left', point(30, 30)],
+    ['top-right', point(1240, 30)],
+    ['bottom-left', point(30, 770)],
+    ['bottom-right', point(1240, 770)],
+    ['centre', point(640, 400)],
+  ])('never renders off-canvas for a %s anchor', (_label, anchor) => {
+    const rect = cardRect(computeCoachMarkPlacement(anchor, CARD, VIEWPORT));
+
+    expect(rect.left).toBeGreaterThanOrEqual(COACH_MARK_MARGIN);
+    expect(rect.top).toBeGreaterThanOrEqual(COACH_MARK_MARGIN);
+    expect(rect.right).toBeLessThanOrEqual(VIEWPORT.width - COACH_MARK_MARGIN);
+    expect(rect.bottom).toBeLessThanOrEqual(VIEWPORT.height - COACH_MARK_MARGIN);
+  });
+
+  it('pins to the top-left margin when the card is larger than the viewport', () => {
+    const tiny: Size = { width: 320, height: 240 };
+    const placement = computeCoachMarkPlacement(point(160, 120), { width: 400, height: 400 }, tiny);
+
+    expect(placement.left).toBe(COACH_MARK_MARGIN);
+    expect(placement.top).toBe(COACH_MARK_MARGIN);
+  });
+
+  it('moves the card as the anchor moves (zoom / pan)', () => {
+    const before = computeCoachMarkPlacement(point(400, 300), CARD, VIEWPORT);
+    const after = computeCoachMarkPlacement(point(700, 300), CARD, VIEWPORT);
+
+    expect(after.left - before.left).toBeCloseTo(300, 5);
+    expect(after.arrowX2 - before.arrowX2).toBeCloseTo(300, 5);
+  });
+});
diff --git a/packages/app/src/lib/nudges/anchor.ts b/packages/app/src/lib/nudges/anchor.ts
new file mode 100644
index 000000000..444414496
--- /dev/null
+++ b/packages/app/src/lib/nudges/anchor.ts
@@ -0,0 +1,151 @@
+/**
+ * Geometry for anchored (coach-mark) nudges.
+ *
+ * Kept free of DOM APIs so the placement rules are unit-testable: callers pass
+ * plain rects measured with `getBoundingClientRect()` and get viewport
+ * coordinates back. The DOM-facing half — which element to point at — lives
+ * next to each nudge (see `agentic-point-coach-mark.ts`).
+ */
+
+export interface AnchorRect {
+  left: number;
+  top: number;
+  width: number;
+  height: number;
+}
+
+export interface Size {
+  width: number;
+  height: number;
+}
+
+export type CoachMarkSide = 'above' | 'below';
+
+export interface CoachMarkPlacement {
+  /** Card top-left, in viewport coordinates. */
+  left: number;
+  top: number;
+  /** Pointer start — on the card edge facing the anchor. */
+  arrowX1: number;
+  arrowY1: number;
+  /** Pointer end — the anchor's centre. */
+  arrowX2: number;
+  arrowY2: number;
+  side: CoachMarkSide;
+}
+
+/** Gap between the anchor centre and the nearest card edge. */
+export const COACH_MARK_GAP = 52;
+/** Minimum distance the card keeps from every viewport edge. */
+export const COACH_MARK_MARGIN = 12;
+/** How far the pointer's card-edge end stays from the card's corners. */
+const ARROW_INSET = 20;
+/**
+ * A candidate must sit this far inside the viewport to be anchor-worthy — a
+ * point clipped by the plot edge would get a pointer running off-canvas.
+ */
+const ON_SCREEN_INSET = 24;
+
+/**
+ * Current viewport in CSS pixels. The one DOM touch in this module, kept here
+ * so the resolver and the renderer measure against exactly the same box.
+ * `clientWidth` excludes the scrollbar (what a fixed overlay is laid out
+ * against); the `innerWidth` fallback covers environments — jsdom, an
+ * unattached document — that report 0.
+ */
+export function viewportSize(): Size {
+  const root = document.documentElement;
+  return {
+    width: root.clientWidth || window.innerWidth,
+    height: root.clientHeight || window.innerHeight,
+  };
+}
+
+function clamp(value: number, min: number, max: number): number {
+  // When the card is wider/taller than the space available, `min > max`; bias
+  // to `min` so the card stays pinned to the top-left edge rather than
+  // flipping to a negative offset.
+  if (max < min) return min;
+  return Math.min(Math.max(value, min), max);
+}
+
+export function anchorCentre(rect: AnchorRect): { x: number; y: number } {
+  return { x: rect.left + rect.width / 2, y: rect.top + rect.height / 2 };
+}
+
+export interface Bounds {
+  left: number;
+  top: number;
+  right: number;
+  bottom: number;
+}
+
+/**
+ * Does this anchor sit comfortably inside `bounds`? Rejects zero-size rects (a
+ * hidden or not-yet-rendered element) and anything whose centre falls outside,
+ * or within `inset` of, the edge.
+ */
+export function isAnchorWithin(rect: AnchorRect, bounds: Bounds, inset = 0): boolean {
+  if (rect.width <= 0 || rect.height <= 0) return false;
+  const { x, y } = anchorCentre(rect);
+  return (
+    x >= bounds.left + inset &&
+    x <= bounds.right - inset &&
+    y >= bounds.top + inset &&
+    y <= bounds.bottom - inset
+  );
+}
+
+/**
+ * Is this anchor usable right now? Off-screen anchors, and anchors hugging a
+ * viewport edge (where the pointer would run off-canvas), are not.
+ */
+export function isAnchorOnScreen(rect: AnchorRect, viewport: Size): boolean {
+  return isAnchorWithin(
+    rect,
+    { left: 0, top: 0, right: viewport.width, bottom: viewport.height },
+    ON_SCREEN_INSET,
+  );
+}
+
+/**
+ * Place the callout near the anchor without ever leaving the viewport.
+ *
+ * Prefers below the anchor (the chart's x-axis and its caption sit there, so
+ * the card covers less data) and flips above when it would overflow the
+ * bottom. Both axes are clamped, so a card can end up offset from the anchor —
+ * the pointer still connects the two, which is why the arrow is computed from
+ * the *placed* card rather than the desired position.
+ */
+export function computeCoachMarkPlacement(
+  rect: AnchorRect,
+  card: Size,
+  viewport: Size,
+): CoachMarkPlacement {
+  const { x: cx, y: cy } = anchorCentre(rect);
+
+  const belowTop = cy + COACH_MARK_GAP;
+  const fitsBelow = belowTop + card.height <= viewport.height - COACH_MARK_MARGIN;
+  const side: CoachMarkSide = fitsBelow ? 'below' : 'above';
+
+  const desiredTop = fitsBelow ? belowTop : cy - COACH_MARK_GAP - card.height;
+  const top = clamp(
+    desiredTop,
+    COACH_MARK_MARGIN,
+    Math.max(COACH_MARK_MARGIN, viewport.height - card.height - COACH_MARK_MARGIN),
+  );
+  const left = clamp(
+    cx - card.width / 2,
+    COACH_MARK_MARGIN,
+    Math.max(COACH_MARK_MARGIN, viewport.width - card.width - COACH_MARK_MARGIN),
+  );
+
+  const arrowX1 = clamp(
+    cx,
+    left + Math.min(ARROW_INSET, card.width / 2),
+    left + card.width - Math.min(ARROW_INSET, card.width / 2),
+  );
+  const arrowY1 = side === 'below' ? top : top + card.height;
+
+  return { left, top, arrowX1, arrowY1, arrowX2: cx, arrowY2: cy, side };
+}
diff --git a/packages/app/src/lib/nudges/landing-banner.ts b/packages/app/src/lib/nudges/landing-banner.ts
index bd291f883..ab3dd9487 100644
--- a/packages/app/src/lib/nudges/landing-banner.ts
+++ b/packages/app/src/lib/nudges/landing-banner.ts
@@ -1,2 +1,2 @@
-export const LANDING_BANNER_STORAGE_KEY = 'inferencex-kimi-k3-banner-dismissed';
+export const LANDING_BANNER_STORAGE_KEY = 'inferencex-openai-rubin-banner-dismissed';
 export const LANDING_BANNER_DISMISSED_ATTRIBUTE = 'data-landing-banner-dismissed';
diff --git a/packages/app/src/lib/nudges/policy.test.ts b/packages/app/src/lib/nudges/policy.test.ts
index 692c8fa7d..85ea98480 100644
--- a/packages/app/src/lib/nudges/policy.test.ts
+++ b/packages/app/src/lib/nudges/policy.test.ts
@@ -31,6 +31,12 @@ describe('dismissesOnAction', () => {
     expect(dismissesOnAction(makeNudge('modal'))).toBe(true);
   });
 
+  it('returns true for coach-mark by default', () => {
+    // Performing the interaction the coach mark teaches is the whole point —
+    // it must not keep pointing at the chart afterwards.
+    expect(dismissesOnAction(makeNudge('coach-mark'))).toBe(true);
+  });
+
   it('returns false for banner by default', () => {
     expect(dismissesOnAction(makeNudge('banner'))).toBe(false);
   });
@@ -43,6 +49,10 @@ describe('dismissesOnAction', () => {
     expect(dismissesOnAction(makeNudge('modal', false))).toBe(false);
   });
 
+  it('respects explicit dismissOnAction: false on a coach-mark', () => {
+    expect(dismissesOnAction(makeNudge('coach-mark', false))).toBe(false);
+  });
+
   it('respects explicit dismissOnAction: true on a banner', () => {
     expect(dismissesOnAction(makeNudge('banner', true))).toBe(true);
   });
diff --git a/packages/app/src/lib/nudges/registry.test.ts b/packages/app/src/lib/nudges/registry.test.ts
index 7d46c5fb3..a8c3be41c 100644
--- a/packages/app/src/lib/nudges/registry.test.ts
+++ b/packages/app/src/lib/nudges/registry.test.ts
@@ -1,6 +1,14 @@
-import { describe, expect, it } from 'vitest';
+import { afterEach, describe, expect, it, vi } from 'vitest';
 
-import { NUDGE_REGISTRY } from './registry';
+import {
+  AGENTIC_COACH_MARK_STORAGE_KEY,
+  AGENTIC_POINT_ACTION_SELECTOR,
+} from './agentic-point-coach-mark';
+import { dismissesOnAction } from './policy';
+import { NUDGE_REGISTRY, TELEMETRY_TUTORIAL_STORAGE_KEY } from './registry';
+afterEach(() => {
+  vi.unstubAllGlobals();
+});
 
 describe('NUDGE_REGISTRY integrity', () => {
   it('has no duplicate IDs', () => {
@@ -15,13 +23,13 @@ describe('NUDGE_REGISTRY integrity', () => {
 
   it('every entry has a valid type', () => {
     for (const nudge of NUDGE_REGISTRY) {
-      expect(['toast', 'modal', 'banner']).toContain(nudge.type);
+      expect(['toast', 'modal', 'banner', 'coach-mark']).toContain(nudge.type);
     }
   });
 
   it('every entry has a valid scope', () => {
     for (const nudge of NUDGE_REGISTRY) {
-      expect(['dashboard', 'landing', 'evaluation']).toContain(nudge.scope);
+      expect(['dashboard', 'landing', 'evaluation', 'agentic-detail']).toContain(nudge.scope);
     }
   });
 
@@ -62,19 +70,89 @@ describe('NUDGE_REGISTRY integrity', () => {
   it('contains the expected set of migrated nudges', () => {
     const ids = NUDGE_REGISTRY.map((n) => n.id).toSorted();
     expect(ids).toEqual([
+      'agentic-point-detail',
+      'agentx-telemetry-tutorial',
       'eval-samples',
       'export',
       'feedback-modal',
       'filter-hint',
-      'github-star-modal',
       'gradient-label',
-      'kimi-k3-launch-banner',
-      'kimi-k3-launch-modal',
+      'openai-rubin-comparison-banner',
       'reproducibility',
       'star-nudge',
     ]);
   });
 
+  it('renders the telemetry tutorial as an uncentered card on the agentic detail scope', () => {
+    const tutorial = NUDGE_REGISTRY.find((n) => n.id === 'agentx-telemetry-tutorial');
+    expect(tutorial?.scope).toBe('agentic-detail');
+    // A backdrop would cover the very charts the tutorial describes.
+    expect(tutorial?.content.centered).toBeUndefined();
+    // cypress/support/e2e.ts seeds this key so the card can't sit over the
+    // charts under test; a rename here has to be mirrored there.
+    expect(tutorial?.storageKey).toBe(TELEMETRY_TUTORIAL_STORAGE_KEY);
+    expect(TELEMETRY_TUTORIAL_STORAGE_KEY).toBe('inferencex-agentx-telemetry-tutorial-dismissed');
+  });
+
+  it('keeps internal action destinations in the active Chinese route tree', () => {
+    const location = { pathname: '/zh/inference', href: '' };
+    vi.stubGlobal('window', { location });
+
+    const reproducibility = NUDGE_REGISTRY.find((nudge) => nudge.id === 'reproducibility');
+    if (reproducibility?.type !== 'toast') throw new Error('Missing reproducibility toast');
+    reproducibility.content.action?.onClick();
+    expect(location.href).toBe('/zh/about#reproducibility');
+
+    const banner = NUDGE_REGISTRY.find((nudge) => nudge.id === 'openai-rubin-comparison-banner');
+    if (banner?.type !== 'banner') throw new Error('Missing launch banner');
+    expect(banner.storageKey).toBe('inferencex-openai-rubin-banner-dismissed');
+    expect(banner.content.href).toBe(
+      '/inference?g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_outputTputPerMw',
+    );
+    expect(banner.analytics).toEqual({
+      shown: 'inference_rubin_comparison_banner_shown',
+      dismissed: 'inference_rubin_comparison_banner_dismissed',
+      action: 'inference_rubin_comparison_banner_clicked',
+      properties: {
+        banner_id: 'openai-rubin-comparison',
+        scenario: '8k/1k',
+        model: 'DeepSeek-R1-0528',
+        metric: 'y_outputTputPerMw',
+      },
+    });
+    banner.content.onLinkClick?.();
+    expect(location.href).toBe(
+      '/zh/inference?g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_outputTputPerMw',
+    );
+  });
+
+  it('gives every coach mark an anchor to point at', () => {
+    for (const nudge of NUDGE_REGISTRY.filter((n) => n.type === 'coach-mark')) {
+      expect(typeof nudge.content.anchor?.resolve).toBe('function');
+    }
+  });
+
+  it('anchors the agentic coach mark to the inference chart, dismissed for good on a point click', () => {
+    const coachMark = NUDGE_REGISTRY.find((n) => n.id === 'agentic-point-detail');
+
+    expect(coachMark?.type).toBe('coach-mark');
+    expect(coachMark?.scope).toBe('dashboard');
+    expect(coachMark?.dismissal.type).toBe('permanent');
+    expect(coachMark?.storageKey).toBe(AGENTIC_COACH_MARK_STORAGE_KEY);
+    expect(dismissesOnAction(coachMark!)).toBe(true);
+    expect(coachMark?.content.anchor?.actionSelector).toBe(AGENTIC_POINT_ACTION_SELECTOR);
+    expect(typeof coachMark?.content.anchor?.getRect).toBe('function');
+    expect(typeof coachMark?.content.anchor?.getMutationRoot).toBe('function');
+    expect(AGENTIC_COACH_MARK_STORAGE_KEY).toBe('inferencex-agentic-point-coach-mark-dismissed');
+  });
+
+  it('ships a Chinese translation for every user-visible nudge string', () => {
+    for (const nudge of NUDGE_REGISTRY) {
+      expect(nudge.content.titleZh, `${nudge.id} title`).toBeTruthy();
+      expect(nudge.content.descriptionZh, `${nudge.id} description`).toBeTruthy();
+    }
+  });
+
   it('preserves testId for every entry', () => {
     for (const nudge of NUDGE_REGISTRY) {
       expect(nudge.content.testId).toBeTruthy();
diff --git a/packages/app/src/lib/nudges/registry.tsx b/packages/app/src/lib/nudges/registry.tsx
index 1a2b991b9..f42b00f28 100644
--- a/packages/app/src/lib/nudges/registry.tsx
+++ b/packages/app/src/lib/nudges/registry.tsx
@@ -1,6 +1,8 @@
 import {
   ArrowRight,
+  BookOpen,
   Download,
+  MousePointerClick,
   MessageSquareText,
   Palette,
   ShieldCheck,
@@ -13,6 +15,15 @@ import dynamic from 'next/dynamic';
 import { GITHUB_OWNER, GITHUB_REPO } from '@semianalysisai/inferencex-constants';
 
 import { FEEDBACK_SUBMITTED_EVENT } from '@/components/feedback-modal';
+import { isZhPathname, localePath } from '@/lib/i18n';
+import {
+  AGENTIC_COACH_MARK_STORAGE_KEY,
+  AGENTIC_POINT_ACTION_SELECTOR,
+  SCATTER_RENDERED_EVENT,
+  getAgenticPointAnchorMutationRoot,
+  getAgenticPointAnchorRect,
+  resolveAgenticPointAnchor,
+} from '@/lib/nudges/agentic-point-coach-mark';
 import { LANDING_BANNER_STORAGE_KEY } from '@/lib/nudges/landing-banner';
 
 // Keep the ~210-line FeedbackForm out of the landing/dashboard initial JS.
@@ -21,7 +32,7 @@ const FeedbackForm = dynamic(
   { ssr: false },
 );
 import { GitHubIcon } from '@/components/ui/github-icon';
-import { STARRED_EVENT, STARRED_KEY, saveStarred } from '@/lib/star-storage';
+import { STARRED_EVENT, STARRED_KEY } from '@/lib/star-storage';
 import type { NudgeDefinition } from './types';
 
 const GITHUB_REPO_URL = `https://github.com/${GITHUB_OWNER}/${GITHUB_REPO}`;
@@ -43,6 +54,18 @@ function isOnInferenceTab(): boolean {
   return (segments[0] ?? 'inference') === 'inference';
 }
 
+/** Resolve an internal nudge destination in the locale active at click time. */
+function localizedNudgeHref(enPath: string): string {
+  if (typeof window === 'undefined') return enPath;
+  return localePath(enPath, isZhPathname(window.location.pathname) ? 'zh' : 'en');
+}
+
+export const TELEMETRY_TUTORIAL_STORAGE_KEY = 'inferencex-agentx-telemetry-tutorial-dismissed';
+
+function telemetryTutorialHref(): string {
+  return localizedNudgeHref('/agentx/telemetry');
+}
+
 // ---------------------------------------------------------------------------
 // Registry — every engagement nudge in one place
 // ---------------------------------------------------------------------------
@@ -72,7 +95,7 @@ export const NUDGE_REGISTRY: NudgeDefinition[] = [
         label: 'See how',
         labelZh: '了解详情',
         onClick: () => {
-          window.location.href = '/about#reproducibility';
+          window.location.href = localizedNudgeHref('/about#reproducibility');
         },
       },
       testId: 'reproducibility-nudge',
@@ -246,6 +269,55 @@ export const NUDGE_REGISTRY: NudgeDefinition[] = [
     },
   },
 
+  // -------------------------------------------------------------------------
+  // Agentic chart coach mark
+  // -------------------------------------------------------------------------
+  {
+    id: 'agentic-point-detail',
+    type: 'coach-mark',
+    // Three ways in, all retried until an anchor exists (see `isEligible`'s
+    // `requireAnchor`): a short timer for a chart that is already painted,
+    // every subsequent chart render for the usual async-data case, and scroll
+    // — on a laptop viewport the chart starts below the fold, so the first two
+    // fire while there is still nothing on screen to point at. The resolver
+    // rejects an off-screen chart with a single rect read, keeping the scroll
+    // path cheap, and the engine drops these listeners once the tip is up.
+    trigger: [
+      { type: 'timer', delayMs: 1200 },
+      { type: 'event', event: SCATTER_RENDERED_EVENT, delayMs: 700 },
+      { type: 'dom-event', event: 'scroll' },
+    ],
+    dismissal: { type: 'permanent' },
+    storageKey: AGENTIC_COACH_MARK_STORAGE_KEY,
+    // Highest dashboard priority: it is the only nudge tied to a specific
+    // element, so it should claim its slot the moment that element exists.
+    // (It has its own slot, so this only orders it against future coach marks.)
+    priority: 45,
+    scope: 'dashboard',
+    content: {
+      icon: MousePointerClick,
+      iconClassName: 'text-brand',
+      title: 'Every point has a story',
+      titleZh: '每个数据点背后都有细节',
+      description:
+        'Click any point to view server metrics and logs — cache hit rates, queue depth, and the full request timeline for that run.',
+      descriptionZh:
+        '点击任意数据点即可查看服务端指标与日志——cache 命中率、队列深度，以及该次运行的完整请求时间线。',
+      testId: 'agentic-point-coach-mark',
+      anchor: {
+        resolve: resolveAgenticPointAnchor,
+        getRect: getAgenticPointAnchorRect,
+        getMutationRoot: getAgenticPointAnchorMutationRoot,
+        actionSelector: AGENTIC_POINT_ACTION_SELECTOR,
+      },
+    },
+    analytics: {
+      shown: 'inference_agentic_point_coach_mark_shown',
+      dismissed: 'inference_agentic_point_coach_mark_dismissed',
+      action: 'inference_agentic_point_coach_mark_point_clicked',
+    },
+  },
+
   // -------------------------------------------------------------------------
   // Dashboard modals
   // -------------------------------------------------------------------------
@@ -281,118 +353,98 @@ export const NUDGE_REGISTRY: NudgeDefinition[] = [
   },
 
   // -------------------------------------------------------------------------
-  // Landing modals
+  // Landing banner
+  //
+  // Note: the landing scope deliberately has no GitHub star nudge. The footer
+  // grid already carries the persistent star CTA (footer-star-cta) and the
+  // header carries another, so a fixed bottom card here duplicated the control
+  // and covered the footer. The engagement-triggered star-nudge toast on the
+  // dashboard scope is the only overlay star prompt.
   // -------------------------------------------------------------------------
   {
-    id: 'kimi-k3-launch-modal',
-    type: 'modal',
+    id: 'openai-rubin-comparison-banner',
+    type: 'banner',
     trigger: { type: 'immediate' },
     dismissal: { type: 'permanent' },
-    storageKey: 'inferencex-kimi-k3-modal-dismissed',
-    priority: 50,
+    storageKey: LANDING_BANNER_STORAGE_KEY,
+    priority: 60,
     scope: 'landing',
+    renderOnInitialLoad: true,
     content: {
       icon: Sparkles,
       iconClassName: 'text-brand',
-      title: 'Kimi K3 is live',
-      titleZh: 'Kimi K3 已上线',
+      title: "OpenAI's Latest In House Chip verus Rubin NVL72",
+      titleZh: 'OpenAI 最新自研芯片对比 Rubin NVL72',
       description:
-        'Day-zero benchmarks for Kimi K3 are now available across the latest NVIDIA and AMD chips. Results are experimental — see how the new model performs across hardware.',
+        'Compare Jalapeño (Teacup) with Vera Rubin (July) NVL72 on DeepSeek R1 at 8K / 1K.',
       descriptionZh:
-        'Kimi K3 的首日基准测试数据现已覆盖最新的 NVIDIA 和 AMD Chip。结果为实验性数据——来看看新模型在不同硬件上的表现。',
-      testId: 'launch-modal',
-      containerClassName: 'border-brand/40',
+        '对比 Jalapeño (Teacup) 与 Vera Rubin (July) NVL72 在 DeepSeek R1 8K / 1K 工作负载下的表现。',
+      testId: 'launch-banner',
       badge: 'New',
       badgeZh: '最新',
-      dismissLabel: 'Maybe Later',
-      dismissLabelZh: '稍后再看',
-      primaryAction: {
-        label: 'Explore',
-        labelZh: '开始探索',
-        icon: <ArrowRight className="size-4" />,
-        onClick: () => {
-          window.location.href = '/inference?preset=kimi-k3-launch';
-        },
+      href: '/inference?g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_outputTputPerMw',
+      linkLabel: 'View results',
+      linkLabelZh: '查看结果',
+      onLinkClick: () => {
+        window.location.href = localizedNudgeHref(
+          '/inference?g_model=DeepSeek-R1-0528&i_seq=8k%2F1k&i_prec=fp4&i_metric=y_outputTputPerMw',
+        );
       },
     },
     analytics: {
-      shown: 'kimi_k3_modal_shown',
-      dismissed: 'kimi_k3_modal_dismissed',
-      action: 'kimi_k3_modal_explored',
+      shown: 'inference_rubin_comparison_banner_shown',
+      dismissed: 'inference_rubin_comparison_banner_dismissed',
+      action: 'inference_rubin_comparison_banner_clicked',
+      properties: {
+        banner_id: 'openai-rubin-comparison',
+        scenario: '8k/1k',
+        model: 'DeepSeek-R1-0528',
+        metric: 'y_outputTputPerMw',
+      },
     },
   },
+
+  // -------------------------------------------------------------------------
+  // Agentic point-detail modal
+  // -------------------------------------------------------------------------
   {
-    id: 'github-star-modal',
+    id: 'agentx-telemetry-tutorial',
     type: 'modal',
-    trigger: { type: 'immediate' },
-    dismissal: { type: 'timed', durationMs: 7 * 24 * 60 * 60 * 1000 },
-    storageKey: 'inferencex-star-modal-dismissed',
-    permanentSuppressKey: STARRED_KEY,
-    permanentSuppressEvent: STARRED_EVENT,
+    // The detail page is chart-dense and its data arrives asynchronously;
+    // waiting lets the charts paint before the card slides in.
+    trigger: { type: 'timer', delayMs: 2500 },
+    dismissal: { type: 'permanent' },
+    storageKey: TELEMETRY_TUTORIAL_STORAGE_KEY,
     priority: 40,
-    scope: 'landing',
+    scope: 'agentic-detail',
     content: {
-      icon: Star,
-      iconClassName: 'text-yellow-500 fill-yellow-500',
-      title: 'Star InferenceX on GitHub',
-      titleZh: '在 GitHub 上为 InferenceX 加星',
+      icon: BookOpen,
+      iconClassName: 'text-brand',
+      title: 'New to these charts?',
+      titleZh: '第一次看这些图表？',
       description:
-        'Star InferenceX on GitHub to get notified when we publish new benchmark data. We update chip performance comparisons regularly — starring is the easiest way to stay in the loop and help the project grow.',
+        'The telemetry tutorial explains every chart on this page — the sequence-length distributions, the cache and queue series, the request timeline, and the per-conversation flamegraph.',
       descriptionZh:
-        '在 GitHub 上为 InferenceX 加星，以便在我们发布新基准测试数据时收到通知。我们定期更新 Chip 性能对比——加星是保持关注并帮助项目成长的最简单方式。',
-      testId: 'github-star-modal',
-      dismissLabel: 'Maybe Later',
-      dismissLabelZh: '稍后再看',
+        '遥测数据教程会讲解本页的每一张图表——序列长度分布、cache 与队列相关曲线、请求时间线，以及单会话火焰图。',
+      testId: 'telemetry-tutorial-modal',
+      // Deliberately NOT centered: a backdrop here would cover the charts the
+      // tutorial is describing. A bottom-right card leaves the page usable.
+      containerClassName: 'border-brand/40',
+      dismissLabel: 'Not now',
+      dismissLabelZh: '暂不需要',
       primaryAction: {
-        label: 'Star on GitHub',
-        labelZh: '在 GitHub 上加星',
-        icon: <GitHubIcon className="size-4" />,
+        label: 'Read the tutorial',
+        labelZh: '阅读教程',
+        icon: <ArrowRight className="size-4" />,
         onClick: () => {
-          window.open(GITHUB_REPO_URL, '_blank', 'noopener,noreferrer');
-          saveStarred();
+          window.location.href = telemetryTutorialHref();
         },
       },
-      actionClassName: 'star-button-glow',
-    },
-    analytics: {
-      shown: 'star_modal_shown',
-      dismissed: 'star_modal_dismissed',
-      action: 'star_modal_starred',
-    },
-  },
-
-  // -------------------------------------------------------------------------
-  // Landing banner
-  // -------------------------------------------------------------------------
-  {
-    id: 'kimi-k3-launch-banner',
-    type: 'banner',
-    trigger: { type: 'immediate' },
-    dismissal: { type: 'permanent' },
-    storageKey: LANDING_BANNER_STORAGE_KEY,
-    priority: 60,
-    scope: 'landing',
-    renderOnInitialLoad: true,
-    content: {
-      icon: Sparkles,
-      iconClassName: 'text-brand',
-      title: 'Kimi K3 benchmarks are live',
-      titleZh: 'Kimi K3 基准测试已上线',
-      description: 'First inference numbers across NVIDIA and AMD chips, click to explore.',
-      descriptionZh: 'NVIDIA 和 AMD Chip 的首批推理数据，点击探索。',
-      testId: 'launch-banner',
-      badge: 'New',
-      badgeZh: '最新',
-      href: '/inference?preset=kimi-k3-launch',
-      onLinkClick: () => {
-        window.location.href = '/inference?preset=kimi-k3-launch';
-      },
     },
     analytics: {
-      shown: 'launch_banner_shown',
-      dismissed: 'launch_banner_dismissed',
-      action: 'launch_banner_clicked',
-      properties: { banner_id: 'kimi-k3-launch', preset_id: 'kimi-k3-launch' },
+      shown: 'agentx_telemetry_modal_shown',
+      dismissed: 'agentx_telemetry_modal_dismissed',
+      action: 'agentx_telemetry_modal_opened',
     },
   },
 ];
diff --git a/packages/app/src/lib/nudges/types.ts b/packages/app/src/lib/nudges/types.ts
index 6605d1c9f..d3e72fccf 100644
--- a/packages/app/src/lib/nudges/types.ts
+++ b/packages/app/src/lib/nudges/types.ts
@@ -59,6 +59,35 @@ export interface NudgeRenderContext {
   dismiss: () => void;
 }
 
+export interface NudgeAnchor {
+  /**
+   * Resolve the element the coach mark should point at, in the live DOM.
+   * Return `null` when nothing suitable is on screen (chart still loading,
+   * every candidate filtered out, scrolled out of view) — the engine hides
+   * the callout rather than stranding an arrow.
+   */
+  resolve: () => Element | null;
+  /**
+   * Measure and validate a previously resolved anchor. Returning `null` tells
+   * the coach mark to discard it and run the full resolver again. Supplying
+   * this enables anchor reuse during high-frequency geometry updates.
+   */
+  getRect?: (element: Element) => DOMRect | null;
+  /**
+   * Narrow mutation observation to the live surface that owns the anchor.
+   * Defaults to `document.body` for anchors without a stable local root.
+   */
+  getMutationRoot?: (element: Element | null) => Node | null;
+  /**
+   * Clicking an element matching this selector counts as taking the nudge's
+   * action: the user found the affordance, so record engagement and stop
+   * showing it.
+   */
+  actionSelector?: string;
+  /** Extra window events that should trigger a reposition (beyond scroll/resize). */
+  repositionEvents?: readonly string[];
+}
+
 export interface NudgeContent {
   icon: ComponentType<{ className?: string }>;
   iconClassName?: string;
@@ -90,12 +119,23 @@ export interface NudgeContent {
   badge?: string;
   badgeZh?: string;
 
+  // -- Coach-mark-specific (ignored by toasts/modals/banners) --
+
+  /**
+   * Which on-screen element the callout points at. Required for `coach-mark`;
+   * without it the engine has nothing to anchor to and skips the nudge.
+   */
+  anchor?: NudgeAnchor;
+
   // -- Banner-specific (ignored by toasts/modals) --
 
   /** href for the banner link (the whole banner is clickable). */
   href?: string;
   /** Called when the banner link is clicked (for analytics). */
   onLinkClick?: () => void;
+  /** Short action label shown at the end of the banner link. */
+  linkLabel?: string;
+  linkLabelZh?: string;
 }
 
 // ---------------------------------------------------------------------------
@@ -114,7 +154,14 @@ export interface NudgeAnalyticsOverrides {
 // NudgeDefinition — a single registry entry
 // ---------------------------------------------------------------------------
 
-export type NudgeType = 'toast' | 'modal' | 'banner';
+export type NudgeType = 'toast' | 'modal' | 'banner' | 'coach-mark';
+
+/**
+ * Which NudgeEngine instance owns a nudge. One engine mounts per scope, so a
+ * new scope needs a matching `<NudgeEngine scope="…" />` on the surface it
+ * names — `agentic-detail` mounts on /inference/agentic/[id].
+ */
+export type NudgeScope = 'dashboard' | 'landing' | 'evaluation' | 'agentic-detail';
 
 export interface NudgeDefinition {
   id: string;
@@ -127,7 +174,7 @@ export interface NudgeDefinition {
   /** Higher priority wins when multiple nudges are eligible simultaneously. */
   priority: number;
   /** Which NudgeEngine instance manages this nudge. */
-  scope: 'dashboard' | 'landing' | 'evaluation';
+  scope: NudgeScope;
   /**
    * Render an immediate banner in the server response so hydration cannot
    * insert it above existing content and cause a layout shift.
@@ -142,7 +189,7 @@ export interface NudgeDefinition {
 
   /**
    * A secondary localStorage key that permanently suppresses the nudge.
-   * Example: `inferencex-starred` suppresses both star-nudge and github-star-modal.
+   * Example: `inferencex-starred` suppresses the star-nudge toast.
    */
   permanentSuppressKey?: string;
   /**
diff --git a/packages/app/src/lib/og-assets.ts b/packages/app/src/lib/og-assets.ts
index e94b875d9..fda7bbd3b 100644
--- a/packages/app/src/lib/og-assets.ts
+++ b/packages/app/src/lib/og-assets.ts
@@ -1,11 +1,13 @@
-// OG image shared assets — tile grid, logo loader, and color constants.
-// compare/compare-per-dollar/blog still carry their own copies (follow-up to migrate).
+// Shared OG tile grid, logo, font loaders, and color tokens.
 import { readFile } from 'node:fs/promises';
 import { join } from 'node:path';
 
 export const BLUE = '#0B86D1';
 export const BG = '#131416';
 export const PANEL_BG = '#0F1214';
+export const TEXT = '#EAEBEC';
+export const MUTED = '#9BA0A6';
+export const TITLE = '#FFFFFF';
 
 export const TILE_GRID: ({ file: string; rotate?: number } | null)[] = [
   { file: 'teal-chevron.png', rotate: 180 },
@@ -54,3 +56,37 @@ export function getTiles(): Promise<({ src: string; rotate?: number } | null)[]>
   }
   return tilesPromise;
 }
+
+export interface OgFontAsset {
+  name: string;
+  data: ArrayBuffer;
+  weight: 400 | 800;
+  style: 'normal';
+}
+
+let cjkFontsPromise: Promise<OgFontAsset[]> | undefined;
+export function getCjkFonts(): Promise<OgFontAsset[]> {
+  if (!cjkFontsPromise) {
+    cjkFontsPromise = Promise.all([
+      readFile(new URL('compare-og-fonts/NotoSansCJKsc-Compare-Regular.otf', import.meta.url)),
+      readFile(new URL('compare-og-fonts/NotoSansCJKsc-Compare-Bold.otf', import.meta.url)),
+    ]).then(([regular, bold]) => [
+      {
+        name: 'Noto Sans SC',
+        data: regular.buffer.slice(
+          regular.byteOffset,
+          regular.byteOffset + regular.byteLength,
+        ) as ArrayBuffer,
+        weight: 400,
+        style: 'normal',
+      },
+      {
+        name: 'Noto Sans SC',
+        data: bold.buffer.slice(bold.byteOffset, bold.byteOffset + bold.byteLength) as ArrayBuffer,
+        weight: 800,
+        style: 'normal',
+      },
+    ]);
+  }
+  return cjkFontsPromise;
+}
diff --git a/packages/app/src/lib/overview-data-cache.server.test.ts b/packages/app/src/lib/overview-data-cache.server.test.ts
new file mode 100644
index 000000000..d9a4aa914
--- /dev/null
+++ b/packages/app/src/lib/overview-data-cache.server.test.ts
@@ -0,0 +1,93 @@
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+const mocks = vi.hoisted(() => ({
+  loadRows: vi.fn(),
+  assemble: vi.fn(
+    (_rows: unknown, tier: number, engineScope: string, referenceHardware: string) => ({
+      tier,
+      engineScope,
+      referenceHardware,
+      comparisonMode: 'hardware',
+      models: [],
+    }),
+  ),
+}));
+
+vi.mock('next/cache', () => ({
+  revalidateTag: vi.fn(),
+  unstable_cache: (fn: (...args: unknown[]) => Promise<unknown>) => {
+    const entries = new Map<string, Promise<unknown>>();
+    return (...args: unknown[]) => {
+      const key = JSON.stringify(args);
+      const existing = entries.get(key);
+      if (existing) return existing;
+      const value = fn(...args);
+      entries.set(key, value);
+      return value;
+    };
+  },
+}));
+
+vi.mock('@semianalysisai/inferencex-constants', () => ({
+  DISPLAY_MODEL_TO_DB: { TestModel: ['test-model'] },
+}));
+
+vi.mock('@semianalysisai/inferencex-db/connection', () => ({ FIXTURES_MODE: false }));
+
+vi.mock('@/lib/benchmark-data.server', () => ({
+  getCachedBenchmarks: mocks.loadRows,
+  getCachedBenchmarksAsOf: vi.fn(),
+}));
+
+vi.mock('@/lib/overview-data', () => ({
+  applyOverviewHardwareRowScope: (data: unknown) => data,
+  applyOverviewRowScope: (data: unknown) => data,
+  assembleOverviewHistoricalPageData: vi.fn(),
+  assembleOverviewPageData: mocks.assemble,
+  OVERVIEW_DEFAULT_COMPARISON_MODE: 'hardware',
+  OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE: 'all',
+  OVERVIEW_DEFAULT_MODEL_SCOPE: 'default',
+  OVERVIEW_DEFAULT_REFERENCE_HARDWARE: 'b200',
+  OVERVIEW_DEFAULT_ROW_SCOPE: 'all',
+  OVERVIEW_PRIMARY_TIER: 50,
+  overviewHistoricalWindow: vi.fn(),
+  overviewModelsForScope: () => ['TestModel'],
+  overviewSnapshotDate: vi.fn(),
+}));
+
+import { getOverviewPageData } from './overview-data.server';
+
+describe('getOverviewPageData derived cache', () => {
+  beforeEach(() => {
+    vi.clearAllMocks();
+  });
+  it('canonicalizes omitted defaults and bypasses repeated source work', async () => {
+    mocks.loadRows.mockResolvedValueOnce([{ id: 1 }]);
+
+    const implicit = await getOverviewPageData();
+    const explicit = await getOverviewPageData(
+      50,
+      'community',
+      'hardware',
+      'b200',
+      'default',
+      'all',
+      'all',
+    );
+
+    expect(explicit).toBe(implicit);
+    expect(mocks.loadRows).toHaveBeenCalledTimes(1);
+    expect(mocks.assemble).toHaveBeenCalledTimes(1);
+  });
+
+  it('keeps distinct selector states in distinct entries', async () => {
+    mocks.loadRows.mockResolvedValue([{ id: 2 }]);
+
+    const tier30 = await getOverviewPageData(30);
+    const tier75 = await getOverviewPageData(75);
+
+    expect(tier30.tier).toBe(30);
+    expect(tier75.tier).toBe(75);
+    expect(mocks.loadRows).toHaveBeenCalledTimes(2);
+  });
+});
diff --git a/packages/app/src/lib/overview-data.server.test.ts b/packages/app/src/lib/overview-data.server.test.ts
index bd1c76d4b..c15a25eeb 100644
--- a/packages/app/src/lib/overview-data.server.test.ts
+++ b/packages/app/src/lib/overview-data.server.test.ts
@@ -1,5 +1,9 @@
 import { afterEach, describe, expect, it, vi } from 'vitest';
 
+vi.mock('next/cache', () => ({
+  revalidateTag: vi.fn(),
+  unstable_cache: (fn: (...args: unknown[]) => unknown) => fn,
+}));
 import type { BenchmarkRow } from './api';
 import { DISPLAY_MODEL_TO_DB } from '@semianalysisai/inferencex-constants';
 
@@ -53,7 +57,9 @@ const rows = [
 ];
 
 function selectedFrameworks(page: OverviewPageData) {
-  const summary = page.models.find((model) => model.model === Model.Qwen3_5);
+  const summary = page.models.find(
+    (model) => model.model === Model.Qwen3_5 && model.scenario === 'single_turn_8k1k',
+  );
   return {
     candidate: summary?.platforms.find(({ hardware }) => hardware === 'mi355x')?.read.config
       ?.framework,
@@ -71,6 +77,59 @@ afterEach(() => {
 });
 
 describe('getOverviewPageData engine scope forwarding', () => {
+  it('returns a renderable empty history payload when no current snapshot exists', async () => {
+    const getCachedBenchmarks = vi.fn(() => Promise.resolve([]));
+    const getCachedBenchmarksAsOf = vi.fn();
+    vi.doMock('@semianalysisai/inferencex-db/connection', () => ({ FIXTURES_MODE: false }));
+    vi.doMock('@/lib/benchmark-data.server', () => ({
+      getCachedBenchmarks,
+      getCachedBenchmarksAsOf,
+    }));
+    vi.doMock('@/lib/test-fixtures', () => ({ loadFixture: vi.fn() }));
+
+    const { getOverviewPageData } = await import('./overview-data.server');
+
+    await expect(
+      getOverviewPageData(75, 'all', '30d', 'b300', 'all', 'changed', 'priced'),
+    ).resolves.toEqual({
+      models: [],
+      tier: 75,
+      engineScope: 'all',
+      comparisonMode: '30d',
+      referenceHardware: 'b300',
+      modelScope: 'all',
+      rowScope: 'changed',
+      hardwareRowScope: 'priced',
+      unchangedRowCount: 0,
+      emptyRowCount: 0,
+      historicalWindow: null,
+    });
+    expect(getCachedBenchmarksAsOf).not.toHaveBeenCalled();
+  });
+
+  it('keeps hardware mode valid when every platform is missing', async () => {
+    const getCachedBenchmarks = vi.fn(() => Promise.resolve([]));
+    const getCachedBenchmarksAsOf = vi.fn();
+    vi.doMock('@semianalysisai/inferencex-db/connection', () => ({ FIXTURES_MODE: false }));
+    vi.doMock('@/lib/benchmark-data.server', () => ({
+      getCachedBenchmarks,
+      getCachedBenchmarksAsOf,
+    }));
+    vi.doMock('@/lib/test-fixtures', () => ({ loadFixture: vi.fn() }));
+
+    const { getOverviewPageData } = await import('./overview-data.server');
+    const page = await getOverviewPageData(50, 'community', 'hardware');
+
+    expect(page.comparisonMode).toBe('hardware');
+    expect(page.models.length).toBeGreaterThan(0);
+    expect(
+      page.models.every((model) =>
+        model.platforms.every((platform) => platform.costPerMtok === null),
+      ),
+    ).toBe(true);
+    expect(getCachedBenchmarksAsOf).not.toHaveBeenCalled();
+  });
+
   it('forwards the selected hardware reference through fixture mode', async () => {
     vi.doMock('@semianalysisai/inferencex-db/connection', () => ({ FIXTURES_MODE: true }));
     vi.doMock('@/lib/benchmark-data.server', () => ({ getCachedBenchmarks: vi.fn() }));
@@ -145,7 +204,9 @@ describe('getOverviewPageData engine scope forwarding', () => {
 
     const { getOverviewPageData } = await import('./overview-data.server');
     const page = await getOverviewPageData(50, 'community', '30d');
-    const qwen = page.models.find((model) => model.model === Model.Qwen3_5);
+    const qwen = page.models.find(
+      (model) => model.model === Model.Qwen3_5 && model.scenario === 'single_turn_8k1k',
+    );
     const mi355x = qwen?.platforms.find(({ hardware }) => hardware === 'mi355x');
     const b300 = qwen?.platforms.find(({ hardware }) => hardware === 'b300');
 
diff --git a/packages/app/src/lib/overview-data.server.ts b/packages/app/src/lib/overview-data.server.ts
index a2696a55d..5dd8787de 100644
--- a/packages/app/src/lib/overview-data.server.ts
+++ b/packages/app/src/lib/overview-data.server.ts
@@ -2,13 +2,19 @@ import { DISPLAY_MODEL_TO_DB } from '@semianalysisai/inferencex-constants';
 import { FIXTURES_MODE } from '@semianalysisai/inferencex-db/connection';
 
 import type { BenchmarkRow } from '@/lib/api';
+import { benchmarkCurveDate } from '@/lib/benchmark-run-selection';
+import { cachedDerivedData } from '@/lib/api-cache';
 import { getCachedBenchmarks, getCachedBenchmarksAsOf } from '@/lib/benchmark-data.server';
 import type { Model } from '@/lib/data-mappings';
 import {
+  applyOverviewHardwareRowScope,
+  applyOverviewRowScope,
   assembleOverviewHistoricalPageData,
   assembleOverviewPageData,
   OVERVIEW_DEFAULT_COMPARISON_MODE,
+  OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
   OVERVIEW_DEFAULT_MODEL_SCOPE,
+  OVERVIEW_DEFAULT_ROW_SCOPE,
   OVERVIEW_PRIMARY_TIER,
   OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
   overviewHistoricalWindow,
@@ -16,9 +22,11 @@ import {
   overviewSnapshotDate,
   type OverviewComparisonMode,
   type OverviewEngineScope,
+  type OverviewHardwareRowScope,
   type OverviewModelScope,
   type OverviewPageData,
   type OverviewReferenceHardware,
+  type OverviewRowScope,
   type OverviewTier,
 } from '@/lib/overview-data';
 import { loadFixture } from '@/lib/test-fixtures';
@@ -37,13 +45,19 @@ async function loadRowsByModel(
   return Object.fromEntries(entries);
 }
 
-export async function getOverviewPageData(
+async function buildOverviewPageData(
   tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
   engineScope: OverviewEngineScope = 'community',
   comparisonMode: OverviewComparisonMode = OVERVIEW_DEFAULT_COMPARISON_MODE,
   referenceHardware: OverviewReferenceHardware = OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
   modelScope: OverviewModelScope = OVERVIEW_DEFAULT_MODEL_SCOPE,
+  rowScope: OverviewRowScope = OVERVIEW_DEFAULT_ROW_SCOPE,
+  hardwareRowScope: OverviewHardwareRowScope = OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
 ): Promise<OverviewPageData> {
+  // Each filter is a no-op outside its own comparison mode, so both run on
+  // every path and the active one is whichever the matrix is actually in.
+  const scopeRows = (data: OverviewPageData): OverviewPageData =>
+    applyOverviewHardwareRowScope(applyOverviewRowScope(data, rowScope), hardwareRowScope);
   const models = overviewModelsForScope(modelScope);
   // Synthetic rows go through the same assemblers as the live path, so a
   // contract drift breaks fixture tests instead of stranding the page.
@@ -51,12 +65,14 @@ export async function getOverviewPageData(
     ? loadFixture<Record<string, BenchmarkRow[]>>('overview-rows')
     : await loadRowsByModel(models, getCachedBenchmarks);
   if (comparisonMode === 'hardware') {
-    return assembleOverviewPageData(
-      currentRowsByModel,
-      tier,
-      engineScope,
-      referenceHardware,
-      modelScope,
+    return scopeRows(
+      assembleOverviewPageData(
+        currentRowsByModel,
+        tier,
+        engineScope,
+        referenceHardware,
+        modelScope,
+      ),
     );
   }
 
@@ -76,14 +92,17 @@ export async function getOverviewPageData(
   const snapshotDate = overviewSnapshotDate(snapshotRows, engineScope);
   if (snapshotDate === null) {
     return {
-      ...assembleOverviewPageData(
-        currentRowsByModel,
-        tier,
-        engineScope,
-        referenceHardware,
-        modelScope,
-      ),
+      models: [],
+      tier,
+      engineScope,
       comparisonMode,
+      referenceHardware,
+      modelScope,
+      rowScope,
+      hardwareRowScope,
+      unchangedRowCount: 0,
+      emptyRowCount: 0,
+      historicalWindow: null,
     };
   }
 
@@ -94,17 +113,49 @@ export async function getOverviewPageData(
   const baselineRowsByModel = Object.fromEntries(
     Object.entries(unboundedBaselineRows).map(([model, rows]) => [
       model,
-      rows.filter((row) => row.date >= window.earliestDate && row.date <= window.targetDate),
+      rows.filter(
+        (row) =>
+          benchmarkCurveDate(row) >= window.earliestDate &&
+          benchmarkCurveDate(row) <= window.targetDate,
+      ),
     ]),
   );
 
-  return assembleOverviewHistoricalPageData(
-    currentRowsByModel,
-    baselineRowsByModel,
-    window,
+  return scopeRows(
+    assembleOverviewHistoricalPageData(
+      currentRowsByModel,
+      baselineRowsByModel,
+      window,
+      tier,
+      engineScope,
+      referenceHardware,
+      modelScope,
+    ),
+  );
+}
+
+// v2: AgentX rows group above 8K/1K rows (#918). The assembled models array is
+// stored in the derived cache, so the display-order change needs a new key —
+// v1 entries keep serving the old order until they expire otherwise.
+const getCachedOverviewPageData = cachedDerivedData(buildOverviewPageData, 'overview-page-v2');
+
+export function getOverviewPageData(
+  tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
+  engineScope: OverviewEngineScope = 'community',
+  comparisonMode: OverviewComparisonMode = OVERVIEW_DEFAULT_COMPARISON_MODE,
+  referenceHardware: OverviewReferenceHardware = OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
+  modelScope: OverviewModelScope = OVERVIEW_DEFAULT_MODEL_SCOPE,
+  rowScope: OverviewRowScope = OVERVIEW_DEFAULT_ROW_SCOPE,
+  hardwareRowScope: OverviewHardwareRowScope = OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
+): Promise<OverviewPageData> {
+  const loader = FIXTURES_MODE ? buildOverviewPageData : getCachedOverviewPageData;
+  return loader(
     tier,
     engineScope,
+    comparisonMode,
     referenceHardware,
     modelScope,
+    rowScope,
+    hardwareRowScope,
   );
 }
diff --git a/packages/app/src/lib/overview-data.test.ts b/packages/app/src/lib/overview-data.test.ts
index eae5bf74a..a46307b84 100644
--- a/packages/app/src/lib/overview-data.test.ts
+++ b/packages/app/src/lib/overview-data.test.ts
@@ -13,9 +13,15 @@ import {
   Precision,
 } from './data-mappings';
 import {
+  applyOverviewHardwareRowScope,
+  applyOverviewRowScope,
   assembleOverviewHistoricalPageData,
   assembleOverviewPageData,
   buildOverviewModelSummary,
+  OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
+  OVERVIEW_DEFAULT_ROW_SCOPE,
+  overviewRowHasAnyCost,
+  overviewRowHasHistoricalChange,
   overviewCostPerMtok,
   overviewHistoricalWindow,
   overviewScenarioForModel,
@@ -23,8 +29,10 @@ import {
   overviewTierEvidenceDate,
   resolveOverviewComparisonMode,
   resolveOverviewEngineScope,
+  resolveOverviewHardwareRowScope,
   resolveOverviewModelScope,
   resolveOverviewReferenceHardware,
+  resolveOverviewRowScope,
   resolveOverviewTier,
   type OverviewModelSummary,
   type OverviewPageData,
@@ -1167,17 +1175,20 @@ describe('overview platform selection', () => {
     });
   });
 
-  it('restricts AgentX points to the E2E frontier on total throughput', () => {
-    // The slower-E2E point wins on output tokens but loses on total tokens, so
-    // the total-token frontier drops it and the tier read becomes unreachable.
+  it('reads AgentX tiers from every measured point, not just the E2E frontier', () => {
+    // The second point is dominated on total tokens (8100 < 9000) and slower on
+    // E2E (25 > 20). Until #736 the E2E frontier dropped it, leaving only the
+    // interactivity-40 point, which cannot reach the tier-50 read — so the tier
+    // came back null. That gating was removed deliberately, so the point now
+    // stands and the read lands on it exactly.
     const summary = buildOverviewModelSummary(Model.GLM_5_2, [
       agenticRow(40, 20, 9000, 500, { hardware: 'b200', conc: 8 }),
       agenticRow(50, 25, 8100, 900, { hardware: 'b200', conc: 12 }),
     ]);
 
     const b200 = summary.platforms.find(({ hardware }) => hardware === 'b200')!;
-    expect(b200.read.value).toBeNull();
-    expect(b200.missingReason).toBe('cannot_reach_at_tier');
+    expect(b200.read.value).toBe(8100);
+    expect(b200.missingReason).toBeNull();
   });
 
   it('reports scenario-level missing coverage when AgentX rows lack usable P90 metrics', () => {
@@ -1248,7 +1259,7 @@ describe('tier-parameterized overview', () => {
     );
 
     const pair = headlinePairOf(
-      page.models.find((m) => m.model === Model.Qwen3_5)!,
+      page.models.find((m) => m.model === Model.Qwen3_5 && m.scenario === 'single_turn_8k1k')!,
       'mi355x-vs-b200',
     );
     expect(pair?.candidate.read).toMatchObject({ tier: 100, value: 3600 });
@@ -1271,7 +1282,7 @@ describe('tier-parameterized overview', () => {
 
     const at50 = headlinePairOf(
       assembleOverviewPageData({ [Model.Qwen3_5]: rows }).models.find(
-        (m) => m.model === Model.Qwen3_5,
+        (m) => m.model === Model.Qwen3_5 && m.scenario === 'single_turn_8k1k',
       )!,
       'mi355x-vs-b200',
     );
@@ -1279,7 +1290,7 @@ describe('tier-parameterized overview', () => {
 
     const at30 = headlinePairOf(
       assembleOverviewPageData({ [Model.Qwen3_5]: rows }, 30).models.find(
-        (m) => m.model === Model.Qwen3_5,
+        (m) => m.model === Model.Qwen3_5 && m.scenario === 'single_turn_8k1k',
       )!,
       'mi355x-vs-b200',
     );
@@ -1301,7 +1312,7 @@ describe('tier-parameterized overview', () => {
           ],
         },
         tier,
-      ).models.find((m) => m.model === Model.Qwen3_5)!;
+      ).models.find((m) => m.model === Model.Qwen3_5 && m.scenario === 'single_turn_8k1k')!;
 
     const at50 = headlinePairOf(page(), 'mi355x-vs-b200');
     expect(at50?.candidate.precision).toBe(Precision.FP4);
@@ -1325,18 +1336,29 @@ describe('assembleOverviewPageData over the overview-rows fixture', () => {
       overviewRowsFixture as unknown as Record<string, BenchmarkRow[]>,
     );
 
-    // Curated scenarios: DeepSeek, MiniMax and Qwen each get both rows, Kimi
+    // Curated scenarios: DeepSeek, MiniMax and Qwen3.5 each get both rows, Kimi
     // K3 and GLM are AgentX-only. Kimi K2.5 is absent — deprecated models are
     // not default models, and the matrix is built from DEFAULT_MODELS.
+    // Qwen3.8-Flash-Next is curated AgentX-only, like Kimi K3 and GLM: the
+    // model is benchmarked on agentic traces, so it must not claim a
+    // fixed-sequence row it will never fill. AgentX rows group above the 8K/1K
+    // rows, each group in MODEL_CONFIG declaration order.
     expect(page.models.map((m) => `${m.model}/${m.scenario}`)).toEqual([
-      `${Model.DeepSeek_V4_Pro}/single_turn_8k1k`,
       `${Model.DeepSeek_V4_Pro}/agentx`,
       `${Model.Kimi_K3}/agentx`,
-      `${Model.MiniMax_M3}/single_turn_8k1k`,
       `${Model.MiniMax_M3}/agentx`,
       `${Model.GLM_5_2}/agentx`,
-      `${Model.Qwen3_5}/single_turn_8k1k`,
       `${Model.Qwen3_5}/agentx`,
+      `${Model.Qwen3_8_Flash_Next}/agentx`,
+      `${Model.DeepSeek_V4_Pro}/single_turn_8k1k`,
+      `${Model.MiniMax_M3}/single_turn_8k1k`,
+      `${Model.Qwen3_5}/single_turn_8k1k`,
+      // Fork-only default models (dsv2lite, gemma4 — see
+      // .claude/skills/fork-changelog/SKILL.md) have no rows in this fixture,
+      // so the real builder serves them as empty-state 8k1k cells, in
+      // MODEL_CONFIG declaration order.
+      `${Model.DeepSeek_Coder_V2_Lite}/single_turn_8k1k`,
+      `${Model.Gemma4_31B}/single_turn_8k1k`,
     ]);
     expect(page.models.length).toBeGreaterThan(DEFAULT_MODELS.size);
     expect(page).not.toHaveProperty('datasetThroughDate');
@@ -1346,7 +1368,9 @@ describe('assembleOverviewPageData over the overview-rows fixture', () => {
     // therefore have no exact @50 read; GB200's independent FP8 remains visible.
     // GB300's points are single-node and multi-node aggregate deployments, so
     // they must not be interpolated into one synthetic serving curve.
-    const deepseek = page.models.find((m) => m.model === Model.DeepSeek_V4_Pro)!;
+    const deepseek = page.models.find(
+      (m) => m.model === Model.DeepSeek_V4_Pro && m.scenario === 'single_turn_8k1k',
+    )!;
     const dsB300 = headlinePairOf(deepseek, 'b300-vs-b200')!;
     expect(dsB300.baseline.read.value).toBeCloseTo(8101.968);
     expect(dsB300.baseline.costPerMtok).toBeCloseTo(
@@ -1403,7 +1427,9 @@ describe('assembleOverviewPageData over the overview-rows fixture', () => {
 
     // MiniMax: the platform result remains visible when B200 has no 8K/1K data
     // — priced, but with no percentage baseline (the UI's ∞ badge state).
-    const minimax = page.models.find((m) => m.model === Model.MiniMax_M3)!;
+    const minimax = page.models.find(
+      (m) => m.model === Model.MiniMax_M3 && m.scenario === 'single_turn_8k1k',
+    )!;
     const mmGb300 = headlinePairOf(minimax, 'gb300-vs-b200')!;
     expect(mmGb300.baseline.missingReason).toBe('no_scenario_data');
     expect(mmGb300.candidate.read.value).toBe(6510);
@@ -1414,7 +1440,9 @@ describe('assembleOverviewPageData over the overview-rows fixture', () => {
     expect(mmGb300.candidate.costVsReferencePct).toBeNull();
 
     // Qwen: MI355X independently falls back to FP8 while B200 and B300 use FP4.
-    const qwen = page.models.find((m) => m.model === Model.Qwen3_5)!;
+    const qwen = page.models.find(
+      (m) => m.model === Model.Qwen3_5 && m.scenario === 'single_turn_8k1k',
+    )!;
     const qwenMi = headlinePairOf(qwen, 'mi355x-vs-b200')!;
     expect(qwenMi.candidate.precision).toBe(Precision.FP8);
     expect(qwenMi.candidate.read.value).toBe(6688);
@@ -1516,3 +1544,183 @@ describe('overview model scope', () => {
     expect(page.modelScope).toBe('all');
   });
 });
+
+describe('overview row scope', () => {
+  const window = {
+    key: '30d',
+    snapshotDate: '2026-08-03',
+    targetDate: '2026-07-04',
+    earliestDate: '2026-06-04',
+  } as const;
+  // Only Qwen3.5 posts rows, so exactly one summary can be comparable and every
+  // other default row is dead weight in the 30-day matrix.
+  const currentRows = {
+    [Model.Qwen3_5]: frontier([12000, 10000, 8000, 6000], {
+      hardware: 'mi355x',
+      framework: 'sglang',
+      date: '2026-08-01',
+      precision: Precision.FP4,
+    }),
+  };
+  const baselineRows = {
+    [Model.Qwen3_5]: frontier([9600, 8000, 6400, 4800], {
+      hardware: 'mi355x',
+      framework: 'sglang',
+      date: '2026-06-20',
+      precision: Precision.FP4,
+    }),
+  };
+
+  function historyPage() {
+    return assembleOverviewHistoricalPageData(currentRows, baselineRows, window, 50, 'community');
+  }
+
+  it("resolves 'changed' as opt-in and defaults everything else to all rows", () => {
+    expect(resolveOverviewRowScope('changed')).toBe('changed');
+    expect(resolveOverviewRowScope(['changed', 'all'])).toBe('changed');
+    expect(resolveOverviewRowScope('all')).toBe('all');
+    expect(resolveOverviewRowScope('moved')).toBe('all');
+    expect(resolveOverviewRowScope('')).toBe('all');
+    expect(resolveOverviewRowScope(undefined)).toBe('all');
+  });
+
+  it('shows every row by default, so no current cost is hidden without asking', () => {
+    const page = historyPage();
+    const scoped = applyOverviewRowScope(page, OVERVIEW_DEFAULT_ROW_SCOPE);
+
+    expect(OVERVIEW_DEFAULT_ROW_SCOPE).toBe('all');
+    expect(scoped.rowScope).toBe('all');
+    expect(scoped.models).toHaveLength(page.models.length);
+    expect(scoped.unchangedRowCount).toBe(page.models.length - 1);
+  });
+
+  it("keeps only rows with a comparable platform under opt-in scope 'changed'", () => {
+    const page = historyPage();
+    const scoped = applyOverviewRowScope(page, 'changed');
+
+    expect(scoped.rowScope).toBe('changed');
+    expect(scoped.models).toHaveLength(1);
+    expect(scoped.models[0]?.model).toBe(Model.Qwen3_5);
+    expect(scoped.unchangedRowCount).toBe(page.models.length - 1);
+    expect(
+      scoped.models[0]?.platforms.some((p) => p.historicalComparison?.status === 'comparable'),
+    ).toBe(true);
+  });
+
+  it('never drops a row that still prices a platform without a 30-day baseline', () => {
+    const page = historyPage();
+    const dropped = page.models.filter((m) => !overviewRowHasHistoricalChange(m));
+    const stillPriced = dropped.filter((m) => m.platforms.some((p) => p.costPerMtok !== null));
+
+    // The default must keep them; only the opt-in scope may remove them.
+    expect(applyOverviewRowScope(page, 'all').models).toEqual(expect.arrayContaining(stillPriced));
+    for (const summary of stillPriced) {
+      expect(applyOverviewRowScope(page, 'changed').models).not.toContain(summary);
+    }
+  });
+
+  it('filters nothing in hardware mode but still carries the reader’s answer', () => {
+    const page = assembleOverviewPageData(currentRows, 50, 'community');
+    const scoped = applyOverviewRowScope(page, 'changed');
+
+    expect(scoped.models).toHaveLength(page.models.length);
+    // Carried, not cleared: the page rebuilds its own URL from this data, so
+    // clearing it here is what dropped the 30-day filter on a tab switch back.
+    expect(scoped.rowScope).toBe('changed');
+    expect(scoped.unchangedRowCount).toBe(0);
+  });
+
+  it('falls back to the full matrix when no row moved, rather than rendering nothing', () => {
+    const page = assembleOverviewHistoricalPageData({}, {}, window, 50, 'community');
+    const scoped = applyOverviewRowScope(page, 'changed');
+
+    expect(scoped.models).toHaveLength(page.models.length);
+    expect(scoped.rowScope).toBe('all');
+    expect(scoped.unchangedRowCount).toBe(0);
+  });
+});
+
+describe('overview hardware row scope', () => {
+  const window = {
+    key: '30d',
+    snapshotDate: '2026-08-03',
+    targetDate: '2026-07-04',
+    earliestDate: '2026-06-04',
+  } as const;
+  // Only Qwen3.5 posts rows, and only on MI355X. That makes it the row the
+  // filter has to reason about carefully: it prices a platform while the
+  // default B200 reference prices nothing, so it has a cost but no delta.
+  const currentRows = {
+    [Model.Qwen3_5]: frontier([12000, 10000, 8000, 6000], {
+      hardware: 'mi355x',
+      framework: 'sglang',
+      date: '2026-08-01',
+      precision: Precision.FP4,
+    }),
+  };
+
+  function hardwarePage() {
+    return assembleOverviewPageData(currentRows, 50, 'community');
+  }
+
+  it("resolves 'priced' as opt-in and defaults everything else to all rows", () => {
+    expect(resolveOverviewHardwareRowScope('priced')).toBe('priced');
+    expect(resolveOverviewHardwareRowScope(['priced', 'all'])).toBe('priced');
+    expect(resolveOverviewHardwareRowScope('all')).toBe('all');
+    expect(resolveOverviewHardwareRowScope('empty')).toBe('all');
+    expect(resolveOverviewHardwareRowScope('')).toBe('all');
+    expect(resolveOverviewHardwareRowScope(undefined)).toBe('all');
+  });
+
+  it('shows every row by default, so no current cost is hidden without asking', () => {
+    const page = hardwarePage();
+    const scoped = applyOverviewHardwareRowScope(page, OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE);
+
+    expect(OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE).toBe('all');
+    expect(scoped.hardwareRowScope).toBe('all');
+    expect(scoped.models).toHaveLength(page.models.length);
+    expect(scoped.emptyRowCount).toBe(page.models.filter((m) => !overviewRowHasAnyCost(m)).length);
+  });
+
+  it("keeps only rows pricing a platform under opt-in scope 'priced'", () => {
+    const page = hardwarePage();
+    const priced = page.models.filter(overviewRowHasAnyCost);
+    const scoped = applyOverviewHardwareRowScope(page, 'priced');
+
+    expect(priced.length).toBeGreaterThan(0);
+    expect(priced.length).toBeLessThan(page.models.length);
+    expect(scoped.hardwareRowScope).toBe('priced');
+    expect(scoped.models).toEqual(priced);
+    expect(scoped.emptyRowCount).toBe(page.models.length - priced.length);
+  });
+
+  it('keeps a row priced only on a platform the reference never reached', () => {
+    // The tempting predicate — "has a delta against the reference" — would drop
+    // this row, deleting a cost that appears nowhere else, and would hide a
+    // different set of rows for every reference the reader picks.
+    const scoped = applyOverviewHardwareRowScope(hardwarePage(), 'priced');
+    const kept = scoped.models.filter((m) => m.model === Model.Qwen3_5);
+
+    expect(kept.length).toBeGreaterThan(0);
+    expect(kept.some((m) => m.platforms.some((p) => p.costPerMtok !== null))).toBe(true);
+    expect(kept.every((m) => m.platforms.every((p) => p.costVsReferencePct === null))).toBe(true);
+  });
+
+  it('filters nothing in history mode but still carries the reader’s answer', () => {
+    const page = assembleOverviewHistoricalPageData(currentRows, {}, window, 50, 'community');
+    const scoped = applyOverviewHardwareRowScope(page, 'priced');
+
+    expect(scoped.models).toHaveLength(page.models.length);
+    expect(scoped.hardwareRowScope).toBe('priced');
+    expect(scoped.emptyRowCount).toBe(0);
+  });
+
+  it('falls back to the full matrix when no row prices anything', () => {
+    const page = assembleOverviewPageData({}, 50, 'community');
+    const scoped = applyOverviewHardwareRowScope(page, 'priced');
+
+    expect(scoped.models).toHaveLength(page.models.length);
+    expect(scoped.hardwareRowScope).toBe('all');
+    expect(scoped.emptyRowCount).toBe(0);
+  });
+});
diff --git a/packages/app/src/lib/overview-data.ts b/packages/app/src/lib/overview-data.ts
index 1c9c8df13..f6be441bf 100644
--- a/packages/app/src/lib/overview-data.ts
+++ b/packages/app/src/lib/overview-data.ts
@@ -1,6 +1,5 @@
 import { resolveFrameworkPartLabel } from '@semianalysisai/inferencex-constants';
 
-import { restrictAgenticPointsToE2eFrontier } from './agentic-frontier';
 import type { BenchmarkRow } from './api';
 import { rowToAggDataEntry } from './benchmark-transform';
 import { buildAvailabilityHwKey } from './chart-utils';
@@ -16,6 +15,11 @@ import {
   type CategoryTag,
 } from './data-mappings';
 import { frameworkFamily } from './framework-family';
+import {
+  benchmarkCurveDate,
+  benchmarkCurveRunStartedAt,
+  benchmarkCurveWorkflowRunId,
+} from './benchmark-run-selection';
 import {
   computeTierReads,
   singleTurnInteractivity,
@@ -47,9 +51,29 @@ export type OverviewComparisonMode = 'hardware' | OverviewHistoryWindowKey;
 export const OVERVIEW_DEFAULT_COMPARISON_MODE: OverviewComparisonMode = 'hardware';
 export type OverviewModelScope = 'default' | 'all';
 export const OVERVIEW_DEFAULT_MODEL_SCOPE: OverviewModelScope = 'default';
+/** History mode only: `changed` narrows the matrix to rows that moved in the
+ *  window. It is opt-in — the default shows every row, because an unchanged row
+ *  still carries current cost the reader came to audit. Ignored in hardware
+ *  mode, where every row carries a comparison. */
+export type OverviewRowScope = 'changed' | 'all';
+export const OVERVIEW_DEFAULT_ROW_SCOPE: OverviewRowScope = 'all';
+/** Hardware mode only: `priced` drops rows that quote no platform at all, which
+ *  carry neither a cost nor a comparison and exist purely to say "not measured".
+ *  Deliberately not "rows without a delta against the reference": that would
+ *  delete rows pricing three chips just because the reference happens to miss
+ *  this scenario, and the count would swing with the chosen reference. */
+export type OverviewHardwareRowScope = 'priced' | 'all';
+export const OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE: OverviewHardwareRowScope = 'all';
 export type OverviewScenario = 'single_turn_8k1k' | 'agentx';
-/** Row order within a model: the single-turn workload first, AgentX below it. */
+/** Canonical scenario list. Its order decides headline selection on stat-led
+ *  surfaces (single-turn first), NOT matrix row order — the matrix groups
+ *  AgentX rows first via OVERVIEW_SCENARIO_ROW_ORDER below. */
 export const OVERVIEW_SCENARIOS = ['single_turn_8k1k', 'agentx'] as const;
+/** Matrix display order: the AgentX rows lead and the fixed-sequence 8K/1K
+ *  rows follow, so the workload the overview headlines is not buried under
+ *  legacy single-turn rows. Deliberately separate from OVERVIEW_SCENARIOS so
+ *  /run and /rankings keep quoting single-turn where it exists. */
+export const OVERVIEW_SCENARIO_ROW_ORDER = ['agentx', 'single_turn_8k1k'] as const;
 
 export function resolveOverviewEngineScope(
   raw: string | string[] | undefined,
@@ -85,9 +109,25 @@ export function resolveOverviewModelScope(
   return candidate === 'all' ? 'all' : OVERVIEW_DEFAULT_MODEL_SCOPE;
 }
 
+export function resolveOverviewRowScope(
+  raw: string | readonly string[] | undefined,
+): OverviewRowScope {
+  const candidate = Array.isArray(raw) ? raw[0] : raw;
+  return candidate === 'changed' ? 'changed' : OVERVIEW_DEFAULT_ROW_SCOPE;
+}
+
+export function resolveOverviewHardwareRowScope(
+  raw: string | readonly string[] | undefined,
+): OverviewHardwareRowScope {
+  const candidate = Array.isArray(raw) ? raw[0] : raw;
+  return candidate === 'priced' ? 'priced' : OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE;
+}
+
 // Note (wenyao): row order is a contract — defaults, then maintenance, then
 // deprecated, each in MODEL_CONFIG declaration order; the overview e2e asserts
-// inactive rows always sit below the default rows.
+// inactive rows always sit below the default rows. Within each category band
+// the matrix additionally groups AgentX rows above 8K/1K rows (see
+// sortOverviewSummaries), which never reorders across bands.
 export function overviewModelsForScope(scope: OverviewModelScope): Model[] {
   return scope === 'all'
     ? [...DEFAULT_MODELS, ...MAINTENANCE_MODELS, ...DEPRECATED_MODELS]
@@ -198,6 +238,15 @@ export interface OverviewPageData {
   comparisonMode: OverviewComparisonMode;
   referenceHardware: OverviewReferenceHardware;
   modelScope: OverviewModelScope;
+  rowScope: OverviewRowScope;
+  hardwareRowScope: OverviewHardwareRowScope;
+  /** Rows with no 30-day change, counted over the full matrix regardless of the
+   *  active scope, so the toggle can name the same number in both directions.
+   *  Zero outside history mode and whenever filtering would change nothing. */
+  unchangedRowCount: number;
+  /** Rows quoting no platform at all, counted the same way for hardware mode.
+   *  Zero outside hardware mode and whenever filtering would change nothing. */
+  emptyRowCount: number;
   historicalWindow: OverviewHistoricalWindow | null;
 }
 
@@ -241,7 +290,7 @@ export function overviewSnapshotDate(
           scenarios.includes(scenario)
         );
       })
-      .map((row) => row.date);
+      .map(benchmarkCurveDate);
   });
   return dates.length === 0 ? null : (dates.toSorted().at(-1) ?? null);
 }
@@ -293,7 +342,9 @@ export function overviewScenarioForModel(
     return 'single_turn_8k1k';
   }
   if (rows.some((row) => row.benchmark_type === 'agentic_traces')) return 'agentx';
-  return model === Model.Kimi_K3 || model === Model.GLM_5_2 ? 'agentx' : 'single_turn_8k1k';
+  return model === Model.Kimi_K3 || model === Model.GLM_5_2 || model === Model.Qwen3_8_Flash_Next
+    ? 'agentx'
+    : 'single_turn_8k1k';
 }
 
 /**
@@ -308,6 +359,7 @@ const OVERVIEW_MODEL_SCENARIOS: Partial<Record<Model, readonly OverviewScenario[
   [Model.Qwen3_5]: ['single_turn_8k1k', 'agentx'],
   [Model.Kimi_K3]: ['agentx'],
   [Model.GLM_5_2]: ['agentx'],
+  [Model.Qwen3_8_Flash_Next]: ['agentx'],
 };
 
 /** The scenarios this model gets a row for. Unlisted models keep the single
@@ -324,6 +376,24 @@ export function overviewScenariosForModel(
       OVERVIEW_SCENARIOS.filter((scenario) => curated.includes(scenario));
 }
 
+/**
+ * The single scenario a stat-led SEO surface quotes for a model: the first
+ * curated scenario that actually has rows, then the first curated scenario.
+ * Keeps /run and /rankings from headlining a workload the overview matrix
+ * does not show for the model (Kimi K3 and GLM 5.2 stay on AgentX even when
+ * single-turn rows exist).
+ */
+export function overviewHeadlineScenarioForModel(
+  model: Model,
+  rows: readonly BenchmarkRow[] = [],
+): OverviewScenario {
+  const scenarios = overviewScenariosForModel(model, rows);
+  return (
+    scenarios.find((scenario) => rows.some((row) => overviewScenarioOfRow(row) === scenario)) ??
+    scenarios[0]
+  );
+}
+
 function overviewEngineRows(
   rows: readonly BenchmarkRow[],
   engineScope: OverviewEngineScope,
@@ -388,21 +458,26 @@ function buildConfigs(
   const configs: OverviewConfigResult[] = [];
   for (const [key, configRows] of rowsByConfig) {
     const latestDate = configRows.reduce(
-      (latest, row) => (row.date > latest ? row.date : latest),
-      configRows[0].date,
+      (latest, row) => (benchmarkCurveDate(row) > latest ? benchmarkCurveDate(row) : latest),
+      benchmarkCurveDate(configRows[0]),
     );
-    let latestRows = configRows.filter((row) => row.date === latestDate);
-    if (scenario === 'agentx' && latestRows.some((row) => row.workflow_run_id !== undefined)) {
+    let latestRows = configRows.filter((row) => benchmarkCurveDate(row) === latestDate);
+    if (
+      scenario === 'agentx' &&
+      latestRows.some((row) => benchmarkCurveWorkflowRunId(row) !== undefined)
+    ) {
       const winningRow = latestRows.reduce((winner, row) => {
-        const startedAt = row.run_started_at ?? '';
-        const winnerStartedAt = winner.run_started_at ?? '';
+        const startedAt = benchmarkCurveRunStartedAt(row) ?? '';
+        const winnerStartedAt = benchmarkCurveRunStartedAt(winner) ?? '';
         if (startedAt !== winnerStartedAt) return startedAt > winnerStartedAt ? row : winner;
-        return (row.workflow_run_id ?? Number.NEGATIVE_INFINITY) >
-          (winner.workflow_run_id ?? Number.NEGATIVE_INFINITY)
+        return (benchmarkCurveWorkflowRunId(row) ?? Number.NEGATIVE_INFINITY) >
+          (benchmarkCurveWorkflowRunId(winner) ?? Number.NEGATIVE_INFINITY)
           ? row
           : winner;
       });
-      latestRows = latestRows.filter((row) => row.workflow_run_id === winningRow.workflow_run_id);
+      latestRows = latestRows.filter(
+        (row) => benchmarkCurveWorkflowRunId(row) === benchmarkCurveWorkflowRunId(winningRow),
+      );
     }
     const config = buildConfigResult(model, scenario, latestRows[0].precision, key, latestRows);
     if (config) configs.push(config);
@@ -468,7 +543,13 @@ function nonComparableAsMissing(
   if (read === undefined) return nullTierRead(tier);
   return isInRangeTierRead(read)
     ? read
-    : { ...read, value: null, estimated: false, evidenceDate: null, evidenceTopologies: [] };
+    : {
+        ...read,
+        value: null,
+        estimated: false,
+        evidenceDate: null,
+        evidenceTopologies: [],
+      };
 }
 
 function configPriorityIndex(config: OverviewConfigView): number {
@@ -634,12 +715,12 @@ function buildAgenticTierReads(rows: readonly BenchmarkRow[]): TcoTierRead[] {
         interactivity,
         e2eLatency,
         throughput: totalThroughput,
-        date: row.date,
+        date: benchmarkCurveDate(row),
         evidenceLabel: topologyEvidence(row),
       },
     ];
   });
-  return computeTierReads(restrictAgenticPointsToE2eFrontier(points), OVERVIEW_TIERS);
+  return computeTierReads(points, OVERVIEW_TIERS);
 }
 
 /** Single-turn 8K/1K: frontier points at the chart's stored interactivity,
@@ -655,7 +736,7 @@ function buildSingleTurnTierReads(rows: readonly BenchmarkRow[]): TcoTierRead[]
       {
         interactivity,
         throughput: totalTput * deployedGpuFactor(row),
-        date: row.date,
+        date: benchmarkCurveDate(row),
         evidenceLabel: topologyEvidence(row),
       },
     ];
@@ -744,11 +825,53 @@ export function buildOverviewModelSummary(
   };
 }
 
-/** DEFAULT_MODELS fixes the row order, and a model benchmarked on both
- *  scenarios contributes one row per scenario; a rowless model still renders
- *  all platforms with missing reasons. Live and fixture paths both feed this.
- *  Scenario presence reads the unscoped rows so switching the engine scope
- *  changes cell contents, never the shape of the matrix. */
+/**
+ * Tier reads for one hardware SKU without the OVERVIEW_HARDWARE restriction.
+ * Same config building and read selection as the overview matrix, so a /run
+ * page for H100-class hardware quotes the number the overview would show if
+ * its matrix listed that SKU.
+ */
+export function overviewHardwareTierReader(
+  model: Model,
+  rows: BenchmarkRow[],
+  scenario: OverviewScenario = overviewScenarioForModel(model, rows),
+  engineScope: OverviewEngineScope = 'community',
+): (
+  hardware: string,
+  tier: OverviewTier,
+) => { read: OverviewTierRead; costPerMtok: number | null } {
+  const scopedRows = overviewEngineRows(rows, engineScope);
+  const scenarioRows = overviewScenarioRows(scenario, scopedRows);
+  const configs = buildConfigs(model, scenario, scenarioRows);
+  return (hardware, tier) => {
+    const read = selectPlatformRead(configs, hardware, tier);
+    return { read, costPerMtok: overviewCostPerMtok(hardware, read.value) };
+  };
+}
+
+const overviewCategoryRank = (category: CategoryTag): number =>
+  category === 'default' ? 0 : category === 'maintenance' ? 1 : 2;
+const overviewScenarioRank = (scenario: OverviewScenario): number =>
+  OVERVIEW_SCENARIO_ROW_ORDER.indexOf(scenario);
+
+/** Category bands stay in place (defaults, then maintenance, then deprecated)
+ *  while AgentX rows lead within each band — 8K/1K rows follow — and each
+ *  (band, scenario) group keeps MODEL_CONFIG declaration order. The sort is
+ *  stable, so ties never shuffle. */
+function sortOverviewSummaries(summaries: OverviewModelSummary[]): OverviewModelSummary[] {
+  return summaries.toSorted(
+    (a, b) =>
+      overviewCategoryRank(a.category) - overviewCategoryRank(b.category) ||
+      overviewScenarioRank(a.scenario) - overviewScenarioRank(b.scenario),
+  );
+}
+
+/** DEFAULT_MODELS fixes the row order within each (category, scenario) group,
+ *  AgentX rows sit above 8K/1K rows (see sortOverviewSummaries), and a model
+ *  benchmarked on both scenarios contributes one row per scenario; a rowless
+ *  model still renders all platforms with missing reasons. Live and fixture
+ *  paths both feed this. Scenario presence reads the unscoped rows so switching
+ *  the engine scope changes cell contents, never the shape of the matrix. */
 export function assembleOverviewPageData(
   rowsByModel: Record<string, BenchmarkRow[]>,
   tier: OverviewTier = OVERVIEW_PRIMARY_TIER,
@@ -761,9 +884,11 @@ export function assembleOverviewPageData(
     rows: rowsByModel[model] ?? [],
   }));
   return {
-    models: perModel.flatMap(({ model, rows }) =>
-      overviewScenariosForModel(model, rows).map((scenario) =>
-        buildOverviewModelSummary(model, rows, tier, engineScope, scenario, referenceHardware),
+    models: sortOverviewSummaries(
+      perModel.flatMap(({ model, rows }) =>
+        overviewScenariosForModel(model, rows).map((scenario) =>
+          buildOverviewModelSummary(model, rows, tier, engineScope, scenario, referenceHardware),
+        ),
       ),
     ),
     tier,
@@ -771,10 +896,90 @@ export function assembleOverviewPageData(
     comparisonMode: OVERVIEW_DEFAULT_COMPARISON_MODE,
     referenceHardware,
     modelScope,
+    rowScope: 'all',
+    hardwareRowScope: 'all',
+    unchangedRowCount: 0,
+    emptyRowCount: 0,
     historicalWindow: null,
   };
 }
 
+/** A row earns its place in the 30-day matrix when at least one platform has a
+ *  baseline to compare against. `comparable` is the only status carrying a
+ *  `costDeltaPct`, so it is the same predicate the cells render from. */
+export function overviewRowHasHistoricalChange(model: OverviewModelSummary): boolean {
+  return model.platforms.some((platform) => platform.historicalComparison?.status === 'comparable');
+}
+
+/**
+ * Counts the rows that did not move, and narrows the matrix to the ones that
+ * did when the reader opted in.
+ *
+ * Hardware mode passes straight through. So does a window in which nothing is
+ * comparable — filtering to nothing tells the reader less than the unfiltered
+ * matrix, and the empty state is indistinguishable from a data outage.
+ *
+ * Note: an unchanged row is not an empty row. It routinely carries current
+ * costs that exist nowhere else on the page — on the live site Kimi K3's only
+ * row has no 30-day baseline yet still prices three platforms. That is why
+ * narrowing is opt-in rather than the default.
+ */
+export function applyOverviewRowScope(
+  data: OverviewPageData,
+  rowScope: OverviewRowScope,
+): OverviewPageData {
+  // Hardware mode filters on its own terms, but the reader's answer here is
+  // still carried so a tab switch can restore it. Only the count is zeroed:
+  // there is no control to label while this mode is off screen.
+  if (data.comparisonMode === 'hardware') {
+    return { ...data, rowScope, unchangedRowCount: 0 };
+  }
+
+  const changed = data.models.filter(overviewRowHasHistoricalChange);
+  const unchangedRowCount = data.models.length - changed.length;
+  if (changed.length === 0 || unchangedRowCount === 0) {
+    return { ...data, rowScope: 'all', unchangedRowCount: 0 };
+  }
+
+  return rowScope === 'changed'
+    ? { ...data, models: changed, rowScope: 'changed', unchangedRowCount }
+    : { ...data, rowScope: 'all', unchangedRowCount };
+}
+
+/** A row earns its place in the hardware matrix as soon as one platform quotes
+ *  a cost. One price is still a fact about the row; zero prices is a row of
+ *  dashes saying only that nothing was measured. */
+export function overviewRowHasAnyCost(model: OverviewModelSummary): boolean {
+  return model.platforms.some((platform) => platform.costPerMtok !== null);
+}
+
+/**
+ * The hardware-mode counterpart of {@link applyOverviewRowScope}: counts the
+ * rows that price nothing, and drops them when the reader opts in.
+ *
+ * History mode passes straight through, as does a matrix where every row is
+ * empty — filtering to nothing would read as an outage rather than a filter.
+ */
+export function applyOverviewHardwareRowScope(
+  data: OverviewPageData,
+  hardwareRowScope: OverviewHardwareRowScope,
+): OverviewPageData {
+  // Carried rather than cleared, for the same reason as the history scope.
+  if (data.comparisonMode !== 'hardware') {
+    return { ...data, hardwareRowScope, emptyRowCount: 0 };
+  }
+
+  const priced = data.models.filter(overviewRowHasAnyCost);
+  const emptyRowCount = data.models.length - priced.length;
+  if (priced.length === 0 || emptyRowCount === 0) {
+    return { ...data, hardwareRowScope: 'all', emptyRowCount: 0 };
+  }
+
+  return hardwareRowScope === 'priced'
+    ? { ...data, models: priced, hardwareRowScope: 'priced', emptyRowCount }
+    : { ...data, hardwareRowScope: 'all', emptyRowCount };
+}
+
 function overviewPlatformKey(
   model: OverviewModelSummary,
   platform: OverviewPlatformResult,
diff --git a/packages/app/src/lib/overview-links.test.ts b/packages/app/src/lib/overview-links.test.ts
index 566173e6e..661f75281 100644
--- a/packages/app/src/lib/overview-links.test.ts
+++ b/packages/app/src/lib/overview-links.test.ts
@@ -12,6 +12,7 @@ import {
   buildOverviewHistoryDashboardHref,
   detailHref,
   mergeOverviewControlHref,
+  OVERVIEW_CLIENT_ONLY_KEYS,
   overviewEngineScopeHref,
   overviewHref,
   overviewTierHref,
@@ -270,6 +271,24 @@ describe('overviewHref', () => {
 });
 
 describe('overview switch links', () => {
+  it('treats presentation intent as client-only navigation state', () => {
+    expect(OVERVIEW_CLIENT_ONLY_KEYS).toContain('present');
+  });
+
+  it('merges presentation intent canonically without dropping query state or fragments', () => {
+    expect(
+      mergeOverviewControlHref('/overview?tier=75&utm_source=deck#matrix', '/overview?present=1', [
+        'present',
+      ]),
+    ).toBe('/overview?tier=75&present=1&utm_source=deck#matrix');
+
+    expect(
+      mergeOverviewControlHref('/overview?tier=75&present=1&utm_source=deck#matrix', '/overview', [
+        'present',
+      ]),
+    ).toBe('/overview?tier=75&utm_source=deck#matrix');
+  });
+
   it('merges rapid control changes into the latest pending overview URL', () => {
     const tierHref = mergeOverviewControlHref('/overview', '/overview?tier=75', ['tier']);
     const engineHref = mergeOverviewControlHref(tierHref, '/overview?engine=all', ['engine']);
@@ -372,3 +391,147 @@ describe('overview model scope links', () => {
     );
   });
 });
+
+describe('overview row scope links', () => {
+  it('leaves the canonical 30-day URL untouched and appends rows=changed last', () => {
+    expect(overviewHref('en', 50, 'community', '30d', 'b200', 'default', 'all')).toBe(
+      '/overview?compare=30d',
+    );
+    expect(overviewHref('en', 50, 'community', '30d', 'b200', 'default', 'changed')).toBe(
+      '/overview?compare=30d&rows=changed',
+    );
+    expect(overviewHref('en', 100, 'all', '30d', 'b300', 'all', 'changed')).toBe(
+      '/overview?tier=100&engine=all&ref=b300&compare=30d&models=all&rows=changed',
+    );
+    expect(overviewHref('zh', 50, 'community', '30d', 'b200', 'default', 'changed')).toBe(
+      '/zh/overview?compare=30d&rows=changed',
+    );
+  });
+
+  it('keeps the row scope on the URL in hardware mode, where it is dormant', () => {
+    // The 30-day tab is off screen, so the key changes nothing on the rendered
+    // matrix. It is still the only record of that tab's answer, and dropping it
+    // here is what used to lose the filter on refresh or a shared link.
+    expect(overviewHref('en', 50, 'community', 'hardware', 'b200', 'default', 'changed')).toBe(
+      '/overview?rows=changed',
+    );
+    expect(overviewHref('en', 75, 'community', 'hardware', 'b200', 'all', 'changed')).toBe(
+      '/overview?tier=75&models=all&rows=changed',
+    );
+  });
+
+  it('preserves the row scope when changing tiers and engine scope', () => {
+    expect(overviewTierHref('en', 75, 'community', '30d', 'b200', 'default', 'changed')).toBe(
+      '/overview?tier=75&compare=30d&rows=changed',
+    );
+    expect(overviewEngineScopeHref('en', 'all', 50, '30d', 'b200', 'default', 'changed')).toBe(
+      '/overview?engine=all&compare=30d&rows=changed',
+    );
+  });
+
+  it('merges the rows control into pending overview URLs', () => {
+    const narrowed = mergeOverviewControlHref(
+      '/overview?compare=30d',
+      '/overview?compare=30d&rows=changed',
+      ['rows'],
+    );
+    expect(narrowed).toBe('/overview?compare=30d&rows=changed');
+    expect(mergeOverviewControlHref(narrowed, '/overview?tier=75', ['tier'])).toBe(
+      '/overview?tier=75&compare=30d&rows=changed',
+    );
+    expect(mergeOverviewControlHref(narrowed, '/overview?compare=30d', ['rows'])).toBe(
+      '/overview?compare=30d',
+    );
+  });
+});
+
+describe('overview hardware row scope links', () => {
+  it('leaves the canonical URL untouched and appends hwrows=priced last', () => {
+    expect(overviewHref('en', 50, 'community', 'hardware', 'b200', 'default', 'all', 'all')).toBe(
+      '/overview',
+    );
+    expect(
+      overviewHref('en', 50, 'community', 'hardware', 'b200', 'default', 'all', 'priced'),
+    ).toBe('/overview?hwrows=priced');
+    expect(overviewHref('en', 100, 'all', 'hardware', 'b300', 'all', 'all', 'priced')).toBe(
+      '/overview?tier=100&engine=all&ref=b300&models=all&hwrows=priced',
+    );
+    expect(
+      overviewHref('zh', 50, 'community', 'hardware', 'b200', 'default', 'all', 'priced'),
+    ).toBe('/zh/overview?hwrows=priced');
+  });
+
+  it('keeps the hardware row scope on the URL in history mode, where it is dormant', () => {
+    expect(overviewHref('en', 50, 'community', '30d', 'b200', 'default', 'all', 'priced')).toBe(
+      '/overview?compare=30d&hwrows=priced',
+    );
+  });
+
+  it('writes both row scopes at once, so a URL rebuilt from page data keeps them', () => {
+    // Every href the page builds comes from the current page data, including
+    // the one the client router starts from on first load. Emitting only the
+    // active mode's key there erased the other tab's filter on refresh and on
+    // any shared link carrying both.
+    expect(overviewHref('en', 50, 'community', '30d', 'b200', 'default', 'changed', 'priced')).toBe(
+      '/overview?compare=30d&rows=changed&hwrows=priced',
+    );
+    expect(
+      overviewHref('en', 50, 'community', 'hardware', 'b200', 'default', 'changed', 'priced'),
+    ).toBe('/overview?rows=changed&hwrows=priced');
+  });
+
+  it('preserves the hardware row scope when changing tiers and engine scope', () => {
+    expect(
+      overviewTierHref('en', 75, 'community', 'hardware', 'b200', 'default', 'all', 'priced'),
+    ).toBe('/overview?tier=75&hwrows=priced');
+    expect(
+      overviewEngineScopeHref('en', 'all', 50, 'hardware', 'b200', 'default', 'all', 'priced'),
+    ).toBe('/overview?engine=all&hwrows=priced');
+  });
+
+  it('lets each mode keep its own filter across a tab switch', () => {
+    // The comparison switcher owns only `compare`, so the scope belonging to the
+    // mode being left stays in the URL and is waiting when the reader returns.
+    const hardware = mergeOverviewControlHref('/overview', '/overview?hwrows=priced', ['hwrows']);
+    expect(hardware).toBe('/overview?hwrows=priced');
+
+    const history = mergeOverviewControlHref(hardware, '/overview?compare=30d', ['compare']);
+    expect(history).toBe('/overview?compare=30d&hwrows=priced');
+
+    const narrowed = mergeOverviewControlHref(history, '/overview?compare=30d&rows=changed', [
+      'rows',
+    ]);
+    expect(narrowed).toBe('/overview?compare=30d&rows=changed&hwrows=priced');
+
+    expect(mergeOverviewControlHref(narrowed, '/overview', ['compare'])).toBe(
+      '/overview?rows=changed&hwrows=priced',
+    );
+  });
+
+  it('lets the models control reset both scopes without clearing the dormant one', () => {
+    // The models toggle owns `models` plus both row keys, because revealing
+    // inactive models resets them. A merged key the target href does not carry
+    // is deleted, so hiding — which resets nothing — has to restate both.
+    const current = '/overview?compare=30d&models=all&rows=changed&hwrows=priced';
+    const keys = ['models', 'rows', 'hwrows'] as const;
+
+    const hiding = overviewHref(
+      'en',
+      50,
+      'community',
+      '30d',
+      'b200',
+      'default',
+      'changed',
+      'priced',
+    );
+    expect(mergeOverviewControlHref(current, hiding, keys)).toBe(
+      '/overview?compare=30d&rows=changed&hwrows=priced',
+    );
+
+    const revealing = overviewHref('en', 50, 'community', '30d', 'b200', 'all', 'all', 'all');
+    expect(mergeOverviewControlHref(current, revealing, keys)).toBe(
+      '/overview?compare=30d&models=all',
+    );
+  });
+});
diff --git a/packages/app/src/lib/overview-links.ts b/packages/app/src/lib/overview-links.ts
index a61f38df8..d41049f2e 100644
--- a/packages/app/src/lib/overview-links.ts
+++ b/packages/app/src/lib/overview-links.ts
@@ -1,19 +1,31 @@
 import { runIdFromRunUrl } from './known-issues';
 import {
   OVERVIEW_DEFAULT_COMPARISON_MODE,
+  OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
   OVERVIEW_DEFAULT_MODEL_SCOPE,
   OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
+  OVERVIEW_DEFAULT_ROW_SCOPE,
   OVERVIEW_PRIMARY_TIER,
   type OverviewComparisonMode,
   type OverviewConfigView,
   type OverviewEngineScope,
+  type OverviewHardwareRowScope,
   type OverviewModelScope,
   type OverviewModelSummary,
   type OverviewReferenceHardware,
+  type OverviewRowScope,
   type OverviewTier,
 } from './overview-data';
 
-export type OverviewSearchKey = 'tier' | 'engine' | 'ref' | 'compare' | 'models';
+export type OverviewSearchKey =
+  | 'tier'
+  | 'engine'
+  | 'ref'
+  | 'compare'
+  | 'models'
+  | 'rows'
+  | 'hwrows'
+  | 'present';
 
 export const OVERVIEW_SEARCH_ORDER: readonly OverviewSearchKey[] = [
   'tier',
@@ -21,12 +33,17 @@ export const OVERVIEW_SEARCH_ORDER: readonly OverviewSearchKey[] = [
   'ref',
   'compare',
   'models',
+  'rows',
+  'hwrows',
+  'present',
 ];
 
 /** Params the client resolves without asking the server. `ref` only picks which
- *  column the percentages are measured against, and every cost that needs is
- *  already in the payload — so it must not vary the data cache key. */
-export const OVERVIEW_CLIENT_ONLY_KEYS: readonly OverviewSearchKey[] = ['ref'];
+ *  column the percentages are measured against, and `present` only records the
+ *  requested layout. Neither changes the Overview payload or its cache key. */
+export const OVERVIEW_CLIENT_ONLY_KEYS = ['ref', 'present'] as const;
+
+export type OverviewClientOnlySearchKey = (typeof OVERVIEW_CLIENT_ONLY_KEYS)[number];
 
 /** Apply one control's destination to the latest pending overview URL.
  * This prevents a second, fast selection from rebuilding from stale server
@@ -197,6 +214,8 @@ export function overviewHref(
   comparisonMode: OverviewComparisonMode = OVERVIEW_DEFAULT_COMPARISON_MODE,
   referenceHardware: OverviewReferenceHardware = OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
   modelScope: OverviewModelScope = OVERVIEW_DEFAULT_MODEL_SCOPE,
+  rowScope: OverviewRowScope = OVERVIEW_DEFAULT_ROW_SCOPE,
+  hardwareRowScope: OverviewHardwareRowScope = OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
 ): string {
   const base = locale === 'zh' ? '/zh/overview' : '/overview';
   const query = new URLSearchParams();
@@ -207,6 +226,17 @@ export function overviewHref(
   }
   if (comparisonMode !== 'hardware') query.set('compare', comparisonMode);
   if (modelScope !== OVERVIEW_DEFAULT_MODEL_SCOPE) query.set('models', modelScope);
+  // Each mode filters rows on its own terms and carries its own key. Both are
+  // written whenever they are set, including the one whose mode is off screen:
+  // that dormant key is the only record of the other tab's answer, and a URL
+  // rebuilt from page data — on first load, on refresh, or from a shared link —
+  // has nothing else to restore it from.
+  if (rowScope !== OVERVIEW_DEFAULT_ROW_SCOPE) query.set('rows', rowScope);
+  if (hardwareRowScope !== OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE) {
+    query.set('hwrows', hardwareRowScope);
+  }
+  // `rows=changed` and `hwrows=priced` are the only values ever emitted; `all`
+  // is the default and stays out of the URL so the canonical link is unchanged.
   const search = query.toString();
   return search === '' ? base : `${base}?${search}`;
 }
@@ -219,8 +249,19 @@ export function overviewTierHref(
   comparisonMode: OverviewComparisonMode = OVERVIEW_DEFAULT_COMPARISON_MODE,
   referenceHardware: OverviewReferenceHardware = OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
   modelScope: OverviewModelScope = OVERVIEW_DEFAULT_MODEL_SCOPE,
+  rowScope: OverviewRowScope = OVERVIEW_DEFAULT_ROW_SCOPE,
+  hardwareRowScope: OverviewHardwareRowScope = OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
 ): string {
-  return overviewHref(locale, tier, engineScope, comparisonMode, referenceHardware, modelScope);
+  return overviewHref(
+    locale,
+    tier,
+    engineScope,
+    comparisonMode,
+    referenceHardware,
+    modelScope,
+    rowScope,
+    hardwareRowScope,
+  );
 }
 
 /** Engine-scope switch preserving the active service tier. */
@@ -231,6 +272,17 @@ export function overviewEngineScopeHref(
   comparisonMode: OverviewComparisonMode = OVERVIEW_DEFAULT_COMPARISON_MODE,
   referenceHardware: OverviewReferenceHardware = OVERVIEW_DEFAULT_REFERENCE_HARDWARE,
   modelScope: OverviewModelScope = OVERVIEW_DEFAULT_MODEL_SCOPE,
+  rowScope: OverviewRowScope = OVERVIEW_DEFAULT_ROW_SCOPE,
+  hardwareRowScope: OverviewHardwareRowScope = OVERVIEW_DEFAULT_HARDWARE_ROW_SCOPE,
 ): string {
-  return overviewHref(locale, tier, engineScope, comparisonMode, referenceHardware, modelScope);
+  return overviewHref(
+    locale,
+    tier,
+    engineScope,
+    comparisonMode,
+    referenceHardware,
+    modelScope,
+    rowScope,
+    hardwareRowScope,
+  );
 }
diff --git a/packages/app/src/lib/overview-metadata.test.ts b/packages/app/src/lib/overview-metadata.test.ts
new file mode 100644
index 000000000..cdedd2f94
--- /dev/null
+++ b/packages/app/src/lib/overview-metadata.test.ts
@@ -0,0 +1,44 @@
+import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+import { describe, expect, it } from 'vitest';
+
+import { ZH_OG_LOCALE } from './i18n';
+import { buildOverviewMetadata } from './overview-route.server';
+
+const EXPECTED_DESCRIPTIONS = {
+  en: 'Compare hyperscaler cost per million total tokens across B200, MI355X, B300, GB200 NVL72, and GB300 NVL72 for the AgentX long-context, multi-turn coding scenario and fixed-sequence scenarios where data is available.',
+  zh: '对比 B200、MI355X、B300、GB200 NVL72 和 GB300 NVL72 在 AgentX 长上下文、多轮编码场景及固定序列场景下，按超大规模云厂商口径计算的每百万总 token 成本；仅展示具备对应数据的模型。',
+} as const;
+
+describe('buildOverviewMetadata', () => {
+  it.each(['en', 'zh'] as const)('locks the curated Overview platform wording in %s', (locale) => {
+    const metadata = buildOverviewMetadata(locale);
+
+    expect(metadata.description).toBe(EXPECTED_DESCRIPTIONS[locale]);
+  });
+
+  it('preserves the English title, canonical URL, and OpenGraph identity', () => {
+    const metadata = buildOverviewMetadata('en');
+
+    expect(metadata.title).toBe('Agentic Inference Costs');
+    expect(metadata.alternates).toMatchObject({ canonical: `${SITE_URL}/overview` });
+    expect(metadata.openGraph).toMatchObject({
+      title: `Agentic Inference Costs | ${SITE_NAME}`,
+      url: `${SITE_URL}/overview`,
+      type: 'website',
+    });
+    expect(metadata.openGraph).not.toHaveProperty('locale');
+  });
+
+  it('preserves the Chinese title, canonical URL, and OpenGraph locale', () => {
+    const metadata = buildOverviewMetadata('zh');
+
+    expect(metadata.title).toBe('智能体推理成本');
+    expect(metadata.alternates).toMatchObject({ canonical: `${SITE_URL}/zh/overview` });
+    expect(metadata.openGraph).toMatchObject({
+      title: `智能体推理成本 | ${SITE_NAME}`,
+      url: `${SITE_URL}/zh/overview`,
+      type: 'website',
+      locale: ZH_OG_LOCALE,
+    });
+  });
+});
diff --git a/packages/app/src/lib/overview-route.server.test.tsx b/packages/app/src/lib/overview-route.server.test.tsx
new file mode 100644
index 000000000..94be8cf19
--- /dev/null
+++ b/packages/app/src/lib/overview-route.server.test.tsx
@@ -0,0 +1,80 @@
+import { isValidElement } from 'react';
+import { beforeEach, describe, expect, it, vi } from 'vitest';
+
+import type { OverviewPageData } from './overview-data';
+
+const { mockGetOverviewPageData } = vi.hoisted(() => ({
+  mockGetOverviewPageData: vi.fn(),
+}));
+
+vi.mock('@/lib/overview-data.server', () => ({
+  getOverviewPageData: mockGetOverviewPageData,
+}));
+
+import { renderOverviewPage } from './overview-route.server';
+
+const PAGE_DATA = { models: [] } as unknown as OverviewPageData;
+
+beforeEach(() => {
+  vi.clearAllMocks();
+  mockGetOverviewPageData.mockResolvedValue(PAGE_DATA);
+});
+
+describe('renderOverviewPage', () => {
+  it.each(['en', 'zh'] as const)(
+    'resolves every query control once and propagates the %s locale',
+    async (locale) => {
+      const element = await renderOverviewPage({
+        locale,
+        searchParams: Promise.resolve({
+          tier: ['100', '75'],
+          engine: ['all', 'community'],
+          compare: ['60d', 'hardware'],
+          ref: ['gb300', 'b200'],
+          models: ['all', 'default'],
+          rows: ['changed', 'all'],
+          hwrows: ['priced', 'all'],
+        }),
+      });
+
+      expect(mockGetOverviewPageData).toHaveBeenCalledOnce();
+      expect(mockGetOverviewPageData).toHaveBeenCalledWith(
+        100,
+        'all',
+        '60d',
+        'gb300',
+        'all',
+        'changed',
+        'priced',
+      );
+      expect(isValidElement(element)).toBe(true);
+      expect(element.props).toMatchObject({ data: PAGE_DATA, locale });
+    },
+  );
+
+  it('normalizes invalid and array values with the canonical overview defaults', async () => {
+    await renderOverviewPage({
+      locale: 'en',
+      searchParams: Promise.resolve({
+        tier: ['999', '75'],
+        engine: ['vendor', 'all'],
+        compare: ['weekly', '30d'],
+        ref: ['h100', 'b300'],
+        models: ['inactive', 'all'],
+        rows: ['some', 'changed'],
+        hwrows: ['blank', 'priced'],
+      }),
+    });
+
+    expect(mockGetOverviewPageData).toHaveBeenCalledOnce();
+    expect(mockGetOverviewPageData).toHaveBeenCalledWith(
+      50,
+      'community',
+      'hardware',
+      'b200',
+      'default',
+      'all',
+      'all',
+    );
+  });
+});
diff --git a/packages/app/src/lib/overview-route.server.tsx b/packages/app/src/lib/overview-route.server.tsx
new file mode 100644
index 000000000..b917bd995
--- /dev/null
+++ b/packages/app/src/lib/overview-route.server.tsx
@@ -0,0 +1,77 @@
+import type { Metadata } from 'next';
+
+import { SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
+
+import { OverviewPageContent } from '@/components/overview/overview-page';
+import { enAlternates, type Locale, ZH_OG_LOCALE, zhAlternates } from '@/lib/i18n';
+import {
+  OVERVIEW_HARDWARE,
+  overviewHardwareLabel,
+  resolveOverviewComparisonMode,
+  resolveOverviewEngineScope,
+  resolveOverviewHardwareRowScope,
+  resolveOverviewModelScope,
+  resolveOverviewReferenceHardware,
+  resolveOverviewRowScope,
+  resolveOverviewTier,
+} from '@/lib/overview-data';
+import { getOverviewPageData } from '@/lib/overview-data.server';
+
+export type OverviewSearchParams = Record<string, string | string[] | undefined>;
+
+export interface OverviewRoutePageProps {
+  searchParams: Promise<OverviewSearchParams>;
+}
+
+interface OverviewRouteProps extends OverviewRoutePageProps {
+  locale: Locale;
+}
+
+function overviewPlatformList(locale: Locale): string {
+  const labels = OVERVIEW_HARDWARE.map((hardware) => overviewHardwareLabel(hardware));
+  if (locale === 'zh') return `${labels.slice(0, -1).join('、')} 和 ${labels.at(-1)}`;
+  return `${labels.slice(0, -1).join(', ')}, and ${labels.at(-1)}`;
+}
+
+export function buildOverviewMetadata(locale: Locale): Metadata {
+  const platforms = overviewPlatformList(locale);
+  const title = locale === 'zh' ? '智能体推理成本' : 'Agentic Inference Costs';
+  const description =
+    locale === 'zh'
+      ? `对比 ${platforms} 在 AgentX 长上下文、多轮编码场景及固定序列场景下，按超大规模云厂商口径计算的每百万总 token 成本；仅展示具备对应数据的模型。`
+      : `Compare hyperscaler cost per million total tokens across ${platforms} for the AgentX long-context, multi-turn coding scenario and fixed-sequence scenarios where data is available.`;
+  const path = locale === 'zh' ? '/zh/overview' : '/overview';
+
+  return {
+    title,
+    description,
+    alternates: locale === 'zh' ? zhAlternates('/overview') : enAlternates('/overview'),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url: `${SITE_URL}${path}`,
+      type: 'website',
+      ...(locale === 'zh' ? { locale: ZH_OG_LOCALE } : {}),
+    },
+    twitter: {
+      card: 'summary_large_image',
+      title: `${title} | ${SITE_NAME}`,
+      description,
+    },
+  };
+}
+
+export async function renderOverviewPage({ locale, searchParams }: OverviewRouteProps) {
+  const params = await searchParams;
+  const data = await getOverviewPageData(
+    resolveOverviewTier(params.tier),
+    resolveOverviewEngineScope(params.engine),
+    resolveOverviewComparisonMode(params.compare),
+    resolveOverviewReferenceHardware(params.ref),
+    resolveOverviewModelScope(params.models),
+    resolveOverviewRowScope(params.rows),
+    resolveOverviewHardwareRowScope(params.hwrows),
+  );
+
+  return <OverviewPageContent data={data} locale={locale} />;
+}
diff --git a/packages/app/src/lib/png-chart.tsx b/packages/app/src/lib/png-chart.tsx
index 6ad4d558a..bc6cceb9a 100644
--- a/packages/app/src/lib/png-chart.tsx
+++ b/packages/app/src/lib/png-chart.tsx
@@ -1,7 +1,9 @@
-// Shared chart primitives for variant-compare PNG routes.
-// compare/compare-per-dollar/blog still carry their own copies (follow-up to migrate).
-import type { SsrInterpolatedRow } from '@/lib/compare-variant-ssr';
+import { ImageResponse } from 'next/og';
 
+import type { SsrInterpolatedRow } from '@/lib/compare-ssr';
+
+// Render natively at high DPI. Scaling the CSS output would bitmap-upsample
+// Satori's text and SVG paths instead of rasterizing them at the target size.
 export const R = 2;
 export const SIZE = { width: 1200 * R, height: 675 * R };
 export const CHART_FRAME = { left: 0, top: 18 * R, width: 746 * R, height: 382 * R };
@@ -28,6 +30,20 @@ export interface TargetedPoint extends Point {
   target: number;
 }
 
+interface ComparePngChartOptions {
+  curveRows: SsrInterpolatedRow[];
+  plottedRows: SsrInterpolatedRow[];
+  logoSrc: string | null;
+  aLabel: string;
+  bLabel: string;
+  eyebrow: string;
+  title: string;
+  subtitle: string;
+  workload: string;
+  rangeNote: string;
+  footer: string;
+}
+
 export function money(value: number): string {
   if (value >= 10) return `$${value.toFixed(1)}`;
   if (value >= 1) return `$${value.toFixed(2)}`;
@@ -81,15 +97,19 @@ export function buildSeriesPoints(
 ): TargetedPoint[] {
   return curveRows
     .map((row) => ({ target: row.target, cost: getCost(row) }))
-    .filter((p): p is { target: number; cost: number } => p.cost !== null)
-    .map((p) => ({ x: scaleX(p.target), y: scaleY(p.cost), target: p.target }));
+    .filter((point): point is { target: number; cost: number } => point.cost !== null)
+    .map((point) => ({
+      x: scaleX(point.target),
+      y: scaleY(point.cost),
+      target: point.target,
+    }));
 }
 
 export function splitByMatchRange(points: TargetedPoint[], matchedMin: number, matchedMax: number) {
   return {
-    matched: points.filter((p) => p.target >= matchedMin && p.target <= matchedMax),
-    leftExt: points.filter((p) => p.target <= matchedMin),
-    rightExt: points.filter((p) => p.target >= matchedMax),
+    matched: points.filter((point) => point.target >= matchedMin && point.target <= matchedMax),
+    leftExt: points.filter((point) => point.target <= matchedMin),
+    rightExt: points.filter((point) => point.target >= matchedMax),
   };
 }
 
@@ -108,3 +128,398 @@ export function renderSeriesPath(points: Point[], stroke: string, dashed: boolea
     />
   );
 }
+
+/** Shared Satori shell for all indexed compare PNG routes. */
+export function renderComparePngChart({
+  curveRows,
+  plottedRows,
+  logoSrc,
+  aLabel,
+  bLabel,
+  eyebrow,
+  title,
+  subtitle,
+  workload,
+  rangeNote,
+  footer,
+}: ComparePngChartOptions): ImageResponse {
+  const costs = curveRows
+    .flatMap((row) => [row.a?.cost, row.b?.cost])
+    .filter((cost): cost is number => typeof cost === 'number' && Number.isFinite(cost));
+  const costMin = costs.length > 0 ? Math.min(...costs) : 0;
+  const costMax = costs.length > 0 ? Math.max(...costs) : 1;
+  const yAxis = niceAxis(Math.min(0, costMin), costMax);
+  const xMin = curveRows.at(0)?.target ?? 0;
+  const xMax = curveRows.at(-1)?.target ?? 100;
+  const matchedMin = plottedRows.at(0)?.target ?? xMin;
+  const matchedMax = plottedRows.at(-1)?.target ?? xMax;
+  const hasLeftExtension = matchedMin - xMin >= 0.5;
+  const hasRightExtension = xMax - matchedMax >= 0.5;
+  const showRangeEndpoints = hasLeftExtension || hasRightExtension;
+  const scaleX = (value: number) =>
+    CHART.left + (xMax === xMin ? CHART.width / 2 : ((value - xMin) / (xMax - xMin)) * CHART.width);
+  const scaleY = (value: number) =>
+    CHART.top +
+    CHART.height -
+    (yAxis.max === yAxis.min
+      ? CHART.height / 2
+      : ((value - yAxis.min) / (yAxis.max - yAxis.min)) * CHART.height);
+  const aSeries = splitByMatchRange(
+    buildSeriesPoints(curveRows, (row) => row.a?.cost ?? null, scaleX, scaleY),
+    matchedMin,
+    matchedMax,
+  );
+  const bSeries = splitByMatchRange(
+    buildSeriesPoints(curveRows, (row) => row.b?.cost ?? null, scaleX, scaleY),
+    matchedMin,
+    matchedMax,
+  );
+  const aHighlightPoints = plottedRows
+    .filter((row) => row.a)
+    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.a!.cost) }));
+  const bHighlightPoints = plottedRows
+    .filter((row) => row.b)
+    .map((row) => ({ x: scaleX(row.target), y: scaleY(row.b!.cost) }));
+  const svgWidth = 760 * R;
+  const svgHeight = 406 * R;
+
+  return new ImageResponse(
+    <div
+      style={{
+        display: 'flex',
+        flexDirection: 'column',
+        width: SIZE.width,
+        height: SIZE.height,
+        padding: `${38 * R}px ${46 * R}px ${26 * R}px`,
+        background: COLORS.background,
+        color: COLORS.text,
+        fontFamily: 'Arial, sans-serif',
+      }}
+    >
+      <div style={{ display: 'flex', justifyContent: 'space-between', alignItems: 'flex-start' }}>
+        <div style={{ display: 'flex', flexDirection: 'column', gap: 7 * R }}>
+          <div
+            style={{
+              display: 'flex',
+              fontSize: 19 * R,
+              fontWeight: 700,
+              letterSpacing: '0.13em',
+              textTransform: 'uppercase',
+              color: COLORS.blue,
+            }}
+          >
+            {eyebrow}
+          </div>
+          <div style={{ display: 'flex', fontSize: 41 * R, fontWeight: 800 }}>{title}</div>
+          <div style={{ display: 'flex', fontSize: 25 * R, color: COLORS.muted }}>{subtitle}</div>
+        </div>
+        <div
+          style={{
+            display: 'flex',
+            flexDirection: 'column',
+            alignItems: 'flex-end',
+            border: `${R}px solid ${COLORS.border}`,
+            borderRadius: 12 * R,
+            padding: `${13 * R}px ${17 * R}px`,
+            background: COLORS.panel,
+            gap: 5 * R,
+          }}
+        >
+          <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
+            DEFAULT WORKLOAD
+          </div>
+          <div style={{ display: 'flex', fontSize: 21 * R, fontWeight: 700 }}>
+            {workload || 'Default comparison'}
+          </div>
+          <div style={{ display: 'flex', fontSize: 14 * R, color: COLORS.muted }}>
+            Lower cost is better
+          </div>
+        </div>
+      </div>
+
+      <div style={{ display: 'flex', flex: 1, gap: 34 * R, marginTop: 22 * R }}>
+        <div style={{ display: 'flex', position: 'relative', width: svgWidth, height: svgHeight }}>
+          <svg
+            width={svgWidth}
+            height={svgHeight}
+            viewBox={`0 0 ${svgWidth} ${svgHeight}`}
+            style={{ position: 'absolute', left: 0, top: 0 }}
+          >
+            <rect
+              x={CHART_FRAME.left}
+              y={CHART_FRAME.top}
+              width={CHART_FRAME.width}
+              height={CHART_FRAME.height}
+              rx={13 * R}
+              fill={COLORS.panel}
+              stroke={COLORS.border}
+            />
+            {yAxis.ticks.map((tick) => {
+              const y = scaleY(tick);
+              return (
+                <line
+                  key={tick}
+                  x1={CHART.left}
+                  x2={CHART.left + CHART.width}
+                  y1={y}
+                  y2={y}
+                  stroke={COLORS.grid}
+                  strokeWidth={2 * R}
+                />
+              );
+            })}
+            {plottedRows.map((row) => {
+              const x = scaleX(row.target);
+              return (
+                <line
+                  key={`mark-${row.target}`}
+                  x1={x}
+                  x2={x}
+                  y1={CHART.top + CHART.height}
+                  y2={CHART.top + CHART.height + 6 * R}
+                  stroke={COLORS.muted}
+                  strokeWidth={2 * R}
+                />
+              );
+            })}
+            {renderSeriesPath(aSeries.leftExt, COLORS.a, true)}
+            {renderSeriesPath(aSeries.rightExt, COLORS.a, true)}
+            {renderSeriesPath(aSeries.matched, COLORS.a, false)}
+            {renderSeriesPath(bSeries.leftExt, COLORS.b, true)}
+            {renderSeriesPath(bSeries.rightExt, COLORS.b, true)}
+            {renderSeriesPath(bSeries.matched, COLORS.b, false)}
+            {aHighlightPoints.map((point, index) => (
+              <circle
+                key={`a-${index}`}
+                cx={point.x}
+                cy={point.y}
+                r={10 * R}
+                fill={COLORS.a}
+                stroke={COLORS.background}
+                strokeWidth={4 * R}
+              />
+            ))}
+            {bHighlightPoints.map((point, index) => (
+              <circle
+                key={`b-${index}`}
+                cx={point.x}
+                cy={point.y}
+                r={10 * R}
+                fill={COLORS.b}
+                stroke={COLORS.background}
+                strokeWidth={4 * R}
+              />
+            ))}
+          </svg>
+          {logoSrc && (
+            <img
+              src={logoSrc}
+              alt=""
+              height={144 * R}
+              style={{
+                position: 'absolute',
+                left: CHART.left + CHART.width / 2 - 168 * R,
+                top: CHART.top + CHART.height / 2 - 72 * R,
+                opacity: 0.12,
+              }}
+            />
+          )}
+          {yAxis.ticks.map((tick) => (
+            <div
+              key={`y-label-${tick}`}
+              style={{
+                display: 'flex',
+                position: 'absolute',
+                left: CHART_FRAME.left + 14 * R,
+                top: scaleY(tick) - 9 * R,
+                width: CHART.left - CHART_FRAME.left - 28 * R,
+                justifyContent: 'flex-end',
+                color: COLORS.muted,
+                fontSize: 15 * R,
+              }}
+            >
+              {moneyForStep(tick, yAxis.step)}
+            </div>
+          ))}
+          {plottedRows.map((row) => (
+            <div
+              key={`x-label-${row.target}`}
+              style={{
+                display: 'flex',
+                position: 'absolute',
+                left: scaleX(row.target) - 32 * R,
+                top: CHART.top + CHART.height + 15 * R,
+                width: 64 * R,
+                justifyContent: 'center',
+                color: COLORS.muted,
+                fontSize: 16 * R,
+                fontWeight: 600,
+              }}
+            >
+              {row.target}
+            </div>
+          ))}
+          {showRangeEndpoints && hasLeftExtension && (
+            <div
+              style={{
+                display: 'flex',
+                position: 'absolute',
+                left: scaleX(xMin) - 4 * R,
+                top: CHART.top + CHART.height + 16 * R,
+                width: 56 * R,
+                justifyContent: 'flex-start',
+                color: COLORS.faint,
+                fontSize: 13 * R,
+                fontStyle: 'italic',
+              }}
+            >
+              {Math.round(xMin)}
+            </div>
+          )}
+          {showRangeEndpoints && hasRightExtension && (
+            <div
+              style={{
+                display: 'flex',
+                position: 'absolute',
+                left: scaleX(xMax) - 52 * R,
+                top: CHART.top + CHART.height + 16 * R,
+                width: 56 * R,
+                justifyContent: 'flex-end',
+                color: COLORS.faint,
+                fontSize: 13 * R,
+                fontStyle: 'italic',
+              }}
+            >
+              {Math.round(xMax)}
+            </div>
+          )}
+          <div
+            style={{
+              display: 'flex',
+              position: 'absolute',
+              left: CHART.left,
+              top: CHART.top + CHART.height + 38 * R,
+              width: CHART.width,
+              justifyContent: 'center',
+              color: COLORS.muted,
+              fontSize: 15 * R,
+              fontWeight: 600,
+            }}
+          >
+            Interactivity (tok/s/user)
+          </div>
+          {showRangeEndpoints && (
+            <div
+              style={{
+                display: 'flex',
+                position: 'absolute',
+                left: CHART.left,
+                top: CHART.top + CHART.height + 62 * R,
+                width: CHART.width,
+                justifyContent: 'center',
+                color: COLORS.faint,
+                fontSize: 13 * R,
+                fontStyle: 'italic',
+              }}
+            >
+              {rangeNote}
+            </div>
+          )}
+        </div>
+
+        <div
+          style={{
+            display: 'flex',
+            flex: 1,
+            flexDirection: 'column',
+            gap: 17 * R,
+            paddingTop: 18 * R,
+          }}
+        >
+          <div style={{ display: 'flex', fontSize: 18 * R, fontWeight: 700 }}>
+            Matched Interactivity
+          </div>
+          <div style={{ display: 'flex', gap: 20 * R, fontSize: 15 * R, color: COLORS.muted }}>
+            <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
+              <span
+                style={{
+                  display: 'flex',
+                  width: 19 * R,
+                  height: 6 * R,
+                  borderRadius: 3 * R,
+                  background: COLORS.a,
+                }}
+              />
+              {aLabel}
+            </span>
+            <span style={{ display: 'flex', gap: 7 * R, alignItems: 'center' }}>
+              <span
+                style={{
+                  display: 'flex',
+                  width: 19 * R,
+                  height: 6 * R,
+                  borderRadius: 3 * R,
+                  background: COLORS.b,
+                }}
+              />
+              {bLabel}
+            </span>
+          </div>
+          {plottedRows.length > 0 ? (
+            plottedRows.map((row) => (
+              <div
+                key={`row-${row.target}`}
+                style={{
+                  display: 'flex',
+                  flexDirection: 'column',
+                  gap: 6 * R,
+                  border: `${R}px solid ${COLORS.border}`,
+                  borderRadius: 10 * R,
+                  padding: `${11 * R}px ${13 * R}px`,
+                  background: COLORS.panel,
+                }}
+              >
+                <div style={{ display: 'flex', color: COLORS.muted, fontSize: 13 * R }}>
+                  {row.target} tok/s/user
+                </div>
+                <div style={{ display: 'flex', gap: 15 * R, fontSize: 19 * R, fontWeight: 700 }}>
+                  <span style={{ display: 'flex', color: COLORS.a }}>
+                    {row.a ? money(row.a.cost) : 'N/A'}
+                  </span>
+                  <span style={{ display: 'flex', color: COLORS.b }}>
+                    {row.b ? money(row.b.cost) : 'N/A'}
+                  </span>
+                </div>
+              </div>
+            ))
+          ) : (
+            <div style={{ display: 'flex', fontSize: 18 * R, color: COLORS.muted }}>
+              No matched cost data available.
+            </div>
+          )}
+        </div>
+      </div>
+
+      <div
+        style={{
+          display: 'flex',
+          justifyContent: 'space-between',
+          paddingTop: 9 * R,
+          fontSize: 15 * R,
+          color: COLORS.muted,
+        }}
+      >
+        <span style={{ display: 'flex' }}>{footer}</span>
+        <span style={{ display: 'flex', color: COLORS.text, fontWeight: 700 }}>
+          inferencex.semianalysis.com
+        </span>
+      </div>
+    </div>,
+    {
+      ...SIZE,
+      headers: {
+        'Cache-Control': 'public, s-maxage=3600, stale-while-revalidate=86400',
+      },
+    },
+  );
+}
diff --git a/packages/app/src/lib/public-api-errors.ts b/packages/app/src/lib/public-api-errors.ts
new file mode 100644
index 000000000..c10b8d2e2
--- /dev/null
+++ b/packages/app/src/lib/public-api-errors.ts
@@ -0,0 +1,21 @@
+import { NextResponse } from 'next/server';
+
+/** Stable error strings exposed by public read endpoints and documented in OpenAPI. */
+export const PUBLIC_API_ERRORS = {
+  internal: 'Internal server error',
+  notFound: 'Not found',
+  unknownModel: 'Unknown model',
+  invalidDate: 'Invalid date format (YYYY-MM-DD required)',
+  benchmarkHistoryParameters: 'model, isl, and osl are required',
+  idRequired: 'id is required (benchmark_result_id)',
+  idsRequired: 'ids query param is required',
+  noValidIds: 'no valid ids provided',
+} as const;
+
+export function publicApiError(error: string, status: number): NextResponse {
+  return NextResponse.json({ error }, { status });
+}
+
+export function tooManyIdsError(maxIds: number): string {
+  return `too many ids (max ${maxIds})`;
+}
diff --git a/packages/app/src/lib/rankings-zh.ts b/packages/app/src/lib/rankings-zh.ts
new file mode 100644
index 000000000..10cab24f2
--- /dev/null
+++ b/packages/app/src/lib/rankings-zh.ts
@@ -0,0 +1,55 @@
+/**
+ * @file rankings-zh.ts
+ * @description Simplified Chinese copy builders for the `/zh/rankings/<slug>`
+ * pages. Same structure as the English builders in `rankings.ts`; model
+ * names, hardware SKUs, and framework names stay in English per
+ * `docs/chinese-copy.md`.
+ */
+
+import type { RankingPageEntry } from '@/lib/rankings';
+
+export function rankingPageTitleZh(entry: RankingPageEntry): string {
+  return entry.kind === 'fastest-gpu'
+    ? `${entry.model.seoName} 推理最快 GPU 实时排行`
+    : `运行 ${entry.model.seoName} 最省钱的 GPU：每百万 token 成本排行`;
+}
+
+export function rankingPageHeadingZh(entry: RankingPageEntry): string {
+  return entry.kind === 'fastest-gpu'
+    ? `${entry.model.seoName} 最快 GPU 排行`
+    : `${entry.model.seoName} 最低成本 GPU 排行`;
+}
+
+/** Static fallback meta description; the route swaps in a stat-led one when
+ *  live numbers are available. */
+export function rankingPageDescriptionZh(entry: RankingPageEntry): string {
+  return entry.kind === 'fastest-gpu'
+    ? `哪款 GPU 跑 ${entry.model.seoName} 最快？基于实测单 GPU 每秒 token 数对 NVIDIA 与 AMD 硬件持续排名，数据持续更新。`
+    : `哪款 GPU 跑 ${entry.model.seoName} 最省钱？按超大规模云 GPU 价格下实测的每百万 token 成本持续排名，数据持续更新。`;
+}
+
+export function rankingPageKeywordsZh(entry: RankingPageEntry): string[] {
+  const model = entry.model.seoName;
+  if (entry.kind === 'fastest-gpu') {
+    return [
+      `${model} 最快 GPU`,
+      `${model} 最佳 GPU`,
+      `${model} GPU 基准测试`,
+      `${model} 每秒 token 数`,
+      `${model} 推理速度`,
+      `${model} GPU 排行`,
+      `${model} H100 B200 MI355X 对比`,
+      `${model} 推理硬件`,
+    ];
+  }
+  return [
+    `${model} 最便宜 GPU`,
+    `${model} 每百万 token 成本`,
+    `${model} 推理成本`,
+    `运行 ${model} 最省钱的方案`,
+    `${model} GPU 成本对比`,
+    `${model} 部署成本`,
+    `${model} token 单价`,
+    `${model} 推理定价`,
+  ];
+}
diff --git a/packages/app/src/lib/rankings.test.ts b/packages/app/src/lib/rankings.test.ts
new file mode 100644
index 000000000..4b06fa8e8
--- /dev/null
+++ b/packages/app/src/lib/rankings.test.ts
@@ -0,0 +1,188 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  RANKING_KINDS,
+  buildRankingRows,
+  chipForHardware,
+  getAllRankingPageEntries,
+  getRankingPageEntry,
+  rankingPageDescription,
+  rankingPageHeading,
+  rankingPageKeywords,
+  rankingPagePath,
+  rankingPageTitle,
+  scenarioLabel,
+} from '@/lib/rankings';
+import {
+  rankingPageDescriptionZh,
+  rankingPageHeadingZh,
+  rankingPageKeywordsZh,
+  rankingPageTitleZh,
+} from '@/lib/rankings-zh';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
+import type {
+  OverviewModelSummary,
+  OverviewPlatformResult,
+  OverviewTierRead,
+} from '@/lib/overview-data';
+
+const FORBIDDEN_DASHES = /[\u2013\u2014]/;
+const CJK = /[\u4E00-\u9FFF]/;
+
+function syntheticRead(value: number | null): OverviewTierRead {
+  return {
+    tier: 50,
+    value,
+    boundary: null,
+    estimated: false,
+    evidenceDate: null,
+    evidenceTopologies: [],
+    config: null,
+  };
+}
+
+function syntheticPlatform(overrides: Partial<OverviewPlatformResult>): OverviewPlatformResult {
+  return {
+    hardware: 'h100',
+    hardwareLabel: 'H100',
+    precision: 'fp8',
+    read: syntheticRead(100),
+    missingReason: null,
+    costPerMtok: 1,
+    costVsReferencePct: null,
+    historicalComparison: null,
+    ...overrides,
+  };
+}
+
+function syntheticSummary(platforms: OverviewPlatformResult[]): OverviewModelSummary {
+  return {
+    modelLabel: 'Synthetic',
+    scenario: 'single_turn_8k1k',
+    platforms,
+  } as OverviewModelSummary;
+}
+
+describe('rankings registry', () => {
+  it('has one page per (kind, model) pair', () => {
+    const entries = getAllRankingPageEntries();
+    expect(entries.length).toBe(RANKING_KINDS.length * INFERENCE_MODEL_SLUGS.length);
+    // 28 = RANKING_KINDS.length * INFERENCE_MODEL_SLUGS.length. This fork adds
+    // two model slugs upstream doesn't have (dsv2lite, gemma4 — see
+    // .claude/skills/fork-changelog/SKILL.md), so the count is higher than
+    // upstream's.
+    expect(entries.length).toBe(28);
+  });
+
+  it('has unique, well-formed slugs', () => {
+    const entries = getAllRankingPageEntries();
+    expect(new Set(entries.map((entry) => entry.slug)).size).toBe(entries.length);
+    for (const entry of entries) {
+      expect(entry.slug).toBe(`${entry.kind}-for-${entry.model.slug}`);
+      expect(entry.slug).toMatch(/^[a-z0-9-]+$/);
+      expect(rankingPagePath(entry)).toBe(`/rankings/${entry.slug}`);
+      expect(entry.dbKeys.length).toBeGreaterThan(0);
+    }
+  });
+
+  it('resolves slugs case-insensitively and rejects unknown slugs', () => {
+    const [first] = getAllRankingPageEntries();
+    expect(getRankingPageEntry(first.slug)).toBe(first);
+    expect(getRankingPageEntry(first.slug.toUpperCase())).toBe(first);
+    expect(getRankingPageEntry('fastest-gpu-for-nonexistent')).toBeUndefined();
+  });
+
+  it('maps hardware keys to chip pages', () => {
+    expect(chipForHardware('gb200')?.slug).toBe('gb200-nvl72');
+    expect(chipForHardware('not-a-gpu')).toBeNull();
+  });
+});
+
+describe('rankings copy', () => {
+  it('contains no em or en dashes and enough keywords in either locale', () => {
+    for (const entry of getAllRankingPageEntries()) {
+      const copy = [
+        rankingPageTitle(entry),
+        rankingPageHeading(entry),
+        rankingPageDescription(entry),
+        ...rankingPageKeywords(entry),
+        rankingPageTitleZh(entry),
+        rankingPageHeadingZh(entry),
+        rankingPageDescriptionZh(entry),
+        ...rankingPageKeywordsZh(entry),
+      ];
+      for (const text of copy) {
+        expect(text, `dash found in copy for ${entry.slug}: ${text}`).not.toMatch(FORBIDDEN_DASHES);
+        expect(text.length).toBeGreaterThan(0);
+      }
+      expect(rankingPageKeywords(entry).length).toBeGreaterThanOrEqual(6);
+      expect(rankingPageKeywordsZh(entry).length).toBeGreaterThanOrEqual(6);
+    }
+  });
+
+  it('keeps Chinese copy Chinese while model names stay English', () => {
+    for (const entry of getAllRankingPageEntries()) {
+      expect(rankingPageTitleZh(entry)).toMatch(CJK);
+      expect(rankingPageDescriptionZh(entry)).toMatch(CJK);
+      expect(rankingPageTitleZh(entry)).toContain(entry.model.seoName);
+    }
+  });
+
+  it('words every scenario in both locales without dashes', () => {
+    for (const scenario of ['single_turn_8k1k', 'agentx'] as const) {
+      for (const locale of ['en', 'zh'] as const) {
+        const label = scenarioLabel(scenario, locale);
+        expect(label.length).toBeGreaterThan(0);
+        expect(label).not.toMatch(FORBIDDEN_DASHES);
+      }
+    }
+  });
+});
+
+describe('buildRankingRows', () => {
+  const platforms = [
+    syntheticPlatform({
+      hardware: 'h100',
+      hardwareLabel: 'H100',
+      read: syntheticRead(100),
+      costPerMtok: 3,
+    }),
+    syntheticPlatform({
+      hardware: 'mi355x',
+      hardwareLabel: 'MI355X',
+      read: syntheticRead(250),
+      costPerMtok: 1,
+    }),
+    syntheticPlatform({
+      hardware: 'b200',
+      hardwareLabel: 'B200',
+      read: syntheticRead(null),
+      costPerMtok: null,
+    }),
+    syntheticPlatform({
+      hardware: 'h200',
+      hardwareLabel: 'H200',
+      read: syntheticRead(180),
+      costPerMtok: 2,
+    }),
+  ];
+
+  it('ranks fastest-gpu by descending throughput and drops unmeasured platforms', () => {
+    const rows = buildRankingRows(syntheticSummary(platforms), 'fastest-gpu');
+    expect(rows.map((row) => row.hardware)).toEqual(['mi355x', 'h200', 'h100']);
+    expect(rows.map((row) => row.rank)).toEqual([1, 2, 3]);
+    expect(rows[0].throughputPerGpu).toBe(250);
+  });
+
+  it('ranks cheapest-gpu by ascending cost and drops platforms without cost', () => {
+    const rows = buildRankingRows(syntheticSummary(platforms), 'cheapest-gpu');
+    expect(rows.map((row) => row.hardware)).toEqual(['mi355x', 'h200', 'h100']);
+    expect(rows[0].costPerMtok).toBe(1);
+    expect(rows.at(-1)?.costPerMtok).toBe(3);
+  });
+
+  it('links ranked rows to chip pages when the hardware has one', () => {
+    const rows = buildRankingRows(syntheticSummary(platforms), 'fastest-gpu');
+    expect(rows.find((row) => row.hardware === 'mi355x')?.chip?.slug).toBe('mi355x');
+  });
+});
diff --git a/packages/app/src/lib/rankings.ts b/packages/app/src/lib/rankings.ts
new file mode 100644
index 000000000..97082e4ef
--- /dev/null
+++ b/packages/app/src/lib/rankings.ts
@@ -0,0 +1,199 @@
+/**
+ * @file rankings.ts
+ * @description Slug registry, ranking derivation, and copy builders for the
+ * indexable `/rankings/<slug>` pages.
+ *
+ * Two ranking kinds exist per model, matching the two questions people
+ * actually type into a search box:
+ *
+ *   - `fastest-gpu-for-<model>`  — hardware ranked by measured tokens/s per
+ *     GPU at the overview's primary interactivity tier.
+ *   - `cheapest-gpu-for-<model>` — hardware ranked by $ per million total
+ *     tokens at hyperscaler $/GPU/hr pricing, same tier.
+ *
+ * The ranking engine is NOT reimplemented here: rows come from
+ * `buildOverviewModelSummary`, the exact same derivation the /overview
+ * leaderboard renders, so a ranking page can never disagree with the
+ * dashboard. This module only sorts, filters, and words the result.
+ *
+ * Model slugs reuse `INFERENCE_MODEL_SLUGS` — the /inference, /compare, and
+ * /run vocabulary — so every ranking row can deep-link to its pair page.
+ */
+
+import { getAllChipPages, type ChipPageEntry } from '@/lib/chip-pages';
+import { COMPARE_MODEL_SLUGS } from '@/lib/compare-slug';
+import { INFERENCE_MODEL_SLUGS, type InferenceModelSlug } from '@/lib/inference-model-slug';
+import type {
+  OverviewModelSummary,
+  OverviewPlatformResult,
+  OverviewScenario,
+} from '@/lib/overview-data';
+
+export const RANKING_KINDS = ['fastest-gpu', 'cheapest-gpu'] as const;
+export type RankingKind = (typeof RANKING_KINDS)[number];
+
+export interface RankingPageEntry {
+  /** Canonical URL slug under /rankings/, e.g. 'fastest-gpu-for-kimi-k3'. */
+  slug: string;
+  kind: RankingKind;
+  /** Model registry entry — same slug vocabulary as /inference/<model>. */
+  model: InferenceModelSlug;
+  /** DB keys to fetch benchmark rows for, from the compare registry. */
+  dbKeys: string[];
+}
+
+const DB_KEYS_BY_MODEL_SLUG: ReadonlyMap<string, string[]> = new Map(
+  COMPARE_MODEL_SLUGS.map((entry) => [entry.slug, entry.dbKeys]),
+);
+
+function buildEntries(): RankingPageEntry[] {
+  const out: RankingPageEntry[] = [];
+  for (const kind of RANKING_KINDS) {
+    for (const model of INFERENCE_MODEL_SLUGS) {
+      const dbKeys = DB_KEYS_BY_MODEL_SLUG.get(model.slug);
+      if (!dbKeys) continue;
+      out.push({ slug: `${kind}-for-${model.slug}`, kind, model, dbKeys });
+    }
+  }
+  return out;
+}
+
+const ENTRIES: readonly RankingPageEntry[] = buildEntries();
+
+const ENTRY_BY_SLUG: ReadonlyMap<string, RankingPageEntry> = new Map(
+  ENTRIES.map((entry) => [entry.slug, entry]),
+);
+
+export function getAllRankingPageEntries(): readonly RankingPageEntry[] {
+  return ENTRIES;
+}
+
+export function getRankingPageEntry(slug: string): RankingPageEntry | undefined {
+  return ENTRY_BY_SLUG.get(slug.toLowerCase());
+}
+
+export function rankingPagePath(entry: RankingPageEntry): string {
+  return `/rankings/${entry.slug}`;
+}
+
+// ---------------------------------------------------------------------------
+// Ranking derivation (pure — operates on the overview summary shape)
+// ---------------------------------------------------------------------------
+
+export interface RankingRow {
+  rank: number;
+  /** HW_REGISTRY key, e.g. 'gb200'. */
+  hardware: string;
+  hardwareLabel: string;
+  /** /chips/<slug> registry entry when one exists for this hardware. */
+  chip: ChipPageEntry | null;
+  /** Measured tokens/s per GPU at the ranking tier; null when unavailable. */
+  throughputPerGpu: number | null;
+  /** $ per million total tokens at hyperscaler $/GPU/hr; null when unavailable. */
+  costPerMtok: number | null;
+  /** Precision of the winning config, e.g. 'fp4'. */
+  precision: string | null;
+  /** Framework label of the winning config, e.g. 'SGLang'. */
+  framework: string | null;
+  /** Whether the winning config runs disaggregated prefill/decode. */
+  disagg: boolean | null;
+}
+
+const CHIP_BY_HW_KEY: ReadonlyMap<string, ChipPageEntry> = new Map(
+  getAllChipPages().map((chip) => [chip.hwKey, chip]),
+);
+
+export function chipForHardware(hardware: string): ChipPageEntry | null {
+  return CHIP_BY_HW_KEY.get(hardware) ?? null;
+}
+
+function toRankingRow(platform: OverviewPlatformResult, rank: number): RankingRow {
+  return {
+    rank,
+    hardware: platform.hardware,
+    hardwareLabel: platform.hardwareLabel,
+    chip: chipForHardware(platform.hardware),
+    throughputPerGpu: platform.read.value,
+    costPerMtok: platform.costPerMtok,
+    precision: platform.read.config?.precision ?? platform.precision,
+    framework: platform.read.config?.frameworkLabel ?? null,
+    disagg: platform.read.config?.disagg ?? null,
+  };
+}
+
+/** Sort and rank the overview platforms for one ranking kind. Platforms with
+ *  no measurement for the metric are dropped rather than ranked last, so an
+ *  under-swept GPU is never presented as "slowest". */
+export function buildRankingRows(summary: OverviewModelSummary, kind: RankingKind): RankingRow[] {
+  const measurable = summary.platforms.filter((platform) =>
+    kind === 'fastest-gpu' ? platform.read.value !== null : platform.costPerMtok !== null,
+  );
+  const sorted = [...measurable].sort((a, b) => {
+    if (kind === 'fastest-gpu') return (b.read.value ?? 0) - (a.read.value ?? 0);
+    return (
+      (a.costPerMtok ?? Number.POSITIVE_INFINITY) - (b.costPerMtok ?? Number.POSITIVE_INFINITY)
+    );
+  });
+  return sorted.map((platform, index) => toRankingRow(platform, index + 1));
+}
+
+/** Human wording for the overview scenario a ranking or run page is read
+ *  from. EN and ZH variants live together so the two locales cannot drift. */
+export function scenarioLabel(scenario: OverviewScenario, locale: 'en' | 'zh'): string {
+  if (scenario === 'agentx') {
+    return locale === 'zh' ? 'AgentX 智能体编码工作负载' : 'the AgentX agentic coding workload';
+  }
+  return locale === 'zh'
+    ? '单轮对话工作负载（8k 输入 / 1k 输出）'
+    : 'a single-turn chat workload (8k input / 1k output)';
+}
+
+// ---------------------------------------------------------------------------
+// English copy
+// ---------------------------------------------------------------------------
+
+export function rankingPageTitle(entry: RankingPageEntry): string {
+  return entry.kind === 'fastest-gpu'
+    ? `Fastest GPU for ${entry.model.seoName} Inference: Live Rankings`
+    : `Cheapest GPU to Run ${entry.model.seoName}: $ per Million Tokens`;
+}
+
+export function rankingPageHeading(entry: RankingPageEntry): string {
+  return entry.kind === 'fastest-gpu'
+    ? `Fastest GPU for ${entry.model.seoName}`
+    : `Cheapest GPU for ${entry.model.seoName}`;
+}
+
+/** Static fallback meta description; the route swaps in a stat-led one when
+ *  live numbers are available. */
+export function rankingPageDescription(entry: RankingPageEntry): string {
+  return entry.kind === 'fastest-gpu'
+    ? `Which GPU serves ${entry.model.seoName} fastest? Live ranking of NVIDIA and AMD hardware by measured tokens per second per GPU, re-benchmarked continuously.`
+    : `Which GPU serves ${entry.model.seoName} cheapest? Live ranking by measured $ per million tokens at hyperscaler GPU pricing, re-benchmarked continuously.`;
+}
+
+export function rankingPageKeywords(entry: RankingPageEntry): string[] {
+  const model = entry.model.seoName;
+  if (entry.kind === 'fastest-gpu') {
+    return [
+      `fastest GPU for ${model}`,
+      `best GPU for ${model}`,
+      `${model} GPU benchmark`,
+      `${model} tokens per second`,
+      `${model} inference speed`,
+      `${model} GPU ranking`,
+      `${model} H100 B200 MI355X comparison`,
+      `${model} inference hardware`,
+    ];
+  }
+  return [
+    `cheapest GPU for ${model}`,
+    `${model} cost per million tokens`,
+    `${model} inference cost`,
+    `cheapest way to run ${model}`,
+    `${model} GPU cost comparison`,
+    `${model} serving cost`,
+    `${model} $ per token`,
+    `${model} inference pricing`,
+  ];
+}
diff --git a/packages/app/src/lib/run-pages-zh.ts b/packages/app/src/lib/run-pages-zh.ts
new file mode 100644
index 000000000..51e678deb
--- /dev/null
+++ b/packages/app/src/lib/run-pages-zh.ts
@@ -0,0 +1,57 @@
+/**
+ * @file run-pages-zh.ts
+ * @description Simplified Chinese copy builders for the `/zh/run/<pair>`
+ * pages. Same structure as the English builders in `run-pages.ts`; model
+ * names, hardware SKUs, and framework names stay in English per
+ * `docs/chinese-copy.md`.
+ */
+
+import type { RunPageEntry } from '@/lib/run-pages';
+
+export function runPageTitleZh(entry: RunPageEntry): string {
+  const chipLabel = entry.chip.label;
+  return `${entry.model.seoName} 在 ${chipLabel} 上的实测推理吞吐与成本`;
+}
+
+export function runPageHeadingZh(entry: RunPageEntry): string {
+  const chipLabel = entry.chip.label;
+  return `在 ${chipLabel} 上运行 ${entry.model.seoName}`;
+}
+
+/** Static fallback meta description; the route swaps in a stat-led one when
+ *  live numbers are available. */
+export function runPageDescriptionZh(entry: RunPageEntry): string {
+  const chipTitle = entry.chip.title;
+  return `${entry.model.seoName} 在 ${chipTitle} 上的实时基准数据：单 GPU 每秒 token 数、每百万 token 成本，以及产生这些结果的推理配置。`;
+}
+
+export function runPageKeywordsZh(entry: RunPageEntry): string[] {
+  const model = entry.model.seoName;
+  const chipLabel = entry.chip.label;
+  return [
+    `${model} ${chipLabel} 基准测试`,
+    `在 ${chipLabel} 上运行 ${model}`,
+    `${model} ${chipLabel} 吞吐量`,
+    `${model} ${chipLabel} 每秒 token 数`,
+    `${model} ${chipLabel} 每百万 token 成本`,
+    `${model} ${chipLabel} 推理性能`,
+    `${model} 推理硬件`,
+    `${model} 部署 ${chipLabel}`,
+  ];
+}
+
+export function runPageFaqQuestionsZh(entry: RunPageEntry): {
+  throughput: string;
+  cost: string;
+  serving: string;
+  methodology: string;
+} {
+  const model = entry.model.seoName;
+  const chipLabel = entry.chip.label;
+  return {
+    throughput: `${model} 在 ${chipLabel} 上能跑多快？`,
+    cost: `在 ${chipLabel} 上部署 ${model} 的成本是多少？`,
+    serving: `哪些推理引擎可以在 ${chipLabel} 上运行 ${model}？`,
+    methodology: `这些 ${model} 数据是如何测得的？`,
+  };
+}
diff --git a/packages/app/src/lib/run-pages.test.ts b/packages/app/src/lib/run-pages.test.ts
new file mode 100644
index 000000000..e4f1e262e
--- /dev/null
+++ b/packages/app/src/lib/run-pages.test.ts
@@ -0,0 +1,134 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  fmtCostPerMtok,
+  fmtGpuHour,
+  fmtMs,
+  fmtThroughput,
+  pctCheaper,
+  pctFaster,
+} from '@/components/live-seo/format';
+import { getAllChipPages } from '@/lib/chip-pages';
+import { INFERENCE_MODEL_SLUGS } from '@/lib/inference-model-slug';
+import {
+  getAllRunPageEntries,
+  getRunPageEntry,
+  runPageDescription,
+  runPageFaqQuestions,
+  runPageHeading,
+  runPageKeywords,
+  runPagePath,
+  runPageTitle,
+} from '@/lib/run-pages';
+import {
+  runPageDescriptionZh,
+  runPageFaqQuestionsZh,
+  runPageHeadingZh,
+  runPageKeywordsZh,
+  runPageTitleZh,
+} from '@/lib/run-pages-zh';
+
+const FORBIDDEN_DASHES = /[\u2013\u2014]/;
+const CJK = /[\u4E00-\u9FFF]/;
+
+describe('run pages registry', () => {
+  it('has one candidate per (model, chip) pair', () => {
+    const entries = getAllRunPageEntries();
+    expect(entries.length).toBe(INFERENCE_MODEL_SLUGS.length * getAllChipPages().length);
+    // 126 = INFERENCE_MODEL_SLUGS.length * getAllChipPages().length. This fork
+    // adds two model slugs upstream doesn't have (dsv2lite, gemma4 — see
+    // .claude/skills/fork-changelog/SKILL.md), so the count is higher than
+    // upstream's.
+    expect(entries.length).toBe(126);
+  });
+
+  it('has unique, well-formed slugs', () => {
+    const entries = getAllRunPageEntries();
+    expect(new Set(entries.map((entry) => entry.slug)).size).toBe(entries.length);
+    for (const entry of entries) {
+      expect(entry.slug).toBe(`${entry.model.slug}-on-${entry.chip.slug}`);
+      expect(entry.slug).toMatch(/^[a-z0-9-]+$/);
+      expect(runPagePath(entry)).toBe(`/run/${entry.slug}`);
+      expect(entry.dbKeys.length).toBeGreaterThan(0);
+    }
+  });
+
+  it('resolves slugs case-insensitively and rejects unknown slugs', () => {
+    const [first] = getAllRunPageEntries();
+    expect(getRunPageEntry(first.slug)).toBe(first);
+    expect(getRunPageEntry(first.slug.toUpperCase())).toBe(first);
+    expect(getRunPageEntry('kimi-k3-on-abacus')).toBeUndefined();
+  });
+});
+
+describe('run pages copy', () => {
+  it('contains no em or en dashes and enough keywords in either locale', () => {
+    for (const entry of getAllRunPageEntries()) {
+      const faq = runPageFaqQuestions(entry);
+      const faqZh = runPageFaqQuestionsZh(entry);
+      const copy = [
+        runPageTitle(entry),
+        runPageHeading(entry),
+        runPageDescription(entry),
+        ...runPageKeywords(entry),
+        faq.throughput,
+        faq.cost,
+        faq.serving,
+        faq.methodology,
+        runPageTitleZh(entry),
+        runPageHeadingZh(entry),
+        runPageDescriptionZh(entry),
+        ...runPageKeywordsZh(entry),
+        faqZh.throughput,
+        faqZh.cost,
+        faqZh.serving,
+        faqZh.methodology,
+      ];
+      for (const text of copy) {
+        expect(text, `dash found in copy for ${entry.slug}: ${text}`).not.toMatch(FORBIDDEN_DASHES);
+        expect(text.length).toBeGreaterThan(0);
+      }
+      expect(runPageKeywords(entry).length).toBeGreaterThanOrEqual(6);
+      expect(runPageKeywordsZh(entry).length).toBeGreaterThanOrEqual(6);
+    }
+  });
+
+  it('keeps Chinese copy Chinese while model and chip names stay English', () => {
+    for (const entry of getAllRunPageEntries()) {
+      expect(runPageTitleZh(entry)).toMatch(CJK);
+      expect(runPageDescriptionZh(entry)).toMatch(CJK);
+      expect(runPageTitleZh(entry)).toContain(entry.model.seoName);
+      expect(runPageTitleZh(entry)).toContain(entry.chip.label);
+    }
+  });
+
+  it('mentions both the model and the chip in every English title', () => {
+    for (const entry of getAllRunPageEntries()) {
+      expect(runPageTitle(entry)).toContain(entry.model.seoName);
+      expect(runPageTitle(entry)).toContain(entry.chip.label);
+    }
+  });
+});
+
+describe('live-seo number formatting', () => {
+  it('formats throughput with sensible precision', () => {
+    expect(fmtThroughput(1234.56)).toBe('1,235');
+    expect(fmtThroughput(250)).toBe('250');
+    expect(fmtThroughput(42.345)).toBe('42.3');
+  });
+
+  it('formats costs at cents precision with a sub-cent fallback', () => {
+    expect(fmtCostPerMtok(1.234)).toBe('$1.23');
+    expect(fmtCostPerMtok(0.056)).toBe('$0.056');
+    expect(fmtGpuHour(2.5)).toBe('$2.50');
+  });
+
+  it('formats latency and percentage deltas', () => {
+    expect(fmtMs(123.4)).toBe('123 ms');
+    expect(fmtMs(45.67)).toBe('45.7 ms');
+    expect(pctFaster(250, 200)).toBe(25);
+    // A $1 winner against a $2 runner-up is 50% below, not 100%.
+    expect(pctCheaper(1, 2)).toBe(50);
+    expect(pctCheaper(1.5, 2)).toBe(25);
+  });
+});
diff --git a/packages/app/src/lib/run-pages.ts b/packages/app/src/lib/run-pages.ts
new file mode 100644
index 000000000..0e089c536
--- /dev/null
+++ b/packages/app/src/lib/run-pages.ts
@@ -0,0 +1,129 @@
+/**
+ * @file run-pages.ts
+ * @description Slug registry and copy builders for the indexable
+ * `/run/<model>-on-<chip>` pages.
+ *
+ * Each page answers one search-shaped question: what does it actually look
+ * like to serve model X on GPU Y today? The dashboard already measures the
+ * answer daily; these pages give every (model, chip) pair a stable URL with
+ * its own title, meta description, and server-rendered numbers so the
+ * measurement is indexable instead of trapped behind dashboard query params.
+ *
+ * The vocabulary deliberately reuses the two registries that already exist:
+ * model slugs come from `INFERENCE_MODEL_SLUGS` (same slugs as `/inference/…`
+ * and `/compare/…`) and chip slugs come from `chip-pages.ts` (same slugs as
+ * `/chips/…`), so `/run/kimi-k3-on-gb200-nvl72`, `/inference/kimi-k3` and
+ * `/chips/gb200-nvl72` all agree on what the words mean. No third naming
+ * scheme is introduced.
+ *
+ * Which pairs actually get indexed is decided by benchmark availability at
+ * request time (see `run-rankings-data.server.ts`); this module only defines
+ * the static candidate space and the copy.
+ */
+
+import { getAllChipPages, type ChipPageEntry } from '@/lib/chip-pages';
+import { COMPARE_MODEL_SLUGS } from '@/lib/compare-slug';
+import { INFERENCE_MODEL_SLUGS, type InferenceModelSlug } from '@/lib/inference-model-slug';
+
+export interface RunPageEntry {
+  /** Canonical URL slug under /run/, e.g. 'kimi-k3-on-gb200-nvl72'. */
+  slug: string;
+  /** Model registry entry — same slug vocabulary as /inference/<model>. */
+  model: InferenceModelSlug;
+  /** DB keys to fetch benchmark rows for, from the compare registry. */
+  dbKeys: string[];
+  /** Chip registry entry — same slug vocabulary as /chips/<slug>. */
+  chip: ChipPageEntry;
+}
+
+const DB_KEYS_BY_MODEL_SLUG: ReadonlyMap<string, string[]> = new Map(
+  COMPARE_MODEL_SLUGS.map((entry) => [entry.slug, entry.dbKeys]),
+);
+
+function buildEntries(): RunPageEntry[] {
+  const out: RunPageEntry[] = [];
+  for (const model of INFERENCE_MODEL_SLUGS) {
+    const dbKeys = DB_KEYS_BY_MODEL_SLUG.get(model.slug);
+    if (!dbKeys) continue;
+    for (const chip of getAllChipPages()) {
+      out.push({
+        slug: `${model.slug}-on-${chip.slug}`,
+        model,
+        dbKeys,
+        chip,
+      });
+    }
+  }
+  return out;
+}
+
+const ENTRIES: readonly RunPageEntry[] = buildEntries();
+
+const ENTRY_BY_SLUG: ReadonlyMap<string, RunPageEntry> = new Map(
+  ENTRIES.map((entry) => [entry.slug, entry]),
+);
+
+/** Every candidate (model, chip) pair. Availability filtering happens in the
+ *  server layer; unknown slugs 404 at the route. */
+export function getAllRunPageEntries(): readonly RunPageEntry[] {
+  return ENTRIES;
+}
+
+export function getRunPageEntry(slug: string): RunPageEntry | undefined {
+  return ENTRY_BY_SLUG.get(slug.toLowerCase());
+}
+
+export function runPagePath(entry: RunPageEntry): string {
+  return `/run/${entry.slug}`;
+}
+
+// ---------------------------------------------------------------------------
+// English copy
+// ---------------------------------------------------------------------------
+
+export function runPageTitle(entry: RunPageEntry): string {
+  return `${entry.model.seoName} on ${entry.chip.label}: Measured Inference Throughput and Cost`;
+}
+
+export function runPageHeading(entry: RunPageEntry): string {
+  return `Running ${entry.model.seoName} on ${entry.chip.label}`;
+}
+
+/** Static fallback meta description; the route swaps in a stat-led one when
+ *  live numbers are available. Keep ≤160 chars before the supporters line. */
+export function runPageDescription(entry: RunPageEntry): string {
+  return `Live benchmark data for ${entry.model.seoName} inference on ${entry.chip.title}: tokens per second per GPU, cost per million tokens, and the serving configs that produced them.`;
+}
+
+export function runPageKeywords(entry: RunPageEntry): string[] {
+  const model = entry.model.seoName;
+  const chip = entry.chip.label;
+  return [
+    `${model} on ${chip}`,
+    `run ${model} on ${chip}`,
+    `${model} ${chip} benchmark`,
+    `${model} ${chip} throughput`,
+    `${model} ${chip} tokens per second`,
+    `${model} ${chip} cost per million tokens`,
+    `${chip} ${model} inference performance`,
+    `${model} inference hardware`,
+  ];
+}
+
+/** Search-shaped FAQ scaffolding. Answers that need live numbers are filled
+ *  by the route; these are the stable question strings. */
+export function runPageFaqQuestions(entry: RunPageEntry): {
+  throughput: string;
+  cost: string;
+  serving: string;
+  methodology: string;
+} {
+  const model = entry.model.seoName;
+  const chip = entry.chip.label;
+  return {
+    throughput: `How fast is ${model} on ${chip}?`,
+    cost: `How much does it cost to serve ${model} on ${chip}?`,
+    serving: `Which serving engines run ${model} on ${chip}?`,
+    methodology: `How are these ${model} numbers measured?`,
+  };
+}
diff --git a/packages/app/src/lib/run-rankings-data.server.ts b/packages/app/src/lib/run-rankings-data.server.ts
new file mode 100644
index 000000000..9eb155567
--- /dev/null
+++ b/packages/app/src/lib/run-rankings-data.server.ts
@@ -0,0 +1,268 @@
+/**
+ * @file run-rankings-data.server.ts
+ * @description Server-side data derivation for the `/run/<pair>` and
+ * `/rankings/<slug>` pages.
+ *
+ * Both families read through `getCachedBenchmarks` — the same cached query the
+ * dashboard, /compare, and /overview use — and derive their numbers with the
+ * overview's own config building and read selection (`buildOverviewModelSummary`
+ * for rankings, `overviewHardwareTierReader` for run pages, which also covers
+ * SKUs outside the overview matrix), so a ranked or quoted figure on these
+ * pages is always the figure the overview derivation yields for that cell.
+ * Fixtures mode short-circuits inside those shared helpers, so tests and
+ * DB-less builds never touch a connection here.
+ */
+
+import { resolveFrameworkPartLabel } from '@semianalysisai/inferencex-constants';
+
+import type { BenchmarkRow } from '@/lib/api';
+import { cachedDerivedData } from '@/lib/api-cache';
+import { getCachedBenchmarks } from '@/lib/benchmark-data.server';
+import { getChipHw } from '@/lib/chip-pages';
+import { buildDbKeyToSlugMap, getCachedAvailability } from '@/lib/compare-availability';
+import {
+  buildOverviewModelSummary,
+  OVERVIEW_PRIMARY_TIER,
+  OVERVIEW_TIERS,
+  overviewHardwareTierReader,
+  overviewHeadlineScenarioForModel,
+  type OverviewScenario,
+  type OverviewTier,
+} from '@/lib/overview-data';
+import {
+  buildRankingRows,
+  getAllRankingPageEntries,
+  type RankingPageEntry,
+  type RankingRow,
+} from '@/lib/rankings';
+import { getAllRunPageEntries, type RunPageEntry } from '@/lib/run-pages';
+
+// ---------------------------------------------------------------------------
+// Availability: which (model, chip) pairs actually have benchmark rows
+// ---------------------------------------------------------------------------
+
+async function loadAvailableRunEntries(): Promise<RunPageEntry[]> {
+  const rows = await getCachedAvailability();
+  const dbKeyToSlug = buildDbKeyToSlugMap();
+  const availableByModelSlug = new Map<string, Set<string>>();
+  for (const row of rows) {
+    const modelSlug = dbKeyToSlug.get(row.model);
+    if (!modelSlug) continue;
+    let set = availableByModelSlug.get(modelSlug);
+    if (!set) {
+      set = new Set<string>();
+      availableByModelSlug.set(modelSlug, set);
+    }
+    set.add(row.hardware);
+  }
+  return getAllRunPageEntries().filter((entry) =>
+    availableByModelSlug.get(entry.model.slug)?.has(entry.chip.hwKey),
+  );
+}
+
+const getCachedAvailableRunEntries = cachedDerivedData(
+  loadAvailableRunEntries,
+  'run-pages-availability-v1',
+);
+
+/** Candidate pairs narrowed to those with at least one benchmark row. Used by
+ *  the sitemap, the /run index, and related-page link lists. */
+export function getAvailableRunEntries(): Promise<RunPageEntry[]> {
+  return getCachedAvailableRunEntries();
+}
+
+// ---------------------------------------------------------------------------
+// /run/<pair> page data
+// ---------------------------------------------------------------------------
+
+export interface RunTierRead {
+  tier: OverviewTier;
+  /** Tokens/s per GPU at this interactivity tier; null when unreachable. */
+  throughputPerGpu: number | null;
+  /** $ per million total tokens at hyperscaler $/GPU/hr; null when unknown. */
+  costPerMtok: number | null;
+  precision: string | null;
+  framework: string | null;
+  disagg: boolean | null;
+}
+
+export interface RunCostTier {
+  tierLabel: 'hyperscaler' | 'neocloud' | 'retail';
+  costPerGpuHour: number;
+  costPerMtok: number;
+}
+
+export interface RunPageData {
+  hasData: boolean;
+  scenario: OverviewScenario;
+  configCount: number;
+  frameworks: string[];
+  precisions: string[];
+  hasDisagg: boolean;
+  hasMultinode: boolean;
+  bestThroughputPerGpu: number | null;
+  bestMedianTtft: number | null;
+  bestMedianTpot: number | null;
+  primaryTier: OverviewTier;
+  tierLadder: RunTierRead[];
+  /** Cost per million tokens at the primary tier across the three $/GPU/hr
+   *  pricing tiers from HW_REGISTRY. Empty when the tier is unreachable. */
+  costTiers: RunCostTier[];
+  oldest: string | null;
+  newest: string | null;
+}
+
+function costPerMtokAt(costPerGpuHour: number, throughputPerGpu: number): number {
+  return (costPerGpuHour * 1_000_000) / (throughputPerGpu * 3600);
+}
+
+function buildRunPageData(entry: RunPageEntry, rows: BenchmarkRow[]): RunPageData {
+  const hwKey = entry.chip.hwKey;
+  const hwRows = rows.filter((row) => row.hardware === hwKey);
+  const scenario = overviewHeadlineScenarioForModel(entry.model.model, rows);
+
+  const frameworks = new Set<string>();
+  const precisions = new Set<string>();
+  let hasDisagg = false;
+  let hasMultinode = false;
+  let bestThroughputPerGpu: number | null = null;
+  let bestMedianTtft: number | null = null;
+  let bestMedianTpot: number | null = null;
+  let oldest: string | null = null;
+  let newest: string | null = null;
+
+  for (const row of hwRows) {
+    frameworks.add(resolveFrameworkPartLabel(row.model, row.framework));
+    precisions.add(row.precision);
+    if (row.disagg) hasDisagg = true;
+    if (row.is_multinode) hasMultinode = true;
+    const metrics = row.metrics ?? {};
+    const tput = typeof metrics.tput_per_gpu === 'number' ? metrics.tput_per_gpu : null;
+    const ttft = typeof metrics.median_ttft === 'number' ? metrics.median_ttft : null;
+    const tpot = typeof metrics.median_tpot === 'number' ? metrics.median_tpot : null;
+    if (tput !== null && (bestThroughputPerGpu === null || tput > bestThroughputPerGpu)) {
+      bestThroughputPerGpu = tput;
+    }
+    if (ttft !== null && (bestMedianTtft === null || ttft < bestMedianTtft)) {
+      bestMedianTtft = ttft;
+    }
+    if (tpot !== null && (bestMedianTpot === null || tpot < bestMedianTpot)) {
+      bestMedianTpot = tpot;
+    }
+    if (row.date && (oldest === null || row.date < oldest)) oldest = row.date;
+    if (row.date && (newest === null || row.date > newest)) newest = row.date;
+  }
+
+  // Reads come from the same derivation as the overview matrix but without
+  // its fixed platform list, so H100-class SKUs get real ladders too.
+  const readAt = overviewHardwareTierReader(entry.model.model, rows, scenario);
+  const tierLadder: RunTierRead[] = OVERVIEW_TIERS.map((tier) => {
+    const { read, costPerMtok } = readAt(hwKey, tier);
+    return {
+      tier,
+      throughputPerGpu: read.value,
+      costPerMtok,
+      precision: read.config?.precision ?? null,
+      framework: read.config?.frameworkLabel ?? null,
+      disagg: read.config?.disagg ?? null,
+    };
+  });
+
+  const primaryRead = tierLadder.find((read) => read.tier === OVERVIEW_PRIMARY_TIER);
+  const hw = getChipHw(entry.chip);
+  const costTiers: RunCostTier[] =
+    primaryRead?.throughputPerGpu && primaryRead.throughputPerGpu > 0
+      ? (
+          [
+            ['hyperscaler', hw.costh],
+            ['neocloud', hw.costn],
+            ['retail', hw.costr],
+          ] as const
+        )
+          .filter(([, costPerGpuHour]) => costPerGpuHour > 0)
+          .map(([tierLabel, costPerGpuHour]) => ({
+            tierLabel,
+            costPerGpuHour,
+            costPerMtok: costPerMtokAt(costPerGpuHour, primaryRead.throughputPerGpu as number),
+          }))
+      : [];
+
+  return {
+    hasData: hwRows.length > 0,
+    scenario,
+    configCount: hwRows.length,
+    frameworks: [...frameworks].sort(),
+    precisions: [...precisions].sort(),
+    hasDisagg,
+    hasMultinode,
+    bestThroughputPerGpu,
+    bestMedianTtft,
+    bestMedianTpot,
+    primaryTier: OVERVIEW_PRIMARY_TIER,
+    tierLadder,
+    costTiers,
+    oldest,
+    newest,
+  };
+}
+
+async function loadRunPageData(slug: string): Promise<RunPageData | null> {
+  const entry = getAllRunPageEntries().find((candidate) => candidate.slug === slug);
+  if (!entry) return null;
+  const rows = await getCachedBenchmarks(entry.dbKeys);
+  return buildRunPageData(entry, rows);
+}
+
+const getCachedRunPageData = cachedDerivedData(loadRunPageData, 'run-page-data-v2');
+
+export function getRunPageData(slug: string): Promise<RunPageData | null> {
+  return getCachedRunPageData(slug);
+}
+
+// ---------------------------------------------------------------------------
+// /rankings/<slug> page data
+// ---------------------------------------------------------------------------
+
+export interface RankingPageData {
+  scenario: OverviewScenario;
+  tier: OverviewTier;
+  rows: RankingRow[];
+  oldest: string | null;
+  newest: string | null;
+}
+
+async function loadRankingPageData(slug: string): Promise<RankingPageData | null> {
+  const entry: RankingPageEntry | undefined = getAllRankingPageEntries().find(
+    (candidate) => candidate.slug === slug,
+  );
+  if (!entry) return null;
+  const rows = await getCachedBenchmarks(entry.dbKeys);
+  const scenario = overviewHeadlineScenarioForModel(entry.model.model, rows);
+  const summary = buildOverviewModelSummary(
+    entry.model.model,
+    rows,
+    OVERVIEW_PRIMARY_TIER,
+    'community',
+    scenario,
+  );
+  let oldest: string | null = null;
+  let newest: string | null = null;
+  for (const row of rows) {
+    if (!row.date) continue;
+    if (oldest === null || row.date < oldest) oldest = row.date;
+    if (newest === null || row.date > newest) newest = row.date;
+  }
+  return {
+    scenario: summary.scenario,
+    tier: OVERVIEW_PRIMARY_TIER,
+    rows: buildRankingRows(summary, entry.kind),
+    oldest,
+    newest,
+  };
+}
+
+const getCachedRankingPageData = cachedDerivedData(loadRankingPageData, 'ranking-page-data-v2');
+
+export function getRankingPageData(slug: string): Promise<RankingPageData | null> {
+  return getCachedRankingPageData(slug);
+}
diff --git a/packages/app/src/lib/speed-overlay.test.ts b/packages/app/src/lib/speed-overlay.test.ts
deleted file mode 100644
index b69aea49f..000000000
--- a/packages/app/src/lib/speed-overlay.test.ts
+++ /dev/null
@@ -1,43 +0,0 @@
-import { describe, expect, it } from 'vitest';
-
-import { getSpeedOverlayCorners } from './speed-overlay';
-
-describe('getSpeedOverlayCorners', () => {
-  it('places the bus top-left for upper_left rooflines (e.g. tput vs interactivity)', () => {
-    expect(getSpeedOverlayCorners('upper_left')).toEqual({
-      busTop: true,
-      busLeft: true,
-    });
-  });
-
-  it('places the bus top-right for upper_right rooflines (e.g. tput vs e2e or vs TTFT)', () => {
-    expect(getSpeedOverlayCorners('upper_right')).toEqual({
-      busTop: true,
-      busLeft: false,
-    });
-  });
-
-  it('places the bus bottom-right for lower_left rooflines (e.g. cost vs e2e)', () => {
-    // The optimal corner the roofline points to is unreachable for cost charts;
-    // the batchy endpoint sits at the X-flipped corner — high latency, low cost.
-    expect(getSpeedOverlayCorners('lower_left')).toEqual({
-      busTop: false,
-      busLeft: false,
-    });
-  });
-
-  it('places the bus bottom-left for lower_right rooflines (e.g. cost vs interactivity)', () => {
-    // Batchy endpoint = low interactivity (left), low cost (bottom).
-    expect(getSpeedOverlayCorners('lower_right')).toEqual({
-      busTop: false,
-      busLeft: true,
-    });
-  });
-
-  it('falls back to top-left for an undefined roofline direction', () => {
-    expect(getSpeedOverlayCorners(undefined)).toEqual({
-      busTop: true,
-      busLeft: true,
-    });
-  });
-});
diff --git a/packages/app/src/lib/speed-overlay.ts b/packages/app/src/lib/speed-overlay.ts
deleted file mode 100644
index f95f76e8f..000000000
--- a/packages/app/src/lib/speed-overlay.ts
+++ /dev/null
@@ -1,34 +0,0 @@
-/**
- * Bus = batch-friendly endpoint of the Pareto frontier.
- * Car = interactivity-friendly opposite endpoint.
- *
- * The roofline direction in `chartDefinition` is post-flipped by `useChartData`
- * when the X axis is swapped (e.g. interactivity chart input metrics showing
- * TTFT), so it always reflects the effective axis directions:
- *   upper_left  → bus top-left      (e.g. tput vs interactivity)
- *   upper_right → bus top-right     (e.g. tput vs e2e or vs TTFT)
- *   lower_left  → bus bottom-right  (e.g. cost vs e2e — batchy = high latency)
- *   lower_right → bus bottom-left   (e.g. cost vs interactivity — batchy = low int)
- *
- * Y matches the upper/lower part directly. X matches the left/right part for
- * throughput-style metrics ("upper" rooflines) and is flipped for cost / energy
- * metrics ("lower" rooflines), because the batchy endpoint sits opposite the
- * unreachable "optimal" corner the roofline name points to.
- */
-export type RooflineDirection = 'upper_left' | 'upper_right' | 'lower_left' | 'lower_right';
-
-export interface SpeedOverlayCorners {
-  busTop: boolean;
-  busLeft: boolean;
-}
-
-export function getSpeedOverlayCorners(
-  rooflineDirection: RooflineDirection | undefined,
-): SpeedOverlayCorners {
-  const isUpperY = rooflineDirection?.startsWith('upper') ?? true;
-  const rooflineLeft = rooflineDirection?.endsWith('left') ?? true;
-  return {
-    busTop: isUpperY,
-    busLeft: isUpperY ? rooflineLeft : !rooflineLeft,
-  };
-}
diff --git a/packages/app/src/lib/star-storage.ts b/packages/app/src/lib/star-storage.ts
index eaf1872c8..6357449c5 100644
--- a/packages/app/src/lib/star-storage.ts
+++ b/packages/app/src/lib/star-storage.ts
@@ -1,7 +1,5 @@
 export const STARRED_KEY = 'inferencex-starred';
 export const STARRED_EVENT = 'inferencex:starred';
-export const DISMISS_KEY = 'inferencex-star-modal-dismissed';
-export const DISMISS_DURATION_MS = 7 * 24 * 60 * 60 * 1000; // 1 week
 
 export function saveStarred(): void {
   try {
@@ -11,11 +9,3 @@ export function saveStarred(): void {
   }
   window.dispatchEvent(new Event(STARRED_EVENT));
 }
-
-export function saveDismissTimestamp(): void {
-  try {
-    localStorage.setItem(DISMISS_KEY, String(Date.now()));
-  } catch {
-    // localStorage unavailable
-  }
-}
diff --git a/packages/app/src/lib/submissions-types.ts b/packages/app/src/lib/submissions-types.ts
index ff840419c..038d1dfc3 100644
--- a/packages/app/src/lib/submissions-types.ts
+++ b/packages/app/src/lib/submissions-types.ts
@@ -1,30 +1,9 @@
-export interface SubmissionSummaryRow {
-  model: string;
-  hardware: string;
-  framework: string;
-  precision: string;
-  spec_method: string;
-  disagg: boolean;
-  is_multinode: boolean;
-  num_prefill_gpu: number;
-  num_decode_gpu: number;
-  prefill_tp: number;
-  prefill_ep: number;
-  decode_tp: number;
-  decode_ep: number;
-  date: string;
-  total_datapoints: number;
-  distinct_sequences: number;
-  distinct_concurrencies: number;
-  max_concurrency: number;
-  image: string | null;
-}
+import type {
+  SubmissionSummaryRow,
+  SubmissionVolumeRow,
+} from '@semianalysisai/inferencex-db/queries/submissions';
 
-export interface SubmissionVolumeRow {
-  date: string;
-  hardware: string;
-  datapoints: number;
-}
+export type { SubmissionSummaryRow, SubmissionVolumeRow };
 
 export interface SubmissionsResponse {
   summary: SubmissionSummaryRow[];
diff --git a/packages/app/src/lib/supplemental-benchmarks.test.ts b/packages/app/src/lib/supplemental-benchmarks.test.ts
new file mode 100644
index 000000000..caca92c71
--- /dev/null
+++ b/packages/app/src/lib/supplemental-benchmarks.test.ts
@@ -0,0 +1,84 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  SUPPLEMENTAL_BENCHMARK_ROWS,
+  supportsChartTokenMetric,
+  supportsTokenMetric,
+  withSupplementalAvailability,
+  withSupplementalBenchmarkHistory,
+  withSupplementalBenchmarks,
+} from './supplemental-benchmarks';
+
+describe('supplemental benchmark snapshots', () => {
+  it('ships every supplied Jalapeño and July VR200 point', () => {
+    expect(SUPPLEMENTAL_BENCHMARK_ROWS).toHaveLength(50);
+    expect(SUPPLEMENTAL_BENCHMARK_ROWS.filter((row) => row.hardware === 'jalapeno')).toHaveLength(
+      36,
+    );
+    expect(SUPPLEMENTAL_BENCHMARK_ROWS.filter((row) => row.hardware === 'vr200')).toHaveLength(14);
+    expect(
+      new Set(
+        SUPPLEMENTAL_BENCHMARK_ROWS.filter((row) => row.hardware === 'vr200').map(
+          (row) => row.run_url,
+        ),
+      ),
+    ).toEqual(
+      new Set([
+        'https://www.coreweave.com/blog/nvidia-vera-rubin-nvl72-on-coreweave-10x-more-tokens-per-megawatt-than-blackwell',
+      ]),
+    );
+  });
+
+  it('resolves the newest snapshot independently per hardware curve', () => {
+    const latest = withSupplementalBenchmarks([], { model: 'dsr1' });
+    expect(latest.filter((row) => row.hardware === 'jalapeno')).toHaveLength(9);
+    expect(latest.filter((row) => row.hardware === 'vr200')).toHaveLength(14);
+
+    const august17 = withSupplementalBenchmarks([], {
+      model: 'dsr1',
+      date: '2026-08-17',
+      exact: true,
+    });
+    expect(august17).toHaveLength(8);
+    expect(new Set(august17.map((row) => row.date))).toEqual(new Set(['2026-08-17']));
+  });
+
+  it('merges idempotently and exposes complete history and availability', () => {
+    const history = withSupplementalBenchmarkHistory([], { model: 'dsr1', isl: 8192, osl: 1024 });
+    expect(history).toHaveLength(31);
+    expect(
+      withSupplementalBenchmarkHistory(history, { model: 'dsr1', isl: 8192, osl: 1024 }),
+    ).toHaveLength(31);
+    expect(withSupplementalBenchmarkHistory([], { model: 'dsr1', isl: 1024, osl: 1024 })).toEqual(
+      [],
+    );
+
+    const availability = withSupplementalAvailability([]);
+    expect(availability).toHaveLength(5);
+    expect(withSupplementalAvailability(availability)).toHaveLength(5);
+  });
+
+  it('limits only the July VR200 snapshot to output-token metrics', () => {
+    const julyVr = SUPPLEMENTAL_BENCHMARK_ROWS.find((row) => row.hardware === 'vr200');
+    expect(julyVr).toBeDefined();
+    expect(supportsTokenMetric(julyVr!, 'output')).toBe(true);
+    expect(supportsTokenMetric(julyVr!, 'total')).toBe(false);
+    expect(supportsTokenMetric(julyVr!, 'input')).toBe(false);
+    expect(supportsChartTokenMetric('vr200_rubin-july', '2026-07-01', 'output')).toBe(true);
+    expect(supportsChartTokenMetric('vr200_coreweave-vera-rubin', '2026-07-01', 'total')).toBe(
+      false,
+    );
+  });
+
+  it('leaves future VR200 snapshots unrestricted until metadata says otherwise', () => {
+    const futureVr = {
+      ...SUPPLEMENTAL_BENCHMARK_ROWS.find((row) => row.hardware === 'vr200')!,
+      date: '2026-09-01',
+    };
+    expect(supportsTokenMetric(futureVr, 'total')).toBe(true);
+    expect(supportsTokenMetric(futureVr, 'input')).toBe(true);
+    expect(supportsChartTokenMetric('vr200_coreweave-vera-rubin', '2026-09-01', 'total')).toBe(
+      true,
+    );
+  });
+});
diff --git a/packages/app/src/lib/supplemental-benchmarks.ts b/packages/app/src/lib/supplemental-benchmarks.ts
new file mode 100644
index 000000000..414d19e66
--- /dev/null
+++ b/packages/app/src/lib/supplemental-benchmarks.ts
@@ -0,0 +1,445 @@
+import type { AvailabilityRow, BenchmarkRow } from '@/lib/api';
+
+export type TokenMetricType = 'total' | 'input' | 'output';
+
+type PointTuple = readonly [
+  tp: number,
+  ep: number,
+  chips: number,
+  concurrency: number,
+  totalThroughputPerGpu: number,
+  outputThroughputPerGpu: number,
+  inputThroughputPerGpu: number,
+  medianInteractivity: number,
+  medianTtft: number,
+  medianE2el: number,
+];
+
+interface SupplementalDataset {
+  id: string;
+  model: string;
+  modelAliases: readonly string[];
+  date: string;
+  hardware: string;
+  framework: string;
+  /** Framework reported by the source when a display-only snapshot key is used. */
+  sourceFramework?: string;
+  runUrl?: string;
+  points: readonly PointTuple[];
+  /** Omitted means the snapshot is valid for every token metric. */
+  supportedTokenMetrics?: readonly TokenMetricType[];
+}
+
+const dsr1August17: readonly PointTuple[] = [
+  [
+    8, 1, 8, 1, 653.3902134366368, 72.99208842404813, 580.3981250125887, 701.906635545054,
+    0.26185680300113745, 1.5685025695420336,
+  ],
+  [
+    8, 1, 8, 2, 719.5361536453969, 80.74272299874106, 638.7934306466558, 403.43094610437316,
+    0.5063663404725958, 2.8306995255115908,
+  ],
+  [
+    8, 1, 8, 4, 898.0597363189065, 99.90203743397365, 798.1576988849329, 272.2814518471643,
+    1.0072254485276062, 4.324981361016398,
+  ],
+  [
+    8, 1, 8, 8, 1534.6072863182922, 172.51574204983257, 1362.0915442684595, 209.44326150211057,
+    0.5763577754842117, 4.98794364547939,
+  ],
+  [
+    8, 1, 8, 64, 5048.231029483235, 560.0571750236073, 4488.173854459627, 115.21578894050089,
+    4.323833600996295, 12.300118638988351,
+  ],
+  [
+    8, 1, 8, 128, 6709.483077637271, 743.1792481177831, 5966.303829519487, 91.24036469462176,
+    8.597232194500975, 18.66097929701209,
+  ],
+  [
+    8, 1, 8, 256, 8161.618704995656, 907.1073300236875, 7254.5113749719685, 66.7159336839182,
+    17.12080494550173, 30.95121552349883,
+  ],
+  [
+    8, 1, 8, 512, 9313.852074081971, 1034.864099128346, 8278.987974953625, 45.247588625114666,
+    34.22868340747664, 54.62980595699628,
+  ],
+];
+
+const dsr1August22: readonly PointTuple[] = [
+  [
+    8, 1, 8, 1, 628.9710598745098, 70.26415497877618, 558.7069048957336, 703.1202284701499,
+    0.32318450049933745, 1.6413020644977223,
+  ],
+  [
+    8, 1, 8, 2, 702.7172255982504, 78.8553876069523, 623.8618379912981, 404.3055549907138,
+    0.5780675065034302, 2.9056949085061206,
+  ],
+  [
+    32, 1, 32, 32, 1920.4598362575473, 214.7608426420231, 1705.6989936155242, 264.4765740689175,
+    0.7742451535305008, 4.281998021528125,
+  ],
+  [
+    32, 1, 32, 64, 3178.9840936206497, 352.68054106874285, 2826.3035525519067, 230.35672017720918,
+    1.1666485004825518, 5.184136075986316,
+  ],
+  [
+    32, 1, 32, 128, 5145.186589589582, 569.9091654052471, 4575.277424184334, 206.5912495153872,
+    1.9487936009827536, 6.399220873485319,
+  ],
+  [
+    32, 1, 32, 256, 7793.101906141766, 866.1492429627697, 6926.952663178997, 182.86398603089907,
+    3.3710869964852463, 8.456233689998044,
+  ],
+  [
+    32, 1, 32, 512, 10210.271688700279, 1134.465474536103, 9075.806214164175, 141.59599290224634,
+    6.451180875505088, 12.923200229502982,
+  ],
+  [
+    32, 1, 32, 1024, 12380.114048554029, 1375.8503936714096, 11004.26365488262, 101.64418129373612,
+    12.28093515749788, 21.280333061004058,
+  ],
+  [
+    32, 1, 32, 2048, 13748.393106670877, 1528.9868997017807, 12219.406206969097, 65.25477348593513,
+    23.823017231508857, 38.11095889197895,
+  ],
+];
+
+const gptOssAugust22: readonly PointTuple[] = [
+  [
+    4, 1, 4, 1, 2016.798016173172, 225.30227129616, 1791.4957448770122, 1462.120791184631,
+    0.3949238510103896, 1.02815705849207,
+  ],
+  [
+    2, 1, 2, 1, 2939.497127863335, 328.37962654921716, 2611.1175013141183, 1234.6147538275834,
+    0.6590746759902686, 1.4149531370203476,
+  ],
+  [
+    1, 1, 1, 1, 6503.9223998890475, 726.5717624065494, 5777.350637482498, 832.4994039144032,
+    0.090197357989382, 1.1803411214932567,
+  ],
+  [
+    1, 1, 1, 2, 10183.144438915819, 1142.7011784216138, 9040.443260494205, 694.4545223177703,
+    0.18758124650048558, 1.5426225175178843,
+  ],
+  [
+    1, 8, 8, 64, 19940.018979223085, 2212.1710821471593, 17727.847897075924, 431.06024646996104,
+    0.9552501139987726, 3.103124447487062,
+  ],
+  [
+    1, 8, 8, 128, 29671.80995053509, 3286.612866322006, 26385.197084213083, 370.35158862710585,
+    1.7109653410152532, 4.182030279975152,
+  ],
+  [
+    1, 8, 8, 256, 40678.2706249172, 4521.107734910936, 36157.162890006264, 304.44917501143885,
+    3.126444708497729, 6.147738145489711,
+  ],
+  [
+    1, 8, 8, 512, 50120.27941471215, 5568.87498233962, 44551.40443237253, 231.6343673144642,
+    6.035206183005357, 10.032349768007407,
+  ],
+  [
+    1, 8, 8, 1024, 56208.878247329485, 6246.712022289421, 49962.16622504006, 157.04782949676618,
+    12.126750790514052, 17.998166643985314,
+  ],
+  [
+    1, 8, 8, 2048, 59813.91166996877, 6652.027379034343, 53161.88429093443, 97.15511241159633,
+    24.430557979969308, 33.96740277999197,
+  ],
+];
+
+const kimiAugust22: readonly PointTuple[] = [
+  [
+    8, 1, 8, 1, 663.8650823392342, 74.1622659710108, 589.7028163682234, 694.585687233185,
+    0.2223843564861454, 1.541665184951853,
+  ],
+  [
+    8, 1, 8, 2, 717.1141136395695, 80.4709338686305, 636.643179770939, 392.84456394354004,
+    0.4723154735402204, 2.8553318615304306,
+  ],
+  [
+    8, 1, 8, 4, 906.2857383609748, 100.8171150515352, 805.4686233094396, 269.0823232269083,
+    1.088551064953208, 4.511279692465905,
+  ],
+  [
+    32, 1, 32, 64, 2882.197362494171, 319.75464341302455, 2562.4427190811466, 205.59720404498236,
+    1.2084655475337058, 5.713703882531263,
+  ],
+  [
+    32, 1, 32, 128, 4643.181506753474, 514.3043213812836, 4128.877185372191, 183.2102977635607,
+    2.0762837954680435, 7.0896255720290355,
+  ],
+  [
+    32, 1, 32, 256, 6999.716208139327, 777.9699236135996, 6221.746284525728, 160.5242615595079,
+    3.6558798734331504, 9.40580550848972,
+  ],
+  [
+    32, 1, 32, 512, 9340.25110530568, 1037.79730114271, 8302.453804162971, 127.58406890202056,
+    6.881976359465625, 14.142705057514831,
+  ],
+  [
+    32, 1, 32, 1024, 11341.76404400966, 1260.4545049971903, 10081.309539012469, 92.56077732029165,
+    13.268606896046549, 23.27447518700501,
+  ],
+  [
+    32, 1, 32, 2048, 12767.050719647294, 1419.8498069346924, 11347.2009127126, 60.27802288265687,
+    25.96076739806449, 41.372758659534156,
+  ],
+];
+
+const vrJuly: readonly PointTuple[] = [
+  [72, 1, 72, 52, 2793, 4438.21, 0, 52, 0, 0],
+  [72, 1, 72, 70, 2730, 4338.1, 0, 70, 0, 0],
+  [72, 1, 72, 90, 2478, 3937.66, 0, 90, 0, 0],
+  [72, 1, 72, 110, 2205, 3503.85, 0, 110, 0, 0],
+  [72, 1, 72, 130, 1932, 3070.04, 0, 130, 0, 0],
+  [72, 1, 72, 148, 1680, 2669.6, 0, 148, 0, 0],
+  [72, 1, 72, 170, 1344, 2135.68, 0, 170, 0, 0],
+  [72, 1, 72, 185, 1050, 1668.5, 0, 185, 0, 0],
+  [72, 1, 72, 200, 630, 1001.1, 0, 200, 0, 0],
+  [72, 1, 72, 215, 462, 734.14, 0, 215, 0, 0],
+  [72, 1, 72, 235, 346.5, 550.605, 0, 235, 0, 0],
+  [72, 1, 72, 270, 252, 400.44, 0, 270, 0, 0],
+  [72, 1, 72, 320, 178.5, 283.645, 0, 320, 0, 0],
+  [72, 1, 72, 400, 115.5, 183.535, 0, 400, 0, 0],
+];
+
+const DATASETS: readonly SupplementalDataset[] = [
+  {
+    id: 'jalapeno-dsr1-2026-08-17',
+    model: 'dsr1',
+    modelAliases: ['dsr1', 'DeepSeek-R1-0528'],
+    date: '2026-08-17',
+    hardware: 'jalapeno',
+    framework: 'teacup',
+    points: dsr1August17,
+  },
+  {
+    id: 'jalapeno-dsr1-2026-08-22',
+    model: 'dsr1',
+    modelAliases: ['dsr1', 'DeepSeek-R1-0528'],
+    date: '2026-08-22',
+    hardware: 'jalapeno',
+    framework: 'teacup',
+    points: dsr1August22,
+  },
+  {
+    id: 'jalapeno-gptoss-2026-08-22',
+    model: 'gptoss120b',
+    modelAliases: ['gptoss120b', 'gpt-oss-120b'],
+    date: '2026-08-22',
+    hardware: 'jalapeno',
+    framework: 'teacup',
+    points: gptOssAugust22,
+  },
+  {
+    id: 'jalapeno-kimi-2026-08-22',
+    model: 'kimik2.5',
+    modelAliases: ['kimik2.5', 'Kimi-K2.5'],
+    date: '2026-08-22',
+    hardware: 'jalapeno',
+    framework: 'teacup',
+    points: kimiAugust22,
+  },
+  {
+    id: 'vr200-dsr1-2026-07-01',
+    model: 'dsr1',
+    modelAliases: ['dsr1', 'DeepSeek-R1-0528'],
+    date: '2026-07-01',
+    hardware: 'vr200',
+    // A snapshot-specific key keeps "(July)" attached only to these rows. A
+    // later DB-ingested `coreweave-vera-rubin` run receives its own identity.
+    framework: 'rubin-july',
+    sourceFramework: 'coreweave-vera-rubin',
+    runUrl:
+      'https://www.coreweave.com/blog/nvidia-vera-rubin-nvl72-on-coreweave-10x-more-tokens-per-megawatt-than-blackwell',
+    points: vrJuly,
+    // This July snapshot published output throughput against interactivity. Its
+    // stored total rate is not semantically comparable to the dashboard's total
+    // or input-token metrics. New snapshots are unrestricted unless they opt in.
+    supportedTokenMetrics: ['output'],
+  },
+];
+
+function toBenchmarkRow(dataset: SupplementalDataset, point: PointTuple): BenchmarkRow {
+  const [tp, ep, chips, conc, total, output, input, intvty, ttft, e2el] = point;
+  const reciprocalLatency = intvty > 0 ? 1 / intvty : 0;
+  return {
+    id: 0,
+    hardware: dataset.hardware,
+    framework: dataset.framework,
+    model: dataset.model,
+    precision: 'fp4',
+    spec_method: 'none',
+    disagg: false,
+    is_multinode: chips > 8,
+    prefill_tp: tp,
+    prefill_ep: ep,
+    prefill_dp_attention: false,
+    prefill_num_workers: 1,
+    decode_tp: tp,
+    decode_ep: ep,
+    decode_dp_attention: false,
+    decode_num_workers: 1,
+    num_prefill_gpu: chips,
+    num_decode_gpu: chips,
+    benchmark_type: 'single_turn',
+    isl: 8192,
+    osl: 1024,
+    conc,
+    offload_mode: 'off',
+    image: null,
+    metrics: {
+      tput_per_gpu: total,
+      output_tput_per_gpu: output,
+      input_tput_per_gpu: input,
+      median_intvty: intvty,
+      mean_intvty: intvty,
+      p90_intvty: intvty,
+      p95_intvty: intvty,
+      p99_intvty: intvty,
+      median_tpot: reciprocalLatency,
+      mean_tpot: reciprocalLatency,
+      p90_tpot: reciprocalLatency,
+      p95_tpot: reciprocalLatency,
+      p99_tpot: reciprocalLatency,
+      median_ttft: ttft,
+      mean_ttft: ttft,
+      p90_ttft: ttft,
+      p95_ttft: ttft,
+      p99_ttft: ttft,
+      median_e2el: e2el,
+      mean_e2el: e2el,
+      p90_e2el: e2el,
+      p95_e2el: e2el,
+      p99_e2el: e2el,
+    },
+    date: dataset.date,
+    run_url: dataset.runUrl ?? null,
+  };
+}
+
+export const SUPPLEMENTAL_BENCHMARK_ROWS: readonly BenchmarkRow[] = DATASETS.flatMap((dataset) =>
+  dataset.points.map((point) => toBenchmarkRow(dataset, point)),
+);
+
+const benchmarkIdentity = (row: BenchmarkRow) =>
+  [
+    row.model,
+    row.hardware,
+    row.framework,
+    row.precision,
+    row.spec_method,
+    row.date,
+    row.isl,
+    row.osl,
+    row.conc,
+    row.prefill_tp,
+    row.prefill_ep,
+    row.decode_tp,
+    row.decode_ep,
+  ].join('|');
+
+function mergeUnique(rows: BenchmarkRow[], additions: readonly BenchmarkRow[]): BenchmarkRow[] {
+  const identities = new Set(rows.map(benchmarkIdentity));
+  const unique = additions.filter((row) => !identities.has(benchmarkIdentity(row)));
+  return unique.length > 0 ? [...rows, ...unique] : rows;
+}
+
+function datasetsForLatestQuery(model: string, date: string, exact: boolean) {
+  const matching = DATASETS.filter((dataset) => dataset.modelAliases.includes(model));
+  if (exact && date) return matching.filter((dataset) => dataset.date === date);
+  const eligible = date ? matching.filter((dataset) => dataset.date <= date) : matching;
+  // Resolve each supplemental curve independently so a newer Jalapeño snapshot
+  // never suppresses the older-but-still-current VR200 curve.
+  const latestByCurve = new Map<string, SupplementalDataset>();
+  for (const dataset of eligible) {
+    const key = `${dataset.hardware}|${dataset.framework}`;
+    const current = latestByCurve.get(key);
+    if (!current || current.date < dataset.date) latestByCurve.set(key, dataset);
+  }
+  return [...latestByCurve.values()];
+}
+
+export function withSupplementalBenchmarks(
+  rows: BenchmarkRow[],
+  query: {
+    model: string;
+    date?: string;
+    exact?: boolean;
+    runId?: string;
+    view?: { type: 'calculator'; sequence: string };
+  },
+): BenchmarkRow[] {
+  if (query.runId || (query.view && query.view.sequence !== '8k/1k')) return rows;
+  const datasets = datasetsForLatestQuery(query.model, query.date ?? '', query.exact ?? false);
+  return mergeUnique(
+    rows,
+    datasets.flatMap((dataset) => dataset.points.map((point) => toBenchmarkRow(dataset, point))),
+  );
+}
+
+export function withSupplementalBenchmarkHistory(
+  rows: BenchmarkRow[],
+  query: { model: string; isl: number; osl: number; benchmarkType?: 'agentic_traces' },
+): BenchmarkRow[] {
+  if (query.benchmarkType || query.isl !== 8192 || query.osl !== 1024) return rows;
+  const datasets = DATASETS.filter((dataset) => dataset.modelAliases.includes(query.model));
+  return mergeUnique(
+    rows,
+    datasets.flatMap((dataset) => dataset.points.map((point) => toBenchmarkRow(dataset, point))),
+  );
+}
+
+const availabilityIdentity = (row: AvailabilityRow) =>
+  [
+    row.model,
+    row.hardware,
+    row.framework,
+    row.precision,
+    row.spec_method,
+    row.isl,
+    row.osl,
+    row.benchmark_type,
+    row.date,
+  ].join('|');
+
+export function withSupplementalAvailability(rows: AvailabilityRow[]): AvailabilityRow[] {
+  const existing = new Set(rows.map(availabilityIdentity));
+  const additions = DATASETS.map<AvailabilityRow>((dataset) => ({
+    model: dataset.model,
+    isl: 8192,
+    osl: 1024,
+    precision: 'fp4',
+    hardware: dataset.hardware,
+    framework: dataset.framework,
+    spec_method: 'none',
+    disagg: false,
+    benchmark_type: 'single_turn',
+    date: dataset.date,
+  })).filter((row) => !existing.has(availabilityIdentity(row)));
+  return additions.length > 0 ? [...rows, ...additions] : rows;
+}
+
+function capabilityFor(hardware: string, framework: string, date: string) {
+  return DATASETS.find(
+    (dataset) =>
+      dataset.hardware === hardware &&
+      dataset.date === date &&
+      (dataset.framework === framework || dataset.sourceFramework === framework),
+  )?.supportedTokenMetrics;
+}
+
+/** Snapshot-scoped capability check for raw rows (official and unofficial). */
+export function supportsTokenMetric(row: BenchmarkRow, tokenType: TokenMetricType): boolean {
+  return capabilityFor(row.hardware, row.framework, row.date)?.includes(tokenType) ?? true;
+}
+
+/** Snapshot-scoped capability check after a row has become chart data. */
+export function supportsChartTokenMetric(
+  hwKey: string,
+  date: string,
+  tokenType: TokenMetricType,
+): boolean {
+  const [hardware, ...frameworkParts] = hwKey.split('_');
+  return capabilityFor(hardware, frameworkParts.join('_'), date)?.includes(tokenType) ?? true;
+}
diff --git a/packages/app/src/lib/tab-meta-zh.test.ts b/packages/app/src/lib/tab-meta-zh.test.ts
index 404871aec..6368ec68c 100644
--- a/packages/app/src/lib/tab-meta-zh.test.ts
+++ b/packages/app/src/lib/tab-meta-zh.test.ts
@@ -1,31 +1,47 @@
 import { describe, expect, it } from 'vitest';
 
 import { SITE_URL } from '@semianalysisai/inferencex-constants';
+import { DASHBOARD_ROUTES } from './dashboard-routes';
 
 import { isValidTab, TAB_META } from './tab-meta';
 import {
   isZhTab,
+  LANDING_META_ZH,
   TAB_INTRO_ZH,
   TAB_LABELS_ZH,
   TAB_META_ZH,
   tabMetadataZh,
-  ZH_TAB_KEYS,
 } from './tab-meta-zh';
 
 const HAN_REGEX = /\p{Script=Han}/u;
 
-describe('ZH_TAB_KEYS', () => {
-  it.each(ZH_TAB_KEYS)('mirrors a valid English tab "%s"', (tab) => {
-    expect(isValidTab(tab)).toBe(true);
-    expect(TAB_META[tab]).toBeDefined();
+describe('Chinese agentic inference positioning', () => {
+  it('uses the category name for titles and AgentX for the scenario', () => {
+    expect(LANDING_META_ZH.title).toContain('智能体推理基准测试');
+    expect(LANDING_META_ZH.title).not.toContain('AgentX');
+    expect(LANDING_META_ZH.description).toMatch(/AgentX.*场景/u);
+    expect(LANDING_META_ZH.description).toContain('固定序列');
+    expect(TAB_META_ZH.inference.title).toContain('智能体推理基准测试');
+    expect(TAB_META_ZH.inference.title).not.toContain('AgentX');
+    expect(TAB_META_ZH.inference.description).toMatch(/AgentX.*工作负载/u);
+    expect(TAB_INTRO_ZH.inference).toContain('固定序列');
+  });
+});
+
+describe('mirrored dashboard routes', () => {
+  const mirroredRoutes = DASHBOARD_ROUTES.filter((route) => route.localeMirrored);
+
+  it.each(mirroredRoutes)('mirrors a valid English tab "$key"', ({ key }) => {
+    expect(isValidTab(key)).toBe(true);
+    expect(TAB_META[key]).toBeDefined();
   });
 
-  it.each(ZH_TAB_KEYS)('has complete Chinese meta, intro, and label for "%s"', (tab) => {
+  it.each(mirroredRoutes)('has complete Chinese content for "$key"', ({ key }) => {
     // Actual Chinese text, not an English placeholder that slipped through.
-    expect(TAB_META_ZH[tab].title).toMatch(HAN_REGEX);
-    expect(TAB_META_ZH[tab].description).toMatch(HAN_REGEX);
-    expect(TAB_INTRO_ZH[tab]).toMatch(HAN_REGEX);
-    expect(TAB_LABELS_ZH[tab]).toMatch(HAN_REGEX);
+    expect(TAB_META_ZH[key].title).toMatch(HAN_REGEX);
+    expect(TAB_META_ZH[key].description).toMatch(HAN_REGEX);
+    expect(TAB_INTRO_ZH[key]).toMatch(HAN_REGEX);
+    expect(TAB_LABELS_ZH[key]).toMatch(HAN_REGEX);
   });
 });
 
diff --git a/packages/app/src/lib/tab-meta-zh.ts b/packages/app/src/lib/tab-meta-zh.ts
index 55d57ab09..e04319ecf 100644
--- a/packages/app/src/lib/tab-meta-zh.ts
+++ b/packages/app/src/lib/tab-meta-zh.ts
@@ -2,39 +2,26 @@ import type { Metadata } from 'next';
 
 import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
 import { ZH_OG_LOCALE, zhAlternates, zhPath } from '@/lib/i18n';
+import type { ModelRoute, ModelRouteTab } from '@/lib/model-routes';
+import {
+  getDashboardRoute,
+  isDashboardRouteKey,
+  type DashboardRouteKey,
+} from '@/lib/dashboard-routes';
 
 export const LANDING_META_ZH = {
-  title: '开源 AI 推理基准测试',
+  title: '开源智能体推理基准测试',
   description:
-    '跨 Chip 与推理框架对比 AI 推理性能。基于 NVIDIA GB200、B200、AMD MI355X 等硬件的真实基准测试。免费、开源、持续更新。',
+    '比较不同芯片和推理框架在 AgentX 与固定序列场景下的推理性能。AgentX 是 InferenceX 面向长上下文、多轮编码的测试场景；NVIDIA 和 AMD 的公开运行结果会随配置变化更新。',
 };
 
-export const ZH_TAB_KEYS = [
-  'inference',
-  'evaluation',
-  'historical',
-  'calculator',
-  'reliability',
-  'gpu-specs',
-  'gpu-metrics',
-  'collectivex',
-  'submissions',
-  'ai-chart',
-  'current-inferencex-image',
-  'feedback',
-] as const;
+export const isZhTab = isDashboardRouteKey;
 
-export type ZhTabKey = (typeof ZH_TAB_KEYS)[number];
-
-export function isZhTab(tab: string): tab is ZhTabKey {
-  return (ZH_TAB_KEYS as readonly string[]).includes(tab);
-}
-
-export const TAB_META_ZH: Record<ZhTabKey, { title: string; description: string }> = {
+export const TAB_META_ZH: Record<DashboardRouteKey, { title: string; description: string }> = {
   inference: {
-    title: 'AI 推理基准测试',
+    title: '智能体推理基准测试',
     description:
-      '跨 Chip 与云服务商对比 AI 推理延迟、吞吐量与首 token 延迟（TTFT）。基于 NVIDIA GB200、H100、AMD MI355X 等硬件的真实基准测试。',
+      '跨芯片与推理框架，对比智能体推理和固定序列 AI 推理的延迟、吞吐量、成本与首 token 延迟（TTFT）。AgentX 提供长上下文多轮编码工作负载。',
   },
   evaluation: {
     title: 'LLM 评估结果',
@@ -43,44 +30,49 @@ export const TAB_META_ZH: Record<ZhTabKey, { title: string; description: string
   historical: {
     title: '历史推理性能趋势',
     description:
-      '跟踪 AI 推理性能随时间的变化。历史基准测试数据展示各 Chip 与服务商在延迟、吞吐量和成本上的改进。',
+      '跟踪 AI 推理性能随时间的变化。历史基准测试数据展示各芯片与服务商在延迟、吞吐量和成本上的改进。',
   },
   calculator: {
     title: '吞吐量与 TCO 计算器',
     description:
-      '计算 AI 推理吞吐量与总拥有成本（TCO）。跨硬件配置对比 LLM 推理服务的 Chip 成本效益。',
+      '计算 AI 推理吞吐量与总拥有成本（TCO）。跨硬件配置对比 LLM 推理服务的芯片成本效益。',
+  },
+  fleet: {
+    title: '集群生命周期经济性',
+    description:
+      '测算固定 AI 推理集群在整个生命周期内的经济性：按设施功率预算确定集群规模，并跟踪实测软件配置随时间改进带来的收入、成本与利润率变化。',
   },
   reliability: {
     title: '服务商可靠性指标',
-    description: 'AI 推理服务商可靠性与可用性跟踪。对比各 Chip 云服务商的错误率与可用性。',
+    description: 'AI 推理服务商可靠性与可用性跟踪。对比各芯片云服务商的错误率与可用性。',
   },
   'gpu-specs': {
-    title: 'Chip 规格与对比',
+    title: '芯片规格与对比',
     description:
-      '面向 AI 推理的详细 Chip 规格。对比 NVIDIA、AMD 与 Intel Chip 的显存带宽、FLOPS、互连与拓扑。',
+      '面向 AI 推理的详细芯片规格。对比 NVIDIA、AMD 与 Intel 芯片的显存带宽、FLOPS、互连与拓扑。',
   },
   'gpu-metrics': {
-    title: 'Chip 功耗与能效指标',
-    description: 'AI 推理负载下的 Chip 功耗与能效指标。跨硬件对比每瓦 token 数。',
+    title: '芯片功耗与能效指标',
+    description: 'AI 推理负载下的芯片功耗与能效指标。跨硬件对比每瓦 token 数。',
   },
   collectivex: {
     title: 'CollectiveX 通信基准测试',
     description:
-      '跨 NVIDIA 与 AMD Chip 平台对比专家并行（EP）通信性能，包括混合专家（MoE）分发（dispatch）、合并（combine）延迟与逻辑载荷速率。',
+      '跨 NVIDIA 与 AMD 芯片平台对比专家并行（EP）通信性能，包括混合专家（MoE）分发（dispatch）、合并（combine）延迟与逻辑载荷速率。',
   },
   submissions: {
     title: '基准测试提交记录',
     description:
-      '提交到 InferenceX 的全部基准测试配置。查看各 Chip 厂商的提交历史、活动趋势与数据点数量。',
+      '提交到 InferenceX 的全部基准测试配置。查看各芯片厂商的提交历史、活动趋势与数据点数量。',
   },
   'ai-chart': {
     title: 'AI 驱动的图表生成',
-    description: '使用自然语言提示生成自定义推理基准测试图表。借助 AI 对比 Chip、成本与性能。',
+    description: '使用自然语言提示生成自定义推理基准测试图表。借助 AI 对比芯片、成本与性能。',
   },
   'current-inferencex-image': {
     title: 'InferenceX 当前镜像',
     description:
-      '各模型、Chip SKU 和配置的当前 InferenceX Docker 镜像标签。对比已部署镜像与最新 vLLM 和 SGLang 发布版本，标记过期标签。',
+      '各模型、芯片 SKU 和配置的当前 InferenceX Docker 镜像标签。对比已部署镜像与最新 vLLM 和 SGLang 发布版本，标记过期标签。',
   },
   feedback: {
     title: '用户反馈',
@@ -93,43 +85,45 @@ export const TAB_META_ZH: Record<ZhTabKey, { title: string; description: string
  * /zh tab page. The charts themselves render in English; this block gives
  * crawlers and readers genuine Chinese content describing the page.
  */
-export const TAB_INTRO_ZH: Record<ZhTabKey, string> = {
+export const TAB_INTRO_ZH: Record<DashboardRouteKey, string> = {
   inference:
-    '本页面展示 InferenceX 的 AI 推理基准测试结果：跨 Chip、推理框架与模型对比吞吐量（token/s/Chip）、交互性（token/s/用户）、首 token 延迟（TTFT）等指标。每个数据点都来自公开的 GitHub Actions 工作流，可复现、可审计。',
+    '本页面展示 InferenceX 的智能体推理与固定序列 AI 推理基准测试结果：跨芯片、推理框架与模型对比吞吐量（token/s/chip）、交互性（token/s/用户）、首 token 延迟（TTFT）等指标。智能体推理数据来自 AgentX；该场景对公开智能体编码轨迹衍生出的长上下文、多轮、含 subagent 工作负载进行回放。每个数据点都来自公开的 GitHub Actions 工作流，可复现、可审计。',
   evaluation:
     '本页面展示 LLM 评估（evaluation）结果：使用标准化评估集对比各模型与部署配置的准确率，验证推理优化不会损害模型质量。',
   historical:
-    '本页面展示历史趋势图表：跟踪各 Chip、框架与模型的推理性能随时间的演进，量化软件栈优化带来的收益。',
+    '本页面展示历史趋势图表：跟踪各芯片、框架与模型的推理性能随时间的演进，量化软件栈优化带来的收益。',
   calculator:
-    '本页面提供吞吐量与总拥有成本（TCO）计算器：基于真实基准测试数据，估算不同 Chip 配置下 LLM 推理服务的每百万 token 成本与性价比。',
+    '本页面提供吞吐量与总拥有成本（TCO）计算器：基于真实基准测试数据，估算不同芯片配置下 LLM 推理服务的每百万 token 成本与性价比。',
+  fleet:
+    '本页面提供集群生命周期经济性分析：按设施功率预算确定固定集群的规模，从模型发布之日起，基于历史基准测试中实测的软件配置改进，测算收入、成本、利润率与回本时间。',
   reliability:
-    '本页面展示基准测试基础设施的可靠性指标：各 Chip 集群与服务商的运行成功率、错误率与可用性。',
+    '本页面展示基准测试基础设施的可靠性指标：各芯片集群与服务商的运行成功率、错误率与可用性。',
   'gpu-specs':
-    '本页面提供 Chip 规格对比：NVIDIA、AMD 等厂商加速器的显存容量、显存带宽、FLOPS、互连拓扑与功耗规格。',
+    '本页面提供芯片规格对比：NVIDIA、AMD 等厂商加速器的显存容量、显存带宽、FLOPS、互连拓扑与功耗规格。',
   'gpu-metrics':
-    '本页面展示 Chip 功耗与能效指标（PowerX）：推理负载下的实测功耗、每瓦 token 数与每兆瓦 token 产出。',
+    '本页面展示芯片功耗与能效指标（PowerX）：推理负载下的实测功耗、每瓦 token 数与每兆瓦 token 产出。',
   collectivex:
     '本页面展示 CollectiveX 专家并行（EP）通信基准测试结果：在统一工作负载、正确性校验与采样协议下，对比 DeepEP、MoRI、UCCL 及 NCCL/RCCL 参考实现的分发（dispatch）、合并（combine）与完整往返延迟。跨芯片速率均按逻辑载荷计算；只有发布器确认完整且稳定的官方队列才会生成排名与推荐。',
   submissions:
-    '本页面列出提交到 InferenceX 的全部基准测试配置：按 Chip 厂商查看提交历史、活动趋势与数据点数量。',
+    '本页面列出提交到 InferenceX 的全部基准测试配置：按芯片厂商查看提交历史、活动趋势与数据点数量。',
   'ai-chart':
     '本页面提供 AI 驱动的图表生成工具：用自然语言描述您想查看的图表，系统会根据 InferenceX 基准测试数据自动生成可视化结果。',
   'current-inferencex-image':
-    '本页面展示 InferenceX 当前使用的 Docker 镜像标签：按模型、Chip SKU 和配置列出已部署版本，并与上游 vLLM、SGLang 最新发布版本对比，方便排查过期镜像。',
+    '本页面展示 InferenceX 当前使用的 Docker 镜像标签：按模型、芯片 SKU 和配置列出已部署版本，并与上游 vLLM、SGLang 最新发布版本对比，方便排查过期镜像。',
   feedback:
     '本页面为内部反馈查看器：使用解密密钥在浏览器中解密并查阅用户提交的反馈内容，密钥不会离开此页面。',
 };
 
 /** Chinese labels for the dashboard tab bar (TabNav) on /zh pages. */
-export const TAB_LABELS_ZH: Record<string, string> = {
-  overview: '总览',
+export const TAB_LABELS_ZH: Record<DashboardRouteKey, string> = {
   inference: '推理性能',
   evaluation: '准确率评估',
   historical: '历史趋势',
   calculator: 'TCO 计算器',
+  fleet: '集群生命周期',
   reliability: '可靠性',
-  'gpu-specs': 'Chip 规格',
-  'gpu-metrics': 'Chip 功耗',
+  'gpu-specs': '芯片规格',
+  'gpu-metrics': '芯片功耗',
   collectivex: 'CollectiveX 通信',
   submissions: '提交记录',
   'ai-chart': 'AI 图表',
@@ -137,22 +131,87 @@ export const TAB_LABELS_ZH: Record<string, string> = {
   feedback: '反馈',
 };
 
-/** Chinese labels for the site header nav on /zh pages, keyed by English href. */
-export const NAV_LABELS_ZH: Record<string, string> = {
+export type HeaderNavHref =
+  | '/'
+  | '/agentx'
+  | '/overview'
+  | '/inference'
+  | '/compare'
+  | '/blog'
+  | '/about';
+
+/** Chinese labels for the site header nav, keyed by its exact English href set. */
+export const NAV_LABELS_ZH: Record<HeaderNavHref, string> = {
   '/': '首页',
   '/overview': '总览',
   '/inference': '仪表板',
-  '/compare': 'Chip 对比',
+  '/compare': '性能对比',
+  '/blog': '技术文章',
+  '/agentx': 'AgentX',
   '/about': '关于',
 };
 
 const TITLE_SUFFIX = `${SITE_NAME} by ${AUTHOR_NAME}`;
 
+/** Chinese copy for the per-model tab routes (/zh/calculator/<slug>,
+ *  /zh/historical/<slug>). Model names stay in English per site convention. */
+export const MODEL_TAB_META_ZH: Record<
+  ModelRouteTab,
+  {
+    title: (seoName: string) => string;
+    description: (seoName: string) => string;
+    intro: (seoName: string) => string;
+  }
+> = {
+  historical: {
+    title: (seoName) => `${seoName} 历史推理性能趋势`,
+    description: (seoName) =>
+      `跟踪 ${seoName} 推理性能随时间的变化。历史基准测试数据展示各芯片与服务商运行 ${seoName} 时在延迟、吞吐量和成本上的改进。`,
+    intro: (seoName) =>
+      `本页面展示 ${seoName} 的历史趋势图表：跟踪各芯片与框架运行 ${seoName} 时推理性能随时间的演进，量化软件栈优化带来的收益。`,
+  },
+  calculator: {
+    title: (seoName) => `${seoName} 吞吐量与 TCO 计算器`,
+    description: (seoName) =>
+      `计算 ${seoName} 推理的吞吐量与总拥有成本（TCO）。跨硬件配置对比 ${seoName} 推理服务的芯片成本效益。`,
+    intro: (seoName) =>
+      `本页面提供 ${seoName} 吞吐量与总拥有成本（TCO）计算器：基于真实基准测试数据，估算不同芯片配置下 ${seoName} 推理服务的每百万 token 成本与性价比。`,
+  },
+};
+
+/** Generate Next.js Metadata for a per-model /zh tab page. Canonical (and
+ *  hreflang) mirror the English rule in `modelTabCanonicalPath`: the default
+ *  model's page points at the bare tab path, others are self-canonical. */
+export function modelTabMetadataZh(
+  tab: ModelRouteTab,
+  route: ModelRoute,
+  canonicalEnPath: string,
+): Metadata {
+  const meta = MODEL_TAB_META_ZH[tab];
+  const title = meta.title(route.seoName);
+  const description = meta.description(route.seoName);
+  const url = `${SITE_URL}${zhPath(canonicalEnPath)}`;
+  return {
+    title,
+    description,
+    alternates: zhAlternates(canonicalEnPath),
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+      locale: ZH_OG_LOCALE,
+    },
+    twitter: {
+      title: `${title} | ${TITLE_SUFFIX}`,
+      description,
+    },
+  };
+}
+
 /** Generate Next.js Metadata for a /zh tab page (mirrors `tabMetadata`). */
-export function tabMetadataZh(tab: ZhTabKey): Metadata {
+export function tabMetadataZh(tab: DashboardRouteKey): Metadata {
   const meta = TAB_META_ZH[tab];
-  // The English inference tab canonicalizes to the site root; mirror that.
-  const enPath = tab === 'inference' ? '/' : `/${tab}`;
+  const enPath = getDashboardRoute(tab).canonicalPath;
   const url = `${SITE_URL}${zhPath(enPath)}`;
   return {
     title: meta.title,
diff --git a/packages/app/src/lib/tab-meta.test.ts b/packages/app/src/lib/tab-meta.test.ts
index 743e0ac3f..a8d3514ff 100644
--- a/packages/app/src/lib/tab-meta.test.ts
+++ b/packages/app/src/lib/tab-meta.test.ts
@@ -1,9 +1,32 @@
 import { describe, expect, it } from 'vitest';
 
-import { getTabTitle, isValidTab, TAB_META, VALID_TABS } from './tab-meta';
+import { DASHBOARD_ROUTE_KEYS } from './dashboard-routes';
+import { getTabTitle, isValidTab, LANDING_META, TAB_META, tabMetadata } from './tab-meta';
+
+describe('agentic inference positioning', () => {
+  it('uses agentic inference for the category and AgentX for the scenario', () => {
+    expect(LANDING_META.title).toContain('Agentic Inference Benchmark');
+    expect(LANDING_META.title).not.toContain('AgentX');
+    expect(LANDING_META.description).toMatch(/AgentX.*scenario/u);
+    expect(LANDING_META.description).toContain('fixed-sequence');
+    expect(TAB_META.inference.title).toContain('Agentic Inference');
+    expect(TAB_META.inference.title).not.toContain('AgentX');
+    expect(TAB_META.inference.description).toMatch(/AgentX.*workload/u);
+    expect(TAB_META.inference.description).toContain('fixed-sequence');
+  });
+});
+
+describe('current image metadata', () => {
+  it('uses the canonical tab copy and bilingual alternates', () => {
+    const meta = tabMetadata('current-inferencex-image');
+    expect(meta.description).toBe(TAB_META['current-inferencex-image'].description);
+    expect(meta.openGraph?.description).toBe(TAB_META['current-inferencex-image'].description);
+    expect(meta.alternates?.languages).toBeDefined();
+  });
+});
 
 describe('isValidTab', () => {
-  it.each(VALID_TABS)('returns true for valid tab "%s"', (tab) => {
+  it.each(DASHBOARD_ROUTE_KEYS)('returns true for valid tab "%s"', (tab) => {
     expect(isValidTab(tab)).toBe(true);
   });
 
@@ -16,7 +39,7 @@ describe('isValidTab', () => {
 });
 
 describe('getTabTitle', () => {
-  it.each(VALID_TABS)('returns formatted title for "%s"', (tab) => {
+  it.each(DASHBOARD_ROUTE_KEYS)('returns formatted title for "%s"', (tab) => {
     const title = getTabTitle(tab);
     expect(title).toContain(TAB_META[tab].title);
     expect(title).toContain('|');
diff --git a/packages/app/src/lib/tab-meta.ts b/packages/app/src/lib/tab-meta.ts
index 0449e1dc5..d959a97d6 100644
--- a/packages/app/src/lib/tab-meta.ts
+++ b/packages/app/src/lib/tab-meta.ts
@@ -1,36 +1,30 @@
 import type { Metadata } from 'next';
 
 import { AUTHOR_NAME, SITE_NAME, SITE_URL } from '@semianalysisai/inferencex-constants';
-import { hasZhSibling, languageAlternates } from '@/lib/i18n';
+import {
+  getDashboardRoute,
+  isDashboardRouteKey,
+  type DashboardRouteKey,
+} from '@/lib/dashboard-routes';
+import { languageAlternates } from '@/lib/i18n';
+import {
+  DEFAULT_ROUTE_MODEL,
+  modelRoutePath,
+  type ModelRoute,
+  type ModelRouteTab,
+} from '@/lib/model-routes';
 
 export const LANDING_META = {
-  title: 'Open Source AI Inference Benchmark',
+  title: 'Open-Source Agentic Inference Benchmark',
   description:
-    'Compare AI inference performance across chips and frameworks. Real benchmarks on NVIDIA GB200, B200, AMD MI355X, and more. Free, open-source, continuously updated.',
+    "Compare AgentX, InferenceX's long-context, multi-turn coding scenario, with fixed-sequence AI inference across chips and frameworks. Public NVIDIA and AMD runs update when configurations change.",
 };
 
-export const VALID_TABS = [
-  'inference',
-  'evaluation',
-  'historical',
-  'calculator',
-  'reliability',
-  'gpu-specs',
-  'collectivex',
-  'ai-chart',
-  'gpu-metrics',
-  'submissions',
-  'current-inferencex-image',
-  'feedback',
-] as const;
-
-export type TabKey = (typeof VALID_TABS)[number];
-
-export const TAB_META: Record<TabKey, { title: string; description: string }> = {
+export const TAB_META: Record<DashboardRouteKey, { title: string; description: string }> = {
   inference: {
-    title: 'AI Inference Benchmarks',
+    title: 'Agentic Inference Benchmarks',
     description:
-      'Compare AI inference latency, throughput, and time-to-first-token across chips and providers. Real benchmarks on NVIDIA GB200, H100, AMD MI355X, and more.',
+      'Compare latency, throughput, cost, and time-to-first-token for agentic and fixed-sequence AI inference across chips and serving frameworks. AgentX supplies the long-context, multi-turn coding workload.',
   },
   evaluation: {
     title: 'LLM Evaluation Results',
@@ -47,6 +41,11 @@ export const TAB_META: Record<TabKey, { title: string; description: string }> =
     description:
       'Calculate AI inference throughput and total cost of ownership. Compare chip cost-efficiency for LLM serving across hardware configurations.',
   },
+  fleet: {
+    title: 'Fleet Lifecycle Economics',
+    description:
+      'Project a fixed AI inference fleet across its life: size it against a facility power budget, then track revenue, cost, and margin as measured software configs improve over time.',
+  },
   reliability: {
     title: 'Provider Reliability Metrics',
     description:
@@ -90,27 +89,83 @@ export const TAB_META: Record<TabKey, { title: string; description: string }> =
 
 const TITLE_SUFFIX = `${SITE_NAME} by ${AUTHOR_NAME}`;
 
-export function isValidTab(value: string): value is TabKey {
-  return (VALID_TABS as readonly string[]).includes(value);
-}
+export const isValidTab = isDashboardRouteKey;
 
 export function getTabTitle(tab: string): string {
-  const meta = TAB_META[tab as TabKey];
+  const meta = isDashboardRouteKey(tab) ? TAB_META[tab] : undefined;
   return meta ? `${meta.title} | ${TITLE_SUFFIX}` : TITLE_SUFFIX;
 }
 
+/** Model-specific copy for the per-model tab routes (/calculator/<slug>,
+ *  /historical/<slug>). Same shape as TAB_META but parameterized on the
+ *  model's SEO name. */
+export const MODEL_TAB_META: Record<
+  ModelRouteTab,
+  { title: (seoName: string) => string; description: (seoName: string) => string }
+> = {
+  historical: {
+    title: (seoName) => `${seoName} Historical Inference Trends`,
+    description: (seoName) =>
+      `Track ${seoName} inference performance over time. Historical benchmark data showing chip and provider improvements in latency, throughput, and cost for ${seoName}.`,
+  },
+  calculator: {
+    title: (seoName) => `${seoName} Throughput & TCO Calculator`,
+    description: (seoName) =>
+      `Calculate ${seoName} inference throughput and total cost of ownership. Compare chip cost-efficiency for serving ${seoName} across hardware configurations.`,
+  },
+};
+
+/**
+ * English path a per-model tab page canonicalizes to. The default model's
+ * page shows exactly what the bare tab route shows, so it canonicalizes to
+ * the bare path instead of competing with it; every other model is
+ * self-canonical.
+ */
+export function modelTabCanonicalPath(tab: ModelRouteTab, route: ModelRoute): string {
+  return route.model === DEFAULT_ROUTE_MODEL
+    ? getDashboardRoute(tab).canonicalPath
+    : modelRoutePath(tab, route.slug);
+}
+
+/** Generate Next.js Metadata for a per-model tab page. */
+export function modelTabMetadata(tab: ModelRouteTab, route: ModelRoute): Metadata {
+  const meta = MODEL_TAB_META[tab];
+  const title = meta.title(route.seoName);
+  const description = meta.description(route.seoName);
+  const enPath = modelTabCanonicalPath(tab, route);
+  const url = `${SITE_URL}${enPath}`;
+  return {
+    title,
+    description,
+    alternates: {
+      canonical: url,
+      languages: languageAlternates(enPath),
+    },
+    openGraph: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+      url,
+    },
+    twitter: {
+      title: `${title} | ${SITE_NAME}`,
+      description,
+    },
+  };
+}
+
 /** Generate Next.js Metadata for a tab page. */
-export function tabMetadata(tab: TabKey): Metadata {
+export function tabMetadata(tab: DashboardRouteKey): Metadata {
   const meta = TAB_META[tab];
-  const enPath = tab === 'inference' ? '/' : `/${tab}`;
-  const url = tab === 'inference' ? SITE_URL : `${SITE_URL}/${tab}`;
+  const route = getDashboardRoute(tab);
+  const enPath = route.canonicalPath;
+  const url = enPath === '/' ? SITE_URL : `${SITE_URL}${enPath}`;
   return {
     title: meta.title,
     description: meta.description,
     alternates: {
       canonical: url,
       // hreflang to the Chinese sibling page, for tabs mirrored under /zh.
-      ...(hasZhSibling(enPath) && { languages: languageAlternates(enPath) }),
+      ...(route.localeMirrored && { languages: languageAlternates(enPath) }),
     },
     openGraph: {
       title: `${meta.title} | InferenceX`,
diff --git a/packages/app/src/lib/test-fixture-manifest.ts b/packages/app/src/lib/test-fixture-manifest.ts
new file mode 100644
index 000000000..cb834227c
--- /dev/null
+++ b/packages/app/src/lib/test-fixture-manifest.ts
@@ -0,0 +1,68 @@
+import { createHash } from 'node:crypto';
+
+export const FIXTURE_MANIFEST_FILENAME = '_manifest.json';
+export const FIXTURE_MANIFEST_SCHEMA_VERSION = 1;
+
+export type FixtureTopLevel = 'array' | 'object';
+
+export interface FixtureManifestEntry {
+  bytes: number;
+  capturedAt: string | null;
+  sha256: string;
+  source: string | null;
+  topLevel: FixtureTopLevel;
+}
+
+export interface FixtureManifest {
+  schemaVersion: typeof FIXTURE_MANIFEST_SCHEMA_VERSION;
+  generatedAt: string;
+  fixtures: Record<string, FixtureManifestEntry>;
+}
+
+export function fixtureTopLevel(value: unknown): FixtureTopLevel {
+  if (Array.isArray(value)) return 'array';
+  if (value !== null && typeof value === 'object') return 'object';
+  throw new Error('Cypress API fixtures must contain a top-level array or object');
+}
+
+/** Fixture-specific invariants required for substantive E2E coverage. */
+export function assertFixtureContent(name: string, value: unknown): void {
+  fixtureTopLevel(value);
+  if (name !== 'benchmarks-history') return;
+  if (!Array.isArray(value)) {
+    throw new TypeError('Fixture benchmarks-history must contain an array');
+  }
+  const dates = new Set(
+    value.flatMap((row) => {
+      if (row === null || typeof row !== 'object' || !('date' in row)) return [];
+      const date = row.date;
+      return typeof date === 'string' ? [date] : [];
+    }),
+  );
+  if (dates.size < 2) {
+    throw new Error('Fixture benchmarks-history must contain at least two replayable dates');
+  }
+}
+
+export function fixtureSha256(body: string): string {
+  return createHash('sha256').update(body).digest('hex');
+}
+
+export function assertFixtureMatchesManifest(
+  name: string,
+  body: string,
+  value: unknown,
+  entry: FixtureManifestEntry | undefined,
+): void {
+  assertFixtureContent(name, value);
+  if (!entry) throw new Error(`Fixture manifest has no entry for ${name}`);
+  if (Buffer.byteLength(body) !== entry.bytes) {
+    throw new Error(`Fixture ${name} byte length differs from its manifest`);
+  }
+  if (fixtureSha256(body) !== entry.sha256) {
+    throw new Error(`Fixture ${name} checksum differs from its manifest`);
+  }
+  if (fixtureTopLevel(value) !== entry.topLevel) {
+    throw new Error(`Fixture ${name} top-level shape differs from its manifest`);
+  }
+}
diff --git a/packages/app/src/lib/test-fixtures.test.ts b/packages/app/src/lib/test-fixtures.test.ts
new file mode 100644
index 000000000..cb093e61f
--- /dev/null
+++ b/packages/app/src/lib/test-fixtures.test.ts
@@ -0,0 +1,51 @@
+import { readdirSync, readFileSync } from 'node:fs';
+import { resolve } from 'node:path';
+
+import { describe, expect, it } from 'vitest';
+
+import {
+  FIXTURE_MANIFEST_FILENAME,
+  FIXTURE_MANIFEST_SCHEMA_VERSION,
+  type FixtureManifest,
+  assertFixtureContent,
+  assertFixtureMatchesManifest,
+} from '@/lib/test-fixture-manifest';
+
+const fixturesDir = resolve(import.meta.dirname, '../../cypress/fixtures/api');
+const manifest = JSON.parse(
+  readFileSync(resolve(fixturesDir, FIXTURE_MANIFEST_FILENAME), 'utf8'),
+) as FixtureManifest;
+
+describe('Cypress API fixture manifest', () => {
+  it('uses the current schema and records a valid generation timestamp', () => {
+    expect(manifest.schemaVersion).toBe(FIXTURE_MANIFEST_SCHEMA_VERSION);
+    expect(Number.isNaN(Date.parse(manifest.generatedAt))).toBe(false);
+  });
+
+  it('covers every fixture with matching shape, size, and checksum', () => {
+    const fixtureFiles = readdirSync(fixturesDir)
+      .filter((name) => name.endsWith('.json') && name !== FIXTURE_MANIFEST_FILENAME)
+      .toSorted();
+    const fixtureNames = fixtureFiles.map((name) => name.slice(0, -'.json'.length)).toSorted();
+    expect(Object.keys(manifest.fixtures).toSorted()).toEqual(fixtureNames);
+
+    for (const filename of fixtureFiles) {
+      const name = filename.slice(0, -'.json'.length);
+      const body = readFileSync(resolve(fixturesDir, filename), 'utf8');
+      const value = JSON.parse(body) as unknown;
+      expect(() =>
+        assertFixtureMatchesManifest(name, body, value, manifest.fixtures[name]),
+      ).not.toThrow();
+    }
+  });
+
+  it('rejects benchmark history that cannot exercise replay', () => {
+    expect(() => assertFixtureContent('benchmarks-history', [])).toThrow(/two replayable dates/u);
+    expect(() => assertFixtureContent('benchmarks-history', [{ date: '2026-08-20' }])).toThrow(
+      /two replayable dates/u,
+    );
+    expect(() =>
+      assertFixtureContent('benchmarks-history', [{ date: '2026-08-20' }, { date: '2026-08-21' }]),
+    ).not.toThrow();
+  });
+});
diff --git a/packages/app/src/lib/test-fixtures.ts b/packages/app/src/lib/test-fixtures.ts
index 7f7947fdd..6432d5605 100644
--- a/packages/app/src/lib/test-fixtures.ts
+++ b/packages/app/src/lib/test-fixtures.ts
@@ -12,15 +12,32 @@
 
 import { readFileSync } from 'node:fs';
 import { resolve } from 'node:path';
+import {
+  FIXTURE_MANIFEST_FILENAME,
+  FIXTURE_MANIFEST_SCHEMA_VERSION,
+  type FixtureManifest,
+  assertFixtureMatchesManifest,
+} from '@/lib/test-fixture-manifest';
 
 const cache = new Map<string, unknown>();
+let manifestCache: FixtureManifest | undefined;
 
 export function loadFixture<T>(name: string): T {
   const hit = cache.get(name);
   if (hit !== undefined) return hit as T;
   // process.cwd() is packages/app/ when Next.js runs (`bun run dev` or `bun run start`).
-  const path = resolve(process.cwd(), 'cypress', 'fixtures', 'api', `${name}.json`);
-  const data = JSON.parse(readFileSync(path, 'utf8')) as T;
+  const fixturesDir = resolve(process.cwd(), 'cypress', 'fixtures', 'api');
+  if (!manifestCache) {
+    manifestCache = JSON.parse(
+      readFileSync(resolve(fixturesDir, FIXTURE_MANIFEST_FILENAME), 'utf8'),
+    ) as FixtureManifest;
+    if (manifestCache.schemaVersion !== FIXTURE_MANIFEST_SCHEMA_VERSION) {
+      throw new Error(`Unsupported fixture manifest schema ${manifestCache.schemaVersion}`);
+    }
+  }
+  const body = readFileSync(resolve(fixturesDir, `${name}.json`), 'utf8');
+  const data = JSON.parse(body) as T;
+  assertFixtureMatchesManifest(name, body, data, manifestCache.fixtures[name]);
   cache.set(name, data);
   return data;
 }
diff --git a/packages/app/src/lib/toggle-set.ts b/packages/app/src/lib/toggle-set.ts
index 963d3ea4e..ae6fe5dbc 100644
--- a/packages/app/src/lib/toggle-set.ts
+++ b/packages/app/src/lib/toggle-set.ts
@@ -1,32 +1,27 @@
 /**
  * Pure toggle state transition.
  *
- * - If all items are active and one is toggled, only that item remains active (solo).
- * - If only one item is active and it's toggled, all items become active (restore all).
- * - Otherwise, the item is simply added or removed from the active set.
+ * - Clicking an active item while others are active solos it.
+ * - Clicking the only active item restores all.
+ * - Clicking an inactive item adds it.
+ *
+ * Solo used to require EVERY item to be active, which quietly made it
+ * unreachable wherever something starts deselected. On the agentic view the
+ * engine guard switches SGLang off on load, so a fresh chart never had all
+ * items active and the first click on a legend entry removed it instead of
+ * isolating it. Individual removal has its own control — the X that appears on
+ * a legend row on hover — so the label click is free to always mean "show only
+ * this".
  *
  * Kept free of React so server-only modules (`exclusion.ts`, and through it the
  * overview data layer) can reuse it without pulling client hooks into a Server
  * Component graph.
  */
 export function computeToggle(prev: Set<string>, item: string, allItems: Set<string>): Set<string> {
-  const allAreActive = prev.size === allItems.size;
-  const isActive = prev.has(item);
-
-  if (isActive) {
-    if (allAreActive) {
-      // Solo: only keep the clicked item
-      return new Set([item]);
-    } else if (prev.size === 1) {
-      // Restore all: re-enable everything
-      return allItems;
-    }
-    // Remove the clicked item
-    const next = new Set(prev);
-    next.delete(item);
-    return next;
+  if (prev.has(item)) {
+    // Sole survivor -> restore everything, so a second click undoes the solo.
+    if (prev.size <= 1) return allItems;
+    return new Set([item]);
   }
-  // Add the clicked item
-  const next = new Set([...prev, item]);
-  return next;
+  return new Set([...prev, item]);
 }
diff --git a/packages/app/src/lib/typography-gate.test.ts b/packages/app/src/lib/typography-gate.test.ts
new file mode 100644
index 000000000..56237a5da
--- /dev/null
+++ b/packages/app/src/lib/typography-gate.test.ts
@@ -0,0 +1,99 @@
+import { describe, expect, it } from 'vitest';
+
+import { compareToAllowlist, countViolations, isExemptPath, scanSource } from './typography-gate';
+
+describe('scanSource', () => {
+  it('flags arbitrary font sizes with a token suggestion', () => {
+    const source = '<span className="font-mono text-[11px] text-muted-foreground" />';
+    const violations = scanSource('src/components/foo.tsx', source);
+    expect(violations).toHaveLength(1);
+    expect(violations[0]).toMatchObject({
+      rule: 'arbitrary-text-size',
+      line: 1,
+      match: 'text-[11px]',
+    });
+    expect(violations[0].suggestion).toContain('text-2xs');
+  });
+
+  it('counts multiple matches on one line and reports line numbers', () => {
+    const source = ['const a = 1;', '"text-[10px] text-[0.65rem]"'].join('\n');
+    const violations = scanSource('src/components/foo.tsx', source);
+    expect(violations.map((v) => v.line)).toEqual([2, 2]);
+    expect(violations.map((v) => v.match)).toEqual(['text-[10px]', 'text-[0.65rem]']);
+  });
+
+  it('ignores arbitrary text colors and slash line-height modifiers on scale classes', () => {
+    const source = '"text-[#ed1c24] text-2xl/[1.8rem] text-2xs text-3xs"';
+    expect(scanSource('src/components/foo.tsx', source)).toHaveLength(0);
+  });
+
+  it('flags arbitrary tracking but not the named tokens', () => {
+    const source = '"tracking-[0.16em] tracking-eyebrow tracking-heading tracking-widest"';
+    const violations = scanSource('src/components/foo.tsx', source);
+    expect(violations).toHaveLength(1);
+    expect(violations[0].match).toBe('tracking-[0.16em]');
+    expect(violations[0].suggestion).toContain('tracking-eyebrow');
+  });
+
+  it('flags quoted chart font-size literals only inside src/lib/d3-chart', () => {
+    const literal = ".attr('font-size', '12px')";
+    expect(scanSource('src/lib/d3-chart/layers/foo.ts', literal)).toHaveLength(1);
+    expect(scanSource('src/components/foo.ts', literal)).toHaveLength(0);
+  });
+
+  it('lets computed and CHART_TYPE-based chart font sizes through', () => {
+    const source = [
+      ".attr('font-size', px(CHART_TYPE.axisLabel))",
+      `.attr('font-size', \`\${fontSize}px\`)`,
+      ".style('font-size', sizeVar)",
+    ].join('\n');
+    expect(scanSource('src/lib/d3-chart/layers/foo.ts', source)).toHaveLength(0);
+  });
+
+  it('exempts OG renderers, tests, and the gate itself', () => {
+    expect(isExemptPath('src/app/model/[slug]/opengraph-image.tsx')).toBe(true);
+    expect(isExemptPath('src/app/blog/[slug]/og-image-render.tsx')).toBe(true);
+    expect(isExemptPath('src/lib/compare-og.tsx')).toBe(true);
+    expect(isExemptPath('src/lib/typography-gate.ts')).toBe(true);
+    expect(isExemptPath('src/components/foo.test.tsx')).toBe(true);
+    expect(isExemptPath('src/components/foo.tsx')).toBe(false);
+    expect(scanSource('src/lib/compare-og.tsx', '"text-[11px]"')).toHaveLength(0);
+  });
+});
+
+const files = (source: string) => [{ relPath: 'src/components/foo.tsx', source }];
+
+describe('compareToAllowlist', () => {
+  it('passes when counts match the allowlist exactly', () => {
+    const counts = countViolations(files('"text-[11px]"'));
+    const result = compareToAllowlist(counts, {
+      'src/components/foo.tsx': { 'arbitrary-text-size': 1 },
+    });
+    expect(result.errors).toHaveLength(0);
+    expect(result.stale).toHaveLength(0);
+  });
+
+  it('fails when a file grows past its allowance', () => {
+    const counts = countViolations(files('"text-[11px] text-[10px]"'));
+    const result = compareToAllowlist(counts, {
+      'src/components/foo.tsx': { 'arbitrary-text-size': 1 },
+    });
+    expect(result.errors).toHaveLength(1);
+    expect(result.errors[0]).toContain('src/components/foo.tsx');
+  });
+
+  it('fails on offenders in files the allowlist does not know', () => {
+    const counts = countViolations(files('"tracking-[0.5em]"'));
+    const result = compareToAllowlist(counts, {});
+    expect(result.errors).toHaveLength(1);
+  });
+
+  it('reports stale entries when violations were fixed, so the ratchet only shrinks', () => {
+    const result = compareToAllowlist(
+      {},
+      { 'src/components/foo.tsx': { 'arbitrary-text-size': 1 } },
+    );
+    expect(result.errors).toHaveLength(0);
+    expect(result.stale).toHaveLength(1);
+  });
+});
diff --git a/packages/app/src/lib/typography-gate.ts b/packages/app/src/lib/typography-gate.ts
new file mode 100644
index 000000000..ca6a7623e
--- /dev/null
+++ b/packages/app/src/lib/typography-gate.ts
@@ -0,0 +1,151 @@
+/**
+ * Typography ratchet gate — core matching and allowlist logic.
+ *
+ * The repo's typography is governed by a small token set (see
+ * docs/typography.md): arbitrary font sizes / letter-spacings in class
+ * strings and quoted font-size literals in the shared D3 chart library are
+ * blocked for NEW code, while the offenders that existed when the gate landed
+ * are frozen in scripts/typography-allowlist.json and burn down as their
+ * files get touched.
+ *
+ * This module is pure (no fs) so the rules are unit-testable; the CLI entry
+ * point is scripts/check-typography.ts.
+ */
+
+export interface TypographyViolation {
+  rule: string;
+  /** 1-based line number. */
+  line: number;
+  /** The offending token, e.g. "text-[11px]". */
+  match: string;
+  suggestion: string;
+}
+
+interface Rule {
+  name: string;
+  /** Must be a global regex; matched per line. */
+  pattern: RegExp;
+  appliesTo: (relPath: string) => boolean;
+  suggest: (match: string) => string;
+}
+
+const SIZE_SUGGESTIONS: Record<string, string> = {
+  'text-[11px]': 'use text-2xs (11px)',
+  'text-[0.6875rem]': 'use text-2xs (11px)',
+  'text-[10px]': 'use text-3xs (10px)',
+  'text-[0.625rem]': 'use text-3xs (10px)',
+  'text-[0.65rem]': 'use text-3xs (10px)',
+};
+
+const TRACKING_SUGGESTION =
+  'use tracking-eyebrow (0.16em), tracking-eyebrow-wide (0.2em), tracking-heading (-0.04em), or a built-in tracking-* class';
+
+/**
+ * Files the rules never apply to. OG-image renderers draw on a fixed-size
+ * Satori canvas where hardcoded display sizes are correct; this module and
+ * its CLI contain the offending strings by definition.
+ */
+const EXEMPT_PATH =
+  /(?:opengraph-image|og-image-render|compare-og|typography-gate|check-typography)\.tsx?$|\.test\.tsx?$/u;
+
+const RULES: Rule[] = [
+  {
+    name: 'arbitrary-text-size',
+    // Leading digit/dot keeps hex colors (text-[#…]) and var() refs out of scope.
+    pattern: /\btext-\[[\d.][^\]]*\]/gu,
+    appliesTo: () => true,
+    suggest: (match) =>
+      SIZE_SUGGESTIONS[match] ??
+      'use text-2xs (11px) / text-3xs (10px) or the standard text-* scale',
+  },
+  {
+    name: 'arbitrary-tracking',
+    pattern: /\btracking-\[[^\]]+\]/gu,
+    appliesTo: () => true,
+    suggest: () => TRACKING_SUGGESTION,
+  },
+  {
+    name: 'chart-font-literal',
+    // Quoted px literals only: computed values and CHART_TYPE identifiers pass.
+    pattern: /\.(?:attr|style)\(\s*['"]font-size['"],\s*['"][^'"]*['"]/gu,
+    appliesTo: (relPath) => relPath.startsWith('src/lib/d3-chart/'),
+    suggest: () => "size chart text via CHART_TYPE from '@/lib/d3-chart/typography'",
+  },
+];
+
+export const RULE_NAMES = RULES.map((rule) => rule.name);
+
+export function isExemptPath(relPath: string): boolean {
+  return EXEMPT_PATH.test(relPath);
+}
+
+/** Scan one file's source; relPath is POSIX-style, relative to packages/app. */
+export function scanSource(relPath: string, source: string): TypographyViolation[] {
+  if (isExemptPath(relPath)) return [];
+  const violations: TypographyViolation[] = [];
+  const lines = source.split('\n');
+  for (const rule of RULES) {
+    if (!rule.appliesTo(relPath)) continue;
+    for (const [index, line] of lines.entries()) {
+      for (const match of line.matchAll(rule.pattern)) {
+        violations.push({
+          rule: rule.name,
+          line: index + 1,
+          match: match[0],
+          suggestion: rule.suggest(match[0]),
+        });
+      }
+    }
+  }
+  return violations;
+}
+
+/** file → rule → count. Only files with at least one violation appear. */
+export type ViolationCounts = Record<string, Record<string, number>>;
+
+export function countViolations(
+  files: Iterable<{ relPath: string; source: string }>,
+): ViolationCounts {
+  const counts: ViolationCounts = {};
+  for (const { relPath, source } of files) {
+    for (const violation of scanSource(relPath, source)) {
+      counts[relPath] ??= {};
+      counts[relPath][violation.rule] = (counts[relPath][violation.rule] ?? 0) + 1;
+    }
+  }
+  return counts;
+}
+
+export interface GateResult {
+  /** New or grown offenders — always a failure. */
+  errors: string[];
+  /** Counts that dropped below the allowlist — fix by re-running with --update. */
+  stale: string[];
+}
+
+/**
+ * Ratchet comparison: fail on any count above the allowlist (or any offender
+ * in a file the allowlist doesn't know), and report counts that dropped so
+ * the allowlist can only ever shrink.
+ */
+export function compareToAllowlist(
+  counts: ViolationCounts,
+  allowlist: ViolationCounts,
+): GateResult {
+  const errors: string[] = [];
+  const stale: string[] = [];
+
+  const files = new Set([...Object.keys(counts), ...Object.keys(allowlist)]);
+  for (const file of [...files].sort()) {
+    for (const rule of RULE_NAMES) {
+      const actual = counts[file]?.[rule] ?? 0;
+      const allowed = allowlist[file]?.[rule] ?? 0;
+      if (actual > allowed) {
+        errors.push(`${file}: ${actual} ${rule} violation(s), allowlist permits ${allowed}`);
+      } else if (actual < allowed) {
+        stale.push(`${file}: ${rule} dropped to ${actual} (allowlist says ${allowed})`);
+      }
+    }
+  }
+  return { errors, stale };
+}
diff --git a/packages/app/src/lib/url-state.test.ts b/packages/app/src/lib/url-state.test.ts
index 28374f1ee..c10c25c12 100644
--- a/packages/app/src/lib/url-state.test.ts
+++ b/packages/app/src/lib/url-state.test.ts
@@ -1,13 +1,14 @@
 import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
 
-// helper to set up window mocks before module import
-function setupWindow(search = '', pathname = '/inference') {
+// Dynamic imports are intentional: each test resets the module-scope URL snapshot after stubbing window.
+function setupWindow(search = '', pathname = '/inference', hash = '') {
   const location = {
     search,
     pathname,
+    hash,
     origin: 'https://example.com',
   };
-  const history = { replaceState: vi.fn() };
+  const history = { replaceState: vi.fn(), state: { __NA: 1 } };
 
   vi.stubGlobal('window', { location, history });
   return { location, history };
@@ -40,6 +41,17 @@ describe('PARAM_DEFAULTS', () => {
     expect(PARAM_DEFAULTS.i_seq).toBe('');
   });
 
+  it('strips i_metric against the same default the dashboard opens on', async () => {
+    // A share link omits any value equal to PARAM_DEFAULTS. If this drifted
+    // from DEFAULT_Y_AXIS_METRIC, a link captured on the *other* metric would
+    // be written without `i_metric` and reopen on the dashboard default.
+    const { PARAM_DEFAULTS, DEFAULT_Y_AXIS_METRIC } = await import('@/lib/url-state');
+    const { DEFAULT_METRIC_CONFIG_KEY } = await import('@/components/inference/metric-registry');
+    expect(PARAM_DEFAULTS.i_metric).toBe(DEFAULT_Y_AXIS_METRIC);
+    expect(DEFAULT_Y_AXIS_METRIC).toBe('y_tokensPerDollarN');
+    expect(DEFAULT_Y_AXIS_METRIC).toBe(DEFAULT_METRIC_CONFIG_KEY);
+  });
+
   it('has expected default for r_range', async () => {
     const { PARAM_DEFAULTS } = await import('@/lib/url-state');
     expect(PARAM_DEFAULTS.r_range).toBe('last-3-months');
@@ -62,6 +74,11 @@ describe('PARAM_DEFAULTS', () => {
     expect(PARAM_DEFAULTS.i_advlabel).toBe('');
   });
 
+  it('strips the normalized revenue source but preserves OpenRouter as explicit state', async () => {
+    const { PARAM_DEFAULTS } = await import('@/lib/url-state');
+    expect(PARAM_DEFAULTS.i_revenue).toBe('normalized');
+  });
+
   it('has empty string defaults for legend-active params', async () => {
     const { PARAM_DEFAULTS } = await import('@/lib/url-state');
     expect(PARAM_DEFAULTS.i_active).toBe('');
@@ -70,9 +87,13 @@ describe('PARAM_DEFAULTS', () => {
     expect(PARAM_DEFAULTS.r_active).toBe('');
   });
 
-  it('has empty string defaults for calculator fleet-planner params', async () => {
-    const { PARAM_DEFAULTS } = await import('@/lib/url-state');
-    expect(PARAM_DEFAULTS.c_mw).toBe('');
+  it('keeps the fleet defaults aligned with URL-state stripping', async () => {
+    const { DEFAULT_FLEET_MW, DEFAULT_LIFECYCLE_RAMP_MONTHS, PARAM_DEFAULTS } =
+      await import('@/lib/url-state');
+    expect(DEFAULT_FLEET_MW).toBe('10');
+    expect(DEFAULT_LIFECYCLE_RAMP_MONTHS).toBe('0.5');
+    expect(PARAM_DEFAULTS.c_mw).toBe(DEFAULT_FLEET_MW);
+    expect(PARAM_DEFAULTS.c_ramp).toBe(DEFAULT_LIFECYCLE_RAMP_MONTHS);
     expect(PARAM_DEFAULTS.c_costcap).toBe('');
   });
 });
@@ -104,6 +125,12 @@ describe('readUrlParams', () => {
     expect(params.i_advlabel).toBe('1');
   });
 
+  it('reads the token-revenue price source from the URL', async () => {
+    setupWindow('?i_revenue=openrouter');
+    const { readUrlParams } = await import('@/lib/url-state');
+    expect(readUrlParams().i_revenue).toBe('openrouter');
+  });
+
   it('returns empty object when no URL params exist', async () => {
     setupWindow('');
     const { readUrlParams } = await import('@/lib/url-state');
@@ -117,6 +144,106 @@ describe('readUrlParams', () => {
     expect(params.g_model).toBe('test');
     expect(params).not.toHaveProperty('unknown_key');
   });
+
+  it('preserves the locale path, unrelated params, and hash when cleaning initial state', async () => {
+    const { history } = setupWindow('?g_model=test&eval=42', '/zh/evaluation', '#sample-detail');
+    await import('@/lib/url-state');
+    await vi.runAllTimersAsync();
+
+    expect(history.replaceState).toHaveBeenCalledWith(
+      null,
+      '',
+      '/zh/evaluation?eval=42#sample-detail',
+    );
+  });
+});
+
+describe('refreshUrlParams', () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+    vi.resetModules();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+  });
+
+  // A client-side navigation does not reload the module, so the load-time
+  // snapshot still describes the page the user came FROM. Every AgentX card on
+  // the landing page linked to `/inference?g_model=<model>` and opened the
+  // default model instead, because the snapshot from `/` had no g_model.
+  it('picks up params that arrived via a client-side navigation', async () => {
+    const { location } = setupWindow('', '/');
+    const { readUrlParams, refreshUrlParams } = await import('@/lib/url-state');
+    expect(readUrlParams().g_model).toBeUndefined();
+
+    location.search = '?g_model=Qwen-3.5-397B-A17B&i_seq=agentic-traces';
+    location.pathname = '/inference';
+    const refreshed = refreshUrlParams();
+
+    expect(refreshed.g_model).toBe('Qwen-3.5-397B-A17B');
+    expect(refreshed.i_seq).toBe('agentic-traces');
+    // Same object the load-time readers already hold, so they see it too.
+    expect(readUrlParams().g_model).toBe('Qwen-3.5-397B-A17B');
+  });
+
+  it('keeps existing params when the new URL omits them', async () => {
+    const { location } = setupWindow('?g_model=Kimi-K3&i_seq=agentic-traces', '/inference');
+    const { refreshUrlParams } = await import('@/lib/url-state');
+
+    location.search = '?i_seq=8k/1k';
+    const refreshed = refreshUrlParams();
+
+    expect(refreshed.i_seq).toBe('8k/1k');
+    // Not cleared: the provider writes params back as the user filters, and a
+    // partial URL must not wipe state the user did not touch.
+    expect(refreshed.g_model).toBe('Kimi-K3');
+  });
+
+  it('flushes a pending filter before a retained provider refreshes on navigation', async () => {
+    const { location } = setupWindow('?g_model=Kimi-K3', '/inference');
+    const { readUrlParams, refreshUrlParams, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Qwen-3.5-397B-A17B' });
+    location.search = '';
+    location.pathname = '/evaluation';
+    refreshUrlParams();
+
+    expect(readUrlParams().g_model).toBe('Qwen-3.5-397B-A17B');
+  });
+
+  it('tracks explicit navigation intent separately from serialized state', async () => {
+    const { location } = setupWindow('', '/inference');
+    const { hasExplicitUrlParam, refreshUrlParams, writeUrlParams } =
+      await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3' });
+    expect(hasExplicitUrlParam('g_model')).toBe(false);
+
+    location.search = '?g_model=GLM-5.2';
+    refreshUrlParams();
+    expect(hasExplicitUrlParam('g_model')).toBe(true);
+
+    location.search = '';
+    refreshUrlParams();
+    expect(hasExplicitUrlParam('g_model')).toBe(false);
+  });
+
+  it('ignores unknown params', async () => {
+    const { location } = setupWindow('', '/');
+    const { refreshUrlParams } = await import('@/lib/url-state');
+    location.search = '?g_model=GLM-5.2&not_a_real_key=x';
+    const refreshed = refreshUrlParams();
+    expect(refreshed.g_model).toBe('GLM-5.2');
+    expect(refreshed).not.toHaveProperty('not_a_real_key');
+  });
+
+  it('is a no-op without a window (SSR)', async () => {
+    vi.stubGlobal('window', undefined);
+    const { refreshUrlParams } = await import('@/lib/url-state');
+    expect(() => refreshUrlParams()).not.toThrow();
+  });
 });
 
 describe('hasAnyUrlParams', () => {
@@ -183,6 +310,18 @@ describe('writeUrlParams + buildShareUrl', () => {
     expect(url).not.toContain('g_model');
   });
 
+  it('removes the revenue source instead of emitting an empty query param off-metric', async () => {
+    setupWindow('?i_revenue=openrouter', '/inference');
+    const { writeUrlParams, buildShareUrl } = await import('@/lib/url-state');
+
+    // InferenceContext writes the default source when token revenue is not the
+    // active metric, which must clear stale state without serializing i_revenue=.
+    writeUrlParams({ i_revenue: 'normalized' });
+    await vi.advanceTimersByTimeAsync(200);
+
+    expect(buildShareUrl()).not.toContain('i_revenue');
+  });
+
   it('removes params with undefined value', async () => {
     setupWindow('?g_model=custom', '/inference');
     const { writeUrlParams, buildShareUrl } = await import('@/lib/url-state');
@@ -239,6 +378,27 @@ describe('writeUrlParams + buildShareUrl', () => {
     const url = buildShareUrl();
     expect(url).toContain('g_model=immediate');
   });
+
+  it('flushes pending writes synchronously for a remounted provider', async () => {
+    setupWindow('', '/inference');
+    const { readUrlParams, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3', i_seq: 'agentic-traces' });
+
+    expect(readUrlParams()).toMatchObject({
+      g_model: 'Kimi-K3',
+      i_seq: 'agentic-traces',
+    });
+  });
+
+  it('does not restore an initial param after the user resets it to default', async () => {
+    setupWindow('?g_model=Kimi-K3', '/inference');
+    const { readUrlParams, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'DeepSeek-V4-Pro' });
+
+    expect(readUrlParams().g_model).toBeUndefined();
+  });
 });
 
 describe('SSR safety', () => {
@@ -320,7 +480,7 @@ describe('buildShareUrl tab filtering', () => {
       g_model: 'x',
       i_seq: 'y',
       i_pctl: 'p75',
-      c_mw: '10',
+      c_mw: '20',
       c_costcap: '0.5',
       r_range: 'last-7-days',
     });
@@ -330,7 +490,7 @@ describe('buildShareUrl tab filtering', () => {
     expect(url).toContain('g_model=x');
     expect(url).toContain('i_seq=y');
     expect(url).toContain('i_pctl=p75');
-    expect(url).toContain('c_mw=10');
+    expect(url).toContain('c_mw=20');
     expect(url).toContain('c_costcap=0.5');
     expect(url).not.toContain('r_range');
   });
@@ -347,6 +507,30 @@ describe('buildShareUrl tab filtering', () => {
     expect(url).not.toContain('r_range');
   });
 
+  it('preserves the exact locale, slug, scenario, and hash for non-dashboard chart routes', async () => {
+    setupWindow('', '/zh/compare/h100-vs-h200/8k-1k', '#interactive-results');
+    const { writeUrlParams, buildShareUrl } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'x', i_seq: '8k/1k', r_range: 'last-7-days' });
+    await vi.advanceTimersByTimeAsync(200);
+
+    expect(buildShareUrl()).toBe(
+      'https://example.com/zh/compare/h100-vs-h200/8k-1k?g_model=x&i_seq=8k%2F1k#interactive-results',
+    );
+  });
+
+  it('uses the localized dashboard route to select its share scope', async () => {
+    setupWindow('', '/zh/evaluation', '#samples');
+    const { writeUrlParams, buildShareUrl } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'x', e_bench: 'mmlu', i_seq: '8k/1k' });
+    await vi.advanceTimersByTimeAsync(200);
+
+    expect(buildShareUrl()).toBe(
+      'https://example.com/zh/evaluation?g_model=x&e_bench=mmlu#samples',
+    );
+  });
+
   it('omits query string when no non-default params exist', async () => {
     setupWindow('', '/inference');
     const { buildShareUrl } = await import('@/lib/url-state');
@@ -460,3 +644,261 @@ describe('buildShareUrl unofficialrun handling', () => {
     expect(url).not.toContain('unofficialrun');
   });
 });
+
+// ---------------------------------------------------------------------------
+// Carrying chart state across a full-document navigation.
+//
+// Filter changes only ever reach the in-memory store — the address bar is
+// stripped clean after load — so the history entry behind a detail-page link
+// was a bare `/inference` that rebuilt from defaults on Back.
+// ---------------------------------------------------------------------------
+
+describe('currentChartSearch', () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+    vi.resetModules();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+  });
+
+  it('returns the inference params a share link would carry', async () => {
+    setupWindow('', '/inference');
+    const { currentChartSearch, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3', i_active: 'gb200_dynamo-vllm', i_seq: 'agentic-traces' });
+    // Reads through the debounce rather than around it: the click that
+    // navigates away can land inside the 150ms window.
+    const search = currentChartSearch();
+
+    expect(new URLSearchParams(search).get('g_model')).toBe('Kimi-K3');
+    expect(new URLSearchParams(search).get('i_active')).toBe('gb200_dynamo-vllm');
+    expect(new URLSearchParams(search).get('i_seq')).toBe('agentic-traces');
+  });
+
+  it('resolves the tab through the /zh prefix and any trailing segments', async () => {
+    setupWindow('', '/zh/inference/agentic/440106');
+    const { currentChartSearch, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3', r_range: 'all-time' });
+    const params = new URLSearchParams(currentChartSearch());
+
+    // `inference` prefixes, not the reliability tab's — a detail page carries
+    // the chart it was opened from.
+    expect(params.get('g_model')).toBe('Kimi-K3');
+    expect(params.has('r_range')).toBe(false);
+  });
+
+  it('is empty when every value is at its default', async () => {
+    setupWindow('', '/inference');
+    const { currentChartSearch, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'DeepSeek-V4-Pro', r_range: 'last-3-months' });
+    await vi.advanceTimersByTimeAsync(200);
+
+    expect(currentChartSearch()).toBe('');
+  });
+
+  it('is a no-op without a window (SSR)', async () => {
+    vi.stubGlobal('window', undefined);
+    const { currentChartSearch } = await import('@/lib/url-state');
+    expect(currentChartSearch()).toBe('');
+  });
+});
+
+describe('rememberChartStateInUrl', () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+    vi.resetModules();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+  });
+
+  it('writes the chart state into the current history entry', async () => {
+    const { history } = setupWindow('', '/inference');
+    const { rememberChartStateInUrl, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3', i_active: 'gb200_dynamo-vllm' });
+    rememberChartStateInUrl();
+
+    expect(history.replaceState).toHaveBeenCalledTimes(1);
+    const [state] = history.replaceState.mock.calls[0] as [unknown, string, string];
+    const url = (history.replaceState.mock.calls[0] as [unknown, string, string])[2];
+    // Next patches replaceState and keeps its routing bookkeeping in `state`;
+    // dropping it would strand the router on the entry we just rewrote.
+    expect(state).toEqual({ __NA: 1 });
+    expect(url.startsWith('/inference?')).toBe(true);
+    const params = new URLSearchParams(url.slice(url.indexOf('?')));
+    expect(params.get('g_model')).toBe('Kimi-K3');
+    expect(params.get('i_active')).toBe('gb200_dynamo-vllm');
+  });
+
+  it('keeps the pathname and hash intact', async () => {
+    const { history } = setupWindow('', '/zh/inference', '#chart');
+    const { rememberChartStateInUrl, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3' });
+    rememberChartStateInUrl();
+
+    const url = (history.replaceState.mock.calls[0] as [unknown, string, string])[2];
+    expect(url).toBe('/zh/inference?g_model=Kimi-K3#chart');
+  });
+
+  it('leaves a default-only chart on a bare path rather than an empty query', async () => {
+    const { history } = setupWindow('', '/inference');
+    const { rememberChartStateInUrl } = await import('@/lib/url-state');
+
+    expect(rememberChartStateInUrl()).toBe('');
+    const url = (history.replaceState.mock.calls[0] as [unknown, string, string])[2];
+    expect(url).toBe('/inference');
+  });
+
+  it('carries an unofficial run overlay across the navigation', async () => {
+    const { history } = setupWindow('?unofficialruns=987654321', '/inference');
+    const { rememberChartStateInUrl, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3' });
+    rememberChartStateInUrl();
+
+    const url = (history.replaceState.mock.calls[0] as [unknown, string, string])[2];
+    expect(new URLSearchParams(url.slice(url.indexOf('?'))).get('unofficialruns')).toBe(
+      '987654321',
+    );
+  });
+
+  it('is a no-op without a window (SSR)', async () => {
+    vi.stubGlobal('window', undefined);
+    const { rememberChartStateInUrl } = await import('@/lib/url-state');
+    expect(() => rememberChartStateInUrl()).not.toThrow();
+    expect(rememberChartStateInUrl()).toBe('');
+  });
+});
+
+describe('withChartState', () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+    vi.resetModules();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+  });
+
+  it('appends the chart state to an in-app href', async () => {
+    setupWindow('', '/inference');
+    const { withChartState, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3', i_active: 'gb200_dynamo-vllm' });
+    const href = withChartState('/inference/agentic/440106');
+
+    expect(href.startsWith('/inference/agentic/440106?')).toBe(true);
+    expect(new URLSearchParams(href.slice(href.indexOf('?'))).get('g_model')).toBe('Kimi-K3');
+  });
+
+  it('merges into an href that already has a query', async () => {
+    setupWindow('', '/inference');
+    const { withChartState, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3' });
+    const href = withChartState('/inference?view=timeline');
+
+    expect(href).toBe('/inference?view=timeline&g_model=Kimi-K3');
+  });
+
+  it('returns the href untouched when there is no state to carry', async () => {
+    setupWindow('', '/inference');
+    const { withChartState } = await import('@/lib/url-state');
+
+    expect(withChartState('/inference/agentic/440106')).toBe('/inference/agentic/440106');
+  });
+
+  it('returns the href untouched without a window (SSR)', async () => {
+    vi.stubGlobal('window', undefined);
+    const { withChartState } = await import('@/lib/url-state');
+    expect(withChartState('/inference')).toBe('/inference');
+  });
+});
+
+describe('refreshUrlParamsOnNavigation', () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+    vi.resetModules();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+  });
+
+  it('pulls the new URL in once per navigation', async () => {
+    const { location } = setupWindow('', '/');
+    const { readUrlParams, refreshUrlParamsOnNavigation } = await import('@/lib/url-state');
+
+    // First call on a fresh document still refreshes — the load-time snapshot
+    // and the live URL agree, so it is a no-op in effect.
+    expect(refreshUrlParamsOnNavigation('/')).toBe(true);
+
+    location.search = '?g_model=Kimi-K3&i_active=gb200_dynamo-vllm';
+    location.pathname = '/inference';
+    expect(refreshUrlParamsOnNavigation('/inference')).toBe(true);
+    expect(readUrlParams().g_model).toBe('Kimi-K3');
+    expect(readUrlParams().i_active).toBe('gb200_dynamo-vllm');
+  });
+
+  it('does not re-import the URL for a component mounting later on the same path', async () => {
+    const { location } = setupWindow('?g_model=Kimi-K3', '/inference');
+    const { readUrlParams, refreshUrlParamsOnNavigation } = await import('@/lib/url-state');
+
+    expect(refreshUrlParamsOnNavigation('/inference')).toBe(true);
+    expect(readUrlParams().g_model).toBe('Kimi-K3');
+
+    // The user picks another model. Writes go to the in-memory store, never the
+    // address bar, so the stale URL must not be replayed over the new choice
+    // when some unrelated component mounts.
+    location.search = '?g_model=Kimi-K3';
+    expect(refreshUrlParamsOnNavigation('/inference')).toBe(false);
+  });
+});
+
+describe('rememberChartStateInUrl — params this module does not own', () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+    vi.resetModules();
+  });
+
+  afterEach(() => {
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+  });
+
+  it('keeps unrelated query params on the entry it rewrites', async () => {
+    const { history } = setupWindow('?utm_source=twitter', '/inference');
+    const { rememberChartStateInUrl, writeUrlParams } = await import('@/lib/url-state');
+
+    writeUrlParams({ g_model: 'Kimi-K3' });
+    rememberChartStateInUrl();
+
+    const url = (history.replaceState.mock.calls[0] as [unknown, string, string])[2];
+    const params = new URLSearchParams(url.slice(url.indexOf('?')));
+    expect(params.get('utm_source')).toBe('twitter');
+    expect(params.get('g_model')).toBe('Kimi-K3');
+  });
+
+  it('collapses the singular unofficialrun spelling onto the canonical plural', async () => {
+    const { history } = setupWindow('?unofficialrun=987654321', '/inference');
+    const { rememberChartStateInUrl } = await import('@/lib/url-state');
+
+    rememberChartStateInUrl();
+
+    const url = (history.replaceState.mock.calls[0] as [unknown, string, string])[2];
+    const params = new URLSearchParams(url.slice(url.indexOf('?')));
+    expect(params.get('unofficialruns')).toBe('987654321');
+    expect(params.has('unofficialrun')).toBe(false);
+  });
+});
diff --git a/packages/app/src/lib/url-state.ts b/packages/app/src/lib/url-state.ts
index ed5092616..244ee8bd3 100644
--- a/packages/app/src/lib/url-state.ts
+++ b/packages/app/src/lib/url-state.ts
@@ -12,6 +12,8 @@
  *
  * Only non-default values are written to keep URLs short.
  */
+import { dashboardRouteForPathname, getDashboardRoute } from '@/lib/dashboard-routes';
+import { routeModelForPathname } from '@/lib/model-routes';
 
 // All known share-link parameter keys
 const URL_STATE_KEYS = [
@@ -23,6 +25,8 @@ const URL_STATE_KEYS = [
   'i_seq',
   'i_prec',
   'i_metric',
+  // Token-revenue sale-price source: normalized $1/M or live OpenRouter catalog.
+  'i_revenue',
   'i_pctl',
   'i_xmetric',
   'i_e2e_xmetric',
@@ -43,10 +47,9 @@ const URL_STATE_KEYS = [
   'i_log',
   'i_legend',
   'i_advlabel',
+  'i_conclabel',
   'i_gradlabel',
   'i_linelabel',
-  'i_speed',
-  'i_mc',
   'i_active',
   // Quick filters (vendor / framework / deployment / mtp-stp).
   // `i_disagg` keeps its historical name for shared-link compatibility.
@@ -73,12 +76,43 @@ const URL_STATE_KEYS = [
   // Calculator (fleet planner)
   'c_mw',
   'c_costcap',
+  // Calculator (fleet lifecycle) — token price and the ramp/interrupt/horizon
+  // assumptions (the throughput steps are measured; those four are not), plus
+  // which metric the chart plots.
+  'c_price',
+  // Output-token price. Input keeps `c_price`, so a link written before the
+  // split still seeds the field it always seeded.
+  'c_oprice',
+  'c_ramp',
+  // What a cached input token sells for, as a % of the price. Agentic only —
+  // fixed sequences record no cache hits for it to apply to.
+  'c_cache',
+  'c_ly',
+  'c_mtbi',
+  'c_rec',
+  'c_life',
 ] as const;
 
 export type UrlStateKey = (typeof URL_STATE_KEYS)[number];
 export type UrlStateParams = Partial<Record<UrlStateKey, string>>;
 
 /** Default values for each parameter. Params matching their default are omitted from share URLs. */
+/**
+ * Dashboard default y-axis: total tokens purchasable per $1 USD at owning
+ * Neocloud TCO, so the dashboard leads with the economics rather than raw
+ * throughput. `?i_metric=` still wins, so existing shared links are unaffected.
+ *
+ * Lives here rather than in `InferenceContext` because `PARAM_DEFAULTS` below
+ * strips any value equal to the default from share links. If the two drifted,
+ * a link captured on the *other* metric would be written without `i_metric`
+ * and reopen on this one.
+ */
+export const DEFAULT_Y_AXIS_METRIC = 'y_tokensPerDollarN';
+
+/** Shared defaults for the fleet lifecycle and calculator MW controls. */
+export const DEFAULT_FLEET_MW = '10';
+export const DEFAULT_LIFECYCLE_RAMP_MONTHS = '0.5';
+
 export const PARAM_DEFAULTS: Record<UrlStateKey, string> = {
   g_model: 'DeepSeek-V4-Pro',
   g_rundate: '',
@@ -95,7 +129,8 @@ export const PARAM_DEFAULTS: Record<UrlStateKey, string> = {
   // explicitly, so an explicit FP4 selection must survive (not be stripped as a
   // "default") or it would silently revert to the per-model auto default on reload.
   i_prec: '',
-  i_metric: 'y_tpPerGpu',
+  i_metric: DEFAULT_Y_AXIS_METRIC,
+  i_revenue: 'normalized',
   i_pctl: 'p90',
   i_xmetric: 'p90_ttft',
   i_e2e_xmetric: 'p90_ttft',
@@ -113,10 +148,9 @@ export const PARAM_DEFAULTS: Record<UrlStateKey, string> = {
   i_log: '',
   i_legend: '',
   i_advlabel: '',
+  i_conclabel: '',
   i_gradlabel: '',
   i_linelabel: '',
-  i_speed: '',
-  i_mc: '',
   i_active: '',
   i_vendor: '',
   i_fw: '',
@@ -135,18 +169,21 @@ export const PARAM_DEFAULTS: Record<UrlStateKey, string> = {
   r_hc: '',
   r_legend: '',
   r_active: '',
-  c_mw: '',
+  c_mw: DEFAULT_FLEET_MW,
   c_costcap: '',
-};
-
-/** Which param prefixes are relevant per tab. */
-const TAB_PARAM_PREFIXES: Record<string, string[]> = {
-  inference: ['g_', 'i_'],
-  evaluation: ['g_', 'e_'],
-  reliability: ['r_'],
-  // The calculator reuses the global model + inference sequence/precision
-  // params, plus its own c_ scope (fleet planner MW / cost cap).
-  calculator: ['g_', 'i_', 'c_'],
+  // Empty means "use the component's default", which for these two is derived
+  // rather than constant: the price from the cheapest visible chip's break-even,
+  // the horizon from the measured run window. A non-empty default here would be
+  // stripped from share links and silently re-derived on load.
+  c_price: '',
+  c_oprice: '',
+  c_life: '',
+  // Empty means the default y metric (margin).
+  c_ly: '',
+  c_ramp: DEFAULT_LIFECYCLE_RAMP_MONTHS,
+  c_cache: '10',
+  c_mtbi: '24',
+  c_rec: '12',
 };
 
 /** In-memory store of current param values (kept in sync via writeUrlParams). */
@@ -155,12 +192,14 @@ const currentState: Record<string, string> = {};
 // On module load: snapshot share-link params from the URL.
 // Cleanup is deferred so it runs after Next.js hydration finishes.
 const _initialParams: UrlStateParams = {};
+const explicitUrlParams: UrlStateParams = {};
 if (typeof window !== 'undefined') {
   const searchParams = new URLSearchParams(window.location.search);
   for (const key of URL_STATE_KEYS) {
     const value = searchParams.get(key);
     if (value !== null) {
       _initialParams[key] = value;
+      explicitUrlParams[key] = value;
       currentState[key] = value;
     }
   }
@@ -171,15 +210,83 @@ if (typeof window !== 'undefined') {
       sp.delete(key);
     }
     const s = sp.toString();
-    window.history.replaceState(null, '', `${window.location.pathname}${s ? `?${s}` : ''}`);
+    window.history.replaceState(
+      null,
+      '',
+      `${window.location.pathname}${s ? `?${s}` : ''}${window.location.hash}`,
+    );
   }, 0);
 }
 
-/** Returns the share-link params that were in the URL at page load. */
+/** Returns the current share-link state, flushing pending writes for provider remounts. */
 export function readUrlParams(): UrlStateParams {
+  flushPendingParams();
   return _initialParams;
 }
 
+/**
+ * Re-read share-link params from the live URL, replacing the load-time
+ * snapshot.
+ *
+ * The snapshot above is captured once per page load, which is correct for a
+ * hard navigation but wrong for a client-side one: a soft transition to
+ * `/inference?g_model=…` does not remount the provider, so every reader kept
+ * seeing the params of the page the user came FROM (usually none). Callers
+ * must only invoke this on a real router navigation — self-writes go through
+ * `history.replaceState`, which deliberately does not, so re-reading after one
+ * of those would fight the user's own filter changes.
+ *
+ * Also mirrors into `currentState` so the next share-link write starts from
+ * what the URL actually asked for.
+ */
+export function refreshUrlParams(): UrlStateParams {
+  if (typeof window === 'undefined') return _initialParams;
+  flushPendingParams();
+  const searchParams = new URLSearchParams(window.location.search);
+  for (const key of URL_STATE_KEYS) {
+    delete explicitUrlParams[key];
+    const value = searchParams.get(key);
+    if (value === null) continue;
+    explicitUrlParams[key] = value;
+    _initialParams[key] = value;
+    currentState[key] = value;
+  }
+  return _initialParams;
+}
+
+/**
+ * Pathname of the last navigation whose params were pulled into the snapshot.
+ * Module-level (not a hook ref) so only the FIRST `useUrlState` consumer to
+ * render after a navigation does the work: a component that mounts later on
+ * the same path must not re-import the URL over filter changes the user has
+ * made since.
+ */
+let lastRefreshedPathname: string | null = null;
+
+/**
+ * Refresh the snapshot once per client-side navigation.
+ *
+ * Providers read their initial state through `getUrlParam` during render, so
+ * the snapshot has to be current BEFORE they initialise. `useUrlState` calls
+ * this while rendering, which puts it ahead of every provider below the
+ * component that called it first (React renders top-down), and ahead of the
+ * mount effects that read `i_gpus` and friends.
+ *
+ * Returns whether a refresh actually happened, which is what the unit tests
+ * assert on.
+ */
+export function refreshUrlParamsOnNavigation(pathname: string): boolean {
+  if (lastRefreshedPathname === pathname) return false;
+  lastRefreshedPathname = pathname;
+  refreshUrlParams();
+  return true;
+}
+
+/** Whether the current navigation explicitly supplied a share-link parameter. */
+export function hasExplicitUrlParam(key: UrlStateKey): boolean {
+  return explicitUrlParams[key] !== undefined;
+}
+
 /** Check whether the current URL has any share-link params. */
 export function hasAnyUrlParams(): boolean {
   if (typeof window === 'undefined') return false;
@@ -218,8 +325,10 @@ function flushPendingParams(): void {
 
     if (value === undefined || value === defaultValue) {
       delete currentState[urlKey];
+      delete _initialParams[urlKey];
     } else {
       currentState[urlKey] = value;
+      _initialParams[urlKey] = value;
     }
   }
 
@@ -239,12 +348,18 @@ function flushPendingParams(): void {
  */
 const UNOFFICIAL_RUN_PARAM_RE = /^unofficialruns?$/iu;
 
-export function buildShareUrl(): string {
+/**
+ * Current in-memory state for the active route, plus any unofficial-run IDs
+ * in the address bar, as a `URLSearchParams`. Flushes pending writes first so
+ * the caller sees the newest filter change rather than the one before it.
+ */
+function collectTabParams(): URLSearchParams {
   flushPendingParams();
 
-  const pathTab = window.location.pathname.split('/').filter(Boolean)[0] || 'inference';
-  const prefixes = TAB_PARAM_PREFIXES[pathTab] ?? TAB_PARAM_PREFIXES.inference;
-
+  const route = dashboardRouteForPathname(window.location.pathname);
+  // Compare and other chart routes share inference controls but deliberately
+  // stay outside the dashboard registry.
+  const prefixes = route?.shareParamScopes ?? getDashboardRoute('inference').shareParamScopes;
   const filtered = new URLSearchParams();
   for (const [key, value] of Object.entries(currentState)) {
     if (prefixes.some((p) => key.startsWith(p))) {
@@ -252,6 +367,14 @@ export function buildShareUrl(): string {
     }
   }
 
+  // On per-model routes (/calculator/<slug>, /historical/<slug>) the pathname
+  // already encodes the model, so a matching g_model would be redundant in the
+  // share link. A mismatch (possible mid-transition) is kept — explicit
+  // g_model wins over the path on load, so the link still opens correctly.
+  if (routeModelForPathname(window.location.pathname) === filtered.get('g_model')) {
+    filtered.delete('g_model');
+  }
+
   // Carry over any unofficial-run IDs currently reflected in the address bar.
   // Only the first match is forwarded and it's always emitted under the plural
   // `unofficialruns` key — the canonical form the app writes on dismiss/load
@@ -264,6 +387,71 @@ export function buildShareUrl(): string {
     }
   }
 
-  const search = filtered.toString();
-  return `${window.location.origin}/${pathTab}${search ? `?${search}` : ''}`;
+  return filtered;
+}
+
+export function buildShareUrl(): string {
+  const search = collectTabParams().toString();
+  return `${window.location.origin}${window.location.pathname}${search ? `?${search}` : ''}${window.location.hash}`;
+}
+
+/**
+ * The chart state this page would share, as a query string (no leading `?`).
+ * Empty when every value is at its default.
+ */
+export function currentChartSearch(): string {
+  if (typeof window === 'undefined') return '';
+  return collectTabParams().toString();
+}
+
+/**
+ * Write the chart state into the CURRENT history entry, so Back restores it.
+ *
+ * Filter changes only ever reach the in-memory `currentState` — the address bar
+ * is deliberately stripped clean after load — so the entry the browser returns
+ * to is a bare `/inference` that rebuilds from defaults. Detail-page links are
+ * plain `<a href>` (they must support open-in-new-tab), i.e. full-document
+ * navigations that also destroy this module, so the URL is the only channel
+ * that survives. Call this immediately before navigating away from a chart.
+ *
+ * `history.state` is passed straight through so the Next router's own entry
+ * bookkeeping is not clobbered.
+ */
+export function rememberChartStateInUrl(): string {
+  if (typeof window === 'undefined') return '';
+  const { pathname, hash, search: liveSearch } = window.location;
+  const chartParams = collectTabParams();
+
+  // Start from whatever is already in the address bar so params this module
+  // does not own (campaign tags and the like) survive the rewrite, then layer
+  // the chart state on top. Both unofficial-run spellings are cleared first —
+  // `collectTabParams` re-emits the canonical plural form.
+  const merged = new URLSearchParams(liveSearch);
+  for (const key of URL_STATE_KEYS) merged.delete(key);
+  const unofficialKeys: string[] = [];
+  for (const key of merged.keys()) {
+    if (UNOFFICIAL_RUN_PARAM_RE.test(key)) unofficialKeys.push(key);
+  }
+  // Deleting inside the iteration above would skip entries.
+  for (const key of unofficialKeys) merged.delete(key);
+  for (const [key, value] of chartParams) merged.set(key, value);
+
+  const search = merged.toString();
+  window.history.replaceState(
+    window.history.state,
+    '',
+    `${pathname}${search ? `?${search}` : ''}${hash}`,
+  );
+  return chartParams.toString();
+}
+
+/**
+ * Append the current chart state to an outbound in-app href, so the page it
+ * opens can link back to the chart the user left. Used for the agentic
+ * point-detail links, which are full-document navigations.
+ */
+export function withChartState(href: string): string {
+  const search = currentChartSearch();
+  if (!search) return href;
+  return `${href}${href.includes('?') ? '&' : '?'}${search}`;
 }
diff --git a/packages/app/src/lib/utils.test.ts b/packages/app/src/lib/utils.test.ts
index 13de88b72..bca384444 100644
--- a/packages/app/src/lib/utils.test.ts
+++ b/packages/app/src/lib/utils.test.ts
@@ -137,25 +137,24 @@ describe('updateRepoUrl', () => {
 });
 
 // ===========================================================================
-// calculateCostsForGpus
+// calculateCostsForGpus (produces both cost-per-million and tokens-per-dollar)
 // costh=2.8, costn=1.4, costr=0.7 (from mock, but this fn uses userCosts arg)
 // ===========================================================================
 describe('calculateCostsForGpus', () => {
-  it('computes cost correctly: tpPerGpu=1000, userCost=$5/hr → 1.389', () => {
-    // tokensPerHour = (1000 * 3600) / 1_000_000 = 3.6
-    // costPerMillion = 5 / 3.6 = 1.388... → 1.389
+  it('computes purchasing power: tpPerGpu=1000, userCost=$5/hr → 720,000 tok/$', () => {
     const data = [pt({ hwKey: 'h100' as any, tpPerGpu: { y: 1000, roof: false } })];
     const result = calculateCostsForGpus(data, { h100: 5 });
     expect(result[0].costUser?.y).toBe(1.389);
+    expect(result[0].tokensPerDollarUser?.y).toBe(720_000);
     expect(result[0].y).toBe(1.389);
     expect(result[0].costUser?.roof).toBe(false);
   });
 
-  it('computes cost correctly: tpPerGpu=1000, userCost=$3/hr → 0.833', () => {
-    // costPerMillion = 3 / 3.6 = 0.833...
+  it('computes purchasing power: tpPerGpu=1000, userCost=$3/hr → 1,200,000 tok/$', () => {
     const data = [pt({ hwKey: 'h100' as any, tpPerGpu: { y: 1000, roof: false } })];
     const result = calculateCostsForGpus(data, { h100: 3 });
     expect(result[0].costUser?.y).toBe(0.833);
+    expect(result[0].tokensPerDollarUser?.y).toBe(1_200_000);
     expect(result[0].y).toBe(0.833);
   });
 
@@ -170,6 +169,7 @@ describe('calculateCostsForGpus', () => {
     const data = [pt({ hwKey: 'h100_trt' as any, tpPerGpu: { y: 1000, roof: false } })];
     const result = calculateCostsForGpus(data, { h100: 5 });
     expect(result[0].costUser?.y).toBe(1.389);
+    expect(result[0].tokensPerDollarUser?.y).toBe(720_000);
   });
 
   it('processes multiple items independently', () => {
@@ -178,10 +178,17 @@ describe('calculateCostsForGpus', () => {
       pt({ hwKey: 'a100' as any, tpPerGpu: { y: 2000, roof: false } }),
     ];
     const result = calculateCostsForGpus(data, { h100: 5, a100: 5 });
-    // h100: 5/3.6 = 1.389
     expect(result[0].costUser?.y).toBe(1.389);
-    // a100: tokensPerHour=(2000*3600)/1e6=7.2; cost=5/7.2=0.694
     expect(result[1].costUser?.y).toBe(0.694);
+    expect(result[0].tokensPerDollarUser?.y).toBe(720_000);
+    expect(result[1].tokensPerDollarUser?.y).toBe(1_440_000);
+  });
+
+  it('returns zero purchasing power for a zero hourly cost', () => {
+    const data = [pt({ hwKey: 'h100' as any, tpPerGpu: { y: 1000, roof: false } })];
+    const result = calculateCostsForGpus(data, { h100: 0 });
+    expect(result[0].costUser?.y).toBe(0);
+    expect(result[0].tokensPerDollarUser?.y).toBe(0);
   });
 });
 
@@ -239,9 +246,7 @@ describe('computeOutputCostFields', () => {
     expect(result[0].costrOutput?.y).toBe(99);
   });
 
-  it('computes output cost fields from outputTputPerGpu when provided: tput=1000', () => {
-    // outputTokensPerHour = (1000 * 3600) / 1_000_000 = 3.6
-    // costhOutput = 2.8 / 3.6 = 0.778, costnOutput = 1.4/3.6 = 0.389, costrOutput = 0.7/3.6 = 0.194
+  it('computes both output cost and output tokens per dollar', () => {
     const item = pt({
       outputTputPerGpu: { y: 1000, roof: false },
     });
@@ -249,18 +254,18 @@ describe('computeOutputCostFields', () => {
     expect(result[0].costhOutput?.y).toBe(0.778);
     expect(result[0].costnOutput?.y).toBe(0.389);
     expect(result[0].costrOutput?.y).toBe(0.194);
+    expect(result[0].outputTokensPerDollarH?.y).toBeCloseTo(1_285_714.286, 3);
     expect(result[0].costhOutput?.roof).toBe(false);
   });
 
   it('falls back to tpPerGpu * 0.875 when outputTputPerGpu is absent: tpPerGpu=1000', () => {
     // fallback tput = 1000 * 0.875 = 875
-    // outputTokensPerHour = (875 * 3600) / 1_000_000 = 3.15
-    // costhOutput = 2.8/3.15 = 0.889, costnOutput = 1.4/3.15 = 0.444, costrOutput = 0.7/3.15 = 0.222
     const item = pt({ tpPerGpu: { y: 1000, roof: false } });
     const result = computeOutputCostFields([item]);
     expect(result[0].costhOutput?.y).toBe(0.889);
     expect(result[0].costnOutput?.y).toBe(0.444);
     expect(result[0].costrOutput?.y).toBe(0.222);
+    expect(result[0].outputTokensPerDollarH?.y).toBeCloseTo(1_125_000, 3);
   });
 
   it('does not overwrite an existing costhOutput field (partial override)', () => {
@@ -299,9 +304,7 @@ describe('computeInputCostFields', () => {
     expect(result[0].costri?.y).toBe(99);
   });
 
-  it('computes input cost fields from inputTputPerGpu when provided: tput=200', () => {
-    // inputTokensPerHour = (200 * 3600) / 1_000_000 = 0.72
-    // costhi = 2.8/0.72 = 3.889, costni = 1.4/0.72 = 1.944, costri = 0.7/0.72 = 0.972
+  it('computes both input cost and input tokens per dollar', () => {
     const item = pt({
       costhi: undefined as any,
       costni: undefined as any,
@@ -312,13 +315,12 @@ describe('computeInputCostFields', () => {
     expect(result[0].costhi?.y).toBe(3.889);
     expect(result[0].costni?.y).toBe(1.944);
     expect(result[0].costri?.y).toBe(0.972);
+    expect(result[0].inputTokensPerDollarH?.y).toBeCloseTo(257_142.857, 3);
     expect(result[0].costhi?.roof).toBe(false);
   });
 
   it('falls back to tpPerGpu * 0.125 when inputTputPerGpu is absent: tpPerGpu=1000', () => {
     // fallback tput = 1000 * 0.125 = 125
-    // inputTokensPerHour = (125 * 3600) / 1_000_000 = 0.45
-    // costhi = 2.8/0.45 = 6.222, costni = 1.4/0.45 = 3.111, costri = 0.7/0.45 = 1.556
     const item = pt({
       costhi: undefined as any,
       costni: undefined as any,
@@ -329,6 +331,7 @@ describe('computeInputCostFields', () => {
     expect(result[0].costhi?.y).toBe(6.222);
     expect(result[0].costni?.y).toBe(3.111);
     expect(result[0].costri?.y).toBe(1.556);
+    expect(result[0].inputTokensPerDollarH?.y).toBeCloseTo(160_714.286, 3);
   });
 
   it('does not overwrite an existing costhi field (partial override)', () => {
diff --git a/packages/app/src/lib/utils.ts b/packages/app/src/lib/utils.ts
index 7a2472137..6a252f4bb 100644
--- a/packages/app/src/lib/utils.ts
+++ b/packages/app/src/lib/utils.ts
@@ -1,10 +1,23 @@
 import { type ClassValue, clsx } from 'clsx';
-import { twMerge } from 'tailwind-merge';
+import { extendTailwindMerge } from 'tailwind-merge';
 
 import type { AggDataEntry, InferenceData, RunInfo } from '@/components/inference/types';
-import { FRAMEWORK_LABELS } from '@semianalysisai/inferencex-constants';
+import { FRAMEWORK_LABELS, USD_TO_CNY } from '@semianalysisai/inferencex-constants';
 
 import { getGpuSpecs } from './constants';
+
+// Custom letter-spacing tokens from globals.css. tailwind-merge only knows the
+// built-in tracking scale (tighter…widest), so without this, e.g.
+// cn('tracking-eyebrow', 'tracking-tight') would keep both classes.
+// (text-2xs/text-3xs need no config: t-shirt sizes are recognized as font sizes.)
+const twMerge = extendTailwindMerge({
+  extend: {
+    theme: {
+      tracking: ['eyebrow', 'eyebrow-wide', 'heading'],
+    },
+  },
+});
+
 /**
  * Combines Tailwind CSS classes and other class values into a single string.
  * This utility helps in conditionally applying classes and merging them efficiently,
@@ -70,7 +83,7 @@ export function formatNumber(tickItem: number) {
 }
 
 /**
- * Calculate costs for each GPU using the formula
+ * Calculate both cost-per-million and tokens-per-dollar values from a user-provided hourly cost.
  * GPUs with prefixes (TRT, MTP) will inherit values from their base parent
  */
 export function calculateCostsForGpus(
@@ -88,16 +101,21 @@ export function calculateCostsForGpus(
     }
     if (userCostPerHour !== undefined) {
       const tputPerGpu = item.tpPerGpu.y;
-      const tokensPerHour = (tputPerGpu * 3600) / 1000000;
-      const costPerMillion = userCostPerHour / tokensPerHour;
+      const millionTokensPerHour = (tputPerGpu * 3600) / 1_000_000;
+      const costPerMillion = millionTokensPerHour > 0 ? userCostPerHour / millionTokensPerHour : 0;
+      const tokensPerDollar = userCostPerHour > 0 ? (tputPerGpu * 3600) / userCostPerHour : 0;
       const costRounded = parseFloat(costPerMillion.toFixed(3));
+      const tokensRounded = parseFloat(tokensPerDollar.toFixed(3));
 
-      // Return the data with costUser property and updated y value
       return {
         ...item,
-        y: costRounded, // Update the main y value for chart rendering
+        y: costRounded,
         costUser: {
           y: costRounded,
+          roof: false,
+        },
+        tokensPerDollarUser: {
+          y: tokensRounded,
           roof: false, // Always false for user-calculated values
         },
       };
@@ -183,18 +201,24 @@ export function getDisplayLabel(config: { label: string; suffix?: string }): str
 }
 
 /**
- * Computes missing output cost fields (costhOutput, costnOutput, costrOutput) for data points.
+ * Computes missing output cost and tokens-per-dollar fields for historical data points.
  * This handles backwards compatibility with historical data that doesn't have these fields.
  *
- * The calculation is: costPerHour / outputTokensPerHour
- * where outputTokensPerHour = (outputTputPerGpu * 3600) / 1000000
- *
  * If outputTputPerGpu is not available, falls back to using the total throughput ratio.
  */
 export function computeOutputCostFields(data: InferenceData[]): InferenceData[] {
   return data.map((item) => {
-    // If output cost fields already exist, return as-is
-    if (item.costhOutput && item.costnOutput && item.costrOutput) {
+    if (
+      item.costhOutput &&
+      item.costnOutput &&
+      item.costrOutput &&
+      item.outputTokensPerDollarH &&
+      item.outputTokensPerDollarN &&
+      item.outputTokensPerDollarR &&
+      item.outputTokensPerRmbH &&
+      item.outputTokensPerRmbN &&
+      item.outputTokensPerRmbR
+    ) {
       return item;
     }
 
@@ -204,12 +228,14 @@ export function computeOutputCostFields(data: InferenceData[]): InferenceData[]
     // Get output throughput - either from outputTputPerGpu or estimate from total throughput
     // For sequence pairs like 1k/8k (ISL/OSL), output tokens dominate, typically ~87.5% of total
     const outputTputPerGpu = item.outputTputPerGpu?.y ?? item.tpPerGpu.y * 0.875;
-    const outputTokensPerHour = (outputTputPerGpu * 3600) / 1000000;
-
-    // Calculate output cost per million tokens for each cost type
-    const costhOutput = outputTokensPerHour > 0 ? specs.costh / outputTokensPerHour : 0;
-    const costnOutput = outputTokensPerHour > 0 ? specs.costn / outputTokensPerHour : 0;
-    const costrOutput = outputTokensPerHour > 0 ? specs.costr / outputTokensPerHour : 0;
+    const outputTokensPerHour = outputTputPerGpu * 3600;
+    const millionOutputTokensPerHour = outputTokensPerHour / 1_000_000;
+    const costhOutput =
+      millionOutputTokensPerHour > 0 ? specs.costh / millionOutputTokensPerHour : 0;
+    const costnOutput =
+      millionOutputTokensPerHour > 0 ? specs.costn / millionOutputTokensPerHour : 0;
+    const costrOutput =
+      millionOutputTokensPerHour > 0 ? specs.costr / millionOutputTokensPerHour : 0;
 
     return {
       ...item,
@@ -225,17 +251,38 @@ export function computeOutputCostFields(data: InferenceData[]): InferenceData[]
         y: parseFloat(costrOutput.toFixed(3)),
         roof: false,
       },
+      outputTokensPerDollarH: item.outputTokensPerDollarH ?? {
+        y: specs.costh > 0 ? outputTokensPerHour / specs.costh : 0,
+        roof: false,
+      },
+      outputTokensPerDollarN: item.outputTokensPerDollarN ?? {
+        y: specs.costn > 0 ? outputTokensPerHour / specs.costn : 0,
+        roof: false,
+      },
+      outputTokensPerDollarR: item.outputTokensPerDollarR ?? {
+        y: specs.costr > 0 ? outputTokensPerHour / specs.costr : 0,
+        roof: false,
+      },
+      outputTokensPerRmbH: item.outputTokensPerRmbH ?? {
+        y: specs.costh > 0 ? outputTokensPerHour / (specs.costh * USD_TO_CNY) : 0,
+        roof: false,
+      },
+      outputTokensPerRmbN: item.outputTokensPerRmbN ?? {
+        y: specs.costn > 0 ? outputTokensPerHour / (specs.costn * USD_TO_CNY) : 0,
+        roof: false,
+      },
+      outputTokensPerRmbR: item.outputTokensPerRmbR ?? {
+        y: specs.costr > 0 ? outputTokensPerHour / (specs.costr * USD_TO_CNY) : 0,
+        roof: false,
+      },
     };
   });
 }
 
 /**
- * Computes missing input cost fields (costhi, costni, costri) at runtime.
+ * Computes missing input cost and tokens-per-dollar fields at runtime.
  * This handles backwards compatibility with historical data that doesn't have these fields.
  *
- * The calculation is: costPerHour / inputTokensPerHour
- * where inputTokensPerHour = (inputTputPerGpu * 3600) / 1000000
- *
  * If inputTputPerGpu is not available, falls back to using a portion of total throughput.
  */
 /**
@@ -359,8 +406,17 @@ export function computeEnergyFields(data: InferenceData[]): InferenceData[] {
 
 export function computeInputCostFields(data: InferenceData[]): InferenceData[] {
   return data.map((item) => {
-    // If input cost fields already exist, return as-is
-    if (item.costhi && item.costni && item.costri) {
+    if (
+      item.costhi &&
+      item.costni &&
+      item.costri &&
+      item.inputTokensPerDollarH &&
+      item.inputTokensPerDollarN &&
+      item.inputTokensPerDollarR &&
+      item.inputTokensPerRmbH &&
+      item.inputTokensPerRmbN &&
+      item.inputTokensPerRmbR
+    ) {
       return item;
     }
 
@@ -370,12 +426,11 @@ export function computeInputCostFields(data: InferenceData[]): InferenceData[] {
     // Get input throughput - either from inputTputPerGpu or estimate from total throughput
     // For sequence pairs like 1k/8k (ISL/OSL), input tokens are typically ~12.5% of total
     const inputTputPerGpu = item.inputTputPerGpu?.y ?? item.tpPerGpu.y * 0.125;
-    const inputTokensPerHour = (inputTputPerGpu * 3600) / 1000000;
-
-    // Calculate input cost per million tokens for each cost type
-    const costhi = inputTokensPerHour > 0 ? specs.costh / inputTokensPerHour : 0;
-    const costni = inputTokensPerHour > 0 ? specs.costn / inputTokensPerHour : 0;
-    const costri = inputTokensPerHour > 0 ? specs.costr / inputTokensPerHour : 0;
+    const inputTokensPerHour = inputTputPerGpu * 3600;
+    const millionInputTokensPerHour = inputTokensPerHour / 1_000_000;
+    const costhi = millionInputTokensPerHour > 0 ? specs.costh / millionInputTokensPerHour : 0;
+    const costni = millionInputTokensPerHour > 0 ? specs.costn / millionInputTokensPerHour : 0;
+    const costri = millionInputTokensPerHour > 0 ? specs.costr / millionInputTokensPerHour : 0;
 
     return {
       ...item,
@@ -391,6 +446,30 @@ export function computeInputCostFields(data: InferenceData[]): InferenceData[] {
         y: parseFloat(costri.toFixed(3)),
         roof: false,
       },
+      inputTokensPerDollarH: item.inputTokensPerDollarH ?? {
+        y: specs.costh > 0 ? inputTokensPerHour / specs.costh : 0,
+        roof: false,
+      },
+      inputTokensPerDollarN: item.inputTokensPerDollarN ?? {
+        y: specs.costn > 0 ? inputTokensPerHour / specs.costn : 0,
+        roof: false,
+      },
+      inputTokensPerDollarR: item.inputTokensPerDollarR ?? {
+        y: specs.costr > 0 ? inputTokensPerHour / specs.costr : 0,
+        roof: false,
+      },
+      inputTokensPerRmbH: item.inputTokensPerRmbH ?? {
+        y: specs.costh > 0 ? inputTokensPerHour / (specs.costh * USD_TO_CNY) : 0,
+        roof: false,
+      },
+      inputTokensPerRmbN: item.inputTokensPerRmbN ?? {
+        y: specs.costn > 0 ? inputTokensPerHour / (specs.costn * USD_TO_CNY) : 0,
+        roof: false,
+      },
+      inputTokensPerRmbR: item.inputTokensPerRmbR ?? {
+        y: specs.costr > 0 ? inputTokensPerHour / (specs.costr * USD_TO_CNY) : 0,
+        roof: false,
+      },
     };
   });
 }
diff --git a/packages/app/src/lib/vendor-logos.test.ts b/packages/app/src/lib/vendor-logos.test.ts
new file mode 100644
index 000000000..cecffa8ab
--- /dev/null
+++ b/packages/app/src/lib/vendor-logos.test.ts
@@ -0,0 +1,36 @@
+import { describe, expect, it } from 'vitest';
+
+import { VENDOR_LOGO_ICONS, getLineLabelVendorIcon } from './vendor-logos';
+
+describe('vendor logo icons', () => {
+  it('maps NVIDIA hardware keys to the full-color NVIDIA mark', () => {
+    for (const key of ['gb300', 'gb200', 'b300', 'b200', 'h100', 'vr200']) {
+      expect(getLineLabelVendorIcon(key)).toBe(VENDOR_LOGO_ICONS.NVIDIA);
+    }
+  });
+
+  it('maps AMD hardware keys to the full-color AMD mark', () => {
+    for (const key of ['mi300x', 'mi325x', 'mi355x']) {
+      expect(getLineLabelVendorIcon(key)).toBe(VENDOR_LOGO_ICONS.AMD);
+    }
+  });
+
+  it('resolves suffixed hardware keys through the base key', () => {
+    expect(getLineLabelVendorIcon('mi355x_dsv4')).toBe(VENDOR_LOGO_ICONS.AMD);
+    expect(getLineLabelVendorIcon('gb200_dynamo')).toBe(VENDOR_LOGO_ICONS.NVIDIA);
+  });
+
+  it('maps Jalapeño (Teacup/OpenAI) to the OpenAI mark', () => {
+    expect(getLineLabelVendorIcon('jalapeno')).toBe(VENDOR_LOGO_ICONS.Teacup);
+  });
+
+  it('returns no icon for unknown hardware', () => {
+    expect(getLineLabelVendorIcon('unknown-hw')).toBeUndefined();
+  });
+
+  it('inlines brand colors in the data URIs', () => {
+    expect(decodeURIComponent(VENDOR_LOGO_ICONS.NVIDIA.href)).toContain('#76B900');
+    expect(decodeURIComponent(VENDOR_LOGO_ICONS.AMD.href)).toContain('#000000');
+    expect(decodeURIComponent(VENDOR_LOGO_ICONS.Teacup.href)).toContain('#ffffff');
+  });
+});
diff --git a/packages/app/src/lib/vendor-logos.ts b/packages/app/src/lib/vendor-logos.ts
new file mode 100644
index 000000000..7ad60d4ff
--- /dev/null
+++ b/packages/app/src/lib/vendor-logos.ts
@@ -0,0 +1,86 @@
+import { GPU_VENDORS } from '@semianalysisai/inferencex-constants';
+
+/**
+ * Maps `HW_REGISTRY` vendor names to logo files under `public/logos/`.
+ *
+ * NVIDIA uses its official full-color mark (NVIDIA green, #76B900), which is
+ * legible on both themes as-is. AMD's corporate mark is monochrome by design —
+ * the brand guidelines only permit the standard black logo or the reversed-out
+ * white logo (the legacy green-arrow mark is retired), so it relies on the
+ * shared dark-mode invert from `isMonochromeLogo` instead of a color variant.
+ *
+ * Vendors without an entry (e.g. Teacup) simply render no logo — surfaces
+ * treat the vendor mark as optional decoration beside the hardware label.
+ */
+export const HW_VENDOR_LOGOS: Record<string, string> = {
+  NVIDIA: 'nvidia-color.svg',
+  AMD: 'amd.svg',
+};
+
+/** Logo filename under `/logos/` for a hardware vendor, if one exists. */
+export function getHwVendorLogo(vendor: string | undefined): string | undefined {
+  return vendor ? HW_VENDOR_LOGOS[vendor] : undefined;
+}
+
+/**
+ * Full-color hardware-vendor marks for chart line labels.
+ *
+ * The SVGs are inlined as `data:` URIs (rather than referencing
+ * `public/logos/*`) so the marks render synchronously inside the D3 SVG —
+ * no network fetch, no flicker on zoom re-renders — and survive the
+ * html-to-image chart export path without any resource embedding.
+ *
+ * Color notes (this is deliberate — do not swap in the monochrome
+ * `public/logos/{nvidia,amd}.svg` simple-icons assets here):
+ * - NVIDIA: official brand green `#76B900` eye mark.
+ * - AMD: the arrow symbol from the official AMD logo, in brand black —
+ *   symbol only, no "AMD" wordmark, so it reads as a square mark like the
+ *   other vendor icons at the 10px label size.
+ * - OpenAI (Jalapeño): the OpenAI mark is monochrome by design; the white
+ *   reversed variant is the official usage on colored backgrounds.
+ * The marks are drawn with a transparent background directly on the pill,
+ * so the area behind each logo stays the exact shade of the line-label fill.
+ */
+export interface VendorLogoIcon {
+  /** `data:image/svg+xml` URI for an SVG `<image href>`. */
+  href: string;
+  /** Rendered mark width in px (chart/user units). */
+  width: number;
+  /** Rendered mark height in px (chart/user units). */
+  height: number;
+}
+
+const svgDataUri = (svg: string): string => `data:image/svg+xml,${encodeURIComponent(svg)}`;
+
+/** NVIDIA eye mark (simple-icons geometry) in NVIDIA brand green. */
+const NVIDIA_LOGO_SVG =
+  '<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 24 24" fill="#76B900"><path d="M8.948 8.798v-1.43a6.7 6.7 0 0 1 .424-.018c3.922-.124 6.493 3.374 6.493 3.374s-2.774 3.851-5.75 3.851c-.398 0-.787-.062-1.158-.185v-4.346c1.528.185 1.837.857 2.747 2.385l2.04-1.714s-1.492-1.952-4-1.952a6.016 6.016 0 0 0-.796.035m0-4.735v2.138l.424-.027c5.45-.185 9.01 4.47 9.01 4.47s-4.08 4.964-8.33 4.964c-.37 0-.733-.035-1.095-.097v1.325c.3.035.61.062.91.062 3.957 0 6.82-2.023 9.593-4.408.459.371 2.34 1.263 2.73 1.652-2.633 2.208-8.772 3.984-12.253 3.984-.335 0-.653-.018-.971-.053v1.864H24V4.063zm0 10.326v1.131c-3.657-.654-4.673-4.46-4.673-4.46s1.758-1.944 4.673-2.262v1.237H8.94c-1.528-.186-2.73 1.245-2.73 1.245s.68 2.412 2.739 3.11M2.456 10.9s2.164-3.197 6.5-3.533V6.201C4.153 6.59 0 10.653 0 10.653s2.35 6.802 8.948 7.42v-1.237c-4.84-.6-6.492-5.936-6.492-5.936z"/></svg>';
+
+/** AMD arrow symbol (extracted from the official logo geometry) in brand black. */
+const AMD_LOGO_SVG =
+  '<svg xmlns="http://www.w3.org/2000/svg" viewBox="18.277 9.137 5.723 5.726" fill="#000000"><path d="M18.324 9.137l1.559 1.56h2.556v2.557L24 14.814V9.137z"/><path d="M19.881 11.01L18.277 12.613L18.277 14.863L20.523 14.863L22.127 13.256L19.881 13.256z"/></svg>';
+
+/** OpenAI mark (repo `public/logos/openai.svg` geometry) in reversed white. */
+const OPENAI_LOGO_SVG =
+  '<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 256 260" fill="#ffffff"><path d="M239.184 106.203a64.716 64.716 0 0 0-5.576-53.103C219.452 28.459 191 15.784 163.213 21.74A65.586 65.586 0 0 0 52.096 45.22a64.716 64.716 0 0 0-43.23 31.36c-14.31 24.602-11.061 55.634 8.033 76.74a64.665 64.665 0 0 0 5.525 53.102c14.174 24.65 42.644 37.324 70.446 31.36a64.72 64.72 0 0 0 48.754 21.744c28.481.025 53.714-18.361 62.414-45.481a64.767 64.767 0 0 0 43.229-31.36c14.137-24.558 10.875-55.423-8.083-76.483Zm-97.56 136.338a48.397 48.397 0 0 1-31.105-11.255l1.535-.87 51.67-29.825a8.595 8.595 0 0 0 4.247-7.367v-72.85l21.845 12.636c.218.111.37.32.409.563v60.367c-.056 26.818-21.783 48.545-48.601 48.601Zm-104.466-44.61a48.345 48.345 0 0 1-5.781-32.589l1.534.921 51.722 29.826a8.339 8.339 0 0 0 8.441 0l63.181-36.425v25.221a.87.87 0 0 1-.358.665l-52.335 30.184c-23.257 13.398-52.97 5.431-66.404-17.803ZM23.549 85.38a48.499 48.499 0 0 1 25.58-21.333v61.39a8.288 8.288 0 0 0 4.195 7.316l62.874 36.272-21.845 12.636a.819.819 0 0 1-.767 0L41.353 151.53c-23.211-13.454-31.171-43.144-17.804-66.405v.256Zm179.466 41.695-63.08-36.63L161.73 77.86a.819.819 0 0 1 .768 0l52.233 30.184a48.6 48.6 0 0 1-7.316 87.635v-61.391a8.544 8.544 0 0 0-4.4-7.213Zm21.742-32.69-1.535-.922-51.619-30.081a8.39 8.39 0 0 0-8.492 0L99.98 99.808V74.587a.716.716 0 0 1 .307-.665l52.233-30.133a48.652 48.652 0 0 1 72.236 50.391v.205ZM88.061 139.097l-21.845-12.585a.87.87 0 0 1-.41-.614V65.685a48.652 48.652 0 0 1 79.757-37.346l-1.535.87-51.67 29.825a8.595 8.595 0 0 0-4.246 7.367l-.051 72.697Zm11.868-25.58 28.138-16.217 28.188 16.218v32.434l-28.086 16.218-28.188-16.218-.052-32.434Z"/></svg>';
+
+/**
+ * Vendor display name → mark. Vendors without a known mark are absent;
+ * labels for those series render exactly as before, with no icon.
+ */
+export const VENDOR_LOGO_ICONS: Record<string, VendorLogoIcon> = {
+  NVIDIA: { href: svgDataUri(NVIDIA_LOGO_SVG), width: 10, height: 10 },
+  AMD: { href: svgDataUri(AMD_LOGO_SVG), width: 10, height: 10 },
+  // Jalapeño (Teacup) is OpenAI silicon — it carries the OpenAI mark.
+  Teacup: { href: svgDataUri(OPENAI_LOGO_SVG), width: 10, height: 10 },
+};
+
+/**
+ * Full-color vendor mark for a hardware key (`gb200`, `mi355x_dsv4`, ...).
+ * Uses the same base-key convention as `quickFilters`: everything before the
+ * first `_` is the registry key.
+ */
+export function getLineLabelVendorIcon(hwKey: string): VendorLogoIcon | undefined {
+  const vendor = GPU_VENDORS[hwKey.split('_')[0]];
+  return vendor ? VENDOR_LOGO_ICONS[vendor] : undefined;
+}
diff --git a/packages/app/src/lib/zh-copy-mechanical-regressions.jsonl b/packages/app/src/lib/zh-copy-mechanical-regressions.jsonl
new file mode 100644
index 000000000..42c02153e
--- /dev/null
+++ b/packages/app/src/lib/zh-copy-mechanical-regressions.jsonl
@@ -0,0 +1,19 @@
+{"rule": "chip-untranslated", "kind": "regression", "before": "- 在目标延迟下，每颗 Chip 能提供多少 output 吞吐量？", "after": "- 在目标延迟下，每颗芯片能提供多少 output 吞吐量？", "note": "blog prose, content/blog/zh/agentic-benchmark-agent-benchmark-guide.mdx"}
+{"rule": "chip-untranslated", "kind": "regression", "before": "'分离式配置按解码 Chip 或预填充 Chip 报告吞吐量，而非按 Chip 总数。'", "after": "'分离式配置按解码芯片或预填充芯片报告吞吐量，而非按芯片总数。'", "note": "disagg caveat banner; three occurrences on one line"}
+{"rule": "chip-untranslated", "kind": "exemption", "text": "`${i.aLabel} 吞吐量为 ${i.aValue.toFixed(0)} tok/s/chip，每百万 token 成本 …`", "note": "unit — AGENTS.md rule 6 keeps units in English"}
+{"rule": "chip-untranslated", "kind": "exemption", "text": "'吞吐量（token/s/chip）'", "note": "expanded spelling of the same unit also stays English"}
+{"rule": "chip-untranslated", "kind": "exemption", "text": "'$/M = TCO（$/chip-hour）× 1,000,000 /（3600 × tok/s/chip）'", "note": "unit inside a formula"}
+{"rule": "chip-untranslated", "kind": "exemption", "text": "const powerLabel = locale === 'zh' ? '全含功率/芯片：' : 'All in Power/Chip:';", "note": "the English sibling on a bilingual ternary must not be rewritten — this exact line was broken once by a line-level scan, which is why segmentation is per string literal"}
+{"rule": "chip-untranslated", "kind": "exemption", "text": "今年 6 月，[OpenAI 公布了这项芯片计划](https://openai.com/index/openai-broadcom-jalapeno-inference-chip/)，与 Broadcom 合作。", "note": "URL path — the word sits in a link target, and rewriting it would break the link"}
+{"rule": "hardcoded-english-label", "kind": "regression", "before": "`<div><strong>Concurrency:</strong> ${d.concurrency}</div>`", "after": "`<div><strong>${t.concurrency}：</strong> ${d.concurrency}</div>`", "note": "chart tooltip shipped 16 of these; a /zh reader saw them beside 投机解码：关闭"}
+{"rule": "hardcoded-english-label", "kind": "exemption", "text": "`<div><strong>TP:</strong> ${d.tp}</div>`", "note": "acronym — stays English"}
+{"rule": "hardcoded-english-label", "kind": "exemption", "text": "`<div><strong>tok/s/MW:</strong> ${d.perMw}</div>`", "note": "unit — stays English"}
+{"rule": "duplicated-technical-loanword", "kind": "regression", "before": "'正式测量前通过 warmup 预热深层 prefix。'", "after": "'正式测量前通过 warmup 建立深层 prefix。'", "note": "warmup was immediately translated a second time in AgentX FAQ copy"}
+{"rule": "duplicated-technical-loanword", "kind": "regression", "before": "'固定 seed 随机种子后可以复现采样。'", "after": "'固定 seed 后可以复现采样。'", "note": "seed should not be followed by a redundant Chinese translation"}
+{"rule": "duplicated-technical-loanword", "kind": "regression", "before": "'启用 KV offload 卸载后会显示后端。'", "after": "'启用 KV offload 后会显示后端。'", "note": "offload should not be followed by a redundant Chinese translation"}
+{"rule": "duplicated-technical-loanword", "kind": "regression", "before": "'warmup（预热）发生在正式测量之前。'", "after": "'warmup 发生在正式测量之前。'", "note": "English-first full-width parenthetical translation is still redundant"}
+{"rule": "duplicated-technical-loanword", "kind": "regression", "before": "'固定 seed (随机种子) 后可以复现采样。'", "after": "'固定 seed 后可以复现采样。'", "note": "English-first ASCII parenthetical translation is still redundant"}
+{"rule": "duplicated-technical-loanword", "kind": "regression", "before": "'KV offload（卸载）会显示后端。'", "after": "'KV offload 会显示后端。'", "note": "offload followed by a parenthetical Chinese translation is redundant"}
+{"rule": "duplicated-technical-loanword", "kind": "exemption", "text": "'这里所说的预热（warmup）发生在正式测量之前。'", "note": "Chinese-first parenthetical explanation is intentional for a broader audience"}
+{"rule": "duplicated-technical-loanword", "kind": "exemption", "text": "'随机种子（seed）用于固定会话采样。'", "note": "Chinese-first parenthetical explanation remains allowed"}
+{"rule": "duplicated-technical-loanword", "kind": "exemption", "text": "'将 KV cache 卸载（offload）到 DRAM。'", "note": "Chinese-first parenthetical explanation remains allowed"}
diff --git a/packages/app/src/lib/zh-copy.test.ts b/packages/app/src/lib/zh-copy.test.ts
new file mode 100644
index 000000000..87abbf8a0
--- /dev/null
+++ b/packages/app/src/lib/zh-copy.test.ts
@@ -0,0 +1,271 @@
+/**
+ * Guard for the hand-authored Simplified Chinese copy.
+ *
+ * The /zh side has no i18n framework — every Chinese string is written by hand
+ * next to its English sibling, so a site-wide refactor can rewrite Chinese prose
+ * while presenting itself as a rename. PR #668 did exactly that.
+ *
+ * Every rule here decides a question that has one right answer without choosing
+ * a tone, register, or context-dependent rendering. Those editorial decisions
+ * belong to human review and to the advisory `review-zh-copy` skill.
+ */
+import fs from 'node:fs';
+import path from 'node:path';
+
+import { describe, expect, it } from 'vitest';
+
+const APP_DIR = path.resolve(import.meta.dirname, '..', '..');
+const SCAN_ROOTS = [path.join(APP_DIR, 'src'), path.join(APP_DIR, 'content')];
+const SCAN_EXTENSIONS = new Set(['.ts', '.tsx', '.mdx', '.json']);
+
+const HAN = String.raw`\p{Script=Han}`;
+const hasHan = new RegExp(HAN, 'u');
+
+/** This file quotes every banned span it looks for, so it must not scan itself. */
+const SELF = 'src/lib/zh-copy.test.ts';
+
+interface Line {
+  file: string;
+  line: number;
+  text: string;
+}
+
+interface Violation {
+  file: string;
+  line: number;
+  span: string;
+  fix: string;
+}
+
+/**
+ * A mechanical rule. `find` takes one Chinese segment and returns the offending
+ * spans in it, so the same function serves the tree scan and the fixture cases
+ * below — a rule cannot pass its regression case and then behave differently on
+ * real source.
+ */
+interface Rule {
+  id: string;
+  fix: string;
+  /**
+   * `chinese` rules read one Han-carrying segment at a time. `bilingual-file`
+   * rules read raw lines of any file that carries a `zh:` dictionary, because
+   * the spans they look for are English and sit inside template literals whose
+   * own line has no Han at all.
+   */
+  corpus: 'chinese' | 'bilingual-file';
+  find: (text: string) => string[];
+}
+
+function matcher(pattern: RegExp, ignore?: RegExp): (text: string) => string[] {
+  return (text) => {
+    const probe = ignore ? text.replaceAll(ignore, '') : text;
+    return [...probe.matchAll(pattern)].map((match) => match[0].trim());
+  };
+}
+
+/** Quoted spans, so a `locale === 'zh' ? '…' : '…'` line yields two segments. */
+const STRING_LITERAL = /'(?:[^'\\\n]|\\.)*'|"(?:[^"\\\n]|\\.)*"|`(?:[^`\\]|\\.)*`/gsu;
+
+/**
+ * Split a line into the pieces a rule may look at. Code is split literal by
+ * literal so a rule never sees the English half of a bilingual ternary — a
+ * line-level scan once rewrote `'All in Power/Chip:'` while translating the
+ * Chinese beside it. Prose has no literals, so it is taken whole.
+ */
+function segment(raw: string, isProse: boolean): string[] {
+  if (!hasHan.test(raw)) return [];
+  if (isProse) return [raw];
+  const literals = (raw.match(STRING_LITERAL) ?? []).filter((text) => hasHan.test(text));
+  // JSX text and template-literal continuations carry no quotes of their own.
+  return literals.length > 0 ? literals : [raw];
+}
+
+// "GPU" is a naturalized loanword that Chinese technical writing uses verbatim.
+// "Chip" is an ordinary English noun whose Chinese equivalent (芯片) is what
+// readers actually use, so it cannot inherit GPU's exemption. Units keep the
+// English form per AGENTS.md rule 6 — that is the one recorded exception.
+const CHIP_UNITS = /(?:tok|tokens?)\/s\/chip|\$\/chip[/-](?:hr|hour)|[A-Za-z]Chip\b|\bChip[A-Z]/giu;
+
+// A URL is an address, not prose. `…/openai-broadcom-jalapeno-inference-chip/`
+// carries the word without stating any translation decision, and rewriting it
+// would break the link, so URLs leave the corpus before any rule reads it.
+const URL_LIKE = /https?:\/\/\S+/giu;
+
+const ignoring = (...patterns: RegExp[]) =>
+  new RegExp(patterns.map((pattern) => pattern.source).join('|'), 'giu');
+
+const RULES: Rule[] = [
+  {
+    id: 'chip-untranslated',
+    corpus: 'chinese',
+    fix: '写作「芯片」（单位 tok/s/chip 等除外）',
+    find: matcher(/\b[Cc]hip\b/gu, ignoring(URL_LIKE, CHIP_UNITS)),
+  },
+  {
+    id: 'hardcoded-english-label',
+    corpus: 'bilingual-file',
+    fix: '把标签接入本地化字典，勿硬编码英文',
+    // Acronyms (TP, EP, DPA) and units (tok/s/MW) stay English per AGENTS.md
+    // rule 6, so a label is only suspect if it reads as prose.
+    find: (text) =>
+      [...text.matchAll(/<strong>(?<label>[A-Za-z][^<${}]{2,45}):<\/strong>/gu)]
+        .filter(({ groups }) => {
+          const label = groups?.label ?? '';
+          return !/\//u.test(label) && !/^[A-Z0-9]+$/u.test(label);
+        })
+        .map((match) => match[0]),
+  },
+  {
+    id: 'duplicated-technical-loanword',
+    corpus: 'chinese',
+    fix: '保留技术借词，删除紧随其后的重复中文翻译',
+    find: matcher(
+      /\b(?:warmup(?:\s+预热|\s*[（(]\s*预热\s*[）)])|seed(?:\s+随机种子|\s*[（(]\s*随机种子\s*[）)])|offload(?:\s+卸载|\s*[（(]\s*卸载\s*[）)]))/giu,
+    ),
+  },
+];
+
+function walk(directory: string): string[] {
+  return fs
+    .readdirSync(directory, { withFileTypes: true })
+    .flatMap((entry) => {
+      const full = path.join(directory, entry.name);
+      if (entry.isDirectory()) return entry.name === 'node_modules' ? [] : walk(full);
+      return SCAN_EXTENSIONS.has(path.extname(entry.name)) ? [full] : [];
+    })
+    .sort();
+}
+
+const sourceFiles: string[] = SCAN_ROOTS.filter((root) => fs.existsSync(root)).flatMap(walk);
+
+/**
+ * Han-carrying prose, split finely enough that a rule never sees an English
+ * sibling as context. JSON-LD is dropped: structured data is JSON, so its quotes
+ * and colons are syntax and "fixing" them breaks the schema.org payload.
+ */
+function chineseLinesFromSource(relative: string, source: string): Line[] {
+  const isProse = path.extname(relative) === '.mdx';
+  let inJsonLd = false;
+  return source.split('\n').flatMap((raw, index) => {
+    if (isProse) {
+      const trimmed = raw.trim();
+      if (trimmed.startsWith('<JsonLd') || trimmed.startsWith('{{')) inJsonLd = true;
+      const skip = inJsonLd;
+      if (inJsonLd && (trimmed.endsWith('/>') || trimmed.endsWith('</JsonLd>') || trimmed === '}}'))
+        inJsonLd = false;
+      if (skip) return [];
+    }
+    return segment(raw, isProse).map((text) => ({ file: relative, line: index + 1, text }));
+  });
+}
+
+const chineseLines: Line[] = sourceFiles.flatMap((file) => {
+  const relative = path.relative(APP_DIR, file);
+  if (relative === SELF) return [];
+  return chineseLinesFromSource(relative, fs.readFileSync(file, 'utf8'));
+});
+
+function report(violations: Violation[]): string {
+  return violations
+    .slice(0, 40)
+    .map((v) => `  ${v.file}:${v.line}  «${v.span}»  → ${v.fix}`)
+    .join('\n');
+}
+
+function expectClean(violations: Violation[]): void {
+  expect(violations.length, `\n${report(violations)}\n`).toBe(0);
+}
+
+describe('zh copy — the tree obeys every mechanical rule', () => {
+  const bilingualLines: Line[] = sourceFiles.flatMap((file) => {
+    const relative = path.relative(APP_DIR, file);
+    if (relative === SELF || relative.includes('.test.')) return [];
+    const source = fs.readFileSync(file, 'utf8');
+    if (!/\bzh:\s*\{/u.test(source)) return [];
+    return source.split('\n').map((text, index) => ({ file: relative, line: index + 1, text }));
+  });
+
+  it.each(RULES.map((rule) => [rule.id, rule] as const))('%s', (_id, rule) => {
+    const corpus = rule.corpus === 'bilingual-file' ? bilingualLines : chineseLines;
+    expectClean(
+      corpus.flatMap(({ file, line, text }) =>
+        rule.find(text).map((span) => ({ file, line, span, fix: rule.fix })),
+      ),
+    );
+  });
+});
+
+describe('zh copy — every rule still catches what it was written for', () => {
+  // Real spans this repo shipped, trimmed to the offending clause. Each row is
+  // hand-checked; the file is deliberately small, because a fixture nobody has
+  // read is not evidence. `.jsonl` is outside SCAN_EXTENSIONS, so the banned
+  // spans quoted here are never scanned as source.
+  interface Fixture {
+    rule: string;
+    kind: 'regression' | 'exemption';
+    before?: string;
+    after?: string;
+    text?: string;
+    note: string;
+  }
+
+  const fixtures: Fixture[] = fs
+    .readFileSync(path.join(APP_DIR, 'src/lib/zh-copy-mechanical-regressions.jsonl'), 'utf8')
+    .split('\n')
+    .filter(Boolean)
+    .map((line) => JSON.parse(line) as Fixture);
+
+  const byId = new Map(RULES.map((rule) => [rule.id, rule]));
+  /** Fixture text is a source line, so it enters the rule exactly as the tree does. */
+  const run = (rule: Rule, text: string) =>
+    rule.corpus === 'bilingual-file' ? rule.find(text) : segment(text, false).flatMap(rule.find);
+
+  it('covers every rule', () => {
+    const covered = new Set(fixtures.filter((f) => f.kind === 'regression').map((f) => f.rule));
+    expect([...byId.keys()].filter((id) => !covered.has(id))).toEqual([]);
+  });
+
+  it.each(fixtures.map((f) => [`${f.rule} — ${f.kind} — ${f.note}`, f] as const))(
+    '%s',
+    (_name, fixture) => {
+      const rule = byId.get(fixture.rule);
+      expect(rule, `unknown rule id "${fixture.rule}"`).toBeDefined();
+      if (!rule) return;
+
+      if (fixture.kind === 'exemption') {
+        expect(run(rule, fixture.text ?? '')).toEqual([]);
+        return;
+      }
+      // Catching `before` is what proves the rule works; passing `after` is what
+      // proves the accepted fix actually clears it.
+      expect(run(rule, fixture.before ?? '').length).toBeGreaterThan(0);
+      expect(run(rule, fixture.after ?? '')).toEqual([]);
+    },
+  );
+});
+
+describe('zh copy — coverage', () => {
+  it('resumes scanning Chinese prose after an MDX JsonLd block', () => {
+    const lines = chineseLinesFromSource(
+      'content/blog/zh/example.mdx',
+      [
+        'JSON-LD 之前的中文。',
+        '<JsonLd>{`{',
+        '  "description": "结构化数据中的 Chip 不应扫描"',
+        '}`}</JsonLd>',
+        'JSON-LD 之后的中文 Chip 应继续扫描。',
+      ].join('\n'),
+    );
+
+    expect(lines.map(({ line, text }) => ({ line, text }))).toEqual([
+      { line: 1, text: 'JSON-LD 之前的中文。' },
+      { line: 5, text: 'JSON-LD 之后的中文 Chip 应继续扫描。' },
+    ]);
+  });
+
+  it('scans a plausible amount of Chinese copy', () => {
+    // A refactor that guts the /zh tree should fail loudly rather than let every
+    // rule above pass vacuously.
+    expect(chineseLines.length).toBeGreaterThan(1000);
+  });
+});
diff --git a/packages/app/src/test/setup-local-storage.ts b/packages/app/src/test/setup-local-storage.ts
index c65de976c..77bf3634a 100644
--- a/packages/app/src/test/setup-local-storage.ts
+++ b/packages/app/src/test/setup-local-storage.ts
@@ -38,3 +38,24 @@ Object.defineProperty(globalThis, 'localStorage', {
   configurable: true,
   value: new MemoryStorage(),
 });
+
+// jsdom ships no `matchMedia`. Components that gate rendering on a breakpoint
+// need it to exist; default to "matches" so a component under test behaves as
+// it does on a desktop viewport unless a test overrides the stub.
+if (globalThis.window !== undefined && !globalThis.window.matchMedia) {
+  Object.defineProperty(globalThis.window, 'matchMedia', {
+    configurable: true,
+    writable: true,
+    value: (query: string): MediaQueryList =>
+      ({
+        matches: true,
+        media: query,
+        onchange: null,
+        addEventListener: () => {},
+        removeEventListener: () => {},
+        addListener: () => {},
+        removeListener: () => {},
+        dispatchEvent: () => false,
+      }) as unknown as MediaQueryList,
+  });
+}
diff --git a/packages/app/timings.json b/packages/app/timings.json
index 29003b6b4..e9236e6a3 100644
--- a/packages/app/timings.json
+++ b/packages/app/timings.json
@@ -1,156 +1,248 @@
 {
   "durations": [
+    {
+      "spec": "cypress/e2e/agentic-catalog.cy.ts",
+      "duration": 446
+    },
+    {
+      "spec": "cypress/e2e/agentic-detail-back-nav.cy.ts",
+      "duration": 2000
+    },
+    {
+      "spec": "cypress/e2e/agentic-point-coach-mark.cy.ts",
+      "duration": 8000
+    },
     {
       "spec": "cypress/e2e/agentic-point-time-series.cy.ts",
-      "duration": 2222
+      "duration": 1961
+    },
+    {
+      "spec": "cypress/e2e/agentic-quick-filters.cy.ts",
+      "duration": 1000
+    },
+    {
+      "spec": "cypress/e2e/agentx-optimizations.cy.ts",
+      "duration": 775
+    },
+    {
+      "spec": "cypress/e2e/agentx-telemetry.cy.ts",
+      "duration": 14767
+    },
+    {
+      "spec": "cypress/e2e/ai-chart.cy.ts",
+      "duration": 3000
+    },
+    {
+      "spec": "cypress/e2e/api-documentation.cy.ts",
+      "duration": 310
     },
     {
       "spec": "cypress/e2e/blog.cy.ts",
-      "duration": 1052
+      "duration": 1537
+    },
+    {
+      "spec": "cypress/e2e/fleet-lifecycle.cy.ts",
+      "duration": 18000
+    },
+    {
+      "spec": "cypress/e2e/calculator-overlay.cy.ts",
+      "duration": 5318
+    },
+    {
+      "spec": "cypress/e2e/chart-overflow-continuation.cy.ts",
+      "duration": 731
+    },
+    {
+      "spec": "cypress/e2e/collectivex.cy.ts",
+      "duration": 8913
     },
     {
       "spec": "cypress/e2e/compare-per-dollar-redirect.cy.ts",
-      "duration": 2471
+      "duration": 1379
     },
     {
       "spec": "cypress/e2e/compare-per-dollar-table.cy.ts",
-      "duration": 657
+      "duration": 397
     },
     {
       "spec": "cypress/e2e/compare-precision.cy.ts",
-      "duration": 1519
+      "duration": 1140
     },
     {
       "spec": "cypress/e2e/compare-redirect.cy.ts",
-      "duration": 1726
+      "duration": 1169
+    },
+    {
+      "spec": "cypress/e2e/compare-scenario-routes.cy.ts",
+      "duration": 1091
     },
     {
       "spec": "cypress/e2e/compare-spec-decode.cy.ts",
-      "duration": 1123
+      "duration": 657
     },
     {
       "spec": "cypress/e2e/compare-table.cy.ts",
-      "duration": 1460
+      "duration": 1163
+    },
+    {
+      "spec": "cypress/e2e/csv-export-overlay.cy.ts",
+      "duration": 730
     },
     {
       "spec": "cypress/e2e/csv-export.cy.ts",
-      "duration": 1883
+      "duration": 1492
     },
     {
       "spec": "cypress/e2e/custom-user-values.cy.ts",
-      "duration": 12004
+      "duration": 12254
     },
     {
       "spec": "cypress/e2e/datasets-distributions.cy.ts",
-      "duration": 387
+      "duration": 272
     },
     {
       "spec": "cypress/e2e/datasets-flamegraph-time.cy.ts",
-      "duration": 336
+      "duration": 245
     },
     {
-      "spec": "cypress/e2e/drill-down-trend.cy.ts",
-      "duration": 4647
+      "spec": "cypress/e2e/datasets-methodology.cy.ts",
+      "duration": 439
     },
     {
       "spec": "cypress/e2e/dropdown-switching.cy.ts",
-      "duration": 1824
+      "duration": 1328
     },
     {
       "spec": "cypress/e2e/evaluation-chart.cy.ts",
-      "duration": 1299
+      "duration": 2496
+    },
+    {
+      "spec": "cypress/e2e/fixed-sequence-logs.cy.ts",
+      "duration": 1000
     },
     {
       "spec": "cypress/e2e/gpu-compare-agentic-detail.cy.ts",
-      "duration": 1310
+      "duration": 1104
     },
     {
       "spec": "cypress/e2e/gpu-power.cy.ts",
-      "duration": 4516
+      "duration": 3036
     },
     {
       "spec": "cypress/e2e/gpu-specs.cy.ts",
-      "duration": 5642
+      "duration": 4711
     },
     {
       "spec": "cypress/e2e/gradient-labels.cy.ts",
-      "duration": 3702
+      "duration": 2875
     },
     {
       "spec": "cypress/e2e/historical-trends.cy.ts",
-      "duration": 5761
+      "duration": 3559
     },
     {
       "spec": "cypress/e2e/inference-chart.cy.ts",
-      "duration": 865
+      "duration": 565
     },
     {
       "spec": "cypress/e2e/inference-replay.cy.ts",
-      "duration": 3742
+      "duration": 8153
+    },
+    {
+      "spec": "cypress/e2e/jalapeno-preview-notice.cy.ts",
+      "duration": 36000
     },
     {
       "spec": "cypress/e2e/landing-performance.cy.ts",
-      "duration": 5744
+      "duration": 5609
     },
     {
       "spec": "cypress/e2e/line-labels.cy.ts",
-      "duration": 9171
+      "duration": 8476
     },
     {
       "spec": "cypress/e2e/measured-power-overlay.cy.ts",
-      "duration": 1723
+      "duration": 1468
     },
     {
       "spec": "cypress/e2e/model-architecture.cy.ts",
-      "duration": 5340
+      "duration": 4570
+    },
+    {
+      "spec": "cypress/e2e/model-route-slugs.cy.ts",
+      "duration": 2000
     },
     {
       "spec": "cypress/e2e/navigation.cy.ts",
-      "duration": 3296
+      "duration": 3442
     },
     {
       "spec": "cypress/e2e/nudge-system.cy.ts",
-      "duration": 26875
+      "duration": 25075
+    },
+    {
+      "spec": "cypress/e2e/overlay-legend-remove.cy.ts",
+      "duration": 810
+    },
+    {
+      "spec": "cypress/e2e/overlay-optimal-only.cy.ts",
+      "duration": 482
+    },
+    {
+      "spec": "cypress/e2e/overview-a11y.cy.ts",
+      "duration": 2000
+    },
+    {
+      "spec": "cypress/e2e/overview.cy.ts",
+      "duration": 9977
+    },
+    {
+      "spec": "cypress/e2e/performance-over-time-removal.cy.ts",
+      "duration": 564
     },
     {
       "spec": "cypress/e2e/performance.cy.ts",
-      "duration": 1733
+      "duration": 1424
     },
     {
       "spec": "cypress/e2e/reliability-chart.cy.ts",
-      "duration": 1724
+      "duration": 1504
     },
     {
-      "spec": "cypress/e2e/sanity.cy.ts",
-      "duration": 1042
+      "spec": "cypress/e2e/resident-sequence-lengths.cy.ts",
+      "duration": 1000
     },
     {
-      "spec": "cypress/e2e/speed-overlay.cy.ts",
-      "duration": 3664
+      "spec": "cypress/e2e/sanity.cy.ts",
+      "duration": 715
     },
     {
       "spec": "cypress/e2e/throughput-calculator.cy.ts",
-      "duration": 8551
+      "duration": 15093
+    },
+    {
+      "spec": "cypress/e2e/tokens-per-currency.cy.ts",
+      "duration": 651
     },
     {
       "spec": "cypress/e2e/ttft-x-axis-toggle.cy.ts",
-      "duration": 2592
+      "duration": 3778
     },
     {
       "spec": "cypress/e2e/unofficial-watermark.cy.ts",
-      "duration": 1304
+      "duration": 923
     },
     {
       "spec": "cypress/e2e/url-params.cy.ts",
-      "duration": 7335
+      "duration": 7000
     },
     {
       "spec": "cypress/e2e/yaxis-metrics-render.cy.ts",
-      "duration": 2990
+      "duration": 3894
     },
     {
       "spec": "cypress/e2e/zh-pages.cy.ts",
-      "duration": 1562
+      "duration": 881
     }
   ]
 }
diff --git a/packages/app/tsconfig.json b/packages/app/tsconfig.json
index 3044b60ce..0ac2c3953 100644
--- a/packages/app/tsconfig.json
+++ b/packages/app/tsconfig.json
@@ -1,16 +1,9 @@
 {
+  "extends": "../../tsconfig.base.json",
   "compilerOptions": {
     "target": "ES2018",
     "lib": ["dom", "dom.iterable", "esnext"],
     "allowJs": true,
-    "skipLibCheck": true,
-    "strict": true,
-    "noEmit": true,
-    "esModuleInterop": true,
-    "module": "esnext",
-    "moduleResolution": "bundler",
-    "resolveJsonModule": true,
-    "isolatedModules": true,
     "jsx": "react-jsx",
     "incremental": true,
     "plugins": [
diff --git a/packages/constants/src/currency.ts b/packages/constants/src/currency.ts
new file mode 100644
index 000000000..f6ccb3b42
--- /dev/null
+++ b/packages/constants/src/currency.ts
@@ -0,0 +1,20 @@
+/**
+ * Pinned FX rate for the tokens-per-currency chart metrics.
+ *
+ * Every cost in `HW_REGISTRY` (`costh`, `costn`, `costr`) is $/GPU/hr from the
+ * SemiAnalysis AI Cloud TCO Model, so the ¥ metrics are the $ metrics divided
+ * by this rate.
+ *
+ * Deliberately a constant rather than a live feed: these charts compare
+ * hardware, and a floating rate would make a chip's plotted ¥ value move on
+ * days when nothing about the chip or the benchmark changed. Two runs of the
+ * same recipe a week apart must land on the same point. Refresh it on a normal
+ * cadence with the TCO figures it sits beside, and note the date here.
+ *
+ * Source: 1 USD = 6.7240 CNY on 2026-08-20, rounded to 6.72.
+ * https://www.exchange-rates.org/exchange-rate-history/usd-cny-2026
+ */
+export const USD_TO_CNY = 6.72;
+
+/** Date the rate above was last checked, surfaced in the chart disclaimer. */
+export const USD_TO_CNY_AS_OF = '2026-08-20';
diff --git a/packages/constants/src/framework-aliases.test.ts b/packages/constants/src/framework-aliases.test.ts
index 3acee7b0d..469ef3cca 100644
--- a/packages/constants/src/framework-aliases.test.ts
+++ b/packages/constants/src/framework-aliases.test.ts
@@ -31,6 +31,10 @@ describe('MODEL_SPEC_METHOD_LABELS', () => {
   it('maps MiniMax-M3 mtp to "M3 EAGLE"', () => {
     expect(MODEL_SPEC_METHOD_LABELS['MiniMax-M3']?.mtp).toBe('M3 EAGLE');
   });
+
+  it('maps Kimi-K3 mtp to "DSpark"', () => {
+    expect(MODEL_SPEC_METHOD_LABELS['Kimi-K3']?.mtp).toBe('DSpark');
+  });
 });
 
 describe('resolveFrameworkPartLabel', () => {
@@ -38,6 +42,10 @@ describe('resolveFrameworkPartLabel', () => {
     expect(resolveFrameworkPartLabel('MiniMax-M3', 'mtp')).toBe('M3 EAGLE');
   });
 
+  it('renders K3 mtp as "DSpark"', () => {
+    expect(resolveFrameworkPartLabel('Kimi-K3', 'mtp')).toBe('DSpark');
+  });
+
   it('keeps the generic MTP label for other models', () => {
     expect(resolveFrameworkPartLabel('DeepSeek-R1-0528', 'mtp')).toBe('MTP');
   });
@@ -46,6 +54,10 @@ describe('resolveFrameworkPartLabel', () => {
     expect(resolveFrameworkPartLabel(undefined, 'mtp')).toBe('MTP');
   });
 
+  it('keeps the generic MTP label for sibling Kimi models', () => {
+    expect(resolveFrameworkPartLabel('Kimi-K2.5', 'mtp')).toBe('MTP');
+  });
+
   it('falls back to FRAMEWORK_LABELS for non-overridden parts even for M3', () => {
     expect(resolveFrameworkPartLabel('MiniMax-M3', 'vllm')).toBe('vLLM');
   });
diff --git a/packages/constants/src/framework-aliases.ts b/packages/constants/src/framework-aliases.ts
index 076f5cb2e..9d139b9c9 100644
--- a/packages/constants/src/framework-aliases.ts
+++ b/packages/constants/src/framework-aliases.ts
@@ -6,14 +6,17 @@ export interface FwEntry {
 /** Single source of truth for framework metadata. Add new frameworks here. */
 export const FW_REGISTRY: Record<string, FwEntry> = {
   atom: { label: 'ATOM¹' },
+  'coreweave-vera-rubin': { label: 'Vera Rubin' },
   'dynamo-sglang': { label: 'Dynamo SGLang' },
   'dynamo-trt': { label: 'Dynamo TRTLLM' },
   'dynamo-vllm': { label: 'Dynamo vLLM' },
   'llmd-vllm': { label: 'llm-d vLLM' },
   'mooncake-atom': { label: 'Mooncake ATOMesh¹' },
   'mori-sglang': { label: 'MoRI SGLang' },
+  'rubin-july': { label: 'July' },
   sglang: { label: 'SGLang' },
   tilert: { label: 'TileRT' },
+  teacup: { label: 'Teacup' },
   trt: { label: 'TRTLLM' },
   vllm: { label: 'vLLM' },
 };
@@ -55,12 +58,13 @@ export const FRAMEWORK_LABELS: Record<string, string> = {
  * Per-model display overrides for hwKey suffix parts (framework / spec_method
  * tokens), keyed by frontend display model name → token → label.
  *
- * M3's speculative-decoding runs are ingested under the generic `mtp` token but
- * actually use EAGLE, so for MiniMax-M3 the suffix reads "EAGLE" while every
- * other model keeps the generic "MTP" label.
+ * Some models are ingested under the generic `mtp` token but ship a
+ * differently-named method: M3 uses EAGLE, and K3 uses DSpark. Those models get
+ * their own label while every other model keeps the generic "MTP" label.
  */
 export const MODEL_SPEC_METHOD_LABELS: Record<string, Record<string, string>> = {
   'MiniMax-M3': { mtp: 'M3 EAGLE' },
+  'Kimi-K3': { mtp: 'DSpark' },
 };
 
 /**
diff --git a/packages/constants/src/gpu-keys.ts b/packages/constants/src/gpu-keys.ts
index 8837dca0b..a3e8dc96f 100644
--- a/packages/constants/src/gpu-keys.ts
+++ b/packages/constants/src/gpu-keys.ts
@@ -5,6 +5,8 @@ export interface HwEntry {
   arch: string;
   /** Display label (e.g. "H100", "GB200 NVL72") */
   label: string;
+  /** Optional compact label for power/TCO badges. Defaults to the uppercase key. */
+  badgeLabel?: string;
   /** Chart sort order (lower = first) */
   sort: number;
   /** Thermal design power in watts */
@@ -22,6 +24,17 @@ export interface HwEntry {
 
 /** Single source of truth for GPU metadata. Add new GPUs here. */
 export const HW_REGISTRY: Record<string, HwEntry> = {
+  vr200: {
+    vendor: 'NVIDIA',
+    arch: 'Vera Rubin',
+    label: 'Vera Rubin',
+    sort: -1,
+    tdp: 1800,
+    power: 3.3,
+    costh: 3.61,
+    costn: 3.61,
+    costr: 3.61,
+  },
   h100: {
     vendor: 'NVIDIA',
     arch: 'Hopper',
@@ -136,6 +149,18 @@ export const HW_REGISTRY: Record<string, HwEntry> = {
     costn: 0.75,
     costr: 0.52,
   },
+  jalapeno: {
+    vendor: 'Teacup',
+    arch: 'Jalapeño',
+    label: 'Jalapeño',
+    badgeLabel: 'Jalapeño (Teacup)',
+    sort: 10,
+    tdp: 700,
+    power: 1.125,
+    costh: 1.47,
+    costn: 1.56,
+    costr: 1.79,
+  },
   // NVIDIA GeForce RTX 5090 (Blackwell, consumer, 32 GB GDDR7). Test-rig
   // hardware, not a fleet GPU -- no hyperscaler offers it and there's no real
   // all-in-power basis for a standalone rig, so costh/power are explicit 9.99
@@ -151,7 +176,7 @@ export const HW_REGISTRY: Record<string, HwEntry> = {
     vendor: 'NVIDIA',
     arch: 'Blackwell (consumer)',
     label: 'RTX 5090',
-    sort: 10,
+    sort: 11,
     tdp: 575,
     power: 9.99,
     costh: 9.99,
@@ -194,6 +219,7 @@ export const VENDOR_OKLCH_ZONES: Record<
 > = {
   amd: { start: 12, end: 42, chroma: { light: 0.18, dark: 0.22 } },
   nvidia: { start: 120, end: 170, chroma: { light: 0.15, dark: 0.15 } },
+  teacup: { start: 235, end: 270, chroma: { light: 0.14, dark: 0.16 } },
   unknown: { start: 275, end: 330, chroma: { light: 0.14, dark: 0.16 } },
 };
 
@@ -208,16 +234,18 @@ export const VENDOR_OKLCH_ZONES: Record<
  * Layout (360° wheel):
  *   NVIDIA:  60–195  (135°) — greens through cyans
  *   AMD:     300–360 + 0–60  (120°, wraps) — magentas through oranges
- *   unknown: 195–300 (105°) — blues/purples
+ *   Teacup:  195–240 (45°) — cyan/blues
+ *   unknown: 240–300 (60°) — blues/purples
  *
  * Each entry is an array of linear {start, span} segments (wrapping bands
  * are split into two segments).
  */
 export const VENDOR_HSL_ZONES: Record<string, { start: number; span: number }[]> = {
   nvidia: [{ start: 60, span: 135 }],
+  teacup: [{ start: 195, span: 45 }],
   amd: [
     { start: 300, span: 60 },
     { start: 0, span: 60 },
   ],
-  unknown: [{ start: 195, span: 105 }],
+  unknown: [{ start: 240, span: 60 }],
 };
diff --git a/packages/constants/src/index.ts b/packages/constants/src/index.ts
index e767e500c..e574f28e3 100644
--- a/packages/constants/src/index.ts
+++ b/packages/constants/src/index.ts
@@ -6,3 +6,6 @@ export * from './precision-keys';
 export * from './models';
 export * from './seo';
 export * from './tables';
+export * from './currency';
+export * from './tco';
+export * from './token-length-sketch';
diff --git a/packages/constants/src/metric-keys.ts b/packages/constants/src/metric-keys.ts
index 94c5b6134..370ea7f50 100644
--- a/packages/constants/src/metric-keys.ts
+++ b/packages/constants/src/metric-keys.ts
@@ -131,26 +131,32 @@ export const METRIC_KEYS = new Set([
   // server_metrics.kv_cache.gpu_usage_pct in v3)
   'gpu_kv_cache_usage_pct',
   // measured power / energy (emitted by runner's aggregate_power.py)
+  // power_valid: numeric 1/0 publication verdict; explicit 0 withholds power
+  // power_metric_schema_version: version 2 defines every unprefixed
+  //                              joules_per_* field as whole-deployment energy
   // avg_power_w:             mean per-GPU draw (W) during the load window
+  // joules_per_successful_query: whole-deployment energy / successful requests
   // joules_per_output_token: energy / total_output_tokens. CLUSTER-WIDE on
-  //                          single-node / non-disagg (total_system_energy);
-  //                          PER-STAGE decode_energy on disagg (decode GPUs only),
-  //                          symmetric with joules_per_input_token below.
+  //                          schema-version-2 rows, including disaggregated runs.
   // joules_per_total_token:  total_system_energy / (total_input + total_output)
   //                          — cluster-wide; workload-shape-fair view that
   //                          doesn't treat prompt as free.
+  'power_valid',
+  'power_metric_schema_version',
   'avg_power_w',
+  'joules_per_successful_query',
   'joules_per_output_token',
   'joules_per_total_token',
   // multinode / disagg role splits (emitted only when the deployment has
   // distinct prefill / decode workers)
   // prefill_avg_power_w / decode_avg_power_w:  mean per-GPU draw within each role
-  // joules_per_input_token:  prefill_energy / total_input_tokens (prefill GPUs only).
-  //   The disagg output counterpart is joules_per_output_token above (decode GPUs
-  //   only) — there is no separate _decode key.
+  // Explicit role-local energy remains separate from the version-2 unprefixed
+  // whole-deployment fields.
   'prefill_avg_power_w',
   'decode_avg_power_w',
   'joules_per_input_token',
+  'prefill_joules_per_input_token',
+  'decode_joules_per_output_token',
   // cluster-wide GPU telemetry beyond power (emitted by aggregate_power.py when
   // the perfmon CSVs include temperature, utilization, or memory samples).
   // avg_temp_c:        mean per-GPU temperature (Celsius) during load window
@@ -171,6 +177,10 @@ export const METRIC_KEYS = new Set([
   // (rowToAggDataEntry in benchmark-transform.ts).
   'prefill_pp',
   'decode_pp',
+  // Aggregate artifacts emit one context-parallel width, while disaggregated
+  // artifacts may emit a separate value for each role.
+  'dcp_size',
+  'pcp_size',
   'prefill_dcp_size',
   'decode_dcp_size',
   'prefill_pcp_size',
diff --git a/packages/constants/src/models.test.ts b/packages/constants/src/models.test.ts
index 34d50270f..3f5a87221 100644
--- a/packages/constants/src/models.test.ts
+++ b/packages/constants/src/models.test.ts
@@ -2,6 +2,8 @@ import { describe, it, expect } from 'vitest';
 import {
   DB_MODEL_TO_DISPLAY,
   DISPLAY_MODEL_TO_DB,
+  MODEL_RELEASE_DATES,
+  getModelReleaseDate,
   sequenceToIslOsl,
   islOslToSequence,
 } from './models';
@@ -36,6 +38,45 @@ describe('DB_MODEL_TO_DISPLAY / DISPLAY_MODEL_TO_DB consistency', () => {
   });
 });
 
+describe('MODEL_RELEASE_DATES', () => {
+  it('stores every date as a real calendar day in YYYY-MM-DD', () => {
+    for (const [model, date] of Object.entries(MODEL_RELEASE_DATES)) {
+      expect(date, model).toMatch(/^\d{4}-\d{2}-\d{2}$/u);
+      // `new Date` happily accepts '2026-02-30' and rolls it into March, so
+      // round-tripping is the only way to catch a day that does not exist.
+      expect(new Date(`${date}T00:00:00Z`).toISOString().slice(0, 10), model).toBe(date);
+    }
+  });
+
+  it('has no model releasing in the future', () => {
+    // A future date is the signature of a typo'd year, and it would put the
+    // Fleet Lifecycle anchor after every sweep it is meant to precede.
+    const today = new Date().toISOString().slice(0, 10);
+    for (const [model, date] of Object.entries(MODEL_RELEASE_DATES)) {
+      expect(
+        date.localeCompare(today),
+        `${model} released ${date}, after today`,
+      ).toBeLessThanOrEqual(0);
+    }
+  });
+
+  it('is keyed by display name, never by DB model key', () => {
+    // Keying by 'dsv4' instead of 'DeepSeek-V4-Pro' does not fail loudly — the
+    // lookup just returns null forever and the caller silently falls back.
+    const dbKeys = new Set(Object.keys(DB_MODEL_TO_DISPLAY));
+    for (const model of Object.keys(MODEL_RELEASE_DATES)) {
+      expect(dbKeys.has(model), `${model} is a DB key, not a display name`).toBe(false);
+    }
+  });
+
+  it('returns null rather than throwing for a model with no sourced date', () => {
+    // Every model shipped today has a date, so this documents the fallback path
+    // that a future model arrives on rather than an existing gap.
+    expect(getModelReleaseDate('not-a-model')).toBeNull();
+    expect(getModelReleaseDate('')).toBeNull();
+  });
+});
+
 describe('sequenceToIslOsl', () => {
   it('parses 1k/1k to 1024/1024', () => {
     expect(sequenceToIslOsl('1k/1k')).toEqual({ isl: 1024, osl: 1024 });
diff --git a/packages/constants/src/models.ts b/packages/constants/src/models.ts
index 9f9f5babb..13818046f 100644
--- a/packages/constants/src/models.ts
+++ b/packages/constants/src/models.ts
@@ -12,6 +12,9 @@ export const DB_MODEL_TO_DISPLAY: Record<string, string> = {
   gptoss120b: 'gpt-oss-120b',
   llama70b: 'Llama-3.3-70B-Instruct-FP8',
   'qwen3.5': 'Qwen-3.5-397B-A17B',
+  // Qwen4-architecture preview, not a Qwen3.5 point release (GatedDeltaNet plus
+  // Qwen Sparse Attention, 512 experts), so it gets its own display bucket.
+  'qwen3.8next': 'Qwen3.8-Flash-Next',
   'kimik2.5': 'Kimi-K2.5',
   'kimik2.6': 'Kimi-K2.5',
   'kimik2.7-code': 'Kimi-K2.5',
@@ -79,3 +82,116 @@ export function rowToSequence(row: {
   if (row.isl === null || row.osl === null) return null;
   return islOslToSequence(row.isl, row.osl);
 }
+
+/**
+ * Model release dates, keyed by frontend display name (YYYY-MM-DD, UTC).
+ *
+ * **This is the only release-date table in the repo, and it must stay that way.**
+ * These dates were per-entry `releaseDate` fields on `MODEL_ARCHITECTURES`, which
+ * is the wrong home for them: a date that captions a diagram is the same fact as
+ * a date that anchors a revenue axis, and two copies of one fact drift. Callers
+ * go through `getModelReleaseDate`.
+ *
+ * **What the date means:** the day the weights became publicly downloadable — not
+ * the announcement, and not the day InferenceX started benchmarking the model.
+ * Inference optimisation cannot begin before the weights are out, so that
+ * publication is the honest zero for any "how far has this come since it
+ * shipped?" axis. Where a display name groups point releases (`GLM-5` covers
+ * both GLM-5 and GLM-5.1), the date is the *earliest* release in the bucket,
+ * because that is when the bucket's weights first existed.
+ *
+ * A date here must never postdate the model's first InferenceX sweep — a model
+ * cannot be benchmarked before it exists. That is the check that catches a wrong
+ * entry, and it is why these are worth sourcing individually rather than
+ * defaulting to whatever `MODELS.md` lists as the date the model was added.
+ *
+ * Only add an entry you can source, and cite the source in a comment.
+ * `getModelReleaseDate` returns null otherwise, so callers fall back to the
+ * earliest date they actually have data for.
+ */
+export const MODEL_RELEASE_DATES: Record<string, string> = {
+  // Newest first. `sweep:` is the model's "Date added" from MODELS.md — the first
+  // InferenceX benchmark — recorded so the "cannot predate its own weights"
+  // invariant above is checkable by eye at review time.
+  //
+  // Weights published on Hugging Face 2026-07-27 under the Kimi K3 License,
+  // eleven days after the 2026-07-16 product launch; the gap was to let vLLM,
+  // NVIDIA and AMD prepare day-zero support. sweep: 2026-07-27 — day zero.
+  'Kimi-K3': '2026-07-27',
+  // Zhipu launched GLM-5.2 on 2026-06-13 via the GLM Coding Plan with MIT
+  // weights at zai-org/GLM-5.2; standalone API and provider support followed
+  // over the next few days. One source dates the Hugging Face upload 2026-06-16
+  // instead, so this is ±3 days — immaterial on a multi-month axis, but it is an
+  // announcement-day date rather than a confirmed upload day. sweep: 2026-07-18.
+  'GLM-5.2': '2026-06-13',
+  // Weights and the official MXFP8 quant live on Hugging Face at
+  // MiniMaxAI/MiniMax-M3 by 2026-06-07, under the custom minimax-community
+  // license. The 2026-06-01 launch shipped API access only; the arXiv report
+  // followed on 2026-06-11. sweep: 2026-06-12.
+  'MiniMax-M3': '2026-06-07',
+  // V4 Preview — V4-Pro (1.6T total / 49B active) and V4-Flash (284B/13B)
+  // published together under MIT on Hugging Face, the API and chat.deepseek.com.
+  // V4-Pro-0813 went GA on 2026-08-13, but the Hugging Face repo still hosts the
+  // April preview build, so the weights date is unchanged. sweep: 2026-04-25.
+  'DeepSeek-V4-Pro': '2026-04-24',
+  // Bucket covers GLM-5 and GLM-5.1, so the date is GLM-5's: released
+  // 2026-02-13, 744B/40B active, MIT, trained entirely on Huawei Ascend. Sources
+  // put the weights in "mid-February" without a day, and reporting clusters on
+  // 2026-02-13 and 2026-02-17; the Hugging Face commit history at zai-org/GLM-5
+  // is the authority if this ever needs to be exact. GLM-5.1 followed in April
+  // 2026. sweep: 2026-03-06.
+  'GLM-5': '2026-02-13',
+  // Apache 2.0 weights, plus official FP8 and GPTQ-Int4 quants, published at
+  // Qwen/Qwen3.5-397B-A17B on 2026-02-16 — the same day InferenceX first swept
+  // it. sweep: 2026-02-16 — day zero.
+  'Qwen-3.5-397B-A17B': '2026-02-16',
+  // Qwen announced the open-sourcing of Qwen3.8-Flash-Next and its FP8 sibling
+  // for 23:00 Beijing time on 2026-08-26, which is also the day the SGLang
+  // bring-up image tag was published. Some coverage puts the Hugging Face repo
+  // live on 08-24; the announced date is the one used here, and either way it
+  // precedes the first sweep on 08-27. The model card itself states no date.
+  // sweep: 2026-08-27 — day zero.
+  'Qwen3.8-Flash-Next': '2026-08-26',
+  // Bucket covers M2.5 and M2.7, so the date is M2.5's: announced 2026-02-12
+  // with weights on Hugging Face, architecturally unchanged from M2 (230B/10B).
+  // Was 2025-10-25, which is M2's launch, not M2.5's — `model-architectures.ts`
+  // still points its sourceUrl at MiniMax-M2, which is where that came from.
+  // sweep: 2026-02-18.
+  'MiniMax-M2.5': '2026-02-12',
+  // Weights at moonshotai/Kimi-K2.5 on 2026-01-27 under a modified MIT license,
+  // 1.04T total / 32B active. sweep: 2026-02-17.
+  'Kimi-K2.5': '2026-01-27',
+  // Weights landed on Hugging Face 2025-08-05 under Apache 2.0, alongside
+  // gpt-oss-20b — OpenAI's first open-weight release since GPT-2. AWS Bedrock
+  // lists the same launch date; the model-card paper followed on 2025-08-08. Was
+  // 2025-06-13, which precedes the release by seven weeks and matches no
+  // published event. sweep: 2025-09-09.
+  'gpt-oss-120b': '2025-08-05',
+  // Announced by WeChat post and pushed to Hugging Face on 2025-05-28 under MIT
+  // — the date the model name encodes. The model card commit landed a day later,
+  // weights first. sweep: 2025-08-13.
+  'DeepSeek-R1-0528': '2025-05-28',
+  // Meta published Llama 3.3 70B Instruct on 2024-12-06. sweep: 2025-08-12,
+  // shipped as workflow templates in the initial repo import.
+  'Llama-3.3-70B-Instruct-FP8': '2024-12-06',
+  // Llama 3.1 was published 2024-07-23. Hidden in the model selector, but the
+  // architecture diagram still captions it. No InferenceX sweep of its own.
+  'Llama-3.1-70B-Instruct-FP8-KV': '2024-07-23',
+  // Fork-only model (see .claude/skills/fork-changelog/SKILL.md), registered
+  // for a vngcloud/InferenceX smoke-test run rather than an upstream sweep.
+  // Apache 2.0 weights published on Hugging Face at
+  // deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct on 2024-06-17, alongside the
+  // full DeepSeek-Coder-V2 release.
+  'DeepSeek-Coder-V2-Lite-Instruct': '2024-06-17',
+  // Fork-only test-rig model (see .claude/skills/fork-changelog/SKILL.md),
+  // registered for a vngcloud/InferenceX smoke test (RedHatAI/gemma-4-31B-it-FP8-block
+  // checkpoint, run 31818878927). Not an upstream-tracked model, so there is
+  // no independently verifiable public release date; dated to the day this
+  // fork first registered the checkpoint, which cannot postdate the sweep.
+  'Gemma-4-31B': '2026-08-15',
+};
+
+/** Release date for a display model name, or null when we have no sourced date. */
+export function getModelReleaseDate(displayModel: string): string | null {
+  return MODEL_RELEASE_DATES[displayModel] ?? null;
+}
diff --git a/packages/constants/src/seo.test.ts b/packages/constants/src/seo.test.ts
new file mode 100644
index 000000000..6c25667c3
--- /dev/null
+++ b/packages/constants/src/seo.test.ts
@@ -0,0 +1,19 @@
+import { describe, expect, it } from 'vitest';
+
+import { DESCRIPTION, DESCRIPTION_ZH, SITE_TITLE, SITE_TITLE_ZH } from './seo';
+
+describe('site positioning', () => {
+  it('uses agentic inference for the category and AgentX for the scenario in English', () => {
+    expect(SITE_TITLE).toContain('Agentic Inference Benchmark');
+    expect(SITE_TITLE).not.toContain('AgentX');
+    expect(DESCRIPTION).toMatch(/agentic inference benchmark.*AgentX.*scenario/u);
+    expect(DESCRIPTION).toContain('fixed-sequence');
+  });
+
+  it('mirrors the category and scenario distinction in Simplified Chinese', () => {
+    expect(SITE_TITLE_ZH).toContain('智能体推理基准测试');
+    expect(SITE_TITLE_ZH).not.toContain('AgentX');
+    expect(DESCRIPTION_ZH).toMatch(/智能体推理基准测试.*AgentX.*场景/u);
+    expect(DESCRIPTION_ZH).toContain('固定序列');
+  });
+});
diff --git a/packages/constants/src/seo.ts b/packages/constants/src/seo.ts
index f50074d49..ec05ad4f9 100644
--- a/packages/constants/src/seo.ts
+++ b/packages/constants/src/seo.ts
@@ -3,9 +3,9 @@ export const SITE_URL = 'https://inferencex.semianalysis.com';
 export const AUTHOR_NAME = 'SemiAnalysis';
 export const AUTHOR_URL = 'https://semianalysis.com';
 export const AUTHOR_HANDLE = '@SemiAnalysis_';
-export const SITE_TITLE = `${SITE_NAME} by ${AUTHOR_NAME} — AI Inference Benchmark`;
+export const SITE_TITLE = `${SITE_NAME} by ${AUTHOR_NAME} — Agentic Inference Benchmark`;
 export const DESCRIPTION =
-  'InferenceX is the open-source AI inference benchmark that matches the rapid pace of modern AI development. Powered by one of the largest open-source chip CI/CD fleets with NVIDIA GB200, AMD MI355X & many more.';
+  'InferenceX is an open-source agentic inference benchmark. It compares the AgentX long-context, multi-turn coding scenario with fixed-sequence serving on NVIDIA, AMD, and other accelerators.';
 /**
  * Social-proof line woven into page meta descriptions to lift search CTR. The
  * named supporters mirror the published /quotes supporters page so the copy
@@ -21,7 +21,7 @@ export const OG_IMAGE = `${SITE_URL}/og-image.png`;
  * names (InferenceX, SemiAnalysis, GPU SKUs) stay in English per the
  * translation quality bar in AGENTS.md.
  */
-export const SITE_TITLE_ZH = `${SITE_NAME} by ${AUTHOR_NAME} — AI 推理基准测试`;
+export const SITE_TITLE_ZH = `${SITE_NAME} by ${AUTHOR_NAME} — 智能体推理基准测试`;
 export const DESCRIPTION_ZH =
-  'InferenceX 是紧跟现代 AI 发展节奏的开源 AI 推理基准测试，由规模领先的开源 Chip CI/CD 集群持续驱动，涵盖 NVIDIA GB200、AMD MI355X 等众多硬件。';
+  'InferenceX 是开源智能体推理基准测试平台，对比 AgentX 长上下文多轮编码场景与固定序列服务在 NVIDIA、AMD 等加速器上的性能。';
 export const SUPPORTERS_LINE_ZH = '获得 OpenAI、Microsoft 与 PyTorch 基金会的支持。';
diff --git a/packages/constants/src/tco.ts b/packages/constants/src/tco.ts
new file mode 100644
index 000000000..935e546da
--- /dev/null
+++ b/packages/constants/src/tco.ts
@@ -0,0 +1,13 @@
+/** Canonical public source for the TCO inputs used across InferenceX. */
+export const TCO_SOURCE_TITLE =
+  'SemiAnalysis Market July 2026 Pricing Surveys & AI Cloud TCO Model';
+
+export const TCO_SOURCE_URL = 'https://semianalysis.com/ai-cloud-tco-model/';
+
+/** Short product name for surfaces that plug the model itself rather than cite it as a data source. */
+export const TCO_MODEL_TITLE = 'AI Cloud TCO Model';
+
+/** SemiAnalysis institutional model tracking accelerator shipments, ASPs, specs and HBM supply. */
+export const ACCELERATOR_MODEL_TITLE = 'Accelerator & HBM Model';
+
+export const ACCELERATOR_MODEL_URL = 'https://semianalysis.com/accelerator-hbm-model/';
diff --git a/packages/constants/src/token-length-sketch.test.ts b/packages/constants/src/token-length-sketch.test.ts
new file mode 100644
index 000000000..3220c76fb
--- /dev/null
+++ b/packages/constants/src/token-length-sketch.test.ts
@@ -0,0 +1,41 @@
+import { describe, expect, it } from 'vitest';
+
+import {
+  buildTokenLengthSketch,
+  mergeTokenLengthSketches,
+  tokenLengthPercentiles,
+} from './token-length-sketch';
+
+describe('token length sketch', () => {
+  it('keeps percentile error below one percent', () => {
+    const samples = Array.from({ length: 100_000 }, (_, index) => index + 1);
+    const percentiles = tokenLengthPercentiles(buildTokenLengthSketch(samples));
+
+    expect(percentiles?.n).toBe(100_000);
+    for (const [actual, expected] of [
+      [percentiles?.p50, 50_000.5],
+      [percentiles?.p75, 75_000.25],
+      [percentiles?.p90, 90_000.1],
+      [percentiles?.p95, 95_000.05],
+      [percentiles?.p99, 99_000.01],
+    ] as const) {
+      expect(Math.abs((actual ?? 0) - expected) / expected).toBeLessThan(0.004);
+    }
+  });
+
+  it('merges point sketches with request weighting', () => {
+    const smallPoint = buildTokenLengthSketch([10, 20]);
+    const largePoint = buildTokenLengthSketch(Array.from({ length: 8 }, () => 1_000));
+    const percentiles = tokenLengthPercentiles(mergeTokenLengthSketches([smallPoint, largePoint]));
+
+    expect(percentiles?.n).toBe(10);
+    expect(Math.abs((percentiles?.p50 ?? 0) - 1_000)).toBeLessThanOrEqual(1);
+    expect(Math.abs((percentiles?.p90 ?? 0) - 1_000)).toBeLessThanOrEqual(1);
+  });
+
+  it('ignores invalid samples and preserves zero', () => {
+    const percentiles = tokenLengthPercentiles(buildTokenLengthSketch([Number.NaN, -1, 0, 10]));
+    expect(percentiles?.n).toBe(2);
+    expect(percentiles?.p50).toBe(5);
+  });
+});
diff --git a/packages/constants/src/token-length-sketch.ts b/packages/constants/src/token-length-sketch.ts
new file mode 100644
index 000000000..cd5ced5f1
--- /dev/null
+++ b/packages/constants/src/token-length-sketch.ts
@@ -0,0 +1,171 @@
+/**
+ * Mergeable, bounded-size histogram for positive integer token lengths.
+ *
+ * Each power-of-two interval is split into 256 linear buckets. This keeps the
+ * relative bucket width below 0.4% while bounding a full 1..2^31 distribution
+ * to fewer than 8,000 counters. Sketches can be merged by adding aligned
+ * counters, which lets the dashboard pool request distributions across an
+ * arbitrary set of benchmark points without reading their request timelines.
+ */
+
+export const TOKEN_LENGTH_SKETCH_VERSION = 1;
+export const TOKEN_LENGTH_BINS_PER_OCTAVE = 256;
+
+export interface TokenLengthSketch {
+  version: number;
+  binsPerOctave: number;
+  /** Number of finite, non-negative samples represented by this sketch. */
+  n: number;
+  /** Zero-length samples are kept outside the logarithmic bins. */
+  zeroCount: number;
+  /** Global index represented by counts[0]. */
+  minBin: number;
+  /** Dense counts from minBin through minBin + counts.length - 1. */
+  counts: number[];
+}
+
+export interface TokenLengthPercentiles {
+  p50: number;
+  p75: number;
+  p90: number;
+  p95: number;
+  p99: number;
+  n: number;
+}
+
+function binIndex(value: number, binsPerOctave: number): number {
+  const exponent = Math.floor(Math.log2(value));
+  const base = 2 ** exponent;
+  const subBin = Math.min(binsPerOctave - 1, Math.floor(((value - base) / base) * binsPerOctave));
+  return exponent * binsPerOctave + subBin;
+}
+
+function binValue(index: number, binsPerOctave: number): number {
+  const exponent = Math.floor(index / binsPerOctave);
+  const subBin = index - exponent * binsPerOctave;
+  const base = 2 ** exponent;
+  const lower = base * (1 + subBin / binsPerOctave);
+  const upper = base * (1 + (subBin + 1) / binsPerOctave);
+
+  // Token lengths are integers. When a bucket spans integers, use the middle
+  // integer rather than a fractional geometric representative.
+  const firstInteger = Math.ceil(lower);
+  const lastInteger = Math.ceil(upper) - 1;
+  if (firstInteger <= lastInteger) return (firstInteger + lastInteger) / 2;
+  return (lower + upper) / 2;
+}
+
+export function buildTokenLengthSketch(samples: readonly number[]): TokenLengthSketch | null {
+  const countsByBin = new Map<number, number>();
+  let zeroCount = 0;
+  let n = 0;
+  for (const raw of samples) {
+    if (!Number.isFinite(raw) || raw < 0) continue;
+    const value = Math.round(raw);
+    n += 1;
+    if (value === 0) {
+      zeroCount += 1;
+      continue;
+    }
+    const index = binIndex(value, TOKEN_LENGTH_BINS_PER_OCTAVE);
+    countsByBin.set(index, (countsByBin.get(index) ?? 0) + 1);
+  }
+  if (n === 0) return null;
+
+  const indices = [...countsByBin.keys()];
+  const minBin = indices.length > 0 ? Math.min(...indices) : 0;
+  const maxBin = indices.length > 0 ? Math.max(...indices) : minBin - 1;
+  const counts = Array.from({ length: Math.max(0, maxBin - minBin + 1) }, () => 0);
+  for (const [index, count] of countsByBin) counts[index - minBin] = count;
+
+  return {
+    version: TOKEN_LENGTH_SKETCH_VERSION,
+    binsPerOctave: TOKEN_LENGTH_BINS_PER_OCTAVE,
+    n,
+    zeroCount,
+    minBin,
+    counts,
+  };
+}
+
+function isCompatibleSketch(
+  sketch: TokenLengthSketch | null | undefined,
+): sketch is TokenLengthSketch {
+  return Boolean(
+    sketch &&
+    sketch.version === TOKEN_LENGTH_SKETCH_VERSION &&
+    sketch.binsPerOctave === TOKEN_LENGTH_BINS_PER_OCTAVE &&
+    Number.isFinite(sketch.n) &&
+    sketch.n >= 0 &&
+    Array.isArray(sketch.counts),
+  );
+}
+
+export function mergeTokenLengthSketches(
+  sketches: readonly (TokenLengthSketch | null | undefined)[],
+): TokenLengthSketch | null {
+  const valid = sketches.filter(isCompatibleSketch);
+  if (valid.length === 0) return null;
+
+  const withBins = valid.filter((sketch) => sketch.counts.length > 0);
+  const minBin = withBins.length > 0 ? Math.min(...withBins.map((sketch) => sketch.minBin)) : 0;
+  const maxBin =
+    withBins.length > 0
+      ? Math.max(...withBins.map((sketch) => sketch.minBin + sketch.counts.length - 1))
+      : minBin - 1;
+  const counts = Array.from({ length: Math.max(0, maxBin - minBin + 1) }, () => 0);
+
+  let n = 0;
+  let zeroCount = 0;
+  for (const sketch of valid) {
+    n += sketch.n;
+    zeroCount += sketch.zeroCount;
+    for (let i = 0; i < sketch.counts.length; i += 1) {
+      counts[sketch.minBin + i - minBin] += sketch.counts[i] ?? 0;
+    }
+  }
+  if (n === 0) return null;
+
+  return {
+    version: TOKEN_LENGTH_SKETCH_VERSION,
+    binsPerOctave: TOKEN_LENGTH_BINS_PER_OCTAVE,
+    n,
+    zeroCount,
+    minBin,
+    counts,
+  };
+}
+
+function valueAtRank(sketch: TokenLengthSketch, rank: number): number {
+  if (rank < sketch.zeroCount) return 0;
+  let cumulative = sketch.zeroCount;
+  for (let i = 0; i < sketch.counts.length; i += 1) {
+    cumulative += sketch.counts[i] ?? 0;
+    if (rank < cumulative) return binValue(sketch.minBin + i, sketch.binsPerOctave);
+  }
+  return binValue(sketch.minBin + sketch.counts.length - 1, sketch.binsPerOctave);
+}
+
+function quantile(sketch: TokenLengthSketch, q: number): number {
+  if (sketch.n === 1) return valueAtRank(sketch, 0);
+  const position = (sketch.n - 1) * q;
+  const lowerRank = Math.floor(position);
+  const upperRank = Math.ceil(position);
+  const lower = valueAtRank(sketch, lowerRank);
+  const upper = valueAtRank(sketch, upperRank);
+  return lower + (upper - lower) * (position - lowerRank);
+}
+
+export function tokenLengthPercentiles(
+  sketch: TokenLengthSketch | null | undefined,
+): TokenLengthPercentiles | null {
+  if (!isCompatibleSketch(sketch) || sketch.n <= 0) return null;
+  return {
+    p50: quantile(sketch, 0.5),
+    p75: quantile(sketch, 0.75),
+    p90: quantile(sketch, 0.9),
+    p95: quantile(sketch, 0.95),
+    p99: quantile(sketch, 0.99),
+    n: sketch.n,
+  };
+}
diff --git a/packages/constants/tsconfig.json b/packages/constants/tsconfig.json
index 75eba9f99..cfc7a1058 100644
--- a/packages/constants/tsconfig.json
+++ b/packages/constants/tsconfig.json
@@ -1,6 +1,7 @@
 {
-  "extends": "../../tsconfig.json",
+  "extends": "../../tsconfig.base.json",
   "compilerOptions": {
+    "target": "ES2018",
     "rootDir": "src",
     "outDir": "dist"
   },
diff --git a/packages/db/migrations/011_append_only_curves.sql b/packages/db/migrations/011_append_only_curves.sql
new file mode 100644
index 000000000..486bdc56c
--- /dev/null
+++ b/packages/db/migrations/011_append_only_curves.sql
@@ -0,0 +1,128 @@
+-- ============================================================
+-- APPEND-ONLY CURVES
+-- ============================================================
+--
+-- A normal sweep remains a complete immutable line snapshot. An explicitly
+-- append-only sweep contributes only new generated points, so the latest curve
+-- may span a consecutive chain of append-only runs plus the nearest full snapshot.
+
+alter table workflow_runs
+  add column append_only boolean not null default false;
+
+alter table changelog_entries
+  add column append_only boolean not null default false;
+
+drop materialized view if exists latest_benchmarks;
+
+-- `select *` in a view is expanded at creation time, so recreate it to expose
+-- workflow_runs.append_only to downstream queries.
+create or replace view latest_workflow_runs as
+select distinct on (github_run_id) *
+from workflow_runs
+order by github_run_id, run_attempt desc;
+
+create materialized view latest_benchmarks as
+with recursive run_lines as (
+  select
+    c.model,
+    c.hardware,
+    c.framework,
+    c.precision,
+    c.disagg,
+    case when br.benchmark_type = 'agentic_traces' then '' else c.spec_method end as line_spec_method,
+    br.benchmark_type,
+    br.isl,
+    br.osl,
+    br.offload_mode,
+    br.workflow_run_id,
+    br.date,
+    wr.run_started_at,
+    wr.append_only,
+    min(br.image) as image,
+    count(distinct br.image) as image_count,
+    bool_and(br.image is not null) as images_complete
+  from benchmark_results br
+  join configs c on c.id = br.config_id
+  join latest_workflow_runs wr on wr.id = br.workflow_run_id
+  where br.error is null
+  group by
+    c.model, c.hardware, c.framework, c.precision, c.disagg,
+    case when br.benchmark_type = 'agentic_traces' then '' else c.spec_method end,
+    br.benchmark_type, br.isl, br.osl, br.offload_mode,
+    br.workflow_run_id, br.date, wr.run_started_at, wr.append_only
+), ranked_runs as (
+  select
+    run_lines.*,
+    row_number() over (
+      partition by
+        model, hardware, framework, precision, disagg, line_spec_method,
+        benchmark_type, isl, osl, offload_mode
+      order by date desc, run_started_at desc nulls last, workflow_run_id desc
+    ) as run_rank
+  from run_lines
+), curve_runs as (
+  select
+    ranked_runs.*,
+    ranked_runs.image as root_image,
+    ranked_runs.date as snapshot_date,
+    ranked_runs.workflow_run_id as snapshot_workflow_run_id
+  from ranked_runs
+  where run_rank = 1
+
+  union all
+
+  select
+    older.*,
+    current.root_image,
+    current.snapshot_date,
+    current.snapshot_workflow_run_id
+  from curve_runs current
+  join ranked_runs older
+    on older.model = current.model
+    and older.hardware = current.hardware
+    and older.framework = current.framework
+    and older.precision = current.precision
+    and older.disagg = current.disagg
+    and older.line_spec_method = current.line_spec_method
+    and older.benchmark_type = current.benchmark_type
+    and older.isl is not distinct from current.isl
+    and older.osl is not distinct from current.osl
+    and older.offload_mode = current.offload_mode
+    and older.run_rank = current.run_rank + 1
+  where current.append_only
+    and current.image_count = 1
+    and current.images_complete
+    and older.image_count = 1
+    and older.images_complete
+    and older.image = current.root_image
+)
+select distinct on (
+  br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode, br.conc
+)
+  br.*,
+  cr.snapshot_date,
+  cr.snapshot_workflow_run_id
+from curve_runs cr
+join benchmark_results br
+  on br.workflow_run_id = cr.workflow_run_id
+  and br.benchmark_type = cr.benchmark_type
+  and br.isl is not distinct from cr.isl
+  and br.osl is not distinct from cr.osl
+  and br.offload_mode = cr.offload_mode
+join configs point_c
+  on point_c.id = br.config_id
+  and point_c.model = cr.model
+  and point_c.hardware = cr.hardware
+  and point_c.framework = cr.framework
+  and point_c.precision = cr.precision
+  and point_c.disagg = cr.disagg
+  and case when br.benchmark_type = 'agentic_traces' then '' else point_c.spec_method end = cr.line_spec_method
+where br.error is null
+order by
+  br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode, br.conc,
+  cr.run_rank;
+
+create unique index latest_benchmarks_pk
+  on latest_benchmarks (config_id, conc, isl, osl, benchmark_type, offload_mode)
+  nulls not distinct;
+create index latest_benchmarks_model_idx on latest_benchmarks (config_id);
diff --git a/packages/db/migrations/012_benchmark_recipe_fingerprint.sql b/packages/db/migrations/012_benchmark_recipe_fingerprint.sql
new file mode 100644
index 000000000..fe1216ece
--- /dev/null
+++ b/packages/db/migrations/012_benchmark_recipe_fingerprint.sql
@@ -0,0 +1,154 @@
+-- ============================================================
+-- BENCHMARK RECIPE FINGERPRINTS
+-- ============================================================
+--
+-- A generated recipe can differ in fields that are not normalized into configs
+-- (for example pipeline parallelism or launcher settings). Keep those variants
+-- distinct even when they share topology, scenario, and concurrency. Historical
+-- rows remain NULL and retain the previous one-point-per-natural-key behavior.
+
+alter table benchmark_results
+  add column recipe_fingerprint text;
+
+alter table benchmark_results
+  drop constraint benchmark_results_unique;
+
+alter table benchmark_results
+  add constraint benchmark_results_unique unique nulls not distinct (
+    workflow_run_id,
+    config_id,
+    benchmark_type,
+    isl,
+    osl,
+    conc,
+    offload_mode,
+    recipe_fingerprint
+  );
+
+drop materialized view if exists latest_benchmarks;
+
+create materialized view latest_benchmarks as
+with recursive run_lines as (
+  select
+    c.model,
+    c.hardware,
+    c.framework,
+    c.precision,
+    c.disagg,
+    case when br.benchmark_type = 'agentic_traces' then '' else c.spec_method end as line_spec_method,
+    br.benchmark_type,
+    br.isl,
+    br.osl,
+    br.offload_mode,
+    br.workflow_run_id,
+    br.date,
+    wr.run_started_at,
+    wr.append_only,
+    min(br.image) as image,
+    count(distinct br.image) as image_count,
+    bool_and(br.image is not null) as images_complete
+  from benchmark_results br
+  join configs c on c.id = br.config_id
+  join latest_workflow_runs wr on wr.id = br.workflow_run_id
+  where br.error is null
+  group by
+    c.model, c.hardware, c.framework, c.precision, c.disagg,
+    case when br.benchmark_type = 'agentic_traces' then '' else c.spec_method end,
+    br.benchmark_type, br.isl, br.osl, br.offload_mode,
+    br.workflow_run_id, br.date, wr.run_started_at, wr.append_only
+), ranked_runs as (
+  select
+    run_lines.*,
+    row_number() over (
+      partition by
+        model, hardware, framework, precision, disagg, line_spec_method,
+        benchmark_type, isl, osl, offload_mode
+      order by date desc, run_started_at desc nulls last, workflow_run_id desc
+    ) as run_rank
+  from run_lines
+), curve_runs as (
+  select
+    ranked_runs.*,
+    ranked_runs.image as root_image,
+    ranked_runs.date as snapshot_date,
+    ranked_runs.workflow_run_id as snapshot_workflow_run_id
+  from ranked_runs
+  where run_rank = 1
+
+  union all
+
+  select
+    older.*,
+    current.root_image,
+    current.snapshot_date,
+    current.snapshot_workflow_run_id
+  from curve_runs current
+  join ranked_runs older
+    on older.model = current.model
+    and older.hardware = current.hardware
+    and older.framework = current.framework
+    and older.precision = current.precision
+    and older.disagg = current.disagg
+    and older.line_spec_method = current.line_spec_method
+    and older.benchmark_type = current.benchmark_type
+    and older.isl is not distinct from current.isl
+    and older.osl is not distinct from current.osl
+    and older.offload_mode = current.offload_mode
+    and older.run_rank = current.run_rank + 1
+  where current.append_only
+    and current.image_count = 1
+    and current.images_complete
+    and older.image_count = 1
+    and older.images_complete
+    and older.image = current.root_image
+)
+select distinct on (
+  br.config_id,
+  br.benchmark_type,
+  br.isl,
+  br.osl,
+  br.offload_mode,
+  br.recipe_fingerprint,
+  br.conc
+)
+  br.*,
+  cr.snapshot_date,
+  cr.snapshot_workflow_run_id
+from curve_runs cr
+join benchmark_results br
+  on br.workflow_run_id = cr.workflow_run_id
+  and br.benchmark_type = cr.benchmark_type
+  and br.isl is not distinct from cr.isl
+  and br.osl is not distinct from cr.osl
+  and br.offload_mode = cr.offload_mode
+join configs point_c
+  on point_c.id = br.config_id
+  and point_c.model = cr.model
+  and point_c.hardware = cr.hardware
+  and point_c.framework = cr.framework
+  and point_c.precision = cr.precision
+  and point_c.disagg = cr.disagg
+  and case when br.benchmark_type = 'agentic_traces' then '' else point_c.spec_method end = cr.line_spec_method
+where br.error is null
+order by
+  br.config_id,
+  br.benchmark_type,
+  br.isl,
+  br.osl,
+  br.offload_mode,
+  br.recipe_fingerprint,
+  br.conc,
+  cr.run_rank;
+
+create unique index latest_benchmarks_pk
+  on latest_benchmarks (
+    config_id,
+    conc,
+    isl,
+    osl,
+    benchmark_type,
+    offload_mode,
+    recipe_fingerprint
+  )
+  nulls not distinct;
+create index latest_benchmarks_model_idx on latest_benchmarks (config_id);
diff --git a/packages/db/migrations/013_server_log_files.sql b/packages/db/migrations/013_server_log_files.sql
new file mode 100644
index 000000000..006819d87
--- /dev/null
+++ b/packages/db/migrations/013_server_log_files.sql
@@ -0,0 +1,16 @@
+-- Preserve every text log emitted inside server-log artifacts while keeping
+-- the existing server_logs.server_log column as the primary/legacy file.
+
+alter table server_logs
+  add column file_name text not null default 'server.log',
+  add column files_complete boolean not null default false,
+  add constraint server_logs_file_name_nonempty check (file_name <> '');
+
+create table server_log_files (
+  server_log_id bigint not null references server_logs(id) on delete cascade,
+  file_name     text   not null,
+  log_text      text   not null,
+
+  constraint server_log_files_file_name_nonempty check (file_name <> ''),
+  constraint server_log_files_unique unique (server_log_id, file_name)
+);
diff --git a/packages/db/package.json b/packages/db/package.json
index 7f676998a..378edda6a 100644
--- a/packages/db/package.json
+++ b/packages/db/package.json
@@ -28,6 +28,7 @@
     "db:backfill-full-response-interactivity": "bun --env-file=../../.env src/backfill-full-response-interactivity.ts",
     "db:backfill-request-timeline": "bun --env-file=../../.env src/backfill-request-timeline.ts",
     "db:backfill-runtime-metadata": "bun --env-file=../../.env src/backfill-runtime-metadata.ts",
+    "db:backfill-server-log-files": "bun --env-file=../../.env src/backfill-server-log-files.ts",
     "db:reset": "bun --env-file=../../.env src/reset-db.ts",
     "db:verify": "bun --env-file=../../.env src/verify-db.ts",
     "test:unit": "vitest run",
@@ -35,7 +36,7 @@
   },
   "dependencies": {
     "@neondatabase/serverless": "^1.1.0",
-    "@noble/ciphers": "^2.2.0",
+    "@noble/ciphers": "^2.3.0",
     "@semianalysisai/inferencex-constants": "workspace:*",
     "postgres": "^3.4.9",
     "stream-chain": "^4.2.5",
@@ -43,7 +44,7 @@
   },
   "devDependencies": {
     "@types/adm-zip": "^0.5.8",
-    "@types/node": "^26.1.1",
+    "@types/node": "^26.2.0",
     "@types/stream-json": "^1.7.8",
     "@vitest/coverage-v8": "^4.1.10",
     "adm-zip": "^0.6.0",
diff --git a/packages/db/src/apply-overrides.ts b/packages/db/src/apply-overrides.ts
index 562d1faac..63c518822 100644
--- a/packages/db/src/apply-overrides.ts
+++ b/packages/db/src/apply-overrides.ts
@@ -1,9 +1,11 @@
 /**
  * Enforce all run-overrides.ts entries against the DB:
  *   1. Patch conclusions for CONCLUSION_OVERRIDES
- *   2. Purge runs listed in PURGED_RUNS
- *   3. Purge specific attempts listed in PURGED_RUN_ATTEMPTS
- *   4. Purge individual benchmark points listed in PURGED_BENCHMARK_POINTS
+ *   2. Patch changelogs listed in CHANGELOG_BACKFILLS
+ *   3. Patch benchmark points listed in BENCHMARK_POINT_BACKFILLS
+ *   4. Purge runs listed in PURGED_RUNS
+ *   5. Purge specific attempts listed in PURGED_RUN_ATTEMPTS
+ *   6. Purge individual benchmark points listed in PURGED_BENCHMARK_POINTS
  * Previews changes (read-only), then confirms before writing.
  *
  * Usage:
@@ -14,14 +16,23 @@
  * Use this command directly only for local preview or manual recovery.
  */
 
+import { isDeepStrictEqual } from 'node:util';
+
 import { confirm, hasNoSslFlag, hasYesFlag } from './cli-utils.js';
+import { configCacheKey } from './etl/config-cache.js';
 import { type Sql, createAdminSql, refreshLatestBenchmarks } from './etl/db-utils.js';
+import { jsonbParam } from './lib/backfill-runner.js';
 import {
+  type BenchmarkPointBackfill,
+  type ChangelogBackfill,
   type PurgedBenchmarkPoint,
+  BENCHMARK_POINT_BACKFILLS,
+  CHANGELOG_BACKFILLS,
   CONCLUSION_OVERRIDES,
   PURGED_BENCHMARK_POINTS,
   PURGED_RUN_ATTEMPTS,
   PURGED_RUNS,
+  validateRunBackfills,
 } from './etl/run-overrides.js';
 
 const sql = createAdminSql({
@@ -73,6 +84,279 @@ async function applyConclusions(stale: StaleRow[]): Promise<void> {
   }
 }
 
+// ── Audited backfills ────────────────────────────────────────────────────────
+
+interface ChangelogBackfillTarget {
+  backfill: ChangelogBackfill;
+  changelogId: number;
+  workflowRunId: number;
+}
+
+function changelogBackfillIsApplied(
+  row: Record<string, unknown>,
+  backfill: ChangelogBackfill,
+): boolean {
+  const { set } = backfill;
+  return (
+    (set.configKeys === undefined || isDeepStrictEqual(row.config_keys, [...set.configKeys])) &&
+    (set.description === undefined || row.description === set.description) &&
+    (set.prLink === undefined || row.pr_link === set.prLink) &&
+    (set.appendOnly === undefined ||
+      (row.changelog_append_only === set.appendOnly && row.run_append_only === set.appendOnly))
+  );
+}
+
+/** Resolve one exact changelog row and return it only when its patch is stale. */
+async function previewChangelogBackfill(
+  backfill: ChangelogBackfill,
+): Promise<ChangelogBackfillTarget | null> {
+  const rows = await sql`
+    SELECT cl.id, cl.workflow_run_id, cl.config_keys, cl.description, cl.pr_link,
+           cl.append_only AS changelog_append_only, wr.append_only AS run_append_only
+    FROM changelog_entries cl
+    JOIN workflow_runs wr ON wr.id = cl.workflow_run_id
+    WHERE wr.github_run_id = ${backfill.githubRunId}
+      AND wr.run_attempt = ${backfill.runAttempt}
+      AND cl.base_ref = ${backfill.baseRef}
+      AND cl.head_ref = ${backfill.headRef}
+  `;
+  if (rows.length !== 1) {
+    throw new Error(
+      `${backfill.id}: expected exactly one changelog row, found ${rows.length} ` +
+        `(run ${backfill.githubRunId} attempt ${backfill.runAttempt}, ` +
+        `${backfill.baseRef}..${backfill.headRef})`,
+    );
+  }
+  const [row] = rows;
+  console.log(
+    `    ${backfill.id}: run ${backfill.githubRunId} attempt ${backfill.runAttempt}, ` +
+      `${backfill.baseRef}..${backfill.headRef}`,
+  );
+  console.log(`      reason: ${backfill.reason}`);
+  if (changelogBackfillIsApplied(row, backfill)) {
+    console.log('      already applied.');
+    return null;
+  }
+  console.log(`      set ${JSON.stringify(backfill.set)}`);
+  return {
+    backfill,
+    changelogId: row.id as number,
+    workflowRunId: row.workflow_run_id as number,
+  };
+}
+
+async function applyChangelogBackfill(target: ChangelogBackfillTarget): Promise<void> {
+  const { backfill, changelogId, workflowRunId } = target;
+  await sql.begin(async (_tx) => {
+    const tx = _tx as unknown as Sql;
+    if (backfill.set.configKeys !== undefined) {
+      await tx`
+        UPDATE changelog_entries
+        SET config_keys = ${[...backfill.set.configKeys]}
+        WHERE id = ${changelogId}
+      `;
+    }
+    if (backfill.set.description !== undefined) {
+      await tx`
+        UPDATE changelog_entries SET description = ${backfill.set.description}
+        WHERE id = ${changelogId}
+      `;
+    }
+    if (backfill.set.prLink !== undefined) {
+      await tx`
+        UPDATE changelog_entries SET pr_link = ${backfill.set.prLink}
+        WHERE id = ${changelogId}
+      `;
+    }
+    if (backfill.set.appendOnly !== undefined) {
+      await tx`
+        UPDATE changelog_entries SET append_only = ${backfill.set.appendOnly}
+        WHERE id = ${changelogId}
+      `;
+      await tx`
+        UPDATE workflow_runs SET append_only = ${backfill.set.appendOnly}
+        WHERE id = ${workflowRunId}
+      `;
+    }
+  });
+  console.log(`    ${backfill.id} patched.`);
+}
+
+interface BenchmarkPointBackfillTarget {
+  backfill: BenchmarkPointBackfill;
+  resultId: number;
+  metrics: Record<string, unknown>;
+}
+
+function asMetricsRecord(value: unknown): Record<string, unknown> | null {
+  return value !== null && typeof value === 'object' && !Array.isArray(value)
+    ? (value as Record<string, unknown>)
+    : null;
+}
+
+function benchmarkPointMetricsPatch(backfill: BenchmarkPointBackfill): Record<string, unknown> {
+  const desiredOffloadMode = backfill.set.offloadMode ?? backfill.offloadMode;
+  return {
+    ...backfill.set.metricsMerge,
+    ...(backfill.set.offloadMode === undefined ? {} : { offload_mode: desiredOffloadMode }),
+  };
+}
+
+/**
+ * Recognize the exact malformed value written by the first benchmark-point
+ * backfill implementation. It concatenated a metrics object with a JSONB
+ * string, producing `[originalMetrics, "{...patch}"]`.
+ */
+function recoverMalformedBenchmarkPointMetrics(
+  value: unknown,
+  backfill: BenchmarkPointBackfill,
+): Record<string, unknown> | null {
+  if (!Array.isArray(value) || value.length !== 2) return null;
+  const originalMetrics = asMetricsRecord(value[0]);
+  if (!originalMetrics || typeof value[1] !== 'string') return null;
+
+  try {
+    const appendedPatch = JSON.parse(value[1]) as unknown;
+    if (!isDeepStrictEqual(appendedPatch, benchmarkPointMetricsPatch(backfill))) return null;
+  } catch {
+    return null;
+  }
+  return originalMetrics;
+}
+
+function patchedBenchmarkPointMetrics(
+  sourceMetrics: Record<string, unknown>,
+  backfill: BenchmarkPointBackfill,
+): Record<string, unknown> {
+  const metrics = { ...sourceMetrics };
+  for (const key of backfill.set.metricsRemove ?? []) delete metrics[key];
+  Object.assign(metrics, benchmarkPointMetricsPatch(backfill));
+  return metrics;
+}
+
+function benchmarkPointBackfillIsApplied(
+  row: Record<string, unknown>,
+  backfill: BenchmarkPointBackfill,
+): boolean {
+  const desiredOffloadMode = backfill.set.offloadMode ?? backfill.offloadMode;
+  if (row.offload_mode !== desiredOffloadMode) return false;
+  const metrics = asMetricsRecord(row.metrics);
+  if (!metrics) return false;
+  if (backfill.set.offloadMode !== undefined && metrics.offload_mode !== desiredOffloadMode) {
+    return false;
+  }
+  for (const [key, value] of Object.entries(backfill.set.metricsMerge ?? {})) {
+    if (!isDeepStrictEqual(metrics[key], value)) return false;
+  }
+  for (const key of backfill.set.metricsRemove ?? []) {
+    if (Object.hasOwn(metrics, key)) return false;
+  }
+  return true;
+}
+
+function benchmarkPointDescription(backfill: BenchmarkPointBackfill): string {
+  return (
+    `run ${backfill.githubRunId} attempt ${backfill.runAttempt}, ` +
+    `config ${configCacheKey(backfill.config)} (production id ${backfill.productionConfigId}), ` +
+    `${backfill.benchmarkType}, isl ${backfill.isl}, osl ${backfill.osl}, ` +
+    `conc ${backfill.conc}, offload ${backfill.offloadMode}, ` +
+    `recipe ${backfill.recipeFingerprint ?? 'legacy'}`
+  );
+}
+
+/** Resolve one source/desired point identity and fail closed on ambiguity. */
+async function previewBenchmarkPointBackfill(
+  backfill: BenchmarkPointBackfill,
+): Promise<BenchmarkPointBackfillTarget | null> {
+  const desiredOffloadMode = backfill.set.offloadMode ?? backfill.offloadMode;
+  const offloadModes = [...new Set([backfill.offloadMode, desiredOffloadMode])];
+  const { config } = backfill;
+  const rows = await sql`
+    SELECT br.id, br.offload_mode, br.metrics
+    FROM benchmark_results br
+    JOIN workflow_runs wr ON wr.id = br.workflow_run_id
+    JOIN configs c ON c.id = br.config_id
+    WHERE wr.github_run_id = ${backfill.githubRunId}
+      AND wr.run_attempt = ${backfill.runAttempt}
+      AND c.hardware = ${config.hardware}
+      AND c.framework = ${config.framework}
+      AND c.model = ${config.model}
+      AND c.precision = ${config.precision}
+      AND c.spec_method = ${config.specMethod}
+      AND c.disagg = ${config.disagg}
+      AND c.is_multinode = ${config.isMultinode}
+      AND c.prefill_tp = ${config.prefillTp}
+      AND c.prefill_ep = ${config.prefillEp}
+      AND c.prefill_dp_attention = ${config.prefillDpAttn}
+      AND c.prefill_num_workers = ${config.prefillNumWorkers}
+      AND c.decode_tp = ${config.decodeTp}
+      AND c.decode_ep = ${config.decodeEp}
+      AND c.decode_dp_attention = ${config.decodeDpAttn}
+      AND c.decode_num_workers = ${config.decodeNumWorkers}
+      AND c.num_prefill_gpu = ${config.numPrefillGpu}
+      AND c.num_decode_gpu = ${config.numDecodeGpu}
+      AND br.benchmark_type = ${backfill.benchmarkType}
+      AND br.isl IS NOT DISTINCT FROM ${backfill.isl}
+      AND br.osl IS NOT DISTINCT FROM ${backfill.osl}
+      AND br.conc = ${backfill.conc}
+      AND br.offload_mode = ANY(${offloadModes})
+      AND br.recipe_fingerprint IS NOT DISTINCT FROM ${backfill.recipeFingerprint ?? null}
+  `;
+  if (rows.length !== 1) {
+    throw new Error(
+      `${backfill.id}: expected exactly one source or desired benchmark row, ` +
+        `found ${rows.length} (${benchmarkPointDescription(backfill)})`,
+    );
+  }
+  const [row] = rows;
+  console.log(`    ${backfill.id}: ${benchmarkPointDescription(backfill)}`);
+  console.log(`      reason: ${backfill.reason}`);
+  if (benchmarkPointBackfillIsApplied(row, backfill)) {
+    console.log('      already applied.');
+    return null;
+  }
+
+  const malformedMetrics = recoverMalformedBenchmarkPointMetrics(row.metrics, backfill);
+  if (
+    row.offload_mode === desiredOffloadMode &&
+    desiredOffloadMode !== backfill.offloadMode &&
+    malformedMetrics === null
+  ) {
+    throw new Error(
+      `${backfill.id}: source identity is missing and the desired identity has unexpected data`,
+    );
+  }
+  const sourceMetrics = asMetricsRecord(row.metrics) ?? malformedMetrics;
+  if (!sourceMetrics) {
+    throw new Error(`${backfill.id}: benchmark metrics have an unexpected JSON shape`);
+  }
+  if (malformedMetrics) {
+    console.log('      repair malformed metrics written by the previous backfill attempt');
+  }
+  console.log(`      set ${JSON.stringify(backfill.set)}`);
+  return {
+    backfill,
+    resultId: row.id as number,
+    metrics: patchedBenchmarkPointMetrics(sourceMetrics, backfill),
+  };
+}
+
+async function applyBenchmarkPointBackfillToDatabase(
+  target: BenchmarkPointBackfillTarget,
+): Promise<void> {
+  const { backfill, resultId, metrics } = target;
+  const desiredOffloadMode = backfill.set.offloadMode ?? backfill.offloadMode;
+  const [updated] = await sql`
+    UPDATE benchmark_results
+    SET offload_mode = ${desiredOffloadMode},
+        metrics = ${jsonbParam(sql, metrics)}
+    WHERE id = ${resultId}
+    RETURNING id
+  `;
+  if (!updated) throw new Error(`${backfill.id}: benchmark row disappeared before update`);
+  console.log(`    ${backfill.id} patched.`);
+}
+
 // ── Purge ─────────────────────────────────────────────────────────────────────
 
 interface PurgeTarget {
@@ -290,10 +574,12 @@ async function previewBenchmarkPointPurge(
       AND br.osl IS NOT DISTINCT FROM ${point.osl}
       AND br.conc = ${point.conc}
       AND br.offload_mode = ${point.offloadMode}
+      AND br.recipe_fingerprint IS NOT DISTINCT FROM ${point.recipeFingerprint ?? null}
   `;
   const description =
     `config ${point.configId}, ${point.benchmarkType}, isl ${point.isl}, ` +
-    `osl ${point.osl}, conc ${point.conc}, offload ${point.offloadMode}`;
+    `osl ${point.osl}, conc ${point.conc}, offload ${point.offloadMode}, ` +
+    `recipe ${point.recipeFingerprint ?? 'legacy'}`;
   if (rows.length === 0) {
     console.log(`    ${description}, not in DB, skipping.`);
     return null;
@@ -312,6 +598,7 @@ async function purgeBenchmarkPoints(resultIds: number[]): Promise<void> {
 // ── Main ──────────────────────────────────────────────────────────────────────
 
 async function main(): Promise<void> {
+  validateRunBackfills();
   console.log('=== apply-overrides ===');
 
   // Phase 1: preview (read-only)
@@ -328,6 +615,30 @@ async function main(): Promise<void> {
     console.log('\n  Conclusions: all up to date.');
   }
 
+  const changelogBackfillTargets: ChangelogBackfillTarget[] = [];
+  if (CHANGELOG_BACKFILLS.length > 0) {
+    console.log(`\n  Changelog backfills (${CHANGELOG_BACKFILLS.length}):`);
+    for (const backfill of CHANGELOG_BACKFILLS) {
+      const target = await previewChangelogBackfill(backfill);
+      if (target) changelogBackfillTargets.push(target);
+    }
+  } else {
+    console.log('\n  Changelog backfills: none registered.');
+  }
+  if (changelogBackfillTargets.length > 0) hasWork = true;
+
+  const benchmarkPointBackfillTargets: BenchmarkPointBackfillTarget[] = [];
+  if (BENCHMARK_POINT_BACKFILLS.length > 0) {
+    console.log(`\n  Benchmark point backfills (${BENCHMARK_POINT_BACKFILLS.length}):`);
+    for (const backfill of BENCHMARK_POINT_BACKFILLS) {
+      const target = await previewBenchmarkPointBackfill(backfill);
+      if (target) benchmarkPointBackfillTargets.push(target);
+    }
+  } else {
+    console.log('\n  Benchmark point backfills: none registered.');
+  }
+  if (benchmarkPointBackfillTargets.length > 0) hasWork = true;
+
   const purgeTargets = [...PURGED_RUNS];
   const found: PurgeTarget[] = [];
   if (purgeTargets.length > 0) {
@@ -393,6 +704,20 @@ async function main(): Promise<void> {
     await applyConclusions(stale);
   }
 
+  if (changelogBackfillTargets.length > 0) {
+    console.log('\n  Patching changelogs...');
+    for (const target of changelogBackfillTargets) {
+      await applyChangelogBackfill(target);
+    }
+  }
+
+  if (benchmarkPointBackfillTargets.length > 0) {
+    console.log('\n  Patching benchmark points...');
+    for (const target of benchmarkPointBackfillTargets) {
+      await applyBenchmarkPointBackfillToDatabase(target);
+    }
+  }
+
   if (found.length > 0) {
     console.log('\n  Purging runs...');
     for (const { wrIds } of found) {
@@ -407,7 +732,17 @@ async function main(): Promise<void> {
     }
   }
 
-  // Phase 4: refresh mat view
+  // Phase 4: verify every requested backfill reached its declared state.
+  for (const backfill of CHANGELOG_BACKFILLS) {
+    const staleTarget = await previewChangelogBackfill(backfill);
+    if (staleTarget) throw new Error(`${backfill.id}: changelog verification failed`);
+  }
+  for (const backfill of BENCHMARK_POINT_BACKFILLS) {
+    const staleTarget = await previewBenchmarkPointBackfill(backfill);
+    if (staleTarget) throw new Error(`${backfill.id}: benchmark point verification failed`);
+  }
+
+  // Phase 5: refresh mat view
   await refreshLatestBenchmarks(sql);
 
   console.log('\n=== apply-overrides complete ===');
diff --git a/packages/db/src/backfill-aggregate-stats.ts b/packages/db/src/backfill-aggregate-stats.ts
index a50f6ea02..c1ce8f2f0 100644
--- a/packages/db/src/backfill-aggregate-stats.ts
+++ b/packages/db/src/backfill-aggregate-stats.ts
@@ -24,67 +24,91 @@
 import { hasNoSslFlag } from './cli-utils.js';
 import {
   computeAggregateStats,
+  computeProfileAggregateStatsFromCompressedChunks,
   mergeProfileStatsUpgrade,
   STATS_VERSION,
   type AggregateStats,
 } from './etl/compute-aggregate-stats.js';
 import { createAdminSql } from './etl/db-utils.js';
 import {
-  confirmProceed,
   jsonbParam,
   parseLimitForceFlags,
   runBackfillMain,
-  runPerIdBackfill,
+  runCandidateIdBackfill,
 } from './lib/backfill-runner.js';
 
 const flags = parseLimitForceFlags();
 
+// Neon's HTTP response and JS drivers should not receive a 100+ MB bytea as
+// one value. Slice oversized TOAST values into independent bounded reads and
+// feed the compressed bytes directly into the streaming gzip parser.
+const MAX_INLINE_PROFILE_BYTES = 64 * 1024 * 1024;
+const PROFILE_CHUNK_BYTES = 8 * 1024 * 1024;
+
 const sql = createAdminSql({
   noSsl: hasNoSslFlag(),
   max: 1,
   onnotice: () => {},
 });
 
+async function* readProfileChunks(id: number, totalBytes: number): AsyncGenerator<Buffer> {
+  for (let offset = 0; offset < totalBytes; offset += PROFILE_CHUNK_BYTES) {
+    const length = Math.min(PROFILE_CHUNK_BYTES, totalBytes - offset);
+    const [row] = await sql<{ chunk: Buffer }[]>`
+      select substring(profile_export_jsonl_gz from ${offset + 1} for ${length}) as chunk
+      from agentic_trace_replay
+      where id = ${id}
+    `;
+    if (!row?.chunk) throw new Error(`profile blob chunk missing at byte ${offset}`);
+    yield row.chunk;
+  }
+}
+
 async function main(): Promise<void> {
   console.log('=== backfill-aggregate-stats ===');
   console.log(`  STATS_VERSION = ${STATS_VERSION}`);
   console.log(`  force = ${flags.force}`);
   console.log(`  limit = ${flags.limit ?? 'none'}`);
 
-  // Find candidates: rows missing stats, or whose stored version is stale.
-  // Using >>'version'::int comparison would error on null; coalesce to -1 so
-  // null-stats rows always count as stale.
-  const candidates = flags.force
-    ? await sql<{ id: number }[]>`
-        select id
-        from agentic_trace_replay
-        order by id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `
-    : await sql<{ id: number }[]>`
-        select id
-        from agentic_trace_replay
-        where aggregate_stats is null
-           or coalesce((aggregate_stats->>'version')::int, -1) <> ${STATS_VERSION}
-        order by id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `;
-
-  if (candidates.length === 0) {
-    console.log('\n  Nothing to do — all rows up to date.');
-    return;
-  }
-
-  if (!(await confirmProceed(`${candidates.length} candidate row(s).`))) return;
-
-  await runPerIdBackfill(
-    candidates.map((c) => c.id),
+  await runCandidateIdBackfill(
+    async () => {
+      // Find candidates: rows missing stats, or whose stored version is stale.
+      // Using >>'version'::int comparison would error on null; coalesce to -1 so
+      // null-stats rows always count as stale.
+      const candidates = flags.force
+        ? await sql<{ id: number }[]>`
+            select id
+            from agentic_trace_replay
+            order by id
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `
+        : await sql<{ id: number }[]>`
+            select id
+            from agentic_trace_replay
+            where aggregate_stats is null
+               or coalesce((aggregate_stats->>'version')::int, -1) <> ${STATS_VERSION}
+            order by id
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `;
+      return candidates.map((candidate) => candidate.id);
+    },
     async (id) => {
       // Fetch one row at a time — the json_gz blob is the heavy field.
       const [row] = await sql<
-        { profile_export_jsonl_gz: Buffer | null; aggregate_stats: AggregateStats | null }[]
+        {
+          profile_export_jsonl_gz: Buffer | null;
+          profile_blob_bytes: number;
+          aggregate_stats: AggregateStats | null;
+        }[]
       >`
-        select profile_export_jsonl_gz, aggregate_stats
+        select
+          case
+            when pg_column_size(profile_export_jsonl_gz) <= ${MAX_INLINE_PROFILE_BYTES}
+              then profile_export_jsonl_gz
+            else null
+          end as profile_export_jsonl_gz,
+          coalesce(pg_column_size(profile_export_jsonl_gz), 0)::bigint as profile_blob_bytes,
+          aggregate_stats
         from agentic_trace_replay
         where id = ${id}
       `;
@@ -93,16 +117,22 @@ async function main(): Promise<void> {
         return 'skipped';
       }
 
+      const profileStats =
+        Number(row.profile_blob_bytes) > MAX_INLINE_PROFILE_BYTES
+          ? await computeProfileAggregateStatsFromCompressedChunks(
+              readProfileChunks(id, Number(row.profile_blob_bytes)),
+            )
+          : await computeAggregateStats({
+              profileBlob: row.profile_export_jsonl_gz,
+              serverBlob: null,
+            });
+
       let stats: AggregateStats;
       // v3 onwards → current is a profile-only change (the server-derived
       // fields haven't changed since v3), so skip re-reading the huge server
       // blob and carry its KV/prefix distributions forward.
       const storedVersion = row.aggregate_stats?.version;
       if (storedVersion !== undefined && storedVersion >= 3 && storedVersion < STATS_VERSION) {
-        const profileStats = await computeAggregateStats({
-          profileBlob: row.profile_export_jsonl_gz,
-          serverBlob: null,
-        });
         stats = mergeProfileStatsUpgrade(row.aggregate_stats!, profileStats);
       } else {
         const [serverRow] = await sql<{ server_metrics_json_gz: Buffer | null }[]>`
@@ -110,10 +140,11 @@ async function main(): Promise<void> {
           from agentic_trace_replay
           where id = ${id}
         `;
-        stats = await computeAggregateStats({
-          profileBlob: row.profile_export_jsonl_gz,
+        const serverStats = await computeAggregateStats({
+          profileBlob: null,
           serverBlob: serverRow?.server_metrics_json_gz ?? null,
         });
+        stats = mergeProfileStatsUpgrade(serverStats, profileStats);
       }
 
       await sql`
diff --git a/packages/db/src/backfill-chart-series.ts b/packages/db/src/backfill-chart-series.ts
index 98161c773..800499534 100644
--- a/packages/db/src/backfill-chart-series.ts
+++ b/packages/db/src/backfill-chart-series.ts
@@ -19,6 +19,8 @@
  *   bun run --cwd packages/db db:backfill-chart-series
  *     [--limit N]   only process the first N candidate rows
  *     [--force]     recompute every row, even if version already matches
+ *     [--shard-count N] split work across N independent processes
+ *     [--shard-index N] zero-based shard handled by this process
  *     [--yes]       skip the confirmation prompt
  */
 
@@ -26,11 +28,10 @@ import { hasNoSslFlag } from './cli-utils.js';
 import { CHART_SERIES_VERSION, computeChartSeries } from './etl/compute-chart-series.js';
 import { createAdminSql } from './etl/db-utils.js';
 import {
-  confirmProceed,
   jsonbParam,
   parseLimitForceFlags,
   runBackfillMain,
-  runPerIdBackfill,
+  runCandidateIdBackfill,
 } from './lib/backfill-runner.js';
 
 const flags = parseLimitForceFlags();
@@ -46,41 +47,42 @@ async function main(): Promise<void> {
   console.log(`  CHART_SERIES_VERSION = ${CHART_SERIES_VERSION}`);
   console.log(`  force = ${flags.force}`);
   console.log(`  limit = ${flags.limit ?? 'none'}`);
+  console.log(`  shard = ${flags.shardIndex + 1}/${flags.shardCount}`);
 
-  // Only rows that actually have a server_metrics blob can produce a
-  // chart_series. Rows without the blob legitimately keep `chart_series`
-  // null and the API serves them via the slow path (which also returns
-  // null because there's no blob to parse — so the page falls into the
-  // "no stored trace_replay blob" branch).
-  const candidates = flags.force
-    ? await sql<{ id: number }[]>`
-        select id
-        from agentic_trace_replay
-        where server_metrics_json_gz is not null
-        order by id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `
-    : await sql<{ id: number }[]>`
-        select id
-        from agentic_trace_replay
-        where server_metrics_json_gz is not null
-          and (
-            chart_series is null
-            or coalesce((chart_series->>'version')::int, -1) <> ${CHART_SERIES_VERSION}
-          )
-        order by id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `;
-
-  if (candidates.length === 0) {
-    console.log('\n  Nothing to do — all rows up to date.');
-    return;
-  }
-
-  if (!(await confirmProceed(`${candidates.length} candidate row(s).`))) return;
-
-  await runPerIdBackfill(
-    candidates.map((c) => c.id),
+  await runCandidateIdBackfill(
+    async () => {
+      // Only rows that actually have a server_metrics blob can produce a
+      // chart_series. Rows without the blob legitimately keep `chart_series`
+      // null and the API serves them via the slow path (which also returns
+      // null because there's no blob to parse — so the page falls into the
+      // "no stored trace_replay blob" branch).
+      const candidates = flags.force
+        ? await sql<{ id: number }[]>`
+            select id
+            from agentic_trace_replay
+            where server_metrics_json_gz is not null
+              and mod(id, ${flags.shardCount}) = ${flags.shardIndex}
+            -- Restore the newest, most actively viewed runs first. The backfill is
+            -- idempotent, so an interrupted pass resumes with only stale rows.
+            order by id desc
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `
+        : await sql<{ id: number }[]>`
+            select id
+            from agentic_trace_replay
+            where server_metrics_json_gz is not null
+              and mod(id, ${flags.shardCount}) = ${flags.shardIndex}
+              and (
+                chart_series is null
+                or coalesce((chart_series->>'version')::int, -1) <> ${CHART_SERIES_VERSION}
+              )
+            -- Restore the newest, most actively viewed runs first. The backfill is
+            -- idempotent, so an interrupted pass resumes with only stale rows.
+            order by id desc
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `;
+      return candidates.map((candidate) => candidate.id);
+    },
     async (id) => {
       const [row] = await sql<
         {
diff --git a/packages/db/src/backfill-full-response-interactivity.ts b/packages/db/src/backfill-full-response-interactivity.ts
index 74a3cd8fb..6a59dfcfa 100644
--- a/packages/db/src/backfill-full-response-interactivity.ts
+++ b/packages/db/src/backfill-full-response-interactivity.ts
@@ -16,11 +16,10 @@ import { hasNoSslFlag } from './cli-utils.js';
 import { createAdminSql, refreshLatestBenchmarks } from './etl/db-utils.js';
 import { fullResponseMetricsFromGzip } from './etl/full-response-interactivity.js';
 import {
-  confirmProceed,
   jsonbParam,
   parseLimitForceFlags,
   runBackfillMain,
-  runPerIdBackfill,
+  runCandidateIdBackfill,
 } from './lib/backfill-runner.js';
 
 const flags = parseLimitForceFlags();
@@ -31,35 +30,30 @@ async function main(): Promise<void> {
   console.log(`  force = ${flags.force}`);
   console.log(`  limit = ${flags.limit ?? 'none'}`);
 
-  const candidates = flags.force
-    ? await sql<{ id: number }[]>`
-        select br.id
-        from benchmark_results br
-        join agentic_trace_replay atr on atr.id = br.trace_replay_id
-        where br.benchmark_type = 'agentic_traces'
-          and atr.profile_export_jsonl_gz is not null
-        order by br.id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `
-    : await sql<{ id: number }[]>`
-        select br.id
-        from benchmark_results br
-        join agentic_trace_replay atr on atr.id = br.trace_replay_id
-        where br.benchmark_type = 'agentic_traces'
-          and atr.profile_export_jsonl_gz is not null
-          and not (br.metrics ? 'median_full_response_itl')
-        order by br.id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `;
-
-  if (candidates.length === 0) {
-    console.log('\n  Nothing to do — all rows up to date.');
-    return;
-  }
-  if (!(await confirmProceed(`${candidates.length} candidate benchmark row(s).`))) return;
-
-  await runPerIdBackfill(
-    candidates.map((candidate) => candidate.id),
+  const processedCandidates = await runCandidateIdBackfill(
+    async () => {
+      const candidates = flags.force
+        ? await sql<{ id: number }[]>`
+            select br.id
+            from benchmark_results br
+            join agentic_trace_replay atr on atr.id = br.trace_replay_id
+            where br.benchmark_type = 'agentic_traces'
+              and atr.profile_export_jsonl_gz is not null
+            order by br.id
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `
+        : await sql<{ id: number }[]>`
+            select br.id
+            from benchmark_results br
+            join agentic_trace_replay atr on atr.id = br.trace_replay_id
+            where br.benchmark_type = 'agentic_traces'
+              and atr.profile_export_jsonl_gz is not null
+              and not (br.metrics ? 'median_full_response_itl')
+            order by br.id
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `;
+      return candidates.map((candidate) => candidate.id);
+    },
     async (id) => {
       const [row] = await sql<{ profile_export_jsonl_gz: Buffer | null }[]>`
         select atr.profile_export_jsonl_gz
@@ -85,9 +79,10 @@ async function main(): Promise<void> {
       `;
       return 'ok';
     },
+    (count) => `${count} candidate benchmark row(s).`,
   );
 
-  if (process.exitCode !== 1) await refreshLatestBenchmarks(sql);
+  if (processedCandidates && process.exitCode !== 1) await refreshLatestBenchmarks(sql);
 }
 
 runBackfillMain('backfill-full-response-interactivity', sql, main);
diff --git a/packages/db/src/backfill-request-timeline.ts b/packages/db/src/backfill-request-timeline.ts
index 580756338..2c0e1209a 100644
--- a/packages/db/src/backfill-request-timeline.ts
+++ b/packages/db/src/backfill-request-timeline.ts
@@ -17,78 +17,126 @@ import { hasNoSslFlag } from './cli-utils.js';
 import {
   REQUEST_TIMELINE_VERSION,
   computeRequestTimeline,
+  type RequestTimeline,
 } from './etl/compute-request-timeline.js';
 import { createAdminSql } from './etl/db-utils.js';
 import {
-  confirmProceed,
-  jsonbParam,
+  TRACE_REPLAY_UPLOAD_CHUNK_BYTES,
+  uploadTraceReplayPayloadChunks,
+} from './etl/trace-replay-ingest.js';
+import {
   parseLimitForceFlags,
   runBackfillMain,
-  runPerIdBackfill,
+  runCandidateIdBackfill,
 } from './lib/backfill-runner.js';
 
 const flags = parseLimitForceFlags();
 
+const PROFILE_DOWNLOAD_CHUNK_BYTES = TRACE_REPLAY_UPLOAD_CHUNK_BYTES;
+
 const sql = createAdminSql({
   noSsl: hasNoSslFlag(),
   max: 1,
   onnotice: () => {},
 });
 
-async function main(): Promise<void> {
-  console.log('=== backfill-request-timeline ===');
-  console.log(`  REQUEST_TIMELINE_VERSION = ${REQUEST_TIMELINE_VERSION}`);
-  console.log(`  force = ${flags.force}`);
-  console.log(`  limit = ${flags.limit ?? 'none'}`);
+/** Avoid Neon's encoded-response limit and postgres.js's large-value copies. */
+async function readProfileBlob(id: number): Promise<Buffer | null> {
+  const [sizeRow] = await sql<{ size: number | null }[]>`
+    select octet_length(profile_export_jsonl_gz)::int as size
+    from agentic_trace_replay
+    where id = ${id}
+  `;
+  const size = sizeRow?.size;
+  if (size === null || size === undefined) return null;
 
-  // Only rows with a profile_export blob can produce a timeline. Rows
-  // without the blob keep `request_timeline` null and the API serves them
-  // as "no timeline data".
-  const candidates = flags.force
-    ? await sql<{ id: number }[]>`
-        select id
-        from agentic_trace_replay
-        where profile_export_jsonl_gz is not null
-        order by id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `
-    : await sql<{ id: number }[]>`
-        select id
-        from agentic_trace_replay
-        where profile_export_jsonl_gz is not null
-          and (
-            request_timeline is null
-            or coalesce((request_timeline->>'version')::int, -1) <> ${REQUEST_TIMELINE_VERSION}
-          )
-        order by id
-        ${flags.limit ? sql`limit ${flags.limit}` : sql``}
-      `;
+  const chunks: Buffer[] = [];
+  for (let offset = 0; offset < size; offset += PROFILE_DOWNLOAD_CHUNK_BYTES) {
+    const [chunkRow] = await sql<{ data: Buffer }[]>`
+      select substring(
+        profile_export_jsonl_gz
+        from ${offset + 1}
+        for ${PROFILE_DOWNLOAD_CHUNK_BYTES}
+      ) as data
+      from agentic_trace_replay
+      where id = ${id}
+    `;
+    if (!chunkRow) return null;
+    chunks.push(chunkRow.data);
+  }
+  return Buffer.concat(chunks, size);
+}
 
-  if (candidates.length === 0) {
-    console.log('\n  Nothing to do — all rows up to date.');
+/** Upload the JSONB through the same bounded protocol used by trace ingest. */
+async function writeRequestTimeline(id: number, timeline: RequestTimeline | null): Promise<void> {
+  if (timeline === null) {
+    await sql`update agentic_trace_replay set request_timeline = null where id = ${id}`;
     return;
   }
 
-  if (!(await confirmProceed(`${candidates.length} candidate row(s).`))) return;
+  const payload = Buffer.from(JSON.stringify(timeline));
+  await sql.begin(async (tx) => {
+    await tx`
+      create temporary table trace_replay_upload_parts (
+        field text not null,
+        part integer not null,
+        data bytea not null,
+        primary key (field, part)
+      ) on commit drop
+    `;
+    await uploadTraceReplayPayloadChunks(tx, 'request_timeline', payload);
+    await tx`
+      update agentic_trace_replay
+      set request_timeline = (
+        select convert_from(string_agg(data, ''::bytea order by part), 'UTF8')::jsonb
+        from pg_temp.trace_replay_upload_parts
+        where field = 'request_timeline'
+      )
+      where id = ${id}
+    `;
+  });
+}
+
+async function main(): Promise<void> {
+  console.log('=== backfill-request-timeline ===');
+  console.log(`  REQUEST_TIMELINE_VERSION = ${REQUEST_TIMELINE_VERSION}`);
+  console.log(`  force = ${flags.force}`);
+  console.log(`  limit = ${flags.limit ?? 'none'}`);
 
-  await runPerIdBackfill(
-    candidates.map((c) => c.id),
+  await runCandidateIdBackfill(
+    async () => {
+      // Only rows with a profile_export blob can produce a timeline. Rows
+      // without the blob keep `request_timeline` null and the API serves them
+      // as "no timeline data".
+      const candidates = flags.force
+        ? await sql<{ id: number }[]>`
+            select id
+            from agentic_trace_replay
+            where profile_export_jsonl_gz is not null
+            order by id
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `
+        : await sql<{ id: number }[]>`
+            select id
+            from agentic_trace_replay
+            where profile_export_jsonl_gz is not null
+              and (
+                request_timeline is null
+                or coalesce((request_timeline->>'version')::int, -1) <> ${REQUEST_TIMELINE_VERSION}
+              )
+            order by id
+            ${flags.limit ? sql`limit ${flags.limit}` : sql``}
+          `;
+      return candidates.map((candidate) => candidate.id);
+    },
     async (id) => {
-      const [row] = await sql<{ profile_export_jsonl_gz: Buffer | null }[]>`
-        select profile_export_jsonl_gz
-        from agentic_trace_replay
-        where id = ${id}
-      `;
-      if (!row) {
+      const profileBlob = await readProfileBlob(id);
+      if (profileBlob === null) {
         console.warn(`  id=${id}: row vanished, skipping`);
         return 'skipped';
       }
-      const timeline = computeRequestTimeline(row.profile_export_jsonl_gz);
-      await sql`
-        update agentic_trace_replay
-        set request_timeline = ${timeline === null ? null : jsonbParam(sql, timeline)}
-        where id = ${id}
-      `;
+      const timeline = await computeRequestTimeline(profileBlob);
+      await writeRequestTimeline(id, timeline);
       return 'ok';
     },
   );
diff --git a/packages/db/src/backfill-server-log-files.ts b/packages/db/src/backfill-server-log-files.ts
new file mode 100644
index 000000000..8047e0db5
--- /dev/null
+++ b/packages/db/src/backfill-server-log-files.ts
@@ -0,0 +1,375 @@
+/**
+ * Backfill every .log/.out file from GitHub and the historical GCS backup.
+ *
+ * The script pairs server_logs_/multinode_server_logs_ artifacts with their
+ * bmk[_agentic]_ sibling, maps the raw benchmark rows through the production
+ * mapper, and attaches the log bundle to the exact persisted points. Inserts
+ * use filename conflicts plus files_complete metadata, so reruns are safe.
+ *
+ * Usage:
+ *   bun run --cwd packages/db db:backfill-server-log-files
+ *   bun run --cwd packages/db db:backfill-server-log-files --run 31415828111 --yes
+ *   bun run --cwd packages/db db:backfill-server-log-files --all --from-run 26606969606 --yes
+ *   bun run --cwd packages/db db:backfill-server-log-files --all --source gcs --dry-run
+ *   bun run --cwd packages/db db:backfill-server-log-files --all --source auto --yes
+ */
+
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+import { hasNoSslFlag } from './cli-utils.js';
+import { mapBenchmarkRow, type BenchmarkParams } from './etl/benchmark-mapper.js';
+import { insertServerLogFilePaths } from './etl/benchmark-ingest.js';
+import { createAdminSql } from './etl/db-utils.js';
+import { listServerLogFilePaths, serverLogArtifactRoot } from './etl/server-log-artifacts.js';
+import { createSkipTracker } from './etl/skip-tracker.js';
+import { downloadArtifact, listRunArtifacts } from './lib/github-artifacts.js';
+import {
+  downloadGcsArtifact,
+  listGcsServerLogArtifacts,
+  type GcsArtifactMeta,
+} from './lib/gcs-artifacts.js';
+import { confirmProceed, parseLimitForceFlags, runBackfillMain } from './lib/backfill-runner.js';
+import { retryArtifactOperation } from './lib/artifact-retry.js';
+import { repositoryFromRunUrl } from './lib/runtime-metadata-artifacts.js';
+import {
+  pairServerLogArtifacts,
+  resolveServerLogResultCandidates,
+} from './lib/server-log-backfill.js';
+
+const DEFAULT_REPO = 'SemiAnalysisAI/InferenceX';
+const RETENTION_DAYS = 90;
+const sql = createAdminSql({ noSsl: hasNoSslFlag(), max: 2, onnotice: () => {} });
+
+interface CandidateRun {
+  github_run_id: number;
+  run_attempt: number;
+  html_url: string | null;
+  date: string;
+}
+
+type ArtifactSource = 'auto' | 'github' | 'gcs';
+
+interface BackfillFlags {
+  all: boolean;
+  dryRun: boolean;
+  source: ArtifactSource;
+}
+
+function parseRunFilter(): number | null {
+  const index = process.argv.indexOf('--run');
+  if (index === -1) return null;
+  const raw = process.argv[index + 1];
+  if (!raw || !/^\d+$/u.test(raw) || Number(raw) <= 0) {
+    throw new Error('--run requires a positive GitHub Actions run ID');
+  }
+  return Number(raw);
+}
+
+function parseFromRunFilter(): number | null {
+  const index = process.argv.indexOf('--from-run');
+  if (index === -1) return null;
+  const raw = process.argv[index + 1];
+  if (!raw || !/^\d+$/u.test(raw) || Number(raw) <= 0) {
+    throw new Error('--from-run requires a positive GitHub Actions run ID');
+  }
+  return Number(raw);
+}
+
+function parseBackfillFlags(): BackfillFlags {
+  let source: ArtifactSource = 'auto';
+  for (let i = 2; i < process.argv.length; i++) {
+    if (process.argv[i] !== '--source') continue;
+    const value = process.argv[++i];
+    if (value !== 'auto' && value !== 'github' && value !== 'gcs') {
+      throw new Error('--source requires auto, github, or gcs');
+    }
+    source = value;
+  }
+  return {
+    all: process.argv.includes('--all'),
+    dryRun: process.argv.includes('--dry-run'),
+    source,
+  };
+}
+
+function findJsonFiles(root: string): string[] {
+  const files: string[] = [];
+  for (const entry of fs.readdirSync(root, { withFileTypes: true })) {
+    const pathname = path.join(root, entry.name);
+    if (entry.isDirectory()) files.push(...findJsonFiles(pathname));
+    else if (entry.isFile() && entry.name.endsWith('.json')) files.push(pathname);
+  }
+  return files.toSorted();
+}
+
+function readMappedRows(root: string): BenchmarkParams[] {
+  const tracker = createSkipTracker();
+  const rows: BenchmarkParams[] = [];
+  for (const file of findJsonFiles(root)) {
+    const parsed = JSON.parse(fs.readFileSync(file, 'utf8')) as unknown;
+    const rawRows = Array.isArray(parsed) ? parsed : [parsed];
+    for (const raw of rawRows) {
+      if (!raw || typeof raw !== 'object' || Array.isArray(raw)) continue;
+      const mapped = mapBenchmarkRow(raw as Record<string, unknown>, tracker);
+      if (mapped) rows.push(mapped);
+    }
+  }
+  return rows;
+}
+
+async function findBenchmarkResultIds(
+  run: CandidateRun,
+  rows: readonly BenchmarkParams[],
+): Promise<number[]> {
+  const ids = new Set<number>();
+  for (const row of rows) {
+    const c = row.config;
+    const candidates = await sql<{ id: number; offload_mode: string }[]>`
+      select br.id, br.offload_mode
+      from benchmark_results br
+      join workflow_runs wr on wr.id = br.workflow_run_id
+      join configs cfg on cfg.id = br.config_id
+      where wr.github_run_id = ${run.github_run_id}
+        and wr.run_attempt = ${run.run_attempt}
+        and cfg.hardware = ${c.hardware}
+        and cfg.framework = ${c.framework}
+        and cfg.model = ${c.model}
+        and cfg.precision = ${c.precision}
+        and cfg.spec_method = ${c.specMethod}
+        and cfg.disagg = ${c.disagg}
+        and cfg.is_multinode = ${c.isMultinode}
+        and cfg.prefill_tp = ${c.prefillTp}
+        and cfg.prefill_ep = ${c.prefillEp}
+        and cfg.prefill_dp_attention = ${c.prefillDpAttn}
+        and cfg.prefill_num_workers = ${c.prefillNumWorkers}
+        and cfg.decode_tp = ${c.decodeTp}
+        and cfg.decode_ep = ${c.decodeEp}
+        and cfg.decode_dp_attention = ${c.decodeDpAttn}
+        and cfg.decode_num_workers = ${c.decodeNumWorkers}
+        and cfg.num_prefill_gpu = ${c.numPrefillGpu}
+        and cfg.num_decode_gpu = ${c.numDecodeGpu}
+        and br.benchmark_type = ${row.benchmarkType}
+        and br.isl is not distinct from ${row.isl}
+        and br.osl is not distinct from ${row.osl}
+        and br.conc = ${row.conc}
+        and br.recipe_fingerprint is not distinct from ${row.recipeFingerprint}
+    `;
+    const resolution = resolveServerLogResultCandidates(
+      candidates.map((candidate) => ({
+        id: Number(candidate.id),
+        offloadMode: candidate.offload_mode,
+      })),
+      row.offloadMode,
+    );
+    if (resolution.usedUniqueFallback) {
+      console.warn(
+        `  [WARN] benchmark result ${resolution.ids[0]} uses a historical offload label; ` +
+          `matched uniquely without offload_mode`,
+      );
+    }
+    for (const id of resolution.ids) ids.add(id);
+  }
+  return [...ids];
+}
+
+async function resultLogsAreComplete(resultIds: readonly number[]): Promise<boolean> {
+  if (resultIds.length === 0) return false;
+  const [row] = await sql<{ complete: boolean }[]>`
+    select count(*) = ${resultIds.length} as complete
+    from benchmark_results br
+    join server_logs sl on sl.id = br.server_log_id
+    where br.id = any(${sql.array([...resultIds])}::bigint[])
+      and sl.files_complete
+  `;
+  return row?.complete ?? false;
+}
+
+function gcsCompressedBytes(pairs: readonly ReturnType<typeof pairServerLogArtifacts>[number][]) {
+  return pairs.reduce((total, pair) => total + ((pair.serverLogs as GcsArtifactMeta).size ?? 0), 0);
+}
+
+function isWithinGithubRetention(date: string, now = new Date()): boolean {
+  const cutoff = new Date(now);
+  cutoff.setUTCDate(cutoff.getUTCDate() - RETENTION_DAYS);
+  return new Date(`${date}T23:59:59Z`) >= cutoff;
+}
+
+async function main(): Promise<void> {
+  console.log('=== backfill-server-log-files ===');
+  const runFilter = parseRunFilter();
+  const fromRunFilter = parseFromRunFilter();
+  if (runFilter !== null && fromRunFilter !== null) {
+    throw new Error('--run and --from-run cannot be combined');
+  }
+  const flags = parseBackfillFlags();
+  const { limit } = parseLimitForceFlags();
+  const runs = await sql<CandidateRun[]>`
+    select distinct wr.github_run_id, wr.run_attempt, wr.html_url, wr.date::text as date
+    from workflow_runs wr
+    join benchmark_results br on br.workflow_run_id = wr.id
+    where (${runFilter}::bigint is null or wr.github_run_id = ${runFilter})
+      and (${fromRunFilter}::bigint is null or wr.github_run_id >= ${fromRunFilter})
+      and (
+        ${runFilter}::bigint is not null
+        or ${flags.all}
+        or wr.date >= current_date - ${RETENTION_DAYS}::integer
+      )
+    order by wr.github_run_id, wr.run_attempt
+    ${limit === null ? sql`` : sql`limit ${limit}`}
+  `;
+  if (runs.length === 0) {
+    console.log('  Nothing to do.');
+    return;
+  }
+
+  let gcsByRun = new Map<number, GcsArtifactMeta[]>();
+  if (flags.source !== 'github') {
+    process.stdout.write('  Indexing public GCS artifact backup...');
+    gcsByRun = await listGcsServerLogArtifacts();
+    console.log(` ${gcsByRun.size} run(s) indexed`);
+  }
+
+  if (flags.dryRun) {
+    let pairedRuns = 0;
+    let pairs = 0;
+    let compressedBytes = 0;
+    let missing = 0;
+    for (const run of runs) {
+      const runPairs = pairServerLogArtifacts(gcsByRun.get(Number(run.github_run_id)) ?? []);
+      if (runPairs.length === 0) missing++;
+      else pairedRuns++;
+      pairs += runPairs.length;
+      compressedBytes += gcsCompressedBytes(runPairs);
+    }
+    console.log(
+      `\n=== dry run: ${runs.length} DB run attempt(s), ${pairedRuns} with GCS pairs, ` +
+        `${pairs} pair(s), ${(compressedBytes / 2 ** 30).toFixed(2)} GiB compressed, ` +
+        `${missing} without GCS pairs ===`,
+    );
+    return;
+  }
+
+  if (!(await confirmProceed(`${runs.length} workflow run(s) will be checked for log files.`))) {
+    return;
+  }
+
+  let artifactsProcessed = 0;
+  let filesStored = 0;
+  let pointsLinked = 0;
+  let artifactFailures = 0;
+  let runFailures = 0;
+  let unmatchedArtifacts = 0;
+  let emptyArtifacts = 0;
+  let missingRuns = 0;
+  let completeSkipped = 0;
+  let gcsRuns = 0;
+  let githubRuns = 0;
+  for (const [runIndex, run] of runs.entries()) {
+    const runId = Number(run.github_run_id);
+    const repository = repositoryFromRunUrl(run.html_url) ?? DEFAULT_REPO;
+    const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), `server-log-backfill-${runId}-`));
+    try {
+      let source: Exclude<ArtifactSource, 'auto'> | null = null;
+      let pairs =
+        flags.source === 'github' ? [] : pairServerLogArtifacts(gcsByRun.get(runId) ?? []);
+      if (pairs.length > 0) {
+        source = 'gcs';
+        gcsRuns++;
+      } else if (
+        flags.source !== 'gcs' &&
+        (flags.source === 'github' || isWithinGithubRetention(run.date))
+      ) {
+        const artifacts = await retryArtifactOperation(
+          `listing GitHub artifacts for run ${runId}`,
+          () => listRunArtifacts(repository, String(runId)),
+        );
+        pairs = pairServerLogArtifacts(artifacts.filter((artifact) => !artifact.expired));
+        if (pairs.length > 0) {
+          source = 'github';
+          githubRuns++;
+        }
+      }
+      if (!source) {
+        missingRuns++;
+        console.log(
+          `  [${runIndex + 1}/${runs.length}] ${repository} run ${runId} attempt ${run.run_attempt}: no retained log pairs`,
+        );
+        continue;
+      }
+
+      for (const pair of pairs) {
+        let benchmarkDir: string | null = null;
+        let serverLogDir: string | null = null;
+        try {
+          benchmarkDir =
+            source === 'gcs'
+              ? await retryArtifactOperation(`downloading ${pair.benchmarks.name}`, () =>
+                  downloadGcsArtifact(pair.benchmarks as GcsArtifactMeta, tempDir),
+                )
+              : await retryArtifactOperation(`downloading ${pair.benchmarks.name}`, () =>
+                  downloadArtifact(pair.benchmarks, tempDir),
+                );
+          const mappedRows = readMappedRows(benchmarkDir);
+          const resultIds = await findBenchmarkResultIds(run, mappedRows);
+          if (resultIds.length === 0) {
+            unmatchedArtifacts++;
+            console.warn(`  [WARN] ${pair.serverLogs.name}: no matching benchmark rows`);
+            continue;
+          }
+          if (await resultLogsAreComplete(resultIds)) {
+            completeSkipped++;
+            continue;
+          }
+          serverLogDir =
+            source === 'gcs'
+              ? await retryArtifactOperation(`downloading ${pair.serverLogs.name}`, () =>
+                  downloadGcsArtifact(pair.serverLogs as GcsArtifactMeta, tempDir),
+                )
+              : await retryArtifactOperation(`downloading ${pair.serverLogs.name}`, () =>
+                  downloadArtifact(pair.serverLogs, tempDir),
+                );
+          const root = serverLogArtifactRoot(serverLogDir, pair.serverLogs.name);
+          const logFiles = root ? listServerLogFilePaths(root) : [];
+          if (logFiles.length === 0) {
+            emptyArtifacts++;
+            console.warn(`  [WARN] ${pair.serverLogs.name}: no .log/.out files`);
+            continue;
+          }
+          await insertServerLogFilePaths(sql, resultIds, logFiles);
+          artifactsProcessed++;
+          filesStored += logFiles.length;
+          pointsLinked += resultIds.length;
+        } catch (error) {
+          artifactFailures++;
+          console.error(`  ✗ ${repository} run ${runId} artifact ${pair.serverLogs.name}:`, error);
+        } finally {
+          if (benchmarkDir) fs.rmSync(benchmarkDir, { recursive: true, force: true });
+          if (serverLogDir) fs.rmSync(serverLogDir, { recursive: true, force: true });
+        }
+      }
+      console.log(
+        `  [${runIndex + 1}/${runs.length}] ${repository} run ${runId} attempt ${run.run_attempt}: ` +
+          `${pairs.length} ${source} pair(s)`,
+      );
+    } catch (error) {
+      runFailures++;
+      console.error(`  ✗ ${repository} run ${runId}:`, error);
+    } finally {
+      fs.rmSync(tempDir, { recursive: true, force: true });
+    }
+  }
+
+  console.log(
+    `\n=== backfill complete: ${artifactsProcessed} artifact(s), ${filesStored} file(s), ` +
+      `${pointsLinked} point link(s), ${completeSkipped} complete artifact(s) skipped, ` +
+      `${unmatchedArtifacts} unmatched artifact(s), ${emptyArtifacts} empty artifact(s), ` +
+      `${gcsRuns} GCS run(s), ${githubRuns} GitHub run(s), ${missingRuns} unavailable run(s), ` +
+      `${artifactFailures} failed artifact(s), ${runFailures} failed run(s) ===`,
+  );
+  console.log('  Invalidate API cache after the backfill: bun run admin:cache:invalidate');
+  if (artifactFailures > 0 || runFailures > 0) process.exitCode = 1;
+}
+
+runBackfillMain('backfill-server-log-files', sql, main);
diff --git a/packages/db/src/connection.test.ts b/packages/db/src/connection.test.ts
index 1e7c38ae6..3898177f9 100644
--- a/packages/db/src/connection.test.ts
+++ b/packages/db/src/connection.test.ts
@@ -1,64 +1,122 @@
-import { describe, expect, it } from 'vitest';
+import { afterEach, describe, expect, it, vi } from 'vitest';
 
-import { postgresOptionsForUrl, shouldUseNeon } from './connection';
+import { getCollectiveXDb, getDb, resolveDatabaseConnection } from './connection';
 
-describe('shouldUseNeon', () => {
-  it('defaults to Neon for neon.tech URLs', () => {
-    expect(shouldUseNeon('postgres://user:pass@ep-test-123.us-east-1.aws.neon.tech/db')).toBe(true);
+afterEach(() => {
+  vi.unstubAllEnvs();
+});
+
+describe('resolveDatabaseConnection', () => {
+  it('resolves an environment URL and Neon HTTP defaults', () => {
+    vi.stubEnv('TEST_DATABASE_URL', 'postgres://user:pass@ep-test-123.us-east-1.aws.neon.tech/app');
+
+    expect(resolveDatabaseConnection({ envVar: 'TEST_DATABASE_URL' })).toEqual({
+      url: 'postgres://user:pass@ep-test-123.us-east-1.aws.neon.tech/app',
+      driver: 'neon',
+      ssl: 'require',
+    });
   });
 
-  it('defaults to postgres.js for non-Neon URLs', () => {
-    expect(shouldUseNeon('postgres://user:pass@db.example.com/app')).toBe(false);
+  it('uses an explicit URL instead of the configured environment URL', () => {
+    vi.stubEnv('TEST_DATABASE_URL', 'postgres://user:pass@db.example.com/app');
+
+    expect(
+      resolveDatabaseConnection({
+        envVar: 'TEST_DATABASE_URL',
+        url: 'postgres://postgres:postgres@localhost:5432/local',
+        driver: 'postgres',
+      }),
+    ).toEqual({
+      url: 'postgres://postgres:postgres@localhost:5432/local',
+      driver: 'postgres',
+      ssl: false,
+    });
   });
 
-  it('honors DATABASE_DRIVER=postgres override', () => {
+  it('defaults remote non-Neon URLs to postgres.js with required TLS', () => {
     expect(
-      shouldUseNeon('postgres://user:pass@ep-test-123.us-east-1.aws.neon.tech/db', 'postgres'),
-    ).toBe(false);
+      resolveDatabaseConnection({
+        envVar: 'UNUSED',
+        url: 'postgres://user:pass@db.example.com/app',
+      }),
+    ).toMatchObject({ driver: 'postgres', ssl: 'require' });
   });
 
-  it('honors DATABASE_DRIVER=neon override', () => {
-    expect(shouldUseNeon('postgres://user:pass@db.example.com/app', 'neon')).toBe(true);
+  it('allows each driver to be selected explicitly', () => {
+    const neonUrl = 'postgres://user:pass@ep-test-123.us-east-1.aws.neon.tech/app';
+    const remoteUrl = 'postgres://user:pass@db.example.com/app';
+
+    expect(
+      resolveDatabaseConnection({
+        envVar: 'UNUSED',
+        url: neonUrl,
+        driver: 'postgres',
+      }).driver,
+    ).toBe('postgres');
+    expect(
+      resolveDatabaseConnection({
+        envVar: 'UNUSED',
+        url: remoteUrl,
+        driver: 'neon',
+      }).driver,
+    ).toBe('neon');
   });
-});
 
-describe('postgresOptionsForUrl', () => {
-  it('keeps TLS enabled for remote non-Neon hosts', () => {
-    expect(postgresOptionsForUrl('postgres://user:pass@db.example.com/app')).toEqual({
-      max: 5,
-      ssl: 'require',
-    });
+  it('disables TLS for every loopback URL', () => {
+    for (const url of [
+      'postgres://user:pass@localhost:5432/app',
+      'postgres://user:pass@127.0.0.1:5432/app',
+      'postgres://user:pass@[::1]:5432/app',
+    ]) {
+      expect(resolveDatabaseConnection({ envVar: 'UNUSED', url }).ssl).toBe(false);
+    }
   });
 
-  it('disables TLS for localhost', () => {
-    expect(postgresOptionsForUrl('postgres://user:pass@localhost:5432/app')).toEqual({
-      max: 5,
-      ssl: false,
-    });
+  it('requires TLS for remote hostnames inherited from Object.prototype', () => {
+    for (const hostname of ['constructor', 'toString', '__proto__']) {
+      expect(
+        resolveDatabaseConnection({
+          envVar: 'UNUSED',
+          url: `postgres://user:pass@${hostname}/app`,
+        }).ssl,
+      ).toBe('require');
+    }
   });
 
-  it('disables TLS for loopback IPv4 and IPv6 URLs', () => {
-    expect(postgresOptionsForUrl('postgres://user:pass@127.0.0.1:5432/app')).toEqual({
-      max: 5,
-      ssl: false,
-    });
-    expect(postgresOptionsForUrl('postgres://user:pass@[::1]:5432/app')).toEqual({
-      max: 5,
-      ssl: false,
-    });
+  it('lets explicit TLS inputs override hostname defaults', () => {
+    expect(
+      resolveDatabaseConnection({
+        envVar: 'UNUSED',
+        url: 'postgres://user:pass@db.example.com/app',
+        ssl: false,
+      }).ssl,
+    ).toBe(false);
+    expect(
+      resolveDatabaseConnection({
+        envVar: 'UNUSED',
+        url: 'postgres://user:pass@localhost:5432/app',
+        ssl: 'true',
+      }).ssl,
+    ).toBe('require');
   });
 
-  it('honors DATABASE_SSL=false override for remote hosts', () => {
-    expect(postgresOptionsForUrl('postgres://user:pass@db.example.com/app', 'false')).toEqual({
-      max: 5,
-      ssl: false,
+  it('memoizes clients by effective connection identity rather than environment name', () => {
+    const url = 'postgres://user:pass@ep-memo-test.us-east-1.aws.neon.tech/app';
+    const first = getDb({ url, driver: 'neon' });
+    const sameConnectionFromAnotherClient = getCollectiveXDb({ url, driver: 'neon' });
+    const otherUrl = getDb({
+      url: 'postgres://user:pass@ep-memo-other.us-east-1.aws.neon.tech/app',
+      driver: 'neon',
     });
+
+    expect(sameConnectionFromAnotherClient).toBe(first);
+    expect(otherUrl).not.toBe(first);
   });
 
-  it('honors DATABASE_SSL=true override for loopback hosts', () => {
-    expect(postgresOptionsForUrl('postgres://user:pass@localhost:5432/app', 'true')).toEqual({
-      max: 5,
-      ssl: 'require',
-    });
+  it('reports the selected environment variable when no URL is configured', () => {
+    vi.stubEnv('MISSING_DATABASE_URL', '');
+    expect(() => resolveDatabaseConnection({ envVar: 'MISSING_DATABASE_URL' })).toThrow(
+      'MISSING_DATABASE_URL is not set',
+    );
   });
 });
diff --git a/packages/db/src/connection.ts b/packages/db/src/connection.ts
index 23721d10b..53ac0381a 100644
--- a/packages/db/src/connection.ts
+++ b/packages/db/src/connection.ts
@@ -24,9 +24,24 @@ export const FIXTURES_MODE = process.env.E2E_FIXTURES === '1';
 
 const LOOPBACK_HOSTS = new Set(['localhost', '127.0.0.1', '::1']);
 
-interface PostgresConnectionOptions {
-  max: number;
-  ssl: false | 'require';
+export type DatabaseDriver = 'neon' | 'postgres';
+export type DatabaseSsl = false | 'require';
+
+export interface DatabaseConnectionResolution {
+  url: string;
+  driver: DatabaseDriver;
+  ssl: DatabaseSsl;
+}
+
+export interface DatabaseConnectionOptions {
+  /** Environment variable used when no explicit URL is supplied. */
+  envVar: string;
+  /** Explicit connection URL, primarily for programmatic and test callers. */
+  url?: string;
+  /** Explicit driver, or an environment/default input used to select one. */
+  driver?: string;
+  /** Explicit TLS input. `false` always wins over hostname-based defaults. */
+  ssl?: string | boolean;
 }
 
 function getDbHostname(url: string): string | null {
@@ -37,35 +52,40 @@ function getDbHostname(url: string): string | null {
   }
 }
 
-/**
- * DATABASE_DRIVER=neon  → @neondatabase/serverless HTTP driver (default for *.neon.tech URLs)
- * DATABASE_DRIVER=postgres → postgres.js TCP driver  (default for everything else)
- */
-export function shouldUseNeon(url: string, driver = process.env.DATABASE_DRIVER): boolean {
-  const normalizedDriver = driver?.toLowerCase();
-  const hostname = getDbHostname(url);
+function resolveDriver(url: string, hostname: string | null, input?: string): DatabaseDriver {
+  const normalized = input?.toLowerCase();
+  if (normalized === 'postgres') return 'postgres';
+  if (normalized === 'neon') return 'neon';
+  const useNeon = hostname?.endsWith('.neon.tech') ?? url.includes('.neon.tech');
+  return useNeon ? 'neon' : 'postgres';
+}
 
-  if (normalizedDriver === 'postgres') return false;
-  if (normalizedDriver === 'neon') return true;
-  return hostname?.endsWith('.neon.tech') ?? url.includes('.neon.tech');
+function resolveSsl(hostname: string | null, input?: string | boolean): DatabaseSsl {
+  const normalized = typeof input === 'string' ? input.toLowerCase() : input;
+  if (normalized === false || normalized === 'false') return false;
+  if (normalized === true || normalized === 'true' || normalized === 'require') return 'require';
+  return hostname && LOOPBACK_HOSTS.has(hostname) ? false : 'require';
 }
 
 /**
- * DATABASE_SSL=false disables TLS unconditionally.
- * Otherwise: loopback → no TLS, remote → TLS required.
+ * Resolve the URL, driver, and TLS policy shared by API, admin, and MCP clients.
+ * Explicit inputs take precedence. Otherwise Neon hosts use the HTTP driver,
+ * loopback postgres.js connections disable TLS, and remote connections require it.
  */
-export function postgresOptionsForUrl(
-  url: string,
-  sslEnv = process.env.DATABASE_SSL,
-): PostgresConnectionOptions {
-  const ssl = sslEnv?.toLowerCase();
-  if (ssl === 'false') return { max: 5, ssl: false };
-  if (ssl === 'true') return { max: 5, ssl: 'require' };
+export function resolveDatabaseConnection({
+  envVar,
+  url: urlOverride,
+  driver: driverInput,
+  ssl: sslInput,
+}: DatabaseConnectionOptions): DatabaseConnectionResolution {
+  const url = urlOverride ?? process.env[envVar];
+  if (!url) throw new Error(`${envVar} is not set`);
+
   const hostname = getDbHostname(url);
-  return {
-    max: 5,
-    ssl: hostname && LOOPBACK_HOSTS.has(hostname) ? false : 'require',
-  };
+  const driver = resolveDriver(url, hostname, driverInput);
+  const ssl = resolveSsl(hostname, sslInput);
+
+  return { url, driver, ssl };
 }
 
 /** Wrap postgres.js Sql instance to match DbClient signature. */
@@ -78,45 +98,57 @@ function wrapPostgres(sql: postgres.Sql): DbClient {
 // hot reload, leaking the previous postgres.js TCP connection pool.
 const g = globalThis as unknown as { __dbClients?: Map<string, DbClient> };
 
-function makeDbClient(url: string): DbClient {
-  return shouldUseNeon(url)
-    ? (neon(url) as DbClient)
-    : wrapPostgres(postgres(url, postgresOptionsForUrl(url)));
+function makeDbClient(connection: DatabaseConnectionResolution): DbClient {
+  return connection.driver === 'neon'
+    ? (neon(connection.url) as DbClient)
+    : wrapPostgres(postgres(connection.url, { max: 5, ssl: connection.ssl }));
 }
 
-/** One memoized client per connection env var; throws when the var is unset. */
-function memoizedClient(envVar: string): DbClient {
+export interface DatabaseClientOverrides {
+  url?: string;
+  driver?: DatabaseDriver;
+  ssl?: DatabaseSsl;
+}
+
+/** One memoized client per effective connection identity. */
+function memoizedClient(envVar: string, overrides: DatabaseClientOverrides = {}): DbClient {
+  const connection = resolveDatabaseConnection({
+    envVar,
+    url: overrides.url,
+    driver: overrides.driver ?? process.env.DATABASE_DRIVER,
+    ssl: overrides.ssl ?? process.env.DATABASE_SSL,
+  });
+  const key =
+    connection.driver === 'neon'
+      ? `neon\0${connection.url}`
+      : `postgres\0${connection.ssl}\0${connection.url}`;
   g.__dbClients ??= new Map();
-  const cached = g.__dbClients.get(envVar);
+  const cached = g.__dbClients.get(key);
   if (cached) return cached;
-  const url = process.env[envVar];
-  if (!url) throw new Error(`${envVar} is not set`);
-  const client = makeDbClient(url);
-  g.__dbClients.set(envVar, client);
+  const client = makeDbClient(connection);
+  g.__dbClients.set(key, client);
   return client;
 }
 
 /** Read-only SQL client for API routes. Requires DATABASE_READONLY_URL. */
-export function getDb(): DbClient {
-  return memoizedClient('DATABASE_READONLY_URL');
+export function getDb(overrides?: DatabaseClientOverrides): DbClient {
+  return memoizedClient('DATABASE_READONLY_URL', overrides);
 }
 
 /** Write-capable SQL client for API routes that need to insert (e.g. user feedback). */
-export function getWriteDb(): DbClient {
-  return memoizedClient('DATABASE_WRITE_URL');
+export function getWriteDb(overrides?: DatabaseClientOverrides): DbClient {
+  return memoizedClient('DATABASE_WRITE_URL', overrides);
 }
 
 /**
- * Read-only SQL client for the CollectiveX database — a separate Neon
- * instance from the main benchmark DB, holding raw sweep-run documents.
- * Must point at the same primary as the write URL (not a lagging replica):
- * the lazy-ingest routes read their own writes within a single request.
+ * Ordinary read client for the separate CollectiveX database. Lazy-ingest
+ * paths use the write client for state checks and post-write reads.
  */
-export function getCollectiveXDb(): DbClient {
-  return memoizedClient('DATABASE_COLLECTIVEX_READONLY_URL');
+export function getCollectiveXDb(overrides?: DatabaseClientOverrides): DbClient {
+  return memoizedClient('DATABASE_COLLECTIVEX_READONLY_URL', overrides);
 }
 
-/** Write-capable SQL client for the CollectiveX database (lazy ingest + run deletion). */
-export function getCollectiveXWriteDb(): DbClient {
-  return memoizedClient('DATABASE_COLLECTIVEX_WRITE_URL');
+/** CollectiveX primary client for lazy ingest, consistent reads, and deletion. */
+export function getCollectiveXWriteDb(overrides?: DatabaseClientOverrides): DbClient {
+  return memoizedClient('DATABASE_COLLECTIVEX_WRITE_URL', overrides);
 }
diff --git a/packages/db/src/etl/benchmark-ingest.test.ts b/packages/db/src/etl/benchmark-ingest.test.ts
new file mode 100644
index 000000000..d1a1db6a3
--- /dev/null
+++ b/packages/db/src/etl/benchmark-ingest.test.ts
@@ -0,0 +1,114 @@
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+import { describe, expect, it, vi } from 'vitest';
+
+import type { Sql } from './db-utils';
+import {
+  benchmarkPointIngestKey,
+  insertServerLogFilePaths,
+  insertServerLogFiles,
+} from './benchmark-ingest';
+
+const point = (recipeFingerprint: string | null) => ({
+  configId: 7,
+  benchmarkType: 'single_turn' as const,
+  isl: 8192,
+  osl: 1024,
+  conc: 12,
+  offloadMode: 'off',
+  recipeFingerprint,
+});
+
+describe('benchmarkPointIngestKey', () => {
+  it('keeps recipes at the same config and concurrency distinct', () => {
+    expect(
+      benchmarkPointIngestKey(
+        point('aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa'),
+      ),
+    ).not.toBe(
+      benchmarkPointIngestKey(
+        point('bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb'),
+      ),
+    );
+  });
+
+  it('keeps one stable legacy identity for null fingerprints', () => {
+    expect(benchmarkPointIngestKey(point(null))).toBe(benchmarkPointIngestKey(point(null)));
+  });
+});
+
+function fakeTransactionSql(linkedId: number | null) {
+  const calls: { text: string; values: unknown[] }[] = [];
+  const tag = vi.fn((strings: TemplateStringsArray, ...values: unknown[]) => {
+    const text = strings.join('');
+    calls.push({ text, values });
+    if (text.includes('select id, server_log_id')) {
+      return Promise.resolve([{ id: 42, server_log_id: linkedId }]);
+    }
+    if (text.includes('insert into server_logs')) return Promise.resolve([{ id: 99 }]);
+    return Promise.resolve([]);
+  }) as any;
+  tag.array = (value: unknown) => value;
+  tag.begin = (fn: (tx: typeof tag) => Promise<void>) => fn(tag);
+  return { sql: tag as Sql, calls };
+}
+
+describe('insertServerLogFiles', () => {
+  const files = [
+    { fileName: 'results/benchmark.log', logText: 'benchmark' },
+    { fileName: 'results/router.log', logText: 'router' },
+    { fileName: 'results/server.log', logText: 'server' },
+  ];
+
+  it('stores server.log as primary and every other filename as a child row', async () => {
+    const { sql, calls } = fakeTransactionSql(null);
+    await insertServerLogFiles(sql, [42], files);
+
+    const primaryInsert = calls.find((call) => call.text.includes('insert into server_logs'));
+    expect(primaryInsert?.values).toEqual(['server', 'results/server.log']);
+    const childNames = calls
+      .filter((call) => call.text.includes('insert into server_log_files'))
+      .map((call) => call.values[1]);
+    expect(childNames).toEqual(['results/benchmark.log', 'results/router.log']);
+    expect(calls.some((call) => call.text.includes('set server_log_id'))).toBe(true);
+  });
+
+  it('idempotently adds missing filenames to an existing linked bundle', async () => {
+    const { sql, calls } = fakeTransactionSql(7);
+    await insertServerLogFiles(sql, [42], files);
+
+    expect(calls.filter((call) => call.text.includes('insert into server_logs'))).toHaveLength(0);
+    expect(calls.filter((call) => call.text.includes('insert into server_log_files'))).toHaveLength(
+      2,
+    );
+    expect(calls.some((call) => call.text.includes('files_complete = true'))).toBe(true);
+  });
+
+  it('reads archived file paths lazily and removes null bytes', async () => {
+    const root = fs.mkdtempSync(path.join(os.tmpdir(), 'server-log-ingest-test-'));
+    const serverPath = path.join(root, 'server.log');
+    const routerPath = path.join(root, 'router.log');
+    fs.writeFileSync(serverPath, 'server\u0000');
+    fs.writeFileSync(routerPath, 'router\u0000');
+    try {
+      const { sql, calls } = fakeTransactionSql(null);
+      await insertServerLogFilePaths(
+        sql,
+        [42],
+        [
+          { fileName: 'router.log', path: routerPath },
+          { fileName: 'server.log', path: serverPath },
+        ],
+      );
+
+      const primaryInsert = calls.find((call) => call.text.includes('insert into server_logs'));
+      expect(primaryInsert?.values).toEqual(['server', 'server.log']);
+      const childInsert = calls.find((call) => call.text.includes('insert into server_log_files'));
+      expect(childInsert?.values).toEqual([99, 'router.log', 'router']);
+    } finally {
+      fs.rmSync(root, { recursive: true, force: true });
+    }
+  });
+});
diff --git a/packages/db/src/etl/benchmark-ingest.ts b/packages/db/src/etl/benchmark-ingest.ts
index 2a2382c83..617c51efe 100644
--- a/packages/db/src/etl/benchmark-ingest.ts
+++ b/packages/db/src/etl/benchmark-ingest.ts
@@ -2,27 +2,63 @@
  * Bulk DB insert functions for `benchmark_results` and `run_stats`.
  */
 
+import fs from 'node:fs';
+import path from 'node:path';
+
 import type postgres from 'postgres';
-import type { BenchmarkParams } from './benchmark-mapper';
+import { cleanLogText, type ServerLogFile, type ServerLogFilePath } from './server-log-artifacts';
+import type { BenchmarkType, WorkerPower } from './benchmark-mapper';
 import { kvCachePoolTokensFromServerLog } from './server-log-metrics';
 
 type Sql = ReturnType<typeof postgres>;
 
+export interface BenchmarkPersistenceInput {
+  configId: number;
+  benchmarkType: BenchmarkType;
+  isl: number | null;
+  osl: number | null;
+  conc: number;
+  offloadMode: string;
+  image: string | null;
+  recipeFingerprint: string | null;
+  metrics: Record<string, number>;
+  workers?: WorkerPower[];
+}
+
+type BenchmarkPointIdentity = Pick<
+  BenchmarkPersistenceInput,
+  'configId' | 'benchmarkType' | 'isl' | 'osl' | 'conc' | 'offloadMode' | 'recipeFingerprint'
+>;
+
+/** Stable in-batch identity matching benchmark_results_unique. */
+export function benchmarkPointIngestKey(row: BenchmarkPointIdentity): string {
+  return JSON.stringify([
+    row.configId,
+    row.benchmarkType,
+    row.isl,
+    row.osl,
+    row.conc,
+    row.offloadMode,
+    row.recipeFingerprint,
+  ]);
+}
+
 /**
  * Bulk-insert benchmark results for a single artifact in one DB round-trip using `UNNEST`.
- * Rows are deduplicated within the batch on the conflict key `(config_id, isl, osl, conc)`
- * before sending, because Postgres rejects an `ON CONFLICT DO UPDATE` statement that
+ * Rows are deduplicated within the batch on the persisted point identity, including
+ * the producer's recipe fingerprint when present, before sending because Postgres
+ * rejects an `ON CONFLICT DO UPDATE` statement that
  * would update the same row twice in a single query.
  *
  * @param sql - Active `postgres` connection.
- * @param rows - Mapped benchmark rows with their resolved `configId`.
+ * @param rows - Benchmark persistence fields with their resolved `configId`.
  * @param workflowRunId - DB id of the parent `workflow_runs` row.
  * @param date - ISO date string (`YYYY-MM-DD`) for the `date` column.
  * @returns Counts of newly inserted rows and rows that hit the conflict path.
  */
 export async function bulkIngestBenchmarkRows(
   sql: Sql,
-  rows: (BenchmarkParams & { configId: number })[],
+  rows: BenchmarkPersistenceInput[],
   workflowRunId: number,
   date: string,
 ): Promise<{ newCount: number; dupCount: number; insertedIds: number[] }> {
@@ -30,13 +66,10 @@ export async function bulkIngestBenchmarkRows(
 
   // Postgres rejects ON CONFLICT DO UPDATE if the same conflict key appears
   // more than once in a single batch. Deduplicate within the batch, keeping
-  // the last occurrence (last metrics for each unique config/benchmark_type/isl/osl/conc/offload_mode).
-  const seen = new Map<string, BenchmarkParams & { configId: number }>();
+  // the last occurrence for each unique recipe/config/scenario/concurrency point.
+  const seen = new Map<string, BenchmarkPersistenceInput>();
   for (const r of rows) {
-    seen.set(
-      `${r.configId}-${r.benchmarkType}-${r.isl ?? ''}-${r.osl ?? ''}-${r.conc}-${r.offloadMode}`,
-      r,
-    );
+    seen.set(benchmarkPointIngestKey(r), r);
   }
   const deduped = [...seen.values()];
 
@@ -47,6 +80,7 @@ export async function bulkIngestBenchmarkRows(
   const osls = deduped.map((r) => r.osl);
   const concs = deduped.map((r) => r.conc);
   const images = deduped.map((r) => r.image);
+  const recipeFingerprints = deduped.map((r) => r.recipeFingerprint);
   const metricsJsons = deduped.map((r) => JSON.stringify(r.metrics));
   // workers is optional — encode missing values as JSON null so the JSONB
   // unnest input has a homogeneous type (jsonb[]) and stores SQL NULL in the
@@ -58,7 +92,7 @@ export async function bulkIngestBenchmarkRows(
   const result = await sql<{ inserted: boolean; id: number }[]>`
     insert into benchmark_results (
       workflow_run_id, config_id, benchmark_type, offload_mode, date,
-      isl, osl, conc, image, metrics, workers
+      isl, osl, conc, image, recipe_fingerprint, metrics, workers
     )
     select
       ${workflowRunId},
@@ -70,9 +104,13 @@ export async function bulkIngestBenchmarkRows(
       unnest(${sql.array(osls)}::int[]),
       unnest(${sql.array(concs)}::int[]),
       unnest(${sql.array(images)}),
+      unnest(${sql.array(recipeFingerprints)}),
       unnest(${sql.array(metricsJsons)}::jsonb[]),
       unnest(${sql.array(workersJsons)}::jsonb[])
-    on conflict (workflow_run_id, config_id, benchmark_type, isl, osl, conc, offload_mode)
+    on conflict (
+      workflow_run_id, config_id, benchmark_type, isl, osl, conc, offload_mode,
+      recipe_fingerprint
+    )
     do update set
       -- Replace metrics with the fresh artifact values, but carry over
       -- kv_cache_pool_tokens: it is derived from the server log at
@@ -91,42 +129,145 @@ export async function bulkIngestBenchmarkRows(
 }
 
 /**
- * Insert a server log once and link it to the given benchmark result IDs.
- * Idempotent: skips rows that already have a server_log_id set.
+ * Store every .log/.out file from one server-log artifact and link the bundle
+ * to the given benchmark result IDs. Existing bundles receive only missing
+ * filenames, making both normal ingest and the historical backfill idempotent.
  */
-export async function insertServerLog(
+interface DeferredServerLogFile {
+  fileName: string;
+  readText: () => string;
+}
+
+function primaryDeferredServerLogFile(
+  files: readonly DeferredServerLogFile[],
+): DeferredServerLogFile | null {
+  return (
+    files.find((file) => path.posix.basename(file.fileName).toLowerCase() === 'server.log') ??
+    files[0] ??
+    null
+  );
+}
+
+async function insertDeferredServerLogFiles(
   sql: Sql,
   benchmarkResultIds: number[],
-  serverLog: string,
+  files: readonly DeferredServerLogFile[],
 ): Promise<void> {
-  if (benchmarkResultIds.length === 0) return;
+  if (benchmarkResultIds.length === 0 || files.length === 0) return;
 
-  // Only link rows that don't already have a server log
-  const unlinked = await sql<{ id: number }[]>`
-    select id from benchmark_results
-    where id = any(${sql.array(benchmarkResultIds)}::bigint[])
-      and server_log_id is null
-  `;
-  if (unlinked.length === 0) return;
+  const deduped = [...new Map(files.map((file) => [file.fileName, file])).values()];
+  const primary = primaryDeferredServerLogFile(deduped);
+  if (!primary) return;
+  const additional = deduped.filter((file) => file.fileName !== primary.fileName);
+  const primaryText = primary.readText();
+  const serverLog =
+    primary.fileName.toLowerCase().endsWith('/server.log') ||
+    primary.fileName.toLowerCase() === 'server.log'
+      ? primaryText
+      : null;
+  const kvCachePoolTokens = serverLog ? kvCachePoolTokensFromServerLog(serverLog) : null;
 
-  const [{ id: logId }] = await sql<{ id: number }[]>`
-    insert into server_logs (server_log) values (${serverLog})
-    returning id
-  `;
-  // Derive the KV-cache pool size (tokens) from the log's authoritative
-  // "GPU KV cache size: N tokens" line(s) and stash it on the result's metrics
-  // JSON, mirroring how trace-replay-ingest derives cache-hit rates. The
-  // scraped vllm:cache_config_info metric can't reconstruct this for MLA models.
-  const kvCachePoolTokens = kvCachePoolTokensFromServerLog(serverLog);
-  await sql`
-    update benchmark_results
-    set server_log_id = ${logId}${
-      kvCachePoolTokens === null
-        ? sql``
-        : sql`, metrics = jsonb_set(metrics, '{kv_cache_pool_tokens}', to_jsonb(${kvCachePoolTokens}::bigint))`
+  await sql.begin(async (tx) => {
+    const rows = await tx<{ id: number; server_log_id: number | null }[]>`
+      select id, server_log_id from benchmark_results
+      where id = any(${tx.array(benchmarkResultIds)}::bigint[])
+      for update
+    `;
+    const unlinked = rows.filter((row) => row.server_log_id === null);
+    const bundleIds = new Set(
+      rows.flatMap((row) => (row.server_log_id === null ? [] : [Number(row.server_log_id)])),
+    );
+
+    if (unlinked.length > 0) {
+      const [{ id: logId }] = await tx<{ id: number }[]>`
+        insert into server_logs (server_log, file_name, files_complete)
+        values (${primaryText}, ${primary.fileName}, true)
+        returning id
+      `;
+      bundleIds.add(Number(logId));
+      await tx`
+        update benchmark_results
+        set server_log_id = ${logId}
+        where id = any(${tx.array(unlinked.map((row) => row.id))}::bigint[])
+      `;
     }
-    where id = any(${sql.array(unlinked.map((r) => r.id))}::bigint[])
-  `;
+
+    for (const logId of bundleIds) {
+      // Legacy rows used the synthetic name server.log. Backfill upgrades that
+      // label to the artifact-relative path without rewriting the large text.
+      await tx`
+        update server_logs
+        set file_name = case
+              when file_name = 'server.log' then ${primary.fileName}
+              else file_name
+            end,
+            files_complete = true
+        where id = ${logId}
+      `;
+      for (const file of additional) {
+        const logText = file.readText();
+        await tx`
+          insert into server_log_files (server_log_id, file_name, log_text)
+          values (${logId}, ${file.fileName}, ${logText})
+          on conflict (server_log_id, file_name) do nothing
+        `;
+      }
+    }
+
+    // Derive the KV-cache pool size (tokens) from the authoritative server.log
+    // when the artifact includes one. Multinode bundles without server.log are
+    // still stored in full; they simply cannot contribute this derived metric.
+    if (kvCachePoolTokens !== null) {
+      await tx`
+        update benchmark_results
+        set metrics = jsonb_set(
+          metrics,
+          '{kv_cache_pool_tokens}',
+          to_jsonb(${kvCachePoolTokens}::bigint)
+        )
+        where id = any(${tx.array(rows.map((row) => row.id))}::bigint[])
+      `;
+    }
+  });
+}
+
+export async function insertServerLogFiles(
+  sql: Sql,
+  benchmarkResultIds: number[],
+  files: readonly ServerLogFile[],
+): Promise<void> {
+  await insertDeferredServerLogFiles(
+    sql,
+    benchmarkResultIds,
+    files.map((file) => ({ fileName: file.fileName, readText: () => file.logText })),
+  );
+}
+
+/** Read archived log files one at a time so multinode bundles stay memory-bounded. */
+export async function insertServerLogFilePaths(
+  sql: Sql,
+  benchmarkResultIds: number[],
+  files: readonly ServerLogFilePath[],
+): Promise<void> {
+  await insertDeferredServerLogFiles(
+    sql,
+    benchmarkResultIds,
+    files.map((file) => ({
+      fileName: file.fileName,
+      readText: () => cleanLogText(fs.readFileSync(file.path, 'utf8')),
+    })),
+  );
+}
+
+/** Compatibility wrapper for callers/tests that still provide one legacy stream. */
+export async function insertServerLog(
+  sql: Sql,
+  benchmarkResultIds: number[],
+  serverLog: string,
+): Promise<void> {
+  await insertServerLogFiles(sql, benchmarkResultIds, [
+    { fileName: 'server.log', logText: serverLog },
+  ]);
 }
 
 /**
diff --git a/packages/db/src/etl/benchmark-mapper.test.ts b/packages/db/src/etl/benchmark-mapper.test.ts
index 1985e69de..03f65ba64 100644
--- a/packages/db/src/etl/benchmark-mapper.test.ts
+++ b/packages/db/src/etl/benchmark-mapper.test.ts
@@ -127,6 +127,108 @@ describe('mapBenchmarkRow', () => {
       expect(result!.metrics).not.toHaveProperty('ep');
     });
 
+    it.each([
+      {
+        name: 'boolean verdict and junk-suffixed schema',
+        input: { power_valid: true, power_metric_schema_version: '2garbage' },
+        expectedVerdict: 0,
+        expectedSchema: undefined,
+      },
+      {
+        name: 'garbage verdict and valid numeric schema',
+        input: { power_valid: 'garbage', power_metric_schema_version: 2 },
+        expectedVerdict: 0,
+        expectedSchema: 2,
+      },
+      {
+        name: 'canonical numeric strings',
+        input: { power_valid: '1', power_metric_schema_version: '2' },
+        expectedVerdict: 1,
+        expectedSchema: 2,
+      },
+      {
+        name: 'explicit invalid verdict',
+        input: { power_valid: 0, power_metric_schema_version: 2 },
+        expectedVerdict: 0,
+        expectedSchema: 2,
+      },
+      {
+        name: 'legacy row without power contract fields',
+        input: {},
+        expectedVerdict: undefined,
+        expectedSchema: undefined,
+      },
+    ])(
+      'normalizes power contract discriminators: $name',
+      ({ input, expectedVerdict, expectedSchema }) => {
+        const tracker = createSkipTracker();
+        const result = mapBenchmarkRow(makeV1Row(input), tracker);
+
+        if (expectedVerdict === undefined) {
+          expect(result!.metrics).not.toHaveProperty('power_valid');
+        } else {
+          expect(result!.metrics.power_valid).toBe(expectedVerdict);
+        }
+        if (expectedSchema === undefined) {
+          expect(result!.metrics).not.toHaveProperty('power_metric_schema_version');
+        } else {
+          expect(result!.metrics.power_metric_schema_version).toBe(expectedSchema);
+        }
+      },
+    );
+
+    it.each([true, false, null, 0.5, 2, Number.NaN, Number.POSITIVE_INFINITY])(
+      'fails closed for explicitly malformed power_valid=%j',
+      (powerValid) => {
+        const tracker = createSkipTracker();
+        const result = mapBenchmarkRow(makeV1Row({ power_valid: powerValid }), tracker);
+
+        expect(result!.metrics.power_valid).toBe(0);
+      },
+    );
+
+    it.each([
+      null,
+      true,
+      false,
+      0,
+      -1,
+      1.5,
+      Number.NaN,
+      Number.POSITIVE_INFINITY,
+      Number.MAX_SAFE_INTEGER + 1,
+      '0',
+      '-1',
+      '02',
+      '2.0',
+      ' 2',
+      '2 ',
+      '2garbage',
+      '1e2',
+      String(Number.MAX_SAFE_INTEGER + 1),
+    ])('omits malformed power_metric_schema_version=%j', (schemaVersion) => {
+      const tracker = createSkipTracker();
+      const result = mapBenchmarkRow(
+        makeV1Row({ power_metric_schema_version: schemaVersion }),
+        tracker,
+      );
+
+      expect(result!.metrics).not.toHaveProperty('power_metric_schema_version');
+    });
+
+    it.each([1, 2, Number.MAX_SAFE_INTEGER, '1', '2', String(Number.MAX_SAFE_INTEGER)])(
+      'accepts safe positive integer power_metric_schema_version=%j',
+      (schemaVersion) => {
+        const tracker = createSkipTracker();
+        const result = mapBenchmarkRow(
+          makeV1Row({ power_metric_schema_version: schemaVersion }),
+          tracker,
+        );
+
+        expect(result!.metrics.power_metric_schema_version).toBe(Number(schemaVersion));
+      },
+    );
+
     it('preserves the KV transfer engine for fixed-sequence multinode rows', () => {
       const tracker = createSkipTracker();
       const result = mapBenchmarkRow(
@@ -415,6 +517,28 @@ describe('mapBenchmarkRow', () => {
     });
   });
 
+  describe('recipe fingerprint', () => {
+    it('preserves the producer fingerprint outside metrics', () => {
+      const result = mapBenchmarkRow(
+        makeV1Row({
+          recipe_fingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+        }),
+        createSkipTracker(),
+      );
+
+      expect(result!.recipeFingerprint).toBe(
+        'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+      );
+      expect(result!.metrics).not.toHaveProperty('recipe_fingerprint');
+    });
+
+    it('uses null for legacy artifacts without a fingerprint', () => {
+      const result = mapBenchmarkRow(makeV1Row(), createSkipTracker());
+
+      expect(result!.recipeFingerprint).toBeNull();
+    });
+  });
+
   describe('spec_decoding', () => {
     it('normalizes spec_decoding to lowercase', () => {
       const tracker = createSkipTracker();
diff --git a/packages/db/src/etl/benchmark-mapper.ts b/packages/db/src/etl/benchmark-mapper.ts
index ee71fad25..544c6ab16 100644
--- a/packages/db/src/etl/benchmark-mapper.ts
+++ b/packages/db/src/etl/benchmark-mapper.ts
@@ -39,6 +39,8 @@ const NON_METRIC_KEYS = new Set([
   'osl',
   'conc',
   'image',
+  'recipe_fingerprint',
+  'recipe-fingerprint',
   'disagg',
   'is_multinode',
   'spec_decoding',
@@ -130,6 +132,8 @@ export interface BenchmarkParams {
   /** 'on' | 'off' — KV cache offload to CPU. Defaults to 'off'. */
   offloadMode: string;
   image: string | null;
+  /** Stable producer-generated identity for the complete recipe, excluding concurrency. */
+  recipeFingerprint: string | null;
   metrics: Record<string, number>;
   /**
    * Per-worker measured-power breakdown emitted by the runner's
@@ -243,6 +247,7 @@ export function mapBenchmarkRow(
   // Dynamo artifacts that incorrectly emitted disagg=false.
   const disagg = frameworkDisagg || parallelism.decodeNumWorkers > 0;
   let metrics = captureNumericMetrics(row);
+  normalizePowerContractMetrics(row, metrics);
   if (isAgentic) metrics = preferFullResponseMetrics(metrics);
   if (!disagg) {
     const usePrefill =
@@ -332,6 +337,11 @@ export function mapBenchmarkRow(
 
   // Artifact names encode '/' as '#' to avoid path separators; restore the URI.
   const image = row.image ? String(row.image).replaceAll('#', '/') : null;
+  const rawRecipeFingerprint = row.recipe_fingerprint ?? row['recipe-fingerprint'];
+  const recipeFingerprint =
+    typeof rawRecipeFingerprint === 'string' && rawRecipeFingerprint.trim().length > 0
+      ? rawRecipeFingerprint.trim()
+      : null;
 
   // Per-worker measured-power breakdown. The runner emits this as an array
   // of objects sibling to the scalar metrics; we surface it on a dedicated
@@ -357,6 +367,7 @@ export function mapBenchmarkRow(
     conc,
     offloadMode: offloadModeRaw,
     image,
+    recipeFingerprint,
     metrics,
     workers,
   };
@@ -442,6 +453,42 @@ function captureNumericMetrics(row: Record<string, any>): Record<string, number>
   return metrics;
 }
 
+/**
+ * Re-parse the power publication contract after generic metric capture. These
+ * two fields are semantic discriminators, so loose numeric coercion must never
+ * turn malformed producer output into an affirmative verdict or schema.
+ */
+function normalizePowerContractMetrics(
+  row: Record<string, any>,
+  metrics: Record<string, number>,
+): void {
+  if (Object.hasOwn(row, 'power_valid')) {
+    const verdict = row.power_valid;
+    metrics.power_valid = verdict === 1 || verdict === '1' ? 1 : 0;
+  } else {
+    delete metrics.power_valid;
+  }
+
+  if (!Object.hasOwn(row, 'power_metric_schema_version')) {
+    delete metrics.power_metric_schema_version;
+    return;
+  }
+
+  const rawVersion = row.power_metric_schema_version;
+  const version =
+    typeof rawVersion === 'number'
+      ? rawVersion
+      : // Canonical decimal form: no sign, whitespace, decimal point, or zero padding.
+        typeof rawVersion === 'string' && /^[1-9]\d*$/u.test(rawVersion)
+        ? Number(rawVersion)
+        : undefined;
+  if (typeof version === 'number' && Number.isSafeInteger(version) && version > 0) {
+    metrics.power_metric_schema_version = version;
+  } else {
+    delete metrics.power_metric_schema_version;
+  }
+}
+
 /**
  * Narrow a raw `workers` value from the artifact JSON to `WorkerPower[]` or
  * undefined. Each entry must have a string `role`, a numeric `worker_idx`,
diff --git a/packages/db/src/etl/changelog-ingest.test.ts b/packages/db/src/etl/changelog-ingest.test.ts
index 4907f1a00..a38546fcb 100644
--- a/packages/db/src/etl/changelog-ingest.test.ts
+++ b/packages/db/src/etl/changelog-ingest.test.ts
@@ -1,5 +1,36 @@
 import { describe, expect, it, vi } from 'vitest';
-import { ingestChangelogEntries } from './changelog-ingest';
+import {
+  hasAppendOnlyFlag,
+  ingestChangelogEntries,
+  parseChangelogEntries,
+} from './changelog-ingest';
+
+describe('append-only changelog metadata', () => {
+  it('parses the marker and recognizes an all-append-only run', () => {
+    const entries = parseChangelogEntries([
+      {
+        'config-keys': ['dsv4-fp4-b300-vllm-mtp'],
+        description: ['Add concurrency 192'],
+        'pr-link': 'https://github.com/SemiAnalysisAI/InferenceX/pull/2600',
+        'append-only': true,
+      },
+    ]);
+
+    expect(entries[0]).toMatchObject({ appendOnly: true, evalsOnly: false });
+    expect(hasAppendOnlyFlag([{ entries }])).toBe(true);
+  });
+
+  it('rejects mixing append-only and regular entries in one run', () => {
+    const entries = parseChangelogEntries([
+      { 'config-keys': ['dsv4-fp4-b300-vllm-mtp'], 'append-only': true },
+      { 'config-keys': ['dsv4-fp4-h200-vllm-mtp'] },
+    ]);
+
+    expect(() => hasAppendOnlyFlag([{ entries }])).toThrow(
+      'append-only changelog entries cannot be mixed with regular entries',
+    );
+  });
+});
 
 describe('ingestChangelogEntries', () => {
   it('updates existing metadata for the same workflow and git refs', async () => {
@@ -24,6 +55,7 @@ describe('ingestChangelogEntries', () => {
           description: 'Updated benchmark description',
           prLink: 'https://github.com/SemiAnalysisAI/InferenceX/pull/2174',
           evalsOnly: false,
+          appendOnly: true,
         },
       ],
     );
@@ -31,7 +63,7 @@ describe('ingestChangelogEntries', () => {
     expect(written).toBe(1);
     expect(queries).toHaveLength(1);
     expect(queries[0].replaceAll(/\s+/gu, ' ')).toContain(
-      'on conflict (workflow_run_id, base_ref, head_ref) do update set date = excluded.date, config_keys = excluded.config_keys, description = excluded.description, pr_link = excluded.pr_link',
+      'on conflict (workflow_run_id, base_ref, head_ref) do update set date = excluded.date, config_keys = excluded.config_keys, description = excluded.description, pr_link = excluded.pr_link, append_only = excluded.append_only',
     );
   });
 });
diff --git a/packages/db/src/etl/changelog-ingest.ts b/packages/db/src/etl/changelog-ingest.ts
index 3f158985f..11a295cdc 100644
--- a/packages/db/src/etl/changelog-ingest.ts
+++ b/packages/db/src/etl/changelog-ingest.ts
@@ -12,6 +12,7 @@ export interface ChangelogEntry {
   description: string;
   prLink: string | null;
   evalsOnly: boolean;
+  appendOnly: boolean;
 }
 
 /**
@@ -39,7 +40,8 @@ export function parseChangelogEntries(raw: unknown): ChangelogEntry[] {
       description.match(/\bPR:\s*(?<url>https?:\/\/\S+)/u)?.[1] ??
       null;
     const evalsOnly = item['evals-only'] === true;
-    out.push({ configKeys, description, prLink, evalsOnly });
+    const appendOnly = item['append-only'] === true;
+    out.push({ configKeys, description, prLink, evalsOnly, appendOnly });
   }
   return out;
 }
@@ -52,6 +54,19 @@ export function hasEvalsOnlyFlag(changelogs: { entries: ChangelogEntry[] }[]): b
   return changelogs.some((c) => c.entries.some((e) => e.evalsOnly));
 }
 
+/**
+ * Return whether this run extends existing curves. Mixed metadata is rejected because
+ * append-only is stored and resolved at workflow-run scope.
+ */
+export function hasAppendOnlyFlag(changelogs: { entries: ChangelogEntry[] }[]): boolean {
+  const entries = changelogs.flatMap((changelog) => changelog.entries);
+  const appendOnlyEntries = entries.filter((entry) => entry.appendOnly);
+  if (appendOnlyEntries.length > 0 && appendOnlyEntries.length !== entries.length) {
+    throw new Error('append-only changelog entries cannot be mixed with regular entries');
+  }
+  return appendOnlyEntries.length > 0;
+}
+
 /**
  * Insert changelog entries for a workflow run into the `changelog_entries` table.
  * Uses `ON CONFLICT DO UPDATE` on `(workflow_run_id, base_ref, head_ref)`, so
@@ -77,17 +92,18 @@ export async function ingestChangelogEntries(
   for (const e of entries) {
     const [row] = await sql`
       insert into changelog_entries (
-        workflow_run_id, date, base_ref, head_ref, config_keys, description, pr_link
+        workflow_run_id, date, base_ref, head_ref, config_keys, description, pr_link, append_only
       ) values (
         ${workflowRunId}, ${date}, ${baseRef}, ${headRef},
-        ${sql.array(e.configKeys)}, ${e.description}, ${e.prLink}
+        ${sql.array(e.configKeys)}, ${e.description}, ${e.prLink}, ${e.appendOnly}
       )
       on conflict (workflow_run_id, base_ref, head_ref)
       do update set
         date = excluded.date,
         config_keys = excluded.config_keys,
         description = excluded.description,
-        pr_link = excluded.pr_link
+        pr_link = excluded.pr_link,
+        append_only = excluded.append_only
       returning id
     `;
     if (row) written++;
diff --git a/packages/db/src/etl/compute-aggregate-stats.test.ts b/packages/db/src/etl/compute-aggregate-stats.test.ts
index 8fd6a45b5..d2119c7bc 100644
--- a/packages/db/src/etl/compute-aggregate-stats.test.ts
+++ b/packages/db/src/etl/compute-aggregate-stats.test.ts
@@ -5,17 +5,21 @@ import { describe, expect, it } from 'vitest';
 import {
   STATS_VERSION,
   computeAggregateStats,
+  computeProfileAggregateStatsFromCompressedChunks,
   mergeProfileStatsUpgrade,
 } from './compute-aggregate-stats.js';
 
 /** Build a minimal `profile_export.jsonl` from a few synthetic requests. */
-function makeProfileBlob(requests: { isl: number; osl: number; rl?: number; ttft?: number }[]) {
+function makeProfileBlob(
+  requests: { isl: number; osl: number; rl?: number; ttft?: number; cancelled?: boolean }[],
+) {
   const lines = requests.map((r, i) =>
     JSON.stringify({
       metadata: {
         benchmark_phase: 'profiling',
         conversation_id: `conv-${i}`,
         turn_index: 0,
+        was_cancelled: r.cancelled,
       },
       metrics: {
         input_sequence_length: { value: r.isl, unit: 'tokens' },
@@ -81,6 +85,8 @@ describe('computeAggregateStats', () => {
     expect(stats.isl?.mean).toBeCloseTo(200, 6);
     expect(stats.osl?.n).toBe(3);
     expect(stats.osl?.mean).toBeCloseTo(75, 6);
+    expect(stats.sequenceLengths.isl?.n).toBe(3);
+    expect(stats.sequenceLengths.osl?.n).toBe(3);
 
     // Server-side metrics still null when there's no server blob.
     expect(stats.kvCacheUtil).toBeNull();
@@ -93,6 +99,38 @@ describe('computeAggregateStats', () => {
     expect(stats.e2elPerOsl?.p90).toBeCloseTo(2 / 75 + 0.8 * (0.03 - 2 / 75), 6);
   });
 
+  it('excludes cancelled requests from every streamed profile distribution', async () => {
+    const stats = await computeAggregateStats({
+      profileBlob: makeProfileBlob([
+        { isl: 100, osl: 50, rl: 1000 },
+        { isl: 1_000_000, osl: 1_000_000, rl: 1_000_000, cancelled: true },
+      ]),
+      serverBlob: null,
+    });
+
+    expect(stats.isl?.n).toBe(1);
+    expect(stats.isl?.p50).toBe(100);
+    expect(stats.osl?.n).toBe(1);
+    expect(stats.e2elPerOsl?.n).toBe(1);
+    expect(stats.sequenceLengths.isl?.n).toBe(1);
+    expect(stats.sequenceLengths.osl?.n).toBe(1);
+  });
+
+  it('accepts a gzip profile split across arbitrarily small compressed chunks', async () => {
+    const blob = makeProfileBlob([
+      { isl: 100, osl: 50, rl: 1000 },
+      { isl: 200, osl: 100, rl: 2000 },
+    ]);
+    const chunks = Array.from({ length: Math.ceil(blob.length / 7) }, (_, index) =>
+      blob.subarray(index * 7, Math.min(blob.length, (index + 1) * 7)),
+    );
+
+    const stats = await computeProfileAggregateStatsFromCompressedChunks(chunks);
+    expect(stats.isl?.n).toBe(2);
+    expect(stats.osl?.n).toBe(2);
+    expect(stats.e2elPerOsl?.n).toBe(2);
+  });
+
   it('computes KV util + prefix hit rate from the server blob alone', async () => {
     const stats = await computeAggregateStats({
       profileBlob: null,
@@ -148,7 +186,15 @@ describe('mergeProfileStatsUpgrade', () => {
       version: STATS_VERSION - 1,
       isl: null,
       osl: null,
-      kvCacheUtil: { mean: 0.4, p50: 0.4, p75: 0.5, p90: 0.6, p99: 0.7, n: 3 },
+      kvCacheUtil: {
+        mean: 0.4,
+        p50: 0.4,
+        p75: 0.5,
+        p90: 0.6,
+        p95: 0.65,
+        p99: 0.7,
+        n: 3,
+      },
       prefixCacheHitRate: null,
       normalizedSessionTimeS: 999,
       p90PrefillTpsPerUser: 999,
diff --git a/packages/db/src/etl/compute-aggregate-stats.ts b/packages/db/src/etl/compute-aggregate-stats.ts
index 0ad8a1e6b..07729b59b 100644
--- a/packages/db/src/etl/compute-aggregate-stats.ts
+++ b/packages/db/src/etl/compute-aggregate-stats.ts
@@ -9,16 +9,18 @@
  * computation changes so the backfill script knows which rows to recompute.
  */
 
-import { gunzipSync } from 'node:zlib';
+import { createInterface } from 'node:readline';
+import { Readable } from 'node:stream';
+import { createGunzip } from 'node:zlib';
 
 import { gunzipJsonWithinLimit, streamCollectKeys } from './gzip-json-stream';
-import { computeDerivedFromBlob } from '../queries/derived-agentic-metrics';
 import {
   STATS_VERSION,
-  extractIslOsl,
   extractServerMetricSamples,
   percentilesOf,
+  sequenceLengthSketches,
   type MetricPercentiles,
+  type SequenceLengthSketches,
 } from '../queries/agentic-aggregates';
 
 export { STATS_VERSION };
@@ -35,6 +37,121 @@ export interface AggregateStats {
    * (tok/s/user): pXX E2E Normalized Interactivity = 1 / pXX(E2EL/OSL).
    */
   e2elPerOsl: MetricPercentiles | null;
+  /** Bounded mergeable distributions used by the chart-level subtitle. */
+  sequenceLengths: SequenceLengthSketches;
+}
+
+interface ProfileMetricEnvelope {
+  value?: number;
+}
+
+interface ProfileRecord {
+  metadata?: {
+    benchmark_phase?: string;
+    was_cancelled?: boolean;
+  };
+  metrics?: {
+    input_sequence_length?: ProfileMetricEnvelope | number;
+    output_sequence_length?: ProfileMetricEnvelope | number;
+    request_latency?: ProfileMetricEnvelope | number;
+    time_to_first_token?: ProfileMetricEnvelope | number;
+  };
+}
+
+function profileMetricValue(value: ProfileMetricEnvelope | number | undefined): number | undefined {
+  const number = typeof value === 'number' ? value : value?.value;
+  return typeof number === 'number' && Number.isFinite(number) ? number : undefined;
+}
+
+/**
+ * Stream a profile export line by line so exceptionally large traces never
+ * materialize their multi-gigabyte decompressed JSONL as one string. The
+ * numeric sample arrays are tiny relative to the source and are needed for
+ * exact percentile calculation.
+ */
+async function extractProfileSamples(
+  compressedChunks: Iterable<Uint8Array> | AsyncIterable<Uint8Array>,
+): Promise<{
+  isl: number[];
+  osl: number[];
+  e2elPerOsl: number[];
+}> {
+  const input = Readable.from(compressedChunks).pipe(createGunzip());
+  const lines = createInterface({ input, crlfDelay: Infinity });
+  const isl: number[] = [];
+  const osl: number[] = [];
+  const e2elPerOsl: number[] = [];
+
+  for await (const line of lines) {
+    if (!line) continue;
+    let record: ProfileRecord;
+    try {
+      record = JSON.parse(line) as ProfileRecord;
+    } catch {
+      continue;
+    }
+    if (record.metadata?.benchmark_phase && record.metadata.benchmark_phase !== 'profiling') {
+      continue;
+    }
+    if (record.metadata?.was_cancelled === true) continue;
+
+    const metrics = record.metrics ?? {};
+    const inputLength = profileMetricValue(metrics.input_sequence_length);
+    const outputLength = profileMetricValue(metrics.output_sequence_length);
+    if (inputLength !== undefined) isl.push(inputLength);
+    if (outputLength !== undefined) osl.push(outputLength);
+
+    const requestLatencyMs = profileMetricValue(metrics.request_latency);
+    const ttftMs = profileMetricValue(metrics.time_to_first_token);
+    if (
+      requestLatencyMs !== undefined &&
+      ttftMs !== undefined &&
+      inputLength !== undefined &&
+      outputLength !== undefined &&
+      requestLatencyMs > 0 &&
+      ttftMs > 0 &&
+      inputLength > 0 &&
+      outputLength > 0
+    ) {
+      e2elPerOsl.push(requestLatencyMs / 1000 / outputLength);
+    }
+  }
+
+  return { isl, osl, e2elPerOsl };
+}
+
+/**
+ * Compute the profile-derived half of the aggregate bundle from compressed
+ * chunks. Backfills use this for oversized TOAST values so neither Postgres
+ * nor the JS driver has to materialize one enormous bytea response.
+ */
+export async function computeProfileAggregateStatsFromCompressedChunks(
+  compressedChunks: Iterable<Uint8Array> | AsyncIterable<Uint8Array>,
+): Promise<AggregateStats> {
+  let islPct: MetricPercentiles | null = null;
+  let oslPct: MetricPercentiles | null = null;
+  let e2elPerOsl: MetricPercentiles | null = null;
+  let sequenceLengths: SequenceLengthSketches = { isl: null, osl: null };
+
+  try {
+    const { isl, osl, e2elPerOsl: ratios } = await extractProfileSamples(compressedChunks);
+    islPct = percentilesOf(isl);
+    oslPct = percentilesOf(osl);
+    sequenceLengths = sequenceLengthSketches(isl, osl);
+    e2elPerOsl = percentilesOf(ratios);
+  } catch {
+    // Ignore malformed blobs and leave the profile-derived fields null.
+  }
+
+  return {
+    version: STATS_VERSION,
+    isl: islPct,
+    osl: oslPct,
+    kvCacheUtil: null,
+    prefixCacheHitRate: null,
+    e2elPerOsl,
+    sequenceLengths,
+  };
 }
 
 /**
@@ -125,21 +242,9 @@ export async function computeAggregateStats(args: {
   profileBlob: Buffer | null;
   serverBlob: Buffer | null;
 }): Promise<AggregateStats> {
-  let islPct: MetricPercentiles | null = null;
-  let oslPct: MetricPercentiles | null = null;
-  let e2elPerOsl: MetricPercentiles | null = null;
-
-  if (args.profileBlob) {
-    try {
-      const jsonl = gunzipSync(args.profileBlob).toString('utf8');
-      const { isl, osl } = extractIslOsl(jsonl);
-      islPct = percentilesOf(isl);
-      oslPct = percentilesOf(osl);
-      e2elPerOsl = computeDerivedFromBlob(jsonl).e2el_per_osl;
-    } catch {
-      // ignore malformed blob — leave nulls
-    }
-  }
+  const profile = args.profileBlob
+    ? await computeProfileAggregateStatsFromCompressedChunks([args.profileBlob])
+    : await computeProfileAggregateStatsFromCompressedChunks([]);
 
   let kvPct: MetricPercentiles | null = null;
   let prefixPct: MetricPercentiles | null = null;
@@ -161,11 +266,8 @@ export async function computeAggregateStats(args: {
   }
 
   return {
-    version: STATS_VERSION,
-    isl: islPct,
-    osl: oslPct,
+    ...profile,
     kvCacheUtil: kvPct,
     prefixCacheHitRate: prefixPct,
-    e2elPerOsl,
   };
 }
diff --git a/packages/db/src/etl/compute-chart-series.test.ts b/packages/db/src/etl/compute-chart-series.test.ts
index 749241713..54295cc77 100644
--- a/packages/db/src/etl/compute-chart-series.test.ts
+++ b/packages/db/src/etl/compute-chart-series.test.ts
@@ -119,6 +119,33 @@ function buildDynamoSeries(
   };
 }
 
+/** A kv_cache_usage_perc series for one engine as seen from one endpoint. */
+function kvSeriesFor(
+  endpoint_url: string,
+  labels: Record<string, string>,
+  samples: [startNs: number, avg: number][],
+) {
+  return {
+    endpoint_url,
+    labels,
+    timeslices: samples.map(([start_ns, avg]) => ({ start_ns, end_ns: start_ns + 1e9, avg })),
+  };
+}
+
+/** Gzip a blob carrying only kv_cache_usage_perc, optionally in both phases. */
+function kvBlob(profiling: unknown[], warmup: unknown[] = []) {
+  return gzipSync(
+    Buffer.from(
+      JSON.stringify({
+        metrics: { 'vllm:kv_cache_usage_perc': { series: profiling } },
+        ...(warmup.length > 0
+          ? { warmup_metrics: { 'vllm:kv_cache_usage_perc': { series: warmup } } }
+          : {}),
+      }),
+    ),
+  );
+}
+
 describe('computeChartSeries', () => {
   it('returns null when the blob is null', async () => {
     expect(await computeChartSeries(null)).toBeNull();
@@ -298,10 +325,12 @@ describe('computeChartSeries', () => {
     });
 
     expect(result?.metricSources).toHaveLength(3);
+    // engine/dpRank are null at worker granularity (v15) — a source spans
+    // every engine the worker owns.
     expect(result?.metricSources.map(({ source: s }) => [s.role, s.workerId, s.engine])).toEqual([
-      ['prefill', 'prefill-a', '0'],
-      ['prefill', 'prefill-b', '0'],
-      ['decode', 'decode-a', '0'],
+      ['prefill', 'prefill-a', null],
+      ['prefill', 'prefill-b', null],
+      ['decode', 'decode-a', null],
     ]);
     const prefillA = result?.metricSources.find(({ source: s }) => s.workerId === 'prefill-a');
     const decode = result?.metricSources.find(({ source: s }) => s.role === 'decode');
@@ -316,6 +345,347 @@ describe('computeChartSeries', () => {
     expect(nonDisagg?.metricSources).toEqual([]);
   });
 
+  // ── Per-engine identity (v13) ─────────────────────────────────────────
+  //
+  // The blob stores one series per (scrape endpoint x phase block x label
+  // set), which is not one series per engine. These cover the three ways
+  // that mismatch used to inflate `kvCacheUsageByEngine` and fragment the
+  // cluster average.
+
+  it('collapses API-server frontends that mirror the same engines', async () => {
+    // vLLM with two API servers exposes every DP rank on BOTH /metrics
+    // endpoints, scraped a fraction of a second apart. That is 4 series for
+    // 2 engines, not 4 engines.
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor('http://localhost:8895/metrics', { engine: '0', model_name: 'm' }, [
+          [0, 0.2],
+          [1e9, 0.3],
+        ]),
+        kvSeriesFor('http://localhost:8895/metrics', { engine: '1', model_name: 'm' }, [
+          [0, 0.6],
+          [1e9, 0.7],
+        ]),
+        kvSeriesFor('http://localhost:8896/metrics', { engine: '0', model_name: 'm' }, [
+          [0.176e9, 0.2],
+          [1.176e9, 0.3],
+        ]),
+        kvSeriesFor('http://localhost:8896/metrics', { engine: '1', model_name: 'm' }, [
+          [0.176e9, 0.6],
+          [1.176e9, 0.7],
+        ]),
+      ]),
+    );
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual(['0', '1']);
+    // The mean is over the two real engines, once per scrape tick — not four
+    // ticks averaging one endpoint's subset each.
+    expect(cs?.kvCacheUsage).toEqual([
+      { t: 0, value: 0.4 },
+      { t: 1, value: 0.5 },
+    ]);
+  });
+
+  it('joins an engine warmup and profiling scrapes into one line', async () => {
+    const cs = await computeChartSeries(
+      kvBlob(
+        [
+          kvSeriesFor('http://localhost:8000/metrics', { engine: '0' }, [[10e9, 0.8]]),
+          kvSeriesFor('http://localhost:8000/metrics', { engine: '1' }, [[10e9, 0.4]]),
+        ],
+        [
+          kvSeriesFor('http://localhost:8000/metrics', { engine: '0' }, [[0, 0.2]]),
+          kvSeriesFor('http://localhost:8000/metrics', { engine: '1' }, [[0, 0.1]]),
+        ],
+      ),
+    );
+    expect(cs?.kvCacheUsageByEngine).toEqual([
+      {
+        engineLabel: '0',
+        points: [
+          { t: 0, value: 0.2 },
+          { t: 10, value: 0.8 },
+        ],
+      },
+      {
+        engineLabel: '1',
+        points: [
+          { t: 0, value: 0.1 },
+          { t: 10, value: 0.4 },
+        ],
+      },
+    ]);
+  });
+
+  it('suppresses the per-engine overlay for a single-engine deployment', async () => {
+    // Warmup + profiling is two series but still one engine, so there is
+    // nothing for a per-rank overlay to compare.
+    const cs = await computeChartSeries(
+      kvBlob(
+        [kvSeriesFor('http://localhost:8000/metrics', { engine: '0' }, [[10e9, 0.8]])],
+        [kvSeriesFor('http://localhost:8000/metrics', { engine: '0' }, [[0, 0.2]])],
+      ),
+    );
+    expect(cs?.kvCacheUsageByEngine).toEqual([]);
+    expect(cs?.kvCacheUsage).toEqual([
+      { t: 0, value: 0.2 },
+      { t: 10, value: 0.8 },
+    ]);
+  });
+
+  it('keeps disaggregated workers apart even when their ranks collide', async () => {
+    // Prefill rank 0 and decode rank 0 are different engines; `worker_id`
+    // is what says so.
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor(
+          'http://10.0.0.1:7500/metrics',
+          { dp_rank: '0', engine_type: 'prefill', worker_id: 'aaaaaaaawork0001' },
+          [[0, 0.02]],
+        ),
+        kvSeriesFor(
+          'http://10.0.0.2:7502/metrics',
+          { dp_rank: '0', engine_type: 'decode', worker_id: 'bbbbbbbbwork0002' },
+          [[0, 0.5]],
+        ),
+      ]),
+    );
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual(['prefill 0', 'decode 0']);
+    expect(cs?.kvCacheUsage).toEqual([{ t: 0, value: 0.26 }]);
+  });
+
+  it('collapses tensor-parallel shard ranks that share one KV pool', async () => {
+    // A TP4 SGLang worker reports kv_cache_usage_perc once per tp_rank, but
+    // the four ranks shard a single pool and track each other to ~4dp. They
+    // are one engine, not four.
+    const shards = [0, 1, 2, 3].map((tp) =>
+      kvSeriesFor(
+        'http://10.0.0.1:7500/metrics',
+        {
+          tp_rank: String(tp),
+          pp_rank: '0',
+          moe_ep_rank: String(tp),
+          engine_type: 'prefill',
+          worker_id: 'worker-aaaa',
+        },
+        [
+          [0, 0.4 + tp * 0.0001],
+          [1e9, 0.6 + tp * 0.0001],
+        ],
+      ),
+    );
+    const cs = await computeChartSeries(kvBlob(shards));
+    // One pool -> one engine, so the per-rank overlay stays off entirely.
+    expect(cs?.kvCacheUsageByEngine).toEqual([]);
+    expect(cs?.kvCacheUsage).toEqual([
+      { t: 0, value: 0.40015 },
+      { t: 1, value: 0.60015 },
+    ]);
+  });
+
+  it('keeps DP ranks apart even when shard ranks co-vary with them', async () => {
+    // Same worker, but now dp_rank names a real per-rank pool and tp_rank
+    // happens to move with it. Dropping shard labels must not fuse these.
+    const cs = await computeChartSeries(
+      kvBlob(
+        [0, 1].map((dp) =>
+          kvSeriesFor(
+            'http://10.0.0.1:7500/metrics',
+            {
+              dp_rank: String(dp),
+              tp_rank: String(dp),
+              engine_type: 'prefill',
+              worker_id: 'worker-aaaa',
+            },
+            [[0, dp === 0 ? 0.2 : 0.8]],
+          ),
+        ),
+      ),
+    );
+    // Both engines are prefill, so naming the role would add nothing.
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual(['0', '1']);
+    expect(cs?.kvCacheUsage).toEqual([{ t: 0, value: 0.5 }]);
+  });
+
+  it('keeps DP ranks in rank order regardless of the order the exporter emits', async () => {
+    // Regression: role-qualified labels used to sort by blob position, so a
+    // multi-worker aggregated deployment came out as [decode 3, decode 1,
+    // decode 0, decode 2] with the palette indexed by that scrambled order.
+    const cs = await computeChartSeries(
+      kvBlob(
+        [3, 1, 0, 2].map((rank) =>
+          kvSeriesFor(
+            `http://10.30.1.${100 + rank}:7500/metrics`,
+            { engine: String(rank), dynamo_component: 'backend', worker_id: `w${rank}` },
+            [[0, [0.25, 0.5, 0.75, 1][rank]!]],
+          ),
+        ),
+      ),
+      { framework: 'dynamo-vllm', disagg: false },
+    );
+    // Every engine is a decode worker, so the role is dropped and the bare
+    // ranks render in 0..N order — with each rank's own data still attached.
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual(['0', '1', '2', '3']);
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.points[0]!.value)).toEqual([0.25, 0.5, 0.75, 1]);
+  });
+
+  it('names the role only when engines actually differ in role', async () => {
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor(
+          'http://10.0.0.2:7502/metrics',
+          { dp_rank: '0', engine_type: 'decode', worker_id: 'dec' },
+          [[0, 0.6]],
+        ),
+        kvSeriesFor(
+          'http://10.0.0.1:7500/metrics',
+          { dp_rank: '0', engine_type: 'prefill', worker_id: 'pre' },
+          [[0, 0.2]],
+        ),
+      ]),
+    );
+    // Mixed roles -> role is shown, and prefill sorts ahead of decode.
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual(['prefill 0', 'decode 0']);
+  });
+
+  it('falls through to dynamo_component when engine_type has no role mapping', async () => {
+    // Aggregated dynamo-sglang workers carry engine_type="unified" (no role)
+    // alongside dynamo_component. Stopping at the first present label would
+    // resolve the role to null and lose the prefill/decode distinction.
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor(
+          'http://10.0.0.1:7500/metrics',
+          { engine_type: 'unified', dynamo_component: 'prefill', worker_id: 'pre', dp_rank: '0' },
+          [[0, 0.2]],
+        ),
+        kvSeriesFor(
+          'http://10.0.0.2:7502/metrics',
+          { engine_type: 'unified', dynamo_component: 'backend', worker_id: 'dec', dp_rank: '0' },
+          [[0, 0.6]],
+        ),
+      ]),
+    );
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual(['prefill 0', 'decode 0']);
+  });
+
+  it('qualifies engines whose display label would otherwise collide', async () => {
+    // Two decode workers that each number their ranks from 0. One role, so the
+    // rank alone is the base and the worker id disambiguates.
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor(
+          'http://10.0.0.1:7502/metrics',
+          { dp_rank: '0', engine_type: 'decode', worker_id: 'worker-a01a' },
+          [[0, 0.4]],
+        ),
+        kvSeriesFor(
+          'http://10.0.0.2:7503/metrics',
+          { dp_rank: '0', engine_type: 'decode', worker_id: 'worker-b01b' },
+          [[0, 0.6]],
+        ),
+      ]),
+    );
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual(['0 (a01a)', '0 (b01b)']);
+  });
+
+  it('keeps same-label endpoints apart when their values disagree', async () => {
+    // Two independent replicas behind a router share an identical label set.
+    // Treating them as mirrors and dropping one would silently lose an engine.
+    const labels = { engine_type: 'unified', model_name: 'm', tp_rank: '0' };
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor('http://node-a:8888/metrics', labels, [
+          [0, 0.1],
+          [1e9, 0.1],
+        ]),
+        kvSeriesFor('http://node-b:8888/metrics', labels, [
+          [0, 0.9],
+          [1e9, 0.9],
+        ]),
+      ]),
+    );
+    expect(cs?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual([
+      'node-a:8888',
+      'node-b:8888',
+    ]);
+    expect(cs?.kvCacheUsage).toEqual([
+      { t: 0, value: 0.5 },
+      { t: 1, value: 0.5 },
+    ]);
+  });
+
+  it('prefers the mirror that covers the most wall-clock, not the densest', async () => {
+    // A truncated but high-frequency mirror must not shorten the engine.
+    const labels = { engine: '0', model_name: 'm' };
+    const long: [number, number][] = Array.from({ length: 11 }, (_, i) => [i * 1e9, 0.5]);
+    const dense: [number, number][] = Array.from({ length: 12 }, (_, i) => [i * 1e8, 0.5]);
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor('http://a:8000/metrics', labels, dense),
+        kvSeriesFor('http://b:8000/metrics', labels, long),
+      ]),
+    );
+    // Same values -> mirrors -> one engine, and it must span the full 10 s.
+    expect(cs?.kvCacheUsage).toHaveLength(11);
+    expect(cs?.kvCacheUsage.at(-1)?.t).toBe(10);
+  });
+
+  it('drops an engine from the mean while it stops reporting', async () => {
+    // A carries 1 for t=0..9, goes silent for 300 s, then returns. B reports 0
+    // throughout. During the hole the mean must be B alone, not (1+0)/2.
+    const a: [number, number][] = [];
+    const b: [number, number][] = [];
+    for (let i = 0; i < 10; i++) a.push([i * 1e9, 1]);
+    for (let i = 310; i < 320; i++) a.push([i * 1e9, 1]);
+    for (let i = 0; i < 320; i++) b.push([i * 1e9, 0]);
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor('http://a:8000/metrics', { engine: '0' }, a),
+        kvSeriesFor('http://a:8000/metrics', { engine: '1' }, b),
+      ]),
+    );
+    const at = (t: number) => cs?.kvCacheUsage.find((p) => p.t === t)?.value;
+    expect(at(5)).toBe(0.5); // both reporting
+    expect(at(150)).toBe(0); // A silent -> excluded, B alone
+    expect(at(315)).toBe(0.5); // A back
+  });
+
+  it('averages engines on unaligned scrape grids without sawtoothing', async () => {
+    // A prefill worker and a decode worker on their own sub-second grids.
+    // Grouping on an exact start_ns made every tick "prefill only" (0.0) or
+    // "decode only" (1.0), i.e. a full-scale sawtooth where the real cluster
+    // mean is a flat 0.5.
+    const prefill: [number, number][] = [];
+    const decode: [number, number][] = [];
+    for (let i = 0; i < 10; i++) {
+      prefill.push([i * 1e9, 0]);
+      decode.push([i * 1e9 + 0.5e9, 1]);
+    }
+    const cs = await computeChartSeries(
+      kvBlob([
+        kvSeriesFor(
+          'http://10.0.0.1:7500/metrics',
+          { worker_id: 'p', engine_type: 'prefill' },
+          prefill,
+        ),
+        kvSeriesFor(
+          'http://10.0.0.2:7502/metrics',
+          { worker_id: 'd', engine_type: 'decode' },
+          decode,
+        ),
+      ]),
+    );
+    // The edge ticks fall outside one engine's observed window — t=0 predates
+    // decode's first scrape, t=9.5 postdates prefill's last — so they report
+    // the one engine that was actually running rather than a carried-forward
+    // stale value. Every tick in between averages both and stays flat.
+    expect(cs?.kvCacheUsage.at(0)).toEqual({ t: 0, value: 0 });
+    expect(cs?.kvCacheUsage.at(-1)).toEqual({ t: 9.5, value: 1 });
+    const both = cs!.kvCacheUsage.slice(1, -1);
+    expect(both).toHaveLength(18);
+    expect(both.every((p) => p.value === 0.5)).toBe(true);
+  });
+
   it('does not interpret Dynamo-native labels without selecting the Dynamo adapter', async () => {
     const json = JSON.stringify({
       metrics: {
@@ -339,3 +709,326 @@ describe('computeChartSeries', () => {
     expect(result?.metricSources).toEqual([]);
   });
 });
+
+// ── Summed series on the canonical grid (v14) ───────────────────────────
+//
+// Components of one metric are not scraped in lockstep. Summing them on an
+// exact `start_ns` emitted one point per component per tick, each holding that
+// component's share alone — a comb that `rollingAverage` (a sample-count mean)
+// then read as a fraction of the real cluster total.
+
+/** A counter series for one component, sampled at `1 Hz` from `startNs`. */
+function rateSeriesFor(
+  endpoint_url: string,
+  labels: Record<string, string>,
+  startNs: number,
+  rates: number[],
+) {
+  return {
+    endpoint_url,
+    labels,
+    timeslices: rates.map((rate, i) => ({
+      start_ns: startNs + i * 1e9,
+      end_ns: startNs + (i + 1) * 1e9,
+      rate,
+    })),
+  };
+}
+
+function rateBlob(metricName: string, series: unknown[]) {
+  return gzipSync(Buffer.from(JSON.stringify({ metrics: { [metricName]: { series } } })));
+}
+
+/** Σ of a rate series ≈ total tokens, which is what the cumulative charts read. */
+function total(points: { value: number }[] | undefined): number {
+  return (points ?? []).reduce((a, p) => a + p.value, 0);
+}
+
+// ── Metric-source identity is one WORKER (v15) ──────────────────────────
+//
+// `dp_rank` (sglang) and `engine` (vllm) name engines INSIDE a worker. Keying
+// the source on them split each worker into 4-5 identical-looking dropdown
+// entries — a 6-prefill/1-decode run listed 35 "endpoints" for 7 workers.
+
+/** One worker's series for `metric`, one entry per sub-engine label. */
+function workerSeries(
+  endpoint_url: string,
+  dynamo_component: 'prefill' | 'backend',
+  worker_id: string,
+  subLabel: 'dp_rank' | 'engine',
+  subValues: string[],
+  value: number,
+  field: 'rate' | 'avg' = 'rate',
+) {
+  return subValues.map((v) => ({
+    endpoint_url,
+    labels: { dynamo_component, worker_id, [subLabel]: v },
+    timeslices: [{ start_ns: 0, end_ns: 1e9, [field]: value }],
+  }));
+}
+
+function dynamoBlob(series: unknown[]) {
+  return gzipSync(
+    Buffer.from(
+      JSON.stringify({
+        metrics: {
+          'vllm:prompt_tokens': { series },
+        },
+      }),
+    ),
+  );
+}
+
+describe('metric source identity', () => {
+  it("collapses a worker's sglang dp_rank series into one source", async () => {
+    const cs = await computeChartSeries(
+      dynamoBlob([
+        ...workerSeries(
+          'http://10.0.0.1:7500/metrics',
+          'prefill',
+          'w-pre',
+          'dp_rank',
+          ['0', '1', '2', '3'],
+          25,
+        ),
+        ...workerSeries(
+          'http://10.0.0.2:7502/metrics',
+          'backend',
+          'w-dec',
+          'dp_rank',
+          ['0', '1', '2', '3'],
+          10,
+        ),
+      ]),
+      { framework: 'dynamo-sglang', disagg: true },
+    );
+    expect(cs?.metricSources.map(({ source: s }) => [s.role, s.workerId])).toEqual([
+      ['prefill', 'w-pre'],
+      ['decode', 'w-dec'],
+    ]);
+    // The worker's four ranks are summed into its one source, not listed apart.
+    const prefill = cs?.metricSources.find(({ source: s }) => s.workerId === 'w-pre');
+    expect(prefill?.promptTps).toEqual([{ t: 0, value: 100 }]);
+  });
+
+  it("collapses a worker's vllm engine series into one source", async () => {
+    // vllm labels the sub-engine `engine` and leaves dp_rank unset, so the
+    // duplication arrives under a different label than sglang's.
+    const cs = await computeChartSeries(
+      dynamoBlob([
+        ...workerSeries(
+          'http://10.0.0.1:7500/metrics',
+          'prefill',
+          'w-pre',
+          'engine',
+          ['0', '1', '2', '3'],
+          25,
+        ),
+      ]),
+      { framework: 'dynamo-vllm', disagg: true },
+    );
+    expect(cs?.metricSources).toHaveLength(1);
+    expect(cs?.metricSources[0]!.source.workerId).toBe('w-pre');
+    expect(cs?.metricSources[0]!.promptTps).toEqual([{ t: 0, value: 100 }]);
+  });
+
+  it('keeps a worker whole when it also emits an unlabelled aggregate series', async () => {
+    // sglang emits one series with no dp_rank alongside the per-rank ones,
+    // which used to show up as a fifth entry for the same worker.
+    const cs = await computeChartSeries(
+      dynamoBlob([
+        {
+          endpoint_url: 'http://10.0.0.1:7500/metrics',
+          labels: { dynamo_component: 'prefill', worker_id: 'w-pre' },
+          timeslices: [{ start_ns: 0, end_ns: 1e9, rate: 7 }],
+        },
+        ...workerSeries(
+          'http://10.0.0.1:7500/metrics',
+          'prefill',
+          'w-pre',
+          'dp_rank',
+          ['0', '1'],
+          25,
+        ),
+      ]),
+      { framework: 'dynamo-sglang', disagg: true },
+    );
+    expect(cs?.metricSources).toHaveLength(1);
+    expect(cs?.metricSources[0]!.source.dpRank).toBeNull();
+    expect(cs?.metricSources[0]!.source.engine).toBeNull();
+  });
+
+  it('keeps distinct workers apart even when their ranks collide', async () => {
+    const cs = await computeChartSeries(
+      dynamoBlob([
+        ...workerSeries('http://10.0.0.1:7500/metrics', 'prefill', 'w-a', 'dp_rank', ['0'], 5),
+        ...workerSeries('http://10.0.0.2:7500/metrics', 'prefill', 'w-b', 'dp_rank', ['0'], 5),
+        ...workerSeries('http://10.0.0.3:7500/metrics', 'prefill', 'w-c', 'dp_rank', ['0'], 5),
+      ]),
+      { framework: 'dynamo-vllm', disagg: true },
+    );
+    expect(cs?.metricSources.map(({ source: s }) => s.workerId)).toEqual(['w-a', 'w-b', 'w-c']);
+  });
+});
+
+describe('computeChartSeries summed series', () => {
+  it('sums components scraped on offset grids instead of emitting a comb', async () => {
+    // Two workers at 1 Hz, 16 ms out of phase — the real SGLang/vLLM pattern.
+    // Grouping on an exact start_ns produced 6 points alternating 100 and 20;
+    // the cluster was never once reported as its actual 120 tok/s.
+    const cs = await computeChartSeries(
+      rateBlob('vllm:prompt_tokens', [
+        rateSeriesFor('http://a:8000/metrics', { worker_id: 'a' }, 0, [100, 100, 100]),
+        rateSeriesFor('http://b:8000/metrics', { worker_id: 'b' }, 0.016e9, [20, 20, 20]),
+      ]),
+    );
+    expect(cs?.prefillTps).toEqual([
+      { t: 0, value: 100 },
+      { t: 1, value: 120 },
+      { t: 2, value: 120 },
+    ]);
+  });
+
+  it('ignores a label split that reports nothing rather than halving the rate', async () => {
+    // SGLang splits its token counters by `is_streaming`. When every request
+    // streams, the "false" series is a full-length run of zeros on its own
+    // grid, and interleaving it dragged the rolling average down ~2.2x.
+    const cs = await computeChartSeries(
+      rateBlob('sglang:generation_tokens', [
+        rateSeriesFor('http://a:8000/metrics', { is_streaming: 'false' }, 0, [0, 0, 0, 0]),
+        rateSeriesFor('http://a:8000/metrics', { is_streaming: 'true' }, 0.016e9, [80, 90, 70, 60]),
+      ]),
+    );
+    expect(cs?.decodeTps.map((p) => p.value)).toEqual([0, 80, 90, 70]);
+  });
+
+  it('preserves the token total, which the cumulative charts read as a sum', async () => {
+    // `cumulativeUniqueInputTokens` does `sum += value`, so regridding must not
+    // move Σ — one point per one-second bucket, no more and no less. Components
+    // already on the grid round-trip exactly.
+    const rates = [10, 40, 0, 25, 5];
+    const aligned = await computeChartSeries(
+      rateBlob('vllm:prompt_tokens', [
+        rateSeriesFor('http://a:8000/metrics', { worker_id: 'a' }, 0, rates),
+        rateSeriesFor('http://b:8000/metrics', { worker_id: 'b' }, 0, rates),
+      ]),
+    );
+    expect(total(aligned?.prefillTps)).toBe(160);
+    expect(aligned?.prefillTps.map((p) => p.t)).toEqual([0, 1, 2, 3, 4]);
+
+    // An off-grid component is read at each tick through its own step
+    // function, so Σ can differ by at most the one trailing bucket that falls
+    // past the last whole tick (here worker b's final 5). On a real 4000-tick
+    // row that edge is invisible: points 439817's prefill and decode totals
+    // both came out byte-identical to the pre-v14 values.
+    const offset = await computeChartSeries(
+      rateBlob('vllm:prompt_tokens', [
+        rateSeriesFor('http://a:8000/metrics', { worker_id: 'a' }, 0, rates),
+        rateSeriesFor('http://b:8000/metrics', { worker_id: 'b' }, 0.4e9, rates),
+      ]),
+    );
+    expect(total(offset?.prefillTps)).toBe(155);
+    expect(offset?.prefillTps.map((p) => p.t)).toEqual([0, 1, 2, 3, 4]);
+  });
+
+  it('counts mirrored API-server frontends once, not twice', async () => {
+    // vLLM with two API servers exposes the same engine on both /metrics
+    // endpoints. Summing both double-counted every token — measured at exactly
+    // 2x on the stored rows for points 439201 and 439263.
+    const cs = await computeChartSeries(
+      rateBlob('vllm:prompt_tokens', [
+        rateSeriesFor('http://localhost:8888/metrics', { engine: '0' }, 0, [500, 500, 500]),
+        rateSeriesFor('http://localhost:8889/metrics', { engine: '0' }, 0.019e9, [500, 500, 500]),
+      ]),
+    );
+    expect(cs?.prefillTps.map((p) => p.value)).toEqual([500, 500, 500]);
+  });
+
+  it('keeps endpoints that disagree, which are engines rather than mirrors', async () => {
+    // Same label set on two endpoints but very different levels: a router in
+    // front of two replicas. Dropping one would silently lose half the load.
+    const cs = await computeChartSeries(
+      rateBlob('vllm:prompt_tokens', [
+        rateSeriesFor('http://localhost:8888/metrics', { engine: '0' }, 0, [100, 100, 100]),
+        rateSeriesFor('http://localhost:8889/metrics', { engine: '0' }, 0, [900, 900, 900]),
+      ]),
+    );
+    expect(cs?.prefillTps.map((p) => p.value)).toEqual([1000, 1000, 1000]);
+  });
+
+  it('puts divided metrics on one lattice so the hit rate survives', async () => {
+    // Regression guard: anchoring each metric's grid at its own first sample
+    // put hits on ...x.18 and queries on ...x.00, so the join found no shared
+    // instant and the prefix-cache-hit-rate chart came out empty.
+    const cs = await computeChartSeries(
+      gzipSync(
+        Buffer.from(
+          JSON.stringify({
+            metrics: {
+              'vllm:prefix_cache_hits': {
+                series: [
+                  rateSeriesFor('http://a:8000/metrics', { engine: '0' }, 0.18e9, [75, 75]),
+                  rateSeriesFor('http://a:8000/metrics', { engine: '1' }, 0.34e9, [75, 75]),
+                ],
+              },
+              'vllm:prefix_cache_queries': {
+                series: [
+                  rateSeriesFor('http://a:8000/metrics', { engine: '0' }, 0, [100, 100]),
+                  rateSeriesFor('http://a:8000/metrics', { engine: '1' }, 0.5e9, [100, 100]),
+                ],
+              },
+            },
+          }),
+        ),
+      ),
+    );
+    expect(cs?.prefixCacheHitRate.length).toBeGreaterThan(0);
+    for (const point of cs!.prefixCacheHitRate) {
+      expect(Number.isInteger(point.t)).toBe(true);
+      expect(point.value).toBeCloseTo(0.75, 5);
+    }
+  });
+
+  it('sums queue depth across workers on their own scrape offsets', async () => {
+    const json = JSON.stringify({
+      metrics: {
+        'vllm:num_requests_running': {
+          series: [0, 1, 2, 3].map((w) => ({
+            endpoint_url: `http://10.0.0.${w}:7500/metrics`,
+            labels: { worker_id: `w${w}` },
+            timeslices: [
+              { start_ns: w * 0.01e9, end_ns: w * 0.01e9 + 1e9, avg: 2 },
+              { start_ns: w * 0.01e9 + 1e9, end_ns: w * 0.01e9 + 2e9, avg: 2 },
+            ],
+          })),
+        },
+        'vllm:num_requests_waiting': {
+          series: [0, 1, 2, 3].map((w) => ({
+            endpoint_url: `http://10.0.0.${w}:7500/metrics`,
+            labels: { worker_id: `w${w}` },
+            timeslices: [
+              { start_ns: w * 0.01e9, end_ns: w * 0.01e9 + 1e9, avg: 1 },
+              { start_ns: w * 0.01e9 + 1e9, end_ns: w * 0.01e9 + 2e9, avg: 1 },
+            ],
+          })),
+        },
+      },
+    });
+    const cs = await computeChartSeries(gzipSync(Buffer.from(json)));
+    // Four workers x (2 running + 1 waiting) once the last one has reported.
+    expect(cs?.queueDepth.at(-1)).toEqual({ t: 1, running: 8, waiting: 4, total: 12 });
+  });
+
+  it('leaves a single-component metric on a plain one-per-tick grid', async () => {
+    const cs = await computeChartSeries(
+      rateBlob('vllm:generation_tokens', [
+        rateSeriesFor('http://a:8000/metrics', { engine: '0' }, 0, [10, 20, 30]),
+      ]),
+    );
+    expect(cs?.decodeTps).toEqual([
+      { t: 0, value: 10 },
+      { t: 1, value: 20 },
+      { t: 2, value: 30 },
+    ]);
+  });
+});
diff --git a/packages/db/src/etl/compute-chart-series.ts b/packages/db/src/etl/compute-chart-series.ts
index ba2661296..82878ad86 100644
--- a/packages/db/src/etl/compute-chart-series.ts
+++ b/packages/db/src/etl/compute-chart-series.ts
@@ -8,7 +8,7 @@
  * `CHART_SERIES_VERSION` whenever the extraction algorithm changes.
  */
 
-import { gunzipJsonWithinLimit, streamCollectKeys } from './gzip-json-stream';
+import { collectMetricPhases } from './gzip-json-stream';
 import {
   selectServerMetricsAdapter,
   type MetricSource,
@@ -65,8 +65,63 @@ import {
  * warmup block are unaffected. (v11 was a short-lived, since-reverted attempt to
  * carry kvCachePoolTokens in chart_series; that value now lives in
  * benchmark_results.metrics, derived from the server log — unrelated to this.)
+ *
+ * v13: give the KV-cache series a real per-engine identity instead of "one
+ * entry per raw series". Three independent sources of duplication were
+ * inflating `kvCacheUsageByEngine` (up to 8x — 64 lines for a run with 18
+ * real engines) and fragmenting the cluster-average `kvCacheUsage`:
+ *   1. v12's warmup merge concatenates each engine's warmup and profiling
+ *      series, so every engine appeared at least twice — and single-engine
+ *      deployments started drawing a spurious two-line "per-engine" overlay.
+ *   2. vLLM run with multiple API-server frontends exposes the *same* engine
+ *      set on every `/metrics` endpoint, so an 8-rank DP deployment scraped
+ *      from two frontends yielded 16 series for 8 engines.
+ *   3. Tensor-/pipeline-/expert-parallel ranks each report the one KV pool
+ *      they share, so a TP8 worker looked like 8 engines holding identical
+ *      values.
+ * Series are now grouped by their Prometheus label set (see
+ * `seriesIdentityKey`), and the cluster average is a real mean across those
+ * logical engines (see `averageAcrossEngines`) rather than a mean over
+ * whichever engines happened to share an exact `start_ns`.
+ *
+ * v14: give the SUMMED series the same treatment v13 gave the averaged one.
+ * Components of a metric are not scraped in lockstep — a disaggregated run
+ * puts every worker on its own `/metrics` endpoint and its own sub-second
+ * offset, and even a single-endpoint SGLang run splits its token counters into
+ * `is_streaming="true"`/`"false"` series ~16 ms apart — so summing on an exact
+ * `start_ns` emitted one point per component per tick, each carrying that
+ * component's share alone instead of the cluster total. The detail page reads
+ * those as a comb: `rollingAverage` is a sample-count mean, so it averaged the
+ * real samples together with the other components' structural gaps and drew
+ * roughly 1/N of the actual throughput (measured: 2.2x low on an 8-rank SGLang
+ * DP run, ~7x low on the 7-worker disaggregated rows).
+ *
+ * Every summed series — prefill/decode/prefix-hit rates, queue depth, host KV
+ * usage and the prompt-token source breakdown — is now evaluated on one
+ * canonical grid at the blob's native scrape cadence (`canonicalTickNs`), with
+ * each component holding its last sample between its own scrapes
+ * (`sumOntoGrid`). The grid has to be uniform, not the union of scrape
+ * instants the way `averageAcrossEngines` does it: a mean is scale-free but a
+ * sum is not, and `cumulativeUniqueInputTokens` turns these rates into token
+ * totals with `sum += value`, which only stays correct at one point per
+ * scrape bucket.
+ *
+ * v15: a metric SOURCE is one worker, not one engine inside it. The source
+ * identity carried `dp_rank` (sglang) and `engine` (vllm), which name engines
+ * *within* a worker, so the detail page's endpoint picker listed each worker
+ * once per rank: a 6-prefill/1-decode run offered 35 "endpoints" for 7
+ * workers, reading `Prefill - <id>` five times in a row. SGLang also emits an
+ * unlabelled aggregate series alongside the per-rank ones, which became a
+ * fifth entry for the same worker; vllm leaves `dp_rank` unset and duplicates
+ * under `engine` instead, so the same bug wore a different label per
+ * framework. Identity is now `(role, worker)` — see `seriesIdentityKey`'s
+ * sibling reasoning in `server-metrics-adapters.ts`. Each worker's ranks
+ * aggregate into its one source (throughput summed, KV averaged) and the
+ * per-rank detail stays reachable as that source's own
+ * `kvCacheUsageByEngine`.
+ *
  */
-export const CHART_SERIES_VERSION = 12;
+export const CHART_SERIES_VERSION = 15;
 
 export interface TimeSeriesPoint {
   /** Seconds from benchmark start. */
@@ -111,9 +166,10 @@ export interface ChartSeries {
    */
   hostKvCacheUsage: TimeSeriesPoint[];
   /**
-   * Per-DP-rank KV cache utilization (0..1 each). One entry per engine
-   * series found in the raw metric, ordered by the `engine` label when
-   * present and by series-array index otherwise. Empty for single-engine
+   * Per-DP-rank KV cache utilization (0..1 each). One entry per LOGICAL
+   * engine — one KV pool — not per raw series; see `resolveLogicalEngines`
+   * for how mirrored endpoints, phase blocks and shard ranks collapse.
+   * Ordered by role, then numeric rank, then worker. Empty for single-engine
    * deployments — the average `kvCacheUsage` line covers that case alone.
    * The detail page overlays these on the same chart so DEP load skew is
    * visible without changing the headline number.
@@ -209,32 +265,18 @@ function mergePhaseMetrics(profiling: MetricsMap, warmup: MetricsMap): MetricsMa
   return out;
 }
 
-/**
- * Stream-parse fallback: collect the chart's metric subtrees from both phase
- * blocks and merge (see v11) when the full JSON exceeds the in-memory
- * fast-path ceiling.
- */
-async function streamCollectMetrics(buffer: Buffer): Promise<MetricsMap> {
-  const [profiling, warmup] = await Promise.all([
-    streamCollectKeys<RawMetric>(buffer, 'metrics', CHART_METRIC_KEYS),
-    streamCollectKeys<RawMetric>(buffer, 'warmup_metrics', CHART_METRIC_KEYS),
-  ]);
-  return mergePhaseMetrics(profiling, warmup);
-}
-
 /**
  * Parse the gzipped server_metrics blob into the metric map. Small blobs use
- * the synchronous fast path; oversized blobs use the streaming parser. Merges
- * the warmup block into the profiling one (v11) so the series span both phases.
+ * the synchronous fast path; oversized blobs use one streaming parser pass
+ * that collects both phase blocks. Merges the warmup block into the profiling
+ * one (v11) so the series span both phases.
  */
 async function parseMetrics(buffer: Buffer): Promise<MetricsMap> {
-  const json = gunzipJsonWithinLimit(buffer);
-  if (json === null) return await streamCollectMetrics(buffer);
-  const obj = JSON.parse(json) as {
-    metrics?: MetricsMap;
-    warmup_metrics?: MetricsMap;
-  };
-  return mergePhaseMetrics(obj.metrics ?? {}, obj.warmup_metrics ?? {});
+  const { metrics, warmupMetrics } = await collectMetricPhases<RawMetric>(
+    buffer,
+    CHART_METRIC_KEYS,
+  );
+  return mergePhaseMetrics(metrics, warmupMetrics);
 }
 
 /**
@@ -270,37 +312,619 @@ export function computeChartSeriesFromMetricPhases(
   return buildSeriesFromMetrics(mergePhaseMetrics(profiling, warmup), context);
 }
 
+// Note: v14 removed `aggregateByStart`/`sortedEntries`. Nothing groups on an
+// exact `start_ns` any more — averaged series go through `averageAcrossEngines`
+// and summed ones through `sumOntoGrid`, both of which treat a component's
+// samples as a step function rather than requiring components to share a
+// nanosecond.
+
+// ── Per-engine identity (v13) ───────────────────────────────────────────
+//
+// A "logical engine" is one KV-cache pool: one DP rank of one worker. The blob
+// stores one `RawSeries` per (scrape endpoint × phase block × label set), which
+// is NOT the same thing — see the v13 note at the top of the file.
+
+/**
+ * Ranks that shard ONE engine rather than naming a separate one. A KV cache is
+ * allocated per engine and shared by its tensor-, pipeline- and expert-parallel
+ * ranks, so every such rank reports the same pool: on a TP8 SGLang prefill
+ * worker the eight `tp_rank` series agree to ~4 decimal places on average
+ * (with rare single-scrape transients where one rank spikes alone).
+ * Treating them as separate engines drew eight near-identical lines per pool.
+ *
+ * `engine` / `engine_idx` / `dp_rank` are deliberately NOT here — those DO name
+ * distinct pools (one per DP rank / engine core).
+ */
+const INTRA_ENGINE_SHARD_LABELS = new Set(['tp_rank', 'pp_rank', 'ep_rank', 'moe_ep_rank']);
+
+/**
+ * Identity of a metric series, following Prometheus semantics: the label set
+ * IS the series, and the scrape endpoint is transport rather than identity.
+ * Two `/metrics` endpoints exposing `{engine="3", model_name="X"}` are two
+ * views of one engine — which is exactly what vLLM does when it runs several
+ * API-server frontends over one DP engine group.
+ *
+ * Deployments whose endpoints really are distinct engines say so in the
+ * labels: Dynamo tags every series with `worker_id` (plus `dynamo_component`
+ * / `engine_type`), so prefill worker rank 0 and decode worker rank 0 keep
+ * separate identities here.
+ *
+ * Shard ranks are excluded (see `INTRA_ENGINE_SHARD_LABELS`). Series left with
+ * no labels at all fall back to the endpoint, which keeps label-less workers
+ * apart rather than silently fusing them.
+ */
+function seriesIdentityKey(s: RawSeries): string {
+  const labels = s.labels ?? {};
+  const names = Object.keys(labels)
+    .filter((name) => !INTRA_ENGINE_SHARD_LABELS.has(name))
+    .toSorted();
+  if (names.length === 0) return `@${s.endpoint_url ?? ''}`;
+  // Join on control characters so a name or value containing '=' or ','
+  // cannot forge another label set's key.
+  return names.map((name) => `${name}\u0001${labels[name]}`).join('\u0002');
+}
+
+/** Dynamo/SGLang name their roles differently; normalize for display. */
+const ENGINE_ROLE_BY_NATIVE_LABEL: Record<string, string> = {
+  prefill: 'prefill',
+  decode: 'decode',
+  backend: 'decode',
+};
+
+/** Label lookup that treats blank values as absent. */
+function labelOrNull(labels: Record<string, string>, ...names: string[]): string | null {
+  for (const name of names) {
+    const value = labels[name]?.trim();
+    if (value) return value;
+  }
+  return null;
+}
+
+/** DP-rank-ish label under any of the names the frameworks emit. */
+function engineRankLabel(labels: Record<string, string>): string | null {
+  return labelOrNull(labels, 'engine', 'engine_idx', 'dp_rank');
+}
+
+function engineRoleLabel(labels: Record<string, string>): string | null {
+  // Try each source in turn rather than taking the first that EXISTS: an
+  // aggregated dynamo-sglang worker carries engine_type="unified" (which maps
+  // to no role) alongside dynamo_component="backend", and stopping at the
+  // first present label would resolve the role to null.
+  for (const name of ['engine_type', 'dynamo_component']) {
+    const native = labelOrNull(labels, name);
+    const role = native ? ENGINE_ROLE_BY_NATIVE_LABEL[native] : undefined;
+    if (role) return role;
+  }
+  return null;
+}
+
+/** Rank as a sort key. Only plain digit strings sort numerically. */
+function engineRankSortKey(rank: string | null): number {
+  return rank !== null && /^\d+$/u.test(rank) ? Number(rank) : Number.POSITIVE_INFINITY;
+}
+
+/** Prefill before decode; unroled engines last. Stable across runs. */
+const ROLE_SORT_ORDER: Record<string, number> = { prefill: 0, decode: 1 };
+function engineRoleSortKey(role: string | null): number {
+  return role === null ? 2 : (ROLE_SORT_ORDER[role] ?? 2);
+}
+
+/** `scheme://host:port/path` -> `host:port`, the human-facing part. */
+function endpointHostPort(endpointUrl: string): string | null {
+  const hostPort = /^\w+:\/\/(?<hostPort>[^/]+)/u.exec(endpointUrl)?.groups?.['hostPort'];
+  return hostPort ?? (endpointUrl || null);
+}
+
+/**
+ * Short, human-readable tiebreaker for engines that would otherwise share a
+ * display label (e.g. two decode workers that each number their ranks 0..7).
+ * `preferEndpoint` is set when the engines being separated came from the same
+ * worker, so the worker id cannot tell them apart.
+ */
+function engineDiscriminator(
+  labels: Record<string, string>,
+  endpointUrl: string,
+  preferEndpoint = false,
+): string | null {
+  const worker = preferEndpoint ? null : labelOrNull(labels, 'worker_id');
+  if (worker) return worker.length > 4 ? worker.slice(-4) : worker;
+  return endpointHostPort(endpointUrl);
+}
+
+/**
+ * Absolute gap between two endpoints' whole-run means, on a 0..1 gauge, below
+ * which they are treated as mirrors of one engine rather than two engines
+ * sharing a label set.
+ *
+ * Measured mirrors sit far under this: the three two-endpoint vLLM configs in
+ * the corpus differ by 0.03%-2.19% of their means (under 0.001 absolute, even
+ * on the heavily loaded rows), while a genuinely distinct prefill and decode
+ * worker differ by ~0.08. So both sides have roughly 4x of margin.
+ *
+ * Residual limitation, accepted deliberately: two independent replicas behind
+ * a round-robin router would have similar means BY DESIGN and would still be
+ * fused, showing one line instead of two. That case degrades the per-engine
+ * overlay but not the cluster average — averaging two engines that track each
+ * other gives the same number either way — whereas the case this does catch
+ * (replicas under uneven load) is the one where fusing would make the average
+ * itself wrong. Distinguishing the former needs lag-aligned pointwise
+ * comparison, which no data in the corpus currently justifies.
+ */
+const MIRROR_MEAN_TOLERANCE = 0.02;
+
 /**
- * Aggregate one timeslice field across all series of a metric, indexed by
- * `start_ns`. Multi-engine vllm deployments report one series per engine —
- * the cluster value is the sum (for running/waiting/throughput counters)
- * or the average (for kv_cache_usage_perc, a per-engine fraction).
+ * Same idea as `MIRROR_MEAN_TOLERANCE`, but for counter rates (v14), where an
+ * absolute threshold is meaningless — a token-throughput mean is O(10^5), so
+ * the gauge tolerance would never fire and every mirror would be counted twice.
+ *
+ * Expressed as a fraction of the larger mean. The measured mirrors in the
+ * corpus differ by 0.03%-2.19% of their means, and a router endpoint that
+ * aggregates several workers differs from any single worker by far more than
+ * 5%, so this sits above the observed mirror spread with room to spare while
+ * staying well below a genuine difference.
+ */
+const MIRROR_RATE_RELATIVE_TOLERANCE = 0.05;
+
+/**
+ * Do these endpoints' whole-run means agree closely enough to be two views of
+ * one series rather than two different ones?
+ *
+ * Gauges keep v13's absolute test unchanged — they are 0..1 fractions (or small
+ * request counts) whose mirrors agree almost exactly. Rates need a relative
+ * test because their magnitude is unbounded.
  */
-function aggregateByStart(
+function looksMirrored(means: readonly number[], field: 'avg' | 'rate'): boolean {
+  const spread = Math.max(...means) - Math.min(...means);
+  if (field === 'avg') return spread <= MIRROR_MEAN_TOLERANCE;
+  const scale = Math.max(...means.map((m) => Math.abs(m)));
+  // Two endpoints that both read ~0 are mirrors of an idle series, not two
+  // engines: fall back to the absolute test so the relative one can't divide
+  // by a vanishing scale.
+  if (scale === 0) return true;
+  return spread <= MIRROR_RATE_RELATIVE_TOLERANCE * scale;
+}
+
+interface LogicalEngine {
+  engineLabel: string;
+  points: TimeSeriesPoint[];
+}
+
+/** One endpoint's samples for one identity, keyed by scrape instant. */
+type ScrapeMap = Map<number, { sum: number; count: number }>;
+
+interface EngineGroup {
+  labels: Record<string, string>;
+  byEndpoint: Map<string, ScrapeMap>;
+}
+
+/** An engine after endpoint resolution, before its display label is composed. */
+interface ResolvedEngine {
+  rank: string | null;
+  role: string | null;
+  discriminator: string | null;
+  /** Which worker owns this engine — `worker_id`, else the endpoint. */
+  workerKey: string | null;
+  order: number;
+  points: TimeSeriesPoint[];
+}
+
+function scrapesToPoints(scrapes: ScrapeMap, tOf: (ns: number) => number): TimeSeriesPoint[] {
+  return [...scrapes.entries()]
+    .toSorted((a, b) => a[0] - b[0])
+    .map(([startNs, { sum, count }]) => ({ t: tOf(startNs), value: sum / count }));
+}
+
+function meanOf(scrapes: ScrapeMap): number {
+  let total = 0;
+  for (const { sum, count } of scrapes.values()) total += sum / count;
+  return scrapes.size === 0 ? 0 : total / scrapes.size;
+}
+
+/** Wall-clock span covered, so a dense-but-truncated mirror can't win. */
+function spanOf(scrapes: ScrapeMap): number {
+  let lo = Number.POSITIVE_INFINITY;
+  let hi = Number.NEGATIVE_INFINITY;
+  for (const ns of scrapes.keys()) {
+    if (ns < lo) lo = ns;
+    if (ns > hi) hi = ns;
+  }
+  return hi >= lo ? hi - lo : 0;
+}
+
+/**
+ * Collapse a gauge's raw series into one entry per logical engine.
+ *
+ * The three kinds of duplication need three different treatments:
+ *   - Same endpoint, different phase blocks (v12's warmup merge): keying by
+ *     scrape instant unions them into one series. The blocks' first/last
+ *     bounds can look overlapping (profiling often emits one boundary sample
+ *     then gaps until warmup ends) but they never share an instant, so the
+ *     union neither drops nor double-counts a scrape.
+ *   - Same endpoint, same instant: intra-engine shard ranks reporting the one
+ *     pool they share, so they collapse to their mean (they agree to ~4 decimal
+ *     places on average, with rare single-scrape transients).
+ *   - Different endpoints reporting the same identity: usually mirrored
+ *     API-server frontends carrying the same measurement a few hundred ms
+ *     apart. Merging those would interleave near-duplicate samples and halve
+ *     the effective span of the frontend's fixed-width rolling average, so the
+ *     best-covered endpoint wins and the rest are dropped — but ONLY when their
+ *     values agree. Endpoints that disagree are genuinely different engines
+ *     behind one label (a router in front of several replicas), and dropping
+ *     one would silently lose an engine, so those are kept separate instead.
+ */
+function resolveComponents(
   series: readonly RawSeries[] | undefined,
-  field: 'avg' | 'rate',
-  combine: 'sum' | 'avg',
-): Map<number, number> {
-  const sums = new Map<number, number>();
-  const counts = new Map<number, number>();
+  tOf: (ns: number) => number,
+  field: 'avg' | 'rate' = 'avg',
+): ResolvedEngine[] {
+  const groups = new Map<string, EngineGroup>();
   for (const s of series ?? []) {
+    const key = seriesIdentityKey(s);
+    let group = groups.get(key);
+    if (!group) {
+      group = { labels: s.labels ?? {}, byEndpoint: new Map() };
+      groups.set(key, group);
+    }
+    const endpoint = s.endpoint_url ?? '';
+    let scrapes = group.byEndpoint.get(endpoint);
+    if (!scrapes) {
+      scrapes = new Map();
+      group.byEndpoint.set(endpoint, scrapes);
+    }
     for (const ts of s.timeslices ?? []) {
-      if (typeof ts.start_ns !== 'number') continue;
-      const v = ts[field];
-      if (typeof v !== 'number' || !Number.isFinite(v)) continue;
-      sums.set(ts.start_ns, (sums.get(ts.start_ns) ?? 0) + v);
-      counts.set(ts.start_ns, (counts.get(ts.start_ns) ?? 0) + 1);
+      if (typeof ts.start_ns !== 'number' || !Number.isFinite(ts.start_ns)) continue;
+      const value = ts[field];
+      if (typeof value !== 'number' || !Number.isFinite(value)) continue;
+      const at = scrapes.get(ts.start_ns);
+      if (at) {
+        at.sum += value;
+        at.count++;
+      } else {
+        scrapes.set(ts.start_ns, { sum: value, count: 1 });
+      }
     }
   }
-  if (combine === 'sum') return sums;
-  const out = new Map<number, number>();
-  for (const [t, s] of sums) out.set(t, s / (counts.get(t) ?? 1));
+
+  const resolved: ResolvedEngine[] = [];
+  for (const group of groups.values()) {
+    const endpoints = [...group.byEndpoint]
+      .filter(([, scrapes]) => scrapes.size > 0)
+      // Endpoint URL keeps the walk deterministic before any ranking.
+      .toSorted((a, b) => (a[0] < b[0] ? -1 : a[0] > b[0] ? 1 : 0));
+    if (endpoints.length === 0) continue;
+
+    const rank = engineRankLabel(group.labels);
+    const role = engineRoleLabel(group.labels);
+    const workerId = labelOrNull(group.labels, 'worker_id');
+    const push = (endpointUrl: string, scrapes: ScrapeMap, preferEndpoint: boolean) => {
+      resolved.push({
+        rank,
+        role,
+        discriminator: engineDiscriminator(group.labels, endpointUrl, preferEndpoint),
+        workerKey: workerId ?? (endpointUrl || null),
+        order: resolved.length,
+        points: scrapesToPoints(scrapes, tOf),
+      });
+    };
+
+    if (endpoints.length === 1) {
+      push(endpoints[0]![0], endpoints[0]![1], false);
+      continue;
+    }
+
+    const means = endpoints.map(([, scrapes]) => meanOf(scrapes));
+    const mirrored = looksMirrored(means, field);
+    if (!mirrored) {
+      // Same labels, different measurements: distinct engines, not mirrors.
+      for (const [endpointUrl, scrapes] of endpoints) push(endpointUrl, scrapes, true);
+      continue;
+    }
+    // Mirrors: keep the endpoint that covers the most wall-clock, breaking
+    // ties on sample count and then URL, so a dense but truncated mirror
+    // cannot shorten the engine's series.
+    const best = endpoints.reduce((a, b) => {
+      const sa = spanOf(a[1]);
+      const sb = spanOf(b[1]);
+      if (sb !== sa) return sb > sa ? b : a;
+      if (b[1].size !== a[1].size) return b[1].size > a[1].size ? b : a;
+      return a;
+    });
+    push(best[0], best[1], false);
+  }
+
+  // Sort on the identity's components, never on the composed string: role
+  // first, then numeric rank, then worker, then blob order. Sorting a label
+  // like "decode 3" lexically (or falling back to array index) scrambled DP
+  // ranks on multi-worker runs.
+  return resolved.toSorted(
+    (a, b) =>
+      engineRoleSortKey(a.role) - engineRoleSortKey(b.role) ||
+      engineRankSortKey(a.rank) - engineRankSortKey(b.rank) ||
+      (a.discriminator ?? '').localeCompare(b.discriminator ?? '') ||
+      a.order - b.order,
+  );
+}
+
+/** Collapse a gauge's raw series into one labelled entry per logical engine. */
+function resolveLogicalEngines(
+  series: readonly RawSeries[] | undefined,
+  tOf: (ns: number) => number,
+): LogicalEngine[] {
+  const ordered = resolveComponents(series, tOf, 'avg');
+
+  // Only name the role when there is more than one, otherwise every engine on
+  // an aggregated deployment reads "decode 0", "decode 1", ... for no reason.
+  const roles = new Set(ordered.map((e) => e.role).filter((r) => r !== null));
+  const showRole = roles.size > 1;
+  const withBase = ordered.map((engine, idx) => {
+    const named =
+      showRole && engine.role
+        ? engine.rank === null
+          ? engine.role
+          : `${engine.role} ${engine.rank}`
+        : engine.rank;
+    // Nothing rank- or role-like to go on: the worker/endpoint is the only
+    // thing that names this engine, so lead with it rather than a bare index.
+    return { ...engine, base: named ?? engine.discriminator ?? `#${idx}` };
+  });
+
+  // Qualify collisions (e.g. two decode workers that each number their ranks
+  // 0..7) so every legend entry names exactly one line.
+  const baseCounts = new Map<string, number>();
+  for (const engine of withBase) {
+    baseCounts.set(engine.base, (baseCounts.get(engine.base) ?? 0) + 1);
+  }
+  const used = new Set<string>();
+  return withBase.map((engine) => {
+    let label = engine.base;
+    if ((baseCounts.get(engine.base) ?? 0) > 1 && engine.discriminator) {
+      label = `${engine.base} (${engine.discriminator})`;
+    }
+    // The discriminator isn't guaranteed unique either; fall back to a counter
+    // so a legend entry never stands for two lines.
+    let candidate = label;
+    for (let n = 2; used.has(candidate); n++) candidate = `${label} #${n}`;
+    used.add(candidate);
+    return { engineLabel: candidate, points: engine.points };
+  });
+}
+
+/**
+ * Mean utilization across logical engines on the union of their scrape times.
+ *
+ * Engines are not scraped in lockstep — different workers (and occasionally a
+ * single lagging rank) sit on their own sub-second grid — so grouping on an
+ * exact `start_ns` would average whichever subset happened to share that
+ * nanosecond. On a disaggregated run that means alternating between
+ * "prefill only" and "decode only", which reads as a full-scale sawtooth
+ * rather than a cluster average.
+ *
+ * Each engine therefore holds its last scrape until its next one (a gauge
+ * keeps its value between scrapes) and contributes only inside its own
+ * observed window, so an engine that starts late or stops early neither
+ * pulls the mean toward a stale value nor drops it to zero. A hole in the
+ * middle of that window is bounded too — see `carryLimitSeconds`.
+ */
+function averageAcrossEngines(
+  engines: readonly { points: TimeSeriesPoint[] }[],
+): TimeSeriesPoint[] {
+  const active = engines.filter((engine) => engine.points.length > 0);
+  if (active.length === 0) return [];
+  // Single engine: its own samples already are the cluster average.
+  if (active.length === 1) return active[0]!.points;
+
+  const timeline = [...new Set(active.flatMap((e) => e.points.map((p) => p.t)))].toSorted(
+    (a, b) => a - b,
+  );
+  const cursors: number[] = Array.from({ length: active.length }, () => -1);
+  const lastT = active.map((engine) => engine.points.at(-1)!.t);
+  const carryLimit = active.map((engine) => carryLimitSeconds(engine.points));
+  const out: TimeSeriesPoint[] = [];
+  for (const t of timeline) {
+    let sum = 0;
+    let n = 0;
+    for (const [i, engine] of active.entries()) {
+      const points = engine.points;
+      let cursor = cursors[i]!;
+      while (cursor + 1 < points.length && points[cursor + 1]!.t <= t) cursor++;
+      cursors[i] = cursor;
+      // Before this engine's first scrape or after its last — no value to
+      // carry, so it sits out of this tick's mean entirely.
+      if (cursor < 0 || t > lastT[i]!) continue;
+      // Inside the window but far past the last sample: the engine stopped
+      // reporting for a while, so don't average in a stale reading.
+      if (t - points[cursor]!.t > carryLimit[i]!) continue;
+      sum += points[cursor]!.value;
+      n++;
+    }
+    if (n > 0) out.push({ t, value: sum / n });
+  }
   return out;
 }
 
-/** Stable order: emit one point per unique start_ns, chronologically. */
-function sortedEntries(m: Map<number, number>): [number, number][] {
-  return [...m.entries()].toSorted((a, b) => a[0] - b[0]);
+/**
+ * How long one engine's last sample may stand in for it: 5x its own median
+ * scrape gap. A dropped scrape or two still carries (real runs sit at 1 Hz
+ * with gaps never above ~1 s), but a long reporting hole drops the engine
+ * out of the mean instead of pinning it to a minutes-old value.
+ */
+function carryLimitSeconds(points: readonly TimeSeriesPoint[]): number {
+  if (points.length < 2) return Number.POSITIVE_INFINITY;
+  const gaps: number[] = [];
+  for (let i = 1; i < points.length; i++) {
+    const gap = points[i]!.t - points[i - 1]!.t;
+    if (gap > 0) gaps.push(gap);
+  }
+  if (gaps.length === 0) return Number.POSITIVE_INFINITY;
+  gaps.sort((a, b) => a - b);
+  return 5 * gaps[gaps.length >> 1]!;
+}
+
+// ── Canonical scrape grid for summed series (v14) ───────────────────────
+//
+// `averageAcrossEngines` can evaluate on the union of scrape instants because
+// a mean is scale-free: adding more evaluation points doesn't change the level.
+// A SUM cannot. Its consumers on the detail page treat one point as one
+// one-second bucket — `cumulativeUniqueInputTokens` does `sum += value` to turn
+// a token *rate* into a token *total*, and `rollingAverage` is a sample-count
+// mean — so the emitted grid has to be uniform and at the native scrape
+// cadence, or the totals and the levels both move.
+
+/** Widest interval we will ever call a scrape cadence, in ns (5 minutes). */
+const MAX_TICK_NS = 300 * 1e9;
+/** Hard cap on emitted ticks, so a malformed cadence can't blow up the row. */
+const MAX_GRID_TICKS = 500_000;
+/**
+ * Slack when snapping a component's first/last sample to the lattice, as a
+ * fraction of a tick. A sample sitting a floating-point hair outside the grid
+ * should still get its tick rather than losing an endpoint.
+ */
+const GRID_EPSILON = 1e-6;
+
+/**
+ * The blob's native scrape interval, in nanoseconds.
+ *
+ * Taken as the median gap of the single series with the most timeslices: that
+ * series is by construction the best-sampled view of the run, and a median
+ * shrugs off the one-off boundary gap that the profiling block leaves where
+ * warmup used to be. Returns null when nothing in the blob has two samples to
+ * compare, in which case callers fall back to the pre-v14 exact-instant path.
+ */
+function canonicalTickNs(metrics: MetricsMap): number | null {
+  let best: readonly RawSlice[] | null = null;
+  for (const metricMeta of Object.values(metrics)) {
+    for (const s of metricMeta?.series ?? []) {
+      const ts = s.timeslices ?? [];
+      if (ts.length > (best?.length ?? 0)) best = ts;
+    }
+  }
+  if (!best || best.length < 2) return null;
+  const gaps: number[] = [];
+  for (let i = 1; i < best.length; i++) {
+    const a = best[i - 1]!.start_ns;
+    const b = best[i]!.start_ns;
+    if (typeof a !== 'number' || typeof b !== 'number') continue;
+    const gap = b - a;
+    if (gap > 0 && gap <= MAX_TICK_NS) gaps.push(gap);
+  }
+  if (gaps.length === 0) return null;
+  gaps.sort((a, b) => a - b);
+  const median = gaps[gaps.length >> 1]!;
+  return median > 0 ? median : null;
+}
+
+/**
+ * Evaluate a step-held component at `t`, or null when it has nothing to say.
+ *
+ * `cursor` is carried across calls so the whole sweep stays linear; callers
+ * must visit `t` in ascending order.
+ */
+interface StepCursor {
+  points: readonly TimeSeriesPoint[];
+  carryLimit: number;
+  cursor: number;
+}
+
+function stepValueAt(state: StepCursor, t: number): number | null {
+  const { points } = state;
+  while (state.cursor + 1 < points.length && points[state.cursor + 1]!.t <= t) state.cursor++;
+  // Outside the component's own observed window it contributes nothing, so a
+  // worker that starts late or stops early neither invents load nor forces a
+  // zero into the cluster total.
+  if (state.cursor < 0 || t > points.at(-1)!.t) return null;
+  // Inside the window but long past the last sample: the component stopped
+  // reporting, so don't carry a stale reading across the hole.
+  if (t - points[state.cursor]!.t > state.carryLimit) return null;
+  return points[state.cursor]!.value;
+}
+
+/**
+ * Sum components onto a uniform grid at the native scrape cadence.
+ *
+ * Components of one metric are NOT scraped in lockstep: a disaggregated run
+ * gives every worker its own `/metrics` endpoint on its own sub-second offset,
+ * and even a single-endpoint SGLang run splits its token counters into
+ * `is_streaming="true"`/`"false"` series that sit ~16 ms apart. Summing on an
+ * exact `start_ns` therefore emitted one point per component per tick, each
+ * holding that component's share alone. Downstream that reads as a comb —
+ * `rollingAverage` averages the real samples together with the other
+ * components' structural gaps, so an 8-worker run's throughput chart drew
+ * roughly an eighth of the actual tokens/sec.
+ *
+ * Each component instead holds its last sample between scrapes (its rate is a
+ * step function over its bucket) and the grid samples the sum once per tick.
+ * Returns an empty array when there is nothing to place on a grid.
+ */
+function sumOntoGrid(
+  components: readonly (readonly TimeSeriesPoint[])[],
+  tickS: number | null,
+): TimeSeriesPoint[] {
+  const active = components.filter((points) => points.length > 0);
+  if (active.length === 0) return [];
+  const first = Math.min(...active.map((points) => points[0]!.t));
+  const last = Math.max(...active.map((points) => points.at(-1)!.t));
+  if (!Number.isFinite(first) || !Number.isFinite(last)) return [];
+
+  // Every summed series shares one lattice anchored at t=0, so metrics that
+  // are divided by one another downstream (hit rate = hits/queries, host KV
+  // usage = used/total, queue depth = running+waiting) land on identical `t`
+  // values and can be joined. Anchoring each metric at its OWN first sample
+  // instead silently emptied the hit-rate chart, because `sglang:cached_tokens`
+  // starts ~0.18 s off the grid `sglang:prompt_tokens` starts on and the two
+  // lattices then never share a point.
+  //
+  // With no inferable cadence (every series carrying a single timeslice, as in
+  // the smallest fixtures) fall back to the union of the components' own
+  // instants: still a step-held sum rather than an exact-instant one, just
+  // without a regular grid to place it on.
+  let timeline: number[];
+  if (tickS && tickS > 0) {
+    const firstTick = Math.ceil(first / tickS - GRID_EPSILON);
+    const lastTick = Math.floor(last / tickS + GRID_EPSILON);
+    const ticks = lastTick - firstTick + 1;
+    if (ticks < 1 || ticks > MAX_GRID_TICKS) return [];
+    timeline = Array.from({ length: ticks }, (_, i) => (firstTick + i) * tickS);
+  } else {
+    timeline = [...new Set(active.flatMap((points) => points.map((p) => p.t)))].toSorted(
+      (a, b) => a - b,
+    );
+  }
+
+  const states: StepCursor[] = active.map((points) => ({
+    points,
+    carryLimit: carryLimitSeconds(points),
+    cursor: -1,
+  }));
+  const out: TimeSeriesPoint[] = [];
+  for (const t of timeline) {
+    let sum = 0;
+    let n = 0;
+    for (const state of states) {
+      const value = stepValueAt(state, t);
+      if (value === null) continue;
+      sum += value;
+      n++;
+    }
+    // No component covers this tick — emit nothing rather than a false zero.
+    if (n > 0) out.push({ t, value: sum });
+  }
+  return out;
+}
+
+/** `sumOntoGrid` over a metric's raw series, resolved to logical components. */
+function summedSeries(
+  series: readonly RawSeries[] | undefined,
+  tOf: (ns: number) => number,
+  field: 'avg' | 'rate',
+  tickS: number | null,
+): TimeSeriesPoint[] {
+  const components = resolveComponents(series, tOf, field).map((c) => c.points);
+  return sumOntoGrid(components, tickS);
+}
+
+/** Index a grid series by `t` so two of them can be divided tick-for-tick. */
+function byT(points: readonly TimeSeriesPoint[]): Map<number, number> {
+  return new Map(points.map((p) => [p.t, p.value]));
 }
 
 function buildSeriesFromMetrics(
@@ -329,6 +953,12 @@ function buildSeriesFromMetrics(
   }
   if (!Number.isFinite(startNs)) startNs = 0;
   const tOf = (ns: number) => (ns - (originStartNs ?? startNs)) / 1e9;
+  // Grid spacing for every summed series (v14). Null when the blob has nothing
+  // with two comparable samples, in which case `sumOntoGrid` emits only the
+  // single-component case and multi-component metrics come out empty rather
+  // than fragmented.
+  const tickNs = canonicalTickNs(metrics);
+  const tickS = tickNs === null ? null : tickNs / 1e9;
 
   // Pick the first metric name whose series array has any data; fallback
   // chain lets the same code path serve both vllm:* and sglang:* blobs.
@@ -347,33 +977,15 @@ function buildSeriesFromMetrics(
     'vllm:gpu_cache_usage_perc',
     'sglang:token_usage',
   );
-  const kvCacheUsage: TimeSeriesPoint[] = sortedEntries(
-    aggregateByStart(kvSeries, 'avg', 'avg'),
-  ).map(([t, v]) => ({ t: tOf(t), value: v }));
-  // Per-engine breakdown of the same metric. We only emit it when there's
-  // more than one series — single-engine deployments would just duplicate
-  // the cluster-average line.
-  const kvCacheUsageByEngine: { engineLabel: string; points: TimeSeriesPoint[] }[] = [];
-  if (kvSeries && kvSeries.length > 1) {
-    // Sort by numeric engine label when present so rank 0..N renders in
-    // order; fall back to series-array index otherwise.
-    const decorated = kvSeries.map((s, idx) => {
-      const raw =
-        s.labels?.['engine'] ?? s.labels?.['engine_idx'] ?? s.labels?.['dp_rank'] ?? String(idx);
-      const numeric = Number(raw);
-      return { series: s, idx, label: raw, sortKey: Number.isFinite(numeric) ? numeric : idx };
-    });
-    decorated.sort((a, b) => a.sortKey - b.sortKey);
-    for (const { series, label } of decorated) {
-      const pts: TimeSeriesPoint[] = [];
-      for (const ts of series.timeslices ?? []) {
-        if (typeof ts.start_ns !== 'number' || typeof ts.avg !== 'number') continue;
-        if (!Number.isFinite(ts.avg)) continue;
-        pts.push({ t: tOf(ts.start_ns), value: ts.avg });
-      }
-      if (pts.length > 0) kvCacheUsageByEngine.push({ engineLabel: label, points: pts });
-    }
-  }
+  // One entry per logical engine (v13) — mirrored API-server frontends and the
+  // warmup/profiling phase split are collapsed here rather than showing up as
+  // extra "engines".
+  const engines = resolveLogicalEngines(kvSeries, tOf);
+  const kvCacheUsage: TimeSeriesPoint[] = averageAcrossEngines(engines);
+  // Per-engine breakdown of the same metric. Emitted only for genuinely
+  // multi-engine deployments — with one engine it would just duplicate the
+  // cluster-average line.
+  const kvCacheUsageByEngine = engines.length > 1 ? engines : [];
 
   // Prefix cache hit rate per scrape: Σhits.rate / Σqueries.rate across
   // engines, joined on start_ns. SGLang names: cached_tokens / prompt_tokens.
@@ -383,19 +995,21 @@ function buildSeriesFromMetrics(
     'vllm:prompt_tokens',
     'sglang:prompt_tokens',
   );
-  const hitsByT = aggregateByStart(hitsSeries, 'rate', 'sum');
-  const qsByT = aggregateByStart(qsSeries, 'rate', 'sum');
+  const hitsOnGrid = summedSeries(hitsSeries, tOf, 'rate', tickS);
+  const qsOnGrid = summedSeries(qsSeries, tOf, 'rate', tickS);
+  const qsByT = byT(qsOnGrid);
   const prefixCacheHitRate: TimeSeriesPoint[] = [];
-  for (const [t, h] of sortedEntries(hitsByT)) {
+  for (const { t, value: h } of hitsOnGrid) {
     const q = qsByT.get(t);
-    if (q !== undefined && q > 0) prefixCacheHitRate.push({ t: tOf(t), value: h / q });
+    if (q !== undefined && q > 0) prefixCacheHitRate.push({ t, value: h / q });
   }
 
   // Queue depth: sum running + waiting across engines per timeslice.
   const runSeries = pickSeries('vllm:num_requests_running', 'sglang:num_running_reqs');
   const waitSeries = pickSeries('vllm:num_requests_waiting', 'sglang:num_queue_reqs');
-  const runByT = aggregateByStart(runSeries, 'avg', 'sum');
-  const waitByT = aggregateByStart(waitSeries, 'avg', 'sum');
+  const runOnGrid = summedSeries(runSeries, tOf, 'avg', tickS);
+  const waitByT = byT(summedSeries(waitSeries, tOf, 'avg', tickS));
+  const runByT = byT(runOnGrid);
   const queueDepth: QueueDepthPoint[] = [];
   // Union of timestamps so we surface activity even if one of the gauges
   // didn't report a sample on a given tick.
@@ -403,18 +1017,14 @@ function buildSeriesFromMetrics(
   for (const t of [...allTimes].toSorted((a, b) => a - b)) {
     const running = runByT.get(t) ?? 0;
     const waiting = waitByT.get(t) ?? 0;
-    queueDepth.push({ t: tOf(t), running, waiting, total: running + waiting });
+    queueDepth.push({ t, running, waiting, total: running + waiting });
   }
 
-  // Throughput: sum the counter `rate` (already per-second) across engines.
-  // Takes a fallback chain so vllm:* and sglang:* both work.
-  const counterRate = (...names: string[]): TimeSeriesPoint[] => {
-    const s = pickSeries(...names);
-    return sortedEntries(aggregateByStart(s, 'rate', 'sum')).map(([t, v]) => ({
-      t: tOf(t),
-      value: v,
-    }));
-  };
+  // Throughput: sum the counter `rate` (already per-second) across engines,
+  // on the canonical grid. Takes a fallback chain so vllm:* and sglang:* both
+  // work.
+  const counterRate = (...names: string[]): TimeSeriesPoint[] =>
+    summedSeries(pickSeries(...names), tOf, 'rate', tickS);
   const prefillTps = counterRate('vllm:prompt_tokens', 'sglang:prompt_tokens');
   const decodeTps = counterRate('vllm:generation_tokens', 'sglang:generation_tokens');
   // Tokens served from prefix cache per scrape. Lets the frontend derive
@@ -424,21 +1034,19 @@ function buildSeriesFromMetrics(
   // SGLang hicache: host-pool KV cache utilization as used/total per
   // timeslice. Both metrics are gauges in absolute tokens. Total stays
   // constant (it's the pool size), used fluctuates.
-  const hostUsedByT = aggregateByStart(
-    metrics['sglang:hicache_host_used_tokens']?.series,
-    'avg',
-    'sum',
-  );
-  const hostTotalByT = aggregateByStart(
-    metrics['sglang:hicache_host_total_tokens']?.series,
-    'avg',
-    'sum',
+  const hostTotalByT = byT(
+    summedSeries(metrics['sglang:hicache_host_total_tokens']?.series, tOf, 'avg', tickS),
   );
   const hostKvCacheUsage: TimeSeriesPoint[] = [];
-  for (const [t, used] of sortedEntries(hostUsedByT)) {
+  for (const { t, value: used } of summedSeries(
+    metrics['sglang:hicache_host_used_tokens']?.series,
+    tOf,
+    'avg',
+    tickS,
+  )) {
     const total = hostTotalByT.get(t);
     if (total !== undefined && total > 0) {
-      hostKvCacheUsage.push({ t: tOf(t), value: used / total });
+      hostKvCacheUsage.push({ t, value: used / total });
     }
   }
 
@@ -449,22 +1057,18 @@ function buildSeriesFromMetrics(
   //   sglang: sglang:realtime_tokens uses a `mode` label with values
   //         {prefill_cache, prefill_compute, decode}. Filter to prefill_*
   //         since decode isn't prompt-token volume.
-  const promptBySrcByT = new Map<string, Map<number, number>>();
-  // Sum a series' per-scrape rates into the bucket for `label`. The bucket is
-  // created even when the series has no valid timeslices — the SGLang fallback
-  // below is gated on `promptBySrcByT.size === 0`, so an empty vllm breakdown
-  // must still suppress it.
+  // Collect the raw series per source label first, then sum each source onto
+  // the canonical grid — same reason as the other summed metrics: a source's
+  // per-worker series sit on their own scrape offsets, so adding them by exact
+  // instant would emit one partial-sum point per worker per tick.
+  const promptBySrc = new Map<string, RawSeries[]>();
+  // The bucket is created even when the series has no valid timeslices — the
+  // SGLang fallback below is gated on `promptBySrc.size === 0`, so an empty
+  // vllm breakdown must still suppress it.
   const addSeriesRates = (label: string, series: RawSeries): void => {
-    let byT = promptBySrcByT.get(label);
-    if (!byT) {
-      byT = new Map<number, number>();
-      promptBySrcByT.set(label, byT);
-    }
-    for (const ts of series.timeslices ?? []) {
-      if (typeof ts.rate === 'number' && typeof ts.start_ns === 'number') {
-        byT.set(ts.start_ns, (byT.get(ts.start_ns) ?? 0) + ts.rate);
-      }
-    }
+    const existing = promptBySrc.get(label);
+    if (existing) existing.push(series);
+    else promptBySrc.set(label, [series]);
   };
   for (const series of metrics['vllm:prompt_tokens_by_source']?.series ?? []) {
     const labels = series.labels ?? {};
@@ -477,7 +1081,7 @@ function buildSeriesFromMetrics(
   //     series carries a `cache_source` label ("device" = HBM, "host" = CPU
   //     offload via hicache). Current runs have only `device`; when hicache
   //     runs land, additional series will appear and the chart will split.
-  if (promptBySrcByT.size === 0) {
+  if (promptBySrc.size === 0) {
     for (const series of metrics['sglang:realtime_tokens']?.series ?? []) {
       const labels = series.labels ?? {};
       const mode = labels['mode'] ?? 'unknown';
@@ -505,11 +1109,10 @@ function buildSeriesFromMetrics(
     }
   }
   const promptTokensBySource: Record<string, TimeSeriesPoint[]> = {};
-  for (const [source, byT] of promptBySrcByT) {
-    const arr: TimeSeriesPoint[] = [];
-    for (const [t, v] of sortedEntries(byT)) {
-      if (v > 0) arr.push({ t: tOf(t), value: v });
-    }
+  for (const [source, seriesForSource] of promptBySrc) {
+    // Idle ticks are dropped rather than emitted as zeros: this feeds a
+    // cumulative stacked area, so a zero adds nothing but payload.
+    const arr = summedSeries(seriesForSource, tOf, 'rate', tickS).filter((p) => p.value > 0);
     if (arr.length > 0) promptTokensBySource[source] = arr;
   }
 
diff --git a/packages/db/src/etl/compute-request-timeline.test.ts b/packages/db/src/etl/compute-request-timeline.test.ts
index 1ad9e63b8..3f4a692d1 100644
--- a/packages/db/src/etl/compute-request-timeline.test.ts
+++ b/packages/db/src/etl/compute-request-timeline.test.ts
@@ -7,6 +7,9 @@ import { REQUEST_TIMELINE_VERSION, computeRequestTimeline } from './compute-requ
 interface SyntheticRequest {
   cid: string;
   ti: number;
+  rootCorrelation?: string;
+  parentCorrelation?: string;
+  xCorrelation?: string;
   srcTrace?: string;
   srcOuter?: number;
   srcInner?: number;
@@ -31,6 +34,11 @@ function makeBlob(requests: SyntheticRequest[]) {
     JSON.stringify({
       metadata: {
         conversation_id: r.cid,
+        ...(r.rootCorrelation === undefined ? {} : { root_correlation_id: r.rootCorrelation }),
+        ...(r.parentCorrelation === undefined
+          ? {}
+          : { parent_correlation_id: r.parentCorrelation }),
+        ...(r.xCorrelation === undefined ? {} : { x_correlation_id: r.xCorrelation }),
         turn_index: r.ti,
         ...(r.srcTrace === undefined ? {} : { source_trace_id: r.srcTrace }),
         ...(r.srcOuter === undefined ? {} : { source_outer_idx: r.srcOuter }),
@@ -58,28 +66,28 @@ function makeBlob(requests: SyntheticRequest[]) {
 }
 
 describe('computeRequestTimeline', () => {
-  it('returns null when the blob is null', () => {
-    expect(computeRequestTimeline(null)).toBeNull();
+  it('returns null when the blob is null', async () => {
+    expect(await computeRequestTimeline(null)).toBeNull();
   });
 
-  it('returns null on a malformed (non-gzip) blob', () => {
-    expect(computeRequestTimeline(Buffer.from('not-gzip'))).toBeNull();
+  it('returns null on a malformed (non-gzip) blob', async () => {
+    expect(await computeRequestTimeline(Buffer.from('not-gzip'))).toBeNull();
   });
 
-  it('returns null when the blob has no parseable records', () => {
-    expect(computeRequestTimeline(gzipSync(Buffer.from('\n\n')))).toBeNull();
+  it('returns null when the blob has no parseable records', async () => {
+    expect(await computeRequestTimeline(gzipSync(Buffer.from('\n\n')))).toBeNull();
   });
 
-  it('returns the current REQUEST_TIMELINE_VERSION in the bundle', () => {
-    const tl = computeRequestTimeline(
+  it('returns the current REQUEST_TIMELINE_VERSION in the bundle', async () => {
+    const tl = await computeRequestTimeline(
       makeBlob([{ cid: 'a', ti: 0, credit: 1000, start: 2000, end: 3000 }]),
     );
     expect(tl?.version).toBe(REQUEST_TIMELINE_VERSION);
   });
 
-  it('shifts ns timestamps to be relative to the earliest credit_issued', () => {
+  it('shifts ns timestamps to be relative to the earliest credit_issued', async () => {
     // Two requests with absolute ns starting at 1_000_000_000.
-    const tl = computeRequestTimeline(
+    const tl = await computeRequestTimeline(
       makeBlob([
         { cid: 'a', ti: 0, credit: 1_000_000_000, start: 1_001_000_000, end: 1_010_000_000 },
         { cid: 'a', ti: 1, credit: 1_020_000_000, start: 1_021_000_000, end: 1_030_000_000 },
@@ -93,8 +101,8 @@ describe('computeRequestTimeline', () => {
     expect(tl?.requests[1]?.start).toBe(21_000_000);
   });
 
-  it('sorts requests by start time, regardless of input order', () => {
-    const tl = computeRequestTimeline(
+  it('sorts requests by start time, regardless of input order', async () => {
+    const tl = await computeRequestTimeline(
       makeBlob([
         { cid: 'a', ti: 0, credit: 30, start: 50, end: 60 },
         { cid: 'a', ti: 1, credit: 0, start: 10, end: 20 },
@@ -104,8 +112,8 @@ describe('computeRequestTimeline', () => {
     expect(tl?.requests.map((r) => r.start)).toEqual([10, 50, 90]);
   });
 
-  it('preserves conversation/worker grouping fields', () => {
-    const tl = computeRequestTimeline(
+  it('preserves conversation/worker grouping fields', async () => {
+    const tl = await computeRequestTimeline(
       makeBlob([
         {
           cid: 'conv-A',
@@ -127,8 +135,142 @@ describe('computeRequestTimeline', () => {
     expect(r.phase).toBe('profiling');
   });
 
-  it('preserves raw source provenance fields when present', () => {
-    const tl = computeRequestTimeline(
+  it('assigns separate replay ids when AIPerf concurrently reuses one conversation', async () => {
+    const tl = await computeRequestTimeline(
+      makeBlob([
+        {
+          cid: 'reused-conversation',
+          ti: 26,
+          rootCorrelation: 'root-a',
+          credit: 10,
+          start: 20,
+          end: 100,
+        },
+        {
+          cid: 'reused-conversation',
+          ti: 40,
+          rootCorrelation: 'root-b',
+          credit: 30,
+          start: 40,
+          end: 120,
+        },
+      ]),
+    );
+
+    expect(tl?.requests.map(({ cid, ri }) => ({ cid, ri }))).toEqual([
+      { cid: 'reused-conversation', ri: 0 },
+      { cid: 'reused-conversation', ri: 1 },
+    ]);
+  });
+
+  it('keeps every turn and subagent request in the same replay across phases', async () => {
+    const tl = await computeRequestTimeline(
+      makeBlob([
+        {
+          cid: 'conversation',
+          ti: 0,
+          rootCorrelation: 'root-a',
+          phase: 'warmup',
+          credit: 0,
+          start: 1,
+          end: 10,
+        },
+        {
+          cid: 'conversation',
+          ti: 1,
+          rootCorrelation: 'root-a',
+          phase: 'profiling',
+          credit: 20,
+          start: 21,
+          end: 30,
+        },
+        {
+          cid: 'conversation::sa:subagent_001_abcd',
+          ti: 1,
+          rootCorrelation: 'root-a',
+          phase: 'profiling',
+          credit: 22,
+          start: 23,
+          end: 29,
+        },
+      ]),
+    );
+
+    expect(new Set(tl?.requests.map((record) => record.ri))).toEqual(new Set([0]));
+  });
+
+  it('uses x_correlation_id as a fallback and ranks replay ids by first dispatch', async () => {
+    const tl = await computeRequestTimeline(
+      makeBlob([
+        {
+          cid: 'conversation',
+          ti: 1,
+          xCorrelation: 'later',
+          credit: 100,
+          start: 110,
+          end: 120,
+        },
+        {
+          cid: 'conversation',
+          ti: 0,
+          xCorrelation: 'earlier',
+          credit: 0,
+          start: 10,
+          end: 20,
+        },
+      ]),
+    );
+
+    expect(tl?.requests.map(({ ti, ri }) => ({ ti, ri }))).toEqual([
+      { ti: 0, ri: 0 },
+      { ti: 1, ri: 1 },
+    ]);
+  });
+
+  it('walks parent correlation links for legacy subagent sessions', async () => {
+    const tl = await computeRequestTimeline(
+      makeBlob([
+        {
+          cid: 'conversation',
+          ti: 1,
+          xCorrelation: 'main-session',
+          credit: 0,
+          start: 10,
+          end: 100,
+        },
+        {
+          cid: 'conversation::sa:subagent-a',
+          ti: 1,
+          xCorrelation: 'child-session',
+          parentCorrelation: 'main-session',
+          credit: 20,
+          start: 30,
+          end: 80,
+        },
+        {
+          cid: 'conversation::sa:subagent-b',
+          ti: 1,
+          xCorrelation: 'grandchild-session',
+          parentCorrelation: 'child-session',
+          credit: 40,
+          start: 50,
+          end: 70,
+        },
+      ]),
+    );
+
+    expect(new Set(tl?.requests.map((record) => record.ri))).toEqual(new Set([0]));
+  });
+
+  it('leaves legacy records without correlation ids backward-compatible', async () => {
+    const tl = await computeRequestTimeline(
+      makeBlob([{ cid: 'legacy', ti: 0, credit: 0, start: 10, end: 20 }]),
+    );
+    expect(tl?.requests[0]?.ri).toBeUndefined();
+  });
+
+  it('preserves raw source provenance fields when present', async () => {
+    const tl = await computeRequestTimeline(
       makeBlob([
         {
           cid: 'trace::fa:003',
@@ -153,8 +295,8 @@ describe('computeRequestTimeline', () => {
     });
   });
 
-  it('preserves the cancelled flag and TTFT/TPOT/ISL/OSL metrics', () => {
-    const tl = computeRequestTimeline(
+  it('preserves the cancelled flag and TTFT/TPOT/ISL/OSL metrics', async () => {
+    const tl = await computeRequestTimeline(
       makeBlob([
         {
           cid: 'a',
@@ -178,8 +320,8 @@ describe('computeRequestTimeline', () => {
     expect(r.osl).toBe(256);
   });
 
-  it('accepts time_per_output_token as a TPOT alias', () => {
-    const tl = computeRequestTimeline(
+  it('accepts time_per_output_token as a TPOT alias', async () => {
+    const tl = await computeRequestTimeline(
       makeBlob([
         {
           cid: 'a',
@@ -195,7 +337,7 @@ describe('computeRequestTimeline', () => {
     expect(tl?.requests[0]?.tpotMs).toBeCloseTo(8.25, 6);
   });
 
-  it('skips records missing both credit_issued_ns and request_start_ns', () => {
+  it('skips records missing both credit_issued_ns and request_start_ns', async () => {
     // Build a record with only request_end_ns — the helper rejects it.
     const broken = gzipSync(
       Buffer.from(
@@ -205,6 +347,6 @@ describe('computeRequestTimeline', () => {
         }),
       ),
     );
-    expect(computeRequestTimeline(broken)).toBeNull();
+    expect(await computeRequestTimeline(broken)).toBeNull();
   });
 });
diff --git a/packages/db/src/etl/compute-request-timeline.ts b/packages/db/src/etl/compute-request-timeline.ts
index 2cbe5174b..ab658d20b 100644
--- a/packages/db/src/etl/compute-request-timeline.ts
+++ b/packages/db/src/etl/compute-request-timeline.ts
@@ -3,22 +3,26 @@
  * Gantt view. Output lands in `agentic_trace_replay.request_timeline`
  * and is read directly by the timeline API route.
  *
- * Shape is a thin array — ~150 bytes per request × ~200 requests per
- * point ≈ 30 KB per row before JSONB compression. Trivial vs the raw
- * gzipped JSONL blob (~1-3 MB).
+ * Shape is a thin array — roughly 150 bytes per request before JSONB
+ * compression. The largest throughput runs contain hundreds of thousands of
+ * requests, so extraction must not materialize the decompressed JSONL.
  *
  * Versioned so the backfill script knows which rows are stale — bump
  * `REQUEST_TIMELINE_VERSION` whenever the extraction algorithm changes.
  */
 
-import { gunzipSync } from 'node:zlib';
+import { Readable } from 'node:stream';
+import { StringDecoder } from 'node:string_decoder';
+import { createGunzip } from 'node:zlib';
 
 /** Bump when the extraction algorithm changes — backfill recomputes anything older. */
-export const REQUEST_TIMELINE_VERSION = 5;
+export const REQUEST_TIMELINE_VERSION = 6;
 
 export interface RequestRecord {
   /** Conversation id (groups turns of one agent session). */
   cid: string;
+  /** Compact replay-lane id derived from AIPerf's stable correlation id. */
+  ri?: number;
   /** Zero-based turn index within the conversation. */
   ti: number;
   /** Source trace id from the original raw dataset, when distinct from replay cid. */
@@ -67,6 +71,9 @@ export interface RequestTimeline {
 
 interface RawMetadata {
   conversation_id?: string;
+  root_correlation_id?: string;
+  parent_correlation_id?: string;
+  x_correlation_id?: string;
   turn_index?: number;
   source_trace_id?: string;
   source_outer_idx?: number;
@@ -108,91 +115,203 @@ function readNum(v: unknown): number | undefined {
 }
 
 /**
- * Parse the gzipped `profile_export.jsonl` blob into a chart-ready
- * timeline. Returns null on a missing or malformed blob.
+ * Correlation ancestry needed to reconstruct the root on older AIPerf exports
+ * that do not repeat `root_correlation_id` on subagent records.
  */
-export function computeRequestTimeline(blob: Buffer | null): RequestTimeline | null {
-  if (!blob) return null;
-  let text: string;
+interface ReplayCorrelationMaps {
+  parentByCorrelation: Map<string, string>;
+  rootByCorrelation: Map<string, string>;
+}
+
+function correlationId(value: unknown): string | undefined {
+  return typeof value === 'string' && value.length > 0 ? value : undefined;
+}
+
+function resolveCorrelationRoot(id: string, maps: ReplayCorrelationMaps): string {
+  const seen = new Set<string>();
+  let current = id;
+  while (!seen.has(current)) {
+    seen.add(current);
+    const explicitRoot = maps.rootByCorrelation.get(current);
+    if (explicitRoot !== undefined) return explicitRoot;
+    const parent = maps.parentByCorrelation.get(current);
+    if (parent === undefined) return current;
+    current = parent;
+  }
+  return id;
+}
+
+/**
+ * AIPerf may sample the same source conversation into several concurrently
+ * executing trajectory lanes. `conversation_id` intentionally stays the
+ * source id, while the root correlation id identifies one replay across every
+ * turn and subagent. For older exports, walk `parent_correlation_id` links from
+ * the per-session `x_correlation_id` instead of treating each subagent session
+ * as an independent replay.
+ */
+function replayKey(meta: RawMetadata, maps: ReplayCorrelationMaps): string | undefined {
+  const explicitRoot = correlationId(meta.root_correlation_id);
+  if (explicitRoot !== undefined) return explicitRoot;
+  const session = correlationId(meta.x_correlation_id);
+  return session === undefined ? undefined : resolveCorrelationRoot(session, maps);
+}
+
+/**
+ * Yield complete UTF-8 lines without retaining the decompressed profile.
+ * `Readable.from([blob])` is intentional: without the array wrapper Node may
+ * iterate the Buffer byte-by-byte instead of passing it as one chunk.
+ */
+async function* gunzipLines(blob: Buffer): AsyncGenerator<string> {
+  const decoder = new StringDecoder('utf8');
+  let carry = '';
+
+  for await (const chunk of Readable.from([blob]).pipe(createGunzip())) {
+    const text = carry + decoder.write(chunk as Buffer);
+    let lineStart = 0;
+    let newline = text.indexOf('\n');
+    while (newline !== -1) {
+      yield text.slice(lineStart, newline);
+      lineStart = newline + 1;
+      newline = text.indexOf('\n', lineStart);
+    }
+    carry = text.slice(lineStart);
+  }
+
+  carry += decoder.end();
+  if (carry.length > 0) yield carry;
+}
+
+function parseRawRecord(line: string): RawRecord | null {
+  if (!line) return null;
   try {
-    text = gunzipSync(blob).toString('utf8');
+    return JSON.parse(line) as RawRecord;
   } catch {
     return null;
   }
+}
 
-  // First pass: parse + collect raw turns; find timeline origin.
-  const raw: {
-    meta: RawMetadata;
-    ttftMs: number | null;
-    tpotMs: number | null;
-    isl: number | null;
-    osl: number | null;
-  }[] = [];
+/**
+ * Parse the gzipped `profile_export.jsonl` blob into a chart-ready
+ * timeline. Returns null on a missing or malformed blob.
+ */
+export async function computeRequestTimeline(blob: Buffer | null): Promise<RequestTimeline | null> {
+  if (!blob) return null;
+
+  // First pass: find the timeline bounds and the first dispatch for each
+  // replay. This deliberately stores only one entry per replay, not one entry
+  // per request, so a 400+ MB decompressed profile stays memory-bounded.
   let originNs = Number.POSITIVE_INFINITY;
   let endNs = 0;
+  let recordCount = 0;
+  const explicitRootFirstStart = new Map<string, number>();
+  const legacySessionFirstStart = new Map<string, number>();
+  const correlationMaps: ReplayCorrelationMaps = {
+    parentByCorrelation: new Map(),
+    rootByCorrelation: new Map(),
+  };
+
+  try {
+    for await (const line of gunzipLines(blob)) {
+      const rec = parseRawRecord(line);
+      if (!rec) continue;
+      const meta = rec.metadata ?? {};
+      // Use credit_issued_ns when available (the true start of the request's
+      // lifecycle), falling back to request_start_ns. Skip rows missing both.
+      const cStart = meta.credit_issued_ns ?? meta.request_start_ns;
+      const cEnd = meta.request_end_ns;
+      if (typeof cStart !== 'number' || typeof cEnd !== 'number') continue;
+
+      recordCount += 1;
+      if (cStart < originNs) originNs = cStart;
+      if (cEnd > endNs) endNs = cEnd;
 
-  for (const line of text.split('\n')) {
-    if (!line) continue;
-    let rec: RawRecord;
-    try {
-      rec = JSON.parse(line) as RawRecord;
-    } catch {
-      continue;
+      const session = correlationId(meta.x_correlation_id);
+      const explicitRoot = correlationId(meta.root_correlation_id);
+      const parent = correlationId(meta.parent_correlation_id);
+      if (session !== undefined && explicitRoot !== undefined) {
+        correlationMaps.rootByCorrelation.set(session, explicitRoot);
+      }
+      if (session !== undefined && parent !== undefined) {
+        correlationMaps.parentByCorrelation.set(session, parent);
+      }
+
+      const firstStarts =
+        explicitRoot === undefined ? legacySessionFirstStart : explicitRootFirstStart;
+      const key = explicitRoot ?? session;
+      if (key === undefined) continue;
+      const current = firstStarts.get(key);
+      if (current === undefined || cStart < current) firstStarts.set(key, cStart);
     }
-    const meta = rec.metadata ?? {};
-    // Use credit_issued_ns when available (the true start of the request's
-    // lifecycle), falling back to request_start_ns. Skip rows missing both.
-    const cStart = meta.credit_issued_ns ?? meta.request_start_ns;
-    const cEnd = meta.request_end_ns;
-    if (typeof cStart !== 'number' || typeof cEnd !== 'number') continue;
-
-    if (cStart < originNs) originNs = cStart;
-    if (cEnd > endNs) endNs = cEnd;
-
-    raw.push({
-      meta,
-      ttftMs: readNum(rec.metrics?.time_to_first_token) ?? null,
-      tpotMs:
-        readNum(rec.metrics?.time_per_output_token) ??
-        readNum(rec.metrics?.inter_token_latency) ??
-        null,
-      isl: readNum(rec.metrics?.input_sequence_length) ?? null,
-      osl: readNum(rec.metrics?.output_sequence_length) ?? null,
-    });
+  } catch {
+    return null;
   }
 
-  if (raw.length === 0) return null;
+  if (recordCount === 0) return null;
   if (!Number.isFinite(originNs)) originNs = 0;
 
-  // Second pass: shift timestamps to be relative to originNs (smaller
-  // numbers fit in JSON nicely and the frontend doesn't need bigint math).
+  // Assign compact deterministic replay ids. UUIDs would add substantial
+  // repeated payload to large timelines, so rank distinct correlation ids by
+  // their first dispatch and send only the numeric rank to the frontend.
+  const replayFirstStart = new Map(explicitRootFirstStart);
+  for (const [session, firstStart] of legacySessionFirstStart) {
+    const key = resolveCorrelationRoot(session, correlationMaps);
+    const current = replayFirstStart.get(key);
+    if (current === undefined || firstStart < current) replayFirstStart.set(key, firstStart);
+  }
+
+  const replayIndex = new Map<string, number>();
+  [...replayFirstStart.entries()]
+    .sort(
+      ([aKey, aStart], [bKey, bStart]) =>
+        aStart - bStart || (aKey < bKey ? -1 : aKey > bKey ? 1 : 0),
+    )
+    .forEach(([key], index) => replayIndex.set(key, index));
+
+  // Second gzip pass: build only the final compact request objects. Re-reading
+  // costs some CPU but avoids simultaneously retaining the decompressed text,
+  // parsed source records, and output records — the previous peak-memory
+  // multiplier that made high-throughput backfills fail.
   const requests: RequestRecord[] = [];
-  for (const r of raw) {
-    const m = r.meta;
-    const credit = (m.credit_issued_ns ?? m.request_start_ns ?? originNs) - originNs;
-    const start = (m.request_start_ns ?? m.credit_issued_ns ?? originNs) - originNs;
-    const ack = typeof m.request_ack_ns === 'number' ? m.request_ack_ns - originNs : null;
-    const end = (m.request_end_ns ?? originNs) - originNs;
-    requests.push({
-      cid: m.conversation_id ?? 'unknown',
-      ti: typeof m.turn_index === 'number' ? m.turn_index : 0,
-      srcTrace: typeof m.source_trace_id === 'string' ? m.source_trace_id : undefined,
-      srcOuter: typeof m.source_outer_idx === 'number' ? m.source_outer_idx : undefined,
-      srcInner: typeof m.source_inner_idx === 'number' ? m.source_inner_idx : undefined,
-      srcKind: typeof m.source_kind === 'string' ? m.source_kind : undefined,
-      wid: m.worker_id ?? 'unknown',
-      ad: typeof m.agent_depth === 'number' ? m.agent_depth : 0,
-      phase: m.benchmark_phase ?? 'unknown',
-      credit,
-      start,
-      ack,
-      end,
-      ttftMs: r.ttftMs,
-      tpotMs: r.tpotMs,
-      isl: r.isl,
-      osl: r.osl,
-      cancelled: m.was_cancelled === true,
-    });
+  try {
+    for await (const line of gunzipLines(blob)) {
+      const rec = parseRawRecord(line);
+      if (!rec) continue;
+      const m = rec.metadata ?? {};
+      const cStart = m.credit_issued_ns ?? m.request_start_ns;
+      const cEnd = m.request_end_ns;
+      if (typeof cStart !== 'number' || typeof cEnd !== 'number') continue;
+
+      const credit = cStart - originNs;
+      const start = (m.request_start_ns ?? m.credit_issued_ns ?? originNs) - originNs;
+      const ack = typeof m.request_ack_ns === 'number' ? m.request_ack_ns - originNs : null;
+      const end = cEnd - originNs;
+      requests.push({
+        cid: m.conversation_id ?? 'unknown',
+        ri: replayIndex.get(replayKey(m, correlationMaps) ?? ''),
+        ti: typeof m.turn_index === 'number' ? m.turn_index : 0,
+        srcTrace: typeof m.source_trace_id === 'string' ? m.source_trace_id : undefined,
+        srcOuter: typeof m.source_outer_idx === 'number' ? m.source_outer_idx : undefined,
+        srcInner: typeof m.source_inner_idx === 'number' ? m.source_inner_idx : undefined,
+        srcKind: typeof m.source_kind === 'string' ? m.source_kind : undefined,
+        wid: m.worker_id ?? 'unknown',
+        ad: typeof m.agent_depth === 'number' ? m.agent_depth : 0,
+        phase: m.benchmark_phase ?? 'unknown',
+        credit,
+        start,
+        ack,
+        end,
+        ttftMs: readNum(rec.metrics?.time_to_first_token) ?? null,
+        tpotMs:
+          readNum(rec.metrics?.time_per_output_token) ??
+          readNum(rec.metrics?.inter_token_latency) ??
+          null,
+        isl: readNum(rec.metrics?.input_sequence_length) ?? null,
+        osl: readNum(rec.metrics?.output_sequence_length) ?? null,
+        cancelled: m.was_cancelled === true,
+      });
+    }
+  } catch {
+    return null;
   }
 
   // Stable order so backfill output is deterministic.
diff --git a/packages/db/src/etl/compute-trace-derived.test.ts b/packages/db/src/etl/compute-trace-derived.test.ts
index ca1227611..824a76c58 100644
--- a/packages/db/src/etl/compute-trace-derived.test.ts
+++ b/packages/db/src/etl/compute-trace-derived.test.ts
@@ -97,7 +97,7 @@ describe('computeTraceDerivedPayloads', () => {
     const [aggregateStats, chartSeries, requestTimeline] = await Promise.all([
       computeAggregateStats({ profileBlob, serverBlob }),
       computeChartSeries(serverBlob, context),
-      Promise.resolve(computeRequestTimeline(profileBlob)),
+      computeRequestTimeline(profileBlob),
     ]);
     const optimized = await computeTraceDerivedPayloads(profileBlob, serverBlob, context);
 
@@ -137,6 +137,6 @@ describe('computeTraceDerivedPayloads', () => {
       await computeAggregateStats({ profileBlob, serverBlob: malformedServer }),
     );
     expect(optimized.chartSeries).toBeNull();
-    expect(optimized.requestTimeline).toEqual(computeRequestTimeline(profileBlob));
+    expect(optimized.requestTimeline).toEqual(await computeRequestTimeline(profileBlob));
   });
 });
diff --git a/packages/db/src/etl/compute-trace-derived.ts b/packages/db/src/etl/compute-trace-derived.ts
index 9cde30fd3..b251260ad 100644
--- a/packages/db/src/etl/compute-trace-derived.ts
+++ b/packages/db/src/etl/compute-trace-derived.ts
@@ -56,7 +56,7 @@ export async function computeTraceDerivedPayloads(
   options: TraceDerivedComputeOptions = {},
 ): Promise<TraceDerivedPayloads> {
   const profileStatsPromise = computeAggregateStats({ profileBlob, serverBlob: null });
-  const requestTimeline = computeRequestTimeline(profileBlob);
+  const requestTimelinePromise = computeRequestTimeline(profileBlob);
 
   const phases = serverBlob
     ? await collectMetricPhases<RawMetric>(
@@ -65,7 +65,11 @@ export async function computeTraceDerivedPayloads(
         options.maxInMemoryBytes,
       ).catch(() => null)
     : null;
-  let aggregateStats = await profileStatsPromise;
+  const [initialAggregateStats, requestTimeline] = await Promise.all([
+    profileStatsPromise,
+    requestTimelinePromise,
+  ]);
+  let aggregateStats = initialAggregateStats;
   let chartSeries: ChartSeries | null = null;
 
   if (phases) {
diff --git a/packages/db/src/etl/db-utils.ts b/packages/db/src/etl/db-utils.ts
index 6ff1678da..c534c2efb 100644
--- a/packages/db/src/etl/db-utils.ts
+++ b/packages/db/src/etl/db-utils.ts
@@ -3,32 +3,35 @@
  */
 
 import postgres, { type Options } from 'postgres';
+import { resolveDatabaseConnection } from '../connection.js';
 
 export type Sql = ReturnType<typeof postgres>;
 
 /**
- * Create a postgres client for admin scripts.
+ * Create a postgres.js client for admin scripts.
  * Reads DATABASE_WRITE_URL by default, or DATABASE_READONLY_URL with `readonly: true`.
- * Pass `envVar` to target another database (e.g. DATABASE_COLLECTIVEX_WRITE_URL).
- * Pass `noSsl: true` to disable TLS for local Postgres.
+ * Pass `envVar` or `url` to target a different database. `noSsl` explicitly
+ * disables TLS, while the shared resolver handles environment and host defaults.
  */
 export function createAdminSql(
   opts: Omit<Options<Record<string, postgres.PostgresType>>, 'ssl'> & {
     readonly?: boolean;
     noSsl?: boolean;
     envVar?: string;
+    url?: string;
   } = {},
 ): Sql {
-  const { readonly, noSsl, envVar: envVarOverride, ...pgOpts } = opts;
+  const { readonly, noSsl, envVar: envVarOverride, url, ...pgOpts } = opts;
   const envVar = envVarOverride ?? (readonly ? 'DATABASE_READONLY_URL' : 'DATABASE_WRITE_URL');
-  const url = process.env[envVar];
-  if (!url) {
-    console.error(`${envVar} is required`);
-    process.exit(1);
-  }
-  return postgres(url, {
+  const connection = resolveDatabaseConnection({
+    envVar,
+    url,
+    driver: 'postgres',
+    ssl: noSsl ? false : process.env.DATABASE_SSL,
+  });
+  return postgres(connection.url, {
     ...pgOpts,
-    ssl: noSsl ? false : 'require',
+    ssl: connection.ssl,
   });
 }
 
diff --git a/packages/db/src/etl/eval-ingest.ts b/packages/db/src/etl/eval-ingest.ts
index 475b5b753..369d83c82 100644
--- a/packages/db/src/etl/eval-ingest.ts
+++ b/packages/db/src/etl/eval-ingest.ts
@@ -3,19 +3,26 @@
  */
 
 import type postgres from 'postgres';
-import type { ConfigParams } from './config-cache';
-import type { EvalParams } from './eval-mapper';
 
 type Sql = ReturnType<typeof postgres>;
 
+export interface EvalPersistenceInput {
+  task: string;
+  isl: number | null;
+  osl: number | null;
+  conc: number | null;
+  lmEvalVersion: string | null;
+  metrics: Record<string, number>;
+}
+
 /**
- * Resolve the config id and insert a single `eval_results` row.
+ * Insert a single `eval_results` row for an already resolved config id.
  * On conflict `(workflow_run_id, config_id, task, isl, osl, conc)` the metrics
  * are overwritten with the latest values.
  *
  * @param sql - Active `postgres` connection.
- * @param getOrCreateConfig - Config cache lookup/upsert function from `createConfigCache`.
- * @param p - Mapped eval parameters to insert.
+ * @param configId - Resolved `configs.id` for this result.
+ * @param p - Eval fields persisted on `eval_results`.
  * @param workflowRunId - DB id of the parent `workflow_runs` row.
  * @param date - ISO date string (`YYYY-MM-DD`) for the `date` column.
  * @returns Outcome (`'new'` or `'dup'`) and the inserted/updated row's `id`,
@@ -23,12 +30,11 @@ type Sql = ReturnType<typeof postgres>;
  */
 export async function ingestEvalRow(
   sql: Sql,
-  getOrCreateConfig: (p: ConfigParams) => Promise<number>,
-  p: EvalParams,
+  configId: number,
+  p: EvalPersistenceInput,
   workflowRunId: number,
   date: string,
 ): Promise<{ outcome: 'new' | 'dup'; id: number }> {
-  const configId = await getOrCreateConfig(p.config);
   const [row] = await sql<{ inserted: boolean; id: number }[]>`
     insert into eval_results (
       workflow_run_id, config_id, task, date,
diff --git a/packages/db/src/etl/ingest-summary.ts b/packages/db/src/etl/ingest-summary.ts
new file mode 100644
index 000000000..16806cc47
--- /dev/null
+++ b/packages/db/src/etl/ingest-summary.ts
@@ -0,0 +1,68 @@
+import type { SkipTracker } from './skip-tracker';
+
+export interface IngestDatabaseTotals {
+  configs: number | string | bigint;
+  benchmarkResults: number | string | bigint;
+  runStats: number | string | bigint;
+  evalResults: number | string | bigint;
+  evalSamples: number | string | bigint;
+  changelogEntries: number | string | bigint;
+}
+
+export interface IngestSummaryOptions {
+  includeFailedRuns?: boolean;
+  includeUnmappedPrecisions?: boolean;
+}
+
+/** Print the DB totals and shared skip/unmapped-value footer used by ingest scripts. */
+export function printIngestSummaryFooter(
+  totals: IngestDatabaseTotals,
+  tracker: SkipTracker,
+  options?: IngestSummaryOptions,
+): void {
+  console.log(`\n  DB totals:`);
+  console.log(`    configs           ${totals.configs}`);
+  console.log(`    benchmark_results ${totals.benchmarkResults}`);
+  console.log(`    run_stats         ${totals.runStats}`);
+  console.log(`    eval_results      ${totals.evalResults}`);
+  console.log(`    eval_samples      ${totals.evalSamples}`);
+  console.log(`    changelog_entries ${totals.changelogEntries}`);
+
+  const { skips, unmappedModels, unmappedHws, unmappedPrecisions } = tracker;
+  const skipLines: [string, number][] = [['no isl/osl (old format)', skips.noIslOsl]];
+  if (options?.includeFailedRuns) {
+    skipLines.push(['failed run (0 successful)', skips.failedRun]);
+  }
+  skipLines.push(
+    ['unmapped model', skips.unmappedModel],
+    ['unmapped hw', skips.unmappedHw],
+    ['bad/empty zip', skips.badZip],
+    ['DB errors', skips.dbError],
+  );
+
+  const nonzeroSkipLines = skipLines.filter(([, count]) => count > 0);
+  const totalSkips = nonzeroSkipLines.reduce((total, [, count]) => total + count, 0);
+  if (totalSkips > 0) {
+    console.log(`\n  Skipped: ${totalSkips} rows`);
+    const pad = Math.max(...nonzeroSkipLines.map(([label]) => label.length));
+    for (const [label, count] of nonzeroSkipLines) {
+      console.log(`    ${label.padEnd(pad)}: ${count}`);
+    }
+  }
+
+  if (unmappedModels.size > 0) {
+    console.log(`\n  Unmapped model values (add to MODEL_TO_KEY to ingest):`);
+    [...unmappedModels].slice(0, 20).forEach((value) => console.log(`    ${value}`));
+    if (unmappedModels.size > 20) console.log(`    ... and ${unmappedModels.size - 20} more`);
+  }
+
+  if (unmappedHws.size > 0) {
+    console.log(`\n  Unmapped hw values (add to hwToGpuKey to ingest):`);
+    [...unmappedHws].slice(0, 20).forEach((value) => console.log(`    ${value}`));
+  }
+
+  if (options?.includeUnmappedPrecisions && unmappedPrecisions.size > 0) {
+    console.log(`\n  Unmapped precision values (add to PRECISION_KEYS to ingest):`);
+    [...unmappedPrecisions].forEach((value) => console.log(`    ${value}`));
+  }
+}
diff --git a/packages/db/src/etl/normalizers.ts b/packages/db/src/etl/normalizers.ts
index 03ab2b420..81ac589ca 100644
--- a/packages/db/src/etl/normalizers.ts
+++ b/packages/db/src/etl/normalizers.ts
@@ -99,6 +99,14 @@ export const MODEL_TO_KEY: Record<string, string> = {
   // Qwen3.5
   'Qwen/Qwen3.5-397B-A17B': 'qwen3.5',
   'Qwen/Qwen3.5-397B-A17B-FP8': 'qwen3.5',
+  // Qwen3.8-Flash-Next (Qwen4 architecture preview — distinct bucket from 3.5).
+  // Hopper has no NVFP4 path, so H200 serves the FP8 checkpoint while Blackwell
+  // serves NVFP4. Both paths are taken from executed sweeps rather than inferred:
+  // FP8 from run 33038487711 (H200, PR #2753), NVFP4 from run 33039186365
+  // (B300, PR #2752). Both report `infmax_model_prefix: qwen3.8next`, so no
+  // PREFIX_ALIASES entry is needed.
+  'Qwen/Qwen3.8-Flash-Next-FP8': 'qwen3.8next',
+  'RadixArk/Qwen3.8-Flash-Next-NVFP4': 'qwen3.8next',
   // Kimi-K2.5 / K2.6 / K2.7-Code (same architecture, distinct DB buckets)
   'moonshotai/Kimi-K2.5': 'kimik2.5',
   'moonshotai/Kimi-K2.6': 'kimik2.6',
diff --git a/packages/db/src/etl/run-overrides.test.ts b/packages/db/src/etl/run-overrides.test.ts
index bc3992d30..65684f5b0 100644
--- a/packages/db/src/etl/run-overrides.test.ts
+++ b/packages/db/src/etl/run-overrides.test.ts
@@ -1,14 +1,257 @@
 import { describe, it, expect } from 'vitest';
 import {
+  type BenchmarkPointBackfill,
+  type ChangelogBackfill,
   type PurgedBenchmarkPoint,
+  applyBenchmarkPointBackfill,
+  applyChangelogBackfills,
+  BENCHMARK_POINT_BACKFILLS,
+  CHANGELOG_BACKFILLS,
   CONCLUSION_OVERRIDES,
   PURGED_BENCHMARK_POINTS,
   PURGED_RUN_ATTEMPTS,
   PURGED_RUNS,
   isBenchmarkPointPurged,
   isRunAttemptPurged,
+  recordBackfilledPointIdentity,
+  validateRunBackfills,
 } from './run-overrides';
 
+const EXAMPLE_CONFIG = {
+  hardware: 'gb300',
+  framework: 'dynamo-vllm',
+  model: 'dsv4',
+  precision: 'fp4',
+  specMethod: 'mtp',
+  disagg: true,
+  isMultinode: true,
+  prefillTp: 8,
+  prefillEp: 8,
+  prefillDpAttn: true,
+  prefillNumWorkers: 1,
+  decodeTp: 8,
+  decodeEp: 8,
+  decodeDpAttn: true,
+  decodeNumWorkers: 1,
+  numPrefillGpu: 8,
+  numDecodeGpu: 8,
+} as const;
+
+function examplePointBackfill(
+  overrides: Partial<BenchmarkPointBackfill> = {},
+): BenchmarkPointBackfill {
+  return {
+    id: 'run-123-point-offload',
+    reason: 'Artifact omitted the offload metadata.',
+    githubRunId: 123,
+    runAttempt: 1,
+    productionConfigId: 456,
+    config: EXAMPLE_CONFIG,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 64,
+    offloadMode: 'off',
+    recipeFingerprint: null,
+    set: {
+      offloadMode: 'on',
+      metricsMerge: { kv_offloading: 'dram', kv_offload_backend: 'lmcache' },
+      metricsRemove: ['stale_offload_field'],
+    },
+    ...overrides,
+  };
+}
+
+describe('audited run backfills', () => {
+  it('validates the checked-in registries', () => {
+    expect(() => validateRunBackfills()).not.toThrow();
+  });
+
+  it('requires a stable ID, reason, exact selector, and non-empty patch', () => {
+    expect(() => validateRunBackfills([], [examplePointBackfill({ id: 'Not Valid' })])).toThrow(
+      /kebab-case/u,
+    );
+    expect(() => validateRunBackfills([], [examplePointBackfill({ reason: ' ' })])).toThrow(
+      /reason/u,
+    );
+    expect(() => validateRunBackfills([], [examplePointBackfill({ set: {} })])).toThrow(
+      /at least one field/u,
+    );
+  });
+
+  it('rejects source-to-destination point identity collisions', () => {
+    const first = examplePointBackfill();
+    const second = examplePointBackfill({
+      id: 'run-123-existing-offload-point',
+      offloadMode: 'on',
+      set: { metricsMerge: { note: 'already on' } },
+    });
+    expect(() => validateRunBackfills([], [first, second])).toThrow(/collides/u);
+  });
+
+  it('rejects duplicate stable selectors even when production config IDs differ', () => {
+    const first = examplePointBackfill();
+    const second = examplePointBackfill({
+      id: 'run-123-duplicate-stable-selector',
+      productionConfigId: 999,
+    });
+
+    expect(() => validateRunBackfills([], [first, second])).toThrow(/duplicate.*selector/u);
+  });
+
+  it('applies point corrections during ingest and synchronizes offload metadata', () => {
+    const backfill = examplePointBackfill();
+    const registry = BENCHMARK_POINT_BACKFILLS as BenchmarkPointBackfill[];
+    registry.push(backfill);
+    const point = {
+      configId: 456,
+      config: EXAMPLE_CONFIG,
+      benchmarkType: 'agentic_traces',
+      isl: null,
+      osl: null,
+      conc: 64,
+      offloadMode: 'off',
+      recipeFingerprint: null,
+      metrics: { median_itl: 0.1, stale_offload_field: true },
+    };
+
+    try {
+      const applied = applyBenchmarkPointBackfill(123, 1, point);
+      expect(applied.backfillId).toBe(backfill.id);
+      expect(applied.point.offloadMode).toBe('on');
+      expect(applied.point.metrics).toEqual({
+        median_itl: 0.1,
+        offload_mode: 'on',
+        kv_offloading: 'dram',
+        kv_offload_backend: 'lmcache',
+      });
+      expect(applied.desiredIdentity).not.toBe(applied.sourceIdentity);
+
+      const otherAttempt = applyBenchmarkPointBackfill(123, 2, point);
+      expect(otherAttempt.backfillId).toBeNull();
+      // GCS fallback has no attempt metadata and intentionally matches by run + point.
+      expect(applyBenchmarkPointBackfill(123, undefined, point).backfillId).toBe(backfill.id);
+    } finally {
+      registry.splice(registry.indexOf(backfill), 1);
+    }
+  });
+
+  it('matches point backfills by stable config dimensions across database branches', () => {
+    const backfill = examplePointBackfill();
+    const registry = BENCHMARK_POINT_BACKFILLS as BenchmarkPointBackfill[];
+    registry.push(backfill);
+
+    try {
+      const applied = applyBenchmarkPointBackfill(123, 1, {
+        configId: 999,
+        config: { ...EXAMPLE_CONFIG },
+        benchmarkType: 'agentic_traces',
+        isl: null,
+        osl: null,
+        conc: 64,
+        offloadMode: 'off',
+        recipeFingerprint: null,
+        metrics: { median_itl: 0.1 },
+      });
+
+      expect(applied.backfillId).toBe(backfill.id);
+      expect(applied.point.configId).toBe(999);
+      expect(applied.point.offloadMode).toBe('on');
+    } finally {
+      registry.splice(registry.indexOf(backfill), 1);
+    }
+  });
+
+  it('does not match a different stable config that reuses the production config ID', () => {
+    const backfill = examplePointBackfill();
+    const registry = BENCHMARK_POINT_BACKFILLS as BenchmarkPointBackfill[];
+    registry.push(backfill);
+
+    try {
+      const applied = applyBenchmarkPointBackfill(123, 1, {
+        configId: 456,
+        config: { ...EXAMPLE_CONFIG, decodeTp: 16, numDecodeGpu: 16 },
+        benchmarkType: 'agentic_traces',
+        isl: null,
+        osl: null,
+        conc: 64,
+        offloadMode: 'off',
+        recipeFingerprint: null,
+        metrics: { median_itl: 0.1 },
+      });
+
+      expect(applied.backfillId).toBeNull();
+    } finally {
+      registry.splice(registry.indexOf(backfill), 1);
+    }
+  });
+
+  it('applies changelog corrections to the row that ingest persists', () => {
+    const backfill: ChangelogBackfill = {
+      id: 'run-123-changelog-configs',
+      reason: 'The artifact listed the wrong config key.',
+      githubRunId: 123,
+      runAttempt: 1,
+      baseRef: 'master',
+      headRef: 'feature-sha',
+      set: {
+        configKeys: ['dsv4-fp4-b300-vllm-mtp'],
+        description: 'Corrected description',
+        prLink: null,
+        appendOnly: true,
+      },
+    };
+    const registry = CHANGELOG_BACKFILLS as ChangelogBackfill[];
+    registry.push(backfill);
+
+    try {
+      const applied = applyChangelogBackfills(123, 1, [
+        {
+          baseRef: 'master',
+          headRef: 'feature-sha',
+          entries: [
+            {
+              configKeys: ['old-first'],
+              description: 'First entry is overwritten by ingest',
+              prLink: 'https://example.com/first',
+              evalsOnly: false,
+              appendOnly: false,
+            },
+            {
+              configKeys: ['old-final'],
+              description: 'Final stored entry',
+              prLink: 'https://example.com/final',
+              evalsOnly: false,
+              appendOnly: false,
+            },
+          ],
+        },
+      ]);
+
+      expect(applied.backfillIds).toEqual([backfill.id]);
+      expect(applied.changelogs[0].entries[0].configKeys).toEqual(['old-first']);
+      expect(applied.changelogs[0].entries[0].appendOnly).toBe(true);
+      expect(applied.changelogs[0].entries[1]).toMatchObject({
+        configKeys: ['dsv4-fp4-b300-vllm-mtp'],
+        description: 'Corrected description',
+        prLink: null,
+        appendOnly: true,
+        evalsOnly: false,
+      });
+    } finally {
+      registry.splice(registry.indexOf(backfill), 1);
+    }
+  });
+
+  it('detects two artifact rows collapsing onto one corrected identity', () => {
+    const seen = new Map<string, string>();
+    recordBackfilledPointIdentity(seen, 'source-off', 'desired-on');
+    expect(() => recordBackfilledPointIdentity(seen, 'source-on', 'desired-on')).toThrow(
+      /collision/u,
+    );
+  });
+});
+
 describe('CONCLUSION_OVERRIDES', () => {
   it('all run IDs are positive integers', () => {
     for (const runId of CONCLUSION_OVERRIDES.keys()) {
@@ -95,6 +338,11 @@ describe('PURGED_BENCHMARK_POINTS', () => {
       expect(point.osl === null || point.osl > 0).toBe(true);
       expect(point.conc).toBeGreaterThan(0);
       expect(point.offloadMode).not.toBe('');
+      expect(
+        point.recipeFingerprint === undefined ||
+          point.recipeFingerprint === null ||
+          point.recipeFingerprint !== '',
+      ).toBe(true);
       const identity = [
         point.githubRunId,
         point.runAttempt,
@@ -104,6 +352,7 @@ describe('PURGED_BENCHMARK_POINTS', () => {
         point.osl,
         point.conc,
         point.offloadMode,
+        point.recipeFingerprint ?? null,
       ].join('|');
       expect(unique.has(identity), `duplicate point override: ${identity}`).toBe(false);
       unique.add(identity);
@@ -152,6 +401,7 @@ describe('isBenchmarkPointPurged', () => {
       osl: 1024,
       conc: 1,
       offloadMode: 'none',
+      recipeFingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
     };
     const registry = PURGED_BENCHMARK_POINTS as PurgedBenchmarkPoint[];
     registry.push(point);
@@ -171,6 +421,50 @@ describe('isBenchmarkPointPurged', () => {
           offloadMode: 'cpu',
         }),
       ).toBe(false);
+      expect(
+        isBenchmarkPointPurged(point.githubRunId, point.runAttempt, {
+          ...point,
+          recipeFingerprint: 'bbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbbb',
+        }),
+      ).toBe(false);
+      expect(
+        isBenchmarkPointPurged(point.githubRunId, point.runAttempt, {
+          ...point,
+          recipeFingerprint: null,
+        }),
+      ).toBe(false);
+    } finally {
+      registry.splice(registry.indexOf(point), 1);
+    }
+  });
+
+  it('treats omitted and null fingerprints as the same legacy identity', () => {
+    const point: PurgedBenchmarkPoint = {
+      githubRunId: 1,
+      runAttempt: 1,
+      configId: 1,
+      benchmarkType: 'single_turn',
+      isl: 1024,
+      osl: 1024,
+      conc: 1,
+      offloadMode: 'none',
+    };
+    const registry = PURGED_BENCHMARK_POINTS as PurgedBenchmarkPoint[];
+    registry.push(point);
+
+    try {
+      expect(
+        isBenchmarkPointPurged(point.githubRunId, point.runAttempt, {
+          ...point,
+          recipeFingerprint: null,
+        }),
+      ).toBe(true);
+      expect(
+        isBenchmarkPointPurged(point.githubRunId, point.runAttempt, {
+          ...point,
+          recipeFingerprint: 'aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa',
+        }),
+      ).toBe(false);
     } finally {
       registry.splice(registry.indexOf(point), 1);
     }
diff --git a/packages/db/src/etl/run-overrides.ts b/packages/db/src/etl/run-overrides.ts
index 25afbddcc..18a3379dd 100644
--- a/packages/db/src/etl/run-overrides.ts
+++ b/packages/db/src/etl/run-overrides.ts
@@ -19,11 +19,19 @@
  * run attempt and skip them on every future ingest. Each entry uses the row's durable
  * database natural key, which can be queried from the linked dashboard point.
  *
+ * CHANGELOG_BACKFILLS — correct stored changelog metadata for one exact run attempt
+ * and base/head ref pair.
+ *
+ * BENCHMARK_POINT_BACKFILLS — correct metrics and/or offload identity for one exact
+ * benchmark point. These are applied both during ingest and against existing DB rows.
+ *
  * Note: GitHub deletes old workflow runs over time so these overrides may not be applicable forever,
  *       but we should keep them around for historical reference. You can find these on github (if available) by filling
  *       in the run id into the following link: https://github.com/SemiAnalysisAI/InferenceX/actions/runs/{run_id_here}
  */
 
+import { type ConfigParams, configCacheKey } from './config-cache';
+
 export const CONCLUSION_OVERRIDES: ReadonlyMap<number, string> = new Map([
   [22806827144, 'success'], // 2026-03-07 | dsr1 fp8 h200 SGLang 0.5.7→0.5.9 bump | Reason: database upload step failed
   [22792161490, 'success'], // 2026-03-07 | GLM-5 fp8 mi355x SGLang benchmark add | Reason: database upload step failed
@@ -81,6 +89,8 @@ export const PURGED_RUNS: ReadonlySet<number> = new Set([
   29912027293, // 2026-07-22 | Reason: accidental ingest while testing
   30346826643, // 2026-07-28 | Initial AMD submission for MiniMax M3 used incorrect AgentX harness; MTP/spec decode is AgentX-only. Will update after harness updates.
   30405836523, // 2026-07-28 | Reason: No non-DSpark — kimik3-fp4-b300-vllm-agentic AgentX points run without speculative decoding, and Kimi-K3 agentic coding is published DSpark-only (source run of the PR #2397 sweep-reuse ingest)
+  32695861783, // 2026-08-24 | Reason: dev image — the dsv4-fp4-b300-sglang-agentic-hicache-mtp AgentX sweep ran on lmsysorg/sglang-staging:dev-cu13-pr-35880, built from the unmerged SGLang draft PR sgl-project/sglang#35880 (source run of the PR #2701 sweep-reuse ingest)
+  32863999669, // 2026-08-25 | Reason: incomplete failing run — Run Sweep on main for #2687 ([Power] Require GB multinode telemetry) failed 30 of 125 jobs (GB200/GB300 multi-node dyn-sgl arms), leaving partial data
 ]);
 
 export const PURGED_RUN_ATTEMPTS: ReadonlyMap<number, ReadonlySet<number>> = new Map([
@@ -116,6 +126,8 @@ export interface BenchmarkPointKey {
   osl: number | null;
   conc: number;
   offloadMode: string;
+  /** Producer recipe identity. Omit or set null only for legacy rows. */
+  recipeFingerprint?: string | null;
 }
 
 export interface PurgedBenchmarkPoint extends BenchmarkPointKey {
@@ -126,9 +138,1051 @@ export interface PurgedBenchmarkPoint extends BenchmarkPointKey {
 /**
  * Individual benchmark rows to skip on ingest and delete from the DB.
  * Keep a dated reason comment beside every entry for auditability:
- * `{ githubRunId, runAttempt, configId, benchmarkType, isl, osl, conc, offloadMode }`.
+ * `{ githubRunId, runAttempt, configId, benchmarkType, isl, osl, conc, offloadMode,
+ * recipeFingerprint }`. Omitted fingerprints target only legacy NULL rows.
+ */
+export const PURGED_BENCHMARK_POINTS: readonly PurgedBenchmarkPoint[] = [
+  // 2026-08-03 | Point 438205 | Qwen3.5 FP4 B300 SGLang MTP, TP2/EP2, conc 32.
+  // Reason: the one-hour AgentX profile stopped after 3,102s, then emitted no requests for 538s.
+  {
+    githubRunId: 30730541420,
+    runAttempt: 1,
+    configId: 744,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 32,
+    offloadMode: 'on',
+    recipeFingerprint: null,
+  },
+  // 2026-08-03 | Point 438197 | Qwen3.5 FP4 B300 SGLang MTP, TP2/EP2, conc 34.
+  // Reason: the one-hour AgentX profile stopped after 373s, then emitted no requests for 3,267s.
+  {
+    githubRunId: 30730541420,
+    runAttempt: 1,
+    configId: 744,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 34,
+    offloadMode: 'on',
+    recipeFingerprint: null,
+  },
+  // 2026-08-03 | Point 438193 | Qwen3.5 FP4 B300 SGLang MTP, TP2/EP2, conc 38.
+  // Reason: the one-hour AgentX profile stopped after 2,994s, then emitted no requests for 646s.
+  {
+    githubRunId: 30730541420,
+    runAttempt: 1,
+    configId: 744,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 38,
+    offloadMode: 'on',
+    recipeFingerprint: null,
+  },
+  // 2026-08-03 | Point 438195 | Qwen3.5 FP4 B300 SGLang MTP, TP2/EP2, conc 40.
+  // Reason: the one-hour AgentX profile stopped after 692s, then emitted no requests for 2,948s.
+  {
+    githubRunId: 30730541420,
+    runAttempt: 1,
+    configId: 744,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 40,
+    offloadMode: 'on',
+    recipeFingerprint: null,
+  },
+  // 2026-08-03 | Point 438207 | Qwen3.5 FP4 B300 SGLang MTP, TP2/EP2, conc 48.
+  // Reason: the one-hour AgentX profile stopped after 1,033s, then emitted no requests for 2,607s.
+  {
+    githubRunId: 30730541420,
+    runAttempt: 1,
+    configId: 744,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 48,
+    offloadMode: 'on',
+    recipeFingerprint: null,
+  },
+  // 2026-08-03 | Point 438211 | Qwen3.5 FP4 B300 SGLang MTP, TP2/EP2, conc 56.
+  // Reason: the one-hour AgentX profile stopped after 2,073s, then emitted no requests for 1,568s.
+  {
+    githubRunId: 30730541420,
+    runAttempt: 1,
+    configId: 744,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 56,
+    offloadMode: 'on',
+    recipeFingerprint: null,
+  },
+  // 2026-08-03 | Point 438242 | Qwen3.5 FP4 B200 SGLang MTP, TP4/EP1, conc 76.
+  // Reason: the one-hour AgentX profile stopped after 20s, then emitted no requests for 3,620s.
+  {
+    githubRunId: 30758866378,
+    runAttempt: 1,
+    configId: 754,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc: 76,
+    offloadMode: 'on',
+    recipeFingerprint: null,
+  },
+];
+
+interface AuditedBackfill {
+  /** Stable, descriptive identifier used in logs and review history. */
+  id: string;
+  /** Why artifact data is being corrected instead of re-running the benchmark. */
+  reason: string;
+}
+
+export interface ChangelogBackfill extends AuditedBackfill {
+  githubRunId: number;
+  runAttempt: number;
+  baseRef: string;
+  headRef: string;
+  set: {
+    configKeys?: readonly string[];
+    description?: string;
+    prLink?: string | null;
+    appendOnly?: boolean;
+  };
+}
+
+/**
+ * Audited corrections to changelog rows already produced by workflow artifacts.
+ * Selectors use the table's complete durable identity. `set` is a partial patch.
+ */
+export const CHANGELOG_BACKFILLS: readonly ChangelogBackfill[] = [
+  {
+    id: 'run-32242794988-restore-append-only',
+    reason:
+      'PR #2676 produced an append-only delta sweep, but conflict resolution removed the marker before its artifacts were reused for merge ingestion.',
+    githubRunId: 32242794988,
+    runAttempt: 3,
+    baseRef: '43996b8dc1b3fa53989abb0c86491a7d3e3b4472',
+    headRef: 'de871da96f061b097fef318489462a958240917a',
+    set: {
+      appendOnly: true,
+    },
+  },
+];
+
+export type JsonValue =
+  | boolean
+  | number
+  | string
+  | null
+  | readonly JsonValue[]
+  | { readonly [key: string]: JsonValue };
+
+export interface BenchmarkPointBackfill extends AuditedBackfill {
+  githubRunId: number;
+  runAttempt: number;
+  /** Production ID retained for audit logs only. Never use it to match another DB branch. */
+  productionConfigId: number;
+  /** Stable configuration dimensions shared by production, staging, and rebuilt databases. */
+  config: ConfigParams;
+  benchmarkType: string;
+  isl: number | null;
+  osl: number | null;
+  conc: number;
+  offloadMode: string;
+  /** Producer recipe identity. Omit or set null only for legacy rows. */
+  recipeFingerprint?: string | null;
+  set: {
+    /** Updates both the first-class column and metrics.offload_mode. */
+    offloadMode?: 'on' | 'off';
+    /** Shallow JSONB merge; existing unrelated metrics are preserved. */
+    metricsMerge?: Readonly<Record<string, JsonValue>>;
+    /** Top-level metric keys to remove before metricsMerge is applied. */
+    metricsRemove?: readonly string[];
+  };
+}
+
+/**
+ * Audited corrections to individual benchmark points. The selector is the row's
+ * complete pre-backfill natural identity; `set` contains only the desired changes.
+ *
+ * Example:
+ * {
+ *   id: 'run-123-attempt-1-conc-64-enable-offload',
+ *   reason: 'The artifact omitted offload metadata for this point.',
+ *   githubRunId: 123,
+ *   runAttempt: 1,
+ *   productionConfigId: 456,
+ *   config: {
+ *     hardware: 'gb300', framework: 'dynamo-vllm', model: 'dsv4',
+ *     precision: 'fp4', specMethod: 'mtp', disagg: true, isMultinode: true,
+ *     prefillTp: 8, prefillEp: 8, prefillDpAttn: true, prefillNumWorkers: 1,
+ *     decodeTp: 8, decodeEp: 8, decodeDpAttn: true, decodeNumWorkers: 1,
+ *     numPrefillGpu: 8, numDecodeGpu: 8,
+ *   },
+ *   benchmarkType: 'agentic_traces',
+ *   isl: null,
+ *   osl: null,
+ *   conc: 64,
+ *   offloadMode: 'off',
+ *   recipeFingerprint: null,
+ *   set: {
+ *     offloadMode: 'on',
+ *     metricsMerge: { kv_offloading: 'dram', kv_offload_backend: 'lmcache' },
+ *   },
+ * }
  */
-export const PURGED_BENCHMARK_POINTS: readonly PurgedBenchmarkPoint[] = [];
+function disaggregatedConfig(
+  base: Pick<ConfigParams, 'hardware' | 'framework' | 'model' | 'precision' | 'specMethod'>,
+  prefill: { tp: number; ep: number; dpAttn: boolean; numWorkers: number },
+  decode: { tp: number; ep: number; dpAttn: boolean; numWorkers: number },
+): ConfigParams {
+  return {
+    ...base,
+    disagg: true,
+    isMultinode: true,
+    prefillTp: prefill.tp,
+    prefillEp: prefill.ep,
+    prefillDpAttn: prefill.dpAttn,
+    prefillNumWorkers: prefill.numWorkers,
+    decodeTp: decode.tp,
+    decodeEp: decode.ep,
+    decodeDpAttn: decode.dpAttn,
+    decodeNumWorkers: decode.numWorkers,
+    numPrefillGpu: prefill.tp * prefill.numWorkers,
+    numDecodeGpu: decode.tp * decode.numWorkers,
+  };
+}
+
+const QWEN35_GB300_DYNAMO_TRT = {
+  hardware: 'gb300',
+  framework: 'dynamo-trt',
+  model: 'qwen3.5',
+  precision: 'fp4',
+  specMethod: 'mtp',
+} as const;
+
+const DSV4_GB300_DYNAMO_TRT = {
+  hardware: 'gb300',
+  framework: 'dynamo-trt',
+  model: 'dsv4',
+  precision: 'fp4',
+  specMethod: 'mtp',
+} as const;
+
+const GLM52_GB300_DYNAMO_TRT = {
+  hardware: 'gb300',
+  framework: 'dynamo-trt',
+  model: 'glm5.2',
+  precision: 'fp4',
+  specMethod: 'mtp',
+} as const;
+
+const DSV4_GB200_DYNAMO_VLLM = {
+  hardware: 'gb200',
+  framework: 'dynamo-vllm',
+  model: 'dsv4',
+  precision: 'fp4',
+  specMethod: 'mtp',
+} as const;
+
+export const BENCHMARK_POINT_BACKFILLS: readonly BenchmarkPointBackfill[] = [
+  // The source recipes in run 31633154542 attach MooncakeStoreConnector to
+  // every disaggregated worker and allocate a 180 GB Mooncake segment per
+  // node. The master matrix omitted the corresponding offload annotation.
+  ...(
+    [
+      [
+        2106,
+        256,
+        null,
+        disaggregatedConfig(
+          {
+            hardware: 'gb300',
+            framework: 'dynamo-vllm',
+            model: 'dsv4',
+            precision: 'fp4',
+            specMethod: 'mtp',
+          },
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 1 },
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2334,
+        1152,
+        null,
+        disaggregatedConfig(
+          {
+            hardware: 'gb300',
+            framework: 'dynamo-vllm',
+            model: 'dsv4',
+            precision: 'fp4',
+            specMethod: 'mtp',
+          },
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 2 },
+          { tp: 12, ep: 12, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2336,
+        1024,
+        null,
+        disaggregatedConfig(
+          {
+            hardware: 'gb300',
+            framework: 'dynamo-vllm',
+            model: 'dsv4',
+            precision: 'fp4',
+            specMethod: 'mtp',
+          },
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 2 },
+          { tp: 16, ep: 16, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+    ] as const
+  ).map(([productionConfigId, conc, recipeFingerprint, config]) => ({
+    id: `run-31633154542-config-${productionConfigId}-conc-${conc}-mooncake-offload`,
+    reason:
+      'The runtime recipe enabled MooncakeStoreConnector, but the artifact reported this AgentX point as non-offloaded.',
+    githubRunId: 31633154542,
+    runAttempt: 2,
+    productionConfigId,
+    config,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc,
+    offloadMode: 'off',
+    recipeFingerprint,
+    set: {
+      offloadMode: 'on' as const,
+      metricsRemove: ['allocated_cpu_dram_gb'],
+      metricsMerge: {
+        kv_offloading: 'dram',
+        kv_offload_backend: 'mooncake',
+        kv_offload_backend_version: '0.3.11.post1',
+      },
+    },
+  })),
+
+  // Every TensorRT-LLM recipe in run 31927376673 configures a 128 GiB native
+  // host KV cache on both prefill and decode workers. The master matrix only
+  // declared NIXL transfer, so all six artifacts lost the offload identity.
+  ...(
+    [
+      [
+        2360,
+        7,
+        '5c282408e21b662cf5afbef0d26a63ad2fb19bb66ca3b431763a1c40628f3036',
+        disaggregatedConfig(
+          QWEN35_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 1 },
+          { tp: 8, ep: 8, dpAttn: false, numWorkers: 7 },
+        ),
+      ],
+      [
+        2361,
+        96,
+        'da36b834945785abfa06c44f742684db7a96af49542d9012019063d552dc7393',
+        disaggregatedConfig(
+          QWEN35_GB300_DYNAMO_TRT,
+          { tp: 2, ep: 2, dpAttn: false, numWorkers: 2 },
+          { tp: 8, ep: 8, dpAttn: false, numWorkers: 3 },
+        ),
+      ],
+      [
+        2362,
+        704,
+        '60e4361ec00fd8a94a700b7ae9dbb4d8b6cf4095b1553f098af45551669fcf1f',
+        disaggregatedConfig(
+          QWEN35_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 3 },
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 2 },
+        ),
+      ],
+      [
+        2363,
+        52,
+        'd7fccb3572037431f39757640d090ab1204bb27ac57ff0f9f9a635e9cefb5867',
+        disaggregatedConfig(
+          QWEN35_GB300_DYNAMO_TRT,
+          { tp: 1, ep: 1, dpAttn: true, numWorkers: 2 },
+          { tp: 2, ep: 2, dpAttn: false, numWorkers: 2 },
+        ),
+      ],
+      [
+        2364,
+        565,
+        'a6017a5c8a675fb4f1f74e49bb0091f172e6ea357a1bb1b2d6577eb661c9f1c5',
+        disaggregatedConfig(
+          QWEN35_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 3 },
+          { tp: 16, ep: 16, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2365,
+        44,
+        'ac8406a4d46f3711732aa352c801bbd1eb7c3545b2982c5d2c56520fa8288342',
+        disaggregatedConfig(
+          QWEN35_GB300_DYNAMO_TRT,
+          { tp: 1, ep: 1, dpAttn: true, numWorkers: 1 },
+          { tp: 2, ep: 2, dpAttn: false, numWorkers: 1 },
+        ),
+      ],
+    ] as const
+  ).map(([productionConfigId, conc, recipeFingerprint, config]) => ({
+    id: `run-31927376673-config-${productionConfigId}-conc-${conc}-native-offload`,
+    reason:
+      'The TensorRT-LLM runtime recipe configured a native host KV cache, but the artifact reported this AgentX point as non-offloaded.',
+    githubRunId: 31927376673,
+    runAttempt: 1,
+    productionConfigId,
+    config,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc,
+    offloadMode: 'off',
+    recipeFingerprint,
+    set: {
+      offloadMode: 'on' as const,
+      metricsRemove: ['allocated_cpu_dram_gb'],
+      metricsMerge: {
+        kv_offloading: 'dram',
+        kv_offload_backend: 'native',
+        kv_offload_backend_version: '1.3.0rc24',
+      },
+    },
+  })),
+
+  // The six disaggregated DeepSeek-V4-Pro recipes in run 32403083041
+  // configure a 180 GiB native host KV cache on each prefill worker. The
+  // master matrix omitted the offload annotation, so the artifacts reported
+  // these points as non-offloaded.
+  ...(
+    [
+      [
+        1523,
+        4,
+        '38be5ef96c0cd06d88da9fc176812f805d5a142f618c6c4a98d0e5dd9db6e9b7',
+        disaggregatedConfig(
+          DSV4_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 1 },
+          { tp: 8, ep: 8, dpAttn: false, numWorkers: 4 },
+        ),
+      ],
+      [
+        2415,
+        24,
+        'fb3f411ba6d72aa06658b108212333afde9ea2cb499e377fdfc65de5412592cb',
+        disaggregatedConfig(
+          DSV4_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 1 },
+          { tp: 4, ep: 4, dpAttn: false, numWorkers: 6 },
+        ),
+      ],
+      [
+        2417,
+        388,
+        '5d290a80ed5a6d5acf4b19a425017929cc9fe416161c59a76369d14e14694bac',
+        disaggregatedConfig(
+          DSV4_GB300_DYNAMO_TRT,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 1 },
+          { tp: 32, ep: 32, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2419,
+        736,
+        '3c4563cbcbd0145b101f16672e8c8072b55c0dffc0014876cc52b58556d1997b',
+        disaggregatedConfig(
+          DSV4_GB300_DYNAMO_TRT,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 2 },
+          { tp: 32, ep: 32, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2416,
+        1152,
+        '885edc2b1a2521989da184ecc0520dbbdc962b8813fd5a8a26115c343a54745d',
+        disaggregatedConfig(
+          DSV4_GB300_DYNAMO_TRT,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 3 },
+          { tp: 16, ep: 16, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2418,
+        2626,
+        '6e47ac1cc8896a21dcff448d39454300783009dcd1216009b3466a6fc1d1b730',
+        disaggregatedConfig(
+          DSV4_GB300_DYNAMO_TRT,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 5 },
+          { tp: 16, ep: 16, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+    ] as const
+  ).map(([productionConfigId, conc, recipeFingerprint, config]) => ({
+    id: `run-32403083041-config-${productionConfigId}-conc-${conc}-native-offload`,
+    reason:
+      'The TensorRT-LLM runtime recipe configured a native host KV cache, but the artifact reported this AgentX point as non-offloaded.',
+    githubRunId: 32403083041,
+    runAttempt: 3,
+    productionConfigId,
+    config,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc,
+    offloadMode: 'off',
+    recipeFingerprint,
+    set: {
+      offloadMode: 'on' as const,
+      metricsRemove: ['allocated_cpu_dram_gb'],
+      metricsMerge: {
+        kv_offloading: 'dram',
+        kv_offload_backend: 'native',
+        kv_offload_backend_version: '1.3.0rc24',
+      },
+    },
+  })),
+
+  // The six disaggregated GLM-5.2 recipes in run 32346724519 configure a
+  // 128 GiB native host KV cache on each prefill worker. The aggregate point
+  // has no host cache and is intentionally excluded. The master matrix omitted
+  // the offload annotation for the disaggregated points.
+  ...(
+    [
+      [
+        2426,
+        20,
+        '16c12cc3717783b1029c8e65865043f7d518de4af8d8477572eb594099844263',
+        disaggregatedConfig(
+          GLM52_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 1 },
+          { tp: 8, ep: 8, dpAttn: false, numWorkers: 1 },
+        ),
+      ],
+      [
+        2427,
+        30,
+        'daf6d60e1a5c68f3c00719a50d60d1b77ba603e851eb00f6149ef5989ba42e12',
+        disaggregatedConfig(
+          GLM52_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 1 },
+          { tp: 4, ep: 4, dpAttn: false, numWorkers: 4 },
+        ),
+      ],
+      [
+        2428,
+        60,
+        '454122157cfc71a5252e5ec33d7709cc4f4061c52a7ca5b28f4aea1fa3cc9011',
+        disaggregatedConfig(
+          GLM52_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 3 },
+          { tp: 4, ep: 4, dpAttn: false, numWorkers: 4 },
+        ),
+      ],
+      [
+        2429,
+        152,
+        '02abb316847471b7e7bc68e41b3013dafdaddfa6c464274bdfb11d5f3b386e61',
+        disaggregatedConfig(
+          GLM52_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 5 },
+          { tp: 16, ep: 16, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2431,
+        227,
+        'ac551c93599bd44d074294d3de65ca5aaa6b240c12187bb5aacbef788446c65f',
+        disaggregatedConfig(
+          GLM52_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 6 },
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2430,
+        259,
+        '86114ba7b339cc49098f489d2621c81c119b7fd156f4f100cde228ff3c8326d2',
+        disaggregatedConfig(
+          GLM52_GB300_DYNAMO_TRT,
+          { tp: 4, ep: 4, dpAttn: true, numWorkers: 8 },
+          { tp: 16, ep: 16, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+    ] as const
+  ).map(([productionConfigId, conc, recipeFingerprint, config]) => ({
+    id: `run-32346724519-config-${productionConfigId}-conc-${conc}-native-offload`,
+    reason:
+      'The TensorRT-LLM runtime recipe configured a native host KV cache, but the artifact reported this AgentX point as non-offloaded.',
+    githubRunId: 32346724519,
+    runAttempt: 4,
+    productionConfigId,
+    config,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc,
+    offloadMode: 'off',
+    recipeFingerprint,
+    set: {
+      offloadMode: 'on' as const,
+      metricsRemove: ['allocated_cpu_dram_gb'],
+      metricsMerge: {
+        kv_offloading: 'dram',
+        kv_offload_backend: 'native',
+        kv_offload_backend_version: '1.3.0rc22.post1',
+      },
+    },
+  })),
+
+  // The four disaggregated recipes in run 31965016666 attach
+  // MooncakeStoreConnector to NIXL through MultiConnector and allocate a
+  // 140 GB Mooncake segment per node. The three aggregate points in the same
+  // run do not attach the connector and are intentionally excluded.
+  ...(
+    [
+      [
+        1012,
+        128,
+        'b01cb33e392b60b01e2f498ea7300118c6046e074556199a3fe9a7208cf7929e',
+        disaggregatedConfig(
+          DSV4_GB200_DYNAMO_VLLM,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 1 },
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        1012,
+        256,
+        '856babb1e00e524c7eddca689e1345d1a97a41f6743c5d5c109347444581aeef',
+        disaggregatedConfig(
+          DSV4_GB200_DYNAMO_VLLM,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 1 },
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2374,
+        576,
+        'd7afa7f01be968e02911263a9792bb1200ca27de702b5e79d7907e6d46adfc44',
+        disaggregatedConfig(
+          DSV4_GB200_DYNAMO_VLLM,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 2 },
+          { tp: 12, ep: 12, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+      [
+        2375,
+        512,
+        '47a4969f521268bccc1a3efd97a5ceaf231ff33ab13fa6fe3f9cadf48573f2b1',
+        disaggregatedConfig(
+          DSV4_GB200_DYNAMO_VLLM,
+          { tp: 8, ep: 8, dpAttn: true, numWorkers: 2 },
+          { tp: 16, ep: 16, dpAttn: true, numWorkers: 1 },
+        ),
+      ],
+    ] as const
+  ).map(([productionConfigId, conc, recipeFingerprint, config]) => ({
+    id: `run-31965016666-config-${productionConfigId}-conc-${conc}-mooncake-offload`,
+    reason:
+      'The runtime recipe enabled MooncakeStoreConnector, but the artifact reported this AgentX point as non-offloaded.',
+    githubRunId: 31965016666,
+    runAttempt: 2,
+    productionConfigId,
+    config,
+    benchmarkType: 'agentic_traces',
+    isl: null,
+    osl: null,
+    conc,
+    offloadMode: 'off',
+    recipeFingerprint,
+    set: {
+      offloadMode: 'on' as const,
+      metricsRemove: ['allocated_cpu_dram_gb'],
+      metricsMerge: {
+        kv_offloading: 'dram',
+        kv_offload_backend: 'mooncake',
+        kv_offload_backend_version: '0.3.11.post1',
+      },
+    },
+  })),
+];
+
+function pointIdentity(
+  point: BenchmarkPointKey & { githubRunId: number; runAttempt: number },
+): string {
+  return JSON.stringify([
+    point.githubRunId,
+    point.runAttempt,
+    point.configId,
+    point.benchmarkType,
+    point.isl,
+    point.osl,
+    point.conc,
+    point.offloadMode,
+    point.recipeFingerprint ?? null,
+  ]);
+}
+
+function backfillPointIdentity(
+  backfill: BenchmarkPointBackfill,
+  offloadMode: string = backfill.offloadMode,
+): string {
+  return JSON.stringify([
+    backfill.githubRunId,
+    backfill.runAttempt,
+    configCacheKey(backfill.config),
+    backfill.benchmarkType,
+    backfill.isl,
+    backfill.osl,
+    backfill.conc,
+    offloadMode,
+    backfill.recipeFingerprint ?? null,
+  ]);
+}
+
+function backfillProductionPointIdentity(
+  backfill: BenchmarkPointBackfill,
+  offloadMode: string = backfill.offloadMode,
+): string {
+  return pointIdentity({
+    githubRunId: backfill.githubRunId,
+    runAttempt: backfill.runAttempt,
+    configId: backfill.productionConfigId,
+    benchmarkType: backfill.benchmarkType,
+    isl: backfill.isl,
+    osl: backfill.osl,
+    conc: backfill.conc,
+    offloadMode,
+    recipeFingerprint: backfill.recipeFingerprint,
+  });
+}
+
+function validatePositiveInteger(value: number, label: string, id: string): void {
+  if (!Number.isInteger(value) || value <= 0) {
+    throw new Error(`${id}: ${label} must be a positive integer`);
+  }
+}
+
+function validateBackfillConfig(config: ConfigParams, id: string): void {
+  for (const [label, value] of Object.entries({
+    hardware: config.hardware,
+    framework: config.framework,
+    model: config.model,
+    precision: config.precision,
+    specMethod: config.specMethod,
+  })) {
+    if (value.length === 0) throw new Error(`${id}: config.${label} must not be empty`);
+  }
+  for (const [label, value] of Object.entries({
+    prefillTp: config.prefillTp,
+    prefillEp: config.prefillEp,
+    prefillNumWorkers: config.prefillNumWorkers,
+    decodeTp: config.decodeTp,
+    decodeEp: config.decodeEp,
+    decodeNumWorkers: config.decodeNumWorkers,
+    numPrefillGpu: config.numPrefillGpu,
+    numDecodeGpu: config.numDecodeGpu,
+  })) {
+    validatePositiveInteger(value, `config.${label}`, id);
+  }
+}
+
+/** Validate the backfill ledger before an ingest or database write starts. */
+export function validateRunBackfills(
+  changelogs: readonly ChangelogBackfill[] = CHANGELOG_BACKFILLS,
+  points: readonly BenchmarkPointBackfill[] = BENCHMARK_POINT_BACKFILLS,
+): void {
+  const ids = new Set<string>();
+  const changelogIdentities = new Set<string>();
+  const pointSourceIdentities = new Map<string, string>();
+  const pointDesiredIdentities = new Map<string, string>();
+
+  for (const backfill of [...changelogs, ...points]) {
+    if (!/^[a-z0-9]+(?:-[a-z0-9]+)*$/u.test(backfill.id)) {
+      throw new Error(`${backfill.id || '<empty>'}: id must be lowercase kebab-case`);
+    }
+    if (ids.has(backfill.id)) throw new Error(`duplicate backfill id: ${backfill.id}`);
+    ids.add(backfill.id);
+    if (backfill.reason.trim().length === 0) {
+      throw new Error(`${backfill.id}: reason must not be empty`);
+    }
+    validatePositiveInteger(backfill.githubRunId, 'githubRunId', backfill.id);
+    validatePositiveInteger(backfill.runAttempt, 'runAttempt', backfill.id);
+    if (PURGED_RUNS.has(backfill.githubRunId)) {
+      throw new Error(`${backfill.id}: run is already in PURGED_RUNS`);
+    }
+    if (PURGED_RUN_ATTEMPTS.get(backfill.githubRunId)?.has(backfill.runAttempt)) {
+      throw new Error(`${backfill.id}: run attempt is already in PURGED_RUN_ATTEMPTS`);
+    }
+  }
+
+  for (const backfill of changelogs) {
+    if (backfill.baseRef.length === 0 || backfill.headRef.length === 0) {
+      throw new Error(`${backfill.id}: baseRef and headRef must not be empty`);
+    }
+    if (
+      backfill.set.configKeys === undefined &&
+      backfill.set.description === undefined &&
+      backfill.set.prLink === undefined &&
+      backfill.set.appendOnly === undefined
+    ) {
+      throw new Error(`${backfill.id}: set must change at least one field`);
+    }
+    if (backfill.set.configKeys?.some((key) => key.length === 0)) {
+      throw new Error(`${backfill.id}: configKeys must not contain empty values`);
+    }
+    const identity = JSON.stringify([
+      backfill.githubRunId,
+      backfill.runAttempt,
+      backfill.baseRef,
+      backfill.headRef,
+    ]);
+    if (changelogIdentities.has(identity)) {
+      throw new Error(`${backfill.id}: duplicate changelog selector ${identity}`);
+    }
+    changelogIdentities.add(identity);
+  }
+
+  for (const backfill of points) {
+    validatePositiveInteger(backfill.productionConfigId, 'productionConfigId', backfill.id);
+    validateBackfillConfig(backfill.config, backfill.id);
+    validatePositiveInteger(backfill.conc, 'conc', backfill.id);
+    if (backfill.benchmarkType.length === 0 || backfill.offloadMode.length === 0) {
+      throw new Error(`${backfill.id}: benchmarkType and offloadMode must not be empty`);
+    }
+    if (backfill.isl !== null) validatePositiveInteger(backfill.isl, 'isl', backfill.id);
+    if (backfill.osl !== null) validatePositiveInteger(backfill.osl, 'osl', backfill.id);
+    if (backfill.recipeFingerprint === '') {
+      throw new Error(`${backfill.id}: recipeFingerprint must be null or non-empty`);
+    }
+    const mergeKeys = Object.keys(backfill.set.metricsMerge ?? {});
+    const removeKeys = backfill.set.metricsRemove ?? [];
+    if (
+      backfill.set.offloadMode === undefined &&
+      mergeKeys.length === 0 &&
+      removeKeys.length === 0
+    ) {
+      throw new Error(`${backfill.id}: set must change at least one field`);
+    }
+    if (mergeKeys.includes('offload_mode') || removeKeys.includes('offload_mode')) {
+      throw new Error(`${backfill.id}: use set.offloadMode to change metrics.offload_mode`);
+    }
+    if (new Set(removeKeys).size !== removeKeys.length || removeKeys.some((key) => key === '')) {
+      throw new Error(`${backfill.id}: metricsRemove keys must be unique and non-empty`);
+    }
+    const overlap = mergeKeys.find((key) => removeKeys.includes(key));
+    if (overlap) throw new Error(`${backfill.id}: metric ${overlap} is both merged and removed`);
+
+    const sourceIdentity = backfillPointIdentity(backfill);
+    if (pointSourceIdentities.has(sourceIdentity)) {
+      throw new Error(`${backfill.id}: duplicate benchmark point selector ${sourceIdentity}`);
+    }
+    pointSourceIdentities.set(sourceIdentity, backfill.id);
+
+    const desiredOffloadMode = backfill.set.offloadMode ?? backfill.offloadMode;
+    const desiredIdentity = backfillPointIdentity(backfill, desiredOffloadMode);
+    if (pointDesiredIdentities.has(desiredIdentity)) {
+      throw new Error(`${backfill.id}: desired point identity collides with another backfill`);
+    }
+    pointDesiredIdentities.set(desiredIdentity, backfill.id);
+
+    if (
+      PURGED_BENCHMARK_POINTS.some((purged) => {
+        const purgedIdentity = pointIdentity(purged);
+        return (
+          purgedIdentity === backfillProductionPointIdentity(backfill) ||
+          purgedIdentity === backfillProductionPointIdentity(backfill, desiredOffloadMode)
+        );
+      })
+    ) {
+      throw new Error(`${backfill.id}: source or desired point is already being purged`);
+    }
+  }
+
+  for (const [identity, desiredBy] of pointDesiredIdentities) {
+    const selectedBy = pointSourceIdentities.get(identity);
+    if (selectedBy !== undefined && selectedBy !== desiredBy) {
+      throw new Error(
+        `${desiredBy}: desired point identity collides with selector from ${selectedBy}`,
+      );
+    }
+  }
+}
+
+interface BackfillablePoint extends BenchmarkPointKey {
+  config: ConfigParams;
+  metrics: Record<string, unknown>;
+}
+
+export interface AppliedBenchmarkPointBackfill<T extends BackfillablePoint> {
+  point: T;
+  backfillId: string | null;
+  sourceIdentity: string;
+  desiredIdentity: string;
+}
+
+interface BackfillableChangelogEntry {
+  configKeys: string[];
+  description: string;
+  prLink: string | null;
+  appendOnly: boolean;
+}
+
+interface BackfillableChangelog<T extends BackfillableChangelogEntry> {
+  baseRef: string;
+  headRef: string;
+  entries: T[];
+}
+
+/** Apply changelog corrections before artifact metadata is written to PostgreSQL. */
+export function applyChangelogBackfills<T extends BackfillableChangelogEntry>(
+  githubRunId: number,
+  runAttempt: number | null | undefined,
+  changelogs: readonly BackfillableChangelog<T>[],
+): { changelogs: BackfillableChangelog<T>[]; backfillIds: string[] } {
+  const relevant = CHANGELOG_BACKFILLS.filter(
+    (backfill) =>
+      backfill.githubRunId === githubRunId &&
+      (runAttempt === null || runAttempt === undefined || backfill.runAttempt === runAttempt),
+  );
+  const corrected = changelogs.map((changelog) => ({
+    ...changelog,
+    entries: [...changelog.entries],
+  }));
+  const backfillIds: string[] = [];
+
+  for (const changelog of corrected) {
+    const matches = relevant.filter(
+      (backfill) =>
+        backfill.baseRef === changelog.baseRef && backfill.headRef === changelog.headRef,
+    );
+    if (matches.length > 1) {
+      throw new Error(
+        `changelog matches multiple backfills: ${matches.map((backfill) => backfill.id).join(', ')}`,
+      );
+    }
+    const [backfill] = matches;
+    if (!backfill || changelog.entries.length === 0) continue;
+
+    const { appendOnly } = backfill.set;
+    if (appendOnly !== undefined) {
+      changelog.entries = changelog.entries.map((entry) => ({
+        ...entry,
+        appendOnly,
+      }));
+    }
+    // ingestChangelogEntries upserts every entry onto the same (run, base, head)
+    // identity, so the final artifact entry is the row that persists.
+    const entryIndex = changelog.entries.length - 1;
+    const entry = changelog.entries[entryIndex];
+    changelog.entries[entryIndex] = {
+      ...entry,
+      ...(backfill.set.configKeys === undefined
+        ? {}
+        : { configKeys: [...backfill.set.configKeys] }),
+      ...(backfill.set.description === undefined ? {} : { description: backfill.set.description }),
+      ...(backfill.set.prLink === undefined ? {} : { prLink: backfill.set.prLink }),
+    };
+    backfillIds.push(backfill.id);
+  }
+
+  return { changelogs: corrected, backfillIds };
+}
+
+function matchesBenchmarkPoint(
+  point: BackfillablePoint,
+  selector: BenchmarkPointBackfill,
+): boolean {
+  return (
+    configCacheKey(point.config) === configCacheKey(selector.config) &&
+    point.benchmarkType === selector.benchmarkType &&
+    point.isl === selector.isl &&
+    point.osl === selector.osl &&
+    point.conc === selector.conc &&
+    point.offloadMode === selector.offloadMode &&
+    (point.recipeFingerprint ?? null) === (selector.recipeFingerprint ?? null)
+  );
+}
+
+/** Apply one point patch to an ingest row before it reaches PostgreSQL. */
+export function applyBenchmarkPointBackfill<T extends BackfillablePoint>(
+  githubRunId: number,
+  runAttempt: number | null | undefined,
+  point: T,
+): AppliedBenchmarkPointBackfill<T> {
+  const sourceIdentity = JSON.stringify([
+    point.configId,
+    point.benchmarkType,
+    point.isl,
+    point.osl,
+    point.conc,
+    point.offloadMode,
+    point.recipeFingerprint ?? null,
+  ]);
+  const matches = BENCHMARK_POINT_BACKFILLS.filter(
+    (backfill) =>
+      backfill.githubRunId === githubRunId &&
+      (runAttempt === null || runAttempt === undefined || backfill.runAttempt === runAttempt) &&
+      matchesBenchmarkPoint(point, backfill),
+  );
+  if (matches.length > 1) {
+    throw new Error(
+      `benchmark point matches multiple backfills: ${matches.map((b) => b.id).join(', ')}`,
+    );
+  }
+  const [backfill] = matches;
+  if (!backfill) {
+    return { point, backfillId: null, sourceIdentity, desiredIdentity: sourceIdentity };
+  }
+
+  const metrics = { ...point.metrics };
+  for (const key of backfill.set.metricsRemove ?? []) delete metrics[key];
+  Object.assign(metrics, backfill.set.metricsMerge);
+  const offloadMode = backfill.set.offloadMode ?? point.offloadMode;
+  if (backfill.set.offloadMode !== undefined) metrics.offload_mode = offloadMode;
+  const patched = { ...point, offloadMode, metrics };
+  const desiredIdentity = JSON.stringify([
+    patched.configId,
+    patched.benchmarkType,
+    patched.isl,
+    patched.osl,
+    patched.conc,
+    patched.offloadMode,
+    patched.recipeFingerprint ?? null,
+  ]);
+  return {
+    point: patched,
+    backfillId: backfill.id,
+    sourceIdentity,
+    desiredIdentity,
+  };
+}
+
+/** Fail if two distinct artifact points collapse onto one identity after correction. */
+export function recordBackfilledPointIdentity(
+  seen: Map<string, string>,
+  sourceIdentity: string,
+  desiredIdentity: string,
+): void {
+  const existingSource = seen.get(desiredIdentity);
+  if (existingSource !== undefined && existingSource !== sourceIdentity) {
+    throw new Error(
+      `benchmark point backfill collision: ${existingSource} and ${sourceIdentity} both become ${desiredIdentity}`,
+    );
+  }
+  seen.set(desiredIdentity, sourceIdentity);
+}
 
 /**
  * True when this exact benchmark result is suppressed. When an ingest source
@@ -148,7 +1202,8 @@ export function isBenchmarkPointPurged(
       candidate.isl === point.isl &&
       candidate.osl === point.osl &&
       candidate.conc === point.conc &&
-      candidate.offloadMode === point.offloadMode,
+      candidate.offloadMode === point.offloadMode &&
+      (candidate.recipeFingerprint ?? null) === (point.recipeFingerprint ?? null),
   );
 }
 
diff --git a/packages/db/src/etl/server-log-artifacts.test.ts b/packages/db/src/etl/server-log-artifacts.test.ts
new file mode 100644
index 000000000..78f23588b
--- /dev/null
+++ b/packages/db/src/etl/server-log-artifacts.test.ts
@@ -0,0 +1,62 @@
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+import { afterEach, describe, expect, it } from 'vitest';
+
+import {
+  cleanLogText,
+  listServerLogFilePaths,
+  primaryServerLogFile,
+  readServerLogFiles,
+  serverLogArtifactSuffix,
+} from './server-log-artifacts.js';
+
+const roots: string[] = [];
+
+afterEach(() => {
+  for (const root of roots.splice(0)) fs.rmSync(root, { recursive: true, force: true });
+});
+
+function tempRoot(): string {
+  const root = fs.mkdtempSync(path.join(os.tmpdir(), 'server-log-artifacts-'));
+  roots.push(root);
+  return root;
+}
+
+describe('server-log artifact discovery', () => {
+  it('recursively keeps .log and .out files with relative paths', () => {
+    const root = tempRoot();
+    fs.mkdirSync(path.join(root, 'agentic', 'logs'), { recursive: true });
+    fs.writeFileSync(path.join(root, 'agentic', 'logs', 'router.log'), 'router\u0000\n');
+    fs.writeFileSync(path.join(root, 'worker.out'), 'worker\n');
+    fs.writeFileSync(path.join(root, 'metrics.json'), '{}');
+
+    expect(listServerLogFilePaths(root).map((file) => file.fileName)).toEqual([
+      'agentic/logs/router.log',
+      'worker.out',
+    ]);
+    expect(readServerLogFiles(root)).toEqual([
+      { fileName: 'agentic/logs/router.log', logText: 'router\n' },
+      { fileName: 'worker.out', logText: 'worker\n' },
+    ]);
+  });
+
+  it('prefers a nested server.log as the primary stream', () => {
+    const files = [
+      { fileName: 'benchmark.out', logText: 'benchmark' },
+      { fileName: 'results/server.log', logText: 'server' },
+    ];
+    expect(primaryServerLogFile(files)).toEqual(files[1]);
+  });
+
+  it('recognizes single-node and multinode artifact prefixes', () => {
+    expect(serverLogArtifactSuffix('server_logs_config-a')).toBe('config-a');
+    expect(serverLogArtifactSuffix('multinode_server_logs_config-b')).toBe('config-b');
+    expect(serverLogArtifactSuffix('agentic_config-c')).toBeNull();
+  });
+
+  it('removes PostgreSQL-incompatible null bytes', () => {
+    expect(cleanLogText('a\u0000b')).toBe('ab');
+  });
+});
diff --git a/packages/db/src/etl/server-log-artifacts.ts b/packages/db/src/etl/server-log-artifacts.ts
new file mode 100644
index 000000000..fc9e5f5de
--- /dev/null
+++ b/packages/db/src/etl/server-log-artifacts.ts
@@ -0,0 +1,132 @@
+import { execFileSync } from 'node:child_process';
+import fs from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+export interface ServerLogFile {
+  /** POSIX-style path relative to the extracted server-log artifact root. */
+  fileName: string;
+  logText: string;
+}
+
+export interface ServerLogFilePath {
+  /** POSIX-style path relative to the extracted server-log artifact root. */
+  fileName: string;
+  path: string;
+}
+
+export interface ServerLogArtifact {
+  artifactName: string;
+  artifactDir: string;
+}
+
+const SERVER_LOG_PREFIXES = ['multinode_server_logs_', 'server_logs_'] as const;
+
+export function isLogFileName(fileName: string): boolean {
+  const lower = fileName.toLowerCase();
+  return lower.endsWith('.log') || lower.endsWith('.out');
+}
+
+export function cleanLogText(value: string): string {
+  return value.replaceAll('\u0000', '');
+}
+
+/** Return the shared suffix used to pair a server-log artifact with bmk[_agentic] artifacts. */
+export function serverLogArtifactSuffix(artifactName: string): string | null {
+  for (const prefix of SERVER_LOG_PREFIXES) {
+    if (artifactName.startsWith(prefix)) return artifactName.slice(prefix.length);
+  }
+  return null;
+}
+
+/** Recursively list every regular .log/.out file under an extracted artifact root. */
+export function listServerLogFilePaths(root: string): ServerLogFilePath[] {
+  if (!fs.existsSync(root) || !fs.statSync(root).isDirectory()) return [];
+  const files: ServerLogFilePath[] = [];
+
+  const visit = (directory: string): void => {
+    for (const entry of fs.readdirSync(directory, { withFileTypes: true })) {
+      const pathname = path.join(directory, entry.name);
+      if (entry.isDirectory()) visit(pathname);
+      else if (entry.isFile() && isLogFileName(entry.name)) {
+        files.push({
+          fileName: path.relative(root, pathname).split(path.sep).join('/'),
+          path: pathname,
+        });
+      }
+    }
+  };
+
+  visit(root);
+  return files.toSorted((a, b) => a.fileName.localeCompare(b.fileName));
+}
+
+export function readServerLogFiles(root: string): ServerLogFile[] {
+  return readServerLogFilePaths(listServerLogFilePaths(root));
+}
+
+export function readServerLogFilePaths(files: readonly ServerLogFilePath[]): ServerLogFile[] {
+  return files.map((file) => ({
+    fileName: file.fileName,
+    logText: cleanLogText(fs.readFileSync(file.path, 'utf8')),
+  }));
+}
+
+/** Prefer server.log as the legacy/primary stream, then fall back to the first file. */
+export function primaryServerLogFile(files: readonly ServerLogFile[]): ServerLogFile | null {
+  return (
+    files.find((file) => path.posix.basename(file.fileName).toLowerCase() === 'server.log') ??
+    files[0] ??
+    null
+  );
+}
+
+/**
+ * Resolve the root containing logs for an extracted GitHub artifact.
+ * Multinode uploads wrap their files in multinode_server_logs.tar.gz.
+ */
+export function serverLogArtifactRoot(artifactDir: string, artifactName: string): string | null {
+  if (artifactName.startsWith('server_logs_')) return artifactDir;
+  if (!artifactName.startsWith('multinode_server_logs_')) return null;
+
+  const archivePath = path.join(artifactDir, 'multinode_server_logs.tar.gz');
+  const extractedDir = path.join(artifactDir, 'multinode_server_logs');
+  if (!fs.existsSync(extractedDir) && fs.existsSync(archivePath)) {
+    fs.mkdirSync(extractedDir, { recursive: true });
+    execFileSync('tar', ['-xzf', archivePath, '-C', extractedDir], { stdio: 'ignore' });
+  }
+  return fs.existsSync(extractedDir) && fs.statSync(extractedDir).isDirectory()
+    ? extractedDir
+    : null;
+}
+
+export function readServerLogArtifact(artifact: ServerLogArtifact): ServerLogFile[] {
+  const root = serverLogArtifactRoot(artifact.artifactDir, artifact.artifactName);
+  return root ? readServerLogFiles(root) : [];
+}
+
+/** Read one archived GCS copy without retaining its potentially GiB-scale expansion. */
+export function readServerLogArtifactZip(zipPath: string, artifactName: string): ServerLogFile[] {
+  const tempDir = fs.mkdtempSync(path.join(os.tmpdir(), 'server-log-artifact-'));
+  try {
+    execFileSync('unzip', ['-oq', zipPath, '-d', tempDir], { stdio: 'ignore' });
+    return readServerLogArtifact({ artifactName, artifactDir: tempDir });
+  } finally {
+    fs.rmSync(tempDir, { recursive: true, force: true });
+  }
+}
+
+/** Index all extracted single-node and multinode server-log artifacts by shared suffix. */
+export function discoverServerLogArtifacts(artifactsDir: string): Map<string, ServerLogArtifact> {
+  const discovered = new Map<string, ServerLogArtifact>();
+  if (!fs.existsSync(artifactsDir)) return discovered;
+
+  for (const artifactName of fs.readdirSync(artifactsDir)) {
+    const suffix = serverLogArtifactSuffix(artifactName);
+    if (!suffix) continue;
+    const artifactDir = path.join(artifactsDir, artifactName);
+    if (!fs.statSync(artifactDir).isDirectory()) continue;
+    discovered.set(suffix, { artifactName, artifactDir });
+  }
+  return discovered;
+}
diff --git a/packages/db/src/etl/server-metrics-adapters.ts b/packages/db/src/etl/server-metrics-adapters.ts
index f123d9f82..5e80e8bd6 100644
--- a/packages/db/src/etl/server-metrics-adapters.ts
+++ b/packages/db/src/etl/server-metrics-adapters.ts
@@ -56,17 +56,24 @@ const dynamoAdapter: ServerMetricsAdapter = {
             : 'unknown';
     const endpointUrl = series.endpoint_url ?? labels['dynamo_endpoint'] ?? null;
     const workerId = labels['worker_id'] ?? null;
-    const dpRank = labels['dp_rank'] ?? null;
-    const engine = labels['engine'] ?? labels['engine_idx'] ?? null;
     return {
-      id: stableId('dynamo', [role, endpointUrl, workerId, dpRank, engine]),
+      // A source is one WORKER, not one engine inside it. `dp_rank` (sglang)
+      // and `engine` (vllm) both name engines *within* a worker, so keying on
+      // them split every worker into 4-5 identical-looking entries: a 6p1d run
+      // listed 35 endpoints for 7 workers. The endpoint is only identity when
+      // the worker is anonymous — otherwise one worker reachable on two URLs
+      // would split again.
+      id: stableId('dynamo', [role, workerId ?? endpointUrl]),
       adapter: 'dynamo',
       role,
       endpointUrl,
       nativeRole,
       workerId,
-      dpRank,
-      engine,
+      // Null at worker granularity: this source spans every rank/engine the
+      // worker owns, so naming one of them would be arbitrary. The per-engine
+      // breakdown lives in the source's own `kvCacheUsageByEngine`.
+      dpRank: null,
+      engine: null,
     };
   },
 };
@@ -81,14 +88,14 @@ const genericAdapter: ServerMetricsAdapter = {
     const dpRank = labels['dp_rank'] ?? null;
     const engine = labels['engine'] ?? labels['engine_idx'] ?? null;
     return {
-      id: stableId('generic', [endpointUrl, workerId, dpRank, engine]),
+      id: stableId('generic', [workerId ?? endpointUrl]),
       adapter: 'generic',
       role: endpointUrl || workerId || dpRank || engine ? 'unknown' : 'combined',
       endpointUrl,
       nativeRole: null,
       workerId,
-      dpRank,
-      engine,
+      dpRank: null,
+      engine: null,
     };
   },
 };
diff --git a/packages/db/src/etl/workflow-run.ts b/packages/db/src/etl/workflow-run.ts
index 28d27c87e..22b171c09 100644
--- a/packages/db/src/etl/workflow-run.ts
+++ b/packages/db/src/etl/workflow-run.ts
@@ -148,6 +148,7 @@ export function createWorkflowRunServices(sql: Sql, githubToken?: string) {
     conclusion?: string | null;
     htmlUrl?: string | null;
     runStartedAt?: string | null;
+    appendOnly?: boolean;
     ghInfo?: GithubRunInfo | null;
   }): Promise<number | null> {
     const attempt = params.runAttempt ?? params.ghInfo?.runAttempt ?? 0;
@@ -166,16 +167,18 @@ export function createWorkflowRunServices(sql: Sql, githubToken?: string) {
     const runStartedAt = gh?.runStartedAt ?? params.runStartedAt ?? null;
     const headSha = gh?.headSha ?? params.headSha ?? null;
     const headBranch = gh?.headBranch ?? params.headBranch ?? null;
+    const appendOnly = params.appendOnly ?? false;
 
     const [row] = await sql`
       insert into workflow_runs (
         github_run_id, run_attempt, name, status, conclusion,
-        head_sha, head_branch, html_url, created_at, run_started_at, date
+        head_sha, head_branch, html_url, created_at, run_started_at, date, append_only
       ) values (
         ${params.githubRunId}, ${attempt}, ${name},
         ${status}, ${conclusion},
         ${headSha}, ${headBranch}, ${htmlUrl},
-        ${createdAt}::timestamptz, ${runStartedAt}::timestamptz, ${params.date}::date
+        ${createdAt}::timestamptz, ${runStartedAt}::timestamptz, ${params.date}::date,
+        ${appendOnly}
       )
       on conflict (github_run_id, run_attempt)
       do update set
@@ -186,7 +189,8 @@ export function createWorkflowRunServices(sql: Sql, githubToken?: string) {
         created_at     = excluded.created_at,
         run_started_at = excluded.run_started_at,
         head_sha       = excluded.head_sha,
-        head_branch    = excluded.head_branch
+        head_branch    = excluded.head_branch,
+        append_only    = excluded.append_only
       returning id
     `;
 
diff --git a/packages/db/src/etl/zip-reader.test.ts b/packages/db/src/etl/zip-reader.test.ts
new file mode 100644
index 000000000..a964b7dcd
--- /dev/null
+++ b/packages/db/src/etl/zip-reader.test.ts
@@ -0,0 +1,58 @@
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+const mocks = vi.hoisted(() => ({
+  admZip: vi.fn(),
+  getEntries: vi.fn(),
+}));
+
+vi.mock('adm-zip', () => ({
+  default: class MockAdmZip {
+    constructor(zipPath: string) {
+      mocks.admZip(zipPath);
+    }
+
+    getEntries() {
+      return mocks.getEntries();
+    }
+  },
+}));
+
+import { readZipJsonAndText } from './zip-reader';
+
+afterEach(() => {
+  vi.clearAllMocks();
+});
+
+describe('readZipJsonAndText', () => {
+  it('collects JSON and selected text with one archive traversal', () => {
+    const metaData = vi.fn(() => Buffer.from('{"model":"test"}'));
+    const resultsData = vi.fn(() => Buffer.from('{"results":{"task":{}}}'));
+    const samplesData = vi.fn(() => Buffer.from('{"doc_id":1}\n'));
+    const serverLogData = vi.fn(() => Buffer.from('deferred server log'));
+    mocks.getEntries.mockReturnValue([
+      { isDirectory: false, name: 'meta_env.json', getData: metaData },
+      { isDirectory: false, name: 'results_task.json', getData: resultsData },
+      { isDirectory: false, name: 'samples_task_123.jsonl', getData: samplesData },
+      { isDirectory: false, name: 'server.log', getData: serverLogData },
+    ]);
+
+    const contents = readZipJsonAndText(
+      'eval.zip',
+      (name) => name.startsWith('samples_') && name.endsWith('.jsonl'),
+    );
+
+    expect(mocks.admZip).toHaveBeenCalledTimes(1);
+    expect(mocks.getEntries).toHaveBeenCalledTimes(1);
+    expect(metaData).toHaveBeenCalledTimes(1);
+    expect(resultsData).toHaveBeenCalledTimes(1);
+    expect(samplesData).toHaveBeenCalledTimes(1);
+    expect(serverLogData).not.toHaveBeenCalled();
+    expect(contents?.jsonFiles).toEqual(
+      new Map([
+        ['meta_env.json', { model: 'test' }],
+        ['results_task.json', { results: { task: {} } }],
+      ]),
+    );
+    expect(contents?.textFiles).toEqual(new Map([['samples_task_123.jsonl', '{"doc_id":1}\n']]));
+  });
+});
diff --git a/packages/db/src/etl/zip-reader.ts b/packages/db/src/etl/zip-reader.ts
index b572020e7..e46a4216d 100644
--- a/packages/db/src/etl/zip-reader.ts
+++ b/packages/db/src/etl/zip-reader.ts
@@ -28,27 +28,45 @@ export function readZipText(zipPath: string, name: string): string | null {
   }
 }
 
+export interface ZipJsonAndText {
+  jsonFiles: Map<string, unknown>;
+  textFiles: Map<string, string>;
+}
+
 /**
- * Read all text files in a ZIP whose basenames match `predicate`, keyed by basename.
- * Returns an empty map on ZIP-level errors. Used to pull all `samples_*.jsonl`
- * entries from an eval ZIP in a single pass.
+ * Traverse a ZIP once, parsing every JSON entry and reading only selected text entries.
+ * Returns null on ZIP-level errors; individual JSON parse errors yield null values and
+ * individual text read errors omit that entry.
  */
-export function readZipTextsMatching(
+export function readZipJsonAndText(
   zipPath: string,
-  predicate: (name: string) => boolean,
-): Map<string, string> {
-  const out = new Map<string, string>();
+  selectText?: (name: string) => boolean,
+): ZipJsonAndText | null {
   try {
     const zip = new AdmZip(zipPath);
+    const jsonFiles = new Map<string, unknown>();
+    const textFiles = new Map<string, string>();
+
     for (const entry of zip.getEntries()) {
       if (entry.isDirectory) continue;
-      if (!predicate(entry.name)) continue;
-      try {
-        out.set(entry.name, entry.getData().toString('utf8'));
-      } catch {}
+
+      if (entry.name.endsWith('.json')) {
+        try {
+          jsonFiles.set(entry.name, JSON.parse(entry.getData().toString('utf8')));
+        } catch {
+          jsonFiles.set(entry.name, null);
+        }
+      } else if (selectText?.(entry.name)) {
+        try {
+          textFiles.set(entry.name, entry.getData().toString('utf8'));
+        } catch {}
+      }
     }
-  } catch {}
-  return out;
+
+    return { jsonFiles, textFiles };
+  } catch {
+    return null;
+  }
 }
 
 /**
diff --git a/packages/db/src/ingest-ci-run.ts b/packages/db/src/ingest-ci-run.ts
index f5b178447..41c5bcfc7 100644
--- a/packages/db/src/ingest-ci-run.ts
+++ b/packages/db/src/ingest-ci-run.ts
@@ -35,9 +35,18 @@ import {
   fetchRunAttempt,
   listRunArtifacts,
 } from './lib/github-artifacts';
+import { pairServerLogArtifacts } from './lib/server-log-backfill';
 import { createAdminSql, refreshLatestBenchmarks } from './etl/db-utils';
-import { isBenchmarkPointPurged, isRunAttemptPurged } from './etl/run-overrides';
+import {
+  applyBenchmarkPointBackfill,
+  applyChangelogBackfills,
+  isBenchmarkPointPurged,
+  isRunAttemptPurged,
+  recordBackfilledPointIdentity,
+  validateRunBackfills,
+} from './etl/run-overrides';
 import { createSkipTracker } from './etl/skip-tracker';
+import { printIngestSummaryFooter } from './etl/ingest-summary';
 import { createConfigCache } from './etl/config-cache';
 import { createWorkflowRunServices } from './etl/workflow-run';
 import {
@@ -49,7 +58,7 @@ import {
   bulkIngestBenchmarkRows,
   bulkIngestRunStats,
   bulkUpsertAvailability,
-  insertServerLog,
+  insertServerLogFiles,
 } from './etl/benchmark-ingest';
 import { findUnlinkedTraceReplayIds, persistPreparedTraceReplay } from './etl/trace-replay-ingest';
 import {
@@ -58,6 +67,7 @@ import {
 } from './etl/trace-replay-worker-pool';
 import { AsyncSemaphore } from './etl/async-semaphore';
 import { discoverTraceReplayArtifacts } from './etl/trace-artifact-discovery';
+import { discoverServerLogArtifacts, readServerLogArtifact } from './etl/server-log-artifacts';
 import { datasetSlugFromBenchmarkRow } from './etl/dataset-provenance';
 import { mapAggEvalRow, mapEvalRow } from './etl/eval-mapper';
 import { ingestEvalRow } from './etl/eval-ingest';
@@ -67,6 +77,7 @@ import {
   type ChangelogEntry,
   parseChangelogEntries,
   ingestChangelogEntries,
+  hasAppendOnlyFlag,
   hasEvalsOnlyFlag,
 } from './etl/changelog-ingest';
 
@@ -137,7 +148,29 @@ if (isDownloadMode) {
   // most recent per logical name (see RUNNER_SUFFIX_RE in github-artifacts)
   // so a failed attempt's empty metrics can't overwrite the good one via
   // ON CONFLICT DO UPDATE.
-  const byLogical = dedupeArtifactsByLogicalName(listRunArtifacts(REPO, runIdStr));
+  const artifacts = listRunArtifacts(REPO, runIdStr);
+  const byLogical = dedupeArtifactsByLogicalName(artifacts);
+  // Server-log artifacts from eval and benchmark jobs can share a logical
+  // config but differ by runner suffix. Keep the exact server-log sibling for
+  // each selected bmk artifact instead of letting latest-created eval logs win.
+  for (const [key, artifact] of byLogical) {
+    if (
+      artifact.name.startsWith('server_logs_') ||
+      artifact.name.startsWith('multinode_server_logs_')
+    ) {
+      byLogical.delete(key);
+    }
+  }
+  const selectedBenchmarkNames = new Set(
+    [...byLogical.values()]
+      .filter((artifact) => artifact.name.startsWith('bmk_'))
+      .map((artifact) => artifact.name),
+  );
+  for (const pair of pairServerLogArtifacts(artifacts)) {
+    if (selectedBenchmarkNames.has(pair.benchmarks.name)) {
+      byLogical.set(`server-log:${pair.serverLogs.name}`, pair.serverLogs);
+    }
+  }
 
   for (const artifact of byLogical.values()) {
     console.log(`  ${artifact.name}`);
@@ -232,6 +265,7 @@ function findJsonFiles(dir: string): string[] {
 // ── Main ────────────────────────────────────────────────────────────────────
 
 async function main(): Promise<void> {
+  validateRunBackfills();
   const tracker = createSkipTracker();
   const configCache = createConfigCache(sql);
   const { getOrCreateConfig, preloadConfigs } = configCache;
@@ -289,6 +323,61 @@ async function main(): Promise<void> {
     ? workflowGhInfo.createdAt.split('T')[0]
     : new Date().toISOString().split('T')[0];
 
+  // Parse changelog metadata before creating the workflow row: append-only is a
+  // run-level curve-selection contract used by the latest-benchmark queries.
+  const changelogDir = path.join(artifactsDir, ARTIFACT_NAMES.changelog);
+  const changelogFiles = findJsonFiles(changelogDir);
+  const parsedChangelogs: {
+    baseRef: string;
+    headRef: string;
+    entries: ChangelogEntry[];
+  }[] = [];
+  for (const file of changelogFiles) {
+    const data = readJson(file) as Record<string, any> | null;
+    if (!data || typeof data !== 'object') continue;
+    const baseRef = String(data.base_ref ?? '');
+    const headRef = String(data.head_ref ?? '');
+    if (!baseRef || !headRef) continue;
+    const entries = parseChangelogEntries(data.entries);
+    if (entries.length > 0) parsedChangelogs.push({ baseRef, headRef, entries });
+  }
+  if (parsedChangelogs.length === 0) {
+    const headRef = workflowGhInfo?.headBranch ?? workflowGhInfo?.headSha ?? `run-${runIdStr}`;
+    // Prefer the workflow's display name: it describes the sweep, while the head
+    // commit message often describes an unrelated code change.
+    const fallbackDescription =
+      workflowGhInfo?.name?.trim() ||
+      workflowGhInfo?.headCommitMessage?.trim().split('\n')[0]?.trim() ||
+      `GitHub Actions run ${runIdStr}`;
+
+    parsedChangelogs.push({
+      baseRef: 'unknown',
+      headRef,
+      entries: [
+        {
+          configKeys: [],
+          description: fallbackDescription,
+          prLink: null,
+          evalsOnly: false,
+          appendOnly: false,
+        },
+      ],
+    });
+    console.log(
+      `  No changelog metadata artifact found; using fallback changelog: ${fallbackDescription}`,
+    );
+  }
+  const { changelogs, backfillIds: changelogBackfillIds } = applyChangelogBackfills(
+    runIdNum,
+    runAttemptNum,
+    parsedChangelogs,
+  );
+  for (const backfillId of changelogBackfillIds) {
+    console.log(`  Applied changelog backfill ${backfillId}`);
+  }
+  const appendOnly = hasAppendOnlyFlag(changelogs);
+  const evalsOnly = hasEvalsOnlyFlag(changelogs);
+
   const workflowRunId = await getOrCreateWorkflowRun({
     githubRunId: runId,
     runAttempt: runAttemptNum,
@@ -300,6 +389,7 @@ async function main(): Promise<void> {
     headSha: workflowGhInfo?.headSha,
     htmlUrl: reusedIngestMetadata?.sourceRunUrl,
     createdAt: workflowGhInfo?.createdAt || triggerGhInfo?.createdAt || new Date().toISOString(),
+    appendOnly,
     ghInfo: workflowGhInfo,
   });
   if (workflowRunId === null) {
@@ -337,49 +427,6 @@ async function main(): Promise<void> {
   const missingDatasets = new Set<string>();
 
   // ── Check for evals-only flag in changelog ────────────────────────────
-  const changelogDir = path.join(artifactsDir, ARTIFACT_NAMES.changelog);
-  const changelogFiles = findJsonFiles(changelogDir);
-  const parsedChangelogs: {
-    baseRef: string;
-    headRef: string;
-    entries: ChangelogEntry[];
-  }[] = [];
-  for (const file of changelogFiles) {
-    const data = readJson(file) as Record<string, any> | null;
-    if (!data || typeof data !== 'object') continue;
-    const baseRef = String(data.base_ref ?? '');
-    const headRef = String(data.head_ref ?? '');
-    if (!baseRef || !headRef) continue;
-    const entries = parseChangelogEntries(data.entries);
-    if (entries.length > 0) parsedChangelogs.push({ baseRef, headRef, entries });
-  }
-  if (parsedChangelogs.length === 0) {
-    const headRef = workflowGhInfo?.headBranch ?? workflowGhInfo?.headSha ?? `run-${runIdStr}`;
-    // Prefer the workflow's display name ("e2e Test - B300 DSv4 AgentX vLLM 1h
-    // + 10m warmup") — it describes the sweep; the head commit message usually
-    // describes an unrelated code change.
-    const fallbackDescription =
-      workflowGhInfo?.name?.trim() ||
-      workflowGhInfo?.headCommitMessage?.trim().split('\n')[0]?.trim() ||
-      `GitHub Actions run ${runIdStr}`;
-
-    parsedChangelogs.push({
-      baseRef: 'unknown',
-      headRef,
-      entries: [
-        {
-          configKeys: [],
-          description: fallbackDescription,
-          prLink: null,
-          evalsOnly: false,
-        },
-      ],
-    });
-    console.log(
-      `  No changelog metadata artifact found; using fallback changelog: ${fallbackDescription}`,
-    );
-  }
-  const evalsOnly = hasEvalsOnlyFlag(parsedChangelogs);
   if (evalsOnly) {
     console.log('\n  ⚠ evals-only run detected — skipping benchmark and stats ingest');
   }
@@ -401,23 +448,9 @@ async function main(): Promise<void> {
           .filter((d) => fs.statSync(d).isDirectory())
       : [];
 
-    const serverLogPaths = new Map<string, string>();
-    if (fs.existsSync(artifactsDir)) {
-      for (const d of fs.readdirSync(artifactsDir)) {
-        if (!d.startsWith('server_logs_')) continue;
-        // feat-agentx-v1.0 harness nests the log under `results/server.log`;
-        // older runs keep it at the artifact root. Check both.
-        const logPath = [
-          path.join(artifactsDir, d, 'server.log'),
-          path.join(artifactsDir, d, 'results', 'server.log'),
-        ].find((p) => fs.existsSync(p));
-        if (!logPath) continue;
-        const configKey = d.replace(/^server_logs_/u, '');
-        serverLogPaths.set(configKey, logPath);
-      }
-    }
-    if (serverLogPaths.size > 0) {
-      console.log(`  Found ${serverLogPaths.size} server log artifact(s)`);
+    const serverLogArtifacts = discoverServerLogArtifacts(artifactsDir);
+    if (serverLogArtifacts.size > 0) {
+      console.log(`  Found ${serverLogArtifacts.size} server log artifact(s)`);
     }
 
     // Sibling aiperf artifacts: each `bmk_agentic_<suffix>` is paired with an
@@ -442,6 +475,7 @@ async function main(): Promise<void> {
     }
 
     const allBmkFiles = [...bmkFiles, ...allBmkDirs.flatMap((d) => findJsonFiles(d))];
+    const seenPointIdentities = new Map<string, string>();
     console.log(`  Found ${allBmkFiles.length} benchmark JSON file(s)`);
 
     for (const [fileIndex, file] of allBmkFiles.entries()) {
@@ -480,28 +514,47 @@ async function main(): Promise<void> {
 
       const toInsert = [];
       for (const row of rows) {
+        let configId: number;
         try {
-          const configId = await getOrCreateConfig(row.config);
-          if (
-            isBenchmarkPointPurged(runIdNum, runAttemptNum, {
-              configId,
-              benchmarkType: row.benchmarkType,
-              isl: row.isl,
-              osl: row.osl,
-              conc: row.conc,
-              offloadMode: row.offloadMode,
-            })
-          ) {
-            console.log(
-              `    skipped purged benchmark point: config ${configId}, ${row.benchmarkType}, ` +
-                `isl ${row.isl}, osl ${row.osl}, conc ${row.conc}, offload ${row.offloadMode}`,
-            );
-            continue;
-          }
-          toInsert.push({ ...row, configId });
+          configId = await getOrCreateConfig(row.config);
         } catch (error: any) {
           tracker.recordDbError(`config for ${path.basename(file)}`, error);
+          continue;
+        }
+        if (
+          isBenchmarkPointPurged(runIdNum, runAttemptNum, {
+            configId,
+            benchmarkType: row.benchmarkType,
+            isl: row.isl,
+            osl: row.osl,
+            conc: row.conc,
+            offloadMode: row.offloadMode,
+            recipeFingerprint: row.recipeFingerprint,
+          })
+        ) {
+          console.log(
+            `    skipped purged benchmark point: config ${configId}, ${row.benchmarkType}, ` +
+              `isl ${row.isl}, osl ${row.osl}, conc ${row.conc}, offload ${row.offloadMode}, ` +
+              `recipe ${row.recipeFingerprint ?? 'legacy'}`,
+          );
+          continue;
+        }
+        const applied = applyBenchmarkPointBackfill(runIdNum, runAttemptNum, {
+          ...row,
+          configId,
+        });
+        recordBackfilledPointIdentity(
+          seenPointIdentities,
+          applied.sourceIdentity,
+          applied.desiredIdentity,
+        );
+        if (applied.backfillId) {
+          console.log(
+            `    applied benchmark point backfill ${applied.backfillId}: ` +
+              `config ${configId}, conc ${row.conc}`,
+          );
         }
+        toInsert.push(applied.point);
       }
       console.log(`    rows with resolved configs: ${toInsert.length}`);
 
@@ -544,20 +597,24 @@ async function main(): Promise<void> {
             // prefix), so fall back to the fully-stripped suffix — otherwise
             // agentic rows never get their server log (and KV-pool size) linked.
             const configKey = parentDir.replace(/^bmk_/u, '');
-            const logPath =
-              serverLogPaths.get(configKey) ??
-              serverLogPaths.get(stripBmkAndAgenticPrefix(parentDir));
-            if (logPath) {
+            const logArtifact =
+              serverLogArtifacts.get(configKey) ??
+              serverLogArtifacts.get(stripBmkAndAgenticPrefix(parentDir));
+            if (logArtifact) {
               try {
                 const serverLogStart = Date.now();
+                const logFiles = readServerLogArtifact(logArtifact);
+                const totalBytes = logFiles.reduce(
+                  (bytes, logFile) => bytes + Buffer.byteLength(logFile.logText, 'utf8'),
+                  0,
+                );
                 console.log(
-                  `    server_log ${path.basename(logPath)} (${formatBytes(fileSize(logPath))})`,
+                  `    server_logs ${logFiles.length} .log/.out file(s) (${formatBytes(totalBytes)})`,
                 );
-                const serverLog = fs.readFileSync(logPath, 'utf8').replaceAll('\u0000', '');
-                await insertServerLog(sql, insertedIds, serverLog);
-                console.log(`    server_log linked (${elapsed(serverLogStart)})`);
+                await insertServerLogFiles(sql, insertedIds, logFiles);
+                console.log(`    server_logs linked (${elapsed(serverLogStart)})`);
               } catch (error: any) {
-                tracker.recordDbError(`server_log for ${configKey}`, error);
+                tracker.recordDbError(`server_logs for ${configKey}`, error);
               }
             }
           }
@@ -762,13 +819,8 @@ async function main(): Promise<void> {
       if (!mapped) continue;
 
       try {
-        const { outcome } = await ingestEvalRow(
-          sql,
-          getOrCreateConfig,
-          mapped,
-          workflowRunId,
-          date,
-        );
+        const configId = await getOrCreateConfig(mapped.config);
+        const { outcome } = await ingestEvalRow(sql, configId, mapped, workflowRunId, date);
         if (outcome === 'new') totalEvals++;
       } catch (error: any) {
         tracker.recordDbError('eval row', error);
@@ -828,9 +880,10 @@ async function main(): Promise<void> {
 
     for (const params of evalParamsList) {
       try {
+        const configId = await getOrCreateConfig(params.config);
         const { id: evalResultId } = await ingestEvalRow(
           sql,
-          getOrCreateConfig,
+          configId,
           params,
           workflowRunId,
           date,
@@ -857,7 +910,7 @@ async function main(): Promise<void> {
   // ── Ingest changelog (already parsed above for evals-only check) ─────
 
   console.log('\n--- Changelog ---');
-  for (const { baseRef, headRef, entries } of parsedChangelogs) {
+  for (const { baseRef, headRef, entries } of changelogs) {
     try {
       const written = await ingestChangelogEntries(
         sql,
@@ -893,53 +946,20 @@ async function main(): Promise<void> {
   console.log(`  Eval results:      ${totalEvals} new`);
   console.log(`  Eval samples:      ${totalSamples} new across ${totalSampleFiles} file(s)`);
   console.log(`  Changelog entries: ${totalChangelogs} written`);
-  console.log(`\n  DB totals:`);
-  console.log(`    configs           ${configCount.n}`);
-  console.log(`    benchmark_results ${resultCount.n}`);
-  console.log(`    run_stats         ${statsCount.n}`);
-  console.log(`    eval_results      ${evalCount.n}`);
-  console.log(`    eval_samples      ${sampleCount.n}`);
-  console.log(`    changelog_entries ${changelogCount.n}`);
-
-  const { skips, unmappedModels, unmappedHws, unmappedPrecisions } = tracker;
-  const totalSkips =
-    skips.badZip +
-    skips.unmappedModel +
-    skips.unmappedHw +
-    skips.noIslOsl +
-    skips.failedRun +
-    skips.dbError;
-  if (totalSkips > 0) {
-    console.log(`\n  Skipped: ${totalSkips} rows`);
-    const skipLines: [string, number][] = [
-      ['no isl/osl (old format)', skips.noIslOsl],
-      ['failed run (0 successful)', skips.failedRun],
-      ['unmapped model', skips.unmappedModel],
-      ['unmapped hw', skips.unmappedHw],
-      ['bad/empty zip', skips.badZip],
-      ['DB errors', skips.dbError],
-    ].filter(([, n]) => (n as number) > 0) as [string, number][];
-    const pad = Math.max(...skipLines.map(([label]) => label.length));
-    for (const [label, n] of skipLines) {
-      console.log(`    ${label.padEnd(pad)}: ${n}`);
-    }
-  }
-
-  if (unmappedModels.size > 0) {
-    console.log(`\n  Unmapped model values (add to MODEL_TO_KEY to ingest):`);
-    [...unmappedModels].slice(0, 20).forEach((v) => console.log(`    ${v}`));
-    if (unmappedModels.size > 20) console.log(`    ... and ${unmappedModels.size - 20} more`);
-  }
-
-  if (unmappedHws.size > 0) {
-    console.log(`\n  Unmapped hw values (add to hwToGpuKey to ingest):`);
-    [...unmappedHws].slice(0, 20).forEach((v) => console.log(`    ${v}`));
-  }
+  printIngestSummaryFooter(
+    {
+      configs: configCount.n,
+      benchmarkResults: resultCount.n,
+      runStats: statsCount.n,
+      evalResults: evalCount.n,
+      evalSamples: sampleCount.n,
+      changelogEntries: changelogCount.n,
+    },
+    tracker,
+    { includeFailedRuns: true, includeUnmappedPrecisions: true },
+  );
 
-  if (unmappedPrecisions.size > 0) {
-    console.log(`\n  Unmapped precision values (add to PRECISION_KEYS to ingest):`);
-    [...unmappedPrecisions].forEach((v) => console.log(`    ${v}`));
-  }
+  const { unmappedModels, unmappedHws, unmappedPrecisions } = tracker;
 
   // Write unmapped entities to file so CI workflow can send Slack notifications
   const unmappedOutPath = process.env.UNMAPPED_ENTITIES_OUTPUT;
diff --git a/packages/db/src/ingest-gcs-backup.ts b/packages/db/src/ingest-gcs-backup.ts
index ac3e5b93b..4f0aadcd2 100644
--- a/packages/db/src/ingest-gcs-backup.ts
+++ b/packages/db/src/ingest-gcs-backup.ts
@@ -12,7 +12,8 @@
  *   eval_results_all_*     → compiled eval aggregate (agg_eval_all.json, flat rows)
  *   eval_*                 → single eval result (meta_env.json + results_*.json)
  *   changelog-metadata_*   → changelog entries (base_ref, head_ref, entries[])
- *   server_logs_*          → server log text (server.log inside ZIP)
+ *   server_logs_*          → every recursively nested .log/.out file
+ *   multinode_server_logs_* → same, including files inside multinode_server_logs.tar.gz
  *
  * All inserts are idempotent (ON CONFLICT DO UPDATE/NOTHING), so re-running is safe.
  *
@@ -32,8 +33,16 @@ import path from 'path';
 
 import { confirm, hasNoSslFlag, hasYesFlag } from './cli-utils';
 import { createAdminSql, refreshLatestBenchmarks } from './etl/db-utils';
-import { isBenchmarkPointPurged, PURGED_RUNS } from './etl/run-overrides';
+import {
+  applyBenchmarkPointBackfill,
+  applyChangelogBackfills,
+  isBenchmarkPointPurged,
+  PURGED_RUNS,
+  recordBackfilledPointIdentity,
+  validateRunBackfills,
+} from './etl/run-overrides';
 import { createSkipTracker, type Skips } from './etl/skip-tracker';
+import { printIngestSummaryFooter } from './etl/ingest-summary';
 import { GPU_KEYS, parseIslOsl } from './etl/normalizers';
 import { createConfigCache } from './etl/config-cache';
 import { createWorkflowRunServices, type GithubRunInfo } from './etl/workflow-run';
@@ -42,7 +51,7 @@ import {
   bulkIngestBenchmarkRows,
   bulkIngestRunStats,
   bulkUpsertAvailability,
-  insertServerLog,
+  insertServerLogFiles,
 } from './etl/benchmark-ingest';
 import { mapEvalRow, mapAggEvalRow, type EvalParams } from './etl/eval-mapper';
 import { ingestEvalRow } from './etl/eval-ingest';
@@ -51,9 +60,11 @@ import { bulkIngestEvalSamples } from './etl/eval-samples-ingest';
 import {
   parseChangelogEntries,
   ingestChangelogEntries,
+  hasAppendOnlyFlag,
   hasEvalsOnlyFlag,
 } from './etl/changelog-ingest';
-import { readZipJson, readZipJsonMap, readZipText, readZipTextsMatching } from './etl/zip-reader';
+import { readZipJson, readZipJsonAndText, readZipJsonMap } from './etl/zip-reader';
+import { readServerLogArtifactZip, serverLogArtifactSuffix } from './etl/server-log-artifacts';
 
 const GCS_DIR = path.join(import.meta.dirname, '..', '..', '..', 'gcs');
 const CONCURRENCY = 20;
@@ -79,7 +90,11 @@ interface WorkflowMapResult {
   createdAt: string;
   ghInfo: GithubRunInfo | null;
   /** Per-ZIP benchmark rows, ready for configId lookup + bulk insert in phase 2. */
-  bmkZips: { zipFile: string; rows: BenchmarkParams[]; serverLogPath?: string }[];
+  bmkZips: {
+    zipFile: string;
+    rows: BenchmarkParams[];
+    serverLogArtifact?: { zipPath: string; artifactName: string };
+  }[];
   statsRows: { hardware: string; nSuccess: number; total: number }[];
   /**
    * Each eval row carries the matching `samples_<task>_*.jsonl` text when the
@@ -115,15 +130,17 @@ interface WriteResult {
 
 /**
  * Run `fn` over `items` with at most `concurrency` tasks in-flight at once.
- * Result order matches input order. Per-item errors are caught and returned as
- * `null` (with a logged message) so one bad task doesn't abort the whole run.
+ * Result order matches input order. Per-item errors are logged and returned as
+ * `null`, unless failOnError requests a terminal error after in-flight work ends.
  */
 async function pMap<T, R>(
   items: T[],
   fn: (item: T) => Promise<R | null>,
   concurrency: number,
+  failOnError = false,
 ): Promise<(R | null)[]> {
   const results: (R | null)[] = Array.from({ length: items.length }, () => null);
+  const errors: Error[] = [];
   let next = 0;
   async function worker() {
     while (next < items.length) {
@@ -132,10 +149,14 @@ async function pMap<T, R>(
         results[i] = await fn(items[i]);
       } catch (error: any) {
         console.error(`  [ERROR] mapping task ${i} failed: ${error.message}`);
+        errors.push(error instanceof Error ? error : new Error(String(error)));
       }
     }
   }
   await Promise.all(Array.from({ length: Math.min(concurrency, items.length) }, worker));
+  if (failOnError && errors.length > 0) {
+    throw new Error(`${errors.length} restore task(s) failed; first error: ${errors[0].message}`);
+  }
   return results;
 }
 
@@ -207,7 +228,9 @@ async function mapWorkflowDir(
     (f) => f.startsWith('eval_') && !f.startsWith('eval_results_all_'),
   );
   const changelogZips = zipFiles.filter((f) => f.startsWith('changelog-metadata_'));
-  const serverLogZips = zipFiles.filter((f) => f.startsWith('server_logs_'));
+  const serverLogZips = zipFiles.filter(
+    (f) => f.startsWith('server_logs_') || f.startsWith('multinode_server_logs_'),
+  );
 
   if (
     bmkZipFiles.length +
@@ -235,11 +258,16 @@ async function mapWorkflowDir(
   const warnings: string[] = [];
 
   // ── Index server log ZIPs (deferred read — too large for memory) ─────────
-  // Map configKey → zip file path. Actual text is read lazily in phase 2.
-  const serverLogPaths = new Map<string, string>();
+  // Map configKey → ZIP metadata. .log/.out files are extracted lazily in phase 2.
+  const serverLogArtifacts = new Map<string, { zipPath: string; artifactName: string }>();
   for (const zipFile of serverLogZips) {
-    const configKey = zipFile.replace(/^server_logs_/u, '').replace(/_\d+_\d+\.zip$/u, '');
-    serverLogPaths.set(configKey, path.join(artifactsPath, zipFile));
+    const artifactName = zipFile.replace(/_\d+_\d+\.zip$/u, '');
+    const configKey = serverLogArtifactSuffix(artifactName);
+    if (!configKey) continue;
+    serverLogArtifacts.set(configKey, {
+      zipPath: path.join(artifactsPath, zipFile),
+      artifactName,
+    });
   }
 
   // ── Map benchmark ZIPs ────────────────────────────────────────────────────
@@ -300,12 +328,14 @@ async function mapWorkflowDir(
     }
     if (rows.length > 0) {
       // Match server log by config key (bmk_ files only — results_ compiled zips have no matching logs)
-      let serverLogPath: string | undefined;
+      let serverLogArtifact: { zipPath: string; artifactName: string } | undefined;
       if (zipFile.startsWith('bmk_')) {
         const bmkConfigKey = zipFile.replace(/^bmk_/u, '').replace(/_\d+_\d+\.zip$/u, '');
-        serverLogPath = serverLogPaths.get(bmkConfigKey);
+        serverLogArtifact =
+          serverLogArtifacts.get(bmkConfigKey) ??
+          serverLogArtifacts.get(bmkConfigKey.replace(/^agentic_/u, ''));
       }
-      bmkZips.push({ zipFile, rows, serverLogPath });
+      bmkZips.push({ zipFile, rows, serverLogArtifact });
     }
   }
 
@@ -321,7 +351,11 @@ async function mapWorkflowDir(
     for (const [hwKey, stats] of Object.entries(data as Record<string, any>)) {
       if (!GPU_KEYS.has(hwKey)) continue;
       if (typeof stats?.n_success !== 'number' || typeof stats?.total !== 'number') continue;
-      statsRows.push({ hardware: hwKey, nSuccess: stats.n_success, total: stats.total });
+      statsRows.push({
+        hardware: hwKey,
+        nSuccess: stats.n_success,
+        total: stats.total,
+      });
     }
   }
 
@@ -329,12 +363,16 @@ async function mapWorkflowDir(
   const evalRows: WorkflowMapResult['evalRows'] = [];
   for (const zipFile of evalZips) {
     const zipPath = path.join(artifactsPath, zipFile);
-    const files = readZipJsonMap(zipPath);
-    if (!files) {
+    const contents = readZipJsonAndText(
+      zipPath,
+      (name) => name.startsWith('samples_') && name.endsWith('.jsonl'),
+    );
+    if (!contents) {
       local.skips.badZip++;
       warnings.push(`  [WARN] ${dateDir}/${zipFile}: bad/empty zip — skipped`);
       continue;
     }
+    const { jsonFiles: files, textFiles: sampleTexts } = contents;
     const meta = files.get('meta_env.json') as Record<string, any> | undefined;
     const resultsEntry = [...files.entries()].find(
       ([k]) => k.startsWith('results_') && k.endsWith('.json'),
@@ -366,12 +404,6 @@ async function mapWorkflowDir(
       continue;
     }
 
-    // lm-eval names sample files `samples_<task>_<timestamp>.jsonl`. Pull
-    // them all and key by lowercased task to match `EvalParams.task`.
-    const sampleTexts = readZipTextsMatching(
-      zipPath,
-      (n) => n.startsWith('samples_') && n.endsWith('.jsonl'),
-    );
     const samplesByTask = new Map<string, string>();
     for (const [name, text] of sampleTexts) {
       const m = name.match(/^samples_(?<task>.+?)_[^_]+\.jsonl$/u);
@@ -379,18 +411,22 @@ async function mapWorkflowDir(
     }
 
     for (const params of mapped) {
-      evalRows.push({ params, samplesText: samplesByTask.get(params.task) ?? null });
+      evalRows.push({
+        params,
+        samplesText: samplesByTask.get(params.task) ?? null,
+      });
     }
   }
 
   // ── Map compiled eval ZIPs ────────────────────────────────────────────────
   for (const zipFile of evalAggZips) {
-    const files = readZipJsonMap(path.join(artifactsPath, zipFile));
-    if (!files) {
+    const contents = readZipJsonAndText(path.join(artifactsPath, zipFile));
+    if (!contents) {
       local.skips.badZip++;
       warnings.push(`  [WARN] ${dateDir}/${zipFile}: bad/empty zip — skipped`);
       continue;
     }
+    const files = contents.jsonFiles;
     const agg = files.get('agg_eval_all.json');
     if (!Array.isArray(agg)) {
       local.skips.badZip++;
@@ -416,8 +452,17 @@ async function mapWorkflowDir(
   }
 
   // ── Parse changelog ZIPs ──────────────────────────────────────────────────
+  const newestChangelogZip = [...changelogZips]
+    .toSorted((a, b) => {
+      const idA = a.match(/_(?<artifactId>\d+)\.zip$/u)?.[1];
+      const idB = b.match(/_(?<artifactId>\d+)\.zip$/u)?.[1];
+      const tsA = idA ? (artifactCreatedAt.get(Number(idA)) ?? '') : '';
+      const tsB = idB ? (artifactCreatedAt.get(Number(idB)) ?? '') : '';
+      return tsA.localeCompare(tsB) || Number(idA ?? 0) - Number(idB ?? 0);
+    })
+    .at(-1);
   const changelogs: WorkflowMapResult['changelogs'] = [];
-  for (const zipFile of changelogZips) {
+  for (const zipFile of newestChangelogZip ? [newestChangelogZip] : []) {
     const data = readZipJson(path.join(artifactsPath, zipFile)) as Record<string, any> | null;
     if (!data || typeof data !== 'object') {
       local.skips.badZip++;
@@ -431,7 +476,11 @@ async function mapWorkflowDir(
     if (entries.length > 0) changelogs.push({ baseRef, headRef, entries });
   }
 
-  const evalsOnly = hasEvalsOnlyFlag(changelogs);
+  const correctedChangelogs = applyChangelogBackfills(githubRunId, ghInfo?.runAttempt, changelogs);
+  for (const backfillId of correctedChangelogs.backfillIds) {
+    warnings.push(`  [${dateDir}] applied changelog backfill ${backfillId}`);
+  }
+  const evalsOnly = hasEvalsOnlyFlag(correctedChangelogs.changelogs);
   if (evalsOnly) {
     console.log(
       `  [${dateDir}] evals-only run ${githubRunId} — skipping ${bmkZips.length} benchmark ZIP(s) and ${statsRows.length} stats row(s)`,
@@ -450,7 +499,7 @@ async function mapWorkflowDir(
     bmkZips: evalsOnly ? [] : bmkZips,
     statsRows: evalsOnly ? [] : statsRows,
     evalRows,
-    changelogs,
+    changelogs: correctedChangelogs.changelogs,
     evalsOnly,
     localSkips: {
       badZip: local.skips.badZip,
@@ -470,6 +519,7 @@ async function mapWorkflowDir(
 // ── Main ──────────────────────────────────────────────────────────────────────
 
 async function main(): Promise<void> {
+  validateRunBackfills();
   console.log('=== db:ingest:gcs ===');
   console.log(
     'This will ingest all GCS backup ZIP artifacts into the database.\n' +
@@ -577,37 +627,57 @@ async function main(): Promise<void> {
       headBranch: result.headBranch,
       headSha: result.headSha,
       createdAt: result.createdAt,
+      appendOnly: hasAppendOnlyFlag(result.changelogs),
       ghInfo: result.ghInfo,
     });
     if (workflowRunId === null) return wr;
 
     const allInserted: (BenchmarkParams & { configId: number })[] = [];
-    for (const { zipFile, rows, serverLogPath } of result.bmkZips) {
+    const seenPointIdentities = new Map<string, string>();
+    for (const { zipFile, rows, serverLogArtifact } of result.bmkZips) {
       const toInsert: (BenchmarkParams & { configId: number })[] = [];
       for (const row of rows) {
+        let configId: number;
         try {
-          const configId = await getOrCreateConfig(row.config);
-          if (
-            isBenchmarkPointPurged(result.githubRunId, result.ghInfo?.runAttempt, {
-              configId,
-              benchmarkType: row.benchmarkType,
-              isl: row.isl,
-              osl: row.osl,
-              conc: row.conc,
-              offloadMode: row.offloadMode,
-            })
-          ) {
-            console.log(
-              `  [${result.dateDir}] skipped purged benchmark point: config ${configId}, ` +
-                `${row.benchmarkType}, isl ${row.isl}, osl ${row.osl}, conc ${row.conc}, ` +
-                `offload ${row.offloadMode}`,
-            );
-            continue;
-          }
-          toInsert.push({ ...row, configId });
+          configId = await getOrCreateConfig(row.config);
         } catch (error: any) {
           tracker.recordDbError(`config for ${zipFile}`, error);
+          continue;
         }
+        if (
+          isBenchmarkPointPurged(result.githubRunId, result.ghInfo?.runAttempt, {
+            configId,
+            benchmarkType: row.benchmarkType,
+            isl: row.isl,
+            osl: row.osl,
+            conc: row.conc,
+            offloadMode: row.offloadMode,
+            recipeFingerprint: row.recipeFingerprint,
+          })
+        ) {
+          console.log(
+            `  [${result.dateDir}] skipped purged benchmark point: config ${configId}, ` +
+              `${row.benchmarkType}, isl ${row.isl}, osl ${row.osl}, conc ${row.conc}, ` +
+              `offload ${row.offloadMode}, recipe ${row.recipeFingerprint ?? 'legacy'}`,
+          );
+          continue;
+        }
+        const applied = applyBenchmarkPointBackfill(result.githubRunId, result.ghInfo?.runAttempt, {
+          ...row,
+          configId,
+        });
+        recordBackfilledPointIdentity(
+          seenPointIdentities,
+          applied.sourceIdentity,
+          applied.desiredIdentity,
+        );
+        if (applied.backfillId) {
+          console.log(
+            `  [${result.dateDir}] applied benchmark point backfill ` +
+              `${applied.backfillId}: config ${configId}, conc ${row.conc}`,
+          );
+        }
+        toInsert.push(applied.point);
       }
       if (toInsert.length > 0) {
         try {
@@ -623,14 +693,13 @@ async function main(): Promise<void> {
           // Only track as inserted after successful bulk insert
           allInserted.push(...toInsert);
 
-          // Attach server log (read lazily — too large to hold all in memory during phase 1)
-          if (serverLogPath && insertedIds.length > 0) {
-            const serverLog = readZipText(serverLogPath, 'server.log');
-            if (serverLog) {
-              // Strip null bytes — some logs contain 0x00 which PostgreSQL text columns reject
-              const clean = serverLog.replaceAll('\u0000', '');
-              await insertServerLog(sql, insertedIds, clean);
-            }
+          // Attach every .log/.out file (read lazily to bound phase-1 memory).
+          if (serverLogArtifact && insertedIds.length > 0) {
+            const logFiles = readServerLogArtifactZip(
+              serverLogArtifact.zipPath,
+              serverLogArtifact.artifactName,
+            );
+            await insertServerLogFiles(sql, insertedIds, logFiles);
           }
         } catch (error: any) {
           tracker.recordDbError(zipFile, error);
@@ -688,9 +757,10 @@ async function main(): Promise<void> {
 
     for (const { params, samplesText } of result.evalRows) {
       try {
+        const configId = await getOrCreateConfig(params.config);
         const { outcome, id: evalResultId } = await ingestEvalRow(
           sql,
-          getOrCreateConfig,
+          configId,
           params,
           workflowRunId,
           result.dateDir,
@@ -740,6 +810,7 @@ async function main(): Promise<void> {
       return out;
     },
     DB_CONCURRENCY,
+    true,
   );
 
   // Accumulate totals per date, then print one line per date in sorted order.
@@ -846,42 +917,17 @@ async function main(): Promise<void> {
   console.log(`  Eval results:      ${totalEvals} new`);
   console.log(`  Eval samples:      ${totalEvalSamples} new`);
   console.log(`  Changelog entries: ${totalChangelogs} written`);
-  console.log(`\n  DB totals:`);
-  console.log(`    configs           ${configCount.n}`);
-  console.log(`    benchmark_results ${resultCount.n}`);
-  console.log(`    run_stats         ${statsCount.n}`);
-  console.log(`    eval_results      ${evalCount.n}`);
-  console.log(`    eval_samples      ${sampleCount.n}`);
-  console.log(`    changelog_entries ${changelogCount.n}`);
-
-  const { skips, unmappedModels, unmappedHws } = tracker;
-  const totalSkips =
-    skips.badZip + skips.unmappedModel + skips.unmappedHw + skips.noIslOsl + skips.dbError;
-  if (totalSkips > 0) {
-    console.log(`\n  Skipped: ${totalSkips} rows`);
-    const skipLines: [string, number][] = [
-      ['no isl/osl (old format)', skips.noIslOsl],
-      ['unmapped model', skips.unmappedModel],
-      ['unmapped hw', skips.unmappedHw],
-      ['bad/empty zip', skips.badZip],
-      ['DB errors', skips.dbError],
-    ].filter(([, n]) => (n as number) > 0) as [string, number][];
-    const pad = Math.max(...skipLines.map(([label]) => label.length));
-    for (const [label, n] of skipLines) {
-      console.log(`    ${label.padEnd(pad)}: ${n}`);
-    }
-  }
-
-  if (unmappedModels.size > 0) {
-    console.log(`\n  Unmapped model values (add to MODEL_TO_KEY to ingest):`);
-    [...unmappedModels].slice(0, 20).forEach((v) => console.log(`    ${v}`));
-    if (unmappedModels.size > 20) console.log(`    ... and ${unmappedModels.size - 20} more`);
-  }
-
-  if (unmappedHws.size > 0) {
-    console.log(`\n  Unmapped hw values (add to hwToGpuKey to ingest):`);
-    [...unmappedHws].slice(0, 20).forEach((v) => console.log(`    ${v}`));
-  }
+  printIngestSummaryFooter(
+    {
+      configs: configCount.n,
+      benchmarkResults: resultCount.n,
+      runStats: statsCount.n,
+      evalResults: evalCount.n,
+      evalSamples: sampleCount.n,
+      changelogEntries: changelogCount.n,
+    },
+    tracker,
+  );
 
   console.log('\n=== db:ingest:gcs complete ===');
   console.log('  Invalidate API cache: bun run admin:cache:invalidate');
diff --git a/packages/db/src/ingest-supplemental.ts b/packages/db/src/ingest-supplemental.ts
index a2d762174..4483d4397 100644
--- a/packages/db/src/ingest-supplemental.ts
+++ b/packages/db/src/ingest-supplemental.ts
@@ -22,7 +22,11 @@ import {
   normalizeSpecMethod,
   parseBool,
 } from './etl/normalizers';
-import { bulkIngestBenchmarkRows, bulkUpsertAvailability } from './etl/benchmark-ingest';
+import {
+  bulkIngestBenchmarkRows,
+  bulkUpsertAvailability,
+  type BenchmarkPersistenceInput,
+} from './etl/benchmark-ingest';
 import { ingestEvalRow } from './etl/eval-ingest';
 
 const sql = createAdminSql({
@@ -125,9 +129,8 @@ async function ingestSupplementalEvals(
 
       const { outcome } = await ingestEvalRow(
         sql,
-        () => Promise.resolve(configId),
+        configId,
         {
-          config: {} as any,
           task: entry.task,
           isl: 1024,
           osl: 8192,
@@ -217,16 +220,7 @@ async function ingestSupplementalBmk(
       conclusion: 'success',
     }))!;
 
-    const rows: {
-      configId: number;
-      benchmarkType: 'single_turn' | 'agentic_traces';
-      offloadMode: string;
-      isl: number | null;
-      osl: number | null;
-      conc: number;
-      image: string | null;
-      metrics: Record<string, number>;
-    }[] = [];
+    const rows: BenchmarkPersistenceInput[] = [];
 
     for (const entry of entries) {
       const modelKey = resolveModelKey({ model: entry.model, infmax_model_prefix: undefined });
@@ -279,16 +273,12 @@ async function ingestSupplementalBmk(
         osl: entry.osl,
         conc: entry.conc,
         image: entry.image,
+        recipeFingerprint: null,
         metrics: entry.metrics,
       });
     }
 
-    const { newCount, dupCount } = await bulkIngestBenchmarkRows(
-      sql,
-      rows.map((r) => ({ ...r, config: {} as any })),
-      workflowRunId,
-      date,
-    );
+    const { newCount, dupCount } = await bulkIngestBenchmarkRows(sql, rows, workflowRunId, date);
     totalNew += newCount;
     totalDup += dupCount;
 
diff --git a/packages/db/src/lib/artifact-retry.test.ts b/packages/db/src/lib/artifact-retry.test.ts
new file mode 100644
index 000000000..675355b59
--- /dev/null
+++ b/packages/db/src/lib/artifact-retry.test.ts
@@ -0,0 +1,42 @@
+import { describe, expect, it, vi } from 'vitest';
+
+import { retryArtifactOperation } from './artifact-retry.js';
+
+describe('retryArtifactOperation', () => {
+  it('retries transient failures using the configured delays', async () => {
+    const operation = vi
+      .fn<() => string>()
+      .mockImplementationOnce(() => {
+        throw new Error('temporary DNS failure');
+      })
+      .mockImplementationOnce(() => {
+        throw new Error('temporary TLS failure');
+      })
+      .mockReturnValue('downloaded');
+    const wait = vi.fn<(delayMs: number) => Promise<void>>(() => Promise.resolve());
+    const warn = vi.fn<(message: string) => void>();
+
+    await expect(
+      retryArtifactOperation('artifact', operation, { delaysMs: [5, 15], wait, warn }),
+    ).resolves.toBe('downloaded');
+    expect(operation).toHaveBeenCalledTimes(3);
+    expect(wait.mock.calls).toEqual([[5], [15]]);
+    expect(warn).toHaveBeenCalledTimes(2);
+  });
+
+  it('throws the final error after exhausting retries', async () => {
+    const failure = new Error('still unavailable');
+    const wait = vi.fn<(delayMs: number) => Promise<void>>(() => Promise.resolve());
+
+    await expect(
+      retryArtifactOperation(
+        'artifact',
+        () => {
+          throw failure;
+        },
+        { delaysMs: [1], wait, warn: () => {} },
+      ),
+    ).rejects.toBe(failure);
+    expect(wait).toHaveBeenCalledOnce();
+  });
+});
diff --git a/packages/db/src/lib/artifact-retry.ts b/packages/db/src/lib/artifact-retry.ts
new file mode 100644
index 000000000..e145aa04d
--- /dev/null
+++ b/packages/db/src/lib/artifact-retry.ts
@@ -0,0 +1,39 @@
+const DEFAULT_RETRY_DELAYS_MS = [5_000, 15_000, 30_000, 60_000, 120_000] as const;
+
+interface ArtifactRetryOptions {
+  delaysMs?: readonly number[];
+  wait?: (delayMs: number) => Promise<void>;
+  warn?: (message: string) => void;
+}
+
+const defaultWait = (delayMs: number) =>
+  new Promise<void>((resolve) => {
+    setTimeout(resolve, delayMs);
+  });
+
+/** Retry transient artifact-store operations with bounded exponential backoff. */
+export async function retryArtifactOperation<T>(
+  label: string,
+  operation: () => T | Promise<T>,
+  options: ArtifactRetryOptions = {},
+): Promise<T> {
+  const delaysMs = options.delaysMs ?? DEFAULT_RETRY_DELAYS_MS;
+  const wait = options.wait ?? defaultWait;
+  const warn = options.warn ?? console.warn;
+  let lastError: unknown;
+
+  for (let attempt = 0; attempt < delaysMs.length + 1; attempt++) {
+    try {
+      return await operation();
+    } catch (error) {
+      lastError = error;
+      if (attempt >= delaysMs.length) break;
+      const delayMs = delaysMs[attempt];
+      if (delayMs === undefined) break;
+      warn(`  [WARN] ${label} failed (attempt ${attempt + 1}); retrying in ${delayMs / 1_000}s`);
+      await wait(delayMs);
+    }
+  }
+
+  throw lastError;
+}
diff --git a/packages/db/src/lib/backfill-runner.test.ts b/packages/db/src/lib/backfill-runner.test.ts
index 6da9071fe..cf6b40f9c 100644
--- a/packages/db/src/lib/backfill-runner.test.ts
+++ b/packages/db/src/lib/backfill-runner.test.ts
@@ -1,6 +1,10 @@
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
 
-import { parseLimitForceFlags, runPerIdBackfill } from './backfill-runner.js';
+import {
+  parseLimitForceFlags,
+  runCandidateIdBackfill,
+  runPerIdBackfill,
+} from './backfill-runner.js';
 
 describe('parseLimitForceFlags', () => {
   const originalArgv = process.argv;
@@ -10,12 +14,33 @@ describe('parseLimitForceFlags', () => {
 
   it('defaults to no limit and force off', () => {
     process.argv = ['node', 'script.ts'];
-    expect(parseLimitForceFlags()).toEqual({ limit: null, force: false });
+    expect(parseLimitForceFlags()).toEqual({
+      limit: null,
+      force: false,
+      shardCount: 1,
+      shardIndex: 0,
+    });
   });
 
-  it('parses --limit N and --force', () => {
-    process.argv = ['node', 'script.ts', '--limit', '25', '--force', '--yes'];
-    expect(parseLimitForceFlags()).toEqual({ limit: 25, force: true });
+  it('parses --limit N, --force, and process-shard flags', () => {
+    process.argv = [
+      'node',
+      'script.ts',
+      '--limit',
+      '25',
+      '--force',
+      '--shard-count',
+      '4',
+      '--shard-index',
+      '2',
+      '--yes',
+    ];
+    expect(parseLimitForceFlags()).toEqual({
+      limit: 25,
+      force: true,
+      shardCount: 4,
+      shardIndex: 2,
+    });
   });
 });
 
@@ -53,3 +78,45 @@ describe('runPerIdBackfill', () => {
     expect(vi.mocked(console.error).mock.calls[0]?.[0]).toContain('✗ id=1: boom');
   });
 });
+
+describe('runCandidateIdBackfill', () => {
+  const originalArgv = process.argv;
+
+  beforeEach(() => {
+    process.argv = ['node', 'script.ts', '--yes'];
+    vi.spyOn(console, 'log').mockImplementation(() => {});
+  });
+
+  afterEach(() => {
+    process.argv = originalArgv;
+    vi.restoreAllMocks();
+    process.exitCode = undefined;
+  });
+
+  it('stops after candidate loading when there is nothing to do', async () => {
+    const processRow = vi.fn<(_: number) => Promise<'ok'>>(() => Promise.resolve('ok'));
+
+    const processed = await runCandidateIdBackfill(() => Promise.resolve([]), processRow);
+
+    expect(processRow).not.toHaveBeenCalled();
+    expect(processed).toBe(false);
+    expect(vi.mocked(console.log).mock.calls[0]?.[0]).toContain('Nothing to do');
+  });
+
+  it('confirms and serially processes loaded candidate ids', async () => {
+    const seen: number[] = [];
+
+    const processed = await runCandidateIdBackfill(
+      () => Promise.resolve([4, 9]),
+      (id) => {
+        seen.push(id);
+        return Promise.resolve('ok');
+      },
+      (count) => `${count} custom candidates.`,
+    );
+
+    expect(seen).toEqual([4, 9]);
+    expect(processed).toBe(true);
+    expect(vi.mocked(console.log).mock.calls[0]?.[0]).toContain('2 custom candidates');
+  });
+});
diff --git a/packages/db/src/lib/backfill-runner.ts b/packages/db/src/lib/backfill-runner.ts
index de00bee11..5fde13ede 100644
--- a/packages/db/src/lib/backfill-runner.ts
+++ b/packages/db/src/lib/backfill-runner.ts
@@ -12,12 +12,16 @@ import type { Sql } from '../etl/db-utils.js';
 export interface LimitForceFlags {
   limit: number | null;
   force: boolean;
+  shardCount: number;
+  shardIndex: number;
 }
 
 /** Parse the standard `--limit N` / `--force` backfill flags from argv. */
 export function parseLimitForceFlags(): LimitForceFlags {
   let limit: number | null = null;
   let force = false;
+  let shardCount = 1;
+  let shardIndex = 0;
   for (let i = 2; i < process.argv.length; i++) {
     const arg = process.argv[i]!;
     if (arg === '--force') force = true;
@@ -28,9 +32,29 @@ export function parseLimitForceFlags(): LimitForceFlags {
         process.exit(1);
       }
       limit = Number(next);
+    } else if (arg === '--shard-count') {
+      const next = process.argv[++i];
+      const parsed = Number(next);
+      if (!next || !Number.isInteger(parsed) || parsed < 1) {
+        console.error('--shard-count requires a positive integer');
+        process.exit(1);
+      }
+      shardCount = parsed;
+    } else if (arg === '--shard-index') {
+      const next = process.argv[++i];
+      const parsed = Number(next);
+      if (!next || !Number.isInteger(parsed) || parsed < 0) {
+        console.error('--shard-index requires a non-negative integer');
+        process.exit(1);
+      }
+      shardIndex = parsed;
     }
   }
-  return { limit, force };
+  if (shardIndex >= shardCount) {
+    console.error('--shard-index must be less than --shard-count');
+    process.exit(1);
+  }
+  return { limit, force, shardCount, shardIndex };
 }
 
 /**
@@ -78,6 +102,27 @@ export async function runPerIdBackfill(
   if (failed > 0) process.exitCode = 1;
 }
 
+/**
+ * Load candidate ids, handle the standard empty/confirmation gates, then
+ * process the ids serially. The callbacks deliberately stop at orchestration:
+ * each script still owns its candidate, fetch, and update SQL. Returns true
+ * only when the candidates passed the confirmation gate and were processed.
+ */
+export async function runCandidateIdBackfill(
+  loadCandidateIds: () => Promise<readonly number[]>,
+  processRow: (id: number) => Promise<'ok' | 'skipped'>,
+  formatCandidates: (count: number) => string = (count) => `${count} candidate row(s).`,
+): Promise<boolean> {
+  const ids = await loadCandidateIds();
+  if (ids.length === 0) {
+    console.log('\n  Nothing to do — all rows up to date.');
+    return false;
+  }
+  if (!(await confirmProceed(formatCandidates(ids.length)))) return false;
+  await runPerIdBackfill(ids, processRow);
+  return true;
+}
+
 /**
  * jsonb parameter for a freshly computed value. `structuredClone` strips
  * class instances/prototypes so postgres.js serializes plain data only —
diff --git a/packages/db/src/lib/gcs-artifacts.test.ts b/packages/db/src/lib/gcs-artifacts.test.ts
new file mode 100644
index 000000000..3ca3bf6e2
--- /dev/null
+++ b/packages/db/src/lib/gcs-artifacts.test.ts
@@ -0,0 +1,85 @@
+import { describe, expect, it, vi } from 'vitest';
+
+import {
+  listGcsServerLogArtifacts,
+  parseGcsArtifactObject,
+  publicGcsObjectUrl,
+} from './gcs-artifacts.js';
+
+describe('GCS artifact backup', () => {
+  it('parses benchmark and server-log ZIPs while preserving the runner suffix', () => {
+    expect(
+      parseGcsArtifactObject({
+        name:
+          '2026-03-04/Run_Sweep_-_main_12345678901/artifacts/' +
+          'server_logs_cfg_h200-cw_7_9876543210.zip',
+        size: '4096',
+        updated: '2026-03-05T00:00:00Z',
+      }),
+    ).toMatchObject({
+      runId: 12345678901,
+      artifact: {
+        id: 9876543210,
+        name: 'server_logs_cfg_h200-cw_7',
+        size: 4096,
+      },
+    });
+  });
+
+  it('ignores unrelated artifacts and malformed object paths', () => {
+    expect(
+      parseGcsArtifactObject({
+        name: '2026-03-04/Run_12345678901/artifacts/eval_cfg_9876543210.zip',
+      }),
+    ).toBeNull();
+    expect(parseGcsArtifactObject({ name: 'not-an-artifact.zip' })).toBeNull();
+  });
+
+  it('encodes public object path segments without escaping separators', () => {
+    expect(publicGcsObjectUrl('2026-01-01/a folder/artifacts/a+b.zip')).toBe(
+      'https://storage.googleapis.com/inferencemax-gha-backup/' +
+        '2026-01-01/a%20folder/artifacts/a%2Bb.zip',
+    );
+  });
+
+  it('paginates the public bucket index and groups useful artifacts by run', async () => {
+    const fetchMock = vi
+      .fn<typeof fetch>()
+      .mockResolvedValueOnce(
+        Response.json({
+          items: [
+            {
+              name: '2026-01-01/Sweep_12345678901/artifacts/bmk_cfg_pool_0_1000000001.zip',
+              size: '20',
+              updated: '2026-01-02T00:00:00Z',
+            },
+          ],
+          nextPageToken: 'next',
+        }),
+      )
+      .mockResolvedValueOnce(
+        Response.json({
+          items: [
+            {
+              name: '2026-01-01/Sweep_12345678901/artifacts/server_logs_cfg_pool_0_1000000002.zip',
+              size: '30',
+              updated: '2026-01-02T00:00:01Z',
+            },
+            {
+              name: '2026-01-01/Sweep_12345678901/artifacts/eval_cfg_1000000003.zip',
+              size: '40',
+            },
+          ],
+        }),
+      );
+
+    const byRun = await listGcsServerLogArtifacts(fetchMock);
+
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+    expect(String(fetchMock.mock.calls[1]?.[0])).toContain('pageToken=next');
+    expect(byRun.get(12345678901)?.map((artifact) => artifact.name)).toEqual([
+      'bmk_cfg_pool_0',
+      'server_logs_cfg_pool_0',
+    ]);
+  });
+});
diff --git a/packages/db/src/lib/gcs-artifacts.ts b/packages/db/src/lib/gcs-artifacts.ts
new file mode 100644
index 000000000..f43afadc7
--- /dev/null
+++ b/packages/db/src/lib/gcs-artifacts.ts
@@ -0,0 +1,138 @@
+/** Public GCS backup discovery and download helpers. */
+
+import { execFileSync } from 'node:child_process';
+import fs from 'node:fs';
+import path from 'node:path';
+
+import { GCS_BUCKET_BASE } from '@semianalysisai/inferencex-constants';
+
+import type { ArtifactMeta } from './github-artifacts.js';
+
+const GCS_BUCKET = 'inferencemax-gha-backup';
+const GCS_LIST_URL = `https://storage.googleapis.com/storage/v1/b/${GCS_BUCKET}/o`;
+const ARTIFACT_OBJECT_RE =
+  /^(?<date>\d{4}-\d{2}-\d{2})\/(?<workflow>.+)_(?<runId>\d{10,})\/artifacts\/(?<artifactName>.+)_(?<artifactId>\d+)\.zip$/u;
+
+export interface GcsArtifactMeta extends ArtifactMeta {
+  objectName: string;
+  size: number;
+}
+
+interface GcsObject {
+  name?: unknown;
+  size?: unknown;
+  updated?: unknown;
+}
+
+interface GcsListResponse {
+  items?: GcsObject[];
+  nextPageToken?: string;
+}
+
+export function publicGcsObjectUrl(objectName: string): string {
+  return `${GCS_BUCKET_BASE}/${objectName.split('/').map(encodeURIComponent).join('/')}`;
+}
+
+/** Parse one backed-up GitHub artifact ZIP path into its original artifact metadata. */
+export function parseGcsArtifactObject(object: GcsObject): {
+  runId: number;
+  artifact: GcsArtifactMeta;
+} | null {
+  if (typeof object.name !== 'string') return null;
+  const match = ARTIFACT_OBJECT_RE.exec(object.name);
+  if (!match?.groups) return null;
+
+  const artifactName = match.groups.artifactName!;
+  if (
+    !artifactName.startsWith('bmk_') &&
+    !artifactName.startsWith('server_logs_') &&
+    !artifactName.startsWith('multinode_server_logs_')
+  ) {
+    return null;
+  }
+
+  const artifactId = Number(match.groups.artifactId);
+  const runId = Number(match.groups.runId);
+  const size = Number(object.size ?? 0);
+  if (!Number.isSafeInteger(artifactId) || !Number.isSafeInteger(runId)) return null;
+
+  return {
+    runId,
+    artifact: {
+      id: artifactId,
+      name: artifactName,
+      // Backup upload timestamps do not preserve GitHub creation order. Leave
+      // this empty so pairing uses the monotonically increasing artifact id.
+      created_at: '',
+      archive_download_url: publicGcsObjectUrl(object.name),
+      expired: false,
+      objectName: object.name,
+      size: Number.isFinite(size) && size >= 0 ? size : 0,
+    },
+  };
+}
+
+/**
+ * Build a run-id index using GCS's public JSON API. Only benchmark and
+ * server-log ZIPs are retained in memory; the other backup objects are ignored.
+ */
+export async function listGcsServerLogArtifacts(
+  fetchImpl: typeof fetch = fetch,
+): Promise<Map<number, GcsArtifactMeta[]>> {
+  const byRun = new Map<number, GcsArtifactMeta[]>();
+  let pageToken: string | undefined;
+
+  do {
+    const url = new URL(GCS_LIST_URL);
+    url.searchParams.set('maxResults', '1000');
+    url.searchParams.set('fields', 'items(name,size,updated),nextPageToken');
+    if (pageToken) url.searchParams.set('pageToken', pageToken);
+
+    const response = await fetchImpl(url);
+    if (!response.ok) {
+      throw new Error(`GCS artifact listing failed (${response.status} ${response.statusText})`);
+    }
+    const page = (await response.json()) as GcsListResponse;
+    for (const object of page.items ?? []) {
+      const parsed = parseGcsArtifactObject(object);
+      if (!parsed) continue;
+      const artifacts = byRun.get(parsed.runId) ?? [];
+      artifacts.push(parsed.artifact);
+      byRun.set(parsed.runId, artifacts);
+    }
+    pageToken = page.nextPageToken;
+  } while (pageToken);
+
+  return byRun;
+}
+
+/** Download and extract one public GCS artifact into a temporary run directory. */
+export function downloadGcsArtifact(artifact: GcsArtifactMeta, destRoot: string): string {
+  const safeName = path.basename(artifact.name);
+  if (safeName !== artifact.name || safeName === '.' || safeName === '..') {
+    throw new Error(`Unsafe GCS artifact name: ${artifact.name}`);
+  }
+
+  const zipPath = path.join(destRoot, `${artifact.id ?? 'artifact'}.zip`);
+  const destDir = path.join(destRoot, safeName);
+  execFileSync(
+    'curl',
+    [
+      '-fsSL',
+      '--retry',
+      '3',
+      '--retry-all-errors',
+      '--output',
+      zipPath,
+      artifact.archive_download_url,
+    ],
+    { stdio: ['ignore', 'ignore', 'inherit'] },
+  );
+  try {
+    fs.mkdirSync(destDir, { recursive: true });
+    execFileSync('unzip', ['-oq', zipPath, '-d', destDir], { stdio: 'ignore' });
+    return destDir;
+  } finally {
+    fs.rmSync(zipPath, { force: true });
+  }
+}
diff --git a/packages/db/src/lib/server-log-backfill.test.ts b/packages/db/src/lib/server-log-backfill.test.ts
new file mode 100644
index 000000000..fd875a82f
--- /dev/null
+++ b/packages/db/src/lib/server-log-backfill.test.ts
@@ -0,0 +1,105 @@
+import { describe, expect, it } from 'vitest';
+
+import type { ArtifactMeta } from './github-artifacts.js';
+import { pairServerLogArtifacts, resolveServerLogResultCandidates } from './server-log-backfill.js';
+
+const artifact = (name: string, created_at = '2026-08-12T00:00:00Z'): ArtifactMeta => ({
+  name,
+  created_at,
+  archive_download_url: `https://example.test/${name}`,
+});
+
+describe('pairServerLogArtifacts', () => {
+  it('pairs single-node logs with agentic benchmark artifacts', () => {
+    const pairs = pairServerLogArtifacts([
+      artifact('server_logs_config-a_pool_00'),
+      artifact('bmk_agentic_config-a_pool_00'),
+    ]);
+    expect(pairs).toHaveLength(1);
+    expect(pairs[0]?.benchmarks.name).toBe('bmk_agentic_config-a_pool_00');
+  });
+
+  it('pairs multinode logs without hard-coding their internal filenames', () => {
+    const pairs = pairServerLogArtifacts([
+      artifact('multinode_server_logs_config-b_pool_03'),
+      artifact('bmk_agentic_config-b_pool_03'),
+    ]);
+    expect(pairs).toHaveLength(1);
+    expect(pairs[0]?.serverLogs.name).toContain('multinode_server_logs_');
+  });
+
+  it('ignores eval-only server-log artifacts without a benchmark sibling', () => {
+    expect(
+      pairServerLogArtifacts([
+        artifact('multinode_server_logs_eval-only_pool_00'),
+        artifact('eval_eval-only_pool_00'),
+      ]),
+    ).toEqual([]);
+  });
+
+  it('does not collapse an eval runner suffix onto a benchmark suffix', () => {
+    const pairs = pairServerLogArtifacts([
+      artifact('multinode_server_logs_shared_pool_00', '2026-08-12T00:00:00Z'),
+      artifact('bmk_agentic_shared_pool_00', '2026-08-12T00:00:00Z'),
+      artifact('multinode_server_logs_shared_pool_01', '2026-08-12T01:00:00Z'),
+      artifact('eval_shared_pool_01', '2026-08-12T01:00:00Z'),
+    ]);
+    expect(pairs).toHaveLength(1);
+    expect(pairs[0]?.serverLogs.name).toBe('multinode_server_logs_shared_pool_00');
+  });
+
+  it('keeps only the newest exact benchmark/log pair across runner retries', () => {
+    const pairs = pairServerLogArtifacts([
+      artifact('server_logs_retry_pool_00', '2026-08-12T00:00:00Z'),
+      artifact('bmk_agentic_retry_pool_00', '2026-08-12T00:01:00Z'),
+      artifact('server_logs_retry_pool_01', '2026-08-12T01:00:00Z'),
+      artifact('bmk_agentic_retry_pool_01', '2026-08-12T01:01:00Z'),
+    ]);
+    expect(pairs).toHaveLength(1);
+    expect(pairs[0]?.serverLogs.name).toBe('server_logs_retry_pool_01');
+    expect(pairs[0]?.benchmarks.name).toBe('bmk_agentic_retry_pool_01');
+  });
+
+  it('uses artifact ids to break equal GCS upload timestamps', () => {
+    const pairs = pairServerLogArtifacts([
+      { ...artifact('server_logs_retry_pool_00'), id: 100 },
+      { ...artifact('server_logs_retry_pool_00'), id: 200 },
+      { ...artifact('bmk_retry_pool_00'), id: 300 },
+    ]);
+
+    expect(pairs[0]?.serverLogs.id).toBe(200);
+  });
+});
+
+describe('resolveServerLogResultCandidates', () => {
+  it('prefers exact offload-mode matches', () => {
+    expect(
+      resolveServerLogResultCandidates(
+        [
+          { id: 1, offloadMode: 'off' },
+          { id: 2, offloadMode: 'on' },
+        ],
+        'on',
+      ),
+    ).toEqual({ ids: [2], usedUniqueFallback: false });
+  });
+
+  it('accepts historical offload-label drift for one unique candidate', () => {
+    expect(resolveServerLogResultCandidates([{ id: 439874, offloadMode: 'on' }], 'off')).toEqual({
+      ids: [439874],
+      usedUniqueFallback: true,
+    });
+  });
+
+  it('rejects an ambiguous fallback', () => {
+    expect(
+      resolveServerLogResultCandidates(
+        [
+          { id: 1, offloadMode: 'on' },
+          { id: 2, offloadMode: 'auto' },
+        ],
+        'off',
+      ),
+    ).toEqual({ ids: [], usedUniqueFallback: false });
+  });
+});
diff --git a/packages/db/src/lib/server-log-backfill.ts b/packages/db/src/lib/server-log-backfill.ts
new file mode 100644
index 000000000..483fdb42c
--- /dev/null
+++ b/packages/db/src/lib/server-log-backfill.ts
@@ -0,0 +1,75 @@
+import { RUNNER_SUFFIX_RE, type ArtifactMeta } from './github-artifacts.js';
+import { serverLogArtifactSuffix } from '../etl/server-log-artifacts.js';
+
+export interface ServerLogArtifactPair {
+  serverLogs: ArtifactMeta;
+  benchmarks: ArtifactMeta;
+}
+
+export interface ServerLogResultCandidate {
+  id: number;
+  offloadMode: string;
+}
+
+export interface ServerLogResultResolution {
+  ids: number[];
+  usedUniqueFallback: boolean;
+}
+
+function isNewerArtifact(candidate: ArtifactMeta, existing: ArtifactMeta): boolean {
+  return (
+    candidate.created_at > existing.created_at ||
+    (candidate.created_at === existing.created_at && (candidate.id ?? 0) > (existing.id ?? 0))
+  );
+}
+
+/**
+ * Prefer the persisted point whose offload mode matches the current mapper.
+ * Historical mapper changes left a small number of otherwise-identical points
+ * with a different offload label; accept that drift only when the candidate is
+ * unique, since attaching a log to an ambiguous point would be worse than
+ * leaving it unlinked.
+ */
+export function resolveServerLogResultCandidates(
+  candidates: readonly ServerLogResultCandidate[],
+  offloadMode: string,
+): ServerLogResultResolution {
+  const exact = candidates.filter((candidate) => candidate.offloadMode === offloadMode);
+  if (exact.length > 0) {
+    return { ids: exact.map((candidate) => candidate.id), usedUniqueFallback: false };
+  }
+  return candidates.length === 1
+    ? { ids: [candidates[0]!.id], usedUniqueFallback: true }
+    : { ids: [], usedUniqueFallback: false };
+}
+
+/** Pair server_logs_/multinode_server_logs_ artifacts with their raw bmk sibling. */
+export function pairServerLogArtifacts(
+  artifacts: readonly ArtifactMeta[],
+): ServerLogArtifactPair[] {
+  const byName = new Map<string, ArtifactMeta>();
+  for (const artifact of artifacts) {
+    const existing = byName.get(artifact.name);
+    if (!existing || isNewerArtifact(artifact, existing)) byName.set(artifact.name, artifact);
+  }
+  const pairsByLogicalBenchmark = new Map<string, ServerLogArtifactPair>();
+
+  for (const serverLogs of byName.values()) {
+    const suffix = serverLogArtifactSuffix(serverLogs.name);
+    if (!suffix) continue;
+    // Require the exact runner suffix. Eval and benchmark jobs can share the
+    // same logical config while uploading distinct server-log artifacts.
+    const benchmarks = byName.get(`bmk_agentic_${suffix}`) ?? byName.get(`bmk_${suffix}`);
+    if (!benchmarks) continue;
+
+    const logicalName = benchmarks.name.replace(RUNNER_SUFFIX_RE, '');
+    const existing = pairsByLogicalBenchmark.get(logicalName);
+    if (!existing || isNewerArtifact(benchmarks, existing.benchmarks)) {
+      pairsByLogicalBenchmark.set(logicalName, { serverLogs, benchmarks });
+    }
+  }
+
+  return [...pairsByLogicalBenchmark.values()].toSorted((a, b) =>
+    a.serverLogs.name.localeCompare(b.serverLogs.name),
+  );
+}
diff --git a/packages/db/src/queries/agentic-aggregates.test.ts b/packages/db/src/queries/agentic-aggregates.test.ts
index 566ca65f7..89ec4b914 100644
--- a/packages/db/src/queries/agentic-aggregates.test.ts
+++ b/packages/db/src/queries/agentic-aggregates.test.ts
@@ -29,6 +29,7 @@ describe('percentilesOf', () => {
     // For 100 sorted values, p75 = sorted[0.75 * 99] = sorted[74.25] interp.
     expect(p!.p75).toBeCloseTo(75.25, 6);
     expect(p!.p90).toBeCloseTo(90.1, 6);
+    expect(p!.p95).toBeCloseTo(95.05, 6);
     expect(p!.p99).toBeCloseTo(99.01, 6);
   });
 
@@ -49,6 +50,14 @@ describe('extractIslOsl', () => {
           output_sequence_length: { value: 50, unit: 'tokens' },
         },
       }),
+      // cancelled profiling record — it did not complete and must be ignored
+      JSON.stringify({
+        metadata: { benchmark_phase: 'profiling', was_cancelled: true },
+        metrics: {
+          input_sequence_length: { value: 7777, unit: 'tokens' },
+          output_sequence_length: { value: 7777, unit: 'tokens' },
+        },
+      }),
       JSON.stringify({
         metadata: { benchmark_phase: 'profiling' },
         metrics: {
diff --git a/packages/db/src/queries/agentic-aggregates.ts b/packages/db/src/queries/agentic-aggregates.ts
index bfd4b519a..9511d2484 100644
--- a/packages/db/src/queries/agentic-aggregates.ts
+++ b/packages/db/src/queries/agentic-aggregates.ts
@@ -29,9 +29,11 @@ import {
   extractIslOsl,
   fetchAggregateStatsRows,
   percentilesOf,
+  sequenceLengthSketches,
   STATS_VERSION,
   writeBackTraceReplayJsonb,
   type MetricPercentiles,
+  type SequenceLengthSketches,
 } from './agentic-shared';
 
 // STATS_VERSION, the profile extractor `extractIslOsl`, and the percentile
@@ -41,8 +43,10 @@ import {
 export {
   extractIslOsl,
   percentilesOf,
+  sequenceLengthSketches,
   STATS_VERSION,
   type MetricPercentiles,
+  type SequenceLengthSketches,
 } from './agentic-shared';
 
 export interface AgenticAggregate {
@@ -329,6 +333,7 @@ export async function getAgenticAggregates(
               kvCacheUtil: null,
               prefixCacheHitRate: null,
               e2elPerOsl: derived.e2el_per_osl,
+              sequenceLengths: sequenceLengthSketches(isl, osl),
             },
           });
         } catch {
@@ -414,6 +419,7 @@ interface FullAggregateStats {
   kvCacheUtil: MetricPercentiles | null;
   prefixCacheHitRate: MetricPercentiles | null;
   e2elPerOsl: MetricPercentiles | null;
+  sequenceLengths: SequenceLengthSketches;
 }
 
 function blankAggregate(id: number): AgenticAggregate {
diff --git a/packages/db/src/queries/agentic-catalog.ts b/packages/db/src/queries/agentic-catalog.ts
new file mode 100644
index 000000000..b1050583e
--- /dev/null
+++ b/packages/db/src/queries/agentic-catalog.ts
@@ -0,0 +1,91 @@
+/**
+ * Catalog of AgentX points that have a stored telemetry blob.
+ *
+ * `/inference/agentic/<id>` only renders something useful when the row has an
+ * `agentic_trace_replay` blob behind it, so this is the authoritative list of
+ * detail pages that exist. There are hundreds of such rows — far too many to
+ * list one card each — so the catalog collapses them to one entry per
+ * (model, hardware, framework, precision) config and links to a representative
+ * point. The detail page ships a sibling navigator, so landing on one point of
+ * a config is enough to reach the rest of that config's concurrencies.
+ *
+ * Representative point = latest date, then highest throughput per GPU (the
+ * headline point on that config's curve), with the id as a final tiebreak so
+ * the choice is deterministic across requests.
+ */
+
+import type { DbClient } from '../connection.js';
+
+export interface AgenticCatalogEntry {
+  /** `benchmark_results.id` of the representative point for this config. */
+  id: number;
+  model: string;
+  hardware: string;
+  framework: string;
+  precision: string;
+  /** Trace-backed points in this config, i.e. detail pages reachable from it. */
+  points: number;
+  /** Lowest and highest client concurrency measured with a stored trace. */
+  minConc: number;
+  maxConc: number;
+  /** Most recent benchmark date in this config (YYYY-MM-DD). */
+  latestDate: string;
+}
+
+export async function getAgenticCatalog(sql: DbClient): Promise<AgenticCatalogEntry[]> {
+  const rows = (await sql`
+    with traced as (
+      select
+        br.id,
+        br.conc,
+        br.date::text as date,
+        c.model, c.hardware, c.framework, c.precision,
+        (br.metrics->>'tput_per_gpu')::float8 as tput_per_gpu
+      from benchmark_results br
+      join configs c on c.id = br.config_id
+      join agentic_trace_replay atr on atr.id = br.trace_replay_id
+      where atr.profile_export_jsonl_gz is not null
+    ),
+    ranked as (
+      select
+        traced.*,
+        count(*) over w as points,
+        min(conc) over w as min_conc,
+        max(conc) over w as max_conc,
+        max(date) over w as latest_date,
+        row_number() over (
+          partition by model, hardware, framework, precision
+          order by date desc, tput_per_gpu desc nulls last, id
+        ) as rn
+      from traced
+      window w as (partition by model, hardware, framework, precision)
+    )
+    select id, model, hardware, framework, precision,
+           points, min_conc, max_conc, latest_date
+    from ranked
+    where rn = 1
+    order by model, hardware, framework, precision
+  `) as unknown as {
+    id: number;
+    model: string;
+    hardware: string;
+    framework: string;
+    precision: string;
+    points: number;
+    min_conc: number;
+    max_conc: number;
+    latest_date: string;
+  }[];
+
+  return rows.map((r) => ({
+    id: Number(r.id),
+    model: r.model,
+    hardware: r.hardware,
+    framework: r.framework,
+    precision: r.precision,
+    points: Number(r.points),
+    minConc: Number(r.min_conc),
+    maxConc: Number(r.max_conc),
+    latestDate: r.latest_date,
+  }));
+}
diff --git a/packages/db/src/queries/agentic-shared.ts b/packages/db/src/queries/agentic-shared.ts
index 67b84932f..c242e4b88 100644
--- a/packages/db/src/queries/agentic-shared.ts
+++ b/packages/db/src/queries/agentic-shared.ts
@@ -15,6 +15,11 @@
 
 import type { DbClient } from '../connection.js';
 
+import {
+  buildTokenLengthSketch,
+  type TokenLengthSketch,
+} from '@semianalysisai/inferencex-constants';
+
 /**
  * Bump when the aggregate-stats computation algorithm changes — the backfill
  * script recomputes any row whose stored `aggregate_stats.version` is older,
@@ -39,11 +44,15 @@ import type { DbClient } from '../connection.js';
  * v7: add `e2elPerOsl` — percentiles of per-request E2E latency divided by
  * OSL (seconds per output token), the inverse of the "E2E Normalized Interactivity" x-axis
  * metric.
+ *
+ * v8: add p95 and bounded mergeable ISL/OSL sketches. The dashboard merges
+ * the sketches for all resident chart points instead of loading request-level
+ * timelines or attempting to combine per-point percentiles.
  */
-export const STATS_VERSION = 7;
+export const STATS_VERSION = 8;
 
 interface ProfileRecord {
-  metadata?: { benchmark_phase?: string };
+  metadata?: { benchmark_phase?: string; was_cancelled?: boolean };
   metrics?: {
     input_sequence_length?: { value?: number } | number;
     output_sequence_length?: { value?: number } | number;
@@ -63,6 +72,7 @@ export function extractIslOsl(jsonl: string): { isl: number[]; osl: number[] } {
       continue;
     }
     if (rec.metadata?.benchmark_phase && rec.metadata.benchmark_phase !== 'profiling') continue;
+    if (rec.metadata?.was_cancelled === true) continue;
     const m = rec.metrics ?? {};
     const i = readNum(m.input_sequence_length);
     const o = readNum(m.output_sequence_length);
@@ -77,6 +87,7 @@ export interface MetricPercentiles {
   p50: number;
   p75: number;
   p90: number;
+  p95: number;
   p99: number;
   /** Sample count used to compute the percentiles. */
   n: number;
@@ -110,11 +121,27 @@ export function percentilesOf(samples: number[]): MetricPercentiles | null {
     p50: quantile(sorted, 0.5),
     p75: quantile(sorted, 0.75),
     p90: quantile(sorted, 0.9),
+    p95: quantile(sorted, 0.95),
     p99: quantile(sorted, 0.99),
     n: sorted.length,
   };
 }
 
+export interface SequenceLengthSketches {
+  isl: TokenLengthSketch | null;
+  osl: TokenLengthSketch | null;
+}
+
+export function sequenceLengthSketches(
+  isl: readonly number[],
+  osl: readonly number[],
+): SequenceLengthSketches {
+  return {
+    isl: buildTokenLengthSketch(isl),
+    osl: buildTokenLengthSketch(osl),
+  };
+}
+
 /** Pull a numeric metric out of the {value, unit} envelope (or a bare number). */
 export function readNum(v: unknown): number | undefined {
   if (typeof v === 'number') return v;
diff --git a/packages/db/src/queries/benchmark-siblings.ts b/packages/db/src/queries/benchmark-siblings.ts
index d1b7f890a..a6a6c41f8 100644
--- a/packages/db/src/queries/benchmark-siblings.ts
+++ b/packages/db/src/queries/benchmark-siblings.ts
@@ -15,11 +15,15 @@ export interface BenchmarkSibling {
   decode_tp: number;
   decode_ep: number;
   decode_pp: number | null;
+  decode_dcp_size: number | null;
+  decode_pcp_size: number | null;
   decode_dp_attention: boolean;
   decode_num_workers: number;
   prefill_tp: number;
   prefill_ep: number;
   prefill_pp: number | null;
+  prefill_dcp_size: number | null;
+  prefill_pcp_size: number | null;
   prefill_dp_attention: boolean;
   prefill_num_workers: number;
   num_prefill_gpu: number;
@@ -93,8 +97,24 @@ export async function getBenchmarkSiblings(
     select
       br.id, br.conc, br.offload_mode,
       c.decode_tp, c.decode_ep, (br.metrics->>'decode_pp')::int as decode_pp,
+      coalesce(
+        (br.metrics->>'decode_dcp_size')::int,
+        (br.metrics->>'dcp_size')::int
+      ) as decode_dcp_size,
+      coalesce(
+        (br.metrics->>'decode_pcp_size')::int,
+        (br.metrics->>'pcp_size')::int
+      ) as decode_pcp_size,
       c.decode_dp_attention, c.decode_num_workers,
       c.prefill_tp, c.prefill_ep, (br.metrics->>'prefill_pp')::int as prefill_pp,
+      coalesce(
+        (br.metrics->>'prefill_dcp_size')::int,
+        (br.metrics->>'dcp_size')::int
+      ) as prefill_dcp_size,
+      coalesce(
+        (br.metrics->>'prefill_pcp_size')::int,
+        (br.metrics->>'pcp_size')::int
+      ) as prefill_pcp_size,
       c.prefill_dp_attention, c.prefill_num_workers,
       c.num_prefill_gpu, c.num_decode_gpu, c.disagg, c.is_multinode,
       (br.metrics->>'tput_per_gpu')::float8 as tput_per_gpu,
@@ -120,11 +140,15 @@ export async function getBenchmarkSiblings(
     decode_tp: number;
     decode_ep: number;
     decode_pp: number | null;
+    decode_dcp_size: number | null;
+    decode_pcp_size: number | null;
     decode_dp_attention: boolean;
     decode_num_workers: number;
     prefill_tp: number;
     prefill_ep: number;
     prefill_pp: number | null;
+    prefill_dcp_size: number | null;
+    prefill_pcp_size: number | null;
     prefill_dp_attention: boolean;
     prefill_num_workers: number;
     num_prefill_gpu: number;
@@ -143,11 +167,15 @@ export async function getBenchmarkSiblings(
     decode_tp: r.decode_tp,
     decode_ep: r.decode_ep,
     decode_pp: r.decode_pp === null ? null : Number(r.decode_pp),
+    decode_dcp_size: r.decode_dcp_size === null ? null : Number(r.decode_dcp_size),
+    decode_pcp_size: r.decode_pcp_size === null ? null : Number(r.decode_pcp_size),
     decode_dp_attention: r.decode_dp_attention,
     decode_num_workers: r.decode_num_workers,
     prefill_tp: r.prefill_tp,
     prefill_ep: r.prefill_ep,
     prefill_pp: r.prefill_pp === null ? null : Number(r.prefill_pp),
+    prefill_dcp_size: r.prefill_dcp_size === null ? null : Number(r.prefill_dcp_size),
+    prefill_pcp_size: r.prefill_pcp_size === null ? null : Number(r.prefill_pcp_size),
     prefill_dp_attention: r.prefill_dp_attention,
     prefill_num_workers: r.prefill_num_workers,
     num_prefill_gpu: r.num_prefill_gpu,
diff --git a/packages/db/src/queries/benchmarks.test.ts b/packages/db/src/queries/benchmarks.test.ts
new file mode 100644
index 000000000..bdebe5b82
--- /dev/null
+++ b/packages/db/src/queries/benchmarks.test.ts
@@ -0,0 +1,115 @@
+import { describe, expect, it, vi } from 'vitest';
+
+import { getAllBenchmarksForHistory, getBenchmarksForRun, getLatestBenchmarks } from './benchmarks';
+
+interface CapturedQuery {
+  text: string;
+  values: unknown[];
+}
+
+function captureSql() {
+  let query: CapturedQuery = { text: '', values: [] };
+  const sql = vi.fn((strings: TemplateStringsArray, ...values: unknown[]) => {
+    query = {
+      text: strings.join('?').replaceAll(/\s+/gu, ' ').trim(),
+      values,
+    };
+    return Promise.resolve([]);
+  });
+  return {
+    sql: sql as unknown as Parameters<typeof getBenchmarksForRun>[0],
+    query: () => query,
+    calls: () => sql.mock.calls.length,
+  };
+}
+
+describe('append-only benchmark snapshots', () => {
+  it('seeds a run snapshot from only the requested run, then walks older runs', async () => {
+    const captured = captureSql();
+
+    await getBenchmarksForRun(captured.sql, 'dsv4', 123456);
+    expect(captured.calls()).toBe(1);
+
+    const { text, values } = captured.query();
+    expect(text).toContain('WITH RECURSIVE run_lines AS');
+    expect(text).toContain('FROM ranked_runs WHERE github_run_id = ? UNION ALL');
+    expect(text).not.toContain('seed_runs AS');
+    expect(text).not.toContain('r.date <=');
+    expect(text).toContain('WHERE current.append_only');
+    expect(text).toContain('older.image = current.root_image');
+    expect(text).toContain('older.line_spec_method = current.line_spec_method');
+    expect(text).toContain('point_c.id = br.config_id');
+    expect(text).toContain('br.recipe_fingerprint, br.conc, cr.run_rank');
+    expect(text).toContain('br.workflow_run_id, wr.run_started_at::text');
+    expect(text).toContain('br.snapshot_date::text AS curve_date');
+    expect(text).toContain('snapshot_wr.id = br.snapshot_workflow_run_id');
+    expect(values).toEqual([['dsv4'], 123456]);
+  });
+
+  it('chooses latest seeds under both the date and as-of-run cutoffs', async () => {
+    const captured = captureSql();
+
+    await getLatestBenchmarks(captured.sql, 'dsv4', '2026-08-01', false, '456789');
+
+    const { text, values } = captured.query();
+    expect(text).toContain('seed_runs AS ( SELECT DISTINCT ON');
+    expect(text).toContain('FROM ranked_runs r WHERE r.date <= ?::date');
+    expect(text).toContain('lwr.github_run_id = ?');
+    expect(text).toContain('FROM seed_runs UNION ALL');
+    expect(text).not.toContain('WHERE run_rank = 1');
+    expect(values).toEqual([['dsv4'], '2026-08-01', 456789]);
+  });
+
+  it('keeps exact-date seeds independent from the main-chart run cutoff', async () => {
+    const captured = captureSql();
+
+    await getLatestBenchmarks(captured.sql, 'dsv4', '2026-08-01', true, '456789');
+
+    const { text, values } = captured.query();
+    expect(text).toContain('FROM ranked_runs r WHERE r.date = ?::date');
+    expect(text).not.toContain('lwr.github_run_id = ?');
+    expect(values).toEqual([['dsv4'], '2026-08-01']);
+  });
+
+  it('stamps latest materialized-view rows with a separate logical identity', async () => {
+    const captured = captureSql();
+
+    await getLatestBenchmarks(captured.sql, 'dsv4');
+
+    const { text } = captured.query();
+    expect(text).toContain('FROM latest_benchmarks lb');
+    expect(text).toContain('lb.date::text, lb.workflow_run_id');
+    expect(text).toContain('lb.snapshot_date::text AS curve_date');
+    expect(text).toContain('lb.recipe_fingerprint,');
+    expect(text).toContain('snapshot_wr.id = lb.snapshot_workflow_run_id');
+  });
+
+  it('seeds every filtered sequence run as a distinct historical snapshot', async () => {
+    const captured = captureSql();
+
+    await getAllBenchmarksForHistory(captured.sql, 'dsv4', 8192, 1024);
+
+    const { text, values } = captured.query();
+    expect(text).toContain('AND br.isl = ? AND br.osl = ? AND br.error IS NULL');
+    expect(text).toContain('FROM ranked_runs UNION ALL');
+    expect(text).not.toContain('WHERE github_run_id = ?');
+    expect(text).not.toContain('seed_runs AS');
+    expect(text).toContain('cr.snapshot_workflow_run_id, br.config_id');
+    expect(text).toContain('br.recipe_fingerprint, br.conc, cr.run_rank');
+    expect(text).toContain("br.metrics - '{std_ttft,std_tpot");
+    expect(text).toContain('ORDER BY br.snapshot_date, c.id, br.conc');
+    expect(values).toEqual([['dsv4'], 8192, 1024]);
+  });
+
+  it('filters agentic history by scenario without imposing sequence lengths', async () => {
+    const captured = captureSql();
+
+    await getAllBenchmarksForHistory(captured.sql, 'dsv4', null, null, 'agentic_traces');
+
+    const { text, values } = captured.query();
+    expect(text).toContain("AND br.benchmark_type = 'agentic_traces' AND br.error IS NULL");
+    expect(text).not.toContain('AND br.isl = ?');
+    expect(text).toContain('FROM ranked_runs UNION ALL');
+    expect(values).toEqual([['dsv4']]);
+  });
+});
diff --git a/packages/db/src/queries/benchmarks.ts b/packages/db/src/queries/benchmarks.ts
index d7989e4cc..25aead1f6 100644
--- a/packages/db/src/queries/benchmarks.ts
+++ b/packages/db/src/queries/benchmarks.ts
@@ -1,5 +1,6 @@
 import type { DbClient } from '../connection.js';
 import type { WorkerPower } from '../etl/benchmark-mapper.js';
+export type { WorkerPower } from '../etl/benchmark-mapper.js';
 
 /**
  * One entry in `BenchmarkRow.workers` — mirrors the runner's aggregate_power.py
@@ -36,6 +37,8 @@ export interface BenchmarkRow {
   conc: number;
   offload_mode: string;
   image: string | null;
+  /** Producer-generated identity for the complete recipe; null on legacy rows. */
+  recipe_fingerprint?: string | null;
   metrics: Record<string, number>;
   /**
    * Per-worker measured-power breakdown emitted on multinode / disagg runs.
@@ -45,10 +48,336 @@ export interface BenchmarkRow {
    */
   workers?: BenchmarkWorkerRow[];
   date: string;
-  /** Internal workflow identity used to keep merged agentic curves within one run. */
+  /** Producer identity and timestamp; preserved for per-point provenance. */
   workflow_run_id?: number;
   run_started_at?: string | null;
   run_url: string | null;
+  /** Logical snapshot identity. Set when an append-only run carries older points forward. */
+  curve_date?: string;
+  curve_workflow_run_id?: number;
+  curve_run_started_at?: string | null;
+}
+
+const SQL_PIECE = Symbol('sql-piece');
+
+interface SqlPiece {
+  readonly [SQL_PIECE]: true;
+  readonly strings: string[];
+  readonly values: unknown[];
+}
+
+/**
+ * Compose static SQL fragments before invoking a database driver. The final
+ * DbClient call receives one ordinary TemplateStringsArray plus bound values;
+ * neither Neon HTTP nor postgres.js has to interpret driver-specific fragments.
+ */
+function sqlPiece(strings: TemplateStringsArray, ...values: unknown[]): SqlPiece {
+  const composedStrings = [strings[0] ?? ''];
+  const composedValues: unknown[] = [];
+
+  for (const [index, value] of values.entries()) {
+    if (typeof value === 'object' && value !== null && SQL_PIECE in value) {
+      const piece = value as SqlPiece;
+      composedStrings[composedStrings.length - 1] += piece.strings[0] ?? '';
+      for (const [pieceIndex, pieceValue] of piece.values.entries()) {
+        composedValues.push(pieceValue);
+        composedStrings.push(piece.strings[pieceIndex + 1] ?? '');
+      }
+      composedStrings[composedStrings.length - 1] += strings[index + 1] ?? '';
+    } else {
+      composedValues.push(value);
+      composedStrings.push(strings[index + 1] ?? '');
+    }
+  }
+
+  return {
+    [SQL_PIECE]: true,
+    strings: composedStrings,
+    values: composedValues,
+  };
+}
+
+function executeSqlPiece(sql: DbClient, piece: SqlPiece) {
+  const strings = [...piece.strings] as unknown as TemplateStringsArray;
+  Object.defineProperty(strings, 'raw', { value: [...piece.strings] });
+  return sql(strings, ...piece.values);
+}
+
+type RecursiveBenchmarkVariant =
+  | {
+      kind: 'latest';
+      date: string;
+      exact: boolean;
+      asOfRunId?: number;
+    }
+  | {
+      kind: 'run';
+      githubRunId: number;
+    }
+  | {
+      kind: 'history';
+      isl: number | null;
+      osl: number | null;
+      benchmarkType?: string;
+    };
+
+interface RecursiveBenchmarkPlan {
+  runLineFilter: SqlPiece;
+  seed: SqlPiece;
+  pointIdentity: SqlPiece;
+  pointOrder: SqlPiece;
+  metricsExpression: SqlPiece;
+  resultOrder: SqlPiece;
+}
+
+function curveRunBase(modelKeys: string[], runLineFilter: SqlPiece): SqlPiece {
+  return sqlPiece`
+    run_lines AS (
+      SELECT
+        c.model, c.hardware, c.framework, c.precision, c.disagg,
+        CASE WHEN br.benchmark_type = 'agentic_traces' THEN '' ELSE c.spec_method END AS line_spec_method,
+        br.benchmark_type, br.isl, br.osl, br.offload_mode,
+        br.workflow_run_id, br.date, wr.run_started_at, wr.github_run_id,
+        wr.append_only, min(br.image) AS image,
+        count(DISTINCT br.image) AS image_count,
+        bool_and(br.image IS NOT NULL) AS images_complete
+      FROM benchmark_results br
+      JOIN configs c ON c.id = br.config_id
+      JOIN latest_workflow_runs wr ON wr.id = br.workflow_run_id
+      WHERE c.model = ANY(${modelKeys})
+        ${runLineFilter}
+        AND br.error IS NULL
+      GROUP BY
+        c.model, c.hardware, c.framework, c.precision, c.disagg,
+        CASE WHEN br.benchmark_type = 'agentic_traces' THEN '' ELSE c.spec_method END,
+        br.benchmark_type, br.isl, br.osl, br.offload_mode,
+        br.workflow_run_id, br.date, wr.run_started_at, wr.github_run_id, wr.append_only
+    ), ranked_runs AS (
+      SELECT run_lines.*,
+        row_number() OVER (
+          PARTITION BY
+            model, hardware, framework, precision, disagg, line_spec_method,
+            benchmark_type, isl, osl, offload_mode
+          ORDER BY date DESC, run_started_at DESC NULLS LAST, workflow_run_id DESC
+        ) AS run_rank
+      FROM run_lines
+    )`;
+}
+
+const CURVE_RECURSION = sqlPiece`
+  UNION ALL
+
+  SELECT
+    older.*,
+    current.root_image,
+    current.snapshot_date,
+    current.snapshot_workflow_run_id
+  FROM curve_runs current
+  JOIN ranked_runs older
+    ON older.model = current.model
+    AND older.hardware = current.hardware
+    AND older.framework = current.framework
+    AND older.precision = current.precision
+    AND older.disagg = current.disagg
+    AND older.line_spec_method = current.line_spec_method
+    AND older.benchmark_type = current.benchmark_type
+    AND older.isl IS NOT DISTINCT FROM current.isl
+    AND older.osl IS NOT DISTINCT FROM current.osl
+    AND older.offload_mode = current.offload_mode
+    AND older.run_rank = current.run_rank + 1
+  WHERE current.append_only
+    AND current.image_count = 1
+    AND current.images_complete
+    AND older.image_count = 1
+    AND older.images_complete
+    AND older.image = current.root_image`;
+
+const CURVE_POINT_SOURCE = sqlPiece`
+  FROM curve_runs cr
+  JOIN benchmark_results br
+    ON br.workflow_run_id = cr.workflow_run_id
+    AND br.benchmark_type = cr.benchmark_type
+    AND br.isl IS NOT DISTINCT FROM cr.isl
+    AND br.osl IS NOT DISTINCT FROM cr.osl
+    AND br.offload_mode = cr.offload_mode
+  JOIN configs point_c
+    ON point_c.id = br.config_id
+    AND point_c.model = cr.model
+    AND point_c.hardware = cr.hardware
+    AND point_c.framework = cr.framework
+    AND point_c.precision = cr.precision
+    AND point_c.disagg = cr.disagg
+    AND CASE WHEN br.benchmark_type = 'agentic_traces' THEN '' ELSE point_c.spec_method END = cr.line_spec_method
+  WHERE br.error IS NULL`;
+
+function executeRecursiveBenchmarkQuery(
+  sql: DbClient,
+  modelKeys: string[],
+  plan: RecursiveBenchmarkPlan,
+) {
+  const query = sqlPiece`
+    WITH RECURSIVE
+    ${curveRunBase(modelKeys, plan.runLineFilter)},
+    ${plan.seed}
+    ${CURVE_RECURSION}
+    ), selected_points AS (
+      SELECT DISTINCT ON (
+        ${plan.pointIdentity}
+      ) br.*, cr.snapshot_date, cr.snapshot_workflow_run_id
+      ${CURVE_POINT_SOURCE}
+      ORDER BY
+        ${plan.pointOrder}
+    )
+    SELECT
+      br.id,
+      c.hardware,
+      c.framework,
+      c.model,
+      c.precision,
+      c.spec_method,
+      c.disagg,
+      c.is_multinode,
+      c.prefill_tp,
+      c.prefill_ep,
+      c.prefill_dp_attention,
+      c.prefill_num_workers,
+      c.decode_tp,
+      c.decode_ep,
+      c.decode_dp_attention,
+      c.decode_num_workers,
+      c.num_prefill_gpu,
+      c.num_decode_gpu,
+      br.benchmark_type,
+      br.offload_mode,
+      br.isl,
+      br.osl,
+      br.conc,
+      br.image,
+      br.recipe_fingerprint,
+      ${plan.metricsExpression},
+      br.workers,
+      br.date::text,
+      br.workflow_run_id,
+      wr.run_started_at::text,
+      CASE WHEN wr.html_url IS NOT NULL THEN wr.html_url || '/attempts/' || wr.run_attempt ELSE NULL END AS run_url,
+      br.snapshot_date::text AS curve_date,
+      br.snapshot_workflow_run_id AS curve_workflow_run_id,
+      snapshot_wr.run_started_at::text AS curve_run_started_at
+    FROM selected_points br
+    JOIN configs c ON c.id = br.config_id
+    JOIN latest_workflow_runs wr ON wr.id = br.workflow_run_id
+    JOIN latest_workflow_runs snapshot_wr ON snapshot_wr.id = br.snapshot_workflow_run_id
+    ORDER BY ${plan.resultOrder}
+  `;
+  return executeSqlPiece(sql, query);
+}
+
+function queryRecursiveBenchmarks(
+  sql: DbClient,
+  modelKeys: string[],
+  variant: RecursiveBenchmarkVariant,
+) {
+  const pointIdentity = sqlPiece`
+    br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode,
+    br.recipe_fingerprint, br.conc`;
+  const pointOrder = sqlPiece`
+    br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode,
+    br.recipe_fingerprint, br.conc, cr.run_rank`;
+  const rawMetrics = sqlPiece`br.metrics`;
+
+  switch (variant.kind) {
+    case 'latest': {
+      const datePredicate = variant.exact
+        ? sqlPiece`r.date = ${variant.date}::date`
+        : sqlPiece`r.date <= ${variant.date}::date`;
+      const asOfPredicate =
+        !variant.exact && variant.asOfRunId !== undefined
+          ? sqlPiece`AND (
+              r.run_started_at IS NULL
+              OR r.run_started_at <= COALESCE(
+                (SELECT lwr.run_started_at FROM latest_workflow_runs lwr WHERE lwr.github_run_id = ${variant.asOfRunId}),
+                'infinity'::timestamptz
+              )
+            )`
+          : sqlPiece``;
+      const seed = sqlPiece`
+        seed_runs AS (
+          SELECT DISTINCT ON (
+            r.model, r.hardware, r.framework, r.precision, r.disagg,
+            r.line_spec_method, r.benchmark_type, r.isl, r.osl, r.offload_mode
+          )
+            r.*
+          FROM ranked_runs r
+          WHERE ${datePredicate}
+            ${asOfPredicate}
+          ORDER BY
+            r.model, r.hardware, r.framework, r.precision, r.disagg,
+            r.line_spec_method, r.benchmark_type, r.isl, r.osl, r.offload_mode,
+            r.date DESC, r.run_started_at DESC NULLS LAST, r.workflow_run_id DESC
+        ), curve_runs AS (
+          SELECT
+            seed_runs.*,
+            seed_runs.image AS root_image,
+            seed_runs.date AS snapshot_date,
+            seed_runs.workflow_run_id AS snapshot_workflow_run_id
+          FROM seed_runs`;
+      return executeRecursiveBenchmarkQuery(sql, modelKeys, {
+        runLineFilter: sqlPiece``,
+        seed,
+        pointIdentity,
+        pointOrder,
+        metricsExpression: rawMetrics,
+        resultOrder: sqlPiece`br.config_id, br.conc, br.isl, br.osl`,
+      });
+    }
+    case 'run': {
+      const seed = sqlPiece`
+        curve_runs AS (
+          SELECT
+            ranked_runs.*,
+            ranked_runs.image AS root_image,
+            ranked_runs.date AS snapshot_date,
+            ranked_runs.workflow_run_id AS snapshot_workflow_run_id
+          FROM ranked_runs
+          WHERE github_run_id = ${variant.githubRunId}`;
+      return executeRecursiveBenchmarkQuery(sql, modelKeys, {
+        runLineFilter: sqlPiece``,
+        seed,
+        pointIdentity,
+        pointOrder,
+        metricsExpression: rawMetrics,
+        resultOrder: sqlPiece`br.config_id, br.conc, br.isl, br.osl, br.offload_mode`,
+      });
+    }
+    case 'history': {
+      const runLineFilter =
+        variant.benchmarkType === 'agentic_traces'
+          ? sqlPiece`AND br.benchmark_type = 'agentic_traces'`
+          : sqlPiece`AND br.isl = ${variant.isl} AND br.osl = ${variant.osl}`;
+      const seed = sqlPiece`
+        curve_runs AS (
+          SELECT
+            ranked_runs.*,
+            ranked_runs.image AS root_image,
+            ranked_runs.date AS snapshot_date,
+            ranked_runs.workflow_run_id AS snapshot_workflow_run_id
+          FROM ranked_runs`;
+      return executeRecursiveBenchmarkQuery(sql, modelKeys, {
+        runLineFilter,
+        seed,
+        pointIdentity: sqlPiece`
+          cr.snapshot_workflow_run_id,
+          br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode,
+          br.recipe_fingerprint, br.conc`,
+        pointOrder: sqlPiece`
+          cr.snapshot_workflow_run_id,
+          br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode,
+          br.recipe_fingerprint, br.conc, cr.run_rank`,
+        metricsExpression: sqlPiece`br.metrics - '{std_ttft,std_tpot,std_e2el,std_intvty,std_itl,mean_ttft,mean_tpot,mean_e2el,mean_intvty,mean_itl}'::text[] AS metrics`,
+        resultOrder: sqlPiece`br.snapshot_date, c.id, br.conc`,
+      });
+    }
+  }
 }
 
 /**
@@ -56,12 +385,10 @@ export interface BenchmarkRow {
  * up to a given date. Multiple keys support point-release grouping — e.g. passing
  * `['glm5', 'glm5.1']` unions both buckets under the one display.
  *
- * Selection unit is the LINE, not the point: for each line
- * `(config_id, benchmark_type, isl, osl, offload_mode)` we pick the single newest workflow run that
- * produced data for it (newest date, then latest sweep, then highest run id) and return
- * EVERY concurrency that one run measured — and nothing from any other run. A partial
- * re-sweep therefore truncates the line to its own concurrencies rather than stitching the
- * skipped ones from an older run. This guarantees a line never mixes runs/dates.
+ * Selection unit is the LINE, not the point. Normally each line comes entirely from
+ * its newest workflow run. Runs explicitly marked append-only are the sole exception:
+ * their new points extend the immediately preceding same-image curve, including a
+ * consecutive chain of append-only runs back to the nearest full snapshot.
  *
  * The frontend filters by sequence client-side. This eliminates API round-trips when
  * switching sequences — the data is already cached by React Query.
@@ -82,90 +409,15 @@ export async function getLatestBenchmarks(
 ): Promise<BenchmarkRow[]> {
   const modelKeys = Array.isArray(modelKey) ? modelKey : [modelKey];
   if (date) {
-    // Date-filtered: use the base table (the view only has the absolute latest).
-    // exact=true: only this exact date (GPU comparison); exact=false (default): as of this date.
-    const dateFilter = exact ? sql`br.date = ${date}::date` : sql`br.date <= ${date}::date`;
-    // "As of run" filter (main chart only): keep results whose run started no later
-    // than the selected run. run_started_at is an absolute timestamp, so this also
-    // naturally includes all earlier-date runs. NULLs (pre-migration-003 runs that
-    // lack the timestamp) are kept so old history doesn't blank out; COALESCE to
-    // infinity makes an unknown asOfRunId a no-op rather than excluding everything.
-    const runFilter =
-      !exact && asOfRunId
-        ? sql`AND (
-            wr.run_started_at IS NULL
-            OR wr.run_started_at <= COALESCE(
-              (SELECT lwr.run_started_at FROM latest_workflow_runs lwr WHERE lwr.github_run_id = ${Number(asOfRunId)}),
-              'infinity'::timestamptz
-            )
-          )`
-        : sql``;
-    // winners: the single newest run per LINE
-    // (config_id, benchmark_type, isl, osl, offload_mode) under the
-    // date/run cutoff. br.date is a calendar day, so two same-day sweeps tie on date — break
-    // by wr.run_started_at (latest sweep wins), then br.workflow_run_id so exactly one run wins
-    // even when run_started_at is equal/null. The outer join then pulls EVERY concurrency that
-    // winning run measured for the line, so the line is built from one run only (no carry-forward
-    // of concurrencies a partial re-sweep skipped).
-    const rows = await sql`
-      WITH winners AS (
-        SELECT DISTINCT ON (br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode)
-          br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode,
-          br.workflow_run_id AS winning_run_id
-        FROM benchmark_results br
-        JOIN configs c ON c.id = br.config_id
-        JOIN latest_workflow_runs wr ON wr.id = br.workflow_run_id
-        WHERE c.model = ANY(${modelKeys})
-          AND br.error IS NULL
-          AND ${dateFilter}
-          ${runFilter}
-        ORDER BY br.config_id, br.benchmark_type, br.isl, br.osl, br.offload_mode,
-                 br.date DESC, wr.run_started_at DESC NULLS LAST, br.workflow_run_id DESC
-      )
-      SELECT
-        br.id,
-        c.hardware,
-        c.framework,
-        c.model,
-        c.precision,
-        c.spec_method,
-        c.disagg,
-        c.is_multinode,
-        c.prefill_tp,
-        c.prefill_ep,
-        c.prefill_dp_attention,
-        c.prefill_num_workers,
-        c.decode_tp,
-        c.decode_ep,
-        c.decode_dp_attention,
-        c.decode_num_workers,
-        c.num_prefill_gpu,
-        c.num_decode_gpu,
-        br.benchmark_type,
-        br.offload_mode,
-        br.isl,
-        br.osl,
-        br.conc,
-        br.image,
-        br.metrics,
-        br.workers,
-        br.date::text,
-        br.workflow_run_id,
-        wr.run_started_at::text,
-        CASE WHEN wr.html_url IS NOT NULL THEN wr.html_url || '/attempts/' || wr.run_attempt ELSE NULL END AS run_url
-      FROM benchmark_results br
-      JOIN configs c ON c.id = br.config_id
-      JOIN latest_workflow_runs wr ON wr.id = br.workflow_run_id
-      JOIN winners w
-        ON w.config_id = br.config_id
-        AND w.benchmark_type = br.benchmark_type
-        AND w.isl IS NOT DISTINCT FROM br.isl
-        AND w.osl IS NOT DISTINCT FROM br.osl
-        AND w.offload_mode = br.offload_mode
-        AND w.winning_run_id = br.workflow_run_id
-      WHERE br.error IS NULL
-      ORDER BY br.config_id, br.conc, br.isl, br.osl
-    `;
+    // The latest variant chooses one seed per line under the requested cutoff.
+    // Its recursive history remains unrestricted so an append-only seed can
+    // still reach the nearest earlier full same-image snapshot.
+    const rows = await queryRecursiveBenchmarks(sql, modelKeys, {
+      kind: 'latest',
+      date,
+      exact: exact ?? false,
+      asOfRunId: !exact && asOfRunId ? Number(asOfRunId) : undefined,
+    });
     return rows as unknown as BenchmarkRow[];
   }
 
@@ -196,15 +448,20 @@ export async function getLatestBenchmarks(
       lb.osl,
       lb.conc,
       lb.image,
+      lb.recipe_fingerprint,
       lb.metrics,
       lb.workers,
       lb.date::text,
       lb.workflow_run_id,
       wr.run_started_at::text,
-      CASE WHEN wr.html_url IS NOT NULL THEN wr.html_url || '/attempts/' || wr.run_attempt ELSE NULL END AS run_url
+      CASE WHEN wr.html_url IS NOT NULL THEN wr.html_url || '/attempts/' || wr.run_attempt ELSE NULL END AS run_url,
+      lb.snapshot_date::text AS curve_date,
+      lb.snapshot_workflow_run_id AS curve_workflow_run_id,
+      snapshot_wr.run_started_at::text AS curve_run_started_at
     FROM latest_benchmarks lb
     JOIN configs c ON c.id = lb.config_id
     JOIN latest_workflow_runs wr ON wr.id = lb.workflow_run_id
+    JOIN latest_workflow_runs snapshot_wr ON snapshot_wr.id = lb.snapshot_workflow_run_id
     WHERE c.model = ANY(${modelKeys})
     ORDER BY lb.config_id, lb.conc, lb.isl, lb.osl, lb.date DESC
   `;
@@ -212,11 +469,9 @@ export async function getLatestBenchmarks(
 }
 
 /**
- * Fetch the benchmark results produced by ONE specific workflow run (by GitHub
- * run id). Unlike {@link getLatestBenchmarks}, this returns exactly what that run
- * measured — used by the GPU comparison view to plot individual same-day runs as
- * distinct series (e.g. comparing a day-zero sweep against a same-day re-sweep).
- * Returns an empty array if the run produced no results for the model.
+ * Fetch the curve snapshot represented by one workflow run. Ordinary runs return
+ * exactly their own points; append-only runs also include the immediately preceding
+ * same-image curve chain. Used by GPU comparison for same-day run snapshots.
  */
 export async function getBenchmarksForRun(
   sql: DbClient,
@@ -224,54 +479,14 @@ export async function getBenchmarksForRun(
   githubRunId: string | number,
 ): Promise<BenchmarkRow[]> {
   const modelKeys = Array.isArray(modelKey) ? modelKey : [modelKey];
-  const rows = await sql`
-    SELECT DISTINCT ON (br.config_id, br.conc, br.isl, br.osl, br.offload_mode)
-      br.id,
-      c.hardware,
-      c.framework,
-      c.model,
-      c.precision,
-      c.spec_method,
-      c.disagg,
-      c.is_multinode,
-      c.prefill_tp,
-      c.prefill_ep,
-      c.prefill_dp_attention,
-      c.prefill_num_workers,
-      c.decode_tp,
-      c.decode_ep,
-      c.decode_dp_attention,
-      c.decode_num_workers,
-      c.num_prefill_gpu,
-      c.num_decode_gpu,
-      br.benchmark_type,
-      br.offload_mode,
-      br.isl,
-      br.osl,
-      br.conc,
-      br.image,
-      br.metrics,
-      br.workers,
-      br.date::text,
-      br.workflow_run_id,
-      wr.run_started_at::text,
-      CASE WHEN wr.html_url IS NOT NULL THEN wr.html_url || '/attempts/' || wr.run_attempt ELSE NULL END AS run_url
-    FROM benchmark_results br
-    JOIN configs c ON c.id = br.config_id
-    JOIN latest_workflow_runs wr ON wr.id = br.workflow_run_id
-    WHERE c.model = ANY(${modelKeys})
-      AND br.error IS NULL
-      AND wr.github_run_id = ${Number(githubRunId)}
-    ORDER BY br.config_id, br.conc, br.isl, br.osl, br.offload_mode, br.date DESC
-  `;
+  const rows = await queryRecursiveBenchmarks(sql, modelKeys, {
+    kind: 'run',
+    githubRunId: Number(githubRunId),
+  });
   return rows as unknown as BenchmarkRow[];
 }
 
-/**
- * Fetch ALL benchmark results for a model + sequence across ALL dates.
- * No DISTINCT ON — returns every successful result, one per (config, conc, date).
- * Used by Historical Trends and Performance Over Time features.
- */
+/** Fetch every logical curve snapshot across time for historical views. */
 export async function getAllBenchmarksForHistory(
   sql: DbClient,
   modelKey: string | string[],
@@ -280,49 +495,11 @@ export async function getAllBenchmarksForHistory(
   benchmarkType?: string,
 ): Promise<BenchmarkRow[]> {
   const modelKeys = Array.isArray(modelKey) ? modelKey : [modelKey];
-  const sequenceFilter =
-    benchmarkType === 'agentic_traces'
-      ? sql`br.benchmark_type = 'agentic_traces'`
-      : sql`br.isl = ${isl} AND br.osl = ${osl}`;
-  const rows = await sql`
-    SELECT
-      br.id,
-      c.hardware,
-      c.framework,
-      c.model,
-      c.precision,
-      c.spec_method,
-      c.disagg,
-      c.is_multinode,
-      c.prefill_tp,
-      c.prefill_ep,
-      c.prefill_dp_attention,
-      c.prefill_num_workers,
-      c.decode_tp,
-      c.decode_ep,
-      c.decode_dp_attention,
-      c.decode_num_workers,
-      c.num_prefill_gpu,
-      c.num_decode_gpu,
-      br.benchmark_type,
-      br.offload_mode,
-      br.isl,
-      br.osl,
-      br.conc,
-      br.image,
-      br.metrics - '{std_ttft,std_tpot,std_e2el,std_intvty,std_itl,mean_ttft,mean_tpot,mean_e2el,mean_intvty,mean_itl}'::text[] as metrics,
-      br.workers,
-      br.date::text,
-      br.workflow_run_id,
-      wr.run_started_at::text,
-      CASE WHEN wr.html_url IS NOT NULL THEN wr.html_url || '/attempts/' || wr.run_attempt ELSE NULL END AS run_url
-    FROM configs c
-    JOIN benchmark_results br ON br.config_id = c.id
-      AND ${sequenceFilter}
-      AND br.error IS NULL
-    JOIN latest_workflow_runs wr ON wr.id = br.workflow_run_id
-    WHERE c.model = ANY(${modelKeys})
-    ORDER BY br.date, c.id, br.conc
-  `;
+  const rows = await queryRecursiveBenchmarks(sql, modelKeys, {
+    kind: 'history',
+    isl,
+    osl,
+    benchmarkType,
+  });
   return rows as unknown as BenchmarkRow[];
 }
diff --git a/packages/db/src/queries/datasets.ts b/packages/db/src/queries/datasets.ts
index bbcb2ece6..f895cd342 100644
--- a/packages/db/src/queries/datasets.ts
+++ b/packages/db/src/queries/datasets.ts
@@ -9,6 +9,7 @@
 
 import type { DbClient } from '../connection.js';
 import type { ConversationStructure } from '../etl/weka-structure.js';
+export type { ConversationStructure, StructureNode } from '../etl/weka-structure.js';
 
 export interface DatasetSummary {
   blockSize?: number;
diff --git a/packages/db/src/queries/derived-agentic-metrics.ts b/packages/db/src/queries/derived-agentic-metrics.ts
index 2224008d2..5685ef2e7 100644
--- a/packages/db/src/queries/derived-agentic-metrics.ts
+++ b/packages/db/src/queries/derived-agentic-metrics.ts
@@ -29,9 +29,11 @@ import {
   fetchAggregateStatsRows,
   percentilesOf,
   readNum,
+  sequenceLengthSketches,
   STATS_VERSION,
   writeBackTraceReplayJsonb,
   type MetricPercentiles,
+  type SequenceLengthSketches,
 } from './agentic-shared';
 
 export interface DerivedAgenticMetric {
@@ -62,6 +64,7 @@ interface StoredAggregateStats {
   kvCacheUtil: MetricPercentiles | null;
   prefixCacheHitRate: MetricPercentiles | null;
   e2elPerOsl: MetricPercentiles | null;
+  sequenceLengths: SequenceLengthSketches;
 }
 
 /**
@@ -233,6 +236,7 @@ export async function getDerivedAgenticMetrics(
           kvCacheUtil: prior?.kvCacheUtil ?? null,
           prefixCacheHitRate: prior?.prefixCacheHitRate ?? null,
           e2elPerOsl: e2el_per_osl,
+          sequenceLengths: sequenceLengthSketches(isl, osl),
         };
         writeBackTraceReplayJsonb(sql, 'aggregate_stats', Number(row.trace_replay_id), merged);
       }
diff --git a/packages/db/src/queries/request-chart-data.test.ts b/packages/db/src/queries/request-chart-data.test.ts
new file mode 100644
index 000000000..451df3c0e
--- /dev/null
+++ b/packages/db/src/queries/request-chart-data.test.ts
@@ -0,0 +1,87 @@
+import { describe, expect, it } from 'vitest';
+
+import { REQUEST_TIMELINE_VERSION } from '../etl/compute-request-timeline';
+import type { DbClient } from '../connection.js';
+
+import {
+  encodeRequestChartData,
+  getRequestChartData,
+  REQUEST_CHART_DATA_VERSION,
+} from './request-chart-data';
+
+function mockSql(queue: unknown[][]): { sql: DbClient; calls: string[] } {
+  const responses = [...queue];
+  const calls: string[] = [];
+  const sql = ((strings: TemplateStringsArray) => {
+    calls.push(strings.join('?'));
+    return Promise.resolve(responses.shift() ?? []);
+  }) as unknown as DbClient;
+  return { sql, calls };
+}
+
+describe('request chart data', () => {
+  it('dictionary-encodes repeated conversation and phase strings', () => {
+    const encoded = encodeRequestChartData(
+      { version: REQUEST_TIMELINE_VERSION, startNs: 10, endNs: 30, durationS: 2 },
+      [
+        {
+          cid: 'conversation-a',
+          phase: 'profiling',
+          start: 1_000,
+          end: 2_000,
+          ttftMs: 3.123_456,
+          tpotMs: 4.654_321,
+          isl: 5,
+          osl: 6,
+          cancelled: false,
+        },
+        {
+          cid: 'conversation-a',
+          phase: 'profiling',
+          start: 7_000,
+          end: 8_000,
+          ttftMs: null,
+          tpotMs: null,
+          isl: 9,
+          osl: 10,
+          cancelled: true,
+        },
+      ],
+    );
+
+    expect(encoded).toMatchObject({
+      version: REQUEST_CHART_DATA_VERSION,
+      timelineVersion: REQUEST_TIMELINE_VERSION,
+      cids: ['conversation-a'],
+      phases: ['profiling'],
+    });
+    expect(encoded.requests).toEqual([
+      [0, 0, 1, 2, 3.123, 4.654, 5, 6, 0],
+      [0, 0, 7, 8, null, null, 9, 10, 1],
+    ]);
+  });
+
+  it('projects chart-only fields with one JSONB expansion', async () => {
+    const { sql, calls } = mockSql([
+      [
+        {
+          trace_replay_id: 7,
+          has_blob: true,
+          timeline_version: REQUEST_TIMELINE_VERSION,
+          start_ns: 10,
+          end_ns: 30,
+          duration_s: 2,
+          requests: [['c', 'profiling', 1_000, 2_000, 3, 4, 5, 6, false]],
+        },
+      ],
+    ]);
+
+    const result = await getRequestChartData(sql, 42);
+
+    expect(result?.requests).toEqual([[0, 0, 1, 2, 3, 4, 5, 6, 0]]);
+    expect(calls).toHaveLength(1);
+    expect(calls[0]).toContain('jsonb_array_elements');
+    expect(calls[0]).toContain("request->>'ttftMs'");
+    expect(calls[0]).not.toContain('profile_export_jsonl_gz as blob');
+  });
+});
diff --git a/packages/db/src/queries/request-chart-data.ts b/packages/db/src/queries/request-chart-data.ts
new file mode 100644
index 000000000..7b6e4c042
--- /dev/null
+++ b/packages/db/src/queries/request-chart-data.ts
@@ -0,0 +1,181 @@
+/**
+ * Compact per-request data used by the default agentic point charts.
+ *
+ * The full request timeline contains source-trace metadata and Gantt-only
+ * timing fields. Large runs can make that document tens of megabytes, even
+ * though the summary charts need only nine values per request. This query
+ * projects those values once in Postgres and dictionary-encodes repeated
+ * strings before the response crosses the network.
+ */
+
+import {
+  REQUEST_TIMELINE_VERSION,
+  type RequestRecord,
+  type RequestTimeline,
+} from '../etl/compute-request-timeline';
+
+import type { DbClient } from '../connection.js';
+import { getRequestTimeline } from './request-timeline';
+
+const REQUEST_CHART_SHAPE_VERSION = 2;
+export const REQUEST_CHART_DATA_VERSION =
+  REQUEST_TIMELINE_VERSION * 100 + REQUEST_CHART_SHAPE_VERSION;
+
+export type RequestChartTuple = [
+  cid: number,
+  phase: number,
+  startUs: number,
+  endUs: number,
+  ttftMs: number | null,
+  tpotMs: number | null,
+  isl: number | null,
+  osl: number | null,
+  cancelled: 0 | 1,
+];
+
+export interface RequestChartDataWire {
+  version: number;
+  timelineVersion: number;
+  startNs: number;
+  endNs: number;
+  durationS: number;
+  cids: string[];
+  phases: string[];
+  requests: RequestChartTuple[];
+}
+
+interface RawChartRow {
+  trace_replay_id: number;
+  has_blob: boolean;
+  timeline_version: number | null;
+  start_ns: number | null;
+  end_ns: number | null;
+  duration_s: number | null;
+  requests: [
+    cid: string,
+    phase: string,
+    start: number,
+    end: number,
+    ttftMs: number | null,
+    tpotMs: number | null,
+    isl: number | null,
+    osl: number | null,
+    cancelled: boolean,
+  ][];
+}
+
+type ChartRecord = Pick<
+  RequestRecord,
+  'cid' | 'phase' | 'start' | 'end' | 'ttftMs' | 'tpotMs' | 'isl' | 'osl' | 'cancelled'
+>;
+
+function dictionaryIndex(value: string, values: string[], indices: Map<string, number>): number {
+  const existing = indices.get(value);
+  if (existing !== undefined) return existing;
+  const index = values.length;
+  values.push(value);
+  indices.set(value, index);
+  return index;
+}
+
+const nanosecondsToMicroseconds = (value: number): number => Math.round(value / 1_000);
+const roundMilliseconds = (value: number | null): number | null =>
+  value === null ? null : Math.round(value * 1_000) / 1_000;
+
+export function encodeRequestChartData(
+  timeline: Pick<RequestTimeline, 'version' | 'startNs' | 'endNs' | 'durationS'>,
+  records: readonly ChartRecord[],
+): RequestChartDataWire {
+  const cids: string[] = [];
+  const phases: string[] = [];
+  const cidIndices = new Map<string, number>();
+  const phaseIndices = new Map<string, number>();
+  const requests: RequestChartTuple[] = [];
+
+  for (const record of records) {
+    requests.push([
+      dictionaryIndex(record.cid, cids, cidIndices),
+      dictionaryIndex(record.phase, phases, phaseIndices),
+      nanosecondsToMicroseconds(record.start),
+      nanosecondsToMicroseconds(record.end),
+      roundMilliseconds(record.ttftMs),
+      roundMilliseconds(record.tpotMs),
+      record.isl,
+      record.osl,
+      record.cancelled ? 1 : 0,
+    ]);
+  }
+
+  return {
+    version: REQUEST_CHART_DATA_VERSION,
+    timelineVersion: timeline.version,
+    startNs: timeline.startNs,
+    endNs: timeline.endNs,
+    durationS: timeline.durationS,
+    cids,
+    phases,
+    requests,
+  };
+}
+
+export async function getRequestChartData(
+  sql: DbClient,
+  benchmarkResultId: number,
+): Promise<RequestChartDataWire | null> {
+  const rows = (await sql`
+    select
+      atr.id as trace_replay_id,
+      (atr.profile_export_jsonl_gz is not null) as has_blob,
+      (atr.request_timeline->>'version')::int as timeline_version,
+      (atr.request_timeline->>'startNs')::double precision as start_ns,
+      (atr.request_timeline->>'endNs')::double precision as end_ns,
+      (atr.request_timeline->>'durationS')::double precision as duration_s,
+      coalesce((
+        select jsonb_agg(jsonb_build_array(
+          request->>'cid',
+          request->>'phase',
+          (request->>'start')::double precision,
+          (request->>'end')::double precision,
+          (request->>'ttftMs')::double precision,
+          (request->>'tpotMs')::double precision,
+          (request->>'isl')::int,
+          (request->>'osl')::int,
+          coalesce((request->>'cancelled')::boolean, false)
+        ) order by ordinal)
+        from jsonb_array_elements(atr.request_timeline->'requests')
+          with ordinality as expanded(request, ordinal)
+      ), '[]'::jsonb) as requests
+    from benchmark_results br
+    join agentic_trace_replay atr on atr.id = br.trace_replay_id
+    where br.id = ${benchmarkResultId}
+  `) as unknown as RawChartRow[];
+  const row = rows[0];
+  if (!row) return null;
+
+  if (row.timeline_version !== null && Number(row.timeline_version) === REQUEST_TIMELINE_VERSION) {
+    const records: ChartRecord[] = row.requests.map((request) => ({
+      cid: request[0],
+      phase: request[1],
+      start: Number(request[2]),
+      end: Number(request[3]),
+      ttftMs: request[4] === null ? null : Number(request[4]),
+      tpotMs: request[5] === null ? null : Number(request[5]),
+      isl: request[6] === null ? null : Number(request[6]),
+      osl: request[7] === null ? null : Number(request[7]),
+      cancelled: Boolean(request[8]),
+    }));
+    return encodeRequestChartData(
+      {
+        version: Number(row.timeline_version),
+        startNs: Number(row.start_ns ?? 0),
+        endNs: Number(row.end_ns ?? 0),
+        durationS: Number(row.duration_s ?? 0),
+      },
+      records,
+    );
+  }
+
+  if (!row.has_blob) return null;
+  const timeline = await getRequestTimeline(sql, benchmarkResultId);
+  return timeline ? encodeRequestChartData(timeline, timeline.requests) : null;
+}
diff --git a/packages/db/src/queries/request-timeline.test.ts b/packages/db/src/queries/request-timeline.test.ts
index 1f1d58a56..558fe6ecc 100644
--- a/packages/db/src/queries/request-timeline.test.ts
+++ b/packages/db/src/queries/request-timeline.test.ts
@@ -26,19 +26,58 @@ const timeline: RequestTimeline = {
 };
 
 describe('getRequestTimeline', () => {
-  it('returns the current precomputed timeline without selecting the raw profile blob', async () => {
+  it('returns the current precomputed timeline from a bounded text read without selecting the raw profile blob', async () => {
+    const encoded = JSON.stringify(timeline);
     const { sql, calls } = mockSql([
-      [{ trace_replay_id: 870, has_blob: true, request_timeline: timeline }],
+      [
+        {
+          trace_replay_id: 870,
+          has_blob: true,
+          timeline_version: REQUEST_TIMELINE_VERSION,
+        },
+      ],
+      [{ chunk: encoded, chunk_chars: encoded.length }],
     ]);
 
     await expect(getRequestTimeline(sql, 422991)).resolves.toEqual(timeline);
-    expect(calls).toHaveLength(1);
-    expect(calls[0]).not.toContain('profile_export_jsonl_gz as blob');
+    expect(calls).toHaveLength(2);
+    expect(calls[1]).not.toContain('profile_export_jsonl_gz as blob');
+  });
+
+  it('never uses compressed JSONB storage size to decide whether to read bounded chunks', async () => {
+    const encoded = JSON.stringify(timeline);
+    const { sql, calls } = mockSql([
+      [
+        {
+          trace_replay_id: 870,
+          has_blob: true,
+          timeline_version: REQUEST_TIMELINE_VERSION,
+        },
+      ],
+      [{ chunk: encoded, chunk_chars: encoded.length }],
+    ]);
+
+    await expect(getRequestTimeline(sql, 422991)).resolves.toEqual(timeline);
+    expect(calls).toHaveLength(2);
+    expect(calls[0]).not.toContain('pg_column_size');
+    expect(calls[0]).not.toContain('octet_length');
+    expect(calls[1]).toContain('substr(text');
+    expect(calls[1]).not.toContain('profile_export_jsonl_gz as blob');
   });
 
   it('does not fetch a blob when neither a current timeline nor a blob exists', async () => {
     const { sql, calls } = mockSql([
-      [{ trace_replay_id: 870, has_blob: false, request_timeline: null }],
+      [
+        {
+          trace_replay_id: 870,
+          has_blob: false,
+          timeline_version: null,
+          start_ns: null,
+          end_ns: null,
+          duration_s: null,
+          request_count: null,
+        },
+      ],
     ]);
 
     await expect(getRequestTimeline(sql, 422991)).resolves.toBeNull();
@@ -66,11 +105,16 @@ describe('getRequestTimeline', () => {
         }),
       ),
     );
-    const stale = { ...timeline, version: REQUEST_TIMELINE_VERSION - 1 };
-    const { sql, calls } = mockSql([
-      [{ trace_replay_id: 870, has_blob: true, request_timeline: stale }],
-      [{ blob }],
-    ]);
+    const staleHeader = {
+      trace_replay_id: 870,
+      has_blob: true,
+      timeline_version: REQUEST_TIMELINE_VERSION - 1,
+      start_ns: 100,
+      end_ns: 200,
+      duration_s: 0.0000001,
+      request_count: 0,
+    };
+    const { sql, calls } = mockSql([[staleHeader], [{ blob }]]);
 
     const result = await getRequestTimeline(sql, 422991);
 
@@ -84,11 +128,16 @@ describe('getRequestTimeline', () => {
   });
 
   it('does not write back when the blob is missing (never persists a null timeline)', async () => {
-    const stale = { ...timeline, version: REQUEST_TIMELINE_VERSION - 1 };
-    const { sql, calls } = mockSql([
-      [{ trace_replay_id: 870, has_blob: true, request_timeline: stale }],
-      [{ blob: null }],
-    ]);
+    const staleHeader = {
+      trace_replay_id: 870,
+      has_blob: true,
+      timeline_version: REQUEST_TIMELINE_VERSION - 1,
+      start_ns: 100,
+      end_ns: 200,
+      duration_s: 0.0000001,
+      request_count: 0,
+    };
+    const { sql, calls } = mockSql([[staleHeader], [{ blob: null }]]);
 
     await expect(getRequestTimeline(sql, 422991)).resolves.toBeNull();
     // meta read + blob read only — no write-back for a null recompute.
diff --git a/packages/db/src/queries/request-timeline.ts b/packages/db/src/queries/request-timeline.ts
index 9b7cc4b56..cc5a27cbe 100644
--- a/packages/db/src/queries/request-timeline.ts
+++ b/packages/db/src/queries/request-timeline.ts
@@ -2,8 +2,9 @@
  * Per-request timeline for the agentic detail page's Gantt view.
  *
  * Backed by `agentic_trace_replay.request_timeline` (pre-computed at
- * ingest time, see `etl/compute-request-timeline.ts`). The fast path is
- * a single SQL row read; the slow path re-computes from
+ * ingest time, see `etl/compute-request-timeline.ts`). Current payloads are
+ * read as bounded text chunks so Neon's 64 MiB encoded-response limit cannot
+ * reject even highly compressible JSONB. The slow path re-computes from
  * `profile_export_jsonl_gz` and is only taken when the column is missing
  * or the stored `REQUEST_TIMELINE_VERSION` is stale.
  */
@@ -22,13 +23,61 @@ export type { RequestTimeline, RequestRecord } from '../etl/compute-request-time
 interface RawMetaRow {
   trace_replay_id: number;
   has_blob: boolean;
-  request_timeline: RequestTimeline | null;
+  timeline_version: number | null;
 }
 
 interface RawBlobRow {
   blob: Buffer | null;
 }
 
+interface RawTimelineChunkRow {
+  chunk: string | null;
+  chunk_chars: number;
+}
+
+// JSONB can be far smaller on disk than its serialized response due to TOAST
+// compression, so every current timeline uses bounded chunks. Timeline fields
+// are compact identifiers/numbers; 16 MiB leaves ample room for JSON-string
+// escaping under Neon's 64 MiB per-query response cap.
+const TIMELINE_TEXT_CHUNK_CHARS = 16 * 1024 * 1024;
+const MAX_TIMELINE_TEXT_CHUNKS = 64;
+
+async function readChunkedRequestTimeline(
+  sql: DbClient,
+  traceReplayId: number,
+): Promise<RequestTimeline | null> {
+  const pieces: string[] = [];
+  for (let part = 0; part < MAX_TIMELINE_TEXT_CHUNKS; part += 1) {
+    const offset = part * TIMELINE_TEXT_CHUNK_CHARS;
+    const chunkRows = (await sql`
+      with payload as materialized (
+        select request_timeline::text as text
+        from agentic_trace_replay
+        where id = ${traceReplayId}
+      ),
+      piece as materialized (
+        select substr(text, ${offset + 1}::int, ${TIMELINE_TEXT_CHUNK_CHARS}::int) as chunk
+        from payload
+      )
+      select
+        chunk,
+        char_length(chunk)::int as chunk_chars
+      from piece
+    `) as unknown as RawTimelineChunkRow[];
+    const row = chunkRows[0];
+    if (!row || row.chunk === null) return null;
+    pieces.push(row.chunk);
+    if (row.chunk_chars < TIMELINE_TEXT_CHUNK_CHARS) {
+      try {
+        return JSON.parse(pieces.join('')) as RequestTimeline;
+      } catch {
+        return null;
+      }
+    }
+  }
+  return null;
+}
+
 export async function getRequestTimeline(
   sql: DbClient,
   benchmarkResultId: number,
@@ -37,7 +86,7 @@ export async function getRequestTimeline(
     select
       atr.id as trace_replay_id,
       (atr.profile_export_jsonl_gz is not null) as has_blob,
-      atr.request_timeline
+      (atr.request_timeline->>'version')::int as timeline_version
     from benchmark_results br
     join agentic_trace_replay atr on atr.id = br.trace_replay_id
     where br.id = ${benchmarkResultId}
@@ -45,9 +94,11 @@ export async function getRequestTimeline(
   const row = rows[0];
   if (!row) return null;
 
-  // Fast path: pre-computed timeline at the current version.
-  if (row.request_timeline && Number(row.request_timeline.version) === REQUEST_TIMELINE_VERSION) {
-    return row.request_timeline;
+  // Read every current pre-computed timeline through bounded text queries.
+  // Choosing from compressed storage size is unsafe: a small TOAST value can
+  // still serialize past Neon's per-query response limit.
+  if (Number(row.timeline_version) === REQUEST_TIMELINE_VERSION) {
+    return readChunkedRequestTimeline(sql, row.trace_replay_id);
   }
 
   if (!row.has_blob) return null;
@@ -61,7 +112,7 @@ export async function getRequestTimeline(
     from agentic_trace_replay
     where id = ${row.trace_replay_id}
   `) as unknown as RawBlobRow[];
-  const timeline = computeRequestTimeline(blobRows[0]?.blob ?? null);
+  const timeline = await computeRequestTimeline(blobRows[0]?.blob ?? null);
 
   // Self-heal the stored request_timeline so the next request (and the
   // trace-histograms route, which reads the same column) takes the fast path.
diff --git a/packages/db/src/queries/resident-sequence-lengths.test.ts b/packages/db/src/queries/resident-sequence-lengths.test.ts
new file mode 100644
index 000000000..a341ba45e
--- /dev/null
+++ b/packages/db/src/queries/resident-sequence-lengths.test.ts
@@ -0,0 +1,47 @@
+import { describe, expect, it } from 'vitest';
+
+import { buildTokenLengthSketch } from '@semianalysisai/inferencex-constants';
+
+import type { DbClient } from '../connection.js';
+import { STATS_VERSION } from './agentic-shared';
+import { getResidentSequenceLengthSketches } from './resident-sequence-lengths.js';
+
+function mockSql(rows: unknown[]): DbClient {
+  return (() => Promise.resolve(rows)) as unknown as DbClient;
+}
+
+describe('getResidentSequenceLengthSketches', () => {
+  it('merges current sketches and omits stale rows without a blob fallback', async () => {
+    const result = await getResidentSequenceLengthSketches(
+      mockSql([
+        {
+          benchmark_result_id: 1,
+          stats: {
+            version: STATS_VERSION,
+            sequenceLengths: {
+              isl: buildTokenLengthSketch([10, 20]),
+              osl: buildTokenLengthSketch([1, 2]),
+            },
+          },
+        },
+        {
+          benchmark_result_id: 2,
+          stats: {
+            version: STATS_VERSION,
+            sequenceLengths: {
+              isl: buildTokenLengthSketch([1_000]),
+              osl: buildTokenLengthSketch([100]),
+            },
+          },
+        },
+        { benchmark_result_id: 3, stats: { version: STATS_VERSION - 1 } },
+      ]),
+      [1, 2, 3],
+    );
+
+    expect(result.coveredPoints).toBe(2);
+    expect(result.requestedPoints).toBe(3);
+    expect(result.isl?.n).toBe(3);
+    expect(result.osl?.n).toBe(3);
+  });
+});
diff --git a/packages/db/src/queries/resident-sequence-lengths.ts b/packages/db/src/queries/resident-sequence-lengths.ts
new file mode 100644
index 000000000..45e3e365c
--- /dev/null
+++ b/packages/db/src/queries/resident-sequence-lengths.ts
@@ -0,0 +1,52 @@
+import {
+  mergeTokenLengthSketches,
+  type TokenLengthSketch,
+} from '@semianalysisai/inferencex-constants';
+
+import type { DbClient } from '../connection.js';
+import { fetchAggregateStatsRows, STATS_VERSION } from './agentic-shared';
+
+interface StoredStats {
+  version: number;
+  sequenceLengths?: {
+    isl: TokenLengthSketch | null;
+    osl: TokenLengthSketch | null;
+  };
+}
+
+export interface ResidentSequenceLengthSketches {
+  isl: TokenLengthSketch | null;
+  osl: TokenLengthSketch | null;
+  /** Resident benchmark points represented by current-version sketches. */
+  coveredPoints: number;
+  requestedPoints: number;
+}
+
+/**
+ * Merge the precomputed request-length sketches for a bounded set of chart
+ * points. There is deliberately no raw-blob fallback here: a stale row is
+ * omitted until the versioned backfill reaches it rather than making a chart
+ * load decompress multi-megabyte request profiles.
+ */
+export async function getResidentSequenceLengthSketches(
+  sql: DbClient,
+  benchmarkResultIds: number[],
+): Promise<ResidentSequenceLengthSketches> {
+  if (benchmarkResultIds.length === 0) {
+    return { isl: null, osl: null, coveredPoints: 0, requestedPoints: 0 };
+  }
+
+  const rows = await fetchAggregateStatsRows<StoredStats>(sql, benchmarkResultIds);
+  const current = rows.filter(
+    (row) =>
+      row.stats?.version === STATS_VERSION &&
+      Boolean(row.stats.sequenceLengths?.isl || row.stats.sequenceLengths?.osl),
+  );
+
+  return {
+    isl: mergeTokenLengthSketches(current.map((row) => row.stats?.sequenceLengths?.isl)),
+    osl: mergeTokenLengthSketches(current.map((row) => row.stats?.sequenceLengths?.osl)),
+    coveredPoints: current.length,
+    requestedPoints: benchmarkResultIds.length,
+  };
+}
diff --git a/packages/db/src/queries/server-logs.test.ts b/packages/db/src/queries/server-logs.test.ts
new file mode 100644
index 000000000..41478fa61
--- /dev/null
+++ b/packages/db/src/queries/server-logs.test.ts
@@ -0,0 +1,221 @@
+import { describe, expect, it, vi } from 'vitest';
+
+import type { DbClient } from '../connection.js';
+import {
+  escapePostgresRegexLiteral,
+  getServerLogAvailability,
+  getServerLogChunk,
+  getServerLogFileNames,
+  searchServerLogs,
+} from './server-logs.js';
+
+function mockSql(rows: unknown[]) {
+  const sql = vi.fn(() => Promise.resolve(rows)) as unknown as DbClient;
+  return { sql, call: sql as unknown as ReturnType<typeof vi.fn> };
+}
+
+describe('getServerLogChunk', () => {
+  it('returns a bounded chunk and a continuation offset', async () => {
+    const { sql, call } = mockSql([
+      { file_name: 'results/server.log', log_text: 'abc', char_count: 3, has_more: true },
+    ]);
+
+    await expect(getServerLogChunk(sql, 42, 10, 3)).resolves.toEqual({
+      fileName: 'results/server.log',
+      serverLog: 'abc',
+      offset: 10,
+      nextOffset: 13,
+    });
+    expect(call).toHaveBeenCalledOnce();
+    expect(call.mock.calls[0]?.slice(1)).toEqual(expect.arrayContaining([42, 11, 3, 14]));
+    expect(call.mock.calls[0]?.[0].join('')).toContain('from ::integer');
+    expect(call.mock.calls[0]?.[0].join('')).toContain('for ::integer');
+    expect(call.mock.calls[0]?.[0].join('')).not.toContain('sl.server_log as log_text');
+  });
+
+  it('marks an exact-size final chunk as complete', async () => {
+    const { sql } = mockSql([
+      { file_name: 'router.log', log_text: 'abc', char_count: 3, has_more: false },
+    ]);
+    await expect(getServerLogChunk(sql, 42, 0, 3)).resolves.toEqual({
+      fileName: 'router.log',
+      serverLog: 'abc',
+      offset: 0,
+      nextOffset: null,
+    });
+  });
+
+  it('returns null when the benchmark has no linked log', async () => {
+    const { sql } = mockSql([]);
+    await expect(getServerLogChunk(sql, 42, 0, 3)).resolves.toBeNull();
+  });
+
+  it('selects an additional file by its artifact-relative name', async () => {
+    const { sql, call } = mockSql([
+      { file_name: 'workers/decode.out', log_text: 'worker', char_count: 6, has_more: false },
+    ]);
+    await expect(getServerLogChunk(sql, 42, 0, 64, 'workers/decode.out')).resolves.toMatchObject({
+      fileName: 'workers/decode.out',
+      serverLog: 'worker',
+    });
+    expect(call.mock.calls[0]?.slice(1)).toContain('workers/decode.out');
+  });
+
+  it('advances offsets in PostgreSQL characters instead of UTF-16 code units', async () => {
+    const { sql } = mockSql([
+      { file_name: 'unicode.log', log_text: 'a😀b', char_count: 3, has_more: true },
+    ]);
+    await expect(getServerLogChunk(sql, 42, 5, 3, 'unicode.log')).resolves.toMatchObject({
+      serverLog: 'a😀b',
+      nextOffset: 8,
+    });
+  });
+});
+
+describe('getServerLogFileNames', () => {
+  it('returns primary and additional filenames in query order', async () => {
+    const { sql } = mockSql([
+      { file_name: 'results/server.log' },
+      { file_name: 'results/benchmark.log' },
+      { file_name: 'results/router.log' },
+    ]);
+    await expect(getServerLogFileNames(sql, 42)).resolves.toEqual([
+      'results/server.log',
+      'results/benchmark.log',
+      'results/router.log',
+    ]);
+  });
+});
+
+describe('searchServerLogs', () => {
+  it('escapes regex syntax so searches remain literal', () => {
+    expect(escapePostgresRegexLiteral('error[42] + retry?')).toBe(
+      String.raw`error\[42\] \+ retry\?`,
+    );
+  });
+
+  it('maps bounded matches across files and reports truncation', async () => {
+    const call = vi
+      .fn()
+      .mockResolvedValueOnce([{ file_name: 'results/router.log' }])
+      .mockResolvedValueOnce([
+        {
+          file_name: 'results/router.log',
+          match_position: 15,
+          before_text: 'INFO router: ',
+          match_text: 'Ready',
+          after_text: ' for requests',
+          char_count: 1024,
+          has_more: true,
+        },
+      ])
+      .mockResolvedValueOnce([
+        {
+          file_name: 'results/router.log',
+          match_position: 68,
+          before_text: 'worker ',
+          match_text: 'ready',
+          after_text: '',
+          char_count: 1024,
+          has_more: true,
+        },
+      ]);
+    const sql = call as unknown as DbClient;
+
+    await expect(searchServerLogs(sql, 42, 'ready', 1)).resolves.toEqual({
+      matches: [
+        {
+          fileName: 'results/router.log',
+          offset: 14,
+          before: 'INFO router: ',
+          match: 'Ready',
+          after: ' for requests',
+        },
+      ],
+      truncated: true,
+    });
+    expect(call.mock.calls[1]?.[0].join('')).toContain('regexp_instr(chunk_text');
+    expect(call.mock.calls[1]?.[0].join('')).toContain('substring(sl.server_log');
+  });
+
+  it('returns an empty non-truncated result when no file matches', async () => {
+    const call = vi
+      .fn()
+      .mockResolvedValueOnce([{ file_name: 'results/router.log' }])
+      .mockResolvedValueOnce([
+        {
+          file_name: 'results/router.log',
+          match_position: 0,
+          before_text: '',
+          match_text: '',
+          after_text: '',
+          char_count: 128,
+          has_more: false,
+        },
+      ]);
+    const sql = call as unknown as DbClient;
+    await expect(searchServerLogs(sql, 42, 'missing', 50)).resolves.toEqual({
+      matches: [],
+      truncated: false,
+    });
+  });
+
+  it('overlaps slices so a match crossing a chunk boundary is not skipped', async () => {
+    const call = vi
+      .fn()
+      .mockResolvedValueOnce([{ file_name: 'results/router.log' }])
+      .mockResolvedValueOnce([
+        {
+          file_name: 'results/router.log',
+          match_position: 0,
+          before_text: '',
+          match_text: '',
+          after_text: '',
+          char_count: 100,
+          has_more: true,
+        },
+      ])
+      .mockResolvedValueOnce([
+        {
+          file_name: 'results/router.log',
+          match_position: 3,
+          before_text: 'x',
+          match_text: 'ready',
+          after_text: ' now',
+          char_count: 20,
+          has_more: false,
+        },
+      ])
+      .mockResolvedValueOnce([]);
+    const sql = call as unknown as DbClient;
+
+    await expect(searchServerLogs(sql, 42, 'ready', 50)).resolves.toEqual({
+      matches: [
+        {
+          fileName: 'results/router.log',
+          offset: 98,
+          before: 'x',
+          match: 'ready',
+          after: ' now',
+        },
+      ],
+      truncated: false,
+    });
+  });
+});
+
+describe('getServerLogAvailability', () => {
+  it('returns an id-keyed presence map', async () => {
+    const { sql } = mockSql([{ id: 9 }, { id: 4 }]);
+    await expect(getServerLogAvailability(sql, [4, 9, 12])).resolves.toEqual({
+      4: true,
+      9: true,
+    });
+  });
+
+  it('does not query for an empty id set', async () => {
+    const { sql, call } = mockSql([]);
+    await expect(getServerLogAvailability(sql, [])).resolves.toEqual({});
+    expect(call).not.toHaveBeenCalled();
+  });
+});
diff --git a/packages/db/src/queries/server-logs.ts b/packages/db/src/queries/server-logs.ts
index 2033dd862..91b00a470 100644
--- a/packages/db/src/queries/server-logs.ts
+++ b/packages/db/src/queries/server-logs.ts
@@ -1,17 +1,369 @@
 import type { DbClient } from '../connection.js';
 
-/**
- * Fetch a server log by benchmark_result_id. Returns null if not found.
- */
+/** Map of `benchmark_results.id` → true for each id with a linked log bundle. */
+export type ServerLogAvailabilityMap = Record<number, true>;
+
+export interface ServerLogChunk {
+  fileName: string;
+  serverLog: string;
+  offset: number;
+  nextOffset: number | null;
+}
+
+export interface ServerLogSearchMatch {
+  fileName: string;
+  offset: number;
+  before: string;
+  match: string;
+  after: string;
+}
+
+export interface ServerLogSearchResult {
+  matches: ServerLogSearchMatch[];
+  truncated: boolean;
+}
+
+const isPreLogFilesSchema = (error: unknown): boolean => {
+  const code = (error as { code?: unknown })?.code;
+  return code === '42P01' || code === '42703';
+};
+
+/** Escape user text for PostgreSQL's POSIX regex engine. */
+export const escapePostgresRegexLiteral = (value: string): string =>
+  value.replaceAll(/[\\^$.*+?()[\]{}|]/gu, String.raw`\$&`);
+
+/** Fetch a complete log file. Omit fileName for the primary/legacy stream. */
 export async function getServerLog(
   sql: DbClient,
   benchmarkResultId: number,
+  fileName?: string,
 ): Promise<string | null> {
+  try {
+    const requestedFile = fileName ?? null;
+    const rows = (await sql`
+      with selected_log as (
+        select sl.file_name, sl.server_log as log_text, 0 as priority
+        from benchmark_results br
+        join server_logs sl on sl.id = br.server_log_id
+        where br.id = ${benchmarkResultId}
+          and (${requestedFile}::text is null or sl.file_name = ${requestedFile})
+        union all
+        select slf.file_name, slf.log_text, 1 as priority
+        from benchmark_results br
+        join server_log_files slf on slf.server_log_id = br.server_log_id
+        where br.id = ${benchmarkResultId}
+          and ${requestedFile}::text is not null
+          and slf.file_name = ${requestedFile}
+        order by priority
+        limit 1
+      )
+      select log_text from selected_log
+    `) as { log_text: string }[];
+    return rows[0]?.log_text ?? null;
+  } catch (error) {
+    if (!isPreLogFilesSchema(error)) throw error;
+    if (fileName && fileName !== 'server.log') return null;
+    const rows = (await sql`
+      select sl.server_log
+      from benchmark_results br
+      join server_logs sl on sl.id = br.server_log_id
+      where br.id = ${benchmarkResultId}
+    `) as { server_log: string }[];
+    return rows[0]?.server_log ?? null;
+  }
+}
+
+/** List the artifact-relative filenames for one benchmark result, primary first. */
+export async function getServerLogFileNames(
+  sql: DbClient,
+  benchmarkResultId: number,
+): Promise<string[] | null> {
+  try {
+    const rows = (await sql`
+      select file_name
+      from (
+        select sl.file_name, 0 as priority
+        from benchmark_results br
+        join server_logs sl on sl.id = br.server_log_id
+        where br.id = ${benchmarkResultId}
+        union all
+        select slf.file_name, 1 as priority
+        from benchmark_results br
+        join server_logs sl on sl.id = br.server_log_id
+        join server_log_files slf on slf.server_log_id = sl.id
+        where br.id = ${benchmarkResultId}
+          and slf.file_name <> sl.file_name
+      ) files
+      order by priority, file_name
+    `) as { file_name: string }[];
+    return rows.length > 0 ? rows.map((row) => row.file_name) : null;
+  } catch (error) {
+    if (!isPreLogFilesSchema(error)) throw error;
+    return (await getServerLog(sql, benchmarkResultId)) === null ? null : ['server.log'];
+  }
+}
+
+/**
+ * Fetch a bounded character range from one stored log file.
+ *
+ * Some production logs are hundreds of megabytes, so the UI must never read
+ * the complete PostgreSQL text value just to render the first screen. Reading
+ * one extra character reports whether another chunk exists without computing
+ * the full log length.
+ */
+export async function getServerLogChunk(
+  sql: DbClient,
+  benchmarkResultId: number,
+  offset: number,
+  limit: number,
+  fileName?: string,
+): Promise<ServerLogChunk | null> {
+  try {
+    const requestedFile = fileName ?? null;
+    const rows = (await sql`
+      with selected_log as (
+        select
+          sl.file_name,
+          substring(
+            sl.server_log
+            from ${offset + 1}::integer
+            for ${limit}::integer
+          ) as log_text,
+          substring(sl.server_log from ${offset + limit + 1}::integer for 1) <> '' as has_more,
+          0 as priority
+        from benchmark_results br
+        join server_logs sl on sl.id = br.server_log_id
+        where br.id = ${benchmarkResultId}
+          and (${requestedFile}::text is null or sl.file_name = ${requestedFile})
+        union all
+        select
+          slf.file_name,
+          substring(
+            slf.log_text
+            from ${offset + 1}::integer
+            for ${limit}::integer
+          ) as log_text,
+          substring(slf.log_text from ${offset + limit + 1}::integer for 1) <> '' as has_more,
+          1 as priority
+        from benchmark_results br
+        join server_log_files slf on slf.server_log_id = br.server_log_id
+        where br.id = ${benchmarkResultId}
+          and ${requestedFile}::text is not null
+          and slf.file_name = ${requestedFile}
+        order by priority
+        limit 1
+      )
+      select file_name, log_text, char_length(log_text) as char_count, has_more
+      from selected_log
+    `) as { file_name: string; log_text: string; char_count: number; has_more: boolean }[];
+    const row = rows[0];
+    if (!row) return null;
+    return {
+      fileName: row.file_name,
+      serverLog: row.log_text,
+      offset,
+      nextOffset: row.has_more ? offset + Number(row.char_count) : null,
+    };
+  } catch (error) {
+    if (!isPreLogFilesSchema(error)) throw error;
+    if (fileName && fileName !== 'server.log') return null;
+    const rows = (await sql`
+      with selected_log as (
+        select
+          substring(
+            sl.server_log
+            from ${offset + 1}::integer
+            for ${limit}::integer
+          ) as server_log,
+          substring(sl.server_log from ${offset + limit + 1}::integer for 1) <> '' as has_more
+        from benchmark_results br
+        join server_logs sl on sl.id = br.server_log_id
+        where br.id = ${benchmarkResultId}
+      )
+      select server_log, char_length(server_log) as char_count, has_more
+      from selected_log
+    `) as { server_log: string; char_count: number; has_more: boolean }[];
+    const row = rows[0];
+    if (!row) return null;
+    return {
+      fileName: 'server.log',
+      serverLog: row.server_log,
+      offset,
+      nextOffset: row.has_more ? offset + Number(row.char_count) : null,
+    };
+  }
+}
+
+/** Search every stored file without transferring complete logs out of PostgreSQL. */
+export async function searchServerLogs(
+  sql: DbClient,
+  benchmarkResultId: number,
+  searchTerm: string,
+  limit: number,
+): Promise<ServerLogSearchResult> {
+  try {
+    interface SearchRow extends Record<string, unknown> {
+      file_name: string;
+      match_position: number;
+      before_text: string;
+      match_text: string;
+      after_text: string;
+      char_count: number;
+      has_more: boolean;
+    }
+    const chunkSize = 16 * 1024 * 1024;
+    const searchLength = [...searchTerm].length;
+    const overlap = Math.min(Math.max(0, searchLength - 1), 255);
+    const searchPattern = escapePostgresRegexLiteral(searchTerm);
+    const fileNames = (await getServerLogFileNames(sql, benchmarkResultId)) ?? [];
+    const matches: ServerLogSearchMatch[] = [];
+    for (const [fileIndex, fileName] of fileNames.entries()) {
+      let offset = 0;
+      while (matches.length <= limit) {
+        const rows = (fileIndex === 0
+          ? await sql`
+                with selected_chunk as materialized (
+                  select
+                    sl.file_name,
+                    substring(sl.server_log from ${offset + 1}::integer for ${chunkSize}::integer) as chunk_text,
+                    substring(sl.server_log from ${offset + chunkSize + 1}::integer for 1) <> '' as has_more
+                  from benchmark_results br
+                  join server_logs sl on sl.id = br.server_log_id
+                  where br.id = ${benchmarkResultId}
+                    and sl.file_name = ${fileName}
+                  limit 1
+                ), located_match as (
+                  select
+                    file_name,
+                    chunk_text,
+                    char_length(chunk_text) as char_count,
+                    has_more,
+                    regexp_instr(chunk_text, ${searchPattern}, 1, 1, 0, 'i') as match_position
+                  from selected_chunk
+                )
+                select
+                  file_name,
+                  match_position,
+                  case when match_position > 0 then substring(
+                    chunk_text
+                    from greatest(1, match_position - 120)
+                    for match_position - greatest(1, match_position - 120)
+                  ) else '' end as before_text,
+                  case when match_position > 0 then substring(
+                    chunk_text from match_position for ${searchLength}::integer
+                  ) else '' end as match_text,
+                  case when match_position > 0 then substring(
+                    chunk_text from match_position + ${searchLength}::integer for 120
+                  ) else '' end as after_text,
+                  char_count,
+                  has_more
+                from located_match
+              `
+          : await sql`
+                with selected_chunk as materialized (
+                  select
+                    slf.file_name,
+                    substring(slf.log_text from ${offset + 1}::integer for ${chunkSize}::integer) as chunk_text,
+                    substring(slf.log_text from ${offset + chunkSize + 1}::integer for 1) <> '' as has_more
+                  from benchmark_results br
+                  join server_log_files slf on slf.server_log_id = br.server_log_id
+                  where br.id = ${benchmarkResultId}
+                    and slf.file_name = ${fileName}
+                  limit 1
+                ), located_match as (
+                  select
+                    file_name,
+                    chunk_text,
+                    char_length(chunk_text) as char_count,
+                    has_more,
+                    regexp_instr(chunk_text, ${searchPattern}, 1, 1, 0, 'i') as match_position
+                  from selected_chunk
+                )
+                select
+                  file_name,
+                  match_position,
+                  case when match_position > 0 then substring(
+                    chunk_text
+                    from greatest(1, match_position - 120)
+                    for match_position - greatest(1, match_position - 120)
+                  ) else '' end as before_text,
+                  case when match_position > 0 then substring(
+                    chunk_text from match_position for ${searchLength}::integer
+                  ) else '' end as match_text,
+                  case when match_position > 0 then substring(
+                    chunk_text from match_position + ${searchLength}::integer for 120
+                  ) else '' end as after_text,
+                  char_count,
+                  has_more
+                from located_match
+              `) as unknown as SearchRow[];
+        const row = rows[0];
+        if (!row) break;
+        const matchPosition = Number(row.match_position);
+        const charCount = Number(row.char_count);
+        if (matchPosition > 0) {
+          const matchOffset = offset + matchPosition - 1;
+          matches.push({
+            fileName: row.file_name,
+            offset: matchOffset,
+            before: row.before_text,
+            match: row.match_text,
+            after: row.after_text,
+          });
+          offset = matchOffset + searchLength;
+          continue;
+        }
+        if (!row.has_more) break;
+        const advance = charCount - overlap;
+        if (advance <= 0) break;
+        offset += advance;
+      }
+      if (matches.length > limit) break;
+    }
+    return { matches: matches.slice(0, limit), truncated: matches.length > limit };
+  } catch (error) {
+    if (!isPreLogFilesSchema(error)) throw error;
+    const legacyLog = await getServerLog(sql, benchmarkResultId);
+    if (legacyLog === null) return { matches: [], truncated: false };
+
+    const matches: ServerLogSearchMatch[] = [];
+    const searchableLog = legacyLog.toLocaleLowerCase();
+    const needle = searchTerm.toLocaleLowerCase();
+    let offset = 0;
+    while (matches.length <= limit) {
+      const matchOffset = searchableLog.indexOf(needle, offset);
+      if (matchOffset === -1) break;
+      matches.push({
+        fileName: 'server.log',
+        offset: matchOffset,
+        before: legacyLog.slice(Math.max(0, matchOffset - 120), matchOffset),
+        match: legacyLog.slice(matchOffset, matchOffset + searchTerm.length),
+        after: legacyLog.slice(
+          matchOffset + searchTerm.length,
+          matchOffset + searchTerm.length + 120,
+        ),
+      });
+      offset = matchOffset + searchTerm.length;
+    }
+    return { matches: matches.slice(0, limit), truncated: matches.length > limit };
+  }
+}
+
+/** Lightweight bulk lookup used to decide whether a point exposes "View logs". */
+export async function getServerLogAvailability(
+  sql: DbClient,
+  benchmarkResultIds: number[],
+): Promise<ServerLogAvailabilityMap> {
+  if (benchmarkResultIds.length === 0) return {};
+
   const rows = (await sql`
-    select sl.server_log
+    select br.id
     from benchmark_results br
-    join server_logs sl on sl.id = br.server_log_id
-    where br.id = ${benchmarkResultId}
-  `) as { server_log: string }[];
-  return rows[0]?.server_log ?? null;
+    where br.id = any(${benchmarkResultIds}::bigint[])
+      and br.server_log_id is not null
+  `) as { id: number }[];
+
+  const result: ServerLogAvailabilityMap = {};
+  for (const row of rows) result[Number(row.id)] = true;
+  return result;
 }
diff --git a/packages/db/src/queries/trace-histograms.test.ts b/packages/db/src/queries/trace-histograms.test.ts
index c3c6ec8af..2b993b30a 100644
--- a/packages/db/src/queries/trace-histograms.test.ts
+++ b/packages/db/src/queries/trace-histograms.test.ts
@@ -1,6 +1,6 @@
 import { describe, expect, it } from 'vitest';
 
-import { REQUEST_TIMELINE_VERSION, type RequestTimeline } from '../etl/compute-request-timeline';
+import { REQUEST_TIMELINE_VERSION } from '../etl/compute-request-timeline';
 import type { DbClient } from '../connection.js';
 
 import { getTraceHistograms } from './trace-histograms';
@@ -15,55 +15,19 @@ function mockSql(queue: unknown[][]): { sql: DbClient; calls: string[] } {
   return { sql, calls };
 }
 
-const timeline: RequestTimeline = {
-  version: REQUEST_TIMELINE_VERSION,
-  startNs: 0,
-  endNs: 10,
-  durationS: 0.00000001,
-  requests: [
-    {
-      cid: 'session-1',
-      ti: 0,
-      wid: '0',
-      ad: 0,
-      phase: 'profiling',
-      credit: 0,
-      start: 1,
-      ack: 2,
-      end: 3,
-      ttftMs: 1,
-      tpotMs: 2,
-      isl: 4096,
-      osl: 512,
-      cancelled: false,
-    },
-    {
-      cid: 'session-1',
-      ti: 1,
-      wid: '0',
-      ad: 0,
-      phase: 'profiling',
-      credit: 4,
-      start: 5,
-      ack: 6,
-      end: 7,
-      ttftMs: 1,
-      tpotMs: 2,
-      isl: null,
-      osl: 128,
-      cancelled: false,
-    },
-  ],
-};
-
 describe('getTraceHistograms', () => {
   it('builds distributions from the precomputed timeline without selecting the raw blob', async () => {
+    // The query unnests isl/osl server-side, so the driver hands back arrays
+    // rather than the timeline document. Requests with a null isl contribute
+    // to osl only, which is why the two arrays differ in length.
     const { sql, calls } = mockSql([
       [
         {
           benchmark_result_id: 422991,
           trace_replay_id: 870,
-          request_timeline: timeline,
+          timeline_version: REQUEST_TIMELINE_VERSION,
+          isl: [4096],
+          osl: [512, 128],
           has_blob: true,
         },
       ],
@@ -74,5 +38,65 @@ describe('getTraceHistograms', () => {
     });
     expect(calls).toHaveLength(1);
     expect(calls[0]).not.toContain('profile_export_jsonl_gz as blob');
+    // Never ship the whole document — that is what blew the 64 MiB HTTP cap.
+    expect(calls[0]).not.toMatch(/atr\.request_timeline\s*,/);
+    expect(calls[0]).toContain('jsonb_array_elements');
+  });
+
+  it('falls back to the blob when the stored timeline is a stale version', async () => {
+    const { sql, calls } = mockSql([
+      [
+        {
+          benchmark_result_id: 1,
+          trace_replay_id: 2,
+          timeline_version: REQUEST_TIMELINE_VERSION - 1,
+          isl: [1],
+          osl: [2],
+          has_blob: true,
+        },
+      ],
+      [],
+    ]);
+    await expect(getTraceHistograms(sql, [1])).resolves.toEqual({});
+    expect(calls).toHaveLength(2);
+    expect(calls[1]).toContain('profile_export_jsonl_gz as blob');
+  });
+
+  it('yields empty arrays when a timeline has no isl/osl at all', async () => {
+    // array_agg over zero matching rows returns NULL, not an empty array.
+    const { sql } = mockSql([
+      [
+        {
+          benchmark_result_id: 3,
+          trace_replay_id: 4,
+          timeline_version: REQUEST_TIMELINE_VERSION,
+          isl: null,
+          osl: null,
+          has_blob: false,
+        },
+      ],
+    ]);
+    await expect(getTraceHistograms(sql, [3])).resolves.toEqual({
+      3: { id: 3, isl: [], osl: [] },
+    });
+  });
+
+  it('drops non-finite values the driver may hand back as strings', async () => {
+    const { sql } = mockSql([
+      [
+        {
+          benchmark_result_id: 5,
+          trace_replay_id: 6,
+          timeline_version: REQUEST_TIMELINE_VERSION,
+          // numeric columns arrive as strings on some drivers.
+          isl: ['4096', null, 'NaN'] as unknown as number[],
+          osl: ['512'] as unknown as number[],
+          has_blob: false,
+        },
+      ],
+    ]);
+    await expect(getTraceHistograms(sql, [5])).resolves.toEqual({
+      5: { id: 5, isl: [4096], osl: [512] },
+    });
   });
 });
diff --git a/packages/db/src/queries/trace-histograms.ts b/packages/db/src/queries/trace-histograms.ts
index 24b96c355..da4b92131 100644
--- a/packages/db/src/queries/trace-histograms.ts
+++ b/packages/db/src/queries/trace-histograms.ts
@@ -14,7 +14,7 @@
 
 import { gunzipSync } from 'node:zlib';
 
-import { REQUEST_TIMELINE_VERSION, type RequestTimeline } from '../etl/compute-request-timeline';
+import { REQUEST_TIMELINE_VERSION } from '../etl/compute-request-timeline';
 
 import type { DbClient } from '../connection.js';
 
@@ -38,18 +38,22 @@ const MAX_FALLBACK_BLOB_BYTES = 24 * 1024 * 1024;
 interface TimelineRow {
   benchmark_result_id: number;
   trace_replay_id: number;
-  request_timeline: RequestTimeline | null;
+  /** Version of the stored timeline, read WITHOUT shipping the document. */
+  timeline_version: number | null;
+  /** ISL per completed request, unnested server-side (see the query below). */
+  isl: (number | null)[] | null;
+  osl: (number | null)[] | null;
   has_blob: boolean;
 }
 
-function histogramFromTimeline(id: number, timeline: RequestTimeline): TraceHistogramPoint {
-  const isl: number[] = [];
-  const osl: number[] = [];
-  for (const request of timeline.requests) {
-    if (typeof request.isl === 'number' && Number.isFinite(request.isl)) isl.push(request.isl);
-    if (typeof request.osl === 'number' && Number.isFinite(request.osl)) osl.push(request.osl);
+/** Postgres hands back `number[]`; drop anything non-finite defensively. */
+function finiteNumbers(values: (number | null)[] | null): number[] {
+  const out: number[] = [];
+  for (const value of values ?? []) {
+    const n = Number(value);
+    if (value !== null && Number.isFinite(n)) out.push(n);
   }
-  return { id, isl, osl };
+  return out;
 }
 
 export async function getTraceHistograms(
@@ -62,11 +66,28 @@ export async function getTraceHistograms(
   const fallbackRows: TimelineRow[] = [];
   for (let i = 0; i < benchmarkResultIds.length; i += QUERY_CHUNK_SIZE) {
     const chunk = benchmarkResultIds.slice(i, i + QUERY_CHUNK_SIZE);
+    // Project the two arrays out in SQL instead of selecting
+    // `atr.request_timeline` whole. The timeline carries ~18 fields per
+    // request, so shipping the document to extract two integers per request
+    // moved 61 MB of JSON for a 151k-request point — past the Neon HTTP
+    // driver's 64 MiB response cap, which failed the whole request with
+    // `507 response is too large` and left the distribution charts empty on
+    // exactly the biggest runs. Unnesting server-side sends ~600 KB instead.
     const chunkRows = (await sql`
       select
         br.id as benchmark_result_id,
         atr.id as trace_replay_id,
-        atr.request_timeline,
+        (atr.request_timeline->>'version')::int as timeline_version,
+        (
+          select array_agg((r->>'isl')::numeric order by ord)
+          from jsonb_array_elements(atr.request_timeline->'requests') with ordinality x(r, ord)
+          where jsonb_typeof(r->'isl') = 'number'
+        ) as isl,
+        (
+          select array_agg((r->>'osl')::numeric order by ord)
+          from jsonb_array_elements(atr.request_timeline->'requests') with ordinality x(r, ord)
+          where jsonb_typeof(r->'osl') = 'number'
+        ) as osl,
         (atr.profile_export_jsonl_gz is not null) as has_blob
       from benchmark_results br
       join agentic_trace_replay atr on atr.id = br.trace_replay_id
@@ -75,10 +96,10 @@ export async function getTraceHistograms(
     for (const row of chunkRows) {
       const id = Number(row.benchmark_result_id);
       if (
-        row.request_timeline &&
-        Number(row.request_timeline.version) === REQUEST_TIMELINE_VERSION
+        row.timeline_version !== null &&
+        Number(row.timeline_version) === REQUEST_TIMELINE_VERSION
       ) {
-        result[id] = histogramFromTimeline(id, row.request_timeline);
+        result[id] = { id, isl: finiteNumbers(row.isl), osl: finiteNumbers(row.osl) };
       } else if (row.has_blob) {
         fallbackRows.push(row);
       }
diff --git a/packages/db/src/queries/trace-server-metrics.test.ts b/packages/db/src/queries/trace-server-metrics.test.ts
index 50b969062..56e84d9ff 100644
--- a/packages/db/src/queries/trace-server-metrics.test.ts
+++ b/packages/db/src/queries/trace-server-metrics.test.ts
@@ -2,10 +2,14 @@ import { gzipSync } from 'node:zlib';
 
 import { describe, expect, it } from 'vitest';
 
-import { CHART_SERIES_VERSION, type ChartSeries } from '../etl/compute-chart-series';
+import {
+  CHART_SERIES_VERSION,
+  type ChartSeries,
+  type MetricSourceSeries,
+} from '../etl/compute-chart-series';
 import type { DbClient } from '../connection.js';
 
-import { getTraceServerMetrics } from './trace-server-metrics';
+import { getTraceServerMetrics, getTraceServerMetricSource } from './trace-server-metrics';
 
 function currentSeries(): ChartSeries {
   return {
@@ -33,6 +37,7 @@ function metaRow(overrides: Record<string, unknown> = {}) {
     trace_replay_id: 7,
     has_blob: true,
     chart_series: currentSeries(),
+    metric_sources: [],
     hardware: 'gb200',
     framework: 'dynamo-vllm',
     model: 'deepseek-r1-0528',
@@ -97,6 +102,8 @@ describe('getTraceServerMetrics', () => {
     });
     expect(calls).toHaveLength(1);
     expect(calls[0]).not.toContain('server_metrics_json_gz as blob');
+    expect(calls[0]).toContain("atr.chart_series - 'metricSources'");
+    expect(calls[0]).toContain("metric_source->'kvCacheUsageByEngine'");
     expect(calls[0]).toContain("br.metrics ->> 'kv_offload_backend'");
   });
 
@@ -133,3 +140,162 @@ describe('getTraceServerMetrics', () => {
     expect(calls).toHaveLength(1);
   });
 });
+
+describe('getTraceServerMetricSource', () => {
+  it('returns only the selected source from a current precomputed series', async () => {
+    const metricSource = {
+      source: {
+        id: 'decode-0',
+        adapter: 'vllm',
+        role: 'decode' as const,
+        endpointUrl: null,
+        nativeRole: null,
+        workerId: 'worker-0',
+        dpRank: null,
+        engine: null,
+      },
+      kvCacheUsage: [],
+      prefixCacheHitRate: [],
+      queueDepth: [],
+      promptTokensBySource: {},
+      promptTps: [{ t: 0, value: 10 }],
+      generationTps: [{ t: 0, value: 20 }],
+      prefixCacheHitsTps: [],
+      hostKvCacheUsage: [],
+      kvCacheUsageByEngine: [],
+    };
+    const { sql, calls } = mockSql([
+      [
+        {
+          trace_replay_id: 7,
+          has_blob: true,
+          chart_series_version: CHART_SERIES_VERSION,
+          framework: 'dynamo-vllm',
+          disagg: true,
+          metric_source: metricSource,
+        },
+      ],
+    ]);
+
+    await expect(getTraceServerMetricSource(sql, 42, 'decode-0')).resolves.toEqual(metricSource);
+    expect(calls).toHaveLength(1);
+    expect(calls[0]).toContain("metric_source->'source'->>'id'");
+    expect(calls[0]).not.toContain('server_metrics_json_gz as blob');
+  });
+});
+
+// ── All-endpoints KV overlay collapses per worker ───────────────────────
+//
+// `chart_series.kvCacheUsageByEngine` is per DP rank. On a PD-disaggregated
+// run that is unreadable (6 prefill workers x DP4 + decode = 28 lines), so the
+// read layer averages each worker's ranks into one line. Derived here, not in
+// the ETL, so it needs no CHART_SERIES_VERSION bump or backfill.
+
+/** One metric source (worker) owning `ranks` engines, each at a flat value. */
+function workerSource(
+  role: MetricSourceSeries['source']['role'],
+  workerId: string,
+  ranks: number[],
+): MetricSourceSeries {
+  return {
+    source: {
+      id: `dynamo|${role}|${workerId}`,
+      adapter: 'dynamo',
+      role,
+      endpointUrl: `http://10.0.0.1:7500/metrics`,
+      nativeRole: role,
+      workerId,
+      dpRank: null,
+      engine: null,
+    },
+    kvCacheUsage: [],
+    prefixCacheHitRate: [],
+    queueDepth: [],
+    promptTokensBySource: {},
+    promptTps: [],
+    generationTps: [],
+    prefixCacheHitsTps: [],
+    hostKvCacheUsage: [],
+    kvCacheUsageByEngine: ranks.map((value, i) => ({
+      engineLabel: `${role} ${i}`,
+      points: [
+        { t: 0, value },
+        { t: 1, value },
+      ],
+    })),
+  };
+}
+
+describe('getTraceServerMetrics all-endpoints KV overlay', () => {
+  function run(metricSources: unknown[], fallback: unknown[] = []) {
+    const series = {
+      ...currentSeries(),
+      kvCacheUsageByEngine: fallback,
+    } as unknown as ChartSeries;
+    const { sql } = mockSql([[metaRow({ chart_series: series, metric_sources: metricSources })]]);
+    return getTraceServerMetrics(sql, 42);
+  }
+
+  it("averages each worker's ranks into one line", async () => {
+    const result = await run([
+      workerSource('prefill', 'wpre-aaaa', [0.1, 0.3]),
+      workerSource('prefill', 'wpre-bbbb', [0.5, 0.7]),
+      workerSource('decode', 'wdec-cccc', [0.3, 0.5]),
+    ]);
+    expect(result?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual([
+      'prefill aaaa',
+      'prefill bbbb',
+      'decode',
+    ]);
+    expect(result?.kvCacheUsageByEngine.map((e) => e.points[0]!.value)).toEqual([0.2, 0.6, 0.4]);
+  });
+
+  it('keeps the stored per-rank overlay for a single-worker run', async () => {
+    // A plain DP deployment has no per-source data; rank skew is the signal
+    // there, so the stored per-rank lines must survive untouched.
+    const perRank = [
+      { engineLabel: '0', points: [{ t: 0, value: 0.1 }] },
+      { engineLabel: '1', points: [{ t: 0, value: 0.9 }] },
+    ];
+    const result = await run([], perRank);
+    expect(result?.kvCacheUsageByEngine).toEqual(perRank);
+  });
+
+  it('leaves workers that already own one engine each alone', async () => {
+    const perRank = [{ engineLabel: 'prefill 0', points: [{ t: 0, value: 0.2 }] }];
+    const result = await run(
+      [workerSource('prefill', 'wpre-aaaa', [0.2]), workerSource('decode', 'wdec-cccc', [0.4])],
+      perRank,
+    );
+    expect(result?.kvCacheUsageByEngine).toEqual(perRank);
+  });
+
+  it('keeps a single-pool worker when sibling workers own multiple ranks', async () => {
+    const decode = workerSource('decode', 'wdec-cccc', []);
+    decode.kvCacheUsage = [
+      { t: 0, value: 0.4 },
+      { t: 1, value: 0.5 },
+    ];
+    const result = await run([
+      workerSource('prefill', 'wpre-aaaa', [0.1, 0.3]),
+      workerSource('prefill', 'wpre-bbbb', [0.5, 0.7]),
+      decode,
+    ]);
+
+    expect(result?.kvCacheUsageByEngine.map((e) => e.engineLabel)).toEqual([
+      'prefill aaaa',
+      'prefill bbbb',
+      'decode',
+    ]);
+    expect(result?.kvCacheUsageByEngine[2]?.points).toEqual(decode.kvCacheUsage);
+  });
+
+  it('still trims metricSources to descriptors', async () => {
+    const result = await run([
+      workerSource('prefill', 'wpre-aaaa', [0.1, 0.3]),
+      workerSource('decode', 'wdec-cccc', [0.3, 0.5]),
+    ]);
+    expect(result?.metricSources).toHaveLength(2);
+    expect(Object.keys(result!.metricSources[0]!)).toEqual(['source']);
+  });
+});
diff --git a/packages/db/src/queries/trace-server-metrics.ts b/packages/db/src/queries/trace-server-metrics.ts
index 6e8e3d77d..5e3f5f076 100644
--- a/packages/db/src/queries/trace-server-metrics.ts
+++ b/packages/db/src/queries/trace-server-metrics.ts
@@ -22,14 +22,22 @@ import {
 import type { DbClient } from '../connection.js';
 import { writeBackTraceReplayJsonb } from './agentic-shared';
 
-export type { TimeSeriesPoint, QueueDepthPoint } from '../etl/compute-chart-series';
+export type {
+  MetricSourceSeries,
+  QueueDepthPoint,
+  TimeSeriesPoint,
+} from '../etl/compute-chart-series';
 
 // The endpoint payload combines chart_series with separately queried point
 // metadata. Keep a composite response version so metadata-shape changes roll
 // the blob-cache namespace without forcing an expensive chart_series backfill.
-const POINT_META_VERSION = 3;
+const POINT_META_VERSION = 4;
 export const TRACE_SERVER_METRICS_VERSION = CHART_SERIES_VERSION * 100 + POINT_META_VERSION;
 
+export interface MetricSourceDescriptor {
+  source: MetricSourceSeries['source'];
+}
+
 export interface PointMeta {
   id: number;
   hardware: string;
@@ -101,13 +109,20 @@ export interface TraceServerMetrics {
    */
   kvCachePoolTokens: number | null;
   /** Orchestrator-normalized metrics grouped by endpoint/worker. */
-  metricSources: MetricSourceSeries[];
+  metricSources: MetricSourceDescriptor[];
 }
 
+type ChartSeriesSummary = Omit<ChartSeries, 'metricSources'> & {
+  metricSources: MetricSourceDescriptor[];
+};
+
+type KvMetricSource = Pick<MetricSourceSeries, 'source' | 'kvCacheUsage' | 'kvCacheUsageByEngine'>;
+
 interface RawMetaRow extends PointMeta {
   trace_replay_id: number | null;
   has_blob: boolean;
-  chart_series: ChartSeries | null;
+  chart_series: Omit<ChartSeries, 'metricSources'> | null;
+  metric_sources: KvMetricSource[];
   /** Derived at server-log ingest from "GPU KV cache size: N tokens" lines. */
   kv_cache_pool_tokens: string | null;
 }
@@ -148,8 +163,9 @@ function buildMeta(row: RawMetaRow): PointMeta {
 
 function merge(
   meta: PointMeta,
-  series: ChartSeries,
+  series: Omit<ChartSeries, 'metricSources'>,
   kvCachePoolTokens: number | null,
+  metricSources: MetricSourceDescriptor[],
 ): TraceServerMetrics {
   return {
     meta,
@@ -170,7 +186,107 @@ function merge(
     // v8+ field; older chart_series rows lack it → omit per-engine overlay.
     kvCacheUsageByEngine: series.kvCacheUsageByEngine ?? [],
     // v9+ field; old rows are served without a source selector until backfilled.
-    metricSources: series.metricSources ?? [],
+    metricSources,
+  };
+}
+
+/** Mean of several engine series, matched on scrape instant. */
+function meanOfEngines(engines: readonly { points: TimeSeriesPoint[] }[]): TimeSeriesPoint[] {
+  const active = engines.filter((engine) => engine.points.length > 0);
+  if (active.length === 0) return [];
+  if (active.length === 1) return active[0]!.points;
+  // Engines inside one worker are scraped together, so they share instants and
+  // a pointwise mean is exact. A `t` only some of them reported averages those,
+  // rather than counting the absent ones as zero.
+  const byT = new Map<number, { sum: number; n: number }>();
+  for (const engine of active) {
+    for (const point of engine.points) {
+      const at = byT.get(point.t);
+      if (at) {
+        at.sum += point.value;
+        at.n++;
+      } else {
+        byT.set(point.t, { sum: point.value, n: 1 });
+      }
+    }
+  }
+  return [...byT.entries()]
+    .toSorted((a, b) => a[0] - b[0])
+    .map(([t, { sum, n }]) => ({ t, value: sum / n }));
+}
+
+/**
+ * All-endpoints KV overlay: one line per WORKER, averaged over the ranks it
+ * owns.
+ *
+ * `chart_series.kvCacheUsageByEngine` is per rank, which stops being readable
+ * on a PD-disaggregated run — 6 prefill workers at DP4 plus a decode worker is
+ * 28 lines, and ranks inside one worker track each other closely, so the
+ * comparison worth seeing is worker-vs-worker.
+ *
+ * Derived here rather than in the ETL because everything it needs is already
+ * stored: each source carries its own per-rank breakdown, and this layer is
+ * the last place that sees it before `metricSources` is trimmed to
+ * descriptors. Keeping it out of `chart_series` means no version bump and no
+ * backfill for what is a presentation choice.
+ *
+ * Returns null when collapsing would not help — a single worker, or a run with
+ * no per-source data, keeps the per-rank overlay, which is exactly where rank
+ * skew is the signal.
+ */
+function collapseKvByWorker(
+  sources: readonly KvMetricSource[],
+): { engineLabel: string; points: TimeSeriesPoint[] }[] | null {
+  if (sources.length < 2) return null;
+  const entries = sources
+    .map((entry) => ({
+      source: entry.source,
+      // Single-pool workers intentionally omit the per-engine array because
+      // it would duplicate their aggregate line. Preserve that worker when a
+      // sibling with multiple ranks makes the all-endpoints collapse useful.
+      engines:
+        entry.kvCacheUsageByEngine?.length > 0
+          ? entry.kvCacheUsageByEngine
+          : entry.kvCacheUsage.length > 0
+            ? [{ points: entry.kvCacheUsage }]
+            : [],
+    }))
+    .filter((entry) => entry.engines.length > 0);
+  if (entries.length < 2) return null;
+  // Already one engine per worker — nothing to average, so leave the stored
+  // labels alone rather than relabelling for no gain.
+  const totalEngines = entries.reduce((sum, entry) => sum + entry.engines.length, 0);
+  if (totalEngines <= entries.length) return null;
+
+  const workersPerRole = new Map<string, number>();
+  for (const { source } of entries) {
+    workersPerRole.set(source.role, (workersPerRole.get(source.role) ?? 0) + 1);
+  }
+  const used = new Set<string>();
+  return entries.map(({ source, engines }, index) => {
+    const worker = source.workerId;
+    const short = worker && worker.length > 4 ? worker.slice(-4) : worker;
+    // Qualify by worker only when the role actually has several, so a lone
+    // decode worker reads "decode" rather than "decode 0842".
+    const base =
+      (workersPerRole.get(source.role) ?? 0) > 1 && short
+        ? `${source.role} ${short}`
+        : (source.role ?? short ?? `#${index}`);
+    let label = base;
+    for (let n = 2; used.has(label); n++) label = `${base} #${n}`;
+    used.add(label);
+    return { engineLabel: label, points: meanOfEngines(engines) };
+  });
+}
+
+function summarizeSeries(
+  series: Omit<ChartSeries, 'metricSources'>,
+  sources: readonly KvMetricSource[],
+): ChartSeriesSummary {
+  return {
+    ...series,
+    kvCacheUsageByEngine: collapseKvByWorker(sources) ?? series.kvCacheUsageByEngine ?? [],
+    metricSources: sources.map(({ source }) => ({ source })),
   };
 }
 
@@ -182,7 +298,23 @@ export async function getTraceServerMetrics(
     select
       br.trace_replay_id,
       (atr.server_metrics_json_gz is not null) as has_blob,
-      atr.chart_series,
+      case
+        when atr.chart_series is null then null
+        else atr.chart_series - 'metricSources'
+      end as chart_series,
+      coalesce((
+        select jsonb_agg(jsonb_build_object(
+          'source', metric_source->'source',
+          'kvCacheUsage', coalesce(metric_source->'kvCacheUsage', '[]'::jsonb),
+          'kvCacheUsageByEngine', coalesce(
+            metric_source->'kvCacheUsageByEngine',
+            '[]'::jsonb
+          )
+        ))
+        from jsonb_array_elements(
+          coalesce(atr.chart_series->'metricSources', '[]'::jsonb)
+        ) as metric_source
+      ), '[]'::jsonb) as metric_sources,
       br.id, c.hardware, c.framework, c.model, c.precision, c.spec_method,
       c.disagg, c.is_multinode,
       br.conc, br.offload_mode, br.isl, br.osl, br.benchmark_type,
@@ -212,7 +344,8 @@ export async function getTraceServerMetrics(
 
   // Fast path: pre-computed chart_series at the current version.
   if (row.chart_series && Number(row.chart_series.version) === CHART_SERIES_VERSION) {
-    return merge(meta, row.chart_series, kvCachePoolTokens);
+    const summary = summarizeSeries(row.chart_series, row.metric_sources);
+    return merge(meta, summary, kvCachePoolTokens, summary.metricSources);
   }
 
   // Slow path only: fetch the large raw blob after establishing that the
@@ -241,5 +374,60 @@ export async function getTraceServerMetrics(
   // (no-ops on a read-only replica). trace_replay_id is non-null on this path.
   writeBackTraceReplayJsonb(sql, 'chart_series', row.trace_replay_id, series);
 
-  return merge(meta, series, kvCachePoolTokens);
+  const summary = summarizeSeries(series, series.metricSources);
+  return merge(meta, summary, kvCachePoolTokens, summary.metricSources);
+}
+
+interface RawMetricSourceRow {
+  trace_replay_id: number | null;
+  has_blob: boolean;
+  chart_series_version: number | null;
+  framework: string;
+  disagg: boolean;
+  metric_source: MetricSourceSeries | null;
+}
+
+/** Fetch one source's full series only after the user selects it. */
+export async function getTraceServerMetricSource(
+  sql: DbClient,
+  benchmarkResultId: number,
+  sourceId: string,
+): Promise<MetricSourceSeries | null> {
+  const rows = (await sql`
+    select
+      br.trace_replay_id,
+      (atr.server_metrics_json_gz is not null) as has_blob,
+      (atr.chart_series->>'version')::int as chart_series_version,
+      c.framework,
+      c.disagg,
+      (
+        select metric_source
+        from jsonb_array_elements(atr.chart_series->'metricSources') as metric_source
+        where metric_source->'source'->>'id' = ${sourceId}
+        limit 1
+      ) as metric_source
+    from benchmark_results br
+    join configs c on c.id = br.config_id
+    left join agentic_trace_replay atr on atr.id = br.trace_replay_id
+    where br.id = ${benchmarkResultId}
+  `) as unknown as RawMetricSourceRow[];
+  const row = rows[0];
+  if (!row || !row.has_blob || row.trace_replay_id === null) return null;
+
+  if (row.chart_series_version === CHART_SERIES_VERSION) return row.metric_source;
+
+  const blobRows = (await sql`
+    select server_metrics_json_gz as blob
+    from agentic_trace_replay
+    where id = ${row.trace_replay_id}
+  `) as unknown as RawBlobRow[];
+  const blob = blobRows[0]?.blob;
+  if (!blob) return null;
+  const series = await computeChartSeries(blob, {
+    framework: row.framework,
+    disagg: row.disagg,
+  });
+  if (!series) return null;
+  writeBackTraceReplayJsonb(sql, 'chart_series', row.trace_replay_id, series);
+  return series.metricSources.find(({ source }) => source.id === sourceId) ?? null;
 }
diff --git a/packages/db/src/queries/workflow-info.ts b/packages/db/src/queries/workflow-info.ts
index e32401395..d50171d2b 100644
--- a/packages/db/src/queries/workflow-info.ts
+++ b/packages/db/src/queries/workflow-info.ts
@@ -18,6 +18,7 @@ export interface ChangelogRow {
   config_keys: string[];
   description: string;
   pr_link: string | null;
+  append_only: boolean;
 }
 
 export interface DateConfigRow {
@@ -56,7 +57,8 @@ export async function getChangelogByDate(sql: DbClient, date: string): Promise<C
       cl.head_ref,
       cl.config_keys,
       cl.description,
-      cl.pr_link
+      cl.pr_link,
+      cl.append_only
     FROM changelog_entries cl
     JOIN latest_workflow_runs wr ON wr.id = cl.workflow_run_id
     WHERE cl.date = ${date}::date
diff --git a/packages/db/tsconfig.json b/packages/db/tsconfig.json
index 09565c30f..9b2456e03 100644
--- a/packages/db/tsconfig.json
+++ b/packages/db/tsconfig.json
@@ -1,14 +1,7 @@
 {
+  "extends": "../../tsconfig.base.json",
   "compilerOptions": {
-    "target": "ES2022",
-    "module": "esnext",
-    "moduleResolution": "bundler",
-    "strict": true,
-    "noEmit": true,
-    "esModuleInterop": true,
-    "skipLibCheck": true,
-    "resolveJsonModule": true,
-    "isolatedModules": true
+    "target": "ES2022"
   },
   "include": ["src/**/*.ts"],
   "exclude": ["node_modules"]
diff --git a/packages/mcp/README.md b/packages/mcp/README.md
index 31a1f9e32..a583099ad 100644
--- a/packages/mcp/README.md
+++ b/packages/mcp/README.md
@@ -8,6 +8,10 @@ Local MCP (Model Context Protocol) server for querying the InferenceX benchmark
 bun run mcp
 ```
 
+The server requires `DATABASE_READONLY_URL` and deliberately uses the postgres.js TCP
+driver. URL and TLS handling come from the shared database resolver. `DATABASE_SSL=false`
+overrides TLS for any host; loopback hosts disable TLS by default.
+
 Or add to Claude Code:
 
 ```bash
diff --git a/packages/mcp/package.json b/packages/mcp/package.json
index 5fea5f373..664f95a36 100644
--- a/packages/mcp/package.json
+++ b/packages/mcp/package.json
@@ -13,14 +13,14 @@
     "typecheck": "tsc --noEmit"
   },
   "dependencies": {
-    "@modelcontextprotocol/sdk": "^1.29.0",
+    "@modelcontextprotocol/sdk": "^1.30.0",
     "@semianalysisai/inferencex-constants": "workspace:*",
     "@semianalysisai/inferencex-db": "workspace:*",
     "postgres": "^3.4.9",
     "zod": "^4.4.3"
   },
   "devDependencies": {
-    "@types/node": "^26.1.1",
+    "@types/node": "^26.2.0",
     "typescript": "^6.0.3",
     "vitest": "^4.1.10"
   }
diff --git a/packages/mcp/src/server.test.ts b/packages/mcp/src/server.test.ts
index 6ef72bf01..ff5a835e6 100644
--- a/packages/mcp/src/server.test.ts
+++ b/packages/mcp/src/server.test.ts
@@ -27,8 +27,11 @@ const { mockState, mockDb } = vi.hoisted(() => {
 });
 
 vi.mock('@semianalysisai/inferencex-db/connection', () => ({
-  FIXTURES_MODE: false,
-  postgresOptionsForUrl: () => ({ max: 5, ssl: false }),
+  resolveDatabaseConnection: () => ({
+    url: 'postgres://test@localhost/test',
+    driver: 'postgres',
+    ssl: false,
+  }),
 }));
 
 vi.mock('postgres', () => ({
diff --git a/packages/mcp/src/server.ts b/packages/mcp/src/server.ts
index 363f0c8b5..58269ba03 100644
--- a/packages/mcp/src/server.ts
+++ b/packages/mcp/src/server.ts
@@ -6,12 +6,16 @@ import {
   PRECISION_KEYS,
   SPEC_METHOD_KEYS,
 } from '@semianalysisai/inferencex-constants';
-import { postgresOptionsForUrl } from '@semianalysisai/inferencex-db/connection';
+import { resolveDatabaseConnection } from '@semianalysisai/inferencex-db/connection';
 import postgres from 'postgres';
 import { z } from 'zod';
 
-const url = process.env.DATABASE_READONLY_URL!;
-const db = postgres(url, postgresOptionsForUrl(url));
+const connection = resolveDatabaseConnection({
+  envVar: 'DATABASE_READONLY_URL',
+  driver: 'postgres',
+  ssl: process.env.DATABASE_SSL,
+});
+const db = postgres(connection.url, { max: 5, ssl: connection.ssl });
 const MAX_ROWS = 5_000;
 
 const roundMetric = (v: unknown) => (typeof v === 'number' ? Math.round(v * 10000) / 10000 : v);
@@ -44,36 +48,44 @@ const modelMapping = Object.entries(DB_MODEL_TO_DISPLAY)
  */
 const SERVER_INSTRUCTIONS = `InferenceX: ML inference benchmark database. Query GPU performance across hardware and frameworks.
 Models: ${modelMapping}.
-Key tool: get_latest_benchmarks — filters by hardware, model, framework, precision, spec_method, disagg, num_gpu, isl, osl, conc. Returns config details (incl. num_prefill_gpu, num_decode_gpu) and metrics JSONB with keys: median_ttft, p99_ttft, median_tpot, p99_tpot, tput_per_gpu, output_tput_per_gpu, median_itl, median_e2el (all in seconds; throughput in tok/s/GPU).
+Key tool: get_latest_benchmarks filters by hardware, model, framework, precision, spec_method, disagg, num_gpu, isl, osl, and conc. It returns config details (including num_prefill_gpu and num_decode_gpu) plus selected metric keys. Pass metrics=["all"] for the complete metrics JSONB object. Common keys include median_ttft, p99_ttft, median_tpot, p99_tpot, tput_per_gpu, output_tput_per_gpu, median_itl, and median_e2el.
 For aggregations or custom queries use query_sql against the latest_benchmarks view joined to configs.`;
 
 /**
  * Full overview returned by get_overview tool — no length constraint.
  */
-const DOMAIN_OVERVIEW = `InferenceX benchmark database — ML inference performance data across GPU hardware and serving frameworks.
+const DOMAIN_OVERVIEW = `InferenceX benchmark database. It contains ML inference performance data across GPU hardware and serving frameworks.
 
 ## Tables
-- **configs** — Serving configs: (hardware, framework, model, precision, spec_method, disagg) + parallelism (TP/EP/DP per prefill/decode).
-- **benchmark_results** — Perf metrics per config/concurrency/sequence-length/date. \`metrics\` JSONB holds all numbers.
-- **availability** — Denormalized date×config availability.
-- **eval_results** — Eval accuracy (e.g. gsm8k). Joined to configs via config_id.
-- **workflow_runs** — GitHub Actions run metadata.
-- **run_stats** — Per-hardware reliability (n_success/total).
+- **configs** stores serving identity and prefill/decode topology.
+- **benchmark_results** stores one scenario point per workflow, config, recipe, sequence, offload mode, and concurrency. Measurements stay in \`metrics\` JSONB. Per-worker power data stays in \`workers\` JSONB.
+- **workflow_runs**, **changelog_entries**, and **run_stats** store GitHub run metadata, reviewed changelog metadata, and hardware reliability counts.
+- **availability** is the denormalized model/config/scenario/date index used by selectors.
+- **eval_results** and **eval_samples** store aggregate and per-prompt evaluation results.
+- **agentic_trace_replay** stores compressed trace exports and precomputed aggregate, chart, and request-timeline JSONB.
+- **datasets**, **dataset_conversations**, and **run_datasets** store agentic dataset metadata, conversation structures, and workflow-to-dataset links.
+- **server_logs** stores benchmark server logs. **user_feedback** stores encrypted feedback fields.
 
 ## Key Views
-- **latest_benchmarks** (materialized) — Latest successful benchmark per (config, conc, isl, osl). Use this for current data.
+- **latest_workflow_runs** keeps the highest attempt for each GitHub run.
+- **latest_benchmarks** is a materialized view of the newest successful logical curve for each config, benchmark type, sequence, offload mode, recipe fingerprint, and concurrency. It can carry points from same-image append-only history. Producer identity remains in \`workflow_run_id\`; logical snapshot identity is in \`snapshot_workflow_run_id\`.
 
 ## Column Names
 - **configs**: id, hardware, framework, model, precision, spec_method, disagg, is_multinode, prefill_tp, prefill_ep, prefill_dp_attention, prefill_num_workers, decode_tp, decode_ep, decode_dp_attention, decode_num_workers, num_prefill_gpu, num_decode_gpu
-- **benchmark_results**: id, workflow_run_id (FK), config_id (FK), benchmark_type, date, isl, osl, conc, image, metrics (JSONB), error, server_log_id (FK)
-- **latest_benchmarks** (materialized view): config_id, date, isl, osl, conc, image, metrics (JSONB) — latest per (config, conc, isl, osl) where error IS NULL
-- **latest_workflow_runs** (view): id, github_run_id, run_attempt, name, status, conclusion, head_sha, head_branch, html_url, created_at, run_started_at, date
-- **workflow_runs**: id, github_run_id, run_attempt, name, status, conclusion, head_sha, head_branch, html_url, created_at, run_started_at, date
-- **eval_results**: id, workflow_run_id (FK), config_id (FK), task, date, isl, osl, conc, lm_eval_version, metrics (JSONB)
+- **benchmark_results**: id, workflow_run_id (FK), config_id (FK), benchmark_type, date, isl, osl, conc, image, metrics (JSONB), error, server_log_id (FK), workers (JSONB), offload_mode, trace_replay_id (FK), recipe_fingerprint
+- **latest_benchmarks**: all benchmark_results columns plus snapshot_date and snapshot_workflow_run_id
+- **workflow_runs** and **latest_workflow_runs**: id, github_run_id, run_attempt, name, status, conclusion, head_sha, head_branch, html_url, created_at, run_started_at, date, append_only
+- **changelog_entries**: id, workflow_run_id (FK), date, base_ref, head_ref, config_keys (text[]), description, pr_link, append_only
 - **run_stats**: id, workflow_run_id (FK), date, hardware, n_success, total
-- **availability**: model, isl, osl, precision, hardware, framework, spec_method, disagg, date (PK is all columns)
-- **changelog_entries**: id, workflow_run_id (FK), date, base_ref, head_ref, config_keys (text[]), description, pr_link
-- **server_logs**: id, server_log (text)
+- **availability**: model, isl, osl, precision, hardware, framework, spec_method, disagg, date, benchmark_type (unique natural key, NULLS NOT DISTINCT)
+- **eval_results**: id, workflow_run_id (FK), config_id (FK), task, date, isl, osl, conc, lm_eval_version, metrics (JSONB)
+- **eval_samples**: id, eval_result_id (FK), doc_id, prompt, target, response, passed, score, metrics (JSONB), data (JSONB)
+- **agentic_trace_replay**: id, profile_export_jsonl_gz, profile_export_uncompressed_size, server_metrics_csv, server_metrics_csv_size, created_at, server_metrics_json_gz, server_metrics_json_uncompressed_size, aggregate_stats (JSONB), chart_series (JSONB), request_timeline (JSONB)
+- **datasets**: id, slug, label, variant, description, hf_url, license, conversation_count, summary (JSONB), chart_data (JSONB), dataset_version, ingested_at
+- **dataset_conversations**: id, dataset_id (FK), conv_id, models (text[]), num_turns, num_subagent_groups, total_in, total_out, total_cached, structure (JSONB)
+- **run_datasets**: workflow_run_id (FK), dataset_slug, created_at
+- **server_logs**: id, server_log
+- **user_feedback**: id, created_at, doing_well_ciphertext, doing_poorly_ciphertext, want_to_see_ciphertext, user_agent_ciphertext, page_path_ciphertext
 
 ## Enum Values
 - **hardware**: ${HW_ENUM.join(', ')}
@@ -82,13 +94,13 @@ const DOMAIN_OVERVIEW = `InferenceX benchmark database — ML inference performa
 - **precision**: ${PREC_ENUM.join(', ')}
 - **spec_method**: ${SPEC_ENUM.join(', ')}
 
-## Metrics JSONB Keys (seconds; throughput in tok/s/GPU)
-- **Throughput**: tput_per_gpu, output_tput_per_gpu, input_tput_per_gpu
-- **TTFT**: median_ttft, mean_ttft, p99_ttft, std_ttft
-- **TPOT**: median_tpot, mean_tpot, p99_tpot, std_tpot
-- **ITL**: median_itl, mean_itl, p99_itl, std_itl
-- **E2EL**: median_e2el, mean_e2el, p99_e2el, std_e2el
-- **Interactivity**: median_intvty, mean_intvty, p99_intvty, std_intvty
+## Metrics JSONB
+Metrics are artifact-dependent and remain in JSONB. There are no metric-specific columns on benchmark_results or latest_benchmarks.
+- **Throughput**: tput_per_gpu, output_tput_per_gpu, input_tput_per_gpu (tokens/s/GPU); total_tput_tps, output_tput_tps, input_tput_tps (tokens/s)
+- **Latency distributions in seconds**: *_ttft, *_tpot, *_itl, *_e2el, and *_full_response_itl. Available prefixes can include median, mean, std, p75, p90, p95, p99, and p99.9.
+- **Interactivity in tokens/s/user**: *_intvty and *_full_response_intvty
+- **Agentic scalars** can include duration_seconds, *_qps, token-count distributions, cache-hit rates, KV-cache usage, and total request/token counters.
+- Some agentic rows also retain categorical metadata such as offload_mode and cache backend in metrics. Do not assume every JSONB value is numeric or every row contains every key.
 
 ## Common SQL
 \`\`\`sql
@@ -196,7 +208,7 @@ export function createServer(): McpServer {
     {
       title: 'Get Latest Benchmarks',
       description:
-        'Get latest benchmark results with config details and metrics JSONB. This is the primary query tool — use it before falling back to query_sql. All filters are optional; combine any subset. Use sort_by with limit to get top-N results by a metric.',
+        'Get latest benchmark results with config details and selected metric keys. This is the primary query tool to use before query_sql. All filters are optional and can be combined. Pass metrics=["all"] for the complete metrics JSONB object. Use sort_by with limit for top-N results.',
       inputSchema: {
         hardware: z.enum(HW_ENUM).optional().describe('GPU type'),
         model: z.enum(MODEL_ENUM).optional().describe('Model key'),
diff --git a/packages/mcp/tsconfig.json b/packages/mcp/tsconfig.json
index fa2bb5ebb..f699cf349 100644
--- a/packages/mcp/tsconfig.json
+++ b/packages/mcp/tsconfig.json
@@ -1,6 +1,7 @@
 {
-  "extends": "../../tsconfig.json",
+  "extends": "../../tsconfig.base.json",
   "compilerOptions": {
+    "target": "ES2018",
     "rootDir": ".",
     "outDir": "dist",
     "types": ["node"]
diff --git a/tsconfig.base.json b/tsconfig.base.json
new file mode 100644
index 000000000..a2a832dc6
--- /dev/null
+++ b/tsconfig.base.json
@@ -0,0 +1,13 @@
+{
+  "compilerOptions": {
+    "esModuleInterop": true,
+    "isolatedModules": true,
+    "lib": ["ESNext"],
+    "module": "esnext",
+    "moduleResolution": "bundler",
+    "noEmit": true,
+    "resolveJsonModule": true,
+    "skipLibCheck": true,
+    "strict": true
+  }
+}
diff --git a/tsconfig.json b/tsconfig.json
index 3bd6f54d2..e2d32a11d 100644
--- a/tsconfig.json
+++ b/tsconfig.json
@@ -1,16 +1,9 @@
 {
+  "extends": "./tsconfig.base.json",
   "compilerOptions": {
     "target": "ES2018",
     "lib": ["dom", "dom.iterable", "esnext"],
     "allowJs": true,
-    "skipLibCheck": true,
-    "strict": true,
-    "noEmit": true,
-    "esModuleInterop": true,
-    "module": "esnext",
-    "moduleResolution": "bundler",
-    "resolveJsonModule": true,
-    "isolatedModules": true,
     "jsx": "react-jsx",
     "paths": {
       "@/*": ["./packages/app/src/*"]